Datasets:

llmunlearn

/

unlearn_dataset

like 1

Follow

Unlearning of LLMs 6

Apoptosis is widely known as a basic biological event. It is a strictly programmed process for removing superfluous, aged, or damaged cells. Apoptosis is characterized by cell shrinkage, membrane blebbing, and nuclear condensation and fragmentation. During embryonic development, apoptosis is an essential event required for successful organogenesis. In adult tissues, apoptosis counterbalances cell proliferation to maintain cell numbers and tissue homeostasis[@b1][@b2][@b3]. Dysregulation of apoptosis can result in abnormal cell growth and the relevant diseases. A large number of stimuli, both natural and non-natural, have been found to regulate apoptosis. For instance, TGF-β1, TNF-α, ROS, p53, and various chemotherapy drugs[@b4][@b5][@b6][@b7] were found to be potent inducers of apoptosis. TGF-β1 is a pleiotropic cytokine, which is implicated in the control of diverse cellular processes, such as cell cycle state, cell differentiation, chemotaxis and migration[@b8]. In physiological processes, TGF-β1 contributes to cellular homeostasis, embryonic development, immune response, wound healing, and angiogenesis[@b9]. However, in pathological processes, TGF-β1 signaling has been linked with progression of diseases, such as organ fibrosis and cancers[@b10][@b11]. It is known that TGF-β1 can function as pro-tumor factor by increasing cell motility, invasion, and angiogenesis. The role of TGF-β1 depends on cell type, cellular context, cell state, and the cell environment. Although TGF-β1 is a multifunctional molecule, a singular response to TGF-β1 treatment was found in most cell types. However, TGF-β1 can simultaneously induce both apoptosis and EMT, two fundamentally different cell fates, in certain cells. For example, mouse hepatocytes[@b12] and mammary gland epithelial cells[@b13] respond simultaneously to TGF-β1 treatment with apoptosis and EMT. In other words, in a population of the exact same type of cells, some undergo apoptosis and some undergo EMT in response to TGF-β1. It has been shown previously that TGF-β1-induced apoptosis and EMT were cell cycle dependent events[@b14]. Cells synchronized at G1/S phase primarily undergo EMT, while cells synchronized at G2/M phase undergo apoptosis. Strikingly, TGF-β1-mediated growth arrest has been found to be closely associated with its induction of EMT in non-cancer cells. Although TGF-β1 can induce both apoptosis and EMT in certain non-cancer cells, most types of cancer cells of epithelial origin escaped the apoptotic induction effect of TGF-β1. Because of the high heterogeneity of cancer cells, whether TGF-β1 is able to induce EMT in a type of cancer cell depends on the cellular context and the cell plasticity. Further exploration of the molecules implicated in inhibition of TGF-β1-induced apoptosis in normal cells is important for achieving a better understanding of the changes that convert cells from an apoptosis-inducible to an apoptosis-resistant state. Interactions between Hippo and TGF-β signaling pathways has been shown in recent years[@b15][@b16][@b17][@b18]. YAP is a key component of the Hippo pathway, which functions as a transcriptional co-activator. YAP has been shown to play important roles in regulating proliferation, survival, self-renewal and organ size[@b19]. YAP and other core Hippo pathway components were initially identified in Drosophila and are highly conserved in mammals[@b20]. The core kinase cassette in the Hippo pathway consists of a series of serine/threonine kinases of MST and LATS families[@b21]. When the Hippo pathway is on, activated MSTs/LATSs induce the phosphorylation of serine 127 in YAP, which promotes 14-3-3 binding and leads to cytoplasmic retention of YAP[@b22][@b23]. In the absence of activated Hippo, YAP translocates into cell nucleus and promotes downstream transcription networks by interacting with the TEAD family of transcription factors[@b24][@b25]. In mammalian species, functional studies of YAP have been primarily performed in tumor cells. YAP has been shown to play a role in promoting tumorigenesis, cell proliferation, resistance of apoptosis, anchorage-independent growth *in vitro*, migration, and metastasis of tumor cells[@b26][@b27]. Overexpression of YAP inhibits apoptosis induced by cisplatin, paclitaxel, UV irradiation, and ionizing irradiation[@b28][@b29][@b30][@b31]. Increased YAP expression or its nuclear localization has been observed in a variety of human cancers[@b27]. However, there are also reports showing that YAP functions as a pro-apoptotic effector in mammalian cells. YAP overexpression increases p73-mediated apoptosis upon DNA damage[@b32][@b33]. The pro-apoptotic role of YAP is thought to be dependent on different cellular contexts and related downstream molecules[@b34]. Moreover, compared to non-natural inducers in cancer cells, little is known about the effect of YAP on apoptosis induced by physiological stimuli, particularly in normal cells. In addition, in apoptosis and EMT simultaneously induced by TGF-β1, the effect of YAP on these two fundamentally different cell fates has not been investigated. In this study, we examined whether YAP has a regulatory effect on TGF-β1-induced apoptosis and EMT. We found that YAP overexpression can inhibit TGF-β1-induced apoptosis. Although overexpression of YAP alone cannot induce EMT directly, it can increase the number of cells undergoing EMT. YAP knockdown induced apoptosis and dramatically increased the apoptotic response of cells to TGF-β1 treatment, which significantly reduced the portion of survival cells that underwent EMT. In addition, we determined that increased EGFR levels and its activation are involved in YAP-mediated suppression of TGF-β1-induced apoptosis. The data presented suggest that the YAP expression level is implicated in modulating the cellular fates of apoptosis and EMT in response to TGF-β1 stimulation. Results ======= Increases in YAP expression in TGFβ1-induced apoptosis and EMT -------------------------------------------------------------- We first examined the relationship between YAP expression level and TGF-β1 signaling and its association with TGF-β1-mediated apoptosis and EMT in normal mouse mammary epithelial cells (NMuMG). Compared to cancer cell lines that escape the apoptosis-inducing effect of TGF-β1, NMuMG cells simultaneously undergo both apoptosis and EMT, which can be directly observed by cell morphology ([Fig. 1a](#f1){ref-type="fig"}). EMT and apoptosis were also determined by changes in epithelial and mesenchymal marker protein levels and by Annexin V/PI double staining, respectively. As shown, both EMT ([Fig. 1b](#f1){ref-type="fig"}) and apoptosis ([Fig. 1c](#f1){ref-type="fig"}) exhibited time-dependent effects. Additionally, a dose-dependent effect on apoptosis was also observed ([Supplementary Fig. S1](#S1){ref-type="supplementary-material"}). Interestingly, a moderate time-dependent increase in the YAP protein level was found to be associated with TGF-β1-induced apoptosis and EMT ([Fig. 1d](#f1){ref-type="fig"}). Increased YAP expression can clearly be induced by TGF-β1 treatment at a concentration of 1 ng/mL, but no further obvious increases were observed at higher concentrations ([Fig. 1e](#f1){ref-type="fig"}). YAP is a transcriptional cofactor, and it functions primarily in the cell nucleus. Consistently, an increased YAP protein level was found in isolated nuclear fractions after TGF-β1 treatment ([Fig. 1f](#f1){ref-type="fig"}). These results indicate that upregulation of YAP is associated with TGF-β1-induced cell apoptosis and EMT. YAP overexpression inhibits TGF-β1-induced apoptosis but not EMT ---------------------------------------------------------------- As YAP was upregulated during TGF-β1-induced concomitant apoptosis and EMT, we explored the potential contribution of YAP to these two fundamentally different cell fates induced by TGF-β1. NMuMG cells overexpressing HA-tagged YAP and constitutively active YAP^S127A^ were constructed ([Fig. 2a](#f2){ref-type="fig"}). As shown in [Fig. 2b](#f2){ref-type="fig"}, overexpression of YAP was also associated with an increased nuclear localization. Strikingly, overexpression of either wild type or constitutively active YAP suppressed TGF-β1-induced cell death ([Fig. 2c](#f2){ref-type="fig"}). Notably, YAP overexpression had no effect on morphological changes. The inhibitory effect of YAP on apoptosis was confirmed by Annexin V/PI double staining assay, which showed that forced expression of YAP remarkably reduced the number of both early and late apoptotic cells ([Fig. 2d](#f2){ref-type="fig"}). YAP-mediated suppression of apoptosis was also shown by inhibition of TGF-β1-induced caspase-3 cleavage ([Fig. 2e](#f2){ref-type="fig"}). In addition, in stable cell lines expressing two different levels of YAP, we observed that cell survival increased as the expression level of YAP increased ([Supplementary Fig. S2a,b](#S1){ref-type="supplementary-material"}). Consistent with the cell morphology results, overexpression of YAP cannot induce EMT and has no effect on TGF-β1-induced EMT, as evidenced by the protein levels of EMT markers fibronectin and E-cadherin, although a relatively weak increase in N-cadherin was observed ([Supplementary Fig. S2c](#S1){ref-type="supplementary-material"}). Moreover, YAP overexpression has no effect on E-cadherin level and F-actin arrangement induced by TGF-β1, as detected by immunofluorescent assays ([Supplementary Fig. S2d and e](#S1){ref-type="supplementary-material"}). YAP overexpression also has no effect on the expression levels of Snail and PAI-1 genes ([Supplementary Fig. S2f](#S1){ref-type="supplementary-material"}). The above results indicate that YAP has an anti-apoptotic function and its overexpression can substantially suppress TGF-β1-induced cell death. Although YAP has no effect on EMT directly, the number of surviving cells that undergo EMT can be enhanced as the expression level of YAP increases, strongly implying that the expression level of YAP may be involved in controlling the extent of either apoptosis or EMT in response to TGF-β1. Knockdown of YAP induces apoptosis ---------------------------------- To further verify the role of YAP in the determination of cell fate of apoptosis versus EMT, we knocked down YAP by shRNA in NMuMG cells and then determined its possible effect. As examined by an Annexin V/PI double staining assay, YAP knockdown not only induced fundamental cell apoptosis at basal levels ([Fig. 3a](#f3){ref-type="fig"} upper panel) but also enhanced TGF-β1-induced apoptotic cell death ([Fig. 3a](#f3){ref-type="fig"} lower panel). The pro-apoptotic effect of YAP knockdown can be detected by the increased immunoblotting band of cleaved caspase-3 ([Fig. 3b](#f3){ref-type="fig"}). These results further indicate that YAP is a critical survival factor for cells. Although YAP downregulation has no direct effect on TGF-β1-induced EMT, as shown by cell morphology ([Fig. 3c](#f3){ref-type="fig"}) and EMT marker protein levels ([Fig. 3d](#f3){ref-type="fig"}), it greatly reduced the numbers of surviving cells that could undergo EMT in response to TGF-β1. The FACS data analysis ([Fig. 3e](#f3){ref-type="fig"}) showed that, in YAP-knockdown cells, TGF-β1 treatment reduced the percentage of viable cells (Annexin V/PI double negative) from approximately 70% to approximately 20%, while the viability of control cells went from 93% to 70%. The data suggest that, by directly inducing different degrees of apoptosis and thus indirectly influencing the percentage of viable cells which undergo EMT, YAP expression levels can exert a modulatory effect on the balance between cell apoptosis and EMT. In other words, significantly altered expression levels of YAP can break the balance of apoptosis and EMT in response to TGF-β1 stimulation. MCF10A is a human normal breast epithelial cell line that exhibits non-tumorigenic properties. Unlike NMuMG cells, MCF10A cells only underwent EMT ([Fig. 4a](#f4){ref-type="fig"}), not apoptosis ([Fig. 4b](#f4){ref-type="fig"}), in response to TGF-β1 stimulus. Similar to NMuMG cells, TGF-β1 treatment also slightly but obviously increased the expression level of YAP MCF10A cells, as shown in time- and dose-effect assays ([Fig. 4c](#f4){ref-type="fig"} and [d](#f4){ref-type="fig"}). To confirm that YAP is also important for cell survival in non-apoptotic cells upon TGF-β1 treatment, we knocked down YAP and determined its effect on cell viability. As shown in ([Fig. 4e](#f4){ref-type="fig"}), YAP knockdown in MCF10A cells markedly induced spontaneous apoptosis. Interestingly, YAP knockdown also had no direct effect on TGF-β1-induced EMT in MCF10A cells ([Fig. 4f](#f4){ref-type="fig"}), indicating that YAP is not a direct contributor to TGF-β1-induced EMT. YAP is critical for cell survival and can protect cells from TGF-β-induced cell death. Overexpression of YAP increases EGF receptor expression ------------------------------------------------------- YAP and EGFR play similar important roles in cell survival and apoptosis-resistance. To understand the mechanism underlying the YAP-mediated suppression of TGF-β1-induced apoptosis, we examined whether the effect of YAP is associated with the EGFR level and its activation state. Indeed, overexpression of YAP increased the protein level and phosphorylated form of EGFR ([Fig. 5a](#f5){ref-type="fig"}). An increased EGFR mRNA level was also observed in YAP-overexpressing cells ([Fig. 5b](#f5){ref-type="fig"}). In contrast, knockdown of YAP inhibited EGFR expression ([Fig. 5c](#f5){ref-type="fig"}). Since YAP is upregulated upon TGF-β1 treatment, we examined the effect of TGF-β1 on EGFR expression. Interestingly, TGF-β1 treatment also induced an increase in EGFR expression, which was time-dependent ([Fig. 5d](#f5){ref-type="fig"}). Moreover, knock down of YAP abolished the TGF-β1-induced increase in EGFR ([Fig. 5e](#f5){ref-type="fig"}). The data suggest that EGFR is a downstream molecule of YAP. EGFR is required for YAP-mediated anti-apoptotic function --------------------------------------------------------- EGFR is an important survival factor in cells, and its knockdown in NMuMG cells caused marked apoptosis, as shown by Annexin V/PI double staining ([Fig. 6a](#f6){ref-type="fig"}) and increased the level of cleaved caspase-3 ([Fig. 6b](#f6){ref-type="fig"}). To investigate whether EGFR is involved in the YAP-mediated effect, cells were treated with TGF-β1 in the presence of EGF. The addition of exogenous EGF suppressed TGF-β1-induced apoptosis ([Fig. 6c](#f6){ref-type="fig"} and [Supplementary Fig. S3](#S1){ref-type="supplementary-material"}). Inhibition of EGFR by AG1478 abolished the apoptotic resistance effect mediated by YAP, as assessed by Annexin V/PI double staining ([Fig. 6d](#f6){ref-type="fig"}) and cleaved caspase-3 level ([Fig. 6e](#f6){ref-type="fig"}). These results suggest that EGFR is implicated in YAP-mediated anti-apoptotic function in response to TGF-β1. In addition, knockdown of EGFR in YAP overexpressing cells ([Supplementary Fig. S4](#S1){ref-type="supplementary-material"}) also repressed the YAP-mediated suppression of TGF-β1-induced apoptosis, as shown by the level of cleaved caspase-3 ([Fig. 6f](#f6){ref-type="fig"}) and FACS ([Fig. 6g](#f6){ref-type="fig"}). Moreover, EGF rescued the apoptosis induced by YAP knockdown ([Fig. 6h](#f6){ref-type="fig"}). These data suggest that EGFR accounts, at least partly, for the suppression of TGF-β1-induced apoptosis by YAP in NMuMG cells. Discussion ========== TGF-β1 signaling exerts multiple effects depending on different cellular contexts and environments. In non-cancer cells, such as NMuMG and AML12 cells, TGF-β1 can simultaneously induce apoptosis and EMT. The balance between apoptosis and EMT is critical to embryonic development, organogenesis, tissue repair, and regenerative processes. From the view of tumor transformation, the apoptotic effect has often been termed a tumor-suppressing factor, while EMT has been known as a tumor-promoting factor. Cancer cells overcome tumor suppressive barriers by harboring cell migration and metastatic abilities. Exploring the molecules that regulate apoptosis and EMT induced by TGF-β1 and the balance of these two fates is critical for understanding both the physiological and the pathological processes. In the present study, we showed that simultaneous apoptosis and EMT in NMuMG cells could be strongly induced by TGF-β1 treatment. YAP overexpression enabled non-transformed NMuMG cells to overcome TGF-β1 mediated apoptosis; while in YAP-knockdown cells, a large portion of cells underwent spontaneous apoptosis and increased apoptotic response to TGF-β1 stimulation. Our results indicate that YAP is vital in regulating the cell survival rate in response to TGF-β1. Although we found that YAP did not influence the induction of EMT directly, the expression of YAP controlled the numbers of cells that underwent EMT by regulating the survival rate of cells. Thus, YAP is vital for the balance of cell fate in apoptosis and EMT, which may be critical during embryonic development and in maintaining tissue homeostasis in living bodies. In the induction of simultaneous apoptosis and EMT by TGF-β1, although a portion of cells died, another portion of cells survived and underwent EMT. This fact suggests that both pro-apoptotic and anti-apoptotic molecules are involved in the TGF-β1-mediated effect. The balance or the extent of apoptosis or EMT may also be influenced by the levels of pro-apoptotic and anti-apoptotic molecules. It is worth noting that the samples collected for examining YAP and EGFR expression levels were mainly obtained from intact viable cells because apoptotic cells could not be collected effectively due to floating and degradation. In this case, it can be concluded that increased YAP expression was primarily from cells undergoing EMT. Consistent with our results, it has been reported that several classical anti-apoptotic molecules, such as PI3K-AKT, ERK, and c-Src, are activated in response to TGF-β1 in NMuMG cells and hepatocytes, which resist TGF-β1 mediated apoptosis[@b35][@b36][@b37][@b38][@b39]. In addition, MSK1[@b40], Dlx2[@b41], and lncRNA-smad7[@b42] have been shown to have an effect of anti-TGF-β1-induced apoptosis. However, their expression levels are also elevated during TGF-β1-induced apoptosis or EMT. Thus, the levels of anti-apoptotic molecules are essential for a portion of cells to survive and be resistant to the apoptotic effect of TGF-β1 stimulation. YAP downregulation markedly reduced the cell survival rate and a correspondingly decreased proportion of cells underwent EMT, suggesting that sufficient levels of YAP are necessary for cells to be viable and able to undergo EMT. Although TGF-β1-induced simultaneous EMT and apoptosis have been shown to be cell cycle-dependent events[@b14], it still remained unclear what molecular events are linked with different cell cycle phases. Interestingly, a decade ago, Snail has been shown to block cell cycle and confers cell resistance to death in MDCK cells 1[@b43]. Several years after that, it has also been shown that overexpression of snail induced EMT and suppressed TGF-β1-induced apoptosis in mouse hepatocytes[@b44]. These reports strongly support our new finding about the regulatable balance of apoptosis and EMT. In both cases, the anti-apoptotic effect of Snail are most likely associated with or dependent on its EMT inducing effect, because the following two reasons. Firstly, in both MDCK and hepatocytes, EMT can be promoted by TGF-β1 treatment or by Snail overexpression. Moreover, both apoptosis and EMT are the two strong simultaneous responses of hepatocytes to TGF-β1 treatment. Secondly, the induction of EMT generally presupposes cell growth arrest or decelerated proliferation, which can be more prominent in TGF-β1-induced EMT in non-cancer cells. It has been recently reported that TGF-β treatment in pancreatic ductal adenocarcinoma cells induces a lethal EMT that leads to apoptosis[@b45]. Due to the lack of a basic comparison of the time-effect between the apoptosis and EMT induced by TGF-β, the conclusion that EMT preceded apoptosis cannot be made. The relevant conclusion was also not backed up by other assays important for detecting the early apoptotic response of cells to TGF-β. Besides, the parallel assay showing the Snail overexpression induced EMT was not provided to support the assumed apoptosis-inducing effect of Snail. It is important to note that, in TGF-β-induced concomitant apoptosis and EMT in culture dishes, the EMT only temporarily, but not permanently, enabled cell to escape apoptotic inducing effect of TGF-β. YAP overexpression alone was able to induce EMT in MCF10A cells ([Supplementary Fig. S5](#S1){ref-type="supplementary-material"}), which is consistent with published reports[@b31][@b46]. Nevertheless, both overexpression and downregulation of YAP showed no obvious effects on TGF-β1-induced EMT in NMuMG cells. These results suggest that the function of YAP in EMT is cell type-dependent. In addition, the finding that knockdown of YAP in MCF10A cells cannot inhibit TGF-β1-induced EMT indicates that YAP is not required for the induction. In spite of the fact that YAP has no direct effect on TGF-β1-induced EMT, YAP levels are important for cell viability, which indirectly influences the number of cells undergoing EMT. Thus, during the processes of simultaneous apoptosis and EMT stimulated by TGF-β1, YAP plays a modulatory effect in determining the extent or trend of cell death versus EMT. YAP is primarily considered an oncogenic molecule and high levels of YAP predict poor prognosis in primary tumors[@b27]. YAP expression levels in various cancer types are higher than in normal tissues, which may provide an explanation for the fact that cancer cells often escape from the apoptotic effect and acquire the pro-metastasis effect of TGF-β1. Our findings that YAP functions as a potent suppressor of apoptosis and thus can indirectly modulate the concomitant apoptosis and EMT induced by TGF-β1 imply a novel effect of YAP in relevant patho-physiological processes. Epidermal growth factor receptor (EGFR) plays important roles in anti-apoptosis and solid tumor progression. The functions of EGFR and YAP in cell survival and tumor promotion are similar; thus, we examined whether EGFR could be regulated by YAP. TGF-β1 increased the expression of EGFR, and knockdown of YAP inhibited the upregulation of EGFR upon TGF-β1 stimulation. Inhibition of EGFR activity or knockdown of EGFR in YAP-overexpressing cells suggests that EGFR is involved in YAP-mediated cell survival. YAP has been shown to be able to transcriptionally upregulate EGFR in esophageal cancer tissues through TEAD binding sites[@b47]; in ErbB2/EGFR-transgenic mice, increased YAP level in nuclei of mammary glands was observed[@b48]. YAP has also been shown to activate the EGFR by increasing the expression of an EGFR ligand, amphiregulin[@b49][@b50]. These evidences suggest a mutual interaction between the Hippo and EGFR signaling pathways. We did not observe the induction of exogenous EGFR promoter activity by forced YAP expression ([Supplementary Fig. S6](#S1){ref-type="supplementary-material"}). It is possible that the upregulation of EGFR by YAP is not through direct binding of YAP to the EGFR promoter in normal mammary gland cells. However, further exploration of the mechanism by which YAP regulates EGFR expression is important on this aspect. As it has been shown recently that most YAP/TAZ-bound regions coincide with enhancer elements[@b51], one of the possibilities is that EGFR expression is regulated by YAP through an enhancer region. To briefly summarize, the presented results demonstrated that YAP plays a critical role in modulating cell fate in response to TGF-β1-mediated apoptosis and EMT ([Fig. 7](#f7){ref-type="fig"}). Overexpression of YAP suppressed apoptosis and increased the number of cells undergoing EMT. In contrast, downregulation of YAP reduced cell viability, leading to significant spontaneous apoptosis and an enhanced apoptotic response to TGF-β1 treatment, which also led to a pronounced decrease in the proportion of cells that underwent EMT. Further investigation found that YAP overexpression is associated with increased EGFR expression and activation. Inhibition of EGFR expression or activation blocked the anti-apoptotic effect of YAP, suggesting that the modulating effect of YAP on TGF-β1-induced apoptosis and EMT requires EGFR signaling. This study provides an insight into the regulation of cell fate through apoptosis and EMT induced by TGF-β1. Methods ======= Cells and cell culture ---------------------- NMuMG (non-transformed murine mammary gland), MCF10-A and HEK293T cells were purchased from the American Type Culture Collection (ATCC, Manassas, VA, USA). NMuMG cells were cultured in Dulbecco's modified Eagle's medium (Gibco, Grand Island, NY, USA) with 10% fetal bovine serum (Gibco), plus 10 μg/mL insulin (Sigma, St Louis, MO, USA). MCF10A cells were cultured in DMEM/F12 supplemented with 5% horse serum, 10 μg/mL insulin, 20 ng/mL EGF, 100 ng/mL cholera toxin and 0.5 μg/mL hydrocortisone[@b52]. HEK293T cells were cultured in DMEM plus 10% FBS. The cells were incubated at 37 °C in a humidified atmosphere with 5% CO~2~. Experiments were performed when cells reached 40--50% confluence. Reagents and plasmids --------------------- Human recombinant TGF-β1 was purchased from ChemiCon (Millipore, Billerica, MA, USA). Antibodies against Yap, phospho-Yap (Ser 127), H2B, caspase-3, cleaved caspase-3, and phospho-EGFR (Thr 1068) were purchased from Cell Signaling (Danvers, MA, USA). Antibodies against E-cadherin and N-cadherin were obtained from BD Biosciences (Franklin Lakes, NJ, USA). Antibody against ZO-1 was obtained from Invitrogen (Thermo-Fisher, Rockford, IL, USA). The antibodies against tubulin and fibronectin were purchased from Sigma (St Louis, MO, USA). EGF was purchased from Peprotech (Rocky Hill, NJ, USA) and AG1478 from Selleck (Shanghai, China). Cells were transfected using a lenti-virus system. Plasmids pCDH-CMV-MCSEF1-Puro, psPAX2, pMD2.G and pLKO.1-TRC were purchased from Addgene (Cambridge, MA, USA). The shRNAs were constructed in pLKO.1 with a target sequence as follows: Mouse Yap-1: ACTTGGAGGCGCTCTTCAATG, Mouse Yap-2: TGAGAACAATGACAACCAATA, Human Yap-1: GACATCTTCTGGTCAGAGA Human Yap-2: CCACCAAGCTAGATAAAGAA Mouse EGFR-1: GGGAAATGCTCTTTATGAA Mouse EGFR-2: AACTTACAGGAAATCCTGATT Immunoblotting -------------- Cells were lysed in lysis buffer (50 mM HEPES, 5 mM EDTA, 50 mM NaCl, 1% Triton X-100, 50 mM NaF, 10 mM Na~4~P~2~O~7~ · 10 H~2~O, protease inhibitors, pH 7.4). Proteins were separated by SDS-PAGE according to molecular weight and transferred onto nitrocellulose membranes (PALL, Port Washington, New York, USA). Membranes were blocked with 5% skim milk and were incubated with the appropriate primary antibodies diluted 1:2500 and then secondary antibodies conjugated with HRP. Protein bands were visualized by an ECL reagent. Immunofluorescent staining -------------------------- Immunofluorescent staining was performed as previously described[@b53]. RNA isolation and real-time PCR ------------------------------- Total RNA was extracted by TRIzol reagent (Tiangen, Beijing, China) and phenol/chloroform methods were performed according to the manufacturer's instructions. Briefly, 1 μg of RNA was used for reverse transcription by ReverTra Ace (Toyobo, Osaka, Japan). cDNAs were subjected to real-time PCR with a 7500 Fast Real-Time PCR system (ABI, Thermo-Fisher) using Power SYBR Green PCR Master Mix (CWBio, Beijing, China). Specific primer sequences are as follows: mEGFR-f: TGAAGAGGACATGGAGGATG mEGFR-r: CACTCAGAGAACTCAAGAGG mSnail-f: GTGTGTGGAGTTCACCTTCC mSnail-r: AGGAGAGAGTCCCAGATGAG mPAI1-f: TTCCACAAGTCTGATGGCAG mPAI1-r: GGTGGTGAACTCAGTGTAGT Annexin V-FTIC/PI double staining assay --------------------------------------- Cells were trypsinized and washed with PBS twice to remove residual EDTA. Cells were stained with Annexin V-FITC/PI for 10 min in binding buffer (0.01 M HEPES, 0.14 M NaCl; 2.5 mM CaCl~2~, pH 7.4). Stained cells were detected by flow cytometry (FACS Calibur, Becton Dickinson, Franklin Lakes, NJ, USA). To gate the apoptotic region, single-stained cells were used as negative controls. The percentage of apoptotic cells were quantified by the sum of Annexin V^+^PI^−^ and Annexin V^+^PI^+^ cells, as shown in the lower right and upper right quadrants of FACS assay. Stable transfection ------------------- Lenti-viruses were packaged in HEK293T cells by calcium phosphate transfection with plasmid. Viruses were collected 36 h after transfection and filtered through 0.45-μm filters. The host cells were infected by the collected viruses plus 10 μg/mL polybrene and selected with puromycin. Nuclear isolation ----------------- Cells were harvested and washed with PBS and then suspended in hypotonic buffer (10 mM HEPES, 1.5 mM MgCl~2~, 10 mM KCl, 0.5 mM DTT, protease inhibitors, pH 7.9) for 10 min. Cells were scraped and mixed with 10% NP-40 and then centrifuged. The supernatant contained cytosolic components and the precipitate contained the nuclei. The nuclei were resuspended in hypertonic buffer (20 mM HEPES, 1.5 mM MgCl~2~, 420 mM NaCl, 0.2 mM EDTA and 25% glycerol, protease inhibitors, pH 7.9). Luciferase reporter assay ------------------------- A mouse EGFR promoter was constructed in a pGL3-Basic vector and co-transfected into NMuMG cells with a pRL-TK Vector by Lipofectamine 2000 (Invitrogen, Thermo-Fisher, Rockford, IL, USA). Measurements of luciferase activity were performed according to the manufacturer's instructions. Statistical analyses -------------------- Quantitative data from at least three experiments were compared and normalized to control cells. For experiments with two groups, statistical significance was determined by a Student's t-test. For experiments involving more than two groups, statistical analyses were performed by ANOVA followed by pairwise comparison. A P-value \< 0.05 was considered statistically significant. \*P \< 0.05; \*\*P \< 0.01; n.s., non significant. Additional Information ====================== **How to cite this article:** Liu, Y. *et al*. YAP modulates TGF-b1-induced simultaneous apoptosis and EMT through upregulation of the EGF receptor. *Sci. Rep.* **7**, 45523; doi: 10.1038/srep45523 (2017). **Publisher\'s note:** Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Material {#S1} ====================== ###### Supplementary Information We thank Dr. Hongbin Ji for providing Yap and Yap^S127A^ expression plasmids. This study was supported by the Natural Science Foundation of China (81472603, 81672879). The authors declare no competing financial interests. **Author Contributions** J.S. and Y.L. conceived project, designed experiments, analyzed data, and wrote the manuscript. Y.L. performed experiments. X.G. participated in examining the EMT marker protein level. K.H., Y.H., W.S., W.D. and J.L. helped in data analysis. All authors read and approved the final manuscript. {#f1} {#f2} {#f3} {#f4} {#f5} {#f6} {#f7}

GENOME ANNOUNCEMENT {#h0.0} =================== The *Roseobacter* clade (family *Rhodobacteraceae*) is a group of exclusively marine bacteria within the alphaproteobacteria. In sea ice microbial communities (SIMCO), roseobacters are represented by the genus *Octadecabacter*. *O. antarcticus* constitutes up to 1% of the total bacterial community in southern ocean SIMCOs, while *O. arcticus* comprises up to 23% of the total bacterial community in Arctic SIMCOs ([@B1]). The nonpolar clusters are represented only by 16S rRNA gene sequences from uncultured organisms ([@B2]). A recently described mesophilic species has been reclassified as *Pseudooctadecabacter jejudonensis* ([@B3], [@B4]). *Octadecabacter temperatus* SB1 (DSM 26878) was isolated from a surface water sample collected in the harbor of the island of Helgoland in the North Sea ([@B4]). It is the first described nonpolar mesophilic isolate of the genus *Octadecabacter* and the type strain of the species. The strain was grown in marine broth (MB) medium at 20°C ([@B4]). Chromosomal DNA was isolated with the MasterPure complete DNA purification kit as recommended by the manufacturer (Epicentre, Madison, WI, USA). Whole-genome sequencing was performed with a combined approach using the 454 GS-FLX + pyrosequencer (Roche Life Science, Mannheim, Germany) and the Genome Analyzer IIx (Illumina, San Diego, CA, USA). Shotgun libraries for 454 sequencing as well as Nextera XT paired-end libraries for Illumina sequencing were prepared according to the manufacturer's protocols. Sequencing resulted in 54,153 pyrosequencing reads and 5,626,412 paired-end Illumina reads. Hybrid assembly of Illumina and 454 data was performed *de novo* with Mira v3.4.0.1 ([@B5]). Illumina data was also assembled with SPAdes v2.5.0 ([@B6]). Both assemblies were combined resulting in 8 contigs and a 158-fold coverage. Remaining gaps were closed by PCR-based techniques and Sanger sequencing of the recovered products using BigDye 3.0 chemistry on an ABI3730XL capillary sequencer (Applied Biosystems, Life Technologies GmbH, Darmstadt, Germany), and by employing Gap4 v4.11 of the Staden Package ([@B7]). The closed genome sequence of *O. temperatus* SB1 consists of one chromosome (3.23 Mb) and one plasmid (31.6 kb). To date this genome represents the smallest closed genome within the *Roseobacter* clade. Protein-encoding genes were identified and annotated with the Prokka annotation pipeline using Prodigal v2.6 ([@B8]). The closed genome harbored 1 rRNA operon, 39 tRNA genes, and 3,284 potential protein-encoding genes, of which 2,711 have a predicted function. The plasmid encodes a replication system of the RepABC-type with highest similarity to the RepC9-subtype, which is also present in the plasmids of the polar *Octadecabacter* species ([@B9]). Moreover, genes for the biosynthesis and export of lipopolysaccharides are encoded by the plasmid. Like the polar *Octadecabacter* species, *O. temperatus* SB1 lacks a gene cluster for aerobic anoxygenic photosynthesis. In addition, genes encoding xanthorhodopsins are absent in the genome of *O. temperatus*. This is in contrast to the polar *Octadecabacter* species, which harbor such genes as putative adaption for living in icy environments ([@B2]). Nucleotide sequence accession numbers. {#s1} -------------------------------------- This whole-genome shotgun project has been deposited at DDBJ/EMBL/GenBank under the accession numbers [CP012160](CP012160) (chromosome) and [CP012161](CP012161) (plasmid). **Citation** Voget S, Billerbeck S, Simon M, Daniel R. 2015. Closed genome sequence of *Octadecabacter temperatus* SB1, the first mesophilic species of the genus *Octadecabacter*. Genome Announc 3(5):e01051-15. doi:10.1128/genomeA.01051-15. This work was funded by the Deutsche Forschungsgemeinschaft (DFG) as part of the Collaborative Research Center TRR51. We thank Frauke-Dorothee Meyer for technical support.

\section{Introduction} \label{sec:introduction} The goal of hyperparameter and neural architecture search (HNAS) is to automate the process of finding the architecture or hyperparameters $\nestedmathbold{x}_{\star} \in \argmin_{\nestedmathbold{x} \in \mathcal{X}} f(\nestedmathbold{x})$ for a deep learning algorithm which minimizes a validation loss $f(\nestedmathbold{x})$, observed through noise: $y_i = f(\nestedmathbold{x}_i) + \epsilon_i,\quad \epsilon_i \sim \mathcal{N}(0, \sigma^2), i=1,\dots, n$. Bayesian optimization (BO)~\citep{jones-jgo98a,shahriari-ieee16a} is an effective model-based approach for these expensive optimization problems. It constructs a probabilistic surrogate model of the loss function $p(f \mid \mathcal{D})$ based on previous evaluations $\mathcal{D} = \{(\nestedmathbold{x}_i, y_i)\}_{i=1}^n$. Searching for the global minimum of $f$ is then driven by trading off exploration in regions where our model is uncertain and exploitation in regions where the global optimum is assumed to be located. For HNAS problems, standard BO needs to be augmented in order to remain competitive. Training runs for unpromising configurations $\nestedmathbold{x}$ can be stopped early, and thereby serve as low-fidelity approximations of $f(\nestedmathbold{x})$~\citep{Swersky:14,domhan-ijcai15,klein-iclr17}. Also, evaluations of $f$ can be executed in parallel to reduce the wall-clock time required for finding a good solution. Successive halving (SH)~\citep{jamieson-aistats16} and Hyperband (HB)~\cite{li-iclr17} are simple, easily parallelizable multi-fidelity scheduling algorithms. BOHB~\citep{falkner-icml18} combines the efficient any-time performance of HB with a probabilistic model to guide the search. However, both HB and BOHB are {\em synchronous} algorithms: stopping decisions are done only after synchronizing all training jobs at certain resource levels (called {\em rungs}). This can be wasteful when the evaluation of some configurations take longer than others, as it is often the case for HNAS~\citep{ying-arxiv19}, and can substantially delay progress for high-performing configurations (see \figref{teasers} for an illustration). Asynchronous parallel variants of HB have been proposed \citep{li-arxiv18}, but rely on a uniform random sampling of architectures and hyperparameters. \begin{figure}[ht] \centering \begin{subfigure}[t]{0.47\columnwidth} \resizebox{\columnwidth}{!}{\input{plots/hyperband_synchronous.pgf}} \end{subfigure} \begin{subfigure}[t]{0.47\columnwidth} \resizebox{\columnwidth}{!}{\input{plots/hyperband_stopping.pgf}} \end{subfigure} \caption{Evolution of synchronous (left) and asynchronous (right) BOHB on OpenML MLP tuning (see \secref{experiments}). x: wall-clock time (150 secs), y: training epochs done (log scale). Coloured lines trace training jobs for HP configurations, with marker size proportional to validation error (smaller is better). Lines not reaching the maximum number of authorized epochs (27) are stopped early. For synchronous BOHB, rungs (epochs 1, 3, 9) need to be filled completely before any configuration can be promoted, and each synchronization results in some idle time. In asychronous BOHB, configurations proceed to higher epochs faster. Without synchronization overhead, more configurations are promoted to the maximum number of epoch in the same wall-clock time (figure best seen in colours).} \label{fig:teasers} \end{figure} \subsection{Contributions} Dealing with HPO in general and HNAS for neural networks, we would like to make the most efficient use of a given parallel computation budget (e.g., number of GPU cores for a certain amount of time) in order to find high-accuracy solutions in the least amount of wall-clock time. Besides exploiting low-fidelity approximations, we demonstrate that {\em asynchronous parallel} scheduling is a key property for cost-efficient search. Our novel combination of multi-fidelity BO with asynchronous parallel scheduling, dubbed asynchronous BOHB (A-BOHB), can substantially outperform either of them in isolation. More specifically: \begin{itemize} \item We clarify differences between existing asynchronous HB extensions. In our experiments, ASHA, as described by Li et al.~\citep{li-arxiv18} is frequently outperformed by a simpler stopping rule variant, related to the median rule \citep{golovin-kdd17} and first implemented in Ray Tune \citep{liaw2018tune}. \item We introduce a new model-based extension of asynchronous HB which handles multi-fidelity evaluations in a principled way, simultaneously reasoning across resource levels and asynchronously across workers. The underlying surrogate model is a joint Gaussian process (GP) akin the ones used in \citep{Swersky:13, Swersky:14}. \item By sampling from our model, we often achieve the same performance in the same amount of wall-clock time but with half the computational resources as sampling uniformly at random. \item On neural network benchmarks, we show that A-BOHB is more efficient in terms of wall-clock time than other state-of-the-art algorithms, especially if evaluations are expensive. Synchronous BOHB \citep{falkner-icml18} suffers from substantial synchronization overheads, while variants of asynchronous HB fail to approach the global optima. \end{itemize} Next, we relate our work to approaches recently published in the literature. In \secref{schedule}, we review synchronous SH and HB, as well as two asynchronous extensions. Our novel method is presented in \secref{model}. We present empirical evaluations for the HNAS of various neural architecture types in \secref{experiments}, and finish with conclusions and future work in \secref{conclusions}. \subsection{Related Work}\label{sec:related_work} A range of prior Gaussian process (GP) based BO work exploits multiple fidelities of the objective~\citep{Kennedy:00,Klein:17,Poloczek:18,Cutajar:18}. A joint GP model across configurations and tasks was used by Swersky et al. \cite{Swersky:13}, allowing for trade-offs between cheap auxiliaries and the expensive target. Klein et al.~\citep{Klein:17} presented a continuous multi-fidelity BO method, where training set size is an input. It relies on a complex and expensive acquisition function, and experiments with asynchronous scheduling are not provided. The ``freeze-thaw'' approach \citep{Swersky:14} allows for asynchronous parallel evaluations, yet its scheduling is quite different from asynchronous HB. They propose an exponential decay kernel, which we use here as well. No results for deep neural network tuning were presented, and no implementation is publicly available. Previous work on asynchronous BO has demonstrated performance gains over synchronous BO methods~\citep{alvin-icml19, kandasamy-aistats18}, yet this work did not exploit multiple fidelities of the objective. Synchronous HB was combined with BO in the BOHB method~\citep{falkner-icml18}, using TPE~\citep{Bergstra:11} for BO. It combines early speed-ups of HB with fast convergence later on, and shows competitive performance for NAS \citep{ying-arxiv19, dong-arXiv20}. BOHB fits independent surrogate models at each resource level, which hinders learning curve interpolation. Being based on TPE, it is unclear how to extend BOHB to the asynchronous setting. \section{Synchronous and Asynchronous Multi-Fidelity Scheduling} \label{sec:schedule} Multi-fidelity HNAS considers an objective $f(\nestedmathbold{x}, r)$, indexed by the {\em resource level} $r\in \{r_{\text{min}}, \dotsc, r_{\text{max}} \}$. The target function of interest is $f(\nestedmathbold{x}) = f(\nestedmathbold{x}, r_{\text{max}})$, while evaluations of $f(\nestedmathbold{x}, r)$ for $r < r_{\text{max}}$ are lower-fidelity evaluations that may be correlated with $f(\nestedmathbold{x})$ and therefore help guide the search. Here, we consider the number of training epochs ({\it i.e.}, full sweeps over the data) as the resource $r$, though other choices can be sensible as well ({\it e.g.}, training subset ratio\citep{Klein:17}). Parallel multi-fidelity HNAS methods need to make two kinds of decisions: {\em choosing} configurations to evaluate and {\em scheduling} their evaluations (including resource allocation). Scheduling decisions happen at certain resource {\em rungs}, chosen according to the SH principle~\cite{jamieson-aistats16}. Let $\eta\in\{2, 3, 4\}$ be the halving constant, $r_{\text{max}}$ and $r_{\text{min}}$ the maximum and minimum resource level for an evaluation (assuming, for simplicity, that $r_{\text{max}} / r_{\text{min}} = \eta^K$). The full set of rungs is $\mathcal{R} = \{ r_{\text{min}}\eta^k\; |\; k=0,\dots, K \}$. Then, {\em brackets} are (ordered) subsets $B_s = \{ r_{\text{min}}\eta^{k+s}\; |\; k=0,\dots, K-s \} \subseteq \mathcal{R}$, where $s=0,\dots, K$. In the sequel, the term ``rung'' will be overloaded to both mean the resource level at which stop/go decisions are made, and the list of configs (with metric values) which reached this level. We distinguish between {\em synchronous} and {\em asynchronous} scheduling. The former is detailed in \citep{li-iclr17,falkner-icml18}. In a nutshell, each rung of each bracket has an \textit{a priori} fixed size ({\it i.e.}, number of slots for configs evaluated until the rung level), the size ratio between successive rungs being $\eta^{-1}$. Hyperparameter configurations promoted to a rung are evaluated until the corresponding resource level, making use of parallel computation. Once all of them finish, the $\eta^{-1}$ fraction of top performing configurations are promoted to the next rung while the others are terminated. Each rung is a synchronization point for many training tasks: it has to be populated entirely before any configuration can be promoted. In an outer loop, the HB algorithm cycles over brackets $s=0,\dots, K$ periodically until the total budget is spent. The original SH algorithm is obtained as a special case, using only the single bracket $B_0$. We provide a unified perspective on two different asynchronous generalizations of HB, proposed in prior work. For any hyperparameter and architecture configuration, its bracket index $s\in\{0,\dots, K\}$ determines the rung levels $B_s$ at which scheduling decisions are made, sampled from a distribution $P(s)$. Our choice of $P(s)$ is provided in \ifappendix \appref{bracket_distribution}. \else the supplemental material. \fi Once an evaluation of $\nestedmathbold{x}$ reaches a rung $r\in B_s$, its metric value $y$ is recorded there. Then, the binary predicate $\texttt{continue}_s(\nestedmathbold{x}, r, y)$ evaluates to true iff $y$ is in the top $\eta^{-1}$ fraction of records at the rung. \parhead{Stopping variant.} This is a simple extension of the median stopping rule \citep{golovin-kdd17}, first implemented in \href{https://ray.readthedocs.io/en/latest/tune.html}{Ray Tune} \citep{liaw2018tune}. Once a job for $\nestedmathbold{x}$ reaches a rung $r$, if $\texttt{continue}_s(\nestedmathbold{x}, r, y)$ is true, it continues towards the next rung level. Otherwise, it is {\em stopped}, and the worker becomes free to pick up a novel evaluation. As long as fewer than $\eta$ metrics are recorded at a rung, the job is always continued. \parhead{Promotion variant (ASHA).} This variant of asynchronous HB was presented by Li et al. \citep{li-arxiv18} and called {\em ASHA}. Once a job for some $\nestedmathbold{x}$ reaches a rung $r$, it is {\em paused} there, and the worker is released. The evaluation of $\nestedmathbold{x}$ can be {\em promoted} ({\it i.e.}, continued to the next rung) later on. When a worker becomes available, we sample a bracket $s$ and scan the rungs $r\in B_s$ in descending order, running a job to promote the first paused $\nestedmathbold{x}$ for which $\texttt{continue}_s(\nestedmathbold{x}, r, y)$ is true. When no such paused configuration exists, we start a new evaluation from scratch. With fewer than $\eta$ metrics recorded at a rung, no promotions happen there. The promotion variant needs pause-and-resume training. \footnote{ Supporting pause-and-resume by checkpointing is difficult in the multi-node distributed context. It is simpler to start training from scratch whenever a configuration is promoted. This is done in our implementation.} Synchronous scheduling makes less efficient use of parallel computation. Configurations in the same rung may require quite different compute times, for example, if we search over hyperparameters which control network size. At high rungs, few configurations run for much longer, and some workers may just be idle at a synchronization point (such points do not exist in the asynchronous case). Finally, rung levels have fixed sizes, which have to be filled before any configuration can be promoted, while promotions to larger resource levels can happen earlier with asynchronous scheduling. Of course, the latter risks promoting mediocre configurations simply because they were selected earlier. \figref{teasers} illustrates these differences. The stopping and promotion variants can exhibit rather different behavior. Initially, the stopping variant gives most configurations the benefit of doubt, while promotion pauses evaluations until they can be compared against a sufficient number of alternative configurations. \section{Model-based Asynchronous Multi-Fidelity Optimization} \label{sec:model} We now describe our new model-based method that is able to make asynchronous decisions across different fidelities of the objective function. The configurations to be evaluated in HB are drawn at random from the configuration space. Motivated by the fact that in standard sequential settings, Gaussian process (GP) based BO tends to outperform random search~\citep{shahriari-ieee16a}, we aim to equip asynchronous HB with decisions based on a {\em GP surrogate model}. Recall that $\nestedmathbold{x}$ denotes the hyperparameter configuration, $r$ the resource level (see \secref{schedule}). Following a range of prior work on non-parallel multi-fidelity BO \cite{Swersky:13, Klein:17, Poloczek:18, Swersky:14}, we place a GP prior over $f(\nestedmathbold{x}, r)$ to jointly model not only the auto-correlation within each fidelity, but also the cross-correlations between fidelities, \begin{equation*} f(\nestedmathbold{x}, r) \sim \mathcal{GP}(\mu(\nestedmathbold{x}, r), k((\nestedmathbold{x}, r), (\nestedmathbold{x}', r'))). \end{equation*} Compared to synchronous BOHB~\citep{falkner-icml18}, which employs an independent TPE~\citep{bergstra-nips11a} model over $\nestedmathbold{x}$ for each fidelity $r$, our method uses a joint GP model that explicitly captures correlations across all fidelities. Since multiple configurations are evaluated in parallel, we have to deal with {\em pending} evaluations when proposing new configurations to be evaluated. We do so by {\em fantasizing} their outcomes, as detailed by Snoek et al.~\citep{Snoek:12}. In a nutshell, we marginalize the acquisition function over the GP posterior predictive of evaluation outcomes for pending configurations. In practice, we approximate this by averaging the acquisition function over sampled function values. It is worth noting that, as long as kernel hyperparameters remain constant, fantasizing comes almost for free. For GP regression, the posterior covariance matrix does not depend on the function values (only on the inputs), which means that fantasizing does not require additional cubic-complexity computations. Kernel hyperparameters may be updated with empirical Bayes or Markov chain Monte Carlo~\citep{Snoek:12}. {\bf How do pending evaluations and fantasizing generalize to the asynchronous case?} The relevant covariates are now the pairs $(\nestedmathbold{x}, r)$. As a task for $\nestedmathbold{x}$ runs down a bracket $s$, it emits target values at $r\in B_s$ (unless stopped). The following logic is task-local and ensures that each labeled covariate was previously pending. If a task to evaluate $\nestedmathbold{x}$ is started in bracket $s$, we register a pending evaluation for $(\nestedmathbold{x}, r_{\text{min}}\eta^s)$. Once it reaches the next rung $r\in B_s$, we convert the \emph{pending} to a \emph{labeled} evaluation $(\nestedmathbold{x}, r)$. Then, as detailed above, we make a stopping decision. If the configuration survives, we register $(\nestedmathbold{x}, r \eta)$ as \emph{pending}, as $r \eta$ is the next rung it will reach. Details about how posterior distributions over both labeled and labeled plus fantasized evaluations are represented, are given in \ifappendix \appref{posterior}. \else the supplemental material. \fi All linear algebra computations are delayed until the posterior is needed for the next decision. {\bf Choice of the kernel.} We implemented and compared a range of different surrogate models for $f(\nestedmathbold{x}, r)$: a joint Mat\'{e}rn $\frac{5}2$ kernel with automatic relevance determination (ARD) \cite{Snoek:12} over $(\nestedmathbold{x}, r)$, with and without warping of $r$; the additive exponential decay model of \citep{Swersky:14}; a novel non-additive extension of the latter, relaxing underlying assumptions which may not hold in practice. In all cases, the base kernel on $\nestedmathbold{x}$ alone is Mat\'{e}rn $\frac{5}2$ ARD. Details for our new surrogate model are provided in \ifappendix \appref{surrogate}. \else the supplemental material. \fi While A-BOHB can significantly outperform synchronous BOHB or non-model-based variants, we found that the choice of joint surrogate model did not have a consistent impact on these results. {\bf Acquisition strategy.} As in synchronous BOHB~\cite{falkner-icml18}, we employ the GP surrogate model only to choose configurations $\nestedmathbold{x}$ for new tasks, while scheduling decisions are done as in HB. Compared to standard BO without multiple fidelities, it is much less clear how a non-myopically optimal decision would look like in this setting. In this paper, we follow synchronous BOHB in choosing $\nestedmathbold{x}$ by optimizing the expected improvement (EI) acquisition function \cite{shahriari-ieee16a} over the posterior on $f(\cdot, r_{\text{acq}})$, where $r_{\text{acq}}$ is fixed for each decision. We choose $r_{\text{acq}}\in \mathcal{R}$ as the largest rung for which at least $L$ metric values have been observed, where $L$ is typically set to the input dimensionality (\emph{i.e.} number of hyperparameters). After some initial period, $r_{\text{acq}}$ eventually converges to $r_{\text{max}}$. This strategy does not take into account that the choice of $\nestedmathbold{x}$ may have different downstream stopping or promotion consequences. Exploring alternative strategies is a direction left for future work. \section{Experiments}\label{sec:experiments} We compared our asynchronous BOHB (A-BOHB) to a range of state-of-the-art algorithms. For synchronous HB and BOHB, we used the {\tt HpBandSter} package.\footnote{\url{https://github.com/automl/HpBandSter}} The implementation by Falkner et al. \citep{falkner-icml18} uses a simple mechanism to reduce the amount of idling workers: whenever evaluations for a rung are started, if there are more free workers than slots in the rung, left-over workers are used to immediately start the next bracket. However, this requires a logic associating workers with brackets, and still needs synchronization between workers. We implemented A-BOHB, as well as all other asynchronous techniques, in {\tt AutoGluon}.\footnote{\url{https://github.com/awslabs/autogluon}} A number of worker nodes run training jobs as scheduled, publishing metrics after each epoch. One of the workers doubles as master, concurrently running the HP optimization and scheduling. The master node communicates with other workers over the network. All experiments were conducted on \textsc{AWS EC2}, using \texttt{m4.xlarge} instances for CPU, and \texttt{g4dn.xlarge} instances for GPU. Our asynchronously distributed HNAS system will be open sourced in {\tt AutoGluon} along with this publication. In all methods using SH scheduling, the halving rate was $\eta=3$. Our performance metric after wall-clock time $t$ is the {\em immediate regret} $r_t = y_t - y_{\star}$, where $y_{\star}$ is the best observed validation error across all methods, runs and time steps on this dataset. Using the regret in place of just the validation error gives a more detailed picture of progress towards the global optimum. We report mean and the standard error of the mean across multiple runs for each method. We are not aware of a publicly available implementation of Swersky et al. \cite{Swersky:14} (whose complexity is considerable), nor of follow-up work, which is why we could not compare against this technique in our experiments. We run experiments on the following set of benchmarks. A detailed list of all hyperparameters with their corresponding ranges, along with further details, can be found in \ifappendix \appref{hpranges}.\else the supplemental material.\fi {\bf Featurized OpenML Datasets:} For our first benchmark, we optimize the hyperparameters of a multi-layer perceptron (MLP) on tabular classification datasets gathered from OpenML~\citep{vanschoren-sigkdd13a}. Our MLP has two hidden layers with ReLU activations. We tuned 8 hyperparameters: the learning rate of ADAM~\citep{kingma-iclr15}, the batch size, and for each layer the dropout rate, number of units and the scale of a uniform weight initialization distribution. The number of epochs $r$ varies between $r_{\text{min}} = 1$ and $r_{\text{max}} = 27$, and HB is using 4 brackets. Results for each method are averaged over 30 independent runs with different random seeds. Due to space constraints, we only report results on \textit{electricity}, all other results are provided in \ifappendix \appref{results}. \else the supplemental material. \fi {\bf Neural Architecture Search on Tabular Benchmark:} To evaluate our approach in a NAS setting, we used the NASBench201 datasets from Dong et al. \citep{dong-arXiv20} which consists of a offline evaluated grid of neural network architectures for three different datasets: CIFAR-10, CIFAR-100~\citep{krizhevsky-tech09a} and Imagenet16~\citep{chrabaszcz-arXiv17}. Categorical hyperparameters were one-hot encoded. The minimum and maximum number of epochs were $r_{\text{min}} = 1$ and $r_{\text{max}} = 200$, and HB uses 5 brackets. To simulate realistic dynamics of asynchronous NAS, we put each worker to sleep for the time per epoch each real evaluation took. This does not speed up experiments, but decreases cost, since no GPU instances are required. {\bf Language Modelling:} Next, we tune an LSTM~\citep{hochreiter-neural97}, applied to language modelling on the WikiText-2 dataset~\cite{merity-arXiv17} with the default training/validation/test split. We optimized the initial learning rate of SGD, batch size, dropout rate, gradient clipping range, and decay factor for the learning rate which is applied if the validation error has not improved upon the previous epoch. The LSTM consisted of two layers with 1500 hidden units each. For the word embeddings, we also used 1500 hidden units and tied them with the output. The number of epochs $r$ runs between $r_{\text{min}}=1$ and $r_{\text{max}}=81$, and HB uses 5 brackets. {\bf Image Classification:} Lastly, we tune ResNet~\citep{he-cvpr16} on the image classification dataset CIFAR-10~\citep{krizhevsky-tech09a}, tuning the batch size, weight decay, initial learning rate and momentum. We used 40000 images for training, 10000 for validation, and applied standard data augmentation (cropping; horizontal flips). SGD is used for training, the learning rate is multiplied by $0.1$ after 25 epochs. We used $r_{\text{min}}=1$ and $r_{\text{max}}=27$, and 4 brackets for HB. As our goal is to differentiate between the methods, we neglected further tricks of the trade needed to get to state-of-the-art accuracy~\citep{yang-iclr20}. \subsection{Comparing Random- and Model-based Approaches, and their Schedulers} We start with comparing methods using asynchronous scheduling: our BOHB (A-BOHB) and Hyperband (A-HB). Recall that scheduling comes in two variants: stopping and promotion, and that A-HB promotion is known as ASHA~\cite{li-arxiv18}. \figref{async_methods} shows the immediate regret over wall-clock time on the {\it electricity} dataset for 4 workers (left) and the NASBench201 CIFAR-10 dataset for 8 workers (right). Qualitatively similar results for other datasets can be found in \ifappendix \appref{results}.\else the supplemental material.\fi \begin{figure*}[t] \begin{center} \includegraphics[width=.49\textwidth]{plots/traj_time_electricity_nodes_4_async.pdf} \includegraphics[width=.49\linewidth]{plots/nasbench201_cifar10-valid_async.pdf} \caption{Comparison of asynchronous HB and BOHB (stopping and promotion scheduling variants). Left: MLP on {\em electricity} (4 workers). Right: NASBench201 on CIFAR-10 (8 workers). A-BOHB clearly outperforms A-HB. Differences between the scheduling variants are insignificant for MLP, and more pronounced for NAS.} \label{fig:async_methods} \end{center} \end{figure*} \begin{figure*}[t] \begin{center} \includegraphics[width=.49\textwidth]{plots/scaling_num_workers_stopping.pdf} \includegraphics[width=.49\textwidth]{plots/scaling_num_workers_promotion.pdf} \caption{Comparison of stopping (left) and promotion (right) variant for A-HB (dashed lines) and A-BOHB (solid lines) with a variable number of workers on the MLP benchmark. Model-based approaches consistently outperform random-based ones, even as the number of workers increases. } \label{fig:async_methods_scaling} \end{center} \end{figure*} {\bf Model-based versus random-based}. First, we focus on differences between our model-based A-BOHB and A-HB, for both scheduling variants. Initially, the model-based variants perform similar to their random-based counterparts. After some time, they rapidly approach the global optimum, while this would take a lot longer for methods relying on random sampling. In line with typical improvements of BO over random search, a well-fitted model focusses the search to a well-performing part of the space that is less likely to be hit by random sampling. As seen in \figref{async_methods_scaling}, the substantial speed-ups of BOHB over HB are observed consistently with 1, 2, 4, 8, and 16 workers on the MLP benchmark. In fact, in most cases A-BOHB converges to a high-accuracy solution in about the same time as A-HB needs, with twice the number of workers: {\em using a joint surrogate model can save half the resources compared to random search}. {\bf Stopping versus promotion (ASHA) scheduling}. Next, we focus on the two variants of asynchronous scheduling (stopping and promotion), for both A-HB and A-BOHB. On the cheaper MLP benchmark, both scheduling variants perform on-par (\figref{async_methods}, left), while on the more expensive NAS benchmark, the stopping rule achieves a good performance faster than the promotion variant (\figref{async_methods}, right). Recall from \secref{schedule} that for the stopping rule, the initial $\eta$ configurations are not stopped, while the promotion variant pauses all initial configurations at the lowest rung level of their bracket. At least in our expensive experiments, this more careful approach of promoting configurations leads to a worse any-time performance. It is important to note that {\em synchronous} HB and BOHB share this conservative property, along with idle time at synchronization points. At least in situations where initial random configurations, or even well-chosen defaults, provide valuable information at all resource levels, the simpler stopping rule can have an edge. Importantly, in none of our experiments did we witness the initial optimistic scheduling of the stopping rule resulting in worse performance than ASHA at larger wall-clock times. \subsection{Comparison to State-of-the-Art Methods} We now compare our novel A-BOHB (stopping) method against a range of state-of-the-art algorithms: asynchronous HB (ASHA) as proposed by Li et al.~\citep{li-arxiv18}, asynchronous Bayesian optimization~\citep{Snoek:12} (A-BO), asynchronous random search~\citep{Bergstra:11} (A-RS), synchronous BOHB~\citep{falkner-icml18} and synchronous HB~\citep{li-iclr17}. Recall that the differences between synchronous and asynchronous BOHB go much beyond scheduling (see \secref{related_work}). {\bf NASBench201 datasets.} \figref{nasbench} provides results for the three NASBench201 datasets. ASHA performs better than synchronous HB and BOHB, which also has been observed by Li et al.~\citep{li-arxiv18}, but still requires substantially more time to converge to a sensible performance than BO and RS. Note that metrics are reported after each epoch for any method (BO and RS just cannot stop or pause a configuration), so they all can exhibit good any-time performance. We already observed an edge of stopping over promotion scheduling, yet the effect here is stronger. An explanation could be the observation by Dong et al. \citep{dong-arXiv20} that ranking correlations between architectures evaluated for less than 100 epochs versus the full 200 epochs is small, and that even randomly sampled configurations in NAS search spaces often result in decent performance~\citep{dong-arXiv20,yang-iclr20}. Not only is ASHA slower at promoting any configuration to higher resources, it also may initially promote suboptimal ones. This property would just as well adversely affect synchronous BOHB and HB. In contrast, A-BO and A-BOHB run initial configurations for 200 epochs, sampling valuable data at all resource levels in between, allowing them to make better choices further down. At larger wall-clock times, A-BOHB clearly outperforms A-BO, which shows that the information gathered at all resource levels leads to efficient early stopping of configurations. According to results on NASBench201 (CIFAR-10) reported by Dong et al.\citep{dong-arXiv20}, the DARTS-V2 method~\citep{liu-iclr19} achieves a poor validation accuracy of $39.77 \pm 0.00 $ (averaged over 3 runs), the best performing method was regularized evolution \citep{real-icml17a} with validation accuracy of $91.19\pm0.31$ (averaged over 500 runs). Our new proposed method A-BOHB achieves a validation accuracy of $91.3 \pm 0.43$ (averaged over 20 runs). The accuracy $90.82\pm0.53$ reported there for synchronous BOHB matches well with $91.02 \pm 0.0029$ in our experiments. While we provide accuracy as function of wall-clock time for a specified compute budget, the results by Dong et al.\citep{dong-arXiv20} are obtained for vastly different compute and wall-clock time budgets. \begin{figure*}[t] \begin{center} \includegraphics[width=.32\linewidth]{plots/nasbench201_cifar10-valid.pdf} \includegraphics[width=.32\linewidth]{plots/nasbench201_cifar100.pdf} \includegraphics[width=.32\linewidth]{plots/nasbench201_ImageNet16-120.pdf}\\ \caption{Comparison of state-of-the-art distributed HNAS methods on different NASBench201 datasets, run with 8 workers. The synchronous methods clearly require more time to reach good accuracies than the asynchronous ones. ASHA is initially slowed down by its more conservative promotion scheme. At higher wall-clock times, our A-BOHB clearly outperforms all other methods.} \label{fig:nasbench} \end{center} \end{figure*} \begin{figure*}[t] \begin{center} \includegraphics[width=.49\linewidth]{plots/wikitext.pdf} \includegraphics[width=.49\linewidth]{plots/resnet.pdf} \caption{\label{fig:wikitext_resnet} Left: Comparison of state-of-the-art distributed HNAS methods on LSTM language modelling (WikiText-2), using 8 workers. Right: Comparison of A-BOHB (bold) and A-HB (dashed), both stopping variant, on tuning ResNet (CIFAR-10).} \end{center} \end{figure*} {\bf LSTM on WikiText-2.} \figref{wikitext_resnet} left shows results for LSTM tuning on the WikiText-2 dataset, using 8 workers. In order to save computations, we limited the comparison to A-BO, A-BOHB (stopping), ASHA, and synchronous BOHB. Training for the full 81 epochs takes roughly 12 hours, and synchronous BOHB is only able to execute a single bracket in the alloted time budget. Due to limited data, it cannot fit a surrogate model at the highest resource level $r_{max}$ (and lacking a joint model, it cannot interpolate from lower rungs either), which may explain its overall worst performance. ASHA performs better on this problem than on the others, and it becomes on-par with A-BOHB after about 25000 secs (circa 7 hours), yet its any-time performance is worse. {\bf ResNet on CIFAR-10.} In \figref{wikitext_resnet} right, we compare our A-BOHB against A-HB, both with the stopping based variant, for 2, 4, and 8 workers respectively. {\em A-BOHB achieves a competitive regret of $10^{-2}$ at roughly the same time as A-HB with twice as many workers} (a regret of $10^{-2}$ means a difference of $1\%$ in validation accuracy). \section{Conclusions}\label{sec:conclusions} We present a novel methodology for distributed hyperparameter and architecture search, combining {asynchronous} Hyperband (HB) scheduling with multi-fidelity Bayesian optimization (BO). Instead of sampling new configurations at random, they are chosen based on a joint Gaussian process surrogate model over resource levels, and asynchronous scheduling is accommodated by fantasizing the outcome of pending candidates. While we propose a novel surrogate model, we do not find the kernel choice to be particularly relevant, as long as a probabilistic model interpolates across resources and extends over observed and pending data. In a range of tuning experiments, spanning MLPs, tabular NAS benchmarks, neural language modelling and image classification, we compare our proposal to prior synchronous and asynchronous HNAS state-of-the-art, focusing on regret as function of wall-clock time (given same resources), which is ultimately most relevant for practitioners. Asynchronous BOHB significantly outperforms both synchronous BOHB and two different variants of asynchronous HB, especially on very expensive benchmarks. The former is slowed down by synchronization overhead, using its parallel resources inefficiently at higher resource levels. We found that asynchronous HB is competitive when given a lot of parallel resources, but degrades with less parallelism. While it finds good solutions, we often find it to plateau before solutions with state-of-the-art performance are found. The asynchronously distributed HNAS method we implemented for this work will be released as open source. In the future, we will research more principled ways of non-myopic decision-making in the context of asynchronous parallel scheduling, as well as ways to make model-based decisions on scheduling that are more sophisticated than successive halving. We will further investigate advanced bracket sampling strategies and rung specifications. \clearpage \newpage \newpage \section*{Broader Impact} Our work encourages the widespread adoption and democratization of AI and machine learning. Without a high degree of domain-specific knowledge and experience, one typically has to resort to laborious manual tuning or brute-force search which, in addition to consuming time and compute resources, can dramatically increase carbon footprint. By offering an efficient and principled approach to automated hyperparameter tuning, we make it accessible for a broader technical audience to effectively apply and deploy off-the-shelf machine learning methods, while utilizing resources more efficiently. There has been recent massive interest and investment in distributed stochastic optimization applied to neural architecture search (NAS), or more general the automated tuning of expensive deep learning models. While Bayesian optimization (BO) is standard for ML hyperparameter tuning, it has seen surpringly little use in NAS. Most existing open source BO software is limited to single compute instances and mostly sequential search. Asynchronous scheduling in particular, while widely adopted in NAS, is not well served at all. Our work is not only proposing novel methodology for GP-based asynchronous parallel HPO, but also comes with a distributed HPO software system, which will be open sourced along with this publication. Deep learning practitioners with little expertise or interest in BO can use this system in order to automate their workflow, be it tuning hyperparameters or data augmentation, or the search stage of NAS.

\section{Introduction} \begin{figure}[bt!] \centering \hfill \includegraphics[width=\columnwidth]{SharpSepConv.pdf} \caption{\small \label{fig:SharpSepConv} Our \texttt{SharpSepConv} Block configured with 3x3 kernels and stride 1. All convolutions are either 1x1 convolutions or Depthwise Separable Convolutions\cite{chollet2017xception}. The \texttt{c\_mid} hyperparameter can be either proportional to \texttt{c\_out} or set to an arbitrary constant. } \vspace{-0.2cm} \end{figure} Neural Architecture Search (NAS) promises to automatically and efficiently optimize a complex model based on representative data so that it will generalize, and thus make accurate predictions on new examples. Recent NAS results have been impressive, notably in computer vision, but the search process took large amounts of computing infrastructure~\cite{2017nasnet,zoph2017neural,liu2018PNAS}. These methods were followed up with a multiple order of magnitude increase in search efficiency~\cite{2018enas,liu2018darts}. However, key questions remain: What makes a search space worth exploring? Is each model visited in a search space getting a fair shot? We investigate how to improve the search space of DARTS\cite{liu2018darts}, which is one of several based on NASNet\cite{2017nasnet,real2018regularized,2018enas,liu2018PNAS,huang2018gpipe}, propose Differentiable Hyperparameter Grid Search with DARTS, and draw conclusions that apply across architecture search spaces. \noindent To summarize, we make the following contributions: \begin{enumerate} \item Define the novel \texttt{SharpSepConv} block with a more consistent structure of model operations and an adaptable middle filter count in addition to the sharpDARTS architecture search space. This leads to highly parameter-efficient results which match or beat state of the art performance for mobile-scale architectures on CIFAR-10, CIFAR-10.1, and ImageNet with respect to Accuracy, AddMult operations, and GPU search hours. \item Introduce the Cosine Power Annealing learning rate schedule for tuning between Cosine Annealing\cite{loshchilov2016sgdrcosineannealing} and exponential decay, which maintains a more optimal learning rate throughout the training process. \item Introduce Differentiable Hyperparameter Grid Search and the HyperCuboid search space for efficiently evaluating arbitrary discrete choices. \item Demonstrate the low-capacity bias of DARTS search on two search spaces, and introduce Max-W Weight Regularization to correct the problem. \end{enumerate} \section{Related Work} \label{ref:related} Architecture search is the problem of optimizing the structure of a neural network to more accurately solve another underlying problem. In essence, the design steps of a neural network architecture that might otherwise be done by an engineer or graduate student by hand are instead automated and optimized as part of a well defined search space of reasonable layers, connections, outputs, and hyperparameters. In fact, architecture search can itself be defined in terms of hyperparameters\cite{hundt2018hypertree} or as a graph search problem\cite{2017nasnet, 2018enas, cai2018proxylessnas, NIPS2016_ConvNeuralFabrics}. Furthermore, once a search space is defined various tools can be brought to bear on the problem including Bayesian optimization\cite{pmlr-v64-mendoza_towards_2016}, other neural networks\cite{smashHyperNetworks}, reinforcement learning, evolution\cite{real2017large, real2018regularized}, or a wide variety of optimization frameworks. A survey for the topic of Neural Architecture Search (NAS) is available at \cite{2018nassurvey}. Differentiable Architecture Search (DARTS)\cite{liu2018darts} defines a search space in terms of parameters $\alpha$, weights $w_i = softmax(\alpha_i)$, and operation layers $op_i$ which are made differentiable in the form $output_i = w_i * op_i(input_i)$. ProxylessNAS\cite{cai2018proxylessnas} works in a similar manner on a MobileNetv2\cite{s2018mobilenetv2} based search space. It only loads two architectures at a time, updating $w$ based on relative changes between them. This saves GPU memory so that ImageNet size datasets and architectures load directly, but at the cost of shuttling whole architectures between the GPU and main memory. \begin{figure}[btp!] \centering \includegraphics[width=1\columnwidth]{sharpsepconv_and_sharp_darts_labeled.pdf} \caption{\small \label{fig:sharpdartscells} (Top) High-level structure of NASNet\cite{2017nasnet} style architectures like sharpDARTS. Normal cells are stride 1 and Reduction cells are stride 2 (Left) The SharpDARTS and SharpSepConvDARTS normal cell. (Right) The SharpDarts and SharpSepConvDARTS reduction cell. } \vspace{-0.2cm} \end{figure} Depthwise Separable Convolutions (SepConv) are a common building block of these searches, and were described as part of the Xception\cite{chollet2017xception} architecture, which improved efficiency on a per-parameter basis compared to its predecessor Inception-v3\cite{szegedy2016rethinking_incptionv3}. It was subsequently used to great effect in MobileNetV2\cite{s2018mobilenetv2}. A SepConv is where an initial convolution is defined in which the number of groups is equal to the number of input channels, followed by a single 1x1 convolution with a group size of 1. This type of convolution tends to have roughly equivalent or better performance than a standard Conv layer with fewer operations and lower memory utilization.\cite{guo2018network, chollet2017xception} Furthermore, so-called ``bottleneck'' layers or blocks have proven useful to limiting the size and improving the accuracy of neural network models\cite{he2015resnet, s2018mobilenetv2}. Augmentation is another fundamental tactic when optimizing the efficiency of neural networks. For example, on CIFAR-10 Cutout\cite{2017cutout} randomly sets 16x16 squares in an input image to zero; and AutoAugment\cite{cubuk2018autoaugment} is demonstrated on PyramidNet\cite{Han2017PyramidNet}, where it applies reinforcement learning to optimize parameter choices for a set of image transforms, and to the odds of applying each transform during training. \begin{table}\centering \setlength\tabcolsep{3pt} \begin{tabular}{lllllr} \toprule \multicolumn{6}{c}{Operations in the sharpDARTS search space} \\ \cmidrule(r){1-6} Name & K & S & D & CMid & CMidMult \\ \cmidrule(r){1-6} none & -- & -- & -- & -- & -- \\ AvgPool3x3 & 1 & 1 & 1 & -- & -- \\ MaxPool3x3 & 1 & 1 & 1 & -- & -- \\ SkipConnect & 1 & 1 & 1 & -- & -- \\ SepConv3x3 & 3 & 1 & 1 & -- & 1 \\ DilConv3x3 & 3 & 1 & 2 & -- & 1 \\ FloodConv3x3 & 3 & 1 & 1 & -- & 4 \\ DilFloodConv3x3 & 3 & 1 & 2 & -- & 4 \\ ChokeConv3x3 & 3 & 1 & 1 & 32 & -- \\ DilChokeConv3x3 & 3 & 2 & 2 & 32 & -- \\ \bottomrule \end{tabular} \caption{\small \label{table:searchspace} Available operations for connections within cells in the sharpDARTS search space. Columns K, S, and D are kernel, stride, and dilation, respectively. All ``Conv'' operations are configurations of \texttt{SharpSepConv}, shown in Fig. \ref{fig:SharpSepConv} and \ref{fig:pytorch_sharpsepconv}. } \vspace{-0.2cm} \end{table} \begin{figure}[bt!] \includegraphics[width=\columnwidth]{SharpSepConvCode.pdf} \caption{ \small \label{fig:pytorch_sharpsepconv} The complete PyTorch 1.0~\cite{paszke2017pytorch} definition of the \texttt{SharpSepConv} block visualized in Fig. \ref{fig:SharpSepConv}.} \vspace{-0.4cm} \end{figure} \begin{figure*}[bt!] \vspace{-0.2cm} \centering \hfill \includegraphics[width=\textwidth]{annealing_comparison.pdf} \caption{ \label{fig:compare_cosine_annealing} \small A comparison of annealing methods. Cosine Power Annealing is more frequently at an optimal learning rate. These 1000 epoch examples of CIFAR-10 sharpDARTS SGD training were selected to compare progress. Here final validation accuracy differs by 0.04\%, but in the typical case Cosine Power Annealing has better performance. } \vspace{-0.4cm} \end{figure*} \begin{figure}[bt!] \centering \hfill \includegraphics[width=\columnwidth]{cos_power_annealing_imagenet.pdf} \includegraphics[width=\columnwidth]{cos_power_annealing_imagenet_log.pdf} \caption{ \label{fig:cosine_power_annealing} \small A comparison of Cosine Annealing~\cite{loshchilov2016sgdrcosineannealing} (Eq. \ref{eq:cosineannealing}) and Cosine Power Annealing (Eq. \ref{eq:cosinepowerannealing} with $p=10$) learning rate schedules for ImageNet. The bottom chart's logarithmic scale demonstrates how the cosine term is most prominent during early and late epochs, while the exponential term $p$ smooths decay through the middle epochs. } \vspace{-0.2cm} \end{figure} However, improvements via the methods above are only valuable if the results are reproducible and generalize to new data, issues which are a growing concern throughout academia. Recht et al.\cite{recht2018cifar10} have investigated such concerns, creating a new CIFAR-10.1 test set selected from the original tiny images dataset from which CIFAR-10 was itself selected. While the rank ordering of evaluated models remains very consistent, their work demonstrates a significant drop in accuracy on the test set relative to the validation set for all models. Variation implicit in the training process, models, and the dataset itself must be carefully considered when reading results; and so a slightly better score does not guarantee better generalization. For reference, on CIFAR-10.1 ``a conservative confidence interval (Clopper-Pearson at confidence level 95\%) for accuracy 90\% has size about $\pm1\%$ with n = 2,000 (to be precise, [88.6\%, 91.3\%])''\cite{recht2018cifar10}. \section{Search Space and Training Methods} We seek better generalization of architecture search using fewer resources. We begin by analyzing DARTS\cite{liu2018darts}, particularly its search space implementation and training regimen. The search space covers a variety of core operations or ``blocks'' which include pooling, skip connections, no layer, separable convolutions and dilated convolutions. Upon deeper analysis, we note that the \texttt{DilConv} operation contains 2 convolution layers while \texttt{SepConv} contains 4, thus comparing blocks with different scales. This scale imbalance matters, and Sec. \ref{sec:ablation_study}, \ref{sec:hypercuboids_and_regularization}, and \ref{sec:grid_search} will explore why, but first we will go through the design choices and data leading up to that conclusion. Therefore, we start with our \texttt{SharpSepConv} block, which helps to correct the aforementioned imbalance, and the Cosine Power Annealing learning rate schedule. \subsection{Sharp Separable Convolution Block} \label{ssec:goalsandencodings} We first define the \texttt{SharpSepConv} block which consists of 2 Separable Convolutions and 2 bottlenecks to balance the number of layers and to capitalize on these components' computational efficiency, as we described in Related Work (Sec. \ref{ref:related}). All of the parameters of \texttt{SharpSepConv} are visualized in Fig. \ref{fig:SharpSepConv} and the block is defined using the PyTorch 1.0\cite{paszke2017pytorch} framework in Fig. \ref{fig:pytorch_sharpsepconv}. We will refer to these figures and the definitions in the Table \ref{table:cifar10} caption throughout the remaining text. The \texttt{SharpSepConv} block permits variation in the dilation parameters and the number of filters contained in the middle layers, both relative to the input and in absolute terms, without changing the number of convolutions. In effect, \texttt{C\_mid} adds fixed-size bottlenecks directly into the architecture search space which can have a large impact on the number of AddMult operations in an architecture and thus the computational efficiency. Furthermore, \texttt{C\_mid\_mult} makes it possible to incorporate an additional reduction or increase in network size as needed within individual cells as can be seen in Fig. \ref{fig:sharpdartscells}. Using \texttt{SharpSepConv} and other operations we create the sharpDARTS search space defined in Table \ref{table:searchspace}. As the ablation study in Sec. \ref{sec:ablation_study} and the figures in Table \ref{table:cifar10} show, \texttt{SharpSepConv} operations substantially contribute to the improvements inherent to the final sharpDARTS and SharpSepConvDARTS architectures. \begin{table*}\centering \ra{1.3} \setlength\tabcolsep{3pt} \begin{tabular}{@{}lccccccccr@{}}\toprule \multicolumn{1}{l}{Architecture} & \multicolumn{1}{c}{Auto} & \multicolumn{1}{c}{Grad} & \multicolumn{1}{c}{SSC} & \multicolumn{1}{c}{Val Error} & \multicolumn{1}{c}{Test Error} & \multicolumn{1}{c}{Par.} & \multicolumn{1}{c}{$+ \times$} & \multicolumn{1}{c}{GPU} & \multicolumn{1}{c}{Algorithm}\\ \multicolumn{1}{l}{} & \multicolumn{1}{c}{Aug} & \multicolumn{1}{c}{Order} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{CIFAR-10} & \multicolumn{1}{c}{CIFAR-10.1} & \multicolumn{1}{c}{M} & \multicolumn{1}{c}{B} & \multicolumn{1}{c}{Days} & \multicolumn{1}{c}{}\\% \cmidrfule{2-2} \cmidrule{4-4} \cmidrule{6-6} \midrule ShakeShake64d\cite{gastaldi2017shakeshake, recht2018cifar10}&\text{\ding{55}}&&\text{\ding{55}}&2.9$\pm$0.3 &7.0$\pm$1.8&26.2&&--&Man\\ AutoAugment\cite{cubuk2018autoaugment,Han2017PyramidNet}&\checkmark&&\text{\ding{55}}& 1.48 && 26.0 && -- & AugRL+Man \\ GPipe AmoebaNet-B\cite{huang2018gpipe,real2018regularized}&\text{\ding{55}}&&\text{\ding{55}}& 1.0 $\pm$ 0.05 && $\sim$557 &$\sim$50&& Evo+Man \\\midrule NASNet-A \cite{2017nasnet} &\text{\ding{55}}&&\text{\ding{55}}& 2.65 && 3.3 && 1800 & RL \\ NASNet-A \cite{2017nasnet}$^\dagger$ &\text{\ding{55}}&&\text{\ding{55}}& 2.83 && 3.8 &0.624& 3150 & RL \\ AmoebaNet-A \cite{real2018regularized} &\text{\ding{55}}&&\text{\ding{55}}& 3.34 $\pm$ 0.06 && 3.2 && 3150 & Evo \\ AmoebaNet-A \cite{real2018regularized}$^\dagger$&\text{\ding{55}}&&\text{\ding{55}}& 3.12 && 3.1 &0.506& 3150 & Evo \\ AmoebaNet-B \cite{real2018regularized}&\text{\ding{55}}&&\text{\ding{55}}& 2.55 $\pm$ 0.05 && 2.8 && 3150 & Evo \\ ProgressiveNAS \cite{liu2018PNAS}&\text{\ding{55}}&&\text{\ding{55}}& 3.41 $\pm$ 0.09 && 3.2 && 225 & SMBO \\ ENAS \cite{2018enas}&\text{\ding{55}}&&\text{\ding{55}}& 2.89 && 4.6 && 0.5 & RL \\ DARTS random~\cite{liu2018darts}&\text{\ding{55}}&&\text{\ding{55}}& 3.49 && 3.1 && -- & -- \\ DARTS$^\dagger$~\cite{liu2018darts}&\text{\ding{55}}&1&\text{\ding{55}}& 2.94 && 2.9 &0.518& 1.5 &Grad\\ DARTS$^\dagger$~\cite{liu2018darts}&\text{\ding{55}}&2&\text{\ding{55}}& 2.83 $\pm$ 0.06 && 3.4 &0.547& 4 &Grad\\ ProxylessNAS~\cite{cai2018proxylessnas}&\text{\ding{55}}&1&\text{\ding{55}}& 2.08 && 5.7 && 4 & Grad\\ \midrule \textbf{SharpSepConvDARTS}&\checkmark&\textbf{1}&\checkmark& \textbf{1.98$\pm$0.07} &\textbf{5.9$\pm$0.5}& \textbf{3.6} &\textbf{0.579}&\textbf{0.8}&\textbf{Grad}\\ \textbf{DARTS+SSC, no search}&\checkmark&\textbf{2}&\checkmark& \textbf{2.05$\pm$0.14} &\textbf{5.6$\pm$0.8}& \textbf{3.5} &\textbf{0.576}& -- &\textbf{Grad}\\ \textbf{sharpDARTS}&\checkmark&\textbf{1}&\checkmark& \textbf{2.29$\pm$0.06} &\textbf{6.2$\pm$0.4}& \textbf{1.98} &\textbf{0.357}& \textbf{1.8} &\textbf{Grad} \\ SharpSepConvDARTS&\text{\ding{55}}&1&\checkmark& 2.45 && 3.6 &0.579& 0.8 &Grad\\ DARTS+SSC, no search&\text{\ding{55}}&2&\checkmark& 2.55 && 3.5 &0.576& -- &Grad\\ sharpDARTS&\text{\ding{55}}&1&\checkmark& 2.97 && 1.98 &0.357& 1.8 &Grad\\ \bottomrule \bottomrule \end{tabular} \caption{\small \label{table:cifar10} Results for the CIFAR-10 dataset and the CIFAR-10.1\cite{recht2018cifar10} test set, \textbf{bold lines} are trained with AutoAugment\cite{cubuk2018autoaugment} and Cosine Power Annealing (Eq. \ref{eq:cosinepowerannealing}) where we set $p=2$, $\eta^0_{max}=0.025$, $\eta^0_{min}=1\text{e-}8$, and $T_0=2000$ epochs of training for our best results. Our figures with a range in the bottom section are defined as $median \pm range/2$ and include at least 3 runs with Cosine Power Annealing. Dagger$^\dagger$ indicates genotypes available in our accompanying repository. Non-bold numbers have 1000 epochs of Cosine Power Annealing and AutoAugment disabled to better match figures from past work. The code and this table is derived from DARTS\cite{liu2018darts}. The \textbf{GradOrder} column is 1 for standard gradients, 2 for Hessians on DARTS based algorithms. All algorithms utilize \textbf{CutOut}\cite{2017cutout} except for ProgressiveNAS. A \textbf{``genotype''} simply defines an architecture as a graph. It has a list of layers to use, and the previous node, which it should use as an input. A \textbf{``primitive''} is an operation aka block of layers that can be chosen during the search. A \textbf{``node''} takes the output of two instances of an (input, primitive) pair as input and adds them, as seen in Fig. \ref{fig:sharpdartscells}. \textbf{DARTS+SSC, no search} compares the genotype published by DARTS with \texttt{SharpSepConv} layers in the primitives. \textbf{SharpSepConvDARTS} is the same as DARTS + SharpSepConv but uses a genotype determined by a new search incorporating the DARTS primitives with SharpSepConv blocks. sharpDARTS uses the architecture search space described in Table \ref{table:searchspace}.} \vspace{-0.2cm} \end{table*} \begin{table*}\centering \ra{1.3} \setlength\tabcolsep{2pt} \centering \begin{tabular}{@{}lccccccc@{}} \toprule \multirow{2}{*}{\textbf{Architecture}} & \multicolumn{2}{c}{\textbf{\% Test Error}} & \textbf{Params} & \textbf{$+ \times$} & \textbf{Search Cost} & \textbf{Search} \\ \cline{2-3} & top-1 & top-5 & \textbf{M} & \textbf{B} & \textbf{GPU Days} & \textbf{Method} \\ \midrule GPipe AmoebaNet-B \cite{real2018regularized,huang2018gpipe} & 15.7 & 3.0 & 557 & $\sim$50 & 3150 & Evo+Man \\ MobileNetv2 \cite{s2018mobilenetv2} & 28.0 & -- & 3.4 & 0.300 & -- & Man \\ MobileNetv2 (1.4) \cite{s2018mobilenetv2} & 25.3 & -- & 6.9 & 0.585 & -- & Man \\ \midrule NASNet-A \cite{2017nasnet} & 26.0 & 8.4 & 5.3 & 0.564 & 1800 & RL \\ NASNet-B \cite{2017nasnet} & 27.2 & 8.7 & 5.3 & 0.488 & 1800 & RL \\ AmoebaNet-B \cite{real2018regularized} & 26.0 & 8.5 & 5.3 & 0.555 & 3150 & Evolution \\ AmoebaNet-C \cite{real2018regularized} & 24.3 & 7.6 & 6.4 & 0.570 & 3150 & Evolution \\ ProgressiveNAS \cite{liu2018PNAS} & 25.8 & 8.1 & 5.1 & 0.588 & $\sim$225 & SMBO \\ DARTS \cite{liu2018darts} & 26.9 & 9.0 & 4.9 & 0.595 & 4 & Grad \\ ProxylessNAS \cite{cai2018proxylessnas} & 24.9 & 8.1 & 5.1 & 0.581 & $\sim$8 & Grad \\\midrule \textbf{SharpSepConvDARTS} & \textbf{25.1} & \textbf{7.8} & \textbf{4.9} & \textbf{0.573} & \textbf{0.8} & \textbf{Grad} \\ \textbf{sharpDARTS cmid 96} & \textbf{27.8} & \textbf{9.2} & \textbf{3.71} & \textbf{0.481} & \textbf{1.8} & \textbf{Grad} \\ \textbf{sharpDARTS cmid 32} & \textbf{31.3} & \textbf{11.34} & \textbf{3.17} & \textbf{0.370} & \textbf{1.8} & \textbf{Grad} \\ \bottomrule \end{tabular} \caption{\small Mobile ImageNet Architecture Comparison, table updated from DARTS\cite{liu2018darts}. Lower values are better in all columns. A dash indicates data was either not available or not applicable. \label{table:imagenet}} \vspace{-0.2cm} \end{table*} \subsection{Cosine Power Annealing} Cosine Annealing\cite{loshchilov2016sgdrcosineannealing} is a method of adjusting the learning rate over time, reproduced below: \begin{eqnarray} \label{eq:cosineannealing} \eta_t = \eta^i_{min} + \frac{1}{2}(\eta^i_{max} - \eta^i_{min}){ \left(1 + \cos\left(\pi \frac{T_{cur}}{T_i}\right)\right)} \end{eqnarray} Where $\eta^i_{min}$ and $\eta^i_{max}$ are the minimum and maximum learning rates, respectively; $T_i$ is the total number of epochs; $T_{cur}$ is the current epoch; and $i$ is the index into a list of these parameters for a sequence of warm restarts in which $\eta^i_{max}$ typically decays. This schedule has been widely adopted and it is directly implemented in PyTorch 1.0~\cite{paszke2017pytorch} without warm restarts, where $i = 0$. The Cosine Annealing schedule works very well and is employed by DARTS. However, as can be seen in Fig. \ref{fig:compare_cosine_annealing}, the average time between improvements is above 2\% of the total runtime between epochs 300 and 700. This is an imbalance in the learning rate, which is an artifact of the initial slow decay rate of Cosine Annealing followed by the rapid relative decay in learning rate. We mitigate this imbalance by introducing a power curve parameter $p$ into the algorithm which we call \textbf{Cosine Power Annealing:} \begin{equation} \label{eq:cosinepowerannealing} \eta_t = \eta^i_{min} + (\eta^i_{max} - \eta^i_{min})\frac{p^{\frac{1}{2} \left(1 + \cos\left(\pi \frac{T_{cur}}{T_i}\right)\right) + 1}-p}{p^2-p} \end{equation} The introduction of the normalized exponential term $p$ permits tuning of the curve's decay rate such that it maintains a high learning rate for the first few epochs, while simultaneously taking a shallower slope during the final third of epochs. These two elements help reduce the time between epochs in which validation accuracy improves. A comparison is provided in Fig. \ref{fig:compare_cosine_annealing} and \ref{fig:cosine_power_annealing}. In our implementation we also define a special case for the choice of $p=1$ such that it falls back to standard cosine annealing. This algorithm is also compatible with decaying warm restarts, but we leave that schedule out of scope for the purposes of this paper. \subsection{Results} \label{sec:results} \textbf{CIFAR-10 and CIFAR-10.1:} Our absolute top performance is \textbf{SharpSepConvDARTS} with 1.93\% CIFAR-10 top-1 validation error (1.98$\pm$0.07) and 5.925$\pm$0.48 CIFAR-10.1 test error (Table \ref{table:cifar10}) when including our improved training regimen. To the best of our knowledge, this is state of the art performance for mobile-scale ($\sim$600M AddMult ops\cite{liu2018darts}) and ProxylessNAS\cite{cai2018proxylessnas} is in second at 2.08\% val error. We also show a statistically significant\cite{recht2018cifar10} improvement over ShakeShake64d\cite{recht2018cifar10,gastaldi2017shakeshake} which is the best available CIFAR-10.1 model, with 7.0$\pm$1.2 test error. GPipe AmoebaNet-B\cite{huang2018gpipe} remains the best at any scale with 1\% val error. It is scaled up from the original AmoebaNet-B and we expect other models to scale in a similar way. This truly massive model cannot load on typical GPUs due to 557M parameters and billions of AddMult ops. It runs with specialized software across multiple Google TPU hardware devices. Our training time\footnote{This research was conducted on 5 Nvidia GPU types: Titan X, GTX 1080 Ti, GTX 1080, Titan XP, and the RTX 2080Ti.} for a 60 epoch search is 0.8-1.2 days, with 2k epochs of mixed fp16 training of a final model in 1.7-2.8 days, totaling 2.9-3.6 GPU-days end-to-end on one RTX 2080Ti. The discrepancy in totals is because the slower sharpDARTS search finds a faster final model. \textbf{ImageNet:} Our top \textbf{SharpSepConvDARTS} model achieved 25.1\% top-1 and 7.8\% top-5 error, which is competitive with other state of the art mobile-scale models depicted in Table \ref{table:imagenet}, and this translates to relative improvements over DARTS of 7\% in top-1 error, 13\% in top-5 error, and 80\% in search time. Our ImageNet model uses the genotype of the CIFAR-10 search and follows the same cell based architecture of \cite{2017nasnet, liu2018darts, 2018enas} with different operations in our search space. We apply random cropping to 224x224, random horizontal flipping, AutoAugment\cite{cubuk2018autoaugment}, normalization to the dataset mean and std deviation, and finally Cutout\cite{2017cutout} with a cut length of 112x112. Training of final models was done on 2x GTX 2080Ti in 16 bit mixed precision mode and takes 4-6 days, which is 8-12 GPU-days, depending on the model. \section{Towards Better Generalization} \label{sec:towards_better_generalization} DARTS search improved results over random search by 19\% (Table \ref{table:cifar10}), and by adding our training regimen and search space improvements we get an additional 30\% relative improvement over DARTS. Manual changes like those made to the \textbf{SharpSepConv} block and to AmoebaNet-B for GPipe\cite{huang2018gpipe} are not represented in any search space, and yet they directly lead to clear improvements in accuracy. So why aren't they accounted for? Let's assume that it is possible to encode all of these elements and more into a single, broader search space in which virtually every neural network graph imaginable is encoded by hyperparameters. To even imagine tackling a problem of this magnitude, we must first ask ourselves an important question: Does DARTS even generalize to other search domains designed with this challenge in mind? In this section, we provide a preliminary exploration to begin answering these questions. \begin{figure*}[btp!] \centering \hfill \includegraphics[width=0.95\textwidth]{multi_channel_net_2x2_x2x2x2.pdf} \caption{\small \label{fig:multichannel} A HyperCuboid with shape HC = (filter\_scales, normal\_layers, reduction\_layers, layer\_types) = ((2x2)x2x2x2) for Differentiable Hyperparameter Grid Search (Sec. \ref{sec:grid_search}). The data is ordered $(channels, width, height)$. An example single-path model is highlighted with red arrows. Here you can see a small scale n-tuple coordinate system which might serve as the conceptual basis for future large scale search spaces. (Sec. \ref{sec:towards_better_generalization})} \vspace{-0.6cm} \end{figure*} \subsection{Differentiable Grid Search} \label{sec:grid_search} We introduce \textbf{Differentiable Hyperparameter Grid Search} which is run on a \textbf{HyperCuboid Search Space} parameterized by an n-tuple, such as $HC = (\gamma,d_n,d_r,p)$, where $\gamma$ is an arbitrary set of hyperparameters; $d_n$ is the number layers in one block; $d_r$ is the number of layer strides; and $p$ is an arbitrary set of choices, primitives in this case. In the HyperCuboid Search Space many possible paths pass through each node, and all final sequential paths in the Directed Acyclic Graph of architecture weights are of equal length. The number of architecture weights $w$ in the HyperCuboid graph is the product of the size of the tuple elements; and the number of hyperparameters, primitives, and tuple size can vary in this design. We test a specific HyperCuboid called \textbf{MultiChannelNet} with a tuple (filter\_scale\_pairs, normal\_layer\_depth, reduction\_layer\_depth, primitives), visualized at a small ((2x2)x2x2x2) scale in Fig. \ref{fig:multichannel} with SharpSepConv and MaxPool primitives. Here $\gamma=2^n\times2^m|n,m\in[5,6]$ are combinations of possible input and output filter scales. Our actual search has dimension ((4x4)x3x3x2)$|n,m\in[5..8]$ where final linear paths have 14 nodes, since ``Add'' nodes are excluded. In MultiChannelNet, graph nodes determine which filter input, filter output, and layer type should be utilized. Larger weights $w$ imply a better choice of graph node and we therefore search for an optimal path through a sequence of primitive weights which maximizes the total path score. We construct a basic handmade model in one shot with \texttt{SharpSepConv} and 32 filters, doubling the number of output filters at layers of stride 2 up to the limit of 256. We also ran an automated DARTS search for 60 epochs and about 16 GPU-Hours to find an optimal model. To our surprise, the handmade model outperforms DARTS by over 2\%, as Table \ref{table:multichannelnet_search_algorithms} indicates. Why might this be? To answer this we return to our original NASNet based search space to look for discrepancies in an ablation study. \begin{table}\centering \vspace{0.2cm} \setlength\tabcolsep{2pt} \begin{tabular}{lrrcc} \toprule Weights/Path & Par. & $+ \times$ & Val Err \% & Test Err \% \\ \cmidrule(r){1-5} Scalar&0.8M& 23M& 8.62$\pm$0.16& 17.2$\pm$0.8 \\ Handmade& 1.0M&25M& 6.50$\pm$0.18& 13.9$\pm$0.4 \\ \textbf{Max-W}&\textbf{0.9M}&\textbf{31M}&\textbf{5.44$\pm$0.12}&\textbf{12.4$\pm$1.0}\\ \bottomrule \end{tabular} \caption{\small \label{table:multichannelnet_search_algorithms} CIFAR-10 and CIFAR-10.1 results for model paths from our MultiChannelNet (Fig. \ref{fig:multichannel}) example of the HyperCuboid Search Space. There is a statistically significant\cite{recht2018cifar10} improvement from the original Scalar DARTS to our proposed Max-W Weight Regularization DARTS. We also see the hypothesized increase in accuracy, parameters, and AddMult ($+ \times$) flops described in Sec. \ref{sec:hypercuboids_and_regularization} and \ref{sec:grid_search}. Final training is $\sim$0.3 days on an RTX 2080Ti. } \vspace{-0.3cm} \end{table} \subsection{Ablation Study} \label{sec:ablation_study} Figures for our ablation study are in Table \ref{table:cifar10} which also has a description of our preprocessing changes in the caption. We will analyze the 3 main model configurations below: \textbf{(1)} \textbf{DARTS+SSC} directly replaces all convolution primitives in DARTS\cite{liu2018darts} with a \texttt{SharpSepConv} layer where the block parameters, primitives, and the genotype are otherwise held constant; we see a 10\% relative improvement over DARTS val err (2.55\% vs 2.83\%) with only 5\% more AddMult operations and no additional augmentation. \textbf{(2)} \textbf{SharpSepConvDARTS} is the same as item 1 but with a 1st order gradient DARTS search; we see relative improvements of 13\% in val err (2.45\% vs 2.83\%) and 80\% in search time. \textbf{(3)} \textbf{sharpDARTS} (Table \ref{table:searchspace}) has slightly different primitives including \textit{flood} where middle channels expand 4x, \textit{choke} with a fixed 32 middle filters, and one of each conv with a \textit{dilation} of 2. Our sharpDARTS model (Fig. \ref{fig:sharpdartscells}) with our improved training regimen achieved an absolute best error of 2.27\%. Without training enhancements the 1st order gradient search of sharpDARTS has similar accuracy to 1st order DARTS and is definitively more efficient with 32\% fewer parameters, 31\% fewer AddMult operations, and lower memory requirements. However, the absolute accuracy of sharpDARTS is \textit{marginally lower} than the original DARTS, and also suffers from a larger disparity on ImageNet (Table \ref{table:imagenet}). This is startling for two reasons: (1) Substituting the \texttt{SharpSepConv} block improves accuracy in DARTS+SSC and SharpSepConvDARTS. (2) The sharpDARTS search space still contains all primitives needed to represent both the final DARTS+SSC and SharpSepConvDARTS model genotypes perfectly. We've replicated a discrepancy in DARTS behavior across two different search spaces, so the most likely remaining possibility must be a limitation in the DARTS search method itself. \subsection{Max-W Regularization} \label{sec:hypercuboids_and_regularization} GPipe\cite{huang2018gpipe} shows AmoebaNet-B models improving in accuracy as they scale to $>$500M parameters and billions of AddMult flops (Fig. \ref{table:cifar10}, \ref{table:imagenet}). These models are from a search space similar to those used by DARTS, among others\cite{liu2018PNAS,2018enas,2017nasnet}. Assume that the GPipe scaling principle holds for similar training configurations and search spaces, and one might expect that DARTS would tend towards larger models throughout the search process. However, during the early epochs of training, DARTS reliably produces models composed entirely of max pools and skip connects. These are among the smallest primitives in the DARTS architecture search space with respect to parameters and AddMult operations. Higher capacity layers are chosen later in the search process, as visualized in the animation included with the original DARTS source code\footnote{\scriptsize DARTS\cite{liu2018darts} model search animation for reduce cells: \url{https://git.io/fjfTC} normal cells: \url{https://git.io/fjfbT}}. We experimented with removing max pool layers and the undesired behavior simply shifts to skip connects. We posit that the \textbf{DARTS Scalar Weighting} $output_i = w_i * op_i(input_i)$ tends towards the layers with the maximum gradient, and thus models will consist of smaller layers than appropriate. Such bias during early phases of training is inefficient with respect to optimal accuracy, even if models might eventually converge to larger, more accurate models after a long period of training. Therefore, we hypothesize that subtracting the maximum weight in a given layer will regularize weight changes via \textbf{Max-W Weighting:} \begin{equation} output_i = (1 - max(w) + w_i) * op_i(input_i) \label{eqn:max_w} \end{equation} Intuitively, consider the architecture parameters $\alpha$, weights $w=softmax(\alpha) | w\in[0,1]$, highest score index $i_{max}=argmax(w)$, and the highest score weight $w_{max}=w[i_{max}]$ at some arbitrary time during training. If we apply Max-W weighting to the layer corresponding with $w_{max}$, we will have $(1 - w_{max} + w_{max}) * op_i(input_i)$ which reduces to $output_i = 1 * op_i(input_i)|i=i_{max}$. In this case the $\alpha$ underlying $w_{max}$ will remain unchanged, but this is not true for other values $w_i| i\neq i_{max}$. Here it will be incumbent on non-maximum layers $w_i$ to outperform the highest score weight and grow their value $\alpha_i$. As values other than $w_{max}$ grow, $w_{max}$ will naturally drop in accordance with the behavior of $softmax$. This has the net effect of reducing bias corresponding to the highest score layer $op_i$ paired with $w[i_{max}]$. Our search with Max-W weighting on MultiChannelNet found a model which is both larger and significantly more accurate than both the original DARTS Scalar search models and a hand designed model (Table \ref{table:multichannelnet_search_algorithms}). These results indicate that our initial hypothesis holds and Max-W DARTS (Eq. \ref{eqn:max_w}) is an effective approach to regularization when compared to standard Scalar DARTS. Specific models will be released with the code. \section{Future Work} Our investigation also indicates several other areas for future work. The SharpSepConvDARTS and sharpDARTS search spaces might also benefit from Max-W regularization, so it is an interesting topic for an additional ablation study. MultiChannelNet and sharpDARTS indicate that the final DARTS model\cite{liu2018darts} did not fully converge to the optimum due to the scalar weighting bias (Sec. \ref{sec:ablation_study}, \ref{sec:hypercuboids_and_regularization}). We suspect other DARTS algorithms, such as ProxylessNAS\cite{cai2018proxylessnas}, suffer from this same bias. We have also shown how Max-W Regularization correctly chooses larger models when those are more accurate, however this means a search with Max-W DARTS currently exceeds mobile-scale on the DARTS and sharpDARTS search spaces. Adding a resource cost based on time and memory to each node might make it possible to directly optimize cost-accuracy tradeoffs with respect to a specific budget. Arbitrary multi-path subgraphs respecting this budget could be chosen by iterative search with a graph algorithm like network simplex\cite{orlin1997polynomialnetworksimplex}. Other alternatives include a reinforcement learning algorithm or differentiable metrics like the latency loss in ProxylessNAS\cite{cai2018proxylessnas}. \section{Conclusion} In this paper we met or exceeded state of the art mobile-scale architecture search performance on CIFAR-10, CIFAR-10.1 and ImageNet with a new \texttt{SharpSepConv} block. We introduced the Cosine Power Annealing learning rate schedule, which is more often at an optimal learning rate than Cosine Annealing alone, and demonstrated an improved sharpDARTS training regimen. Finally, we introduced Differentiable Hyperparameter Grid Search with a HyperCuboid search space to reproduce bias within the DARTS search method, and demonstrated how Max-W regularization of DARTS corrects that imbalance. Finally, Differentiable Hyperparameter Search and HyperCuboids might be evaluated more broadly in the computer vision space and on other topics such as recurrent networks, reinforcement learning, natural language processing, and robotics. For example, \texttt{SharpSepConv} is manually designed so a new HyperCuboid might be constructed to empirically optimize the number, sequence, type, layers, activations, normalization, and connections within a block. Perhaps a future distributed large scale model might run Differentiable Hyperparameter Search over hundreds of hyperparameters which embed a superset of search spaces, making it possible to efficiently and automatically find new models for deployment to any desired application. {\small \bibliographystyle{style/ieee} \section{Introduction} Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press. This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version. \subsection{Language} All manuscripts must be in English. \subsection{Dual submission} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the policy on dual submissions. \subsection{Paper length} Papers, excluding the references section, must be no longer than eight pages in length. The references section will not be included in the page count, and there is no limit on the length of the references section. For example, a paper of eight pages with two pages of references would have a total length of 10 pages. Overlength papers will simply not be reviewed. This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts. The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven. \subsection{The ruler} The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review. The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution. If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages. The presence or absence of the ruler should not change the appearance of any other content on the page. The camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment the \verb'\iccvfinalcopy' command in the document preamble.) Reviewers: note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly. Just use fractional references (e.g.\ this line is $095.5$), although in most cases one would expect that the approximate location will be adequate. \subsection{Mathematics} Please number all of your sections and displayed equations. It is important for readers to be able to refer to any particular equation. Just because you didn't refer to it in the text doesn't mean some future reader might not need to refer to it. It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''. (Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers). All authors will benefit from reading Mermin's description of how to write mathematics: \url{http://www.pamitc.org/documents/mermin.pdf}. \subsection{Blind review} Many authors misunderstand the concept of anonymizing for blind review. Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available. Blind review means that you do not use the words ``my'' or ``our'' when citing previous work. That is all. (But see below for techreports.) Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith; it says that you are building on her work. If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work. An example of a bad paper just asking to be rejected: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Removed for blind review \end{quote} An example of an acceptable paper: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213. \end{quote} If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work. In such cases, include the anonymized parallel submission~\cite{Authors14} as additional material and cite it as \begin{quote} [1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as additional material {\tt fg324.pdf}. \end{quote} Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report. For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a techreport for further details. Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''. Then submit the techreport as additional material. Again, you may not assume the reviewers will read this material. Sometimes your paper is about a problem which you tested using a tool which is widely known to be restricted to a single institution. For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the ICCV70 audience would like to hear about your solution. The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal. You can handle this paper like any other. Don't write ``We show how to improve our previous work [Anonymous, 1968]. This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''. That would be silly, and would immediately identify the authors. Instead write the following: \begin{quotation} \noindent We describe a system for zero-g frobnication. This system is new because it handles the following cases: A, B. Previous systems [Zeus et al. 1968] didn't handle case B properly. Ours handles it by including a foo term in the bar integral. ... The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know. It displayed the following behaviours which show how well we solved cases A and B: ... \end{quotation} As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors. A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess. He or she would have to be sure that no other authors could have been contracted to solve problem B. \noindent FAQ\medskip\\ {\bf Q:} Are acknowledgements OK?\\ {\bf A:} No. Leave them for the final copy.\medskip\\ {\bf Q:} How do I cite my results reported in open challenges? {\bf A:} To conform with the double blind review policy, you can report results of other challenge participants together with your results in your paper. For your results, however, you should not identify yourself and should not mention your participation in the challenge. Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\ \begin{figure}[t] \begin{center} \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}} \end{center} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{fig:long} \label{fig:onecol} \end{figure} \subsection{Miscellaneous} \noindent Compare the following:\\ \begin{tabular}{ll} \verb'$conf_a$' & $conf_a$ \\ \verb'$\mathit{conf}_a$' & $\mathit{conf}_a$ \end{tabular}\\ See The \TeX book, p165. The space after \eg, meaning ``for example'', should not be a sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided \verb'\eg' macro takes care of this. When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.) However, use it only when there are three or more authors. Thus, the following is correct: `` Frobnication has been trendy lately. It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.'' This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors. If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal. For this citation style, keep multiple citations in numerical (not chronological) order, so prefer \cite{Alpher03,Alpher02,Authors14} to \cite{Alpher02,Alpher03,Authors14}. \begin{figure*} \begin{center} \fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \end{center} \caption{Example of a short caption, which should be centered.} \label{fig:short} \end{figure*} \section{Formatting your paper} All text must be in a two-column format. The total allowable width of the text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the first page) should begin 1.0 inch (2.54 cm) from the top edge of the page. The second and following pages should begin 1.0 inch (2.54 cm) from the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the page. \subsection{Margins and page numbering} All printed material, including text, illustrations, and charts, must be kept within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm) high. Page numbers should be in footer with page numbers, centered and .75 inches from the bottom of the page and make it start at the correct page number rather than the 4321 in the example. To do this fine the line (around line 23) \begin{verbatim} \setcounter{page}{4321} \end{verbatim} where the number 4321 is your assigned starting page. Make sure the first page is numbered by commenting out the first page being empty on line 46 \begin{verbatim} \end{verbatim} \subsection{Type-style and fonts} Wherever Times is specified, Times Roman may also be used. If neither is available on your word processor, please use the font closest in appearance to Times to which you have access. MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of the first page. The title should be in Times 14-point, boldface type. Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs; do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word). Leave two blank lines after the title. AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title and printed in Times 12-point, non-boldface type. This information is to be followed by two blank lines. The ABSTRACT and MAIN TEXT are to be in a two-column format. MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use double-spacing. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422 cm). Make sure your text is fully justified---that is, flush left and flush right. Please do not place any additional blank lines between paragraphs. Figure and table captions should be 9-point Roman type as in Figures~\ref{fig:onecol} and~\ref{fig:short}. Short captions should be centred. \noindent Callouts should be 9-point Helvetica, non-boldface type. Initially capitalize only the first word of section titles and first-, second-, and third-order headings. FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after. SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after. If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line. \subsection{Footnotes} Please use footnotes\footnote {This is what a footnote looks like. It often distracts the reader from the main flow of the argument.} sparingly. Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence). If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced. Use Times 8-point type, single-spaced. \subsection{References} List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Authors14}. Where appropriate, include the name(s) of editors of referenced books. \begin{table} \begin{center} \begin{tabular}{|l|c|} \hline Method & Frobnability \\ \hline\hline Theirs & Frumpy \\ Yours & Frobbly \\ Ours & Makes one's heart Frob\\ \hline \end{tabular} \end{center} \caption{Results. Ours is better.} \end{table} \subsection{Illustrations, graphs, and photographs} All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the paper. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your paper in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage[dvips]{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.eps} \end{verbatim} } \subsection{Color} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the use of color in your document. \section{Final copy} You must include your signed IEEE copyright release form when you submit your finished paper. We MUST have this form before your paper can be published in the proceedings. {\small \bibliographystyle{ieee} \section{Introduction} Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press. This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version. \subsection{Language} All manuscripts must be in English. \subsection{Dual submission} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the policy on dual submissions. \subsection{Paper length} Papers, excluding the references section, must be no longer than eight pages in length. The references section will not be included in the page count, and there is no limit on the length of the references section. For example, a paper of eight pages with two pages of references would have a total length of 10 pages. Overlength papers will simply not be reviewed. This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts. The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven. \subsection{The ruler} The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review. The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution. If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages. The presence or absence of the ruler should not change the appearance of any other content on the page. The camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment the \verb'\iccvfinalcopy' command in the document preamble.) Reviewers: note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly. Just use fractional references (e.g.\ this line is $095.5$), although in most cases one would expect that the approximate location will be adequate. \subsection{Mathematics} Please number all of your sections and displayed equations. It is important for readers to be able to refer to any particular equation. Just because you didn't refer to it in the text doesn't mean some future reader might not need to refer to it. It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''. (Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers). All authors will benefit from reading Mermin's description of how to write mathematics: \url{http://www.pamitc.org/documents/mermin.pdf}. \subsection{Blind review} Many authors misunderstand the concept of anonymizing for blind review. Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available. Blind review means that you do not use the words ``my'' or ``our'' when citing previous work. That is all. (But see below for techreports.) Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith; it says that you are building on her work. If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work. An example of a bad paper just asking to be rejected: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Removed for blind review \end{quote} An example of an acceptable paper: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213. \end{quote} If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work. In such cases, include the anonymized parallel submission~\cite{Authors14} as additional material and cite it as \begin{quote} [1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as additional material {\tt fg324.pdf}. \end{quote} Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report. For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a techreport for further details. Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''. Then submit the techreport as additional material. Again, you may not assume the reviewers will read this material. Sometimes your paper is about a problem which you tested using a tool which is widely known to be restricted to a single institution. For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the ICCV70 audience would like to hear about your solution. The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal. You can handle this paper like any other. Don't write ``We show how to improve our previous work [Anonymous, 1968]. This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''. That would be silly, and would immediately identify the authors. Instead write the following: \begin{quotation} \noindent We describe a system for zero-g frobnication. This system is new because it handles the following cases: A, B. Previous systems [Zeus et al. 1968] didn't handle case B properly. Ours handles it by including a foo term in the bar integral. ... The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know. It displayed the following behaviours which show how well we solved cases A and B: ... \end{quotation} As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors. A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess. He or she would have to be sure that no other authors could have been contracted to solve problem B. \noindent FAQ\medskip\\ {\bf Q:} Are acknowledgements OK?\\ {\bf A:} No. Leave them for the final copy.\medskip\\ {\bf Q:} How do I cite my results reported in open challenges? {\bf A:} To conform with the double blind review policy, you can report results of other challenge participants together with your results in your paper. For your results, however, you should not identify yourself and should not mention your participation in the challenge. Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\ \begin{figure}[t] \begin{center} \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}} \end{center} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{fig:long} \label{fig:onecol} \end{figure} \subsection{Miscellaneous} \noindent Compare the following:\\ \begin{tabular}{ll} \verb'$conf_a$' & $conf_a$ \\ \verb'$\mathit{conf}_a$' & $\mathit{conf}_a$ \end{tabular}\\ See The \TeX book, p165. The space after \eg, meaning ``for example'', should not be a sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided \verb'\eg' macro takes care of this. When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.) However, use it only when there are three or more authors. Thus, the following is correct: `` Frobnication has been trendy lately. It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.'' This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors. If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal. For this citation style, keep multiple citations in numerical (not chronological) order, so prefer \cite{Alpher03,Alpher02,Authors14} to \cite{Alpher02,Alpher03,Authors14}. \begin{figure*} \begin{center} \fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \end{center} \caption{Example of a short caption, which should be centered.} \label{fig:short} \end{figure*} \section{Formatting your paper} All text must be in a two-column format. The total allowable width of the text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the first page) should begin 1.0 inch (2.54 cm) from the top edge of the page. The second and following pages should begin 1.0 inch (2.54 cm) from the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the page. \subsection{Margins and page numbering} All printed material, including text, illustrations, and charts, must be kept within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm) high. \subsection{Type-style and fonts} Wherever Times is specified, Times Roman may also be used. If neither is available on your word processor, please use the font closest in appearance to Times to which you have access. MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of the first page. The title should be in Times 14-point, boldface type. Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs; do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word). Leave two blank lines after the title. AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title and printed in Times 12-point, non-boldface type. This information is to be followed by two blank lines. The ABSTRACT and MAIN TEXT are to be in a two-column format. MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use double-spacing. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422 cm). Make sure your text is fully justified---that is, flush left and flush right. Please do not place any additional blank lines between paragraphs. Figure and table captions should be 9-point Roman type as in Figures~\ref{fig:onecol} and~\ref{fig:short}. Short captions should be centred. \noindent Callouts should be 9-point Helvetica, non-boldface type. Initially capitalize only the first word of section titles and first-, second-, and third-order headings. FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after. SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after. If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line. \subsection{Footnotes} Please use footnotes\footnote {This is what a footnote looks like. It often distracts the reader from the main flow of the argument.} sparingly. Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence). If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced. Use Times 8-point type, single-spaced. \subsection{References} List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Authors14}. Where appropriate, include the name(s) of editors of referenced books. \begin{table} \begin{center} \begin{tabular}{|l|c|} \hline Method & Frobnability \\ \hline\hline Theirs & Frumpy \\ Yours & Frobbly \\ Ours & Makes one's heart Frob\\ \hline \end{tabular} \end{center} \caption{Results. Ours is better.} \end{table} \subsection{Illustrations, graphs, and photographs} All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the paper. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your paper in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage[dvips]{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.eps} \end{verbatim} } \subsection{Color} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the use of color in your document. \section{Final copy} You must include your signed IEEE copyright release form when you submit your finished paper. We MUST have this form before your paper can be published in the proceedings. {\small \bibliographystyle{ieee} \section{Introduction} \begin{figure}[bt!] \centering \hfill \includegraphics[width=\columnwidth]{SharpSepConv.pdf} \caption{\small \label{fig:SharpSepConv} Our \texttt{SharpSepConv} Block configured with 3x3 kernels and stride 1. All convolutions are either 1x1 convolutions or Depthwise Separable Convolutions\cite{chollet2017xception}. The \texttt{c\_mid} hyperparameter can be either proportional to \texttt{c\_out} or set to an arbitrary constant. } \vspace{-0.2cm} \end{figure} Neural Architecture Search (NAS) promises to automatically and efficiently optimize a complex model based on representative data so that it will generalize, and thus make accurate predictions on new examples. Recent NAS results have been impressive, notably in computer vision, but the search process took large amounts of computing infrastructure~\cite{2017nasnet,zoph2017neural,liu2018PNAS}. These methods were followed up with a multiple order of magnitude increase in search efficiency~\cite{2018enas,liu2018darts}. However, key questions remain: What makes a search space worth exploring? Is each model visited in a search space getting a fair shot? We investigate how to improve the search space of DARTS\cite{liu2018darts}, which is one of several based on NASNet\cite{2017nasnet,real2018regularized,2018enas,liu2018PNAS,huang2018gpipe}, propose Differentiable Hyperparameter Grid Search with DARTS, and draw conclusions that apply across architecture search spaces. \noindent To summarize, we make the following contributions: \begin{enumerate} \item Define the novel \texttt{SharpSepConv} block with a more consistent structure of model operations and an adaptable middle filter count in addition to the sharpDARTS architecture search space. This leads to highly parameter-efficient results which match or beat state of the art performance for mobile-scale architectures on CIFAR-10, CIFAR-10.1, and ImageNet with respect to Accuracy, AddMult operations, and GPU search hours. \item Introduce the Cosine Power Annealing learning rate schedule for tuning between Cosine Annealing\cite{loshchilov2016sgdrcosineannealing} and exponential decay, which maintains a more optimal learning rate throughout the training process. \item Introduce Differentiable Hyperparameter Grid Search and the HyperCuboid search space for efficiently evaluating arbitrary discrete choices. \item Demonstrate the low-capacity bias of DARTS search on two search spaces, and introduce Max-W Weight Regularization to correct the problem. \end{enumerate} \section{Related Work} \label{ref:related} Architecture search is the problem of optimizing the structure of a neural network to more accurately solve another underlying problem. In essence, the design steps of a neural network architecture that might otherwise be done by an engineer or graduate student by hand are instead automated and optimized as part of a well defined search space of reasonable layers, connections, outputs, and hyperparameters. In fact, architecture search can itself be defined in terms of hyperparameters\cite{hundt2018hypertree} or as a graph search problem\cite{2017nasnet, 2018enas, cai2018proxylessnas, NIPS2016_ConvNeuralFabrics}. Furthermore, once a search space is defined various tools can be brought to bear on the problem including Bayesian optimization\cite{pmlr-v64-mendoza_towards_2016}, other neural networks\cite{smashHyperNetworks}, reinforcement learning, evolution\cite{real2017large, real2018regularized}, or a wide variety of optimization frameworks. A survey for the topic of Neural Architecture Search (NAS) is available at \cite{2018nassurvey}. Differentiable Architecture Search (DARTS)\cite{liu2018darts} defines a search space in terms of parameters $\alpha$, weights $w_i = softmax(\alpha_i)$, and operation layers $op_i$ which are made differentiable in the form $output_i = w_i * op_i(input_i)$. ProxylessNAS\cite{cai2018proxylessnas} works in a similar manner on a MobileNetv2\cite{s2018mobilenetv2} based search space. It only loads two architectures at a time, updating $w$ based on relative changes between them. This saves GPU memory so that ImageNet size datasets and architectures load directly, but at the cost of shuttling whole architectures between the GPU and main memory. \begin{figure}[btp!] \centering \includegraphics[width=1\columnwidth]{sharpsepconv_and_sharp_darts_labeled.pdf} \caption{\small \label{fig:sharpdartscells} (Top) High-level structure of NASNet\cite{2017nasnet} style architectures like sharpDARTS. Normal cells are stride 1 and Reduction cells are stride 2 (Left) The SharpDARTS and SharpSepConvDARTS normal cell. (Right) The SharpDarts and SharpSepConvDARTS reduction cell. } \vspace{-0.2cm} \end{figure} Depthwise Separable Convolutions (SepConv) are a common building block of these searches, and were described as part of the Xception\cite{chollet2017xception} architecture, which improved efficiency on a per-parameter basis compared to its predecessor Inception-v3\cite{szegedy2016rethinking_incptionv3}. It was subsequently used to great effect in MobileNetV2\cite{s2018mobilenetv2}. A SepConv is where an initial convolution is defined in which the number of groups is equal to the number of input channels, followed by a single 1x1 convolution with a group size of 1. This type of convolution tends to have roughly equivalent or better performance than a standard Conv layer with fewer operations and lower memory utilization.\cite{guo2018network, chollet2017xception} Furthermore, so-called ``bottleneck'' layers or blocks have proven useful to limiting the size and improving the accuracy of neural network models\cite{he2015resnet, s2018mobilenetv2}. Augmentation is another fundamental tactic when optimizing the efficiency of neural networks. For example, on CIFAR-10 Cutout\cite{2017cutout} randomly sets 16x16 squares in an input image to zero; and AutoAugment\cite{cubuk2018autoaugment} is demonstrated on PyramidNet\cite{Han2017PyramidNet}, where it applies reinforcement learning to optimize parameter choices for a set of image transforms, and to the odds of applying each transform during training. \begin{table}\centering \setlength\tabcolsep{3pt} \begin{tabular}{lllllr} \toprule \multicolumn{6}{c}{Operations in the sharpDARTS search space} \\ \cmidrule(r){1-6} Name & K & S & D & CMid & CMidMult \\ \cmidrule(r){1-6} none & -- & -- & -- & -- & -- \\ AvgPool3x3 & 1 & 1 & 1 & -- & -- \\ MaxPool3x3 & 1 & 1 & 1 & -- & -- \\ SkipConnect & 1 & 1 & 1 & -- & -- \\ SepConv3x3 & 3 & 1 & 1 & -- & 1 \\ DilConv3x3 & 3 & 1 & 2 & -- & 1 \\ FloodConv3x3 & 3 & 1 & 1 & -- & 4 \\ DilFloodConv3x3 & 3 & 1 & 2 & -- & 4 \\ ChokeConv3x3 & 3 & 1 & 1 & 32 & -- \\ DilChokeConv3x3 & 3 & 2 & 2 & 32 & -- \\ \bottomrule \end{tabular} \caption{\small \label{table:searchspace} Available operations for connections within cells in the sharpDARTS search space. Columns K, S, and D are kernel, stride, and dilation, respectively. All ``Conv'' operations are configurations of \texttt{SharpSepConv}, shown in Fig. \ref{fig:SharpSepConv} and \ref{fig:pytorch_sharpsepconv}. } \vspace{-0.2cm} \end{table} \begin{figure}[bt!] \includegraphics[width=\columnwidth]{SharpSepConvCode.pdf} \caption{ \small \label{fig:pytorch_sharpsepconv} The complete PyTorch 1.0~\cite{paszke2017pytorch} definition of the \texttt{SharpSepConv} block visualized in Fig. \ref{fig:SharpSepConv}.} \vspace{-0.4cm} \end{figure} \begin{figure*}[bt!] \vspace{-0.2cm} \centering \hfill \includegraphics[width=\textwidth]{annealing_comparison.pdf} \caption{ \label{fig:compare_cosine_annealing} \small A comparison of annealing methods. Cosine Power Annealing is more frequently at an optimal learning rate. These 1000 epoch examples of CIFAR-10 sharpDARTS SGD training were selected to compare progress. Here final validation accuracy differs by 0.04\%, but in the typical case Cosine Power Annealing has better performance. } \vspace{-0.4cm} \end{figure*} \begin{figure}[bt!] \centering \hfill \includegraphics[width=\columnwidth]{cos_power_annealing_imagenet.pdf} \includegraphics[width=\columnwidth]{cos_power_annealing_imagenet_log.pdf} \caption{ \label{fig:cosine_power_annealing} \small A comparison of Cosine Annealing~\cite{loshchilov2016sgdrcosineannealing} (Eq. \ref{eq:cosineannealing}) and Cosine Power Annealing (Eq. \ref{eq:cosinepowerannealing} with $p=10$) learning rate schedules for ImageNet. The bottom chart's logarithmic scale demonstrates how the cosine term is most prominent during early and late epochs, while the exponential term $p$ smooths decay through the middle epochs. } \vspace{-0.2cm} \end{figure} However, improvements via the methods above are only valuable if the results are reproducible and generalize to new data, issues which are a growing concern throughout academia. Recht et al.\cite{recht2018cifar10} have investigated such concerns, creating a new CIFAR-10.1 test set selected from the original tiny images dataset from which CIFAR-10 was itself selected. While the rank ordering of evaluated models remains very consistent, their work demonstrates a significant drop in accuracy on the test set relative to the validation set for all models. Variation implicit in the training process, models, and the dataset itself must be carefully considered when reading results; and so a slightly better score does not guarantee better generalization. For reference, on CIFAR-10.1 ``a conservative confidence interval (Clopper-Pearson at confidence level 95\%) for accuracy 90\% has size about $\pm1\%$ with n = 2,000 (to be precise, [88.6\%, 91.3\%])''\cite{recht2018cifar10}. \section{Search Space and Training Methods} We seek better generalization of architecture search using fewer resources. We begin by analyzing DARTS\cite{liu2018darts}, particularly its search space implementation and training regimen. The search space covers a variety of core operations or ``blocks'' which include pooling, skip connections, no layer, separable convolutions and dilated convolutions. Upon deeper analysis, we note that the \texttt{DilConv} operation contains 2 convolution layers while \texttt{SepConv} contains 4, thus comparing blocks with different scales. This scale imbalance matters, and Sec. \ref{sec:ablation_study}, \ref{sec:hypercuboids_and_regularization}, and \ref{sec:grid_search} will explore why, but first we will go through the design choices and data leading up to that conclusion. Therefore, we start with our \texttt{SharpSepConv} block, which helps to correct the aforementioned imbalance, and the Cosine Power Annealing learning rate schedule. \subsection{Sharp Separable Convolution Block} \label{ssec:goalsandencodings} We first define the \texttt{SharpSepConv} block which consists of 2 Separable Convolutions and 2 bottlenecks to balance the number of layers and to capitalize on these components' computational efficiency, as we described in Related Work (Sec. \ref{ref:related}). All of the parameters of \texttt{SharpSepConv} are visualized in Fig. \ref{fig:SharpSepConv} and the block is defined using the PyTorch 1.0\cite{paszke2017pytorch} framework in Fig. \ref{fig:pytorch_sharpsepconv}. We will refer to these figures and the definitions in the Table \ref{table:cifar10} caption throughout the remaining text. The \texttt{SharpSepConv} block permits variation in the dilation parameters and the number of filters contained in the middle layers, both relative to the input and in absolute terms, without changing the number of convolutions. In effect, \texttt{C\_mid} adds fixed-size bottlenecks directly into the architecture search space which can have a large impact on the number of AddMult operations in an architecture and thus the computational efficiency. Furthermore, \texttt{C\_mid\_mult} makes it possible to incorporate an additional reduction or increase in network size as needed within individual cells as can be seen in Fig. \ref{fig:sharpdartscells}. Using \texttt{SharpSepConv} and other operations we create the sharpDARTS search space defined in Table \ref{table:searchspace}. As the ablation study in Sec. \ref{sec:ablation_study} and the figures in Table \ref{table:cifar10} show, \texttt{SharpSepConv} operations substantially contribute to the improvements inherent to the final sharpDARTS and SharpSepConvDARTS architectures. \begin{table*}\centering \ra{1.3} \setlength\tabcolsep{3pt} \begin{tabular}{@{}lccccccccr@{}}\toprule \multicolumn{1}{l}{Architecture} & \multicolumn{1}{c}{Auto} & \multicolumn{1}{c}{Grad} & \multicolumn{1}{c}{SSC} & \multicolumn{1}{c}{Val Error} & \multicolumn{1}{c}{Test Error} & \multicolumn{1}{c}{Par.} & \multicolumn{1}{c}{$+ \times$} & \multicolumn{1}{c}{GPU} & \multicolumn{1}{c}{Algorithm}\\ \multicolumn{1}{l}{} & \multicolumn{1}{c}{Aug} & \multicolumn{1}{c}{Order} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{CIFAR-10} & \multicolumn{1}{c}{CIFAR-10.1} & \multicolumn{1}{c}{M} & \multicolumn{1}{c}{B} & \multicolumn{1}{c}{Days} & \multicolumn{1}{c}{}\\% \cmidrfule{2-2} \cmidrule{4-4} \cmidrule{6-6} \midrule ShakeShake64d\cite{gastaldi2017shakeshake, recht2018cifar10}&\text{\ding{55}}&&\text{\ding{55}}&2.9$\pm$0.3 &7.0$\pm$1.8&26.2&&--&Man\\ AutoAugment\cite{cubuk2018autoaugment,Han2017PyramidNet}&\checkmark&&\text{\ding{55}}& 1.48 && 26.0 && -- & AugRL+Man \\ GPipe AmoebaNet-B\cite{huang2018gpipe,real2018regularized}&\text{\ding{55}}&&\text{\ding{55}}& 1.0 $\pm$ 0.05 && $\sim$557 &$\sim$50&& Evo+Man \\\midrule NASNet-A \cite{2017nasnet} &\text{\ding{55}}&&\text{\ding{55}}& 2.65 && 3.3 && 1800 & RL \\ NASNet-A \cite{2017nasnet}$^\dagger$ &\text{\ding{55}}&&\text{\ding{55}}& 2.83 && 3.8 &0.624& 3150 & RL \\ AmoebaNet-A \cite{real2018regularized} &\text{\ding{55}}&&\text{\ding{55}}& 3.34 $\pm$ 0.06 && 3.2 && 3150 & Evo \\ AmoebaNet-A \cite{real2018regularized}$^\dagger$&\text{\ding{55}}&&\text{\ding{55}}& 3.12 && 3.1 &0.506& 3150 & Evo \\ AmoebaNet-B \cite{real2018regularized}&\text{\ding{55}}&&\text{\ding{55}}& 2.55 $\pm$ 0.05 && 2.8 && 3150 & Evo \\ ProgressiveNAS \cite{liu2018PNAS}&\text{\ding{55}}&&\text{\ding{55}}& 3.41 $\pm$ 0.09 && 3.2 && 225 & SMBO \\ ENAS \cite{2018enas}&\text{\ding{55}}&&\text{\ding{55}}& 2.89 && 4.6 && 0.5 & RL \\ DARTS random~\cite{liu2018darts}&\text{\ding{55}}&&\text{\ding{55}}& 3.49 && 3.1 && -- & -- \\ DARTS$^\dagger$~\cite{liu2018darts}&\text{\ding{55}}&1&\text{\ding{55}}& 2.94 && 2.9 &0.518& 1.5 &Grad\\ DARTS$^\dagger$~\cite{liu2018darts}&\text{\ding{55}}&2&\text{\ding{55}}& 2.83 $\pm$ 0.06 && 3.4 &0.547& 4 &Grad\\ ProxylessNAS~\cite{cai2018proxylessnas}&\text{\ding{55}}&1&\text{\ding{55}}& 2.08 && 5.7 && 4 & Grad\\ \midrule \textbf{SharpSepConvDARTS}&\checkmark&\textbf{1}&\checkmark& \textbf{1.98$\pm$0.07} &\textbf{5.9$\pm$0.5}& \textbf{3.6} &\textbf{0.579}&\textbf{0.8}&\textbf{Grad}\\ \textbf{DARTS+SSC, no search}&\checkmark&\textbf{2}&\checkmark& \textbf{2.05$\pm$0.14} &\textbf{5.6$\pm$0.8}& \textbf{3.5} &\textbf{0.576}& -- &\textbf{Grad}\\ \textbf{sharpDARTS}&\checkmark&\textbf{1}&\checkmark& \textbf{2.29$\pm$0.06} &\textbf{6.2$\pm$0.4}& \textbf{1.98} &\textbf{0.357}& \textbf{1.8} &\textbf{Grad} \\ SharpSepConvDARTS&\text{\ding{55}}&1&\checkmark& 2.45 && 3.6 &0.579& 0.8 &Grad\\ DARTS+SSC, no search&\text{\ding{55}}&2&\checkmark& 2.55 && 3.5 &0.576& -- &Grad\\ sharpDARTS&\text{\ding{55}}&1&\checkmark& 2.97 && 1.98 &0.357& 1.8 &Grad\\ \bottomrule \bottomrule \end{tabular} \caption{\small \label{table:cifar10} Results for the CIFAR-10 dataset and the CIFAR-10.1\cite{recht2018cifar10} test set, \textbf{bold lines} are trained with AutoAugment\cite{cubuk2018autoaugment} and Cosine Power Annealing (Eq. \ref{eq:cosinepowerannealing}) where we set $p=2$, $\eta^0_{max}=0.025$, $\eta^0_{min}=1\text{e-}8$, and $T_0=2000$ epochs of training for our best results. Our figures with a range in the bottom section are defined as $median \pm range/2$ and include at least 3 runs with Cosine Power Annealing. Dagger$^\dagger$ indicates genotypes available in our accompanying repository. Non-bold numbers have 1000 epochs of Cosine Power Annealing and AutoAugment disabled to better match figures from past work. The code and this table is derived from DARTS\cite{liu2018darts}. The \textbf{GradOrder} column is 1 for standard gradients, 2 for Hessians on DARTS based algorithms. All algorithms utilize \textbf{CutOut}\cite{2017cutout} except for ProgressiveNAS. A \textbf{``genotype''} simply defines an architecture as a graph. It has a list of layers to use, and the previous node, which it should use as an input. A \textbf{``primitive''} is an operation aka block of layers that can be chosen during the search. A \textbf{``node''} takes the output of two instances of an (input, primitive) pair as input and adds them, as seen in Fig. \ref{fig:sharpdartscells}. \textbf{DARTS+SSC, no search} compares the genotype published by DARTS with \texttt{SharpSepConv} layers in the primitives. \textbf{SharpSepConvDARTS} is the same as DARTS + SharpSepConv but uses a genotype determined by a new search incorporating the DARTS primitives with SharpSepConv blocks. sharpDARTS uses the architecture search space described in Table \ref{table:searchspace}.} \vspace{-0.2cm} \end{table*} \begin{table*}\centering \ra{1.3} \setlength\tabcolsep{2pt} \centering \begin{tabular}{@{}lccccccc@{}} \toprule \multirow{2}{*}{\textbf{Architecture}} & \multicolumn{2}{c}{\textbf{\% Test Error}} & \textbf{Params} & \textbf{$+ \times$} & \textbf{Search Cost} & \textbf{Search} \\ \cline{2-3} & top-1 & top-5 & \textbf{M} & \textbf{B} & \textbf{GPU Days} & \textbf{Method} \\ \midrule GPipe AmoebaNet-B \cite{real2018regularized,huang2018gpipe} & 15.7 & 3.0 & 557 & $\sim$50 & 3150 & Evo+Man \\ MobileNetv2 \cite{s2018mobilenetv2} & 28.0 & -- & 3.4 & 0.300 & -- & Man \\ MobileNetv2 (1.4) \cite{s2018mobilenetv2} & 25.3 & -- & 6.9 & 0.585 & -- & Man \\ \midrule NASNet-A \cite{2017nasnet} & 26.0 & 8.4 & 5.3 & 0.564 & 1800 & RL \\ NASNet-B \cite{2017nasnet} & 27.2 & 8.7 & 5.3 & 0.488 & 1800 & RL \\ AmoebaNet-B \cite{real2018regularized} & 26.0 & 8.5 & 5.3 & 0.555 & 3150 & Evolution \\ AmoebaNet-C \cite{real2018regularized} & 24.3 & 7.6 & 6.4 & 0.570 & 3150 & Evolution \\ ProgressiveNAS \cite{liu2018PNAS} & 25.8 & 8.1 & 5.1 & 0.588 & $\sim$225 & SMBO \\ DARTS \cite{liu2018darts} & 26.9 & 9.0 & 4.9 & 0.595 & 4 & Grad \\ ProxylessNAS \cite{cai2018proxylessnas} & 24.9 & 8.1 & 5.1 & 0.581 & $\sim$8 & Grad \\\midrule \textbf{SharpSepConvDARTS} & \textbf{25.1} & \textbf{7.8} & \textbf{4.9} & \textbf{0.573} & \textbf{0.8} & \textbf{Grad} \\ \textbf{sharpDARTS cmid 96} & \textbf{27.8} & \textbf{9.2} & \textbf{3.71} & \textbf{0.481} & \textbf{1.8} & \textbf{Grad} \\ \textbf{sharpDARTS cmid 32} & \textbf{31.3} & \textbf{11.34} & \textbf{3.17} & \textbf{0.370} & \textbf{1.8} & \textbf{Grad} \\ \bottomrule \end{tabular} \caption{\small Mobile ImageNet Architecture Comparison, table updated from DARTS\cite{liu2018darts}. Lower values are better in all columns. A dash indicates data was either not available or not applicable. \label{table:imagenet}} \vspace{-0.2cm} \end{table*} \subsection{Cosine Power Annealing} Cosine Annealing\cite{loshchilov2016sgdrcosineannealing} is a method of adjusting the learning rate over time, reproduced below: \begin{eqnarray} \label{eq:cosineannealing} \eta_t = \eta^i_{min} + \frac{1}{2}(\eta^i_{max} - \eta^i_{min}){ \left(1 + \cos\left(\pi \frac{T_{cur}}{T_i}\right)\right)} \end{eqnarray} Where $\eta^i_{min}$ and $\eta^i_{max}$ are the minimum and maximum learning rates, respectively; $T_i$ is the total number of epochs; $T_{cur}$ is the current epoch; and $i$ is the index into a list of these parameters for a sequence of warm restarts in which $\eta^i_{max}$ typically decays. This schedule has been widely adopted and it is directly implemented in PyTorch 1.0~\cite{paszke2017pytorch} without warm restarts, where $i = 0$. The Cosine Annealing schedule works very well and is employed by DARTS. However, as can be seen in Fig. \ref{fig:compare_cosine_annealing}, the average time between improvements is above 2\% of the total runtime between epochs 300 and 700. This is an imbalance in the learning rate, which is an artifact of the initial slow decay rate of Cosine Annealing followed by the rapid relative decay in learning rate. We mitigate this imbalance by introducing a power curve parameter $p$ into the algorithm which we call \textbf{Cosine Power Annealing:} \begin{equation} \label{eq:cosinepowerannealing} \eta_t = \eta^i_{min} + (\eta^i_{max} - \eta^i_{min})\frac{p^{\frac{1}{2} \left(1 + \cos\left(\pi \frac{T_{cur}}{T_i}\right)\right) + 1}-p}{p^2-p} \end{equation} The introduction of the normalized exponential term $p$ permits tuning of the curve's decay rate such that it maintains a high learning rate for the first few epochs, while simultaneously taking a shallower slope during the final third of epochs. These two elements help reduce the time between epochs in which validation accuracy improves. A comparison is provided in Fig. \ref{fig:compare_cosine_annealing} and \ref{fig:cosine_power_annealing}. In our implementation we also define a special case for the choice of $p=1$ such that it falls back to standard cosine annealing. This algorithm is also compatible with decaying warm restarts, but we leave that schedule out of scope for the purposes of this paper. \subsection{Results} \label{sec:results} \textbf{CIFAR-10 and CIFAR-10.1:} Our absolute top performance is \textbf{SharpSepConvDARTS} with 1.93\% CIFAR-10 top-1 validation error (1.98$\pm$0.07) and 5.925$\pm$0.48 CIFAR-10.1 test error (Table \ref{table:cifar10}) when including our improved training regimen. To the best of our knowledge, this is state of the art performance for mobile-scale ($\sim$600M AddMult ops\cite{liu2018darts}) and ProxylessNAS\cite{cai2018proxylessnas} is in second at 2.08\% val error. We also show a statistically significant\cite{recht2018cifar10} improvement over ShakeShake64d\cite{recht2018cifar10,gastaldi2017shakeshake} which is the best available CIFAR-10.1 model, with 7.0$\pm$1.2 test error. GPipe AmoebaNet-B\cite{huang2018gpipe} remains the best at any scale with 1\% val error. It is scaled up from the original AmoebaNet-B and we expect other models to scale in a similar way. This truly massive model cannot load on typical GPUs due to 557M parameters and billions of AddMult ops. It runs with specialized software across multiple Google TPU hardware devices. Our training time\footnote{This research was conducted on 5 Nvidia GPU types: Titan X, GTX 1080 Ti, GTX 1080, Titan XP, and the RTX 2080Ti.} for a 60 epoch search is 0.8-1.2 days, with 2k epochs of mixed fp16 training of a final model in 1.7-2.8 days, totaling 2.9-3.6 GPU-days end-to-end on one RTX 2080Ti. The discrepancy in totals is because the slower sharpDARTS search finds a faster final model. \textbf{ImageNet:} Our top \textbf{SharpSepConvDARTS} model achieved 25.1\% top-1 and 7.8\% top-5 error, which is competitive with other state of the art mobile-scale models depicted in Table \ref{table:imagenet}, and this translates to relative improvements over DARTS of 7\% in top-1 error, 13\% in top-5 error, and 80\% in search time. Our ImageNet model uses the genotype of the CIFAR-10 search and follows the same cell based architecture of \cite{2017nasnet, liu2018darts, 2018enas} with different operations in our search space. We apply random cropping to 224x224, random horizontal flipping, AutoAugment\cite{cubuk2018autoaugment}, normalization to the dataset mean and std deviation, and finally Cutout\cite{2017cutout} with a cut length of 112x112. Training of final models was done on 2x GTX 2080Ti in 16 bit mixed precision mode and takes 4-6 days, which is 8-12 GPU-days, depending on the model. \section{Towards Better Generalization} \label{sec:towards_better_generalization} DARTS search improved results over random search by 19\% (Table \ref{table:cifar10}), and by adding our training regimen and search space improvements we get an additional 30\% relative improvement over DARTS. Manual changes like those made to the \textbf{SharpSepConv} block and to AmoebaNet-B for GPipe\cite{huang2018gpipe} are not represented in any search space, and yet they directly lead to clear improvements in accuracy. So why aren't they accounted for? Let's assume that it is possible to encode all of these elements and more into a single, broader search space in which virtually every neural network graph imaginable is encoded by hyperparameters. To even imagine tackling a problem of this magnitude, we must first ask ourselves an important question: Does DARTS even generalize to other search domains designed with this challenge in mind? In this section, we provide a preliminary exploration to begin answering these questions. \begin{figure*}[btp!] \centering \hfill \includegraphics[width=0.95\textwidth]{multi_channel_net_2x2_x2x2x2.pdf} \caption{\small \label{fig:multichannel} A HyperCuboid with shape HC = (filter\_scales, normal\_layers, reduction\_layers, layer\_types) = ((2x2)x2x2x2) for Differentiable Hyperparameter Grid Search (Sec. \ref{sec:grid_search}). The data is ordered $(channels, width, height)$. An example single-path model is highlighted with red arrows. Here you can see a small scale n-tuple coordinate system which might serve as the conceptual basis for future large scale search spaces. (Sec. \ref{sec:towards_better_generalization})} \vspace{-0.6cm} \end{figure*} \subsection{Differentiable Grid Search} \label{sec:grid_search} We introduce \textbf{Differentiable Hyperparameter Grid Search} which is run on a \textbf{HyperCuboid Search Space} parameterized by an n-tuple, such as $HC = (\gamma,d_n,d_r,p)$, where $\gamma$ is an arbitrary set of hyperparameters; $d_n$ is the number layers in one block; $d_r$ is the number of layer strides; and $p$ is an arbitrary set of choices, primitives in this case. In the HyperCuboid Search Space many possible paths pass through each node, and all final sequential paths in the Directed Acyclic Graph of architecture weights are of equal length. The number of architecture weights $w$ in the HyperCuboid graph is the product of the size of the tuple elements; and the number of hyperparameters, primitives, and tuple size can vary in this design. We test a specific HyperCuboid called \textbf{MultiChannelNet} with a tuple (filter\_scale\_pairs, normal\_layer\_depth, reduction\_layer\_depth, primitives), visualized at a small ((2x2)x2x2x2) scale in Fig. \ref{fig:multichannel} with SharpSepConv and MaxPool primitives. Here $\gamma=2^n\times2^m|n,m\in[5,6]$ are combinations of possible input and output filter scales. Our actual search has dimension ((4x4)x3x3x2)$|n,m\in[5..8]$ where final linear paths have 14 nodes, since ``Add'' nodes are excluded. In MultiChannelNet, graph nodes determine which filter input, filter output, and layer type should be utilized. Larger weights $w$ imply a better choice of graph node and we therefore search for an optimal path through a sequence of primitive weights which maximizes the total path score. We construct a basic handmade model in one shot with \texttt{SharpSepConv} and 32 filters, doubling the number of output filters at layers of stride 2 up to the limit of 256. We also ran an automated DARTS search for 60 epochs and about 16 GPU-Hours to find an optimal model. To our surprise, the handmade model outperforms DARTS by over 2\%, as Table \ref{table:multichannelnet_search_algorithms} indicates. Why might this be? To answer this we return to our original NASNet based search space to look for discrepancies in an ablation study. \begin{table}\centering \vspace{0.2cm} \setlength\tabcolsep{2pt} \begin{tabular}{lrrcc} \toprule Weights/Path & Par. & $+ \times$ & Val Err \% & Test Err \% \\ \cmidrule(r){1-5} Scalar&0.8M& 23M& 8.62$\pm$0.16& 17.2$\pm$0.8 \\ Handmade& 1.0M&25M& 6.50$\pm$0.18& 13.9$\pm$0.4 \\ \textbf{Max-W}&\textbf{0.9M}&\textbf{31M}&\textbf{5.44$\pm$0.12}&\textbf{12.4$\pm$1.0}\\ \bottomrule \end{tabular} \caption{\small \label{table:multichannelnet_search_algorithms} CIFAR-10 and CIFAR-10.1 results for model paths from our MultiChannelNet (Fig. \ref{fig:multichannel}) example of the HyperCuboid Search Space. There is a statistically significant\cite{recht2018cifar10} improvement from the original Scalar DARTS to our proposed Max-W Weight Regularization DARTS. We also see the hypothesized increase in accuracy, parameters, and AddMult ($+ \times$) flops described in Sec. \ref{sec:hypercuboids_and_regularization} and \ref{sec:grid_search}. Final training is $\sim$0.3 days on an RTX 2080Ti. } \vspace{-0.3cm} \end{table} \subsection{Ablation Study} \label{sec:ablation_study} Figures for our ablation study are in Table \ref{table:cifar10} which also has a description of our preprocessing changes in the caption. We will analyze the 3 main model configurations below: \textbf{(1)} \textbf{DARTS+SSC} directly replaces all convolution primitives in DARTS\cite{liu2018darts} with a \texttt{SharpSepConv} layer where the block parameters, primitives, and the genotype are otherwise held constant; we see a 10\% relative improvement over DARTS val err (2.55\% vs 2.83\%) with only 5\% more AddMult operations and no additional augmentation. \textbf{(2)} \textbf{SharpSepConvDARTS} is the same as item 1 but with a 1st order gradient DARTS search; we see relative improvements of 13\% in val err (2.45\% vs 2.83\%) and 80\% in search time. \textbf{(3)} \textbf{sharpDARTS} (Table \ref{table:searchspace}) has slightly different primitives including \textit{flood} where middle channels expand 4x, \textit{choke} with a fixed 32 middle filters, and one of each conv with a \textit{dilation} of 2. Our sharpDARTS model (Fig. \ref{fig:sharpdartscells}) with our improved training regimen achieved an absolute best error of 2.27\%. Without training enhancements the 1st order gradient search of sharpDARTS has similar accuracy to 1st order DARTS and is definitively more efficient with 32\% fewer parameters, 31\% fewer AddMult operations, and lower memory requirements. However, the absolute accuracy of sharpDARTS is \textit{marginally lower} than the original DARTS, and also suffers from a larger disparity on ImageNet (Table \ref{table:imagenet}). This is startling for two reasons: (1) Substituting the \texttt{SharpSepConv} block improves accuracy in DARTS+SSC and SharpSepConvDARTS. (2) The sharpDARTS search space still contains all primitives needed to represent both the final DARTS+SSC and SharpSepConvDARTS model genotypes perfectly. We've replicated a discrepancy in DARTS behavior across two different search spaces, so the most likely remaining possibility must be a limitation in the DARTS search method itself. \subsection{Max-W Regularization} \label{sec:hypercuboids_and_regularization} GPipe\cite{huang2018gpipe} shows AmoebaNet-B models improving in accuracy as they scale to $>$500M parameters and billions of AddMult flops (Fig. \ref{table:cifar10}, \ref{table:imagenet}). These models are from a search space similar to those used by DARTS, among others\cite{liu2018PNAS,2018enas,2017nasnet}. Assume that the GPipe scaling principle holds for similar training configurations and search spaces, and one might expect that DARTS would tend towards larger models throughout the search process. However, during the early epochs of training, DARTS reliably produces models composed entirely of max pools and skip connects. These are among the smallest primitives in the DARTS architecture search space with respect to parameters and AddMult operations. Higher capacity layers are chosen later in the search process, as visualized in the animation included with the original DARTS source code\footnote{\scriptsize DARTS\cite{liu2018darts} model search animation for reduce cells: \url{https://git.io/fjfTC} normal cells: \url{https://git.io/fjfbT}}. We experimented with removing max pool layers and the undesired behavior simply shifts to skip connects. We posit that the \textbf{DARTS Scalar Weighting} $output_i = w_i * op_i(input_i)$ tends towards the layers with the maximum gradient, and thus models will consist of smaller layers than appropriate. Such bias during early phases of training is inefficient with respect to optimal accuracy, even if models might eventually converge to larger, more accurate models after a long period of training. Therefore, we hypothesize that subtracting the maximum weight in a given layer will regularize weight changes via \textbf{Max-W Weighting:} \begin{equation} output_i = (1 - max(w) + w_i) * op_i(input_i) \label{eqn:max_w} \end{equation} Intuitively, consider the architecture parameters $\alpha$, weights $w=softmax(\alpha) | w\in[0,1]$, highest score index $i_{max}=argmax(w)$, and the highest score weight $w_{max}=w[i_{max}]$ at some arbitrary time during training. If we apply Max-W weighting to the layer corresponding with $w_{max}$, we will have $(1 - w_{max} + w_{max}) * op_i(input_i)$ which reduces to $output_i = 1 * op_i(input_i)|i=i_{max}$. In this case the $\alpha$ underlying $w_{max}$ will remain unchanged, but this is not true for other values $w_i| i\neq i_{max}$. Here it will be incumbent on non-maximum layers $w_i$ to outperform the highest score weight and grow their value $\alpha_i$. As values other than $w_{max}$ grow, $w_{max}$ will naturally drop in accordance with the behavior of $softmax$. This has the net effect of reducing bias corresponding to the highest score layer $op_i$ paired with $w[i_{max}]$. Our search with Max-W weighting on MultiChannelNet found a model which is both larger and significantly more accurate than both the original DARTS Scalar search models and a hand designed model (Table \ref{table:multichannelnet_search_algorithms}). These results indicate that our initial hypothesis holds and Max-W DARTS (Eq. \ref{eqn:max_w}) is an effective approach to regularization when compared to standard Scalar DARTS. Specific models will be released with the code. \section{Future Work} Our investigation also indicates several other areas for future work. The SharpSepConvDARTS and sharpDARTS search spaces might also benefit from Max-W regularization, so it is an interesting topic for an additional ablation study. MultiChannelNet and sharpDARTS indicate that the final DARTS model\cite{liu2018darts} did not fully converge to the optimum due to the scalar weighting bias (Sec. \ref{sec:ablation_study}, \ref{sec:hypercuboids_and_regularization}). We suspect other DARTS algorithms, such as ProxylessNAS\cite{cai2018proxylessnas}, suffer from this same bias. We have also shown how Max-W Regularization correctly chooses larger models when those are more accurate, however this means a search with Max-W DARTS currently exceeds mobile-scale on the DARTS and sharpDARTS search spaces. Adding a resource cost based on time and memory to each node might make it possible to directly optimize cost-accuracy tradeoffs with respect to a specific budget. Arbitrary multi-path subgraphs respecting this budget could be chosen by iterative search with a graph algorithm like network simplex\cite{orlin1997polynomialnetworksimplex}. Other alternatives include a reinforcement learning algorithm or differentiable metrics like the latency loss in ProxylessNAS\cite{cai2018proxylessnas}. \section{Conclusion} In this paper we met or exceeded state of the art mobile-scale architecture search performance on CIFAR-10, CIFAR-10.1 and ImageNet with a new \texttt{SharpSepConv} block. We introduced the Cosine Power Annealing learning rate schedule, which is more often at an optimal learning rate than Cosine Annealing alone, and demonstrated an improved sharpDARTS training regimen. Finally, we introduced Differentiable Hyperparameter Grid Search with a HyperCuboid search space to reproduce bias within the DARTS search method, and demonstrated how Max-W regularization of DARTS corrects that imbalance. Finally, Differentiable Hyperparameter Search and HyperCuboids might be evaluated more broadly in the computer vision space and on other topics such as recurrent networks, reinforcement learning, natural language processing, and robotics. For example, \texttt{SharpSepConv} is manually designed so a new HyperCuboid might be constructed to empirically optimize the number, sequence, type, layers, activations, normalization, and connections within a block. Perhaps a future distributed large scale model might run Differentiable Hyperparameter Search over hundreds of hyperparameters which embed a superset of search spaces, making it possible to efficiently and automatically find new models for deployment to any desired application. {\small \bibliographystyle{style/ieee} \section{Introduction} Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press. This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version. \subsection{Language} All manuscripts must be in English. \subsection{Dual submission} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the policy on dual submissions. \subsection{Paper length} Papers, excluding the references section, must be no longer than eight pages in length. The references section will not be included in the page count, and there is no limit on the length of the references section. For example, a paper of eight pages with two pages of references would have a total length of 10 pages. Overlength papers will simply not be reviewed. This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts. The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven. \subsection{The ruler} The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review. The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution. If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages. The presence or absence of the ruler should not change the appearance of any other content on the page. The camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment the \verb'\iccvfinalcopy' command in the document preamble.) Reviewers: note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly. Just use fractional references (e.g.\ this line is $095.5$), although in most cases one would expect that the approximate location will be adequate. \subsection{Mathematics} Please number all of your sections and displayed equations. It is important for readers to be able to refer to any particular equation. Just because you didn't refer to it in the text doesn't mean some future reader might not need to refer to it. It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''. (Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers). All authors will benefit from reading Mermin's description of how to write mathematics: \url{http://www.pamitc.org/documents/mermin.pdf}. \subsection{Blind review} Many authors misunderstand the concept of anonymizing for blind review. Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available. Blind review means that you do not use the words ``my'' or ``our'' when citing previous work. That is all. (But see below for techreports.) Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith; it says that you are building on her work. If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work. An example of a bad paper just asking to be rejected: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Removed for blind review \end{quote} An example of an acceptable paper: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213. \end{quote} If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work. In such cases, include the anonymized parallel submission~\cite{Authors14} as additional material and cite it as \begin{quote} [1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as additional material {\tt fg324.pdf}. \end{quote} Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report. For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a techreport for further details. Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''. Then submit the techreport as additional material. Again, you may not assume the reviewers will read this material. Sometimes your paper is about a problem which you tested using a tool which is widely known to be restricted to a single institution. For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the ICCV70 audience would like to hear about your solution. The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal. You can handle this paper like any other. Don't write ``We show how to improve our previous work [Anonymous, 1968]. This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''. That would be silly, and would immediately identify the authors. Instead write the following: \begin{quotation} \noindent We describe a system for zero-g frobnication. This system is new because it handles the following cases: A, B. Previous systems [Zeus et al. 1968] didn't handle case B properly. Ours handles it by including a foo term in the bar integral. ... The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know. It displayed the following behaviours which show how well we solved cases A and B: ... \end{quotation} As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors. A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess. He or she would have to be sure that no other authors could have been contracted to solve problem B. \noindent FAQ\medskip\\ {\bf Q:} Are acknowledgements OK?\\ {\bf A:} No. Leave them for the final copy.\medskip\\ {\bf Q:} How do I cite my results reported in open challenges? {\bf A:} To conform with the double blind review policy, you can report results of other challenge participants together with your results in your paper. For your results, however, you should not identify yourself and should not mention your participation in the challenge. Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\ \begin{figure}[t] \begin{center} \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}} \end{center} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{fig:long} \label{fig:onecol} \end{figure} \subsection{Miscellaneous} \noindent Compare the following:\\ \begin{tabular}{ll} \verb'$conf_a$' & $conf_a$ \\ \verb'$\mathit{conf}_a$' & $\mathit{conf}_a$ \end{tabular}\\ See The \TeX book, p165. The space after \eg, meaning ``for example'', should not be a sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided \verb'\eg' macro takes care of this. When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.) However, use it only when there are three or more authors. Thus, the following is correct: `` Frobnication has been trendy lately. It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.'' This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors. If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal. For this citation style, keep multiple citations in numerical (not chronological) order, so prefer \cite{Alpher03,Alpher02,Authors14} to \cite{Alpher02,Alpher03,Authors14}. \begin{figure*} \begin{center} \fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \end{center} \caption{Example of a short caption, which should be centered.} \label{fig:short} \end{figure*} \section{Formatting your paper} All text must be in a two-column format. The total allowable width of the text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the first page) should begin 1.0 inch (2.54 cm) from the top edge of the page. The second and following pages should begin 1.0 inch (2.54 cm) from the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the page. \subsection{Margins and page numbering} All printed material, including text, illustrations, and charts, must be kept within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm) high. Page numbers should be in footer with page numbers, centered and .75 inches from the bottom of the page and make it start at the correct page number rather than the 4321 in the example. To do this fine the line (around line 23) \begin{verbatim} \setcounter{page}{4321} \end{verbatim} where the number 4321 is your assigned starting page. Make sure the first page is numbered by commenting out the first page being empty on line 46 \begin{verbatim} \end{verbatim} \subsection{Type-style and fonts} Wherever Times is specified, Times Roman may also be used. If neither is available on your word processor, please use the font closest in appearance to Times to which you have access. MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of the first page. The title should be in Times 14-point, boldface type. Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs; do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word). Leave two blank lines after the title. AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title and printed in Times 12-point, non-boldface type. This information is to be followed by two blank lines. The ABSTRACT and MAIN TEXT are to be in a two-column format. MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use double-spacing. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422 cm). Make sure your text is fully justified---that is, flush left and flush right. Please do not place any additional blank lines between paragraphs. Figure and table captions should be 9-point Roman type as in Figures~\ref{fig:onecol} and~\ref{fig:short}. Short captions should be centred. \noindent Callouts should be 9-point Helvetica, non-boldface type. Initially capitalize only the first word of section titles and first-, second-, and third-order headings. FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after. SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after. If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line. \subsection{Footnotes} Please use footnotes\footnote {This is what a footnote looks like. It often distracts the reader from the main flow of the argument.} sparingly. Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence). If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced. Use Times 8-point type, single-spaced. \subsection{References} List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Authors14}. Where appropriate, include the name(s) of editors of referenced books. \begin{table} \begin{center} \begin{tabular}{|l|c|} \hline Method & Frobnability \\ \hline\hline Theirs & Frumpy \\ Yours & Frobbly \\ Ours & Makes one's heart Frob\\ \hline \end{tabular} \end{center} \caption{Results. Ours is better.} \end{table} \subsection{Illustrations, graphs, and photographs} All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the paper. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your paper in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage[dvips]{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.eps} \end{verbatim} } \subsection{Color} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the use of color in your document. \section{Final copy} You must include your signed IEEE copyright release form when you submit your finished paper. We MUST have this form before your paper can be published in the proceedings. {\small \bibliographystyle{ieee} \section{Introduction} Please follow the steps outlined below when submitting your manuscript to the IEEE Computer Society Press. This style guide now has several important modifications (for example, you are no longer warned against the use of sticky tape to attach your artwork to the paper), so all authors should read this new version. \subsection{Language} All manuscripts must be in English. \subsection{Dual submission} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the policy on dual submissions. \subsection{Paper length} Papers, excluding the references section, must be no longer than eight pages in length. The references section will not be included in the page count, and there is no limit on the length of the references section. For example, a paper of eight pages with two pages of references would have a total length of 10 pages. Overlength papers will simply not be reviewed. This includes papers where the margins and formatting are deemed to have been significantly altered from those laid down by this style guide. Note that this \LaTeX\ guide already sets figure captions and references in a smaller font. The reason such papers will not be reviewed is that there is no provision for supervised revisions of manuscripts. The reviewing process cannot determine the suitability of the paper for presentation in eight pages if it is reviewed in eleven. \subsection{The ruler} The \LaTeX\ style defines a printed ruler which should be present in the version submitted for review. The ruler is provided in order that reviewers may comment on particular lines in the paper without circumlocution. If you are preparing a document using a non-\LaTeX\ document preparation system, please arrange for an equivalent ruler to appear on the final output pages. The presence or absence of the ruler should not change the appearance of any other content on the page. The camera ready copy should not contain a ruler. (\LaTeX\ users may uncomment the \verb'\iccvfinalcopy' command in the document preamble.) Reviewers: note that the ruler measurements do not align well with lines in the paper --- this turns out to be very difficult to do well when the paper contains many figures and equations, and, when done, looks ugly. Just use fractional references (e.g.\ this line is $095.5$), although in most cases one would expect that the approximate location will be adequate. \subsection{Mathematics} Please number all of your sections and displayed equations. It is important for readers to be able to refer to any particular equation. Just because you didn't refer to it in the text doesn't mean some future reader might not need to refer to it. It is cumbersome to have to use circumlocutions like ``the equation second from the top of page 3 column 1''. (Note that the ruler will not be present in the final copy, so is not an alternative to equation numbers). All authors will benefit from reading Mermin's description of how to write mathematics: \url{http://www.pamitc.org/documents/mermin.pdf}. \subsection{Blind review} Many authors misunderstand the concept of anonymizing for blind review. Blind review does not mean that one must remove citations to one's own work---in fact it is often impossible to review a paper unless the previous citations are known and available. Blind review means that you do not use the words ``my'' or ``our'' when citing previous work. That is all. (But see below for techreports.) Saying ``this builds on the work of Lucy Smith [1]'' does not say that you are Lucy Smith; it says that you are building on her work. If you are Smith and Jones, do not say ``as we show in [7]'', say ``as Smith and Jones show in [7]'' and at the end of the paper, include reference 7 as you would any other cited work. An example of a bad paper just asking to be rejected: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of our previous paper [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Removed for blind review \end{quote} An example of an acceptable paper: \begin{quote} \begin{center} An analysis of the frobnicatable foo filter. \end{center} In this paper we present a performance analysis of the paper of Smith \etal [1], and show it to be inferior to all previously known methods. Why the previous paper was accepted without this analysis is beyond me. [1] Smith, L and Jones, C. ``The frobnicatable foo filter, a fundamental contribution to human knowledge''. Nature 381(12), 1-213. \end{quote} If you are making a submission to another conference at the same time, which covers similar or overlapping material, you may need to refer to that submission in order to explain the differences, just as you would if you had previously published related work. In such cases, include the anonymized parallel submission~\cite{Authors14} as additional material and cite it as \begin{quote} [1] Authors. ``The frobnicatable foo filter'', F\&G 2014 Submission ID 324, Supplied as additional material {\tt fg324.pdf}. \end{quote} Finally, you may feel you need to tell the reader that more details can be found elsewhere, and refer them to a technical report. For conference submissions, the paper must stand on its own, and not {\em require} the reviewer to go to a techreport for further details. Thus, you may say in the body of the paper ``further details may be found in~\cite{Authors14b}''. Then submit the techreport as additional material. Again, you may not assume the reviewers will read this material. Sometimes your paper is about a problem which you tested using a tool which is widely known to be restricted to a single institution. For example, let's say it's 1969, you have solved a key problem on the Apollo lander, and you believe that the ICCV70 audience would like to hear about your solution. The work is a development of your celebrated 1968 paper entitled ``Zero-g frobnication: How being the only people in the world with access to the Apollo lander source code makes us a wow at parties'', by Zeus \etal. You can handle this paper like any other. Don't write ``We show how to improve our previous work [Anonymous, 1968]. This time we tested the algorithm on a lunar lander [name of lander removed for blind review]''. That would be silly, and would immediately identify the authors. Instead write the following: \begin{quotation} \noindent We describe a system for zero-g frobnication. This system is new because it handles the following cases: A, B. Previous systems [Zeus et al. 1968] didn't handle case B properly. Ours handles it by including a foo term in the bar integral. ... The proposed system was integrated with the Apollo lunar lander, and went all the way to the moon, don't you know. It displayed the following behaviours which show how well we solved cases A and B: ... \end{quotation} As you can see, the above text follows standard scientific convention, reads better than the first version, and does not explicitly name you as the authors. A reviewer might think it likely that the new paper was written by Zeus \etal, but cannot make any decision based on that guess. He or she would have to be sure that no other authors could have been contracted to solve problem B. \noindent FAQ\medskip\\ {\bf Q:} Are acknowledgements OK?\\ {\bf A:} No. Leave them for the final copy.\medskip\\ {\bf Q:} How do I cite my results reported in open challenges? {\bf A:} To conform with the double blind review policy, you can report results of other challenge participants together with your results in your paper. For your results, however, you should not identify yourself and should not mention your participation in the challenge. Instead present your results referring to the method proposed in your paper and draw conclusions based on the experimental comparison to other results.\medskip\\ \begin{figure}[t] \begin{center} \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}} \end{center} \caption{Example of caption. It is set in Roman so that mathematics (always set in Roman: $B \sin A = A \sin B$) may be included without an ugly clash.} \label{fig:long} \label{fig:onecol} \end{figure} \subsection{Miscellaneous} \noindent Compare the following:\\ \begin{tabular}{ll} \verb'$conf_a$' & $conf_a$ \\ \verb'$\mathit{conf}_a$' & $\mathit{conf}_a$ \end{tabular}\\ See The \TeX book, p165. The space after \eg, meaning ``for example'', should not be a sentence-ending space. So \eg is correct, {\em e.g.} is not. The provided \verb'\eg' macro takes care of this. When citing a multi-author paper, you may save space by using ``et alia'', shortened to ``\etal'' (not ``{\em et.\ al.}'' as ``{\em et}'' is a complete word.) However, use it only when there are three or more authors. Thus, the following is correct: `` Frobnication has been trendy lately. It was introduced by Alpher~\cite{Alpher02}, and subsequently developed by Alpher and Fotheringham-Smythe~\cite{Alpher03}, and Alpher \etal~\cite{Alpher04}.'' This is incorrect: ``... subsequently developed by Alpher \etal~\cite{Alpher03} ...'' because reference~\cite{Alpher03} has just two authors. If you use the \verb'\etal' macro provided, then you need not worry about double periods when used at the end of a sentence as in Alpher \etal. For this citation style, keep multiple citations in numerical (not chronological) order, so prefer \cite{Alpher03,Alpher02,Authors14} to \cite{Alpher02,Alpher03,Authors14}. \begin{figure*} \begin{center} \fbox{\rule{0pt}{2in} \rule{.9\linewidth}{0pt}} \end{center} \caption{Example of a short caption, which should be centered.} \label{fig:short} \end{figure*} \section{Formatting your paper} All text must be in a two-column format. The total allowable width of the text area is $6\frac78$ inches (17.5 cm) wide by $8\frac78$ inches (22.54 cm) high. Columns are to be $3\frac14$ inches (8.25 cm) wide, with a $\frac{5}{16}$ inch (0.8 cm) space between them. The main title (on the first page) should begin 1.0 inch (2.54 cm) from the top edge of the page. The second and following pages should begin 1.0 inch (2.54 cm) from the top edge. On all pages, the bottom margin should be 1-1/8 inches (2.86 cm) from the bottom edge of the page for $8.5 \times 11$-inch paper; for A4 paper, approximately 1-5/8 inches (4.13 cm) from the bottom edge of the page. \subsection{Margins and page numbering} All printed material, including text, illustrations, and charts, must be kept within a print area 6-7/8 inches (17.5 cm) wide by 8-7/8 inches (22.54 cm) high. \subsection{Type-style and fonts} Wherever Times is specified, Times Roman may also be used. If neither is available on your word processor, please use the font closest in appearance to Times to which you have access. MAIN TITLE. Center the title 1-3/8 inches (3.49 cm) from the top edge of the first page. The title should be in Times 14-point, boldface type. Capitalize the first letter of nouns, pronouns, verbs, adjectives, and adverbs; do not capitalize articles, coordinate conjunctions, or prepositions (unless the title begins with such a word). Leave two blank lines after the title. AUTHOR NAME(s) and AFFILIATION(s) are to be centered beneath the title and printed in Times 12-point, non-boldface type. This information is to be followed by two blank lines. The ABSTRACT and MAIN TEXT are to be in a two-column format. MAIN TEXT. Type main text in 10-point Times, single-spaced. Do NOT use double-spacing. All paragraphs should be indented 1 pica (approx. 1/6 inch or 0.422 cm). Make sure your text is fully justified---that is, flush left and flush right. Please do not place any additional blank lines between paragraphs. Figure and table captions should be 9-point Roman type as in Figures~\ref{fig:onecol} and~\ref{fig:short}. Short captions should be centred. \noindent Callouts should be 9-point Helvetica, non-boldface type. Initially capitalize only the first word of section titles and first-, second-, and third-order headings. FIRST-ORDER HEADINGS. (For example, {\large \bf 1. Introduction}) should be Times 12-point boldface, initially capitalized, flush left, with one blank line before, and one blank line after. SECOND-ORDER HEADINGS. (For example, { \bf 1.1. Database elements}) should be Times 11-point boldface, initially capitalized, flush left, with one blank line before, and one after. If you require a third-order heading (we discourage it), use 10-point Times, boldface, initially capitalized, flush left, preceded by one blank line, followed by a period and your text on the same line. \subsection{Footnotes} Please use footnotes\footnote {This is what a footnote looks like. It often distracts the reader from the main flow of the argument.} sparingly. Indeed, try to avoid footnotes altogether and include necessary peripheral observations in the text (within parentheses, if you prefer, as in this sentence). If you wish to use a footnote, place it at the bottom of the column on the page on which it is referenced. Use Times 8-point type, single-spaced. \subsection{References} List and number all bibliographical references in 9-point Times, single-spaced, at the end of your paper. When referenced in the text, enclose the citation number in square brackets, for example~\cite{Authors14}. Where appropriate, include the name(s) of editors of referenced books. \begin{table} \begin{center} \begin{tabular}{|l|c|} \hline Method & Frobnability \\ \hline\hline Theirs & Frumpy \\ Yours & Frobbly \\ Ours & Makes one's heart Frob\\ \hline \end{tabular} \end{center} \caption{Results. Ours is better.} \end{table} \subsection{Illustrations, graphs, and photographs} All graphics should be centered. Please ensure that any point you wish to make is resolvable in a printed copy of the paper. Resize fonts in figures to match the font in the body text, and choose line widths which render effectively in print. Many readers (and reviewers), even of an electronic copy, will choose to print your paper in order to read it. You cannot insist that they do otherwise, and therefore must not assume that they can zoom in to see tiny details on a graphic. When placing figures in \LaTeX, it's almost always best to use \verb+\includegraphics+, and to specify the figure width as a multiple of the line width as in the example below {\small\begin{verbatim} \usepackage[dvips]{graphicx} ... \includegraphics[width=0.8\linewidth] {myfile.eps} \end{verbatim} } \subsection{Color} Please refer to the author guidelines on the ICCV 2019 web page for a discussion of the use of color in your document. \section{Final copy} You must include your signed IEEE copyright release form when you submit your finished paper. We MUST have this form before your paper can be published in the proceedings. {\small \bibliographystyle{ieee}

Background {#Sec1} ========== Utilities are commonly used to provide a preference weighted estimate of health-related quality of life (HRQL) benefits provided by treatments in the context of assessing the cost effectiveness of those treatments. They are therefore crucial to health technology assessment when using a cost per quality-adjusted life year framework. Utility values can be generated in several ways, including generic preference-based measures, condition-specific preference-based measures, vignettes and directly-elicited patient utility values. However, the use of different instruments (in the same population) has been shown to produce different results, with a range of utility measures having been used in oncology \[[@CR1]-[@CR3]\]. Health-state classifications for economic modelling are frequently based on a patient's progression status, often the primary endpoint in oncology trials (including melanoma). Melanoma is a form of skin cancer and if identified early the prognosis is generally good \[[@CR4]\]; however, for advanced melanoma survival is poor and options for treatment are limited. In oncology progression is generally accepted to be important for both disease burden and HRQL, with a delay in progression an objective of treatment. During our investigation we sought clinical opinion, which indicated that there is a decline in HRQL in the final months of life of advanced melanoma patients, which may not be appropriately captured solely through the use of progression-based heath states. We therefore undertook the study reported here with two objectives in mind: firstly to report and compare pre- and post-progression health state utilities in advanced melanoma, when generated using different methods - via the EORTC (European Organisation for Research and Treatment of Cancer, Brussels, Belgium), and the SF-36 (QualityMetric, Lincoln, USA); secondly, to explore the validity, in advanced melanoma, of progression-based health-state utility modelling compared to modelling based upon time to death health states. Methods {#Sec2} ======= Regulatory approval for ipilimumab in advanced (unresectable or metastatic) melanoma was based on data from the Phase III clinical trial MDX010-20 (Clinicaltrials.gov Identifier: NCT00094653), conducted in accordance with the Declaration of Helsinki, and the laws and regulatory requirements of the countries where the research was performed. Patients (or their legal representatives) were required to give written informed consent before enrolment, and the protocol approved by institutional review boards at each of the 125 sites involved. All patients enrolled in to the study had unresectable advanced melanoma (disease that had spread from the area of skin where it originated), at stage III (the disease having spread to regional lymph nodes, but no evidence of distant metastases), or stage IV (the disease having spread to distant lymph nodes, or other organs). Trial data were extracted from the study database using SAS® (SAS, Cary, USA) and analysed in Stata® 12.0 (StataCorp, Texas, USA) \[[@CR5]\]. The trial included 676 patients, randomised 3:1:1 to ipilimumab + gp100, ipilimumab only, or gp100 only. Gp100 is an experimental peptide-based vaccine, with few side effects, later shown to be ineffective in increasing overall survival. Approximately 60% of patients were male, with the majority of patients having Eastern Cooperative Oncology Group (ECOG) performance status 0 (55%) or 1 (43%). All patients had received prior systemic therapy for their advanced melanoma -- a complete description of the study is available in Hodi et al., 2010 \[[@CR5]\]. In the study, patients were asked to complete both the EORTC QLQ-C30 and SF-36v2 questionnaires on receipt of the first dose of treatment, at the end of treatment and 12 weeks after treatment. These time points were pre-specified for the first 4 doses, however beyond this point quality of life instruments were only administered when patients received retreatment. The EORTC QLQ-C30, administered in MDX010-20, is frequently used in oncology clinical trials as it captures many of the symptoms commonly seen in cancer patients, such as nausea, pain and fatigue, as well as generic aspects of function (including physical, emotional and role). Rowen et al. generated the EORTC-8D -- a condition-specific preference-based measure derived from the EORTC QLQ-C30 \[[@CR6]\]. A sample of health states were valued by 350 members of the UK general public using time trade-off, and these values were modelled to produce utilities for all health states. In order to generate EORTC-8D utilities from the EORTC QLQ-C30 results, Model 3 from the mapping by Rowen et al. was used \[[@CR6]\]. In the MDX010-20 trial, this questionnaire was completed by 616 patients (1,237 observations). The SF-6D (derived from the SF-36 or SF-12) is a widely-used generic preference-based measure of HRQL, which has been valued by 611 members of the UK general public using standard-gamble methodology for a sample of health states. Several methods are available to convert SF-36 results to SF-6D utilities; this study uses the most recent non-parametric Bayesian method \[[@CR7]-[@CR9]\]. In the MDX010-20 trial, 599 patients completed the SF-36 (1,205 observations). The majority of HRQL questionnaires (both EORTC QLQ-C30 and SF-36) were completed within the first 24 weeks of the trial (before and after patients were administered their initial course of ipilimumab), though observations are available for up to 4 years (5% of patients completed questionnaires after 12 months if receiving retreatment). Where multiple observations were available for a patient, all observations were included in calculations. The level of non-response to the EORTC QLQ-C30 and SF-36 was low within the study, at 7.8% and 11.4%, respectively. We did not attempt to impute missing values, as their low number means they would be unlikely to affect the results of the study, nevertheless this remains a limitation of the work performed. To provide context to the analysis, a literature search was completed in May 2013 for utilities in advanced melanoma (including previously untreated disease) with at least 30 patients in the study. This search included the Medline, Embase, NHS Health Economic Evaluation Database, Cochrane Health Technology Assessment, Database of Abstracts of Reviews of Effect (DARE), Cumulative Index to Nursing and Allied Health Literature (CINAHL) and Econlit databases. Progression status versus time to death {#Sec3} --------------------------------------- In order to ascertain whether the time to a patient's death influenced their quality of life, a series of regression models were fitted in Stata. Patients were split into six groups based on their time to death after the HRQL measurement was taken. The groups were based on clinician feedback on the phases of life generally seen in patients before death (over 180 days, 120--180 days, 90--120 days, 60--90 days, 30--60 days, and under 30 days). Patient time to death was then extracted from the study database and grouped according to these time points. Where a patient was still alive at the end of the study, the final measurement point was used as the time of death. In 91% of cases where censoring occurred, this put patients in the health state furthest from death (\>180 days), with only 2.9% of EORTC-8D and 2.5% of SF-6D of observations categorised as censored before the 180 days from death time point. Given this low level of censoring models accounting for censoring were not required to be fitted. The models fitted to the data were generalised estimating equation (GEE) random effects models to account for correlation between repeated measurements from the same individuals. GEE models were chosen instead of generalised linear mixed models, in order to produce population averages as required for the purpose of health technology assessment and economic modelling. To compare model fit the mean absolute error and the root mean squared error were calculated, with tests for significance using Student's *t*-test with a p-value of 0.05. Results {#Sec4} ======= EORTC-8D versus SF-6D values {#Sec5} ---------------------------- In all treatment groups and disease states, mean utilities generated from the EORTC-8D were significantly higher than those seen with the SF-6D (p \< 0.001), the difference in the values being shown in Figure [1](#Fig1){ref-type="fig"}. Of 807 measurements of HRQL at the same time-point for both instruments, the EORTC-8D utility was higher on 777 occasions (Figure [2](#Fig2){ref-type="fig"}). SF-6D values appear to be more concentrated around the mean: 28% of the SF-6D utilities were within 5% of the mean value, compared to only 13% of the EORTC-8D utilities, though no statistical tests were performed to estimate this formally. Despite the differences in magnitude between the two methods, the change in utility on disease progression was similar with both the EORTC-8D utilities (−4.6%) and SF-6D utilities (−3.3%).Figure 1**Plot of EORTC-8D and SF-6D patient reported utilities by progression status and time to death in the MDX010-20 trial.**Figure 2**Plot of EORTC-8D utility versus SF-6D utility when measured at identical time points, by progression status.** Progression status versus time to death {#Sec6} --------------------------------------- Three models were fitted to the patient-level data for both EORTC-8D and SF-6D in order to predict patient utility, these are shown in Table [1](#Tab1){ref-type="table"}. Model 1 used progression status alone to predict utility, Model 2 used time to death health states only, and Model 3 used both progression status and time to death health states.Table 1**Results of regression analyses based for models, including goodness-of-fit statisticsEORTC-8D derived resultsProgression (model 1)Time to death (model 2)Progression + time to death (model 3)**Constant0.8030.8310.837(SE)(0.006)(0.006)(0.006)Progression−0.048−0.029(SE)(0.007)(0.007)120 - 179 days to death−0.060−0.059(SE)(0.012)(0.012)90 - 119 days to death−0.068−0.065(SE)(0.015)(0.015)60 - 89 days to death−0.112−0.106(SE)(0.016)(0.016)30 - 59 days to death−0.153−0.142(SE)(0.015)(0.015)Under 30 days to death−0.178−0.165(SE)(0.019)(0.019)MAERMSEMAERMSEMAERMSEAll values0.1190.1450.1080.1320.1080.1320.9 \< EORTC-8D ≤ 1 (n = 354)0.1670.1710.1370.1450.1390.1470.8 \< EORTC-8D ≤ 0.9 (n = 280)0.0560.0660.0460.0630.0480.0640.65 \< EORTC-8D ≤ 0.8 (n = 332)0.0610.0730.0770.0880.0740.0860.0 \< EORTC-8D ≤ 0.65 (n = 174)0.2350.2460.2060.2250.2040.222**SF-6D derived results**Progression (model 1)Time to death (model 2)Progression + time to death (model 3)Constant0.6420.6670.670(SE)(0.006)(0.005)(0.005)Progression−0.030−0.016(SE)(0.005)(0.006)120 - 179 days to death−0.050−0.050(SE)(0.010)(0.010)90 - 119 days to death−0.059−0.057(SE)(0.013)(0.013)60 - 89 days to death−0.089−0.085(SE)(0.014)(0.013)30 - 59 days to death−0.119−0.113(SE)(0.013)(0.013)Under 30 days to death−0.134−0.126(SE)(0.016)(0.016)MAERMSEMAERMSEMAERMSEAll values0.0910.1240.1020.1340.1010.1330.75 \< SF-6D ≤ 1 (n = 201)0.2010.2100.1990.2120.1990.2120.65 \< SF-6D ≤ 0.75 (n = 228)0.0500.0570.0470.0650.0470.0650.55 \< SF-6D ≤ 0.65 (n = 435)0.0280.0370.0510.0580.0490.0570 \< SF-6D ≤ 0.55 (n = 208)0.1590.1740.1750.1940.1730.192SE = standard error, MAE = mean absolute error, RMSE = root mean squared error. Model 1 showed that patients experienced a reduction in utility of 0.048 (standard error 0.007) following progression using the EORTC-8D and a reduction in utility of 0.030 (standard error 0.005) using the SF-6D. Clinician feedback following discussion of these results was that patients would be treated according to their expected survival period. These health states were specified by clinicians and then used in Model 2 and Model 3, as described in the Methods section. Model 2 used time to death as a categorical variable, grouping patients by survival period, and showed that patient utility fell consistently as a patient approached death in both EORTC-8D and SF-6D utilities. This slightly improved the predictive power of the regression in the EORTC-8D, and slightly worsened it in the SF-6D. Model 3 included whether a patient had progressed as a dummy variable, and their time to death at the point the measurement was taken, and improved on the time to death model for both EORTC-8D and SF-6D. The results of this time to death including progression model showed a similar pattern of decline in utility as a patient approaches death, with an additional decrement to utility if a patient had experienced disease progression. The fit of each of the models can be seen in Table [2](#Tab2){ref-type="table"}, which compares predicted values for each of the health state based models, to observed values in those patient groups. Model fit can also been seen in Table [1](#Tab1){ref-type="table"}, where the mean absolute error and root mean squared error are given for all values, and also reported separately for approximately equal segments of the utility distributions, which allows a comparison of model fit across the distribution of utility values.Table 2**Predicted versus observed mean utility values for models based on health statesModelHealth stateEORTC-8DSF-6DPredictedObservedPredictedObserved**Model 1Pre-progression0.8030.8130.6420.648Post-progression0.7550.7760.6120.626Model 2180 or more days to death0.8310.8400.6670.672120 - 179 days to death0.7710.7670.6160.61090 - 119 days to death0.7630.7560.6130.60060 - 89 days to death0.7200.7230.5850.57430 - 59 days to death0.6790.6700.5570.541Under 30 days to death0.6530.6510.5440.531Model 3*Pre-progression*180 or more days to death0.8370.8480.6700.677120 - 179 days to death0.7790.7770.6200.61590 - 119 days to death0.7720.7590.6130.59160 - 89 days to death0.7310.7370.5850.58830 - 59 days to death0.6950.6900.5570.554Under 30 days to death0.6720.6290.5440.518*Post-progression*180 or more days to death0.8080.8200.6540.661120 - 179 days to death0.7490.7420.6140.59590 - 119 days to death0.7430.7500.6130.62360 - 89 days to death0.7020.6930.5850.54730 - 59 days to death0.6650.6430.5070.521Under 30 days to death0.6420.6750.5340.547 Models including death as a continuous variable, or restricting analysis only to patients who died in the study period did not produce substantially different results (results not presented). Adding further clinically-relevant variables (treatment assignment, prior Interleukin-2 usage, age, ECOG status or melanoma stage: all measured at baseline) did not result in significant coefficients or increase the predictive power of the model. These additional variables were therefore not included in the final models, but are provided in Additional file [1](#MOESM1){ref-type="media"}: Table S1 for completeness. Fitting models separately for pre- and post-progression utilities when including time to death yielded similar coefficients for pre-progression and post-progression - the consistency of parameter estimates and standard errors across models when including progression status (moving from Model 2 to Model 3) indicate low levels of multi-collinearity between progression status and time to death. Utility remained high until approximately 180 days before death with both the EORTC-8D and SF-6D (0.840 and 0.672), similar to values seen in the general population (0.80 for a 55--64 year old using the EQ-5D) \[[@CR10]\]. From this point onwards HRQL continually decreased, with the final 30 days of life having the largest decrease compared to patients 180 days or more from death (EORTC-8D −0.189 and SF-6D −0.141) (Table [1](#Tab1){ref-type="table"}). All time based health state coefficients were significantly lower than the 180 days before death health state (p \< 0.001 for all coefficients), for both EORTC-8D and SF-6D. For the EORTC-8D, the time to death health state approach provides a more accurate estimate of patient utility than progression status according to the mean absolute error and root mean squared error (particularly for values further from the mean). Using both factors further increases accuracy of estimates. For the SF-6D, using time to death as health states resulted in predictions on average slightly worse than the progression based model. Using both time to death and progression improved the accuracy of estimates compared to time to death alone, though did not provide as good a prediction as progression only. Discussion {#Sec7} ========== EORTC-8D versus SF-6D utility values {#Sec8} ------------------------------------ No studies are presently available that directly compare the EORTC-8D to the SF-6D. We observed that EORTC-8D utilities were higher than SF-6D utilities in both the pre- and post-progression states. This difference in utilities derived from the two preference-based measures adds to the body of literature that suggests different measures may produce different utilities. EQ-5D, which like EORTC-8D is valued using time trade-off, has been shown to produce different utilities to SF-6D (both higher and lower) in rheumatoid arthritis, constipation and herniated discs \[[@CR1]-[@CR3],[@CR11]\]. A potential explanation for the difference in utilities generated by the EORTC-8D is the classification system used - this was originally derived using a multiple-myeloma dataset \[[@CR6]\]. When compared to the original data, patients within the MDX010-20 dataset were more likely to respond that they have the best level of health, and less likely to report that they have the lowest level of health for the physical, role functioning, and pain dimensions, than those from the EORTC-8D source dataset. Although this does not in itself mean that the classification system is not appropriate, further research is encouraged to investigate whether the EORTC-8D has desirable psychometric properties when applied to other cancers. Investigation of the properties of EORTC-8D- is particularly important as the EORTC QLQ-C30 is a frequently-used measure in oncology clinical trials, therefore utilities from the EORTC-8D will often be available, whereas a generic instrument (such as the SF-36 or EQ-5D) may not be included in a study. Equally, for health technology assessment purposes, it is likely to be important to understand how the EORTC-8D relates to the SF-6D, Health Utilities Index 3 (HUI-3) and, most notably, the EQ-5D (the preferred measure of the National Institute of Health and Care Excellence \[NICE\], and most frequently used utility measure in health technology assessment studies internationally) \[[@CR12]-[@CR14]\]. At present two studies are available comparing the EORTC-8D and EQ-5D, though only in two types of cancer \[[@CR13],[@CR15]\], however there is the potential for further analyses from a recently completed study of dabrafenib compared to dacarbazine in advanced melanoma, the BREAK-3 study, as this included both EORTC-8D and EQ-5D \[[@CR16]\]. Patient derived data compared to published utility values {#Sec9} --------------------------------------------------------- The literature reviews identified four studies reporting utilities in advanced melanoma. A mapping study by Askew et al. reported a utilities of 0.85 for stage III melanoma (n = 100) and 0.86 for stage IV melanoma (n = 102).\[[@CR17]\] Dixon et al., in a trial of interferon-alpha compared to placebo, collected EQ-5D over time, increasing from 0.76 at baseline (n = 111), to 0.77 at 3 months (n = 80), 0.82 at 12 months (n = 66), and finally to 0.83 at 24 months (n = 31) \[[@CR18]\]; personal communication with the author. Two vignette studies were also available (bespoke descriptions of health states that are valued by the general public) -- one based in Australia/UK (n = 77/63) published by Beusterien et al. \[[@CR19]\], and one based in Canada (n = 87). \[[@CR20]\] These types of studies are often used in health-state valuation for cost-effectiveness analyses \[[@CR21]-[@CR23]\], especially when patient derived utilities are not available \[[@CR24]\]. In the two vignette studies, members of the public were asked to value partial response, giving values of 0.85 and 0.84 in Australia/UK and Canada, stable disease (0.77 and 0.79), progressive disease (0.59 and 0.55), and best supportive care (0.59 and 0.54). The values seen in MDX010-20 were broadly comparable to the values used in the published literature. Although much higher than utilities typically seen in advanced cancer \[[@CR25]\], that values in multiple studies are similar may indicate that patients with advanced melanoma exhibit HRQL similar to age-matched members of the general population \[[@CR10]\]. Of particular interest are the values seen in Dixon et al., which showed the quality of life of patients to increase the longer patients were in the study, potentially indicating a link between proximity to death, and low HRQL. Progression versus time to death approach {#Sec10} ----------------------------------------- Previous reviews have provided evidence for a positive relationship between HRQL measures and the survival duration of cancer patients \[[@CR26],[@CR27]\]. A recent review identified three studies in either advanced or metastatic melanoma, all of which show either overall HRQL or certain domains of HRQL to be predictive of survival \[[@CR28]\]. The analysis we have performed is therefore consistent with this body of work and attempts to quantify this relationship. Utilities generated by both the EORTC-8D and SF-6D showed a substantial decrease in utility in the final 180 days before a patient's death. This was consistent across measures and regression models, and, based on clinician feedback it is possible that this is the case for all metastatic melanoma patients. However, further research is required using different datasets to establish whether this is the case. Progression status is a standard primary endpoint used in clinical trials to define efficacy, and as such has been the obvious choice for defining health states in cost-effectiveness modelling. Oncology models are frequently based on area under the curve models with three health states (pre-progression, post-progression and death), with transitions driven by parametric curves. However, our research suggests that this approach may not be suitable to model patients' HRQL if disease progression is not closely related to HRQL. In this case additional health states may need to be added to model the path of HRQL. If this pattern of HRQL being linked to survival is repeated in other oncology areas, it may be better to replace the current standard method of progression-based utilities, with analysis on a case-by-case basis - including investigation of alternative explanations for HRQL decreases (for example a time to event based approach). Designers of oncology studies should carefully consider the time points at which HRQL is measured, in order to ensure accurate estimates of HRQL can be made both pre- and post-progression. Investigators should also examine utility datasets with a clinical rationale for analyses, and not purely through an economic lens. Whilst it may be that the progression-based model remains a good fit for many cancers, this should be shown with data rather than assumed to be the case. It is important to note that the effect on the results of economic modelling when using a time to death approach (rather than progression-based approach) for HRQL in a model may be dependent upon the length of time patients live in a post-progression state. Within the MDX010-20 dataset, 17% of ipilimumab treated patients were alive at the end of the 56-month trial period - nearly all of these in a post-progression state \[[@CR5]\]. The use of a lower utility than these patients experienced (as would occur when taking the mean utility measurement for progressive disease) over such a substantial time period will substantially worsen the modelled cost effectiveness of a drug. A time-to-death utility may assign a lower utility to patients only shortly before death, thereby increasing the number of modelled quality-adjusted life years, and potentially altering investment decisions made as a result. It should be noted, however, that the cost-effectiveness of treatments would not always improve, and the results seen would depend on individual analyses -- a treatment with worse post-progression survival may become less cost-effective. It is possible that the results seen in MDX010-20 would not be reproduced in other datasets. The MDX010-20 clinical trial was extremely mature, with approximately 90% of enrolled patients having died on completion at 56 months (giving many data points for analysis). Equally the underlying disease of melanoma may result in a different quality of life profile to solid tumours or haematological malignancies, which have been more widely studied. It is also plausible that the results seen were impacted by the treatments used in the MDX010-20 study. In MDX010-20 the control arm was relatively begin (a vaccine with few side effects, later shown to be ineffective). The study drug, ipilimumab, has a defined course of 4 infusions (not on-going treatment, with corresponding adverse events), and a relatively good safety profile. Studies with more toxic chemotherapies including frequent administrations (and more aggressive control arms) may show different results. Finally, the use of ipilimumab may affect pre- and post-progression utility through a delayed response due to the drug's mechanism of action -- as immunotherapy ipilimumab acts indirectly on cancer cells via the immune system. This means that patients may progress before showing a response at which point their symptoms (and hence HRQL) improve, owing to the delayed reaction of treatment \[[@CR27]\]. Although the control arm (gp100) did not show a different pattern to ipilimumab treatment, this may be due to sample size; four-fifths of the patients received ipilimumab, either as monotherapy or in combination with gp100. This delayed reaction therefore may also be sufficient to cause the relatively high quality of life on disease progression. Conclusions {#Sec11} =========== This study adds to the body of evidence that suggests alternative quality of life measures produce different results. This can therefore be seen as a validation of the position of the Scottish Medicines Consortium and NICE in requesting a consistent source of utility values \[[@CR12],[@CR29]\]. The results seen in this study also indicate further research is needed on the relationship between the EORTC-8D, SF-6D and EQ-5D. In addition, this study demonstrated that utilities based on time to death appear to provide a good fit to patient data in the MDX010-20, when compared to utilities based on disease progression. Practitioners should carefully analyse HRQL data prior to constructing economic models based on clinical trials; clinical measures such as disease progression may not explain quality of life changes, and an event-based approach may be more suitable. Additional file {#Sec12} =============== Additional file 1: Table S1.Results of regressions showing the impact of adding melanoma stage, prior interleukin-2, age, and treatment assignment. **Competing interests** AJH, BP and DL work for BresMed, who were reimbursed by Bristol-Myers Squibb Pharmaceuticals Limited (BMS) for time spent performing the analysis and preparing the manuscript. ML is an employee of BMS, and LP was an employee of BMS at the time of the analysis. DR has no competing interests. **Authors' contributions** AJH, DL and LP conceived the study, with access to data co-ordinated by LP and ML. Statistical analysis was performed by AJH, DL and BP. Further analysis and interpretation was provided by DR and ML. The first draft of the manuscript was written by AJH, with input from DL and DR. All authors read and approved the final manuscript. The authors would like to extend our thanks to the journal reviewers and editor, whose insights have prompted further analyses, and made the paper more comprehensive and generalisable. We would also like to recognise the contribution of the clinicians who gave input to the project, Professor Ron Akehurst for aiding the design of the study, Owain Saunders for statistical advice, Professor Simon Dixon for providing additional data, and Susan Moloney in critiquing the analyses performed. Funding for this study was provided by Bristol-Myers Squibb Pharmaceuticals Limited (BMS), the manufacturer of ipilimumab.

INTRODUCTION {#s1} ============ Lipid droplets (LDs) are cytoplasmic lipid storage organelles. They are made of a neutral lipid core, mainly triglycerides and esterified cholesterol, surrounded by a monolayer of polar lipids ([@b29]; [@b81]; [@b93]). Even though the brain has a low amount of LDs under resting conditions ([@b27]; [@b102]), LD content rises in brain tumours ([@b6]; [@b67]; [@b75]). LDs are also reported to accumulate in the brains of Alzheimer\'s patients ([@b33]) and are proposed to be the site where aggregation of α-synuclein occurs in Parkinson\'s patients ([@b23]). The importance of LDs in the brain is also highlighted by the observation that seipin and spartin, proteins whose mutation is associated with neuronal diseases, target to LDs ([@b25]; [@b83]). Deregulated lipid metabolism is also thought to be involved in other central nervous system disorders including Niemann--Pick diseases, schizophrenia and epilepsy ([@b1]). Therefore, understanding the mechanisms controlling LD homeostasis in the brain is essential to determine the role of LDs in the onset and progression of a variety of neurological conditions. Although details are currently lacking, LD formation is thought to start with the accumulation of neutral lipids between the two leaflets of the endoplasmic reticulum (ER) bilayer and the recruitment of LD-associated proteins ([@b17]; [@b31]; [@b80]; [@b88]). Budding and scission would lead to small LDs surrounded by a lipid monolayer that could then travel on microtubules ([@b73]; [@b84]). LD mobility can contribute to lipid homeostasis by changing their cytoplasmic distribution ([@b68]), and by facilitating contact and lipid exchange with other LDs ([@b15]; [@b34]), and with the ER ([@b57]), mitochondria ([@b65]) and peroxisomes ([@b11]). LD homeostasis has principally been studied in adipocytes, which, under resting conditions, have one large LD with a diameter of up to 100 µm. In contrast, non-adipocyte cells generally contain several LDs, with diameters in the low micrometre range ([@b81]). This not only results in a major difference in surface-to-volume ratio, and hence the accessibility to the LD core, but also a difference in the intrinsic curvature of LDs. Distinct mechanisms are thus likely to control LD homeostasis in adipocytes and non-adipocyte cells. Proteins of the PAT family (perilipin, ADRP, TIP47, S3-12 and LSDP5, recently renamed PLIN1--PLIN5, respectively) share sequence similarity and the capacity to bind LDs ([@b9]). Of these, only PLIN2 (ADRP) and PLIN3 (TIP47) are ubiquitously expressed. Both are thought to promote LD growth ([@b18]; [@b19]; [@b47]; [@b60]; [@b82]), but to label somewhat distinct LD populations ([@b66]). Proteins of the GTPase regulator associated with FAK (GRAF) family \[GRAF1 (also called ARHGAP26), GRAF2 (also called ARHGAP10), GRAF3 (also called ARHGAP42) and oligophrenin-1 (OPHN1)\] have N-terminal membrane-binding BAR and PH domains and a central GAP domain, with *in vitro* specificity for small Rho GTPases ([@b5]; [@b10]; [@b36]; [@b76]; [@b87]). GRAF1, GRAF2 and GRAF3 possess a C-terminal SH3 domain, whereas OPHN1 has a proline-rich domain instead. Relatively little is known about their functions. However, they are associated with several diseases, suggesting they are fulfilling important roles *in vivo*. In the case of GRAF1, reduced expression is reported in metastatic brain tumours from primary lung adenocarcinoma ([@b101]), in multiple cases of myelodysplastic syndrome/acute myeloid leukemia ([@b13]; [@b14]; [@b74]; [@b94]), and in patients with X-linked alpha-thalassemia mental retardation syndrome ([@b7]). In addition, anti-GRAF1 autoantibodies are found in patients with inflammatory cerebellar ataxia ([@b41]; [@b40]). Two GRAF1 isoforms, GRAF1a and GRAF1b, are listed in databases, and a third isoform, GRAF1c, is reported ([@b7]). GRAF1b associates with dynamic tubules and vesicles ([@b54]), but the roles of GRAF1a and GRAF1c have not been examined so far. We report here that, whereas GRAF1c shows a similar distribution to GRAF1b, GRAF1a specifically associates with LDs and is a regulator of LD homeostasis. RESULTS {#s2} ======= GRAF1a is a LD-associated protein enriched in neonatal brain {#s2a} ------------------------------------------------------------ Previous studies have shown that GRAF1 is enriched in mammalian brain ([@b36]; [@b86]). We therefore amplified *GRAF1* sequences from a brain cDNA library. We sequenced 20 clones and identified three splice variants, GRAF1a, GRAF1b and GRAF1c. The proteins encoded differed in a region with no predicted secondary structure preceding the terminal SH3 domain ([Fig. 1A](#f01){ref-type="fig"}). The most abundant cDNAs corresponded to the shorter isoforms (11/20 clones for GRAF1b; 8/20 for GRAF1c). In comparison to GRAF1b, GRAF1c lacked a 37-amino-acid stretch enriched in serine and proline residues (S/P, [Fig. 1B](#f01){ref-type="fig"}). In comparison to GRAF1b, GRAF1a contained an additional 55-amino-acid segment highly enriched in hydrophobic residues (Hyd, [Fig. 1A--C](#f01){ref-type="fig"}). Only 1/20 clones corresponded to GRAF1a, suggesting that it is the least abundant in adult brain. ). (G) Single focal plane images of live HeLa cells expressing GFP-tagged GRAF1a, b or c (see corresponding [supplementary material Movies 1--3](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Scale bars: 15 µm. (H,I) Maximum intensity projections of images acquired by structured illumination microscopy (SIM) of HeLa cells expressing Myc--GRAF1a (H) or untagged GRAF1a (I), respectively, stained with anti-Myc or anti-GRAF1 antibodies, and using BODIPY 493/503 to visualize LDs. Scale bars: 10 µm. Boxed areas in G--I are enlarged at the bottom right of each image. (J) Separation of cytoplasmic components of oleic-acid-fed primary glial cells by density gradient centrifugation. The input and odd fractions were analysed by western blotting, showing relative enrichment of endogenous GRAF1a and PLIN3 at the top of the gradient (fraction 1). Actin and Rab11 were used as negative controls.](jcs-127-21-4602-f01){#f01} GRAF1a, GRAF1b and GRAF1c could be discriminated by their sizes ([Fig. 1D](#f01){ref-type="fig"}). In agreement with the PCR data, in adult mouse or rat brain protein extracts, GRAF1 migrated as two major bands, corresponding to the molecular masses of GRAF1b and GRAF1c ([Fig. 1E](#f01){ref-type="fig"}; [supplementary material Fig. S1A,B](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Interestingly, however, the pattern of GRAF1 expression was different in developing mouse brain. Expression of the longer isoform (GRAF1a) started before embryonic day 16 (E16) and increased postnatally, after which there was a switch to shorter isoforms in adults ([Fig. 1E](#f01){ref-type="fig"}). Immunoprecipitation of postnatal day 7 (P7) brain extracts with an anti-GRAF1 antibody resulted in a protein with the same molecular mass as GRAF1a ([supplementary material Fig. S1A](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). GRAF1b and GRAF1c are thus the major GRAF1 isoforms in adult brain, whereas GRAF1a is abundant in neonates. GRAF1a was also predominant in primary cells isolated from the brains of E18 embryos ([Fig. 1F](#f01){ref-type="fig"}). At equal protein amounts, glial cultures devoid of neurones contained more GRAF1 than co-cultures, suggesting that GRAF1a is enriched in glial cells ([Fig. 1F](#f01){ref-type="fig"}). When overexpressed, GFP-tagged GRAF1b and GRAF1c labelled tubules and small vesicles ([Fig. 1G](#f01){ref-type="fig"}). As previously reported for GRAF1b ([@b54]), they were dynamic (data not shown). In striking contrast, GRAF1a mostly labelled the membrane surrounding spherical cytoplasmic inclusions ([Fig. 1G](#f01){ref-type="fig"}), which were relatively immobile (data not shown). Similar structures were labelled when GRAF1a--GFP was transfected in rat pheochromocytoma PC12, human neuroblastoma SH-SY5Y, murine NIH 3T3 fibroblasts and African green monkey epithelial BSC1 cells ([supplementary material Fig. S1C](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Using Myc-tagged GRAF1a and the neutral lipid probe BODIPY 493/503 ([Fig. 1H](#f01){ref-type="fig"}), these organelles were identified as LDs. Similarly, overexpressed untagged GRAF1a was found on LDs in HeLa cells ([Fig. 1I](#f01){ref-type="fig"}), human glioblastoma U-87 MG cells ([supplementary material Fig. S1D](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)) and primary glial cells ([supplementary material Fig. S1E,F](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). A reticular background staining, typical of the ER, was seen upon staining with anti-GRAF1 antibodies when the LD number was low. In order to examine the subcellular distribution of endogenous GRAF1a, cytoplasmic extracts of primary glial cells were fractionated on a sucrose gradient. Similar to what has been previously reported ([@b16]), LDs were at the top of the gradient, as fraction 1 was enriched in PLIN3 ([Fig. 1J](#f01){ref-type="fig"}). GRAF1a was also enriched in fraction 1 ([Fig. 1J](#f01){ref-type="fig"}). Collectively, these experiments show that GRAF1a is a GRAF1 isoform enriched in the brain of neonates that associates with LDs under resting conditions, in a variety of cell lines and primary cells. Specific association of GRAF1a with LDs relies both on an internal hydrophobic linker and its BAR and PH domains {#s2b} ---------------------------------------------------------------------------------------------------------------- Microscopy observation showed that GRAF1a and GRAF1b are both membrane-binding proteins, but they do not target the same intracellular compartments. In addition, unlike GRAF1b, the association of GRAF1a with membranes was resistant to sodium carbonate extraction ([supplementary material Fig. S1G](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.074575/-/DC1)). In comparison to GRAF1b, GRAF1a has a short additional internal linker ([Fig. 1A](#f01){ref-type="fig"}). This linker therefore mediates binding to membranes through hydrophobic interactions. In addition, given that, unlike GRAF1a ([Fig. 2A](#f02){ref-type="fig"}), GRAF1b did not associate with LDs ([Fig. 2B](#f02){ref-type="fig"}), this linker is required for LD targeting. It was also sufficient to drive the SH3 domain of GRAF1 to LDs ([Fig. 2C,D](#f02){ref-type="fig"}), although, in comparison to full-length GRAF1a, the targeting was less specific. {#f02} Deletion of the SH3 or of the GAP domain of GRAF1a did not perturb its distribution ([Fig. 2E,F](#f02){ref-type="fig"}). However, deletion of either membrane-binding region led to background binding to networks of internal membranes ([Fig. 2G,H](#f02){ref-type="fig"}). GRAF1a-ΔPH showed extensive colocalization with the mitochondrial protein COX IV ([supplementary material Fig. S1H](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). GRAF1a-ΔBAR colocalized with the Golgi protein GM130, but also with COX IV and the ER protein calreticulin ([supplementary material Fig. S1I](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). When LDs were enlarged by incubating cells with oleic acid, GRAF1a-ΔPH acquired a more exclusive LD localization ([supplementary material Fig. S1J](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). GRAF1a-ΔBAR, however, still showed residual binding to other membranes, suggesting that it had a substantially reduced affinity for LDs ([supplementary material Fig. S1K](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Western blot analysis of cell extracts confirmed that there was little cleavage of the overexpressed proteins and that the GFP fusions examined thus corresponded to the GRAF1a deletion mutants ([supplementary material Fig. S1L](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). In addition, fractionation of cytoplasmic extracts of transfected U-87 MG cells showed that GRAF1a--GFP, GRAF1a-ΔPH--GFP, GRAF1a-linker-SH3--GFP and GRAF1a-ΔBAR--GFP were all found in the LD-enriched fraction, although the latter two were present to a lesser extent ([supplementary material Fig. S1M](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). These experiments show that no single domain of GRAF1a is sufficient for specific LD binding. Its hydrophobic linker is the only necessary region, but its BAR and PH domains are both required to restrict the distribution of GRAF1a to the LD surface. GRAF1a overexpression induces an increase in LD volume {#s2c} ------------------------------------------------------ Within a cell, the LD population is heterogeneous in terms of lipid and protein composition; this has been proposed to reflect the fact that individual LDs can be in different metabolic states ([@b58]; [@b66]; [@b79]; [@b96]). Under resting conditions, when HeLa cells were transfected with GRAF1a, the protein was found on most LDs, as visualized with lipophilic probes, such as BODIPY 493/503 ([Fig. 1H,I](#f01){ref-type="fig"}; [supplementary material Fig. S1D--F](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)) or LipidTOX Red ([Fig. 3A](#f03){ref-type="fig"}). If LDs were detected with anti-PLIN2 and anti-PLIN3 antibodies, some LDs had the three proteins ([Fig. 3B](#f03){ref-type="fig"}, boxes labelled 1), some were only positive for GRAF1a ([Fig. 3B](#f03){ref-type="fig"}, boxes labelled 2), and a few were devoid of GRAF1a ([Fig. 3B](#f03){ref-type="fig"}, boxes labelled 3). As previously reported, some LDs contained PLIN2 and not PLIN3, and vice-versa ([Fig. 3B](#f03){ref-type="fig"}). PLIN3 was slightly more abundant on GRAF1a-positive LDs than PLIN2 ([supplementary material Fig. S2A](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), but neither protein was found on all GRAF1a LDs. These observations thus extend LD heterogeneity to the presence or absence of GRAF1a. {#f03} In some cells, in addition to LDs, GRAF1a was found on small tubules that sometimes seemed to extend from LDs ([Fig. 2A](#f02){ref-type="fig"}; [Fig. 3C](#f03){ref-type="fig"}). These tubules, however, were not stained with neutral lipid probes ([Fig. 3C](#f03){ref-type="fig"}; [supplementary material Fig. S2B](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), did not contain PLIN2 ([supplementary material Fig. S2C](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)) or PLIN3 ([Fig. 2A](#f02){ref-type="fig"}), and were not induced during phases of LD growth (see below). Whether these tubules were artefacts of the overexpression of GRAF1a leading to its association with endogenous GRAF tubules or a specific structure formed by GRAF1a will require further exploration. When the size and number of LDs, identified by staining with LipidTOX Red, or as PLIN2-positive or PLIN3-positive objects, were quantified and compared between untransfected and GRAF1a--GFP-expressing cells, GRAF1a induced a significant increase in total LD volume ([Fig. 3D](#f03){ref-type="fig"}). In the case of LipidTOX Red and PLIN2, it did so through an increase in the mean volume of LDs ([Fig. 3E](#f03){ref-type="fig"}), whereas it mostly increased the number of PLIN3-positive structures ([Fig. 3F](#f03){ref-type="fig"}). Similar results were obtained when LipidTOX Red, BODIPY 493/503 or PLIN3-positive LD properties were compared between Myc--GRAF1a-transfected and untransfected cells ([supplementary material Fig. S2D--F](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), ruling out an effect of the tag on this phenotype. In addition, as GRAF1a overexpression did not change the efficiency of binding of PLIN2 and PLIN3 to LDs ([supplementary material Fig. S2G](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), nor did overexpressed GRAF1a sterically hinder access to the LD surface. This size increase was specific to the LD-targeting isoform of GRAF1, and was not caused by transfection per se, because GRAF1b expression did not change LD properties ([Fig. 3D--F](#f03){ref-type="fig"}; [supplementary material Fig. S2H](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). GRAF1a overexpression promotes an increase in LD size upon fatty acid loading of cells {#s2d} -------------------------------------------------------------------------------------- In order to follow the behaviour of GRAF1a during LD formation, GRAF1a--GFP-transfected cells were incubated with oleic acid and BODIPY 558/568 C~12~ (hereafter called BODIPY C~12~), a fluorescent fatty acid analogue that incorporates into LDs with similar kinetics to unmodified fatty acids ([@b51]; [@b90]) ([Fig. 4A,B](#f04){ref-type="fig"}; [supplementary material Movie 1](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Within 5--15 min of addition, pre-existing LDs were labelled with BODIPY C~12~ ([Fig. 4B](#f04){ref-type="fig"}, arrows). At this time, new GRAF1a-positive puncta were seen ([Fig. 4B](#f04){ref-type="fig"}, boxed). After 20--30 min, small BODIPY-C~12~-positive LDs appeared, sometimes co-stained with GRAF1a, and sometimes not ([Fig. 4B](#f04){ref-type="fig"}, arrowheads). Thereafter, this large number of small LDs transformed into fewer larger LDs, which, after 1 h, were almost all positive for GRAF1a. Interestingly, in untransfected cells, the density of LDs did not diminish as clearly, suggesting that this decrease was stimulated by GRAF1a ([Fig. 4C](#f04){ref-type="fig"}; [supplementary material Fig. S2I](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). {#f04} When the size and number of LDs was quantified in fixed cells after a 1-h incubation with oleic acid, the mean LD volume was increased in cells overexpressing GRAF1a ([Fig. 4D](#f04){ref-type="fig"}; [supplementary material Fig. S2J](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). However, this did not lead to an increase in total LD volume ([Fig. 4E](#f04){ref-type="fig"}) as the number of distinct LDs was reduced ([Fig. 4F](#f04){ref-type="fig"}). Similarly to resting cells, the efficiency of binding of PLIN3 to LDs was not altered by GRAF1a ([supplementary material Fig. S2K](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Overexpression of GRAF1a therefore does not change the capacity of cells to take up fatty acids, or the efficiency of incorporation into LDs, but rapidly leads to a smaller number of larger LDs. Overexpression of GRAF1a reduces LD motility {#s2e} -------------------------------------------- The precise mechanism responsible for LD growth is not known, but it is thought to proceed through a combination of local lipid synthesis, lipid transfer and LD fusion ([@b97]). The latter two will depend on the ability of a LD to reach its target and tether itself onto it, and are thus tightly linked to LD mobility. Overexpression of GRAF1a induced a significant decrease in mean LD velocity ([Fig. 5A](#f05){ref-type="fig"}; [supplementary material Fig. S2L, Movie 2](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). As above with fixed cells, GRAF1a overexpression also induced an increase in mean LD area, but not of total LD area ([Fig. 5B,C](#f05){ref-type="fig"}). This was specific to GRAF1a, because it was neither induced by GRAF1b nor by GRAF2 ([Fig. 5B,C](#f05){ref-type="fig"}). {#f05} When cells were incubated with fatty acids, LDs were often seen as tight clusters ([Fig. 5D](#f05){ref-type="fig"}). This was particularly striking in cells overexpressing GRAF1a ([supplementary material Fig. S2M](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)) and partially explained the increase in mean LD size. Indeed, LDs were so tightly bound to each other that they sometimes could not be discriminated as independent objects during the analysis of the images. LDs came together as a result of directed movements ([Fig. 5E](#f05){ref-type="fig"}) or of large-scale cytoplasmic rearrangements ([Fig. 5F](#f05){ref-type="fig"}), and remained bound to each other, often without clearly showing signs of fusion. As a result, when cells were incubated with fatty acids for 24 h and imaged, LDs were clustered in the perinuclear region. This was seen both in control and in GRAF1a-overexpressing cells ([Fig. 5G](#f05){ref-type="fig"}; [supplementary material Fig. S2N](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). At this stage, LDs did not disperse upon treatment with nocodazole, suggesting that it could not be explained by tethering to the microtubule network ([supplementary material Fig. S2O](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). These experiments thus show that during phases of LD growth, LDs cluster. GRAF1a overexpression limits LD mobility and potentiates their clustering. GRAF1-knockout mice do not show a strong abnormal phenotype {#s2f} ----------------------------------------------------------- In order to examine whether GRAF1a plays a role in LD growth *in vivo*, a knockout mouse was engineered ([supplementary material Fig. S3A](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1); [Fig. 6A](#f06){ref-type="fig"}). Knockout pups were indistinguishable from their wild-type littermates. GRAF1-knockout mice developed normally, and were fertile and of similar weights as wild-type animals ([supplementary material Fig. S3B](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Histological analysis of various organs revealed no abnormality in tissue architecture ([supplementary material Fig. S3C](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). GRAF1 is therefore not essential for mouse development. Thin-layer chromatography (TLC) analysis showed that overall, there were no significant lipid differences between GRAF1^+/+^ and GRAF1^−/−^ mice, whether looking at tissues from P7 neonates or adults ([Fig. 6B,C](#f06){ref-type="fig"}; [supplementary material Fig. S3D--F](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). In addition, triglyceride levels in their brains were low. In agreement with this, Oil Red O staining of P7 brain sections showed only a few LDs in cells lining the ventricles and no gross difference between knockout and wild-type animals ([Fig. 6D](#f06){ref-type="fig"}). ). As the triglyceride spot intensity of brain samples is at background level, the corresponding mean was not included (\* in B). The spots indicated by \*\* in C co-migrate with fatty acids, but because they are absent from P7 brain and from all other organs ([supplementary material Fig. S3D](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), they probably correspond to sulfatides, a major component of myelin ([@b22]; [@b70]). (D) Oil Red O staining of GRAF1^+/+^ and GRAF1^−/−^ brain sections from P7 pups. Boxed areas are magnified at the top left of each image. Scale bars: 50 µm.](jcs-127-21-4602-f06){#f06} Partial compensation of GRAF1b and GRAF1c by GRAF2, which has a similar domain organization and was also expressed in adult brain ([Fig. 1E](#f01){ref-type="fig"}) could not be ruled out, but there was no change in its expression level ([supplementary material Fig. S3G](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). GRAF2 was, however, unlikely to compensate for GRAF1a, not only because GRAF2 did not target LDs ([supplementary material Fig. S3H](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), but also because it was absent from brain extracts of P7 pups, where GRAF1a was abundant ([Fig. 1E](#f01){ref-type="fig"}). Primary GRAF1^−/−^ glial cells show impaired LD accumulation upon incubation with oleic acid {#s2g} -------------------------------------------------------------------------------------------- We decided to compare LD accumulation in primary glial cultures isolated from GRAF1^+/+^ and GRAF1^−/−^ embryos. In agreement with the histology analysis, GRAF1^+/+^ and GRAF1^−/−^ cells contained few LDs under resting conditions ([supplementary material Fig. S3I](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Upon incubation with oleic acid there was an accumulation of LDs in primary glia ([Fig. 7A](#f07){ref-type="fig"}). When imaged during their growth phase, LDs were strikingly immobile ([Fig. 7A](#f07){ref-type="fig"}) but knocking out GRAF1 increased the mean LD velocity ([Fig. 7B](#f07){ref-type="fig"}). Live imaging showed that in GRAF1^−/−^ cells, LDs had more freedom of movement around their point of anchor ([supplementary material Movie 3](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). The presence of GRAF1 thus restricts LD mobility. {#f07} When primary glia were incubated with oleic acid for 48 h, the total LD volume in GRAF1^−/−^ cells was significantly smaller than in GRAF1^+/+^ cells ([Fig. 7C,D](#f07){ref-type="fig"}). Transfection of untagged GRAF1a in GRAF1^−/−^ cells promoted LD clustering and accumulation ([Fig. 7E](#f07){ref-type="fig"}; [supplementary material Fig. S3J](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). When transfected into GRAF1^+/+^ cells, GRAF1a also induced LD clumping 24 h after oleic acid addition ([Fig. 7F](#f07){ref-type="fig"}), but did not further enhance the clustering seen at 48 h ([supplementary material Fig. S3K](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). These experiments thus show that, similar to what we observed in HeLa cells, GRAF1a regulates LD mobility, facilitates inter-LD contacts and contributes to LD growth in primary glial cells. Overexpression of GRAF1a prevents LD dispersion after the washout of fatty acids {#s2h} -------------------------------------------------------------------------------- After 24 h of oleic acid loading, LDs accumulated *en masse* in a pericentriolar region ([Fig. 5G](#f05){ref-type="fig"}). When fatty acids were removed from the medium, LDs dispersed ([Fig. 8A](#f08){ref-type="fig"}). This is similar to what happens in PLIN1-transfected fibroblasts when lipolysis is stimulated ([@b55]). Strikingly, after a 24-h wash, dispersion did not occur in cells overexpressing GRAF1a--GFP ([Fig. 8A](#f08){ref-type="fig"}). This was also seen with untagged GRAF1a ([supplementary material Fig. S4A](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)) but not with GFP ([supplementary material Fig. S4B](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). Imaging of cells by super-resolution structured illumination microscopy (SIM) showed that GRAF1a proteins were clustered at inter-LD contacts, whereas PLIN3 was distributed on the whole surface of the LDs, suggesting that GRAF1a participates in tethering LDs together ([Fig. 8B,C](#f08){ref-type="fig"}). In comparison to untransfected cells, expression of GRAF1a led to the tight association of LDs into a smaller number of larger clusters, whether LDs were identified as BODIPY-C~12~-positive structures ([Fig. 8D,E](#f08){ref-type="fig"}) or PLIN3-positive objects ([supplementary material Fig. S4C,D](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), but did not alter the binding efficiency of PLIN3 to LDs ([supplementary material Fig. S4F](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). In addition, GRAF1a overexpression attenuated the decrease in total LD volume seen upon fatty acid washout ([Fig. 8F](#f08){ref-type="fig"}; [supplementary material Fig. S4E](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)), suggesting an inhibition of lipolysis. In agreement with this, although there was no difference in the amount of triglycerides stored by cells after the uptake, the triglyceride levels remaining in GRAF1a-overexpressing cells after the chase was higher than in controls ([supplementary material Fig. S4G](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). The magnitude of the effect was relatively small, partly because it was a mixed population of transfected and untransfected cells. Live imaging of cells during the wash showed that, in untransfected cells, LDs were independent and highly mobile, whereas in cells overexpressing GRAF1a they remained as clusters and were relatively immobile ([Fig. 8G](#f08){ref-type="fig"} and [supplementary material Movie 4](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). {#f08} GRAF1a constructs in which the PH, GAP or SH3 domains were deleted still caused LD clumping after the washout of fatty acids ([supplementary material Fig. S4H](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). However, LDs could spread when a GRAF1a-ΔBAR construct was overexpressed ([Fig. 8A,D,E](#f08){ref-type="fig"}). The capacity of GRAF1a to prevent the redistribution of LDs during the washout of fatty acids from overloaded cells is thus specific, and relies on the presence of the BAR domain. During LD growth and during lipolysis, GRAF1a therefore causes the association of LDs and restricts their mobility. This accelerates individual LD growth during fatty acid uptake, and retards lipolysis upon washout. DISCUSSION {#s3} ========== We have identified GRAF1a as the predominant GRAF1 isoform expressed in neonates, whereas GRAF1b and GRAF1c are enriched in adults. In contrast to GRAF1b and GRAF1c, GRAF1a is seldom found on dynamic tubules, but mostly associates with LDs. This is true for tagged and untagged GRAF1a in a variety of cell lines and primary cells. Endogenous GRAF1a expression is, however, highly restricted, because apart from primary glial cells, none of the cell lines we examined (3T3-L1, HeLa, NIH 3T3, U-87 MG, SH-SY5Y, T84, Co115, BSC1 and RPE cells) expressed GRAF1a (our unpublished data). We were, however, able to show enrichment of endogenous GRAF1a in LD fractions isolated from oleic-acid-fed primary glial cells. Our results show that it is the GRAF1a-specific hydrophobic linker that mediates its targeting to LDs. Similarly, sequences of 30--40 residues containing hydrophobic amino acids interspersed with helix-breaking residues (serine, proline, cysteine and glycine) were found in other LD-associated proteins. In recent years, their number has increased to include DGAT2 ([@b46]; [@b59]), NSDHL ([@b20]), lipin-1γ ([@b92]), 17βHSD11 ([@b37]), GPAT4 ([@b95]), LPCAT1 and LPCAT2 ([@b61]), caveolins ([@b30]; [@b69]; [@b72]), Rdh10 and LRAT ([@b42]; [@b43]), AUP1 ([@b44]; [@b77]), UBXD8 ([@b99]), ALDI ([@b89]), AAM-B and CYB5R3 ([@b98]). For all of these proteins, this segment has been shown to be required for LD localization, and when it was examined, it has been shown to insert as a hairpin in the hydrophobic lipid layer ([@b20]; [@b24]; [@b61]; [@b77]; [@b78]; [@b98]). It is, however, generally not sufficient, and adjacent residues also contribute to the targeting ([@b38]; [@b42]; [@b43]; [@b49]; [@b92]; [@b98]). In addition, for most of these monotopic LD proteins, an important fraction of the protein resides on alternative compartments, and efficient recruitment to LDs requires lipid feeding of the cells. This is reminiscent of GRAF1a mutants in which the BAR and/or PH domains have been deleted. However, even under resting conditions, full-length GRAF1a shows minimal binding to intracellular membranes other than LDs, suggesting that the BAR and PH domains raise the affinity of GRAF1a for the LD surface. Although protein-based interactions cannot be ruled out ([@b50]), these domains are mostly known for their capacity to interact with membranes, sometimes showing specificity in terms of lipid composition and/or membrane curvature. In earlier studies, the BAR-PH domain of GRAF1 has been shown to preferentially bind small liposomes (50-nm diameter) with a preference for phosphatidylinositol-4,5- bisphosphate over other phosphoinositides ([@b54]). It has been proposed that phospholipase D1, a phosphatidylinositol-4,5-bisphosphate-binding protein, is important for LD formation in insulin-stimulated NIH 3T3 cells ([@b2]). However, LDs are not known to contain any phosphoinositides ([@b71]), and given that deletion of the PH domain of GRAF1a only marginally perturbs LD localization, it is unlikely to wholly account for the high-affinity binding to LDs. Although the lipid composition of the outer monolayer of LDs is reminiscent of the ER in terms of relative ratio of phospholipid species, detailed analysis shows that it is different from any other intracellular membrane ([@b8]; [@b71]). It is enriched in free cholesterol and in phospholipids containing a high proportion of saturated fatty acyl chains ([@b12]; [@b60]; [@b85]). This might be linked to the metabolic status of the LD ([@b4]; [@b79]), and would allow specific recruitment of different pools of proteins ([@b26]). In order to understand what determines the high-affinity binding of GRAF1a to LDs, it will be essential to further characterize the lipid composition of LDs in non-adipocyte cells and to examine the cholesterol and fatty acyl chain preferences of GRAF1a. We have shown that GRAF1a is enriched at junctions between LDs and that its overexpression induces LD aggregation and inhibits lipolysis upon washout of fatty-acid-loaded cells. This results in an increase in basal LD levels and is associated with a decreased LD mobility. These phenotypes are specific to GRAF1a because they are not induced by overexpression of GRAF proteins that do not target LDs. In addition, they have not been reported for the above-mentioned monotopic LD-associated proteins, showing they are not merely the result of the insertion of a hydrophobic hairpin in the outer LD layer. Although the extent varies with the cell type, clustering of LDs is a common phenotype when cells are incubated with fatty acids ([@b15]; [@b32]; [@b46]; [@b62]; [@b77]; [@b89]; [@b91]). In adipocytes, it is regulated by phosphorylation of PLIN1 ([@b55]; [@b68]). However, most non-adipocyte cells do not express PLIN1. In these, the mechanism leading to LD clumping is not known, although it is prevented if the microtubule network, on which LDs travel, is disrupted ([@b15]; [@b68]). GRAF1a overexpression reduces LD mobility but also potentiates LD clustering. Indeed, in that case, LDs can also contact each other because of cytoplasmic rearrangements, and, in particular, in the presence of GRAF1a, remain tethered, ultimately leading to extensive LD aggregation. The decreased LD mobility observed upon GRAF1a overexpression could also result from the GRAF1a-induced LD clustering because molecular motors associated with each independent LD within a cluster might counteract each other. However, overexpression of Cav^DGV^, a caveolin 3 mutant that constitutively associates with LDs, inhibits LD mobility and lipolysis but does not induce clustering ([@b73]). In addition, small mobile LD clusters are observed in PLIN1-transfected HEK293 cells during lipolysis ([@b68]), further suggesting that GRAF1a could independently inhibit LD mobility and induce clustering. LD clustering would be an efficient means to shield the LD surface from lipases, and allow cells with no role in lipid storage to rapidly retrieve lipids following the dispersion of LDs in the cytosol. In agreement with this, we have shown that overexpression of GRAF1a dramatically inhibits LD spreading when fatty acids are removed from the medium of overloaded cells, and that this goes together with an inhibition of lipolysis. LD fusion is one of the ways in which LDs could grow. Although fusion of LDs of similar sizes seems to be rare ([@b3]; [@b15]; [@b62]), absorption of small LDs into larger ones might occur more frequently. Overexpression of Fsp27 and Cidea, two adipocyte-specific proteins, leads to LD clustering and promotes the formation of enlarged LDs ([@b21]; [@b52]). Fsp27 and Cidea bridge LDs and allow the transfer of triglycerides from the smaller to the larger LD ([@b34]). Interestingly, their LD fusogenic activities reside in domains dispensable for LD clustering, suggesting that the two are distinct ([@b21]; [@b39]). We have shown that GRAF1a is enriched at inter-LD contacts and that GRAF1a-induced LD clustering depends on its BAR domain. This is not surprising given that GRAF1a will dimerize through its membrane interaction BAR domain while still presenting hydrophobic insertion sequences on each molecule. However, the BAR-PH domain of GRAF1 can also alter local membrane curvature because it induces liposome tubulation *in vitro* ([@b54]). The presence of a curvature effector on LDs would be consistent with an effect on droplet fusion, as this might well prime an area of the surface, making it fusogenic, as is observed with synaptotagmin during vesicle fusion ([@b56]). This would be consistent with our observation that GRAF1a also labels intracellular tubules, and that these sometimes extend from the LD surface. Use of lipid monolayers with a composition reflecting the outer LD surface will be needed to examine whether the BAR domain of GRAF1a can alter the biophysical properties of LDs and whether this might contribute to inducing LD clustering/fusion. In addition to studying GRAF1a overexpression in HeLa cells, we have shown that knocking out GRAF1 decreases the capacity of primary glial cells to accumulate LDs upon incubation with oleic acid. Unlike HeLa cells, LDs of primary glial cells are relatively immobile, and LD clustering does not seem to contribute to the first stages of LD growth. GRAF1a could, however, promote LD accumulation by enhancing fusion between a large LD and smaller ones, which we might not resolve in our microscopy analysis. Alternatively, given that knocking out GRAF1a increases LD mobility, GRAF1a might promote the anchoring of LDs to intracellular structures, which might enhance direct lipid transfer ([@b28]; [@b35]; [@b83]). GRAF1a is enriched in the brains of P7 newborn mice. The LD content of various exocrine glands of rats increases shortly after birth ([@b63]; [@b64]). We were, however, unable to correlate GRAF1a expression with an increase in LD levels in the brain. LD accumulation might be very restricted in time and/or space, and future work will hopefully allow understanding of the physiological relevance of GRAF1a expression at this developmental stage. Furthermore, because LDs accumulate in brain tumours and neurodegenerative disorders, it will be interesting to employ mouse models to explore any involvement of GRAF1a under such pathological conditions. GRAF1 is associated with a variety of diseases. Our results now raise the possibility that it might be linked to the involvement of GRAF1a in the regulation of LD homeostasis. The restricted tissue distribution of GRAF1a probably explains the fact that it has not been identified in previous proteomic screens of LDs. However, interestingly, the BAR-domain-containing proteins amphiphysin 2 and SNX1 have been found in LDs isolated from skeletal muscle ([@b100]), and FCHSD2 has been identified in LDs of an insulin-producing β-cell line ([@b48]). Further studies will be required to confirm their binding to LDs, examine their contributions to LD properties and understand how a membrane-sculpting and -sensing BAR domain might contribute to LD homeostasis. MATERIALS AND METHODS {#s4} ===================== Cloning {#s4a} ------- GRAF1 sequences were amplified from a human brain cDNA library (Clontech Laboratories, Mountain View, CA). PCR fragments were recombined in pDONR201 using the Gateway system (Life Technologies, Carlsbad, CA). GRAF1a linker-SH3 \[amino acids 699--814 of human GRAF1a (hGRAF1a)\], GRAF1 SH3 (amino acids 669--759 of hGRAF1b), GRAF1a ΔSH3 (amino acids 1--758 of hGRAF1a), GRAF1a ΔGAP (deletion of amino acids 383--568 of hGRAF1a), GRAF1a ΔPH (deletion of amino acids 265--369 of hGRAF1a), GRAF1a ΔBAR (amino acids 241--814 of hGRAF1a) and GRAF1 BAR-PH (amino acids 1--382 of hGRAF1a) were subcloned from pDONR01 hGRAF1a. hGRAF2 was amplified from IMAGE clone 40027832 and recombined in pDONR201. Two mutations (A390G, A393G) were introduced to abolish unwanted recombinations during the LR Gateway reaction. All constructs were verified by sequencing. For expression in mammalian cells, the entry clones were recombined in modified pCI (Promega, Madison, WI) vectors designed to express proteins with C-terminal EGFP or N-terminal Myc tags. Antibodies and reagents {#s4b} ----------------------- The following commercial antibodies were used: rabbit polyclonal anti-calnexin, anti-PLIN3/TIP47/M6PRBP1 and anti-GFP antibodies from Abcam (Cambridge, UK), and anti-calreticulin antibody from Merck Millipore (Darmstadt, Germany); mouse monoclonal anti-β-tubulin (clone 2-28-33) and anti-c-myc (clone 9E10) antibodies from Sigma (St Louis, MO), anti-Rab8 (clone 4, Rab8), anti-Rab11 (clone 47, Rab11) and anti-GM130 (clone 35, GM130) antibodies from BD Biosciences (Franklin Lakes, NJ), anti-COX IV (clone 1D6E1A8) and anti-transferrin receptor (clone H68.4) antibodies from Life Technologies, anti-β-actin antibody (clone AC-15, Abcam), and anti-PLIN2/ADRP antibody (clone AP 125, PROGEN Biotechnik, Heidelberg, Germany). The secondary reagents for western blotting were horseradish peroxidase (HRP)-conjugated goat anti-mouse-IgG and anti-rabbit-IgG antibodies (Bio-Rad, Hercules, CA) for analysis of extracts, and HRP--Protein-A (Life Technologies) for immunoprecipitations. The secondary antibodies for immunofluorescence were Alexa Fluor 488, 546 or 647 conjugates of goat anti-mouse-IgG and anti-rabbit-IgG antibodies (Life Technologies). The following reagents were used: Nile Red, Nocodazole, DAPI and Oil Red O from Sigma; and BODIPY 493/503, BODIPY 558/568 C~12~ and HCS LipidTOX Red from Life Technologies. Unless otherwise stated, cell culture products were from Life Technologies. Generating anti-GRAF1 and anti-GRAF2 antibodies {#s4c} ----------------------------------------------- GRAF1b GAP-SH3 (amino acids 370--759 of hGRAF1b) and GRAF2 SH3 (amino acids 718--786 of hGRAF2) were cloned in a modified Gateway-compatible pGex 4T2 plasmid (GE Healthcare Life Sciences, Piscataway, NJ). Proteins were expressed in BL21(DE3) pLysS *E. coli* cells (16 h, 18°C). Cells were lysed in 20 mM HEPES pH 7.4, 500 mM NaCl, 1 mM DTT, Complete cocktail of protease inhibitors (Roche, Penzberg, Germany) by freezing--thawing and sonication. Proteins were purified on glutathione--Sepharose beads, cleaved with thrombin, and further purified by Superdex 75 (GRAF1b GAP-SH3) or Superdex 200 (GRAF2 SH3) gel filtration (GE Healthcare Life Sciences). Proteins were injected into rabbits (Eurogentec, Seraing, Belgium). Final bleed sera were purified by affinity chromatography using HiTrap NHS- activated HP columns (GE Healthcare Life Sciences). Specificity was checked using purified proteins and GRAF1^+/+^ and GRAF1^−/−^ brain extracts. Cell lines and transfection {#s4d} --------------------------- HeLa (ECACC 93021013) and BSC1 (ECACC 85011422) cells were grown in DMEM-GlutaMAX supplemented with 10% foetal bovine serum (FBS, HyClone, Thermo Scientific, Waltham, MA). U-87 MG cells (ECACC 89081402) were grown in a 1∶1 mixture of DMEM and Ham\'s F-12 with GlutaMAX, 15 mM HEPES and 10% FBS. NIH 3T3 cells (ECACC 93061524) were grown in DMEM-GlutaMAX with 10% newborn calf serum; PC12 cells in RPMI 1640 with 7.5% FBS and 7.5% heat inactivated horse serum; and SH-SY5Y cells in a 1∶1 mixture of MEM and Ham\'s F-12 with GlutaMAX, 1% non-essential amino acids and 15% FBS. HeLa, U-87 MG and SH-SY5Y cells were transfected using GeneJuice (Merck Millipore) and BSC1 cells using Lipofectamine 2000 (Life Technologies). NIH 3T3 cells and PC12 cells were electroporated using the Neon Transfection system (Life Technologies). Unless otherwise stated, cells were processed 24 h after transfection. Primary cultures of glial cells {#s4e} ------------------------------- Cortices of mouse E18 embryos or P0 neonates were dissected in ice-cold EBSS buffer containing 10 mM HEPES and penicillin-streptomycin. Cortices were minced, incubated at 37°C with papain (2 U/ml, Worthington, Lakewood, NJ) and mechanically homogenized. Cells were plated onto poly-D-lysine (50 µg/ml) coated culture dishes or coverslips. Co-cultures of glial cells and neurones were grown in Phenol-Red-free MEM with 20 mM glucose, 25 mM HEPES, 1 mM sodium pyruvate and 10% FBS; glial cells were grown in DMEM-GlutaMAX with 10% heat- inactivated horse serum. Cells were grown for 7 days before fatty acids were added or cells were transfected using Lipofectamine 2000. The transfection was repeated on two consecutive days. Incubation of cells with fatty acids {#s4f} ------------------------------------ Complexes of oleic acid and BSA were prepared as described previously ([@b57]) and diluted in pre-warmed medium for HeLa cells (1.75 mM), primary glial cells (1.25 mM) or U-87 MG cells (1.25 mM). When required for imaging purposes, BODIPY C~12~ (0.2 µg/ml) or BODIPY 493/503 (0.5 µg/ml) was included in the medium. For feeding of transfected cells, oleic acid was initially added 16 h post-transfection. For chase experiments, the oleic-acid-enriched medium was removed after 24 h, cells were thoroughly washed and then incubated in normal medium for another 24 h. When primary glia were incubated with oleic acid for 48 h, the medium was replaced once at 24 h. Western blot and immunoprecipitation {#s4g} ------------------------------------ Cells and tissues were lysed in 20 mM HEPES pH 7.4, 150 mM NaCl, 2 mM EDTA, 1% IGEPAL CA-630 (Sigma), supplemented with Complete protease inhibitor cocktail. The protein concentration of the cleared supernatants was determined with a Bradford assay (Bio-Rad). Equal protein amounts were either mixed with sample buffer, heated and loaded on 4--12% NuPAGE Novex Bis-Tris gels (Life Technologies) for western blot analysis, or used for immunoprecipitation. For this, cell (100 µg) or tissue (500 µg) extracts were diluted to 100 µl in lysis buffer, and incubated with 1 µl of purified anti-GRAF1 antibody or of pre-immune serum (3 h, 4°C). Protein G Sepharose 4 Fast Flow beads (GE Healthcare Life Sciences) were added for another 30 min. Beads were washed in lysis buffer before adding sample buffer, heating and western blot analysis of a volume equivalent to 100 µg of starting material. Cell fractionation by sucrose density centrifugation {#s4h} ---------------------------------------------------- U-87 MG (five 15-cm dishes per condition, two of which had been transfected 24 h before) or primary glial cells (20 10-cm dishes) were incubated with oleic acid for 24 h. Cells were then scraped, washed in PBS and lysed in hypotonic buffer (20 mM Tris-HCl pH 7.4, 1 mM EDTA, with Complete protease inhibitor cocktail) by gentle pipetting using a wide-mouthed tip. After a 10-min incubation on ice, lysates were carefully homogenized using a loose Dounce homogenizer and centrifuged (1000 ***g***, 10 min). The supernatants were collected (input) and diluted to 3 ml and a final concentration of 10% sucrose in hypotonic buffer. These were then overlaid with 10 ml hypotonic buffer and centrifuged in a swing-out rotor (280,000 ***g***, 4 h, no brakes). 1 ml fractions were collected from the top of the gradient. For U-87 MG cells, fractions were directly diluted in sample buffer, heated at 50°C and examined by western blotting. For primary glial cells, 400 µl of each fraction and a corresponding volume of input were precipitated with CHCl~3~-MeOH as follows: 1 vol. of MeOH was added, samples were vortexed, then 1 vol. of CHCl~3~ was added, and samples were again vortexed and incubated at room temperature for 20 min. Samples were then centrifuged (10,000 ***g***, 5 min) and the top aqueous MeOH layer was removed, leaving the interface. 4 vol. of MeOH were then added, and samples were vortexed and centrifuged (13,000 ***g***, 15 min). The supernatants were discarded, and after air-drying, the pellets were resuspended in sample buffer and heated at 50°C before analysis by western blotting. Sodium carbonate treatment of membranes {#s4i} --------------------------------------- Cells were scraped in medium, spun down and washed in PBS. They were then homogenized in hypotonic buffer (20 mM Tris-HCl pH 7.4, 1 mM EDTA, with Complete protease inhibitor cocktail) using a Dounce homogenizer and incubated on ice for 30 min. Membranes were spun down (100,000 ***g***, 30 min) and resuspended in Na~2~CO~3~ 0.1 M pH 11.5. After a 30 min incubation on ice, samples were centrifuged once again (100,000 g, 30 min), the supernatants were collected, and the membranes were washed once in Na~2~CO~3~. Equivalent volumes of pellets and supernatants were analysed by western blotting. GRAF1^−/−^ mouse generation {#s4j} --------------------------- The *Graf1*/*Arhgap26* gene was disrupted by gene targeting ([supplementary material Fig. S3A](http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1)). A replacement construct was generated using a 2.9-kb BamH1 5′ arm of homology (produced using primers 5′-CTCTGAAGACCTACCTGAGGTAGGGATTTC-3′ and 5′- GACACTCGAGTCTCTTGGTTTTCCAGC-3′) and a 2.7-kb Xba1 3′ arm of homology (produced using primers 5′-TCTAGAGCTGCTCATGAACCACCTGGC-3′ and 5′-TCTAGACACAGGTAGAAAACATC-3′). The targeting vector was linearised and transfected into E14.1 (129Ola) embryonic stem cells by electroporation. Genomic DNA from resultant clones was digested with BamH1 and screened by Southern blotting and hybridisation with a probe generated by PCR using primers 5′-GTGTTCCGGACTCTTGATGGTC-3′ and 5′-CCTACCTCAGGTAGGTCTTCAGAG-3′. Correctly targeted clones were injected into C57BL/6 blastocysts to generate chimaeras, which were then crossed with C57BL/6 mice. Subsequent heterozygotes were interbred to generate GRAF1^−/−^, GRAF1^+/−^ and GRAF1^+/+^ mice. All animal work was conducted under Animals (Scientific Procedures) Act 1986, UK, and subject to local ethical approval by MRC Ethical Review Committee. Histology analysis {#s4k} ------------------ Tissues dissected from 15- to 19-week-old mice were fixed in 10% neutral buffered formalin (Sigma) and paraffin-embedded. Sections (4 µm) were stained with Haematoxylin and Eosin using a Leica Microsystems (Wetzlar, Germany) autostainer ST5020. Pictures were captured with a Nikon (Tokyo, Japan) camera fitted onto a light microscope and using a 5× objective. For Oil Red O staining, P7 newborn mouse brains were fixed in 10% neutral buffered formalin (72 h, 4°C), incubated in 20% sucrose in PBS (24 h, 4°C) and mounted in OCT (VWR, Radnor, PA). Cryosections (16 µm) were stained with Oil Red O (0.3%, 10 min), rinsed in water (5 min), counterstained with Gill\'s Haematoxylin, rinsed in water (5 min) and mounted in glycerol in PVA. Sections were visualized under an Olympus (Shinjuku, Japan) BX41 microscope with a Nikon DS-2MV camera and a 50× objective. Lipid analysis by TLC {#s4l} --------------------- Cells and tissues were homogenized in 20 mM HEPES pH 7.4, 150 mM NaCl, 2 mM EDTA. Lipids were extracted as described previously ([@b53]). Dried lipid films were dissolved in chloroform and volumes equivalent to 100 µg proteins were spotted on TLC Silica Gel 60 plates (Merck Millipore). For major phospholipids, plates were first run in chloroform:ethanol:water:triethylamine (35:40:9:35) until reaching two-thirds of the distance, and were then air-dried; the separation was completed in isohexane:ethylacetate (5:1) ([@b45]). For cholesterol, plates were run in heptane:ether:acetic acid (18:6:2). Plates were air-dried, bathed in a 3% copper (II) acetate solution in 8% H~3~PO~4~ and heated for 5--15 min at 120°C. Pictures were captured and spot intensities were quantified using ImageJ. Individual spots were identified by comparison with commercial standards. Immunofluorescence {#s4m} ------------------ Cells grown on coverslips were fixed in 3.2% paraformaldehyde diluted in culture medium (20 min, 37°C). Immunofluorescence was performed using 0.1% saponin for permeabilisation and 5% goat serum for blocking following standard procedures. When needed, Nile Red (2 µg/ml), BODIPY 493/503 (1 µg/ml) or DAPI (1 µg/ml) were included in the secondary antibody dilution. LipidTOX Red staining was performed after the final washes (1∶250, 30 min). Coverslips were mounted in a buffered PVA Glycerol mountant containing 2.5% DABCO (Sigma). Super-resolution SIM {#s4n} -------------------- SIM was performed on a Zeiss (Goettingen, Germany) ELYRA S.1 using a 63×/1.4 NA oil-immersion objective. The setup consisted of a light-proof incubator, backport ELYRA module, laser rack and PCO (Kelheim, Germany) Edge 4.2 sCMOS camera. Four laser lines were available: 405, 488, 561, and 640 nm. In SIM, a known pattern of low spatial frequency is projected onto the image plane, which, during the recording, is phase-shifted and rotated. The pattern in case of the ELYRA is a grating with constant spatial frequency that varies between the different excitation wavelengths and objectives. We recorded five individual phases and five individual rotations per single plane. Recorded images were then processed with a high-end algorithm to obtain super-resolution SIM images. As indicated in the figure legends, images are either of single focal planes or are presented as maximum intensity projections. For line-profile analysis of GRAF1a and PLIN3 distribution on the surface of LD clusters, pairs of LDs of similar sizes were selected and visualized on a single focal plane going through the middle of the LD pair. A line dissecting each pair was then drawn and the fluorescence profiles in the red, green and far-red channels were quantified using ImageJ. The distance was normalized to the total width of the pair, and the fluorescence intensities were expressed as a ratio to the maximal value measured. Line profile intensities were averaged for ten pairs of LDs selected from two cells. Confocal microscopy {#s4o} ------------------- For live imaging, cells were grown on glass-bottomed culture dishes (MatTek, Ashland, MA). Prior to imaging, medium was replaced with Phenol-Red-free DMEM containing 5% FBS and cells were placed in a temperature controlled chamber on the microscope stage with 5% CO~2~ and 100% humidity. All imaging data were acquired using a fully motorized inverted microscope (Eclipse TE-2000, Nikon) equipped with a CSU-X1 spinning disk confocal head (UltraVIEW VoX, PerkinElmer, Waltham, MA) using a 60× lens (Plan Apochromat V, 1.4 NA, Nikon) under control of Volocity 6.0 (PerkinElmer). 14-bit digital images were acquired with a cooled EMCCD camera (9100-02, Hamamatsu Photonics, Hamamatsu City, Japan). Four 50 mW solid-state lasers (405, 488, 561 and 647 nm, CVI Melles Griot, Rochester, NY) coupled to individual acoustic-optical tunable filters were used as light source. Rapid one- or two-colour time-lapse images were acquired. As indicated in the figure legends, cells are either shown on a single focal plane or as 2D projections of confocal stacks acquired at 0.25-µm intervals. Representative images of live cells and fixed cells collected in independent experiments are shown. For each cell, quantitative analysis was performed using the Volocity software, identifying LDs by staining with BODIPY C~12~ or BODIPY 493/503, or as PLIN2- or PLIN3-positive objects. For each independent experiment, the mean LD volume, total LD volume and mean LD number in transfected and untransfected cells were quantified. These were then averaged for all independent experiments. Details on the numbers of cells analysed are mentioned in the figure legends, and generally correspond to quantifying the properties of a few hundred to a few thousand LDs. To quantify the proportion of GRAF1a-positive LDs also positive for PLIN2 or for PLIN3, the percentage of GRAF1a--GFP objects with a PLIN2 or PLIN3 signal above background was quantified for each transfected cell. This was then averaged within an experiment and repeated independently for statistical analysis. Only cells co-stained with anti-PLIN2 and anti-PLIN3 antibodies were used. For quantification of LD velocity, cells were imaged on a single focal plane over a period of 15 min and images were captured at 5-s intervals. Individual BODIPY-C~12~-positive LDs were tracked in transfected and untransfected cells. Relatively confluent cultures were used to minimize contributions of cell movements to LD mobility. Statistical data analysis {#s4p} ------------------------- For all quantifications provided, the mean±s.e.m. is shown. Details on cell numbers and numbers of independent experiments are provided in the figure legends. Unless otherwise stated statistical significance was assessed by performing two-tailed paired Student\'s *t*-tests (\**P*\<0.05; \*\**P*\<0.01; \*\*\**P*\<0.001). Supplementary Material ====================== ###### Supplementary Material We wish to thank members of the LMB Biomed facility, in particular Annie Mead, Michael Brown and Karen Robinson, and members of the genotyping LMB facility and of the ARES facility, for help with various aspects of the handling and housekeeping of the mice. We would like to thank Mathias Pasche for expert assistance with the SIM. We wish to acknowledge the Human Research Tissue Bank, Cambridge University Hospitals NHS Foundation Trust, and in particular Wanfeng Zhao and Susan Howard, for the H&E histology analysis of tissues. Finally, we wish to thank Emma Evergren and Rohit Mittal and other members of the McMahon laboratory for support and comments throughout this study. **Competing interests** The authors declare no competing interests. **Author contributions** H.J. and A.M. generated the knockout mice. Y.V. prepared the primary cultures of glial cells, performed the Oil Red O staining of mouse brains and imaged cells with the SIM. S.L.-A.H. performed all the other experiments and analysed the data. S.L.-A.H. and H.M.M. designed the experiments and wrote the manuscript. **Funding** This work was funded by the Medical Research Council UK \[grant number U105178805\]. S.L.-A.H. was funded by Swiss National Science Foundation fellowships \[grant numbers PBGE1-121206, PA00P3-124164\]. Deposited in PMC for immediate release. Supplementary material available online at <http://jcs.biologists.org/lookup/suppl/doi:10.1242/jcs.147694/-/DC1>

Dear Editor, Celiac disease (CD) is a chronic inflammatory disease of the gut occurring in genetically susceptible individuals after the ingestion of gluten. It is characterized by a flattened mucosa, villous atrophy, and crypt hyperplasia in the small intestine, by the classic malabsorption syndrome (diarrhea, steatorrhea, weight loss), or by seemingly less severe symptoms such as iron deficiency anemia, osteopenic bone disease, amenorrhea, and infertility \[[@R1]\]. The elimination of gluten from the diet generally leads to a return to normality of the morphological changes \[[@R2]\]. Intestinal damage is caused by an interaction between the deamidated glutamine residues of gliadin and HLA-DQ2 (DQA1\*05/DQB1\*2) or DQ8 (DQA1\*03/DQB1\*0302) molecules \[[@R3]\], with consequent T-cell response and production of autoantibodies against type 2 transglutaminase (anti-tTG2-Ab). HLA phenotype is also considered the most important genetic marker of nonresponders to the hepatitis B (HBV) vaccination. In particular, the immune response to the HBV vaccine is largely determined by the presence of the immunogenetic peptides via the HLA-DR and DQ molecules \[[@R4]\]\[[@R5]\], with the DR3-DQ2 and DR7-DQ2 haplotypes generally having a lower response rate \[[@R6]\]\[[@R7]\]\[[@R8]\]\[[@R9]\]. In 2000, the World Health Organization estimated that 2 billion people worldwide had serological evidence of past or present infection with HBV and that 350 million of these people were chronically infected and at risk for HBV-related liver disease \[[@R10]\]. HCV infection is endemic to most parts of the world, although there is considerable geographic variation \[[@R11]\]. Estimates indicate that 2.2% of the global population is infected with HCV \[[@R12]\]. HCV is the most common chronic blood-borne infection in the U.S. and is a major cause of cirrhosis and hepatocellular carcinoma \[[@R13]\]. An interesting chapter in clinical practice is the possible association between CD and viral hepatitis. Recently, researchers have hypothesized that nonintestinal inflammatory chronic diseases, such as HBV and HCV, may be the immunologic trigger for the development of CD \[[@R14]\]\[[@R15]\]. However, the association between chronic viral hepatitis and the development of CD is still a matter of debate. In a recent study, Leonardi and La Rosa found no cases of CD in a retrospective cohort of patients carrying HBV, and no CD cases appeared during treatment with interferon \[[@R16]\]. This study, although limited by the small size of patients studied, is interesting because it may be representative of what has been observed in Italy. HBV prevalence in Italy is higher than in the rest of Europe \[[@R17]\], and a high prevalence of CD is estimated as well \[[@R18]\]. A more interesting question is: if it does not seem to be an association between CD and viral hepatitis in Italy, what is happening in other regions of the world that have a high incidence of viral hepatitis \[[@R19]\]\[[@R20]\]\[[@R21]\] and an increased incidence of CD \[[@R22]\]\[[@R23]\]? Until further, large, epidemiological studies investigate this question, the answer is still open.We have more answers about what happens in coeliacs after vaccination. Mass immunization of the population has been recommended by the World Health organization since 1991 \[[@R24]\], and it is generally performed in 12-year-old schoolchildren \[[@R25]\]. We know that CD patients have a lower rate of immunization after HBV vaccination \[[@R26]\]\[[@R27]\]. We do not know if CD will increase with the actual rate of vaccination. This may be a problem for the public health system because a great deal of young people may be at high risk of contracting HBV due to a lack of immunization. In a fine paper published in 2008, Nemes et al. demonstrated that the response to HBV vaccination in CD patients is related to the response to a gluten-free diet \[[@R28]\]. In fact, an adequate vaccine response to HBV was found in coeliacs compliant with (Gluten-Free Diets) GFD, whereas nonresponse was a sign of undiagnosed CD or a lack compliance to GFD \[[@R28]\]. Surprisingly, Nemes et al. did not find any association between nonresponse and HLA-DQ2 or DQ8 status \[[@R28]\]. These results have been recently confirmed by Ertem et al., who found that response to HBV vaccine in children with CD who are compliant with GFD did not differ from the response in a healthy population \[[@R29]\]. Therefore, until new epidemiological data shed light on the possible association between CD and HBV infection, good advices seems to be 1\. to screen for CD in schoolchildren before HBV vaccination; 2\. to obtain optimal compliance to GFD in CD patients before HBV vaccination to reduce the risk of unresponsiveness; 3\. to revaccinate during a well-controlled GFD in order to maintain a high level of immunization.

For decades, optical nano-patterning has been the key technique to drive the discoveries in many nanotechnology areas. The long-standing target in the optical nano-patterning development is to achieve both high resolution and high throughput at low costs[@b1]. Particularly, for economically creating periodic structures over a larger area, many effective lithographic techniques have been developed, such as block-copolymer self-assembly lithography, self-assembled nano-particles, and nano-sphere lithography[@b2][@b3][@b4][@b5][@b6]. Among various types of nano-patterning techniques, near-field scanning optical lithography (NSOL) has attracted much attention[@b7][@b8][@b9][@b10][@b11][@b12][@b13][@b14][@b15][@b16][@b17][@b18][@b19][@b20][@b21]. NSOL typically scans a probe with nano-apertures as diffraction-unlimited light sources to direct write surface features in the optical near-field. Without the need of pre-patterned templates or photomasks, an arbitrary pattern can be created with NSOL by raster- or vector-scanning the probe over the photoresist-coated substrate. Sub diffraction-limit feature sizes can be obtained using NSOL technique and feature sizes comparable to electron beam lithography on self-assembled monolayers have been reported[@b14][@b17]. Remarkably high scanning speeds (cm/s) in NSOL have also been demonstrated, albeit with less resolution of around 190 nm[@b16] using a solid immersion lens. A major technical challenge in NSOL has been the precise control of the separation or gap distance between the photoresist and the nano-apertures during the lithography process[@b8][@b11][@b22]. The light emerging from the exit-plane of the nano-aperture is subjected to strong decay and divergence over a distance of tens of nanometers[@b23][@b24]. Precise control of the gap distance within tens of nanometers is crucial for maintaining both high light intensity and a tight light spot during the lithography process. On the other hand, NSOL does not need to prohibitively scale down the working wavelength to improve process resolution, which makes it inexpensive compared to other state-of-the-art lithography methods. Traditional NSOL uses a single aperture/antenna which poses a limit on the throughput. One way to achieve high patterning throughput is to use a large number of nano-apertures in parallel. Massively-parallel scanning lithography has been demonstrated in scanning probe lithography where large arrays of cantilever probes were used to generate lithographic patterns in the contact mode[@b15][@b25][@b26]. Dip-pen lithography, a versatile scanning probe lithography method, is capable of producing massively parallel patterns at the 100 nm scale[@b27][@b28][@b29]. "Snomipede" is a parallel scanning near field lithography method wherein an area as large as 1 mm in size was addressed simultaneously by sixteen probes yielding a minimum feature size of 70 nm[@b15]. On the other hand, there has been no experimental demonstration of massively parallel near-field mask- less optical lithography of a large number of patterns at diffraction-unlimited resolutions, because of the difficulty in precisely controlling the gap distance between a nano-aperture array and the photoresist surface over a large area. In this paper, we report our recent experimental advances in achieving massively parallel near-field maskless optical lithography with diffraction-unlimited resolutions. Our approach was accomplished by integrating a bowtie-aperture-based near-field parallel scanning lithography system with an optical metrology system, named interferometric-spatial-phase-imaging (ISPI), to detect and actively control the near-field gap distance. The ISPI system uses a different wavelength of laser light that is insensitive to the photoresist to avoid interfering with the lithography system. Measured pattern line-width as narrow as 19 nm in the deep sub-wavelength scale was achieved by precisely controlling the working distance of the nano-apertures. Moreover, up to 1,024 nano-patterns were created in parallel. In principal, there is no fundamental limit to further scale up the parallel patterning throughput by the use of an even larger number of nano-antennas. [Figure 1a](#f1){ref-type="fig"} shows the schematic of the parallel NSOL system, which consists of two optical systems: the lithography system and the optical gap detection and alignment system which includes an interferometric method for coarse alignment as the first step that is described elsewhere[@b10] and an interferometric-spatial-phase-imaging (ISPI) setup for subsequent fine alignment and gap control. In the lithography system, a frequency-tripled diode-pumped solid state UV laser (wavelength = 355 nm) is utilized as the light source for exposing photoresist. An optical mask carrying light focusing elements, i.e. an array of bowtie aperture antennas, is mounted on a multi-axis piezoelectric stage and leveled with respect to a photoresist-covered flat substrate mounted onto a scanning stage. [Figure 1b](#f1){ref-type="fig"} shows a schematic sketch of the mask, which is made in a piece of 0.5′′ × 0.5′′ quartz substrate. A square island of 150 μm side length is made at the center of the quartz substrate. A 70-nm thick chromium layer is coated onto top of the island, in which bowtie nano-apertures with a nominal outline dimension of 190 nm by 190 nm, designed to achieve the highest transmission[@b23][@b30], were milled using focused ion beam (FIB). The SEM images of the mask and an array of bowtie apertures are shown in [Fig. 1c,d](#f1){ref-type="fig"}. The pitch of the bowtie aperture array is chosen to be 2 μm, which is significantly larger than the diffraction limit and the surface plasmons wave decay distance. This arrangement allows us to individually address each bowtie aperture antenna using diffraction-limited optics and avoid potential interference between neighboring apertures. On both sides of the island, a series of two-dimensional chirped gratings are made for ISPI gap detection as shown in [Fig. 1b](#f1){ref-type="fig"}, which will be described next. For ISPI gap detection, a red laser beam (660-nm wavelength, which is insensitive to the photoresist) incidents at an oblique angle onto the ISPI gratings and diffracts into different angles due to the chirped period of the gratings. A portion of the diffracted beam reflects back from the substrate surface and re-diffracts on the gratings, creating a set of interference fringes which is captured by a camera. The number and positions of the fringes are sensitive to the distance between the mask and substrate, based on which the gap distance is computed. More details of the ISPI technique were given elsewhere[@b31][@b32][@b33][@b34]. Before lithography, Shipley S1805 photoresist was spun on a quartz substrate at 4500 rpm, giving a thickness of 400 nm. The mask and the resist substrate were held with 5-axis degree-of-freedom, all controlled by piezoelectric stages. The resist substrate can scan in x, y and z directions with a positioning resolution of 0.4 nm. The stage can also adjust its tip-tilt angle to obtain a uniform gap between the mask and resist substrate. At the initial stage of lithography, the working gap between the resist substrate and the mask is calibrated by first bringing the substrate into light contact with the island on the mask and then lifting the mask to a desired working distance. Since the surface roughness of the undeveloped photoresist is less than 1.5 nm, it does not have a significant effect on the ISPI readings. During lithography, the mask is illuminated by the UV exposure laser beam. By modulating the laser beam according to the motion of the resist substrate in the x-y plane vector scan, an arbitrary pattern can be produced. Results and Discussion ====================== The control of the gap between the mask and the substrate contributes greatly to the quality of the lithography result. When the working gap is larger than tens of nanometers, the transmitted optical energy will be too low to expose the photoresist at high throughput and the optical hot spot will diverge significantly, leading to a poor lithography resolution. If the aperture mask and the photoresist are too close, i.e., less than a few nm, the interfacial forces start to play a role in mechanical alignment and scanning processes. When the aperture and the photoresist are in contact, the adhesion and friction may cause unfavorable hysteresis in scanning trajectories, photoresist deformations, and even material damages. The ideal working distance is to bring the mask as close to the substrate as possible, but avoid any contact. Besides the gap distance, the scanning speed and exposure dose are also important parameters which contribute significantly to the quality of the lithographic patterns. For a particular gap distance, increase of scanning speed means less exposure time at a given location which can cause shallower or even no lithography results. Conversely, for a particular scanning speed, increase of exposure dose makes the lithographic lines thicker due to more energy being deposited. It was found experimentally that the required exposure dose increases roughly linearly with the scanning speed. Any possible imperfections of the substrate and the mask will also have some influences on the lithography quality. It was found that the roughness of the photoresist lies within 1.5 nm and the specified flatness of the quartz substrates that were used for the mask and the substrate is λ/20 (λ = 633 nm) = 37 nm over a 0.5′′ x 0.5′′ area, which translates to a flatness better than 1 nm over the 150 μm x 150 μm island where the bowtie apertures are fabricated. As shown in [Fig. 2a](#f2){ref-type="fig"}, we precisely measured and controlled the gap distance at nanoscale and perform lithography at different gap distances varying from 0 to 20 nm while keeping the exposure dose and scan speed unchanged, at 1.1 μJ/cm^2^ and 0.7 μm/s, respectively. The exposure dose is not a strong function of the gap distance within this distance range. The depth of the lines decreases with the increasing distance. All the AFM topographical images in this work have been taken in the contact mode with a Bruker ORC8-10 AFM probe with a nominal tip radius of 15 nm. At a gap distance of 10 nm, the minimum line-width of 19 nm measured at full-width-high-maximum (FWHM) was obtained by optimizing the exposure dose ([Fig. 2b](#f2){ref-type="fig"}) to be 0.67 μJ/cm^2^ and 0.5 μm/s, respectively. Since the feature size is similar to the dimension of the tip radius, the effect of convolution of the tip shape with the feature size could affect the measured FWHM value. Deconvolution of the measured line profile with the geometric model of the tip shows a FWHM value of 25 nm. We believe in this case the transmitted dose through the bowtie aperture is just above the photoresist exposure threshold. During parallel lithography operation, we consistently maintain a parallelism between the mask and substrate better than 30 μrad[@b32]. This ensures that the mask operates at an optimized working distance with a variation less than 5 nm across the entire aperture array of 150 μm by 150 μm in size. A larger variation in working distance caused by uncompensated tilt angles will induce pattern distortions and lead to poor pattern uniformity over the array due to the near-field nature of the optical field emerging from the nano-apertures. [Figure 3a](#f3){ref-type="fig"} shows the lithography results of letters "bnc" when the tilt is controlled below 30 μrad using the ISPI technique. The FWHM of the line is around 56 nm as shown in [Fig. 3b](#f3){ref-type="fig"}. [Figure 3c](#f3){ref-type="fig"} shows a comparison lithography results where the tilt angle was not adjusted using the ISPI technique. It is clearly seen that, in the absence of tilt angle control, variation in working distances caused severe hysteresis in scanning trajectories and pattern distortions. In parallel lithography experiments, we uniformly illuminated the aperture array although they can be individually addressed to generate fully arbitrary patterns by using a micro-scale optical projection system powered by spatial light modulators, such as digital micro-mirror device (DMD). [Figure 4](#f4){ref-type="fig"} shows the parallel lithography results using a 1,024-bowtie aperture array achieved at an optimized gap distance and light power level. [Figure 4a](#f4){ref-type="fig"} shows an array of "N" shape patterns exposed, consisting of two lines and dots in the middle as seen in the inset of [Fig. 4a](#f4){ref-type="fig"}. The mean linewidth of the lines of the "N" shape shown in the inset is 67 nm with a standard deviation of 8 nm. This variation in the linewidth is most likely caused by the variation in the fabrication of each of the bowtie apertures. Characterization of the dimensions of the bowtie apertures from their SEM images showed a variation of about 7 nm in the gap distance between the two arms of the bowtie apertures and the outline dimension. The result in [Fig. 4a](#f4){ref-type="fig"} was obtained using an exposure dose of 1.57 μJ/cm^2^ and a linear scanning speed of 1 μm/s. This higher laser fluence was used to ensure patterns are produced by all the antennas since there is variation between the antennas made using FIB. We also tested using higher scan speeds. With a faster scan speed, the exposure light power intensity needed to be increased accordingly. [Figure 4b](#f4){ref-type="fig"} shows the lithography result obtained at a scan speed of 10 μm/s, where the laser fluence was increased by approximate 10 times to 17 μJ/cm^2^. Laser heating might increase the temperature of the apertures, but the expansion of the metal film caused by a temperature rise of even 100 °C is estimated less than 0.1 nm, hence the performance of the individual bowtie apertures is not likely to be affected. The FWHM line-width of one of the lines is shown in [Fig. 4c](#f4){ref-type="fig"}. Detailed characterization of the linewidth by zoomed-in AFM scans over the whole range of the patterned area show that the mean linewidth is 73 nm with a standard deviation of 15 nm. In summary, we demonstrated massively parallel near-field scanning lithography utilizing a large array of bowtie antenna apertures, with a resulting linewidth of about 70 nm. The capability of massively parallel near-field nano-lithography was enabled by the precise control of the working distance of the apertures. Uniformity of the bowtie aperture fabrication can be the key for achieving uniform lithography patterns with higher resolution. This approach has the potential to renovate the single-aperture NSOL to achieve a much higher throughput. It offers a promising optical nanolithography strategy for nano manufacturing needs. Additional Information ====================== **How to cite this article**: Wen, X. *et al.* High throughput optical lithography by scanning a massive array of bowtie aperture antennas at near-field. *Sci. Rep.* **5**, 16192; doi: 10.1038/srep16192 (2015). Support to this work by the Defense Advanced Research Projects Agency (Grant No. N66001-08-1-2037) and the National Science Foundation (Grant No.CMMI-1120577 and Grant No. CMMI-1405078) are gratefully acknowledged. X.W. also acknowledges the support from the National Basic Research Program (973 Program) of China under Grant No. 2011CB301802 and 2012CB921900. A.D. thanks the support provided by Frederick N. Andrews Fellowship from Purdue University. **Author Contributions** X.X. conceived the idea and designed the experiments. X.W. and A.D. performed the majority of the experiments and data analyses. L.T. fabricated the optical mask and contributed to the experiments. E.E.M. guided the work on ISPI. X.X. and L.P. supervised the overall work. X.W., A.D., L.P. and X.X. co-wrote the manuscript. {#f1} {#f2} {#f3} {#f4} [^1]: These authors contributed equally to this work.

Introduction ============ Major depressive disorder (MDD) will affect up to 25% of the adult population at some point in their lives^[@bib1]^ and represents one of the major causes of morbidity in the Western World^[@bib2]^ with significant economic consequences.^[@bib3]^ Although environmental factors such as stress and inflammatory disease have a strong role in the development of MDD,^[@bib4],\ [@bib5]^ there is also significant evidence of a genetic contribution to susceptibility although estimates to the scale of the contribution vary.^[@bib6]^ To identify contributing genetic loci we previously undertook a genome wide association (GWA) analysis of 1346 individuals displaying recurrent depression fulfilling DSM-IV or ICD-10 criteria of at least moderate severity against 1288 unaffected individuals and identified a number of candidate loci that were strongly associated with MDD (rs9416742, *P*=1.3 × 10^−7^ and rs999845, *P*=1.2 × 10^−7^).^[@bib7]^ Neither of these single-nucleotide polymorphisms (SNPs) occurred within a gene coding region or were found to be in linkage disequilibrium (LD) with any coding SNP. However, both these loci occupied intron 3 of the Bicaudal C Homolog 1 gene (*BICC1*). *BICC1* was originally identified as mouse mutations which truncated the BICC1 protein were shown to result in polycystic kidney disease in mice.^[@bib8]^ Subsequent analyses showed that inactivation of *BICC1* resulted in Wnt^[@bib9]^ and PKA^[@bib10]^ pathway hyperactivation. Biochemical analysis demonstrated that the BICC1 protein contained three K homology domains that are involved in RNA binding and that BICC1 protein is implicated in post-transcriptional regulation of gene expression by antagonising miRNAs.^[@bib11]^ More recent studies have also implicated the involvement of BICC1 protein in modulating bone density.^[@bib12]^ Two mRNA isoforms have been reported for *BICC1*; termed the long and the short isoform. Using microarray analysis we detected the expression of the long isoform of *BICC1* in neuronal tissues but were unable to detect significant levels of expression of the short isoform.^[@bib7]^ Interestingly, the long isoform also displayed significant evidence of differential expression between different brain areas.^[@bib7]^ Furthermore, MRI brain scan studies of MDD patient\'s demonstrated that the T allele of rs999845, which was protective against MDD, was associated with greater hippocampal volumes.^[@bib13]^ Intriguingly, this effect was lost if the patient had been exposed to stressful events during childhood suggesting an epigenetic influence.^[@bib13]^ The finding of disease associated SNPs within the intron 3 of the *BICC1* gene is consistent with the vast majority of other GWA studies where 93% of disease associated SNPs fall within regions of the genome that do not encode protein.^[@bib14]^ The discovery that such a high proportion of associated SNPs occur in non-coding regions has hampered efforts to explore the functional consequences of SNPs within the non-coding genome that still remains poorly understood. Clues as to the functional consequences of many disease associated SNPs come from whole genome functional analyses as typified by the ENCODE consortium.^[@bib15]^ Thus, 73% of disease associated SNPs were detected in regions of the genome that had characteristics of cis-regulatory regions determined using techniques such as FAIRE-seq, DNAseI-seq and ChIP-seq that identify regions of the genome that are less condensed or, which bind transcription factors or specifically modified histones *in vivo*.^[@bib14]^ These cis-regulatory sequences include promoters, enhancers and silencer elements whose integrity is essential to maintaining appropriate levels of the expression of specific genes, in specific cells and in response to the correct stimuli.^[@bib16]^ Indeed, the burden of proof now suggests that polymorphic variation within these cis-regulatory elements represents the main reservoir of heritable disease including MDD.^[@bib14]^ To determine the effects of non-coding SNPs on the tissue specific activity of cis-regulatory elements we have previously used a combination of comparative genomics, reporter assay analysis in transgenic mice and magnetofected primary neurones. In this way we have shown that cis-regulatory regions with strong cell specificity can be identified by virtue of their high degrees of evolutionary conservation.^[@bib17],\ [@bib18],\ [@bib19],\ [@bib20],\ [@bib21]^ Furthermore, we have demonstrated that the activity and tissue specific characteristics of cis-regulatory elements are dependent on the specific promoter used and are best explored using endogenous promoter regions.^[@bib20],\ [@bib21]^ We have also found that the functional consequences of allelic variation only become significant following the activation of specific signalling pathways in specific cell types.^[@bib22]^ On the basis of these previous studies we explored the hypothesis that SNPs in LD with rs9416742 contributed to susceptibility to MDD by altering the activity of highly conserved cis-regulatory elements in intron 3 of the *BICC1* gene on the cell specific and inducible activity of the *BICC1* promoter. We used a combination of LD studies, comparative genomics, *in situ* hybridisation, transgenic analysis and primary cell culture analysis to demonstrate the presence of a novel polymorphic cis-regulatory element in LD with rs9416742 that demonstrated significant allele specific effects on *BICC1* promoter activity in primary amygdala neurones. The significance of these finding in health and in the complex underpinnings of MDD are discussed. Materials and methods ===================== Bioinformatic analysis ---------------------- Genomic analysis of the *BICC1* locus was performed using the UCSC genome browser (<http://genome.ucsc.edu/>). Allele frequency of SNPs was determined using the NCBI SNP database (<http://www.ncbi.nlm.nih.gov/SNP/>). Bioinformatic analysis of the effects of allelic variants on transcription factor binding sites were determined using the RegSNP website (<http://viis.abdn.ac.uk/regsnp/Home.aspx>). PCR amplification of regulatory regions, cloning- and site-directed mutagenesis ------------------------------------------------------------------------------- Because of its high GC content BICCprom was amplified using *BICC1*\_Long_Prom_F and R oligonucleotide primers ([Table 1](#tbl1){ref-type="table"}) from human placental DNA using a GC-rich PCR system (Roche, East Sussex, UK). BICC77 was amplified using ECR rs2393477_F and R oligonucleotide primers ([Table 1](#tbl1){ref-type="table"}) from human placental DNA using the Expand High Fidelity PCR system (Roche). PCR products were cloned into pGEM-T-easy (Promega, Southamton, UK) and sequenced to confirm PCR fidelity and to determine genotype. To reproduce both haplotypes of BICC77 site-directed mutagenesis of BICC77 in pGEM-T Easy was undertaken using BICC1-SDM-61 primers (forward and reverse) and BICC1-SDM-77 (forward and reverse; [Table 1](#tbl1){ref-type="table"}) oligonucleotide primers using QuikChange II XL Site-Directed Mutagenesis Kit (Agilent, Edinburgh, Scotland, UK). DNA constructs -------------- *pBICCprom-LacZ.* BICCprom was digested from pGEM-T easy using NcoI and EcoICRI and ligated into the NcoI and SmaI sites of the SB LacZ reporter plasmid (A kind gift from Sanbing Shen) to form pBICCprom-LacZ. *pBICCprom-Luc.* BICCprom was digested from pGEMT-easy using NcoI and SalI and ligated into the NcoI and XhoI sites of pGL4-23 (Promega). *pBICC77prom-Luc(CG and AA).* BICC77 (CG or AA) was digested out of pGEMT-easy using ZraI and SacI and placed in a ligation containing the pGL4.23 vector linearised with NcoI and Sac1 and BICCprom digested from pGEMT-easy using NcoI and EcoICRI. Transgenic analysis (including lacZ stain) ------------------------------------------ Following linearization and removal of the plasmid backbone, BICCprom-lacZ construct DNA was microinjected into one cell C57/BL6xCBA F1 mouse embryos at a concentration of 2--4 ng μl^−1^ as previously described.^[@bib23]^ Surviving embryos were oviduct transferred into pseudopregnant CD1 host mothers. Transgenic mice from each line were humanely sacrificed by lethal injection of euthatol according to UK Home Office guidelines. These mice were then dissected to recover the brain which underwent submersion fixation in 4% PFA (paraformaldehyde disolved in phosphate buffered saline) for 4 h. All tissues were then stained using 5-bromo-4-chloro-3-indolyl-beta-D-galactopyranoside (X-gal) solution for 4--12 h as previously described.^[@bib23]^ X-gal stained tissues were then prepared for vibratome sectioning by allowing tissue to equilibrate with increasing concentrations of sucrose (to 30%) in phosphate buffered saline at 4 °C. Tissues were equilibrated with bovine serum albumen/Gelatin solution (0.5 gelatin, 30 sucrose, 15.5% bovine serum albumen) overnight at 4 °C. Bovine serum albumen/gelatin embedded tissue was then hardened by addition of one-tenth of a volume of 25% gluteraldehyde. Fifty micron sections were then cut on a Vibratome series 1000 (Leica, Milton Keynes, UK). Sections were mounted on glass slides and photographed under light field or phase contrast illumination. *In situ* hybridisation ----------------------- Radioactive *in situ* hybridisation was carried out as previously described.^[@bib24]^ Mouse cDNA for use as a template for the production of antisense *BICC1* probe for radioactive *in situ* hybridisation was purchased from Source Bioscience (Genbank accession no. AK076154.1). Cell culture and luciferase reporter gene assays ------------------------------------------------ Primary Amygdala cultures were prepared from P1-3 rat neonates as described earlier.^[@bib22]^ Cells were transfected with plasmid DNA using Neuromag according to the manufacturer\'s instructions (OZ Bioscience, Marseilles, France). Briefly, Neuromag reagent was added to a plasmid-media preparation and incubated at room temperature for 15 min allowing 250 ng of plasmid DNA in the media per 150,000 cells to be transfected. The DNA-Media-Neuromag solution was then added to previously prepared single cell cultures in 24-well plates and incubated on a magnetic plate for 15 min in a cell culture incubator at 37 °C and 5% CO~2~. Luciferase activity of plasmid constructs were measured using dual luciferase assay kits (Promega) using lysates from cell cultures transfected with the GAL5.1 luciferase constructs. Cell cultures were lysed and the proteins stabilized using the passive lysis buffer as per the manufacturer\'s instructions (Promega). The dual-luciferase assay analysis was carried out on a Glomax 96-microplate luminometer (Promega) using 20 μl of cell lysate per well of a white 96-well plate. Data analysis ------------- Sample sizes (*n*⩾3) were based on pooled amygdala neurones recovered from at least three individual rat neonates. Statistical significance of data sets were analysed using either two way analysis of variance (ANOVA) analysis with Bonferoni *post hoc* tests or using two tailed unpaired parametric Student *t*-test as indicated using GraphPad PRISM version 5.02 (GraphPad Software, La Jolla, CA, USA). Results ======= *BICC1* displays spatially discrete mRNA expression within components of the limbic system ------------------------------------------------------------------------------------------ Our previous GWA analysis provided a compelling case for examining the involvement of allelic variations within intron 3 of the *BICC1* gene in major depressive disorder.^[@bib7]^ We also provided evidence that the long isoform of *BICC1* (isoform 1), that includes the first seven exons ([Figure 3a](#fig3){ref-type="fig"}), showed strong differential expression in specific nuclei of the human brain.^[@bib7]^ In contrast, the short isoform (isoform 2, lacks the first 7 and last 3 exons), although detectible in kidney and bone tissue, demonstrated consistently low expression in all brain tissues.^[@bib7]^ These observations argue against a role for the short isoform in brain function and, for this reason, we concentrated our analysis on the long isoform. Although our microarray analyses were highly informative they lacked the resolution to determine the specific regions of the brain that expressed *BICC1*. For this reason we used *in situ* hybridisation using probes derived from the full length *BICC1* complementary DNA that reflects the long isoform, on sections of adult male mouse brain to determine whether *BICC1* mRNA was expressed throughout the brain or displayed discrete regions of expression. These *in situ* analyses demonstrated that, although full length *BICC1* transcript expression was present in the majority of tissues at low levels, areas of higher expression were evident within discrete components of the limbic system including the hippocampus, the hypothalamus and the amygdala ([Figure 1](#fig1){ref-type="fig"}). There was also evidence of increased expression within the cingulate cortex and the thalamus ([Figure 1](#fig1){ref-type="fig"}). Because the limbic system plays a critical role in conditioned fear and mood modulation these observations support a role for changes in the expression of *BICC1* long isoform in contributing to mood disorders. Bioinformatic analysis of the *BICC1* promoter ---------------------------------------------- We have established that *BICC1* mRNA demonstrates tissue specific expression patterns within regions of the mouse limbic system. These findings are consistent with evidence of a role for SNP variants within intron 3 of the *BICC1* gene in MDD. Because these SNPs are non-coding we next explored the hypothesis that they influenced mood by altering the transcriptional regulation of *BICC1* within the limbic system. However, we have no knowledge of the systems regulating *BICC1* mRNA expression within the brain. For this reason we used information contained within the UCSC genome browser to direct our studies and provide evidence of the promoter regions supporting *BICC1* expression in the brain. UCSC describes the existence of two mRNA species transcribed and spliced from the human *BICC1* gene locus ([Supplementary Figure 1A](#sup1){ref-type="supplementary-material"} and [Figure 3a](#fig3){ref-type="fig"}). Thus, the short splice form is encoded by 10 exons distributed over 21 kb (GRCh37/hg19, chr10:60,553,182-60,574,087) and the long splice form shares the same 10 exons but has an additional 11 exons distributed over 318 kb (GRCh37/hg19, chr10:60,272,942-60,591,445) (see [Supplementary data](#sup1){ref-type="supplementary-material"} [Supplementary Figure 1a](#sup1){ref-type="supplementary-material"}.). Previous mRNA expression analysis of the short and long splice-forms of *BICC1* suggested that, although expression of both splice-forms was detected in specific human neuronal tissues and cells, the longer splice-form was expressed at considerably higher levels and varied strongly between different tissues suggesting that its expression is maintained by a stronger promoter element.^[@bib7]^ To confirm this observation we used data embedded in the ENCODE data set to explore the promoter characteristic of DNA sequence around the transcriptional start sites of the long and short splice-forms. Using accepted criteria such as histone modification patterns (H3K4me2, H3K4me3, H3K9ac and H3K27ac) as well as FAIRE and DNAse1 hypersensitivity patterns characteristic of promoter regions in a number of different human cell lines (GM2878, HeLa, HepG2 and embryonic stem cells etc.) we recovered data consistent with higher levels of promoter activity around the transcriptional start site of the long splice-form in comparison to the short isoform which demonstrated fewer active promoter characteristics in all cell lines used ([Supplementary Figure 1b--e and Figure 3a](#sup1){ref-type="supplementary-material"}) consistent with its lower activity in microarray analysis reported by Lewis *et al.*^[@bib7]^ Furthermore, the promoter of the long splice-form, was more highly conserved outside of the coding region than the promoter region of the short splice form ([Supplementary Figure 2b and e](#sup1){ref-type="supplementary-material"}). Because of this evidence we focussed our subsequent efforts on the promoter of the long splice form. Cloning and characterisation of the *BICC1* promoter ---------------------------------------------------- To isolate and characterise the long isoform promoter (henceforth referred to as BICC1prom) we designed PCR primers (see [Table 1](#tbl1){ref-type="table"}) to isolate a 691 bp region of the *BICC1* locus that spanned both the GC enriched region (75% GC content) and the region most conserved in vertebrates ([Supplementary Figure 2a](#sup1){ref-type="supplementary-material"}). This sequence also contained the first exon of the long isoform of *BICC1*. Following amplification BICC1prom was cloned into a LacZ reporter plasmid using an NcoI site that coincided with the transcriptional start site of both *BICC1* and the reporter *LacZ* gene ([Figure 2a](#fig2){ref-type="fig"}). Using pronuclear microinjection and oviduct transfer this reporter construct was used to generate a number of transgenic mouse lines (*n*=6). Brain tissues were recovered from these lines and subjected to X-Gal staining. Two of the transgenic mouse lines demonstrated robust and consistent staining for the expression of the *LacZ* gene in areas of the cingulate cortex, hippocampus, thalamus and in specific nuclei of the amygdala including the basolateral, basomedial, central and lateral amygdaloid nuclei ([Figures 2a--h](#fig2){ref-type="fig"}). Because both lines produced similar expression patterns in these areas it is highly unlikely that the patterns of expression generated represented insertional effects but were representative of BICC1prom activity. The activity of the human BICC1prom within these components of the limbic systems was entirely consistent with its role in supporting the expression of mRNA that we describe above. The *BICC1* promoter is active in primary amygdala neurones ----------------------------------------------------------- The analysis of BICC1prom in transgenic lines using a LacZ reporter construct reflected the cell specific distribution of *BICC1* mRNA in limbic regions of the brain ([Figure 2](#fig2){ref-type="fig"}). However, the analysis of LacZ constructs using transgenic mice does not represent an effective method of assaying the quantitative effects of other regulatory regions, SNPs or activation of signal transduction pathways on BICC1prom. For this reason we cloned BICC1prom into the NcoI site of a firefly luciferase plasmid (pGL4.23; Promega) replacing the generic TATA box promoter to form pBICC1prom-Luc ([Figure 3b](#fig3){ref-type="fig"}). Because of the role of the amygdala in mood and fear behaviour we magnetofected BICCprom-luc into cultured primary rat amygdala neurones together with the renilla luciferase vector pGL4.74 and in parallel with the pGL4.23 backbone plasmid. This demonstrated that the *BICC1* promoter was over 40 times more active in amygdala derived neurones that the minimal TATA box promoter of the pGL4.23 plasmid that was used as the basis of our luciferase promoter constructs ([Figure 3b](#fig3){ref-type="fig"}). These observations further support a role for the *BICC1* promoter in the amygdala. BICC1prom is moderately repressed by PKA activation but not PKC activation -------------------------------------------------------------------------- It is well established that modulation of the activity of protein kinase A (PKA) pathways in the amygdala play a critical role in the consolidation of fear memories.^[@bib25],\ [@bib26]^ There is also evidence that components of the protein kinase C (PKC) pathway are activated in the amygdala during fear conditioning^[@bib27]^ where they may also have a role in fear learning.^[@bib28],\ [@bib29]^ Due to the known involvement of PKA and PKC pathways in fear memory and learning we explored the effects of activation of these pathways on the activity of the BICC1 promoter in primary amygdala neurones. We found little evidence that the PKC pathway could influence BICC1prom activity in primary amygdala neurones ([Figure 3c](#fig3){ref-type="fig"}). We found some evidence of negative regulation of BICCprom by activation of the PKA pathway using forskolin but this was not statistically significant ([Figure 3c](#fig3){ref-type="fig"}). Linkage disequilibrium analysis of rs9416742 and rs999845 --------------------------------------------------------- We have provided evidence that BICC1prom is responsible for supporting the discrete expression of *BICC1* mRNA in the limbic system. These studies provide a useful platform for further exploring the possible effects of mood disorder associated allelic variants in *BICC1* intron 3 on the tissue specific activity of BICC1prom. Using genome wide association studies we previously demonstrated that a significant association existed between the rs9416742 and rs999845 polymorphisms (*P*=1.3 × 10^−7^ and 3.14 × 10^−7^, respectively) and susceptibility to MDD.^[@bib7]^ As previously mentioned neither of these SNPs occur within any coding regions of the *BICC1* gene nor were they in LD with coding SNPs. Because evolutionary conservation has been shown to be highly suggestive of functionality we explored the hypothesis that rs9416742 and rs999845 occurred within conserved cis-regulatory regions. Unfortunately, comparative genomics demonstrated that neither SNP occurred within a region of the *BICC1* locus which was highly conserved or displayed markers indicative of cis-regulatory regions. However, both SNPs were found to be in strong linkage disequilibrium with a SNP (rs2393477) that occurred in a 200 base pair region of human BICC1 intron 3 that had been highly conserved in mammalian species as divergent from humans as marsupials (opossum) and monotremes (platypus) ([Figure 3a](#fig3){ref-type="fig"} and [Supplementary Figure 3](#sup1){ref-type="supplementary-material"}). This high degree of deep conservation ('depth\' refers to evolutionary time as opposed to degrees of conservation by pairwise comparison) strongly suggests that this sequence has an important biological function and has served in maintaining species fitness during the 210 million years since placental mammals and monotremes diverged.^[@bib30]^ This region of homology was also 260 kb away from BICC1prom which is well within the known range of influence of many cis-regulatory elements. Other regions of *BICC1* intron 3 between this cluster of SNPs were also found to be highly conserved and also contained polymorphisms but these were not found to be in significant LD with either rs9416742 or rs999845 ([Figure 3a](#fig3){ref-type="fig"} and [Supplementary Figure 3](#sup1){ref-type="supplementary-material"}). BICC77 represses the activity of BICC1prom in amygdala neurones in an allele specific manner -------------------------------------------------------------------------------------------- To determine whether allelic variants of highly conserved regions of *BICC1* intron 3 affected the activity of BICC1prom in amygdala neurones we designed PCR primers (BICC77 for. and BICC77 rev.; [Table 1](#tbl1){ref-type="table"}) to isolate a highly conserved 200 bp region of *BICC1* intron 3 (BICC77) that contained rs2393477 and its close neighbour rs10740761, which was in high LD (*r*^2^\>0.95). BICC77 was then subjected to two rounds of site-directed mutagenesis (SDM; see [Table 1](#tbl1){ref-type="table"} for SDM primers used) to recreate allelic variants of rs2393477 and its close neighbour rs10740761 and sequenced to confirm SDM fidelity. Different BICC77 haplotypes were cloned into BICCprom-luc to form pBICC77AA-Luc and pBICC77CG-Luc ([Figure 3a](#fig3){ref-type="fig"}). Magnetofection of these constructs into primary amygdala cells demonstrated that BICC77 strongly repressed *BICC1* promoter activity in these cells. Interestingly, BICC77AA had a significantly stronger repressive effect that did BICC77GC ([Figures 3a](#fig3){ref-type="fig"} and [4a](#fig4){ref-type="fig"}). BICC177 further represses BICC1prom activity following PKA but not PKC activation. ---------------------------------------------------------------------------------- We have shown that allelic variants of the highly conserved sequence BICC77 are able to differentially down regulate the activity of BICC1prom within primary amygdala neurones. Because of the essential role played by cis-regulatory elements in relaying signal transduction information to promoter regions^[@bib22]^ we next explored the possibility that allelic variants of BICC77 differed in their ability to relay signal transduction signals to BICC1prom. We treated primary amygdala cultures transfected with allelic variants of pBICC77prom-Luc (AA and CG) with forskolin or PMA. PMA had no significant effect on the activity of either allelic variant and forskolin did not significantly alter the ability of BICC77GC to repress BICC1prom activity ([Figure 4b](#fig4){ref-type="fig"}). However, BICC77AA demonstrated significantly higher repression of BICC1prom activity following treatment with the PKA agonist forskolin ([Figures 3a](#fig3){ref-type="fig"} and [4b](#fig4){ref-type="fig"}). Bionformatic analysis of transcription factor binding at rs2393477 and rs10740761. ---------------------------------------------------------------------------------- We asked whether it was possible to use a bioinformatics approach to identify candidate transcription factors whose binding to BICC77 was allele dependent. We used the RegSNP website to determine whether the AA haplotype altered the predicted binding efficiency of known transcription factors compared with the protective GC haplotype. Examination of rs10740761 using RegSNP shows that the C allele potentially binds the transcription factor Deaf1 which is suspected in causing memory deficits and anxiety like behaviour if deleted from the mouse genome.^[@bib31]^ The C allele of rs10740761 was predicted to bind Deaf 1 with a matrix score of 0.7323 wheras the A allele only had a matrix of 0.5. The A allele of rs2393477 was predicted to bind GATA factors (1, 2, 3 and 6) with much higher affinity (matrix score for C allele=0.7 and for T allele 0.92). This is intriguing as overexpression of GATA factors has been associated with depressive behaviour^[@bib32]^ and a highly conserved C/EBP binding site exists just next to this locus. This arrangement of transcription factor binding has been associated with PKA activation on the past.^[@bib33]^ Discussion ========== To locate the presence of a human genomic locus that could contribute to susceptibility to major depressive disorder we carried out a GWA analysis of a large patient cohort and succeeded in identifying two SNPs within intron 3 of the *BICC1* gene with significant association. We observed that the T allele of the most strongly associated SNP; rs9416742, that has a frequency of 0.235 in healthy individuals, was found to have a protective affect against depression. Although neither of these SNPs were found within the coding region of *BICC1*, or within an area of intron 3 that could be identified as a cis-regulatory region, they were found to be in strong linkage disequilibrium with a SNP within a region of intron 3 that had been conserved for 210 million years thus representing the majority of the extent of mammalian evolution. We isolated this conserved region, that we called BICC77, and reproduced both of the allelic forms present in the human population (referred to as AA and GC). Because these SNPs lay within the *BICC1* gene and because subsequent analysis had demonstrated that the allelic variants of *BICC1* in question were associated with greater hippocampal volumes^[@bib13]^ we tested the hypothesis that allelic variants of BICC77 differentially regulating the transcription of the *BICC1* gene in the brain. To test this hypothesis it was necessary to first determine whether *BICC1* was expressed in any areas of the brain known to control mood. We used *in situ* hybridisation to demonstrate that *BICC1* mRNA was more strongly expressed in specific areas of the limbic system including the hippocampus, amygdala and hypothalamus. Subsequent analysis of existing genomic databases succeeded in identifying the active *BICC1* promoter by virtue of its extent of conservation, its accessibility to FAIRE and DNAseI analysis and its interaction with known promoter histone marks. Using magnetofection of primary amygdala cells with luciferase markers and dual-luciferase assays we were able to show that this promoter region was highly active in primary amygdala neurones compared with a generic TATA box promoter. Furthermore, by generating two different lines of transgenic mice with the *BICC1* promoter linked to the *LacZ* gene we also demonstrated pronounced tissue specificity in the activity of the *BICC1* promoter in hippocampus, amygdala and hypothalamus. These analyses established the expression of *BICC1* and the activity of its promoter in regions of the limbic system that have critical roles roles in mood modulation and depressive behaviour. In addition, identifying the tissue specific activity of the *BICC1* promoter provided us with a useful platform for studying the effects of allelic variants of BICC77 on its activity and hence the activity of the *BICC1* gene. We next asked what effect BICC77 would have on the activity of the *BICC1* promoter in primary amygdala neurones and we were able to demonstrate that BICC77 acted as a strong repressor of *BICC1* promoter activity within these cells. We have previously observed this model in the regulation of *BDNF* promoter 4 whereby a highly conserved region of the *BDNF* locus, containing a SNP associated with depression (BE5.2), caused significant repression of *BDNF* promoter 4 within 3 different primary cell types (hippocampus, cortex and amygdala).^[@bib22]^ These analyses of the *BDNF* locus suggested an important role for selective negative regulation in the modulation of inducible gene expression in the brain. In combination with this previous study, the current study highlights the complexities of understanding the specific mechanisms by which depression associated allelic variants may alter pathological susceptibility. Thus, by subtly changing the delicate homoeostatic balance of negative and positive regulatory mechanisms required to maintain levels of gene expression essential for mental health differences in the activity or regulatory elements over a life time, or following trauma, may be sufficient to contribute to mental illness. There is every reason to believe that identifying the effects of allelic variation on negative regulation will be critical for understanding the causes of psychiatric pathologies including depression. With this concept in mind we then asked whether allelic variants of BICC77 differed in their ability to negatively regulate the activity of the *BICC1* promoter in amygdala cells. We recreated the different genotypes of BICC77 present in the human population (AA and CG) and found that the minor haplotype of BICC77(GC), that is in LD with the protective T allele of rs9416742, was not as potent a repressor as the major allele (AA). Whilst there is currently no evidence to suggest that levels of *BICC1* expression within individuals with the CG haplotype of BICC77 are higher than those with the AA haplotype, the observation of a significant difference in the influence of these haplotypes on the activity of the *BICC1* promoter in primary amygdala neurones provides us with a possible insight into a mechanism contributing to mood disorders. These observations reflect the results of a previous study, which demonstrated that the CA haplotype of the GAL5.1 enhancer, that supports expression of the galanin neuropeptide in areas of the brain including the paraventricular nucleus and amygdala, was significantly weaker than the GG haplotype in primary hypothalamic neurones.^[@bib19]^ Subsequent analysis of these haplotypes in human cohorts suggested that the CA haplotype was protective against alcohol and cannabis abuse in certain populations^[@bib34],\ [@bib35]^ It is accepted that one of the main roles of gene regulatory regions is to interpret signal transduction information within the surrounding cell cytoplasm for the transcriptional machinery (RNA polymerase TFIID and mediator) bound at the promoter region of genes. In the specific case of the amygdala, a number of signal transduction pathways have been implicated in modulating the cellular systems that control mood.^[@bib36]^ One of the best studied is the PKA signal transduction pathway that is activated in amygdala neurones by binding of stress released epinephrine to B1 adrenergic receptors.^[@bib36]^ Because of the known contribution of chronic stress in the exacerbation of mood disorders we sought to determine the effects of agonists of signal transduction pathways such as PKA and PKC known to have a role in amygdala function. Treatment of transfected cultures to activate the PKC pathway had no effect on BICC77 and no allelic differences were observed. Activation of PKA pathways in transfected cells also had little or no effect on the ability of the BICC77GC haplotype to repress the *BICC1* promoter. Intriguingly, however, forskolin treatment resulted in further repression of the *BICC1* promoter by BICC77AA. Further bioinformatic analysis suggested that the AA haplotype had an increased affinity with a combination of transcription factors (GATA and C/EBP) that have previously been associated with PKA response.^[@bib33]^ Because the AA haplotype of BICC77 is the major haplotype within the human population and also represents the ancestral haplotype present in all other mammals, it can be hypothesised that the CG allele of BICC77, which is in LD with the protective T allele of rs9416742, is not only less repressive of *BICC1* promoter activity but has also lost its ability to respond to activation by PKA modulated signalling inputs. Indeed, it could be hypothesised that the reduction in its ability to repress the BICC1 promoter or to respond to PKA signalling may be mechanistically related to the protective effects observed for the T allele of rs9416742. Conclusion ========== Major depressive disorder has been shown to be a highly complex disease state to diagnose and this may reflect the large numbers of unreplicated genome wide association studies hits associated with this condition (\>100; HuGE genome navigator). There is also increasing evidence of the possibility of masking or exacerbation of depression phenotypes through epigenetic influences that are altered following early life traumatic events that may have a role in reducing the efficacy of genetic association analysis.^[@bib37],\ [@bib38]^ However, the overwhelming conclusion that has been drawn from the results of these studies is that the genetic and epigenetic influences in susceptibility to mood disorders are unlikely to be solely due to changes in gene coding regions but will instead affect the mechanisms that control the spatial, tissue specific and inducibility of these genes. The current manuscript describes the use of novel combinations of different techniques to determine the mechanisms through which alleles associated with MDD might contribute to susceptibility. Thus, in addition to suggesting a novel contributory mechanism towards MDD susceptibility the current manuscript describes innovative approaches to exploring the mechanisms through which non-coding disease associated SNPs may contribute to disease. This work was funded by The BBSRC (BB/D004659/1) the Wellcome Trust (080980/Z/06/Z) and the Medical Research Council (G0701003). [Supplementary Information](#sup1){ref-type="supplementary-material"} accompanies the paper on the The Pharmacogenomics Journal website (http://www.nature.com/tpj) The authors declare no conflict of interest. Supplementary Material {#sup1} ====================== ###### Click here for additional data file. ###### Click here for additional data file. ###### Click here for additional data file. ###### Click here for additional data file. {#fig1} {#fig2} ![(**a**) A diagrammatic representation (not to scale) of the *BICC1* locus demonstrating the linear relationships between the genomic locus (black line) and the short (light blue lines) and long (dark blue lines) mRNA isoforms. Exons are indicated by heavy vertical lines. Long and short isoform promoters are indicated using bent black arrows. The *BICC1* promoter and BICC77 are represented by a red filled box and a blue filled box respectively and represents a distance of 260 kb. Sites of MDD associated SNPs are represented by thin red vertical lines and conserved SNPs in strong linkage disequilibrium are indicated by thin black vertical lines. The region of the *BICC1* locus in strong LD is represented by a red triangle. The functional relationships suggested in the current manuscript are summarised using broken red lines to indicate the increased repressive effects of BICC77AA on BICC1prom compared with BICC77GC in primary amygdala neurones. (**b**) diagrammatic representation of the luciferase reporter constructs used in subsequent experiments. (**c**) A bar graph representing the results of a dual-luciferase analysis comparing the activity of the pBICCprom-Luc vector (BICCprom) to that of pGL4.23 following magnetotransfection into primary amygdala neurones. (**d**) Effects of treatment of primary amygdala neurones magnetofected with the pBICCprom-Luc plasmid and treated with 10 μ[M]{.smallcaps} forskolin or 10 μ[M]{.smallcaps} PMA for 16 h before dual-luciferase analysis. *n*=15. MDD, major depressive disorder; N.S., not significant; PMA, phorbol 12-myristate 13-acetate; SNP, single-nucleotide polymorphism. Error bars are standard error of the mean, *n*=4, analysis using (**c**) two tailed unpaired *t*-test or (**d**) ANOVA. n.s. not significant, \*\*\**P*\<0.005.](tpj201562f3){#fig3} {#fig4} ###### Names and sequences of oligonucleotides used in the construction of the plasmid reporter constructs described in the manuscript *Oligonucleotide* *Sequence (5\'-3\')* ------------------- --------------------------------------------------- BICC1_Long_Prom_F GTGGCAGTGTGTACGTGGTG BICC1_Long_Prom_R AAAAGGAGTTAGAGGCTCAGTCG ECR rs2393477_F TTCTTAGCCTTTCTGAAAATCAGTAG ECR rs2393477_R TTAACTCTGTTTTCTGCCAGAGTGAG BICC1-SDM-61 F GGAATATTATTTTTAACCTATCTAACGTATGGAGTTGACCCCCTAAAAG BICC1-SDM-61 R CTTTTAGGGGGTCAACTCCATACGTTAGATAGGTTAAAAATAATATTCC BICC1-SDM-77 F CATATTTCTCAAAGGTCCTATAAGCATTACAACACATTTATCAAAG BICC1-SDM-77 R CTTTGATAAATGTGTTGTAATGCTTATAGGACCTTTGAGAAATATG Abbreviation: BICC1, bicaudal C homolog 1.

Direct observation of intracellular processes has the potential to yield insight into fundamental biological pathways and disease mechanisms. Several techniques have been developed to enable high-resolution imaging of live cells; yet, the limited ability to trace intracellular components has hindered progress. Hence, two of the persistent challenges are probe design and cellular delivery with minimal toxicity, pivotal for advances in live-cell imaging technologies. Here we describe an efficient approach to tag and image intracellular components in live mammalian cells. Using the microfluidic cell squeezing platform to deliver small fluorescent *tris*-*N*-nitrilotriacetic acid (*tris*NTA) probes (∼1 kDa), we demonstrate highly efficient, minimally disruptive, light-triggered tracing of native proteins and the subsequent super-resolution imaging of live-cell phenomena. Live-cell microscopy contributed significant knowledge of dynamic processes such as protein trafficking and single-molecule localization-based imaging techniques visualize proteins with high-resolution information (≤50 nm)[@b1][@b2]. All fluorescent imaging techniques require protocols to introduce the label with the need to minimize its influence on the system. Fluorescent proteins, self-labelling tags[@b3][@b4][@b5][@b6] or labelling by enzymatic methods[@b7][@b8] can interfere with protein function, assembly or dynamics. Bulky fusion proteins (\>20 kDa) entail the risk of steric hindrance and functional perturbations, whereas smaller tags (for example, tetracysteine tag) deal with unspecific interactions or require additional experimental steps[@b8] and optimized flanking sequences for each protein target[@b9]. Although synthetic fluorophores have enhanced photostability, quantum yield, spectral range and localization precision, it is difficult to introduce such probes to the cytosolic environment using existing delivery technologies. On the one hand, current transduction strategies such as delivery by cell-penetrating peptides (CPPs), electroporation and so on are suboptimal, suffering from poor and endosomal uptake, rapid degradation by extracellular and endosomal proteases, low *in vivo* efficiency or elaborated chemical synthesis. On the other hand, antibody-based labelling approaches, for example, are limited to chemically arrested (fixed) cells and the availability of specific antibodies for a protein target. Owing to the described limitations of existing labelling and transduction technologies, there is a persistent demand for techniques enabling high-throughput in-cell labelling by minimal tags that are conductive to high-resolution and super-resolution microscopy. Here we demonstrate robust in-cell targeting of native proteins using a labelled multivalent chelator head *tris*NTA[@b10] and a genetically encoded oligohistidine sequence ([Fig. 1a](#f1){ref-type="fig"}). *tris*NTA site specifically recognizes His~6--10~-tagged proteins in the (sub)nanomolar range (*K*~d~ of 0.1--10 nM) even in the crowded cellular environment[@b11]. The minimal size of the tag and the molecular probe allows direct targeting with nanometre precision at subnanomolar concentrations as required for single-molecule localization-based imaging techniques[@b1][@b2][@b12][@b13] with no impact on intracellular trafficking or demand for additional cofactors affecting endogenous processes. We simplified efficient transfer of the *tris*NTA probe into living cells by cell squeezing[@b14], combining precisely controlled cytosolic delivery with high specificity and low cytotoxicity. Briefly, transient cell permeabilization is achieved by rapid viscoelastic deformation of cells as they pass through micrometre-scale constrictions. This facilitates fast uptake of probes into the cytosol before cell-intrinsic repair mechanisms kick in[@b15]. Results ======= High-affinity protein labelling at subnanomolar concentration ------------------------------------------------------------- We first investigated the specificity of the *tris*NTA/His tag targeting in chemically arrested cells. To evaluate precise localization, different proteins resident at distinct subcellular compartments were selected: (i) the transporter associated with antigen processing (TAP) in the membrane of the endoplasmic reticulum[@b16]; (ii) histone 2B (H2B) in the nucleus; and (iii) Lamin A at the nuclear envelope. All proteins of interest (POIs) were fused to a His~10~ tag and a fluorescent protein (TAP1^mVenus-His10^, H2B^mVenus-His10^ and ^His10-mEGFP^Lamin A) for specific targeting and co-localization studies, respectively. For sensitive detection, *tris*NTA was covalently coupled to different fluorescent dyes (*tris*NTA^f^, f=Alexa488, ATTO565, ATTO647N, Alexa647 and ATTO655). Mammalian cells were transiently transfected with the corresponding target genes. His-tagged proteins were specifically stained by *tris*NTA^f^ with excellent co-localization and signal-to-noise ratio (Pearson\'s coefficients between 0.90 and 0.96), using confocal laser scanning microscopy (CLSM; [Fig. 1b](#f1){ref-type="fig"} and [Supplementary Fig. 1](#S1){ref-type="supplementary-material"}). Strikingly, even at 200 pM of *tris*NTA^f^, His-tagged proteins were labelled with high specificity ([Supplementary Fig. 2](#S1){ref-type="supplementary-material"}). By analysing a variety of fluorescent dyes, we noticed that *tris*NTA^ATTO565^ targeting produced a higher background compared with *tris*NTA^ATTO647N^, *tris*NTA^Alexa647^ or *tris*NTA^ATTO655^ ([Supplementary Fig. 3](#S1){ref-type="supplementary-material"}). This was assigned to unspecific binding of the ATTO565 dye. Moreover, the superposition of both fluorescence intensity profiles reflects an excellent correlation between the POI expression level and the labelling density of *tris*NTA^Alexa647^ ([Fig. 1c,d](#f1){ref-type="fig"}, Pearson\'s coefficient *r*=0.95). Notably, using nanomolar concentrations, *tris*NTA^f^ labelling is significantly more efficient within 30 min than SNAP^f^-tag labelling ([Supplementary Fig. 4](#S1){ref-type="supplementary-material"}). In contrast, mammalian cells expressing H2B lacking a His tag showed neither *tris*NTA^f^ labelling nor unspecific staining ([Supplementary Fig. 5](#S1){ref-type="supplementary-material"}). In conclusion, *tris*NTA^f^ targeting at subnanomolar concentrations is highly specific to trace His-tagged proteins. To exploit these benefits further, we combined *tris*NTA^f^ with well-established labelling methods for multiplexed protein modification. Specific *tris*NTA^Alexa488^ labelling of ^His10^LaminA ([Fig. 1e](#f1){ref-type="fig"}, green) was successfully achieved in combination with SNAP^f^-tag labelling of H2B (magenta) and antibody labelling of tubulin (red), as well as the lysosomal-associated membrane protein 1 (blue). Thus, the ultra-small interaction pair complements the toolbox of well-established labelling techniques and the nanomolar concentrations perform various avenues in multiplexed labelling. High-throughput live-cell labelling within mammalian cells ---------------------------------------------------------- Encouraged by these observations, we aimed at protein labelling in living cells. To transfer *tris*NTA^f^ into cells, we applied microfluidic cell squeezing ([Fig. 2a](#f2){ref-type="fig"}). As the *tris*NTA^f^ probes are chemically diverse relating to the used fluorophores, common transduction strategies are unlikely to efficiently deliver nanomolar concentrations of *tris*NTA^f^ into mammalian cells. Specifically, mammalian cells were mechanically pushed ('squeezed\') through micrometre constrictions at elevated pressure of 30 psi. This approach allows for high cell survival (\>90%) and efficient uptake of *tris*NTA^f^ (up to 80%; [Supplementary Figs 6 and 7](#S1){ref-type="supplementary-material"}). Energy-dependent endocytosis, often observed at cargo transfer with supercharged molecules ([Supplementary Fig. 8](#S1){ref-type="supplementary-material"}) or low concentrations of CPPs ([Supplementary Fig. 9](#S1){ref-type="supplementary-material"})[@b17], were prevented by performing cell squeezing at 4 °C ([Supplementary Fig. 10](#S1){ref-type="supplementary-material"}). By squeezing TAP1^mVenus-His10^-transfected HeLa cells in the presence of *tris*NTA^f^ (100 nM), we achieved a high-throughput delivery and a high-density labelling, illustrated by an excellent co-localization between both reporter molecules ([Fig. 2b](#f2){ref-type="fig"}). We noticed that both probe delivery by cell squeezing and protein labelling are highly reproducible (*n*\>20). To quantify the *tris*NTA^f^ delivery, we performed flow cytometry analysis on micromanipulated cells. Thirty-five per cent of transfected cells (54% of total) were effectively transduced with *tris*NTA^ATTO565^ ([Fig. 2c](#f2){ref-type="fig"}), which is \> 30 × more efficient for *tris*NTA^f^ delivery than electroporation ([Supplementary Fig. 11](#S1){ref-type="supplementary-material"}). Squeezing of up to 1,000,000 cells per second enables live-cell labelling at high throughput and reproducibility, and hence largely exceeds the efficiency achieved by other direct transfer methods such as microinjection. The massively parallel and constant cell transduction surpasses the stochastic, while variable efficient uptake by CPPs ([Supplementary Fig. 9](#S1){ref-type="supplementary-material"}) and is more than 1,000-fold below the micromolar probe concentrations used by elegant self-labelling tags, for example, SNAP tag[@b3][@b4][@b5][@b6]. Highly specific protein targeting with minimal disturbance ---------------------------------------------------------- Analogous micromanipulation and live-cell labelling was applied to H2B^mVenus-His10^- and ^His10-mEGFP^Lamin A-transfected cells ([Fig. 2d](#f2){ref-type="fig"}). Cell transduction was analysed 15 min and 1 h after squeezing by CLSM. An excellent co-localization between the *tris*NTA^f^ reporter molecule and the His-tagged POIs was observed (Pearson\'s coefficients range from 0.81 to 0.93), in line with the subcellular localization in chemically arrested cells. Notably, live-cell labelling is independent of cell types, for example, HeLa, HeLa Kyoto, Chinese hamster ovary (CHO-K1) or human embryonic kidney 293 cells ([Figs 1b](#f1){ref-type="fig"} and [2d](#f2){ref-type="fig"}). Beyond that, *tris*NTA^f^ labelling after removal of the bulky fluorescent protein fully exploited the small size of the lock-and-key element and confirmed again specific labelling in living cells with minimal perturbation ([Supplementary Fig. 12](#S1){ref-type="supplementary-material"}). The specificity was validated in cells transfected with H2B^EGFP^ lacking a His tag and untransfected cells ([Supplementary Fig. 13](#S1){ref-type="supplementary-material"}). In both cases, neither co-localization of *tris*NTA^f^ with H2B^EGFP^ nor unspecific binding was detected. Delivery of nickel-free *tris*NTA^f^ or of free Alexa647 dye showed no labelling in TAP1^mVenus-His10^-transfected cells ([Supplementary Figs 14 and 15](#S1){ref-type="supplementary-material"}). In contrast, *tris*NTA^Alexa647^ ([Fig. 2](#f2){ref-type="fig"} and [Supplementary Fig. 14](#S1){ref-type="supplementary-material"}) and *tris*NTA^ATTO565^ ([Supplementary Fig. 16](#S1){ref-type="supplementary-material"}) clearly stain TAP1^mVenus-His10^ at the endoplasmic reticulum membrane after cell squeezing. Noticeably, cell viability of *tris*NTA^f^-transduced cells was negligibly affected 1 and 24 h after labelling. Similar concentrations of unbound nickel ions inside mammalian cells had no significant toxic effects ([Supplementary Fig. 7](#S1){ref-type="supplementary-material"}). In contrast, electroporation entailed more than twofold increased toxicity compared with squeezing ([Supplementary Figs 7 and 11](#S1){ref-type="supplementary-material"}). Collectively, nanomolar delivery of *tris*NTA^f^ fully realized the potential of in-cell protein manipulation with minimal perturbation and modification rates exceeding common approaches. After successful in-cell labelling of different His-tagged proteins, we aimed for *in vivo* multiplexed labelling by combining *tris*NTA^f^ with well-established labelling methods. By *tris*NTA^Alexa647^ delivery via squeezing and subsequent SNAP^f^-tag labelling, we achieved specific and distinct targeting of ^His10^LaminA in the presence of two different SNAP^f^-tagged proteins in live cells ([Fig. 3a](#f3){ref-type="fig"} and [Supplementary Fig. 17](#S1){ref-type="supplementary-material"}). Hence, *tris*NTA^f^ complements the toolbox for *in vivo* multiplexed labelling, offering minimal disturbance due to its small size and simultaneously using low nanomolar concentrations. We next determined the minimal reporter concentration required for specific live-cell labelling. Well-resolved images of TAP1^mVenus-His10^ were obtained even at 1 nM of *tris*NTA^Alexa647^ ([Fig. 3b](#f3){ref-type="fig"}). Based on previous observations, approximately one-third of the cargo provided during squeezing is the effective intracellular concentration[@b14]. Thus, the estimated cytosolic concentration of *tris*NTA^f^ further corroborates the high target sensitivity at subnanomolar concentrations (∼300 pM). These results are in line with the detection limit of ∼200 pM *tris*NTA^Alexa647^ in chemically arrested cells ([Supplementary Fig. 2](#S1){ref-type="supplementary-material"}). Hence, this enables the precise adjustment of the effective, intracellular *tris*NTA^f^ concentration to improve the signal-to-background ratio, hardly realized by alternative approaches at nanomolar probe concentrations (for example, CPPs, SNAP, CLIP and Halo tag; [Supplementary Figs 4 and 9](#S1){ref-type="supplementary-material"})[@b3][@b4][@b6][@b18], and circumvents endocytic uptake observed with supercharged proteins at similar nanomolar concentrations ([Supplementary Fig. 8](#S1){ref-type="supplementary-material"})[@b19]. In-cell protein modification with nanometre precision ----------------------------------------------------- Incited by this observation, we aimed at temporal and spatial control of protein tracing by light, which depends on low probe concentrations for high signal-to-background ratios. Using photoactivatable *tris*NTA^f^ (PA-*tris*NTA^ATTO565^)[@b20] for dynamic cellular imaging on demand, light-activated *in vivo* labelling of ^His10-mEGFP^Lamin A was demonstrated up to 24 h after squeezing ([Fig. 3c](#f3){ref-type="fig"}). Notably, already a 10-s 405-nm light pulse sufficiently activated PA-*tris*NTA at single-cell level and led to excellent co-localization in a dose-adapted manner. This probe enables *in situ* labelling at defined time points such as certain mitotic phases and paves the way for live-cell protein tracing with high temporal resolution. The nanomolar concentrations (≤10 nM) and in particular the small size of the tag and probe are especially beneficial for advanced microscopy techniques, bringing the fluorophore in 1-nm proximity to the target protein. Hence, we performed live-cell super-resolution microscopy with *tris*NTA^ATTO655^ on ^His10-mEGFP^Lamin A-transfected cells. Using direct stochastic optical reconstruction microscopy (*d*STORM)[@b21], Lamin A structures with a high signal-to-noise ratio were obtained in the super-resolved images of live cells. A localization precision of 16.4±3.1 nm was achieved, resulting in a resolution of 40 nm by in-cell *tris*NTA^ATTO655^ tracing with substantially increased resolution compared with diffraction-limited fluorescence microscopy ([Fig. 3d](#f3){ref-type="fig"}). Discussion ========== We established a high-throughput method for protein labelling inside living cells using a minimalistic lock-and-key probe. Our method is versatile in the choice of the molecular probe, cell type and the subcellular localization of the POIs, a persistent challenge in live-cell analysis. The high-affinity *tris*NTA/His tag interaction pair enables fast labelling (≤10 min) at subnanomolar concentrations with tunable labelling density and flexibility of cell-impermeable organic fluorophores. Compared with carrier-mediated transport by CPPs[@b11][@b18], delivery of 1,000-fold lower concentrations (nM versus μM) effectively decreases the fluorescence background. Furthermore, high-throughput analysis with up to 1,000,000 cells per second can be achieved in contrast to microinjection. In addition, *tris*NTA^f^ delivery via squeezing avoids endocytic cargo uptake, frequently observed with low CPP concentrations and supercharged molecules, offering decreased toxicity and a \>30-fold higher efficiency compared with electroporation. The minimal probe complements the toolbox of well-established labelling techniques such as self-labelling enzymes and can be combined with the latter to achieve distinct labelling of different proteins in fixed as well as in living cells. Moreover, *in situ* photoactivation of PA-*tris*NTA allows labelling at defined time points, to trace proteins for dynamic cellular imaging. The achieved close target proximity of the labelling pair substantially improved the localization accuracy in live-cell super-resolution microscopy. Remarkable aspects of our approach are the speed, flexibility and efficiency for high-throughput live-cell targeting of proteins even if assembled in stable and transient macromolecular complexes. This study is exploited via one of the smallest high-affinity lock-and-key recognition pairs known so far and allows even multiple cargos to be delivered simultaneously, displaying diverse chemical properties. The quantity of cargo for in-cell manipulation can be precisely tuned and the biological output can in turn be fine-tuned. As the affinity tag is widely used in life sciences and our delivery platform is broadly applicable across cell types, this live-cell labelling method could potentially be implemented across numerous cell-impermeable probes and prodrugs, as well as translated to difficult cell lines including patient-derived cells and embryonic stem cells, providing the opportunity to use these cells for advanced microscopy techniques and live-cell analysis. Methods ======= Plasmid construction -------------------- The H2B construct H2B^mVenus-His10^ was generated by consecutive insertion of H2B and mVenus-His~10~ into pCDNA3.1(+) (Life Technologies). mVenus-His~10~ was PCR amplified using Phusion High-Fidelity DNA Polymerase (Fermentas) and the primer pair forward (fw) 5′-GCGCGC[GCGGCCGC]{.ul}GTGAGCAAGGGCGAGGAGCTGTTCA-3′ (NotI restriction site underlined) and reverse (rev) 5′-GCGCGC[TCTAGA]{.ul}TTAGTGATGGTGGTGATGATGATG-3′ (XbaI restriction site underlined), and cloned into the pCDNA3.1(+) plasmid using the indicated restriction enzymes (Fermentas). Subsequently, H2B was amplified using the primer pair fw 5′-GCGCGC[GGTACC]{.ul}**ATG**CCAGAGCCAGCGAAGTCTGCTCCCGC-3′ (Acc65I restriction site underlined, start codon bold) and rev 5′-GCGCGC[GCGGCCGC]{.ul}TCTTGG AGCTGGTGTACTTAGTGAC-3′ (NotI restriction site underlined), and cloned into the pCDNA3.1(+) plasmid, amino terminally of mVenus-His~10~. As a template for the amplification of H2B, the plasmid pEGFP-N1 containing the human H2B sequence (plasmid 11680, Addgene) was used, which also served as control for the specificity of *tris*NTA^f^ for His-tagged POIs in living cells ([Supplementary Figs 5 and 13](#S1){ref-type="supplementary-material"}). The plasmid encoding for human Lamin A was generously provided by Dr Sascha Neumann (Institute of Biochemistry, University of Cologne)[@b22] and used as template to amplify Lamin A with the primer pair fw 5′-GCGCGC[CTCGAG]{.ul}CTATGGAGACCCCGTCCCAGCGGCGCGCCACCCG-3′ (XhoI restriction site underlined) and rev 5′-GCGCGC[GATATC]{.ul}**TTA**CATGATGCTGCAGTTCTGGGGGCTCTGGG-3′ (EcoRV restriction site underlined, stop codon bold). Lamin A was inserted into the pcDXC3GMS plasmid (Addgene) already containing His~10~-mEGFP, leading to the fusion gene encoding for ^His10--mEGFP^Lamin A. To generate the plasmid containing ^His10^Lamin A without a fluorescent protein, Lamin A was amplified via PCR, simultaneously introducing a His~10~-tag with the primer pair fw 5′-GCGCGC[GGATCC]{.ul}ACC**ATG***CACCATCATCATCATCATCACCACCATCAC*TCCGGACTCAGATCTCGAGTCATGGAGACCC-3′ (BamHI restriction site underlined, start codon bold, His~10~-tag italic) and rev 5′-GCGCGC[GCGGCCGC]{.ul}**TTA**CATGATGCTGCAGTTCTGGGGGCTCTGGG-3′ (NotI restriction site underlined, stop codon bold). Indicated restriction sites were used to insert this PCR product into the pCDNA3.1 (+) vector. The pSNAP^f^-H2B and pSNAP^f^-Cox8A plasmids (New England Biolabs) were used for SNAP^f^-tag labelling. In addition, a plasmid coding for the core domain of TAP1, tagged with mVenus-His~10~ (TAP1^mVenus-His10^) was used as previously described[@b23]. Cell culture and transfection ----------------------------- HeLa cells, HeLa Kyoto cells, Chinese hamster ovary (CHO-K1) cells and human embryonic kidney 293 cells were maintained in DMEM medium with 4.5 g l^−1^ glucose (Gibco), supplied with 10% (v/v) FCS (Gibco) in T75 cell culture flasks (Greiner). Every 2--3 days, cells were passaged using PBS (Sigma-Aldrich) and 0.5% trypsin/0.02% EDTA/PBS (GE Healthcare). All cell lines were cultivated in a humidified tissue culture incubator at 37 °C and 5% CO~2~. Mycoplasma contamination tests were carried out regularly, following the guidelines described[@b24]. Transient transfection was performed with Lipofectamine 2000 (Life Technologies), following the manufacturer\'s instructions. For fixation and staining, 2 × 10^4^ cells per well were seeded into eight wells on cover glass II slides (Sarstedt) and transfected with 0.2 μg DNA per well. For squeezing experiments, 8 × 10^5^ cells were seeded into six-well cell culture plates (Greiner) and transfected with 2 μg DNA per well. After transfection, cells were incubated 12--48 h at 37 °C and 5% CO~2~ until experiments were performed. Cell viability test ------------------- To analyse cell viability after squeezing, the Sytox Blue Dead Cell stain (Life Technologies) was used to stain cells with a permeable plasma membrane. Cells were squeezed in the presence of 100 nM *tris*NTA^Alexa647^, 100 nM Alexa647 dye or 500 nM NiCl~2~, followed by incubation with 1 μM of Sytox Blue Dead Cell Stain (20 min, room temperature (RT)) at different time points. Cell viability analysis was performed using the Attune flow cytometer (Life Technologies) and data were processed using FlowJo 7.6.5 (Tree Star Inc.). Before detaching the cells, the supernatant was collected to avoid altering the results by removing dead cells during washing. Identical Sytox Blue Dead Cell Staining was conducted with cells after electroporation (described above), followed by flow cytometry analysis of cell viability and uptake of *tris*NTA^ATTO565^. All experiments were performed in triplicates and error bars indicate the s.d. Confocal imaging ---------------- Imaging was performed using the confocal laser scanning TCS SP5 microscope (Leica) and a Plan-Appochromat 63 × 1.4 Oil differential interference contrast objective. Images were acquired sequentially to avoid cross-talk. The following laser lines were used for excitation: 405 nm (diode laser) for 4,6-diamidino-2-phenylindole, Hoechst and Sytox Blue Dead Cell Stain; 488 nm (argon laser) for monomeric enhanced green fluorescent protein (mEGFP), Alexa488, MitoTracker^Green^ and Syto16; 488 or 514 nm (argon laser) for mVenus; 565 nm (diode-pumped solid-state laser) for ATTO565 and TMR-Star; and 633 nm (helium-neon laser) for ATTO647N, Alexa647 and ATTO655. For image analysis, ImageJ[@b25] was used in combination with Fiji[@b26]. Super-resolution imaging ------------------------ A custom-built microscope was used for super-resolution imaging of *tris*NTA^ATTO655^-labelled ^His10-mEGFP^Lamin A[@b27]. Samples were illuminated with 488 nm (Sapphire 488 LP, Coherent) and 643 nm (iBeam smart, Toptica Photonics) laser beams in total internal reflection fluorescence mode. The excitation light was focused on the back focal plane of a 100 × oil objective (PLAPO 100 × , total internal reflection fluorescence mode, numerical aperture 1.45, Olympus) mounted on an inverted microscope (Olympus IX71). The emission was recorded using an electron multiplying charge-coupled device camera (Ixon3, Andor) with frame-transfer mode, 5.1 × pre-amplifier gain and electron multiplying (EM) gain set to 200. For every sample, 40,000 images were recorded at a frame rate of 33 Hz and image reconstruction was performed with rapidSTORM[@b28]. The localization (*σ*~loc~) precision of the *d*STORM images was calculated to 16.4±3.1 nm and a resolution of ≤40 nm (for *d*STORM image see [Fig. 3d](#f3){ref-type="fig"}). Calculations were performed according to Mortensen *et al.*[@b29] Fixation and *tris*NTA^f^ labelling ----------------------------------- Before fixation, cells were washed with PBS (Sigma-Aldrich). Fixation with 4% formaldehyde (Roth)/PBS for 15 min at RT was followed by quenching using 50 mM glycine/PBS (10 min, RT; Roth) and permeabilization with 0.1% Triton X-100/PBS (10 min, RT; Roth). After blocking with 5% (w/v) BSA (Albumin Fraction V, Roth) in PBS (1 h, RT), cells were stained with 100 nM of *tris*NTA^f^ in 1% (w/v) BSA/PBS. Cells were stained with 0.1 μg ml^−1^ 4′,6-diamidino-2-phenylindole (Sigma-Aldrich) in 1% BSA/PBS for 30 min--1 h at RT. After washing with 5% BSA/PBS (3 × ), cells were postfixed with 2% formaldehyde/PBS (15 min, RT) and stored in PBS until confocal imaging was performed. When *tris*NTA^f^ labelling was combined with antibody labelling, 100 nM *tris*NTA was applied together with the primary antibodies for 1 h at RT. After subsequent washing with PBS, cells were incubated with secondary antibodies. Antibodies were all diluted in 1% BSA/PBS as follows: rabbit α-tubulin (Abcam, 1:350), mouse α-human CD107a (LAMP-1; Biolegend, 1:1,000), goat α-rabbit^Cy3^ (Abcam, 1:1,000) and goat α-mouse^PacificBlue^ (Life Technologies, 1:1,000). *tris*NTA was conjugated with the corresponding fluorescent dyes by reacting amino-*tris*NTA with the N-hydroxysuccinimide (NHS)-activated dye, respectively. After RP-C~18~ HPLC chromatography, the multivalent chelator head was loaded with Ni^2+^ (10 mM NiCl~2~ in 20 mM HEPES, pH 7.0) and purified by anion exchange chromatography (HiTrap Q HP; GE Healthcare)[@b11]. SNAP^f^-tag labelling in fixed cells ------------------------------------ HeLa Kyoto cells transfected with H2B^SNAPf^ were labelled with O6-benzyl guanine^Alexa647^ (New England Biolabs) 24 h after transfection. After washing with PBS, cells were fixed with 4% formaldehyde/PBS (10 min, RT), washed twice with PBS and subsequently permeabilized using 0.1% Triton X-100/PBS (10 min, RT). Cells were washed three times with 0.5% BSA/PBS and incubated with 10 or 100 nM of benzyl guanine^Alexa647^ for 30 min at 37 °C. After three washing steps with 0.1% Triton X-100/0.5% BSA/PBS and two washing steps with PBS, imaging was conducted by CLSM. In case of combined *tris*NTA^f^- and SNAP^f^-tag labelling, SNAP^f^-tag labelling was performed first, followed by *tris*NTA^f^ labelling as described above. *tris*NTA^f^ delivery via cell squeezing ---------------------------------------- Squeezing was performed using a chip with constrictions of 7 μm in diameter and 10 μm in length (CellSqueeze 10-(7)x1, SQZbiotech), if not otherwise stated. In all microfluidic experiments, a cell density of 1.5 × 10^6^ cells per ml in 10% (v/v) FCS/PBS were squeezed through the chip at a pressure of 30 psi. Transduction was conducted at 4 °C, to block cargo uptake by endocytosis. During squeezing, the following cargo concentrations were used: 0.1--100 nM of *tris*NTA^f^ (f=ATTO565, ATTO647N, Alexa647 or ATTO655), 200 nM of PA-*tris*NTA^ATTO565^ and 100 nM of free Alexa647 (Molecular Probes). After squeezing, cells were incubated for 5 min at 4 °C, to reseal the plasma membrane[@b15]. Squeezed cells were washed with DMEM containing 10% FCS and 10 mM histidine (Sigma-Aldrich), to remove unspecifically bound *tris*NTA^f^ from the cell surface, seeded into eight wells on cover glass II slides (Sarstedt) in DMEM containing 10% FCS and cultured at 37 °C and 5% CO~2~. Confocal imaging was performed at different time points (0.25, 0.5, 1, 2 and 24 h) after squeezing. As a control for endosomal uptake, cells were incubated with 100 nM of *tris*NTA^f^ at RT and 4 °C without microfluidic cell manipulation ([Supplementary Fig. 10](#S1){ref-type="supplementary-material"}). In case of combined *tris*NTA^f^ and SNAP^f^-tag labelling, cells were first squeezed in the presence of 100 nM *tris*NTA^Alexa647^ and 3 μM SNAP-Cell TMR-Star (New England Biolabs) was added 5 min after squeezing. After 15 min incubation at 37 °C and 5% CO~2~, cells were washed with 10% FCS/DMEM and incubated again 30 min under standard cell culture conditions before CLSM imaging was performed. To visualize mitochondria in case of Cox8A^SNAPf^ labelling MitoTracker^Alexa488^ was added 10 min prior imaging. *tris*NTA^f^ delivery by the CPP Tat~49--57~ -------------------------------------------- TAP^mVenus-His10^-transfected HeLa Kyoto cells were incubated with different concentrations (10 μM and 100, 10 and 1 nM) of a non-covalent complex composed of Tat~49--57~^His6^ and *tris*NTA^Alexa647^ for 30 min. After washing with 10% FCS/DMEM and PBS at 37 °C, *tris*NTA uptake and labelling of His~10~-tagged TAP was analysed by live-cell imaging via CLSM[@b11]. *tris*NTA^f^ delivery via supercharged GFP (GFP^36+^) ----------------------------------------------------- *tris*NTA^ATTO565^ (100 nM) and ^His6^GFP^36+^ (100 nM) were incubated for 30 min at RT to form the *tris*NTA-His tag complex. HeLa cells were washed with PBS and treated with the pre-formed complex at 37 °C. *In vivo* uptake was immediately followed by CLSM. After 20 min, cells were washed three times with PBS and 20 U ml^−1^ heparin/PBS (2 × ), to remove the complex from the plasma membrane. Internalization of *tris*NTA/^His6^GFP^36+^ was analysed by CLSM after washing three times with PBS. ^His6^GFP^36+^ was expressed in BL21(DE3) *Escherichia coli.* After lysis by sonication in 2 M NaCl/PBS, ^His6^GFP^36+^ proteins were purified via immobilized metal ion affinity chromatography using Ni Sepharose 6 Fast Flow (GE Healthcare). Elusion was performed with 500 mM imidazole before desalting of the eluted protein was conducted with PD-10 desalting columns (GE Healthcare)[@b19]. *tris*NTA^f^ delivery via electroporation ----------------------------------------- Trypsinized HeLa Kyoto cells were permeabilized via electroporation in the presence of *tris*NTA^ATTO565^ (100 nM) with a Nucleofector Device (Lonza) using Nucleofector Kit V and program I-013. Transduced cells were washed with 10% FCS/DMEM and plated into eight-well cover glass II slides. To quantify *tris*NTA^f^ uptake, flow cytometry analysis (as described above) and CLSM imaging was performed 1 h after electroporation. Before CLSM imaging, transduced HeLa cells were incubated with Sytox Blue Dead Cell Stain or Syto16 (Life Technologies) live cell stain, to distinguish between dead and live cells. Cell viability was analysed by flow cytometry using Sytox Blue Dead Cell Stain (see above). Photoactivation of PA-*tris*NTA^f^ ---------------------------------- HeLa Kyoto cells, transfected with ^His10-mEGFP^ Lamin A, were squeezed in the presence of PA-*tris*NTA^ATTO565^ (200 nM)[@b20] and incubated at 37 °C and 5% CO~2~. Imaging via CLSM and photoactivation was conducted 2 or 24 h after squeezing, after cells reattached to the glass surface. PA-*tris*NTA^ATTO565^ was photoactivated in single cells by illumination with the 405-nm laser for 10 s. Imaging of both channels (mEGFP and ATTO565) was performed before and directly after photoactivation. Additional information ====================== **How to cite this article:** Kollmannsperger, A. *et al.* Live-cell protein labelling with nanometre precision by cell squeezing. *Nat. Commun.* 7:10372 doi: 10.1038/ncomms10372 (2016). Supplementary Material {#S1} ====================== ###### Supplementary Information Supplementary Figures 1-17 The German Research Foundation (Cluster of Excellence---Macromolecular Complexes to R.W., M.H. and R.T., as well as CRC 807, SPP 1623 and RTG 1986 to R.T. and SFB 807 to M.H.) supported the work. We thank Drs Sascha Neumann (Institute of Biochemistry, University of Cologne, Germany) and Ulrich Rothbauer (The Natural and Medical Sciences Institute, University of Tübingen, Germany) for generously providing us with the original Lamin A construct and the HeLa Kyoto cells, respectively. Furthermore, we thank Valentina Herbring and Dr Peter Mayerhofer for help with flow cytometry, and Markus Braner for helpful suggestions on the manuscript. **Author contributions** A.K. designed and performed the cell squeezing and labelling experiments. A.S. determined the squeezing efficiency. A.S., R.L. and K.F.J. designed and provided the microfluidic devices. A.R. and M.H. performed the *d*STORM imaging and analysis. A.K., R.W. and R.T. wrote the manuscript and analysed the data. R.W. and R.T. conceived the ideas and directed the work. {#f1} {#f2} {#f3}

\section{Introduction} % % % % \IEEEPARstart{W}{ith} the increasing crowd density in public transportation hubs, security inspection is widely adopted to examine the passengers' belongings with X-ray scanners, making sure that no prohibited item is mixed into the luggage. Usually, items in the luggage are intensely occluded since they are randomly stacked and heavily overlapped with each other. Therefore, the prolonged time of inspecting in a myriad of complex X-ray images with unremitting attention may tire the security inspectors to accurately distinguish all the prohibited items, which may lead to public danger. Therefore, it is indispensible to adopt an efficient, precise, and automatic approach to assist inspectors in checking the passengers' belongings in X-ray images. \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{images/x-ray_knife.pdf} \caption{Samples of the five categories of cutters in OPIXray dataset and their corresponding noisy X-ray images.} \label{Samples_five} \end{figure} \begin{table*}[!t] \begin{center} \newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}} \setlength{\tabcolsep}{1.5mm} \caption{Comparison of existing open-source X-ray datasets.} { \small \begin{tabular}{lccccccccc} \toprule \multirow{2}{*}{Dataset} & \multirow{2}{*}{Year} & \multirow{2}{*}{Color} & \multirow{2}{*}{Category} & \multirow{2}{*}{Positive Samples} & \multicolumn{3}{c}{Annotation} & \multirow{2}{*}{Task} & \multirow{2}{*}{Data Source}\\ \cmidrule{6-8} & && && Bounding box & Number & Professional & &\\ \midrule GDXray \cite{mery2015gdxray} & 2015& Gray-scale & 3 & 8,150 & \footnotesize\Checkmark & 8,150 & \footnotesize\XSolidBrush & Detection & Unknown\\ SIXray \cite{miao2019sixray} & 2019 &RGB& 6 & 8,929 & \footnotesize\XSolidBrush& \footnotesize\XSolidBrush& \footnotesize\XSolidBrush & Classification & Subway Station\\ \midrule \textbf{OPIXray} & 2020&RGB & 5 & 8,885 & \footnotesize\Checkmark & 8,885 & \footnotesize\Checkmark & Detection & Airport\\ \bottomrule \end{tabular} } \end{center} \label{dataset-comparison} \end{table*} Fortunately, the innovation of deep learning, especially the convolutional neural network, makes it possible to accomplish the goal by transferring it into object detection task in computer vision \cite{sultana2020unsupervised,wu2020deeper,li2020stacked,ren2020salient,jiang2020cmsalgan}. However, different from traditional detection tasks, in this scenerio, items within a luggage are randomly overlapped where most areas of objects are occluded resulting in noisy X-ray images with heavy occlusions. Thus, this characteristic of these noisy X-ray images leads to strong requirement of high-quality datasets and models with satisfactory performance for this detection task. Regarding dataset, to the best of our knowledge, there are two released X-ray benchmarks for classification task, namely GDXray \cite{mery2015gdxray} and SIXray \cite{miao2019sixray}, which are incommensurate with our target of occluded prohibited item detection. Additionally, all images in GDXray are gray-scale, and less than 1\% pictures in SIXray are annotated with prohibited items, both of which are not of very high quality. In addition, traditional CNN-based models \cite{li2021image,tang2019salient,qiu2020hierarchical,xiao2018deep,fu2018refinet,qin2020forward} trained through common image recognition datasets fail to achieve satisfactory performance in this scenario. This urgently requires researchers to make breakthroughs in both datasets and models. In this work, we contributed the first high-quality dataset named OPIXray, which aims at occluded prohibited items detection in security inspection. Considering that cutter is the most common tool passengers carry, we choose it as the prohibited item to detect. As a result, OPIXray contains 8885 X-ray images of 5 categories of cutters (illustrated in Fig. \ref{Samples_five}). Besides, we provide abundant analysis and discussions of the OPIXray dataset including the construction principles, the quality control procedures, and the potential tasks. In order to simulate the real security inspection scene to the greatest extent, the images of the dataset are gathered from an airport and these prohibited items are annotated manually by professional inspectors, which can be used as a benchmark for model training and further facilitate future research. In addition, to further improve occluded X-ray object detection, we propose an over-sampling de-occlusion attention network (DOAM-O), which consists of a de-occlusion attention module (DOAM) and an new over-sampling training strategy. DOAM is a plug-and-play module, which can be easily inserted to promote most popular detectors. DOAM simultaneously leverages the different appearance information (shape and material) of the prohibited item to generate the attention map, which helps refine feature maps for the general detectors. The over-sampling training strategy retrains these hard samples (i.e., prohibited items containing high-level occlusions) to extract more effective information from the hard X-ray samples, further boosting the performance of the model. To the best of our knowledge, this is the first work that explicitly exploits over-sample technologies to the X-ray prohibited items, which might offer a new way to solve the occlusion detection problem under noisy data scenarios. Note that in this paper, we extend our prior conference publication \cite{WeiOccluded2020} which contributes the rudiment of the OPIXray dataset and proposes a de-occlusion network which mainly consists of the DOAM module with traditional training strategy (DOAM-T). Compared with our previous conference paper, we add more analysis and discussions of the OPIXray dataset including the construction principles, the potential tasks, etc., to make the process of constructing this dataset clearer. We hope these investigations can be beneficial to facilitate the expansion and scale-up of OPIXray in the future research as well as the development of prohibited items detection in X-ray security inspection. Besides, we extend the model of our previous conference paper by replacing the traditional training strategy of the model with a new over-sampling training strategy (DOAM-O). The new strategy could make the model learn to put more emphasis on these hard samples the existed massively under the X-ray detection dataset, which is more suitable for the prohibited items detection scenario. Experiments show that DOAM-O could beat the DOAM-T by a significant margin in most cases, validating our claim that the over-sampling training strategy could help the model put more emphasis on these hard samples. Moreover, we discuss different implementations for the over-sampling training strategy and different attention mechanisms, ensuring our method is theoretically reasonable and yields competitive performance. Our contributions can be listed as follows: \begin{itemize} \item{We contribute the first high-quality dataset, which named OPIXray, which aims at occluded prohibited items detection in security inspection. Besides, we introduce abundant analysis and discussions for the OPIXray dataset. We hope that contributing this high-quality dataset can promote the development of prohibited items detection in noisy X-ray images.} \item{We propose a plug-and-play module, namely DOAM, which simultaneously leverages the different appearance information (shape and material) of items. DOAM helps refine feature maps for the general detectors, promoting the performance of most popular detectors.} \item{We propose an over-sampling training strategy, making models learn to put more emphasis on these hard samples existed massively, which is more suitable for this scenario. To the best of our knowledge, this is the first work exploiting the over-sampling technologies to prohibited items detection, which might provide a new way to solve the occlusion problem in noisy X-ray images.} \item{Extensive experiments were conducted on the published dataset OPIXray, which is the only dataset for X-ray prohibited items detection. The results demonstrate that our method can drastically boost the detection accuracy and achieve the new SOTA performance for this task.} \end{itemize} The structure of the paper is illustrated as follows: Section \ref{Section:relatedwork} introduces the related works; Section \ref{Section:framework} introduces the OPIXray dataset we contributed; Section \ref{Section:model} describes and discusses the proposed de-occlusion methodology and over-sampling training strategy; Section \ref{Section:exp} demonstrates the experiments; and Section \ref{Section:conclusion} summarizes the whole contributions and provides the conclusion. \section{Related Work}\label{Section:relatedwork} \subsection{X-ray Images and Benchmarks} X-ray shows great power in numerous tasks, such as medical imaging analysis \cite{guo2019improved,chaudhary2019diagnosis,lu2019towards} and security inspection \cite{miao2019sixray,huang2019modeling,hassan2020detecting}. Notwithstanding, occlusion between items could disturb the accessibility of the information contained in X-ray images. Several studies in the literature have endeavored for addressing this challenging issue. However, only a few X-ray datasets have been proposed because of the distinctiveness of security inspection. One released benchmark, named GDXray\cite{mery2015gdxray}, contains 19407 images, partial of which embraces prohibited items of three categories, including gun, shuriken and razor blade. However, images in GDXray are gray-scale with a very simple background, which is far away from complex real-world scenario. SIXray\cite{miao2019sixray}, a recently released benchmark, is a large-scale X-ray dataset consisting of 1059231 X-ray images, which is about 100 times larger than the GDXray dataset\cite{mery2015gdxray}. But the positive samples are deficient (less than 1\%) for imitating a comparable real-world testing environment where prohibited items that inspectors target to only show up with low recurrence. Besides, different from ours, SIXray is a dataset for the task of classification, concentrating on the issue of data imbalance. \subsection{Attention Mechanism} Attention is a method which tends to dispose the most instructive parts of signals among all the accessible computational resources. It has been widely-used in numerous scenes, such as action recognition \cite{li2018unified,li2020spatio,hou2017content}, visual question answering \cite{yu2020reasoning,zhang2019frame,huasong2020self}, and adversarial learning \cite{Liu2019Perceptual,hu2020adversarial,Liu2020Spatiotemporal,li2020attentionfgan}. Attention grasps long-range contextual information which can be generally applied in different tasks, including machine translation \cite{vaswani2017attention}, image captioning \cite{chen2017sca}, scene segmentation \cite{fu2019dual} and object recognition \cite{tang2015rgb}. \cite{wang2018non} utilized self-attention module to explore the effectiveness of non-local operation in space-time dimensions for videos and images. \cite{fu2019dual} proposed a dual attention network (DANet) for scene segmentation by capturing contextual dependence based on the self-attention mechanism. Squeeze-and-Excitation Networks (SENet) \cite{hu2018squeeze}, terming the Squeeze-and-Excitation block (SE), models inter-dependencies between channels, which adaptively re-calibrates channel-wise feature. \subsection{Object Detection} In computer vision, object detection is one of important tasks, which underpins a few instance-level recognition tasks and many downstream applications \cite{zhou2020salient,qiu2020hierarchical,fu2018refinet}. Most of the CNN-based methods can be further divided into two general approaches: proposal-free detectors and proposal-based detectors. Recently one-stage methods have gained much attention over two-stage approaches due to their simpler design and competitive performance. Due to that the security inspection has a high demand for time cost, in this paper, we mainly use single-stage models as a comparison. Here we review some work that is the closest to ours. SSD \cite{liu2016ssd} discretizes the output space of bounding boxes into a set of default boxes over different aspect ratios and scales. YOLO \cite{redmon2016you, redmon2017yolo9000, redmon2018yolov3, bochkovskiy2020yolov4} is the collection of a series of well-known methods which values both real-time and accuracy among single-stage detection algorithms. FCOS \cite{tian2019fcos} proposes a fully convolutional one-stage object detector to solve object detection in a per-pixel prediction fashion, analogue to other dense prediction problems such as semantic segmentation. \section{The OPIXray Dataset}\label{Section:framework} The quality of the dataset is an essential problem because more rubust and sophisticated models can hardly be derived from coarse dataset. By exploiting images in high quality dataset, better models and algorithms can be vigorously provoked, reasonably evaluated, and fairly compared. Needless to say, a professional dataset with high-quality annotations is much more critical in special practice scenarious. In this work, we contribute to the first and high-quality dataset aiming at occluded prohibited items detection in security inspection. Next we introduce the construction principles, data properties and potential tasks of the proposed OPIXray dataset. \begin{table}[!t] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{7pt} \caption{The category distribution of the OPIXray dataset.} \label{data_table} \small \begin{tabular}{ccccccc} \hline \multirow{2}{*}{OPIXray} & \multicolumn{5}{c}{Categories} & \multirow{2}{*}{Total} \\ \cline{2-6} & FO & ST & SC & UT & MU & \\ \hline Training & 1589 & 809 & 1494 & 1635 & 1612 & 7109 \\ Testing & 404 & 235 & 369 & 343 & 430 & 1776 \\ \hline Total & 1993 & 1044 & 1863 & 1978 & 2042 & 8885 \\ \hline \end{tabular} \end{table} \subsection{Construction Principles} We construct the OPIXray dataset in accordance with the following two principles: \begin{itemize} \item \textbf{Common Category}. The categories of prohibited items of the images should belong to those frequently seen in daily life, which makes the application more practical. \item \textbf{Professional Annotation}. Objects in X-ray images are quite different from those in natural images, whose color-monotonous and luster-insufficient characteristics make them difficult to be recognized for people without professional training. Therefore, the boundary box of prohibited items in X-ray images should be annotated professionally. \item \textbf{Quality Control Procedures}. The construction of data sets needs to include the necessary quality control procedures, making further data collection in this domain reliable if someone want to scale this work up, or use it in another setting or environment. \item \textbf{Explicit Task}. The dataset should be set according to the problem to be solved, so that the relevant experimental results can verify the effectiveness of the model for solving this problem. \end{itemize} Therefore, We construct our dataset based on the four principles. \textbf{First}, we select five categories of ``cutters'' as the common prohibited items to detect. The background images of the OPIXray dataset are generated by the latest X-ray imaging machine and gathered from daily security inspections in the international airport. \textbf{Second}, professional security inspectors synthesized the samples by embedding the specific prohibited item into the background images. And prohibited item in each sample is annotated manually with a bounding box by professional inspectors from the international airport, which is the procedure to train security inspectors. \textbf{Third}, We followed the similar quality control procedures of annotation as the famous Pascal VOC. Each prohibited item is annotated by professional airport security inspectors. All inspectors followed the same annotation guidelines including what to annotate, how to annotate bounding, how to treat occlusion, etc. In addition, the accuracy of each annotation was checked by another professional inspector, including checking for omitted objects to ensure exhaustive labelling. \textbf{Finally}, for further study of the impact brought by object occlusion, we classify the testing set into three subsets according to the degree of occlusion, and name them in grades: Occlusion Level 1 (OL1), Occlusion Level 2 (OL2) and Occlusion Level 3 (OL3), where higher level implies severer occlusion around prohibited items in image. As illustrated in Fig. \ref{Test_Occ}, there is no or slight occlusion on prohibited items in OL1 and partial occlusion in OL2, and severe or full occlusion in OL3. Tab. \ref{data_cate_table} shows the category statistic of these three subsets with different occlusion levels. \subsection{Data properties} The OPIXray dataset contains a total of 8885 X-ray images, 5 categories of common cutters, namely, ``Folding Knife", ``Straight Knife", ``Scissor'', ``Utility Knife'' and ``Multi-tool Knife'' (illustrated in Fig. \ref{Samples_five}). Some images contain more than one prohibited item and every prohibited item is located with a bounding-box annotation. The statistics of category distribution are shown in Tab. \ref{data_table}. All images are stored in JPG format with the resolution of 1225*954. The dataset is partitioned into a training set and a testing set, where the ratio is about 4 : 1. The statistics of category distribution of training set and testing set are also shown in Tab. \ref{data_table}. Note that due to that some images contain more than one prohibited item, the sum of all items in the different categories is greater than the total number of images. According to our statistics, there are about 35 samples (30 in the training set and 5 in the testing set) containing more than one prohibited items. \begin{figure}[!t] \centering \includegraphics[width=\linewidth]{images/test_occ.pdf} \caption{Samples of different occlusion levels in OPIXray dataset. ``OL'' refers to ``occlusion level''. There is no or slight occlusion on prohibited items in OL1, partial occlusion in OL2 and severe or full occlusion in OL3.} \label{Test_Occ} \end{figure} \subsection{Other Potential Tasks} In addition to occluded object detection, our OPIXray dataset can further serve the evaluation of various detection tasks including small object detection and few-shot object detection. \textbf{Small Object Detection}. Security inspectors often struggle to find small prohibited items in baggage or suitcase. In our OPIXray dataset, there are many small prohibited items. According to the definition of small by SPIE, the size of small object is usually no more than 0.12\% of entire image size. Suppose we define small object as the object whose max rectangle area is no more than 0.12\% of entire image, all the prohibited items in our dataset are small objects while other safe items (such as personal computers, umbrellas, etc.) are large items. \begin{table}[!t] \centering \renewcommand\arraystretch{1.2} \setlength{\tabcolsep}{9pt} \caption{The category distribution of different occlusion levels in the testing set of the OPIXray dataset.$^*$} \label{data_cate_table} \small \begin{threeparttable} \begin{tabular}{ccccccc} \hline \multirow{2}{*}{Testing} & \multicolumn{5}{c}{Categories} & \multirow{2}{*}{Total} \\ \cline{2-6} & FO & ST & SC & UN & MU & \\ \hline OL1 & 206 & 88 & 160 & 214 & 255 & 922 \\ OL2 & 148 & 84 & 126 & 88 & 105 & 548 \\ OL3 & 50 & 63 & 83 & 41 & 70 & 306 \\ \hline Total & 404 & 235 & 369 & 343 & 430 & 1776 \\ \hline \end{tabular} \begin{tablenotes} \footnotesize \item $^*$Note that ``OL'' refers to ``Occlusion Level''. \end{tablenotes} \end{threeparttable} \end{table} \textbf{Few-Shot Object Detection}. In the real security scene, The frequency of samples containing prohibited items is very low. As a result, images of the dataset can be used to construct a new dataset for few-shot object detection, which is more consistent with the real scene settings. Besides, in security inspection, some categories of prohibited items are extremely rare, leading to a very small amount of training data. Therefore, ``few-shot object detection'' can become a research track of X-ray prohibited items detection. Researchers can be inspired by relevant methods of few-shot learning or meta-learning to address the problem that some categories of prohibited items with few samples are hard to detect in security inspection. \begin{figure*}[!t] \centering \includegraphics[width=\linewidth]{images/doam_tip.pdf} \caption{The network architecture of the de-occlusion attention module (DOAM) integrated with a general backbone network. As illustrated, two feature maps are generated by Edge Guidance (EG) and Material Awareness (MA) and fused to generate the attention map in Attention Generation. Further, the attention map is applied to the input image to generate refined feature maps we desire. Finally, the refined feature map can be utilized by the SSD network.} \label{doam_tmm} \end{figure*} \section{Over-sampling de-occlusion attention network}\label{Section:model} This section describes the details of the proposed network, Over-sampling de-occlusion attention network (DOAM-O). As the de-occlusion attention module (DOAM) lies at the core of the proposed method, first, we elaborate the details of the de-occlusion attention module including the edge guidance (EG), material awareness (MA) and attention generation (AG). Second, we introduce the over-sampling strategy for selecting hard samples in detail. Finally, we discuss the essential difference between our method and other congeneric mechanisms. \begin{figure}[!t] \begin{algorithm}[H] \caption{The Procedure of DOAM.} \label{Alg} \begin{algorithmic}[1] \STATE \textbf{Input}: An X-ray image $\mathbf{x}\in\mathbb{R}^{C\times{H}\times{W}}$; \STATE Generate the horizontal edge image $\mathbf{E}^{h}$ and the vertical edge image $\mathbf{E}^{v}$ by the $Sobel$ operator. \STATE Generate the edge image ${E}$ by synthesizing $\mathbf{E}^{h}$ and $\mathbf{E}^{v}$. \FOR {$\mathrm{N}_1$ steps} \STATE Refine the feature map $\mathbf{A}$ through ${\mathcal{F}_\mathbf{e}}(\cdot)$. \ENDFOR \STATE Generate the image $\mathbf{P}$ by concatenating $\mathbf{x}$ and ${E}$. \FOR {$\mathrm{N}_2$ steps} \STATE Refine the feature map $\mathbf{B}_{1}$ through ${\mathcal{F}_\mathbf{m}}(\cdot)$. \ENDFOR \FOR {$k\in\left\{\mathbf{k}_1,\ldots,\mathbf{k}_n\right\}$} \STATE Generate the refined feature map $\mathbf{B}_{2}^k$ through Eq. (\ref{B_2}). \STATE Generate $\mathbf{B}_{3}^k$ by concatenating $\mathbf{B}_{1}$ and $\mathbf{B}_{2}^k$. \STATE Update the feature map set $\LARGE{U}$ = $\LARGE{U} \cup \mathbf{B}_{3}^k$. \ENDFOR \STATE Choose the appropriate feature map $\mathbf{B}$ from $\LARGE{U}$ by drawing the gated convolutional network $\mathcal{G}$. \STATE Generate the fused feature map $\mathbf{C}$ by operating $\mathbf{A}$ and $\mathbf{B}$. \STATE Generate the attention map $\mathcal{M}=\sigma(\mathbf{C})$. \STATE Generate the final feature map $\mathbf{D}$ by performing a matrix multiplication between $\mathcal{M}$ and $\mathbf{P}$ through Eq. (\ref{D}). \STATE \textbf{Output}: the final refined feature map $\mathbf{D}\in\mathbb{R}^{C_h\times{H}\times{W}}$. \end{algorithmic} \end{algorithm} \end{figure} \subsection{De-occlusion Attention Module} The DOAM simultaneously lays particular emphasis on edge information and material information of the prohibited item by utilizing two sub-modules, namely, Edge Guidance (EG) and Material Awareness (MA). Then, our module leverages the two information above to generate an attention distribution map as a high-quality mask for each input sample to generate high-quality feature maps, serving identifiable information for general detectors. \subsubsection{Edge Guidance (EG)} For each input sample $\mathbf{x}\in\LARGE{X}$, we compute the edge images $\textbf{E}^{h}$ and $\textbf{E}^{v}$ in horizontal and vertical directions using the convolutional neural network with $Sobel$ operators $s_h$, $s_v$, which denote the horizontal and vertical kernel respectively. By jointly exploit the above two results $\textbf{E}^{h}$ and $\textbf{E}^{v}$, the entire edge image $\textbf{E}$ of input image $\mathbf{x}$ is further generated. Next, we define $\mathrm{N}_1$ as the Module Operation Intensity of EG, which has a strong correlation with the performance of module. In order to limit EG to only amplify the edge information of the prohibited items, we use $\mathrm{N}_1$ network blocks, each of which consists of a convolutional layer with a $3\times3$ kernel size, a batch normalization layer, and a ReLU layer, to extract the feature map $\mathbf{A}$, which is the final output of the EG module. The operations can be formulated as follows: \begin{equation} \mathcal{F}_{\mathbf{e}}(\mathbf{x})=\mathrm{ReLU}\left(\mathbf{W}_{e}\cdot\mathbf{x}+\mathbf{b}_{e}\right), \end{equation} \begin{equation}\label{F:EG} \mathbf{A}=\left\{\mathcal{F}_{\mathbf{e}}({E})\right\}_{\mathrm{N}_1}, \end{equation} where $\left\{\cdot\right\}_{\mathrm{N}_1}$ means that the operation is repeated $\mathrm{N}_1$ times. $\mathbf{W}_{e}$, $\mathbf{b}_{e}$ are parameters of the convolutional layer. As shown in Eq. \ref{F:EG}, the feature map $\mathbf{A}$ laying emphasis on edge guidance information is extracted, which helps the model to adaptively give more attention to the edge features of the prohibited items. \begin{figure*}[!t] \centering \includegraphics[width=\linewidth]{images/framework_0224.pdf} \caption{The over-sampling strategy of DOAM-O. $\LARGE{X}_i$ refers to the $i$-th epoch training data inputted, $\LARGE{X}_{m}$ refers the images of one batch whose loss is minimum in the ``hard sample pool'' $S$ and $\mathcal{L}_{m}$ refers to their loss value. $\mathcal{L}_i$ refers to the total loss of the batch including regression loss and classification loss and $\mathrm{L}$ refers to the standard threshold (artificial setting).} \label{framework} \end{figure*} \subsubsection{Material Awareness (MA)} Material information is highly reflected in color and texture. In contract to the color information, which can directly represent itself at any position of the picture, texture information can only be recognized and utilized in conjunction with its surroundings. Based on the common knowledge that people identify the material of an object according to both its color and texture, we aggregate these two information as so-called aggregated regional information, which is regarded to represent the material information. To build relations between each position of the concatenated image (the input image $\mathbf{x}$ and its edge image ${E}$ in EG module) and a certain region around it, we utilize $\mathrm{N}_2$ network blocks which is defined as the Module Operation Intensity of MA module. Similar to $\mathrm{N}_1$ in EG, $\mathrm{N}_2$ contains a convolutional layer with the kernel size is $3\times3$, a batch normalization layer, and a ReLU layer, to extract a temporary feature map $\mathbf{B}_{1}$, which is a temperature feature map outputted by the network blocks of MA module. These operations can be formulated as follows: \begin{equation} \mathcal{F}_{\mathbf{m}}(\mathbf{x})=\mathrm{ReLU}\left(\mathbf{W}_{m}\cdot\mathbf{x}+\mathbf{b}_{m}\right), \end{equation} \begin{equation}\label{F:EG} \mathbf{B}_1=\left\{\mathcal{F}_{\mathbf{e}}(\mathbf{x}||{E})\right\}_{\mathrm{N}_2}, \end{equation} where $||$ denotes concatenating operation and $\left\{\cdot\right\}_{\mathrm{N}_2}$ means that the operation is repeated $\mathrm{N}_2$ times. Moreover, we refine $\mathbf{B}_1$ through Region Information Aggregation (RIA) and further generate the refined feature map $\mathbf{B}$ as the final output of MA module. As shown in Fig. \ref{doam_tmm}, RIA operation takes two inputs, specifically the feature map $\mathbf{B}_{1}$ and a parameter $k$, and aggregates them by average pooling and extending operations with a fixed region size of $k\times k$, generating another temperate feature map $\mathbf{B}_{2}^k$. The average pooling and extending operations can be formulated together as follows: \begin{equation}\label{B_2} {\mathbf{B}_{2}}^k_{ij}=\frac{\sum _{m=i-(i\bmod{k})}^{i-(i\bmod{k})+k}\sum _{n=j-(j\bmod{k})}^{j-(j\bmod{k})+k}{\mathbf{B}_{1}}_{mn} }{k^2}, \end{equation} where ``mod'' refers to the modulus operation and ${\mathbf{B}_{2}}^k_{ij}$ refers the feature of the $i$-th row and $j$-th column of feature map $\mathbf{B}_{2}$ when the kernel size for the average pooling layer is $k$. We further concatenate the two feature maps ($\mathbf{B}_{1}$ and $\mathbf{B}_{2}$) in the dimension of channel to generate a new feature map $\mathbf{B}_{3}$ with the dimentions of $2C_r\times{H}\times{W}$, in which every point is able to perceive a certain region around it with size of $k\times k$. In order to be compatible with various region size (due to different values of $k$), the module generates a set of feature maps, $\LARGE{U}$=$\left\{\mathbf{B}_{3}^{k_1},\cdots,\mathbf{B}_{3}^{k_n}\right\}$. Since the scales of prohibited items vary a lot, we design a mechanism to adaptively choose an optimal value of which enables RIA operation to perform well in most situations. We integrate the gated convolutional neural network \cite{yu2019free} $\mathcal{G}$ with $3\times3$ kernels into RIA, to select the proper feature map $\mathbf{B}$ from the set $\LARGE{U}$ as output. The operations can be formulated as follows: \begin{equation}\label{B} \mathbf{B} = \mathcal{G}(U), \end{equation} where $U$=$\left\{\mathbf{B}_{3}^{k_1},\cdots,\mathbf{B}_{3}^{k_n}\right\}$. \subsubsection{Attention Generation} As is illustrated in Algorithm \ref{Alg}, feature map $\mathbf{A}\in \mathbb{R}^{C_e\times{H}\times{W}}$ is outputted by EG module, and $\mathbf{B}\in \mathbb{R}^{C_r\times{H}\times{W}}$ is choosen from a refined feature map set $\LARGE{U}$ in the MA module. Next,in order to fuse the two knids of information, we concatenate the two feature maps genertated by the two core sub-modules and feed them into a convolutional layer with 1$\times$1 kernel size. Therefore, a fused feature map is then generated, $\mathbf{C}\in \mathbb{R}^{(C_e+C_r)\times{H}\times{W}}$, which integrates edge and material information which have been strengthened. These operation can be formulated as follows: \begin{equation}\label{F:mix} \mathbf{C}=\mathbf{W}_{a}\left(\mathbf{A}||\mathbf{B}\right)+\mathbf{b}_{a}, \end{equation} where $||$ represents the operation of concatenating, and $\mathbf{W}_{a}$, $\mathbf{b}_{a}$ are parameters of the convolutional layer. Then we utilize the feature map $\mathbf{C}$ as the input of a sigmoid function to generate the attention map $\mathcal{M}$: \begin{equation}\label{M} \mathcal{M}=\sigma(\mathbf{C}) = \frac{1}{1+e^{-\mathbf{C}}}, \end{equation} where $\mathcal{M}\in\mathbb{R}^{{H}\times{W}}$. Finally, we calculate the inner product of the attention map $\mathcal{M}$ and the concatenated image $\mathbf{P}$ to acquire the final refined feature map $\mathbf{D}$: \begin{equation}\label{D} \mathbf{D}_{j}=\sum_{i=1}^{H\times{W}}\mathcal{M}_{ji}\mathbf{P}_i, \end{equation} where $\mathbf{D}\in\mathbb{R}^{{C}_{h}\times{H}\times{W}}$ and $\mathbf{P}$ refers to the concatenated image of the input image $\mathbf{x}$ and its corresponding edge image ${E}$. The final feature map $\mathbf{D}$ emphasizes the information of the prohibited items and therefore highly contributes to the detection if served to detectors. \begin{figure}[!t] \begin{algorithm}[H] \caption{The Procedure of the over-sampling strategy.} \label{Alg} \begin{algorithmic}[1] \STATE \textbf{Input}: The sample batches $\LARGE{X}_1,\LARGE{X}_2,\dots,\LARGE{X}_\mathcal{N}$, the standard loss threshold $\mathrm{L}$, the hard sample pool $S$ with size $N_S$; \FOR {$i = 0 \to \mathcal{N}$} \STATE Calculate the loss $\mathcal{L}_i$ through Eq. \ref{loss_i}. \IF {$S\,$ isnot$\, full$} \IF {$\mathcal{L}_{i}>\mathrm{L}$} \STATE $S=S \cup X_{i}$. \ENDIF \ELSE \STATE Find the batch ${X}_{m}$ having minimum loss $\mathcal{L}_{m}$ in $S$. \IF {$\mathcal{L}_{i}>\mathcal{L}_{m}$} \STATE $S = \complement_{S}\LARGE{X}_{m} \cup\LARGE{X}_{i}$. \ENDIF \ENDIF \ENDFOR \STATE \textbf{Output}: The hard sample pool $S=\{X_{1},X_{2},\dots,X_{N_S}\}$ \end{algorithmic} \end{algorithm} \end{figure} \subsection{Over-sampling Strategy} Suppose we select $n$ samples from the whole training set for each batch and the size of ``hard sample pool'' $S$ is $N_S$. As illustrated in Fig. \ref{framework}, first, we input the $i$-th epoch training data, $\LARGE{X}_i = \left\{\mathbf{x}_1,\cdots,\mathbf{x}_n\right\}$, into the backbone network to extract features. Second, we calculate the total loss of each batch including regression loss and classification loss. The loss function can be formulated as follows: \begin{equation} \mathcal{L}_i=\frac{\sum_{i=1}^{n}\left(\mathcal{L}_{loc}+\mathcal{L}_{\text {conf }}\right)}{n}. \label{loss_i} \end{equation} Third, we compare $\mathcal{L}_i$ with the standard threshold $\mathrm{L}$ (artificial setting). In the case of that $\mathcal{L}_i$ is smaller than $\mathrm{L}$, the gradient calculation and back propagation are carried out directly and this training iteration is exited. In the case of that $\mathcal{L}_i$ is larger than $\mathrm{L}$, if the number of batches in $S$ is smaller than $N_S$, we regard the images of this batch as ``hard samples'' and put them into $S$. The hard sample selection function $\mathcal{F}_S$ can be regareded as the update process of ``hard samples pool'' $S$. Therefore, in the case of the hard sample pool is not full, the selection function can be formulated as follows: \begin{equation} \mathcal{F}_S: \quad S=\left\{\begin{array}{ll} S \cup X_{i} & , \quad \mathcal{L}_{i}>\mathrm{L} \\ S & , \text { otherwise } \end{array}. \right. \end{equation} If the number of batches in $S$ is larger than $N_S$, we compare $\mathcal{L}_i$ with the minimum loss value of these batch in $S$ already. In the case of that $\mathcal{L}_i$ is larger than the minimum loss value in $S$, we update $S$ by replacing the batch images whose loss is minimum with this batch images. In the case of that $\mathcal{L}_i$ is smaller than the minimum loss value in $S$, this training iteration is exited. Therefore, in the case of the hard sample pool is full, the selection function can be formulated as follows: \begin{equation} \mathcal{F}_S: \quad S=\left\{\begin{array}{ll} \complement_{S}\LARGE{X}_{m} \cup\LARGE{X}_{i} & , \quad \mathcal{L}_{i}>\mathcal{L}_{m} \\ S & , \quad \text { otherwise } \end{array}, \right. \end{equation} where $\LARGE{X}_{m}$ means the images of one batch whose loss is minimum in the ``hard sample pool'' $S$ and $\mathcal{L}_{m}$ refers to their loss value. Finally, after this training iteration is finished, a ``hard sample pool'' $S$ with $N_S$ batches ``hard samples'' is generated. The images in $S$ are retrained to help the model learn more detailed information from ``difficlties''. After these ``hard samples'' are retrained, we clear $S$ and start the next training iteration. By exploiting this over-sampling strategy to retrain these hard samples, the model reduced the performance decrease caused by hard samples in the training set significantly and enhanced the ability of generalization highly, just as that human learns a lot from the difficulties they felt hard to solve. \subsection{Comparison with other attention mechanisms of DOAM} \textbf{Non-local \cite{wang2018non}.} Non-local network captures the long dependence on the feature map by establishing the relationship between each pixel on the feature map and the global pixels, so that the network can better predict the location of the target according to the relationship between the objects in the image. As \cite{wang2018non} introduced, the non-local attention mechanism can be formulated as follows: \begin{equation} \text{Non-local}:\quad y_{i}=\frac{1}{C(x)} \sum_{\forall j} f\left(x_{i}, x_{j}\right) g\left(x_{j}\right), \end{equation} where $j$ refers to all other positions except $i$ in the feature map. However, prohibited items in X-ray images are often small in size, and has little relations with long-distance areas filled of safety items. Too much consideration of the relationship between each position and all other positions on the feature map will distract the attention of the network, causing the performance not satisfactory. Therefore, we narrow down the scope of considering local relations and artificially set several receptive fields and let the gate convolution choose the most appropriate range. This strategy of RIA module of DOAM in Fig. \ref{framework} can be considered as a variant of non-local mechanism. \begin{equation} \text{DOAM}:\quad y_{i}=\frac{1}{k^{2}} \sum_{\forall j \in R_{k}}\left(x_{i}+x_{j}\right), \end{equation} where $R_{k}$ refers to the range of distance $k$ centered on $i$. \textbf{SE \cite{hu2018squeeze}.} SENet focuses on assigning different attention weights to the feature maps of each channel, and then optimizes the channel feature maps which have great influence on the prediction results. This is essentially to make the model pay more attention to the channel feature maps containing key features. As \cite{hu2018squeeze} introduced, this attention mechanism can be formulated as follows: \begin{equation} \text{SE}:\quad \mathbf{s}=\mathbf{F}(\mathbf{z}, \mathbf{W}), \end{equation} where $\mathbf{z}$ refers to the set consisting of all channel feature maps. By analyzing the characteristics of the security image, the edge features of the prohibited items have an important influence on the prediction performance of the model. Therefore, we directly use the edge image (illustrated in Fig. \ref{framework}) to generate the attention map to leading the model to focus on the edge features. In addition, we concatenate the edge image as an individual channel with other feature maps, which can be considered as a variant of SE mechanism that one channel are artificially set to edge image and this mechanism can be formulated as follows: \begin{equation} \text{DOAM}:\quad\mathbf{s}=\mathbf{F}(\mathbf{z} \cup z_e, \mathbf{W}), \end{equation} where $z_e$ refers to the edge image of the original input images in Fig. \ref{framework}. Compared with SE, we integrate prior knowledge into the training process of the model, which can achieve better results for specific problems. \textbf{DA \cite{fu2019dual}.} DA takes into account the relationship between location and channel on the feature map, and adopts the hybrid attention mechanism of space and channel. Similar to the location attention mechanism in non-local, DA construct the relations between each position and all of other position in the whole feature maps. Besides, the channel attention mechanism of DA is similar to SE that modeling long distance dependence between different channels. Inspired by this mechanism, we fuse the features outputted by edge guidance module and material awareness module, and further generate an attention map considering the edge and material characteristics comprehensively. \subsection{Comparison with other congeners of sampling strategy} \textbf{Easy sample over-sampling strategy}. One strategy is simple resampling, which restricts the network to focus more attention on the sample data that is easy to identify. Empirically, this strategy sometimes ensures the location and category of dangerous goods with a higher degree of confidence can be more accurately predicted, and further reduces the false recognition of dangerous goods areas in the security inspection pictures. Similar to the definition above, in the case of the sample pool is not full, the selection function can be formulated as follows: \begin{equation} \mathcal{F}_S:\quad S=\left\{\begin{array}{ll} S \cup X_{i} & , \quad \mathcal{L}_{i}<\mathrm{L} \\ S & , \text { otherwise } \end{array}.\right. \end{equation} Besides, in the case of the sample pool is full, the selection function can be formulated as follows: \begin{equation} \mathcal{F}_S:\quad S=\left\{\begin{array}{ll} \complement_{S}\LARGE{X}_{m} \cup\LARGE{X}_{i} & , \quad \mathcal{L}_{i}<\mathcal{L}_{max} \\ S & , \quad \text { otherwise } \end{array}, \right. \end{equation} where $\mathcal{L}_{max}$ refers to the loss value of the batch whose loss is maximal in $S$ and other variables are the same as the definitions above. However, the usage of this training strategy will miss more difficult sample data, which can not guarantee the principle that the dangerous goods in the security pictures should be checked as much as possible. \textbf{Random sample over-sampling strategy}. Another strategy is enhancing the generalization ability of the model and reduce the biased prediction of the model by introducing random sample data. The selection function can be formulated as follows: \begin{equation} \mathcal{F}_S:\quad S=\left\{X_{1}, X_{2},\dots, X_{N_S}\},\right. \end{equation} where $X_{1}, X_{2},\dots, X_{N_S}$ are the batches randomly selected from the training data. However, random sample over-sampling strategy has no clear purpose and suffers more uncertainty compared to the hard sample resampling method, thereby it is also not a satisfactory training strategy. \textbf{Focal Loss \cite{lin2017focal}}. The Focal Loss strategy uses weighted cross-entropy loss and adaptively enhance the learning ability of hard samples. The loss function of this method is as follows: \begin{equation} \mathcal{L}\left(p_{\mathrm{t}}\right)=-\left(1-p_{\mathrm{t}}\right)^{\gamma} \log \left(p_{\mathrm{t}}\right), \end{equation} where $p_{\mathrm{t}} \in \left[0,1\right]$ and ${\gamma} \ge 0$. Note that this learning method mainly focuses on the hard samples in the training data, and its essence is the same as our hard sample resampling method. However, due to the existence of some outliers in the training data, paying more attention to these extremely hard samples will cause the model to ignore the correct optimization direction to fit these samples. Fortunately, the hard sample resampling method gives similar weights to the data in the sample pool, which will not lead to over learning of outliers. \section{Experiments}\label{Section:exp} In this section, extensive experiments are conducted to evaluate the model we proposed. In our work, the primary goal is to detect occluded prohibited items in X-ray images in security inspection scenario. As far as we know, no dataset focusing on this task has been proposed up to now, so we only employ the OPIXray dataset for evaluation all through the following experiments. \textbf{First}, we demonstrate that DOAM-O outperforms all the attention mechanisms mentioned above, over different categories and distinctive occlusion levels. \textbf{Second}, we perform ablation experiments to completely evaluate the effectiveness of DOAM-O. \textbf{Third}, we verify the overall compatibility and effectiveness of DOAM-O after integrating with various object detection networks. \textbf{Finally}, we visualize the generated attention distribution map and the whole network to give a more intuitive verification. \textbf{Evaluation strategy:} All experiments are proceeded on the OPIXray dataset and trained on training set in Tab. \ref{data_table}. In most experiments, models are tested by the testing set data in Tab. \ref{data_table}. As for comparing different attention mechanisms in terms of occlusion levels, models are tested on OL1, OL2 and OL3 in Tab. \ref{data_cate_table} respectively. \textbf{Baseline Detail:} To make fair comparisons between different attention mechanisms, we plug DOAM and each of the other attention modules into SSD respectively, and evaluate the performance of these integrated networks as well as SSD itself. These attention modules are added to the backbone (VGG16) of SSD. More explicitly, they are embedded behind the max pooling layer where the feature map is refined to half. In ablation study, we plug each sub-module of DOAM into SSD \cite{liu2016ssd} one by one and report the performance of SSD \cite{liu2016ssd} and report the results under different conditions to evaluate the utility of various sub-modules. Lastly, we evaluate the compatibility of DOAM by integrating it into several mainstream detection networks, including YOLOv3, FCOS, as well as SSD. \textbf{Parameter setting:} All through the experiments, models are optimized using SGD optimizer with initial learning rate at 0.0001. The batch size is set to 24 and the momentum and weight decay are set to 0.9 and 0.0005 respectively. We utilize the mean Average Precision (mAP) as the metric to evaluate the model performance and the IOU threshold is set to 0.5. We calculate the AP of each category using the model with the best performance to observe the improvement in different categories. Moreover, to avoid the bias from data transformation while generating edge image, no data augmentation technique are used to modify the pixel value of the original image, which leads to a better analysis of the impact of edge information. \subsection{Comparing with Different Attention Mechanisms} We compare three variants of attention mechanisms, including SE \cite{hu2018squeeze}, Non-local \cite{wang2018non} and DA \cite{fu2019dual}. Tab. \ref{table_attention_catagories} and \ref{Occlusion-result} reports the performances of all models. Note that ``FO'', ``ST'', ``SC'', ``UT'' and ``MU'' refer to ``Folding Knife'', ``Straight Knife'', ``Scissor'', ``Utility Knife'' and ``Multi-tool Knife'', respectively. Besides, ``DOAM-T'' refers to the model ``SSD+DOAM'' with traditioanl training strategy and ``DOAM-O'' refers to the model ``SSD+DOAM'' with over-sampling training strategy. \subsubsection{Object Categories} As is shown in Tab. \ref{table_attention_catagories}, DOAM not only improves the performance of SSD by $3.12\%$, but also outstrips SE, Non-local, and DA methods by $2.16\%$, $2.60\%$, $2.05\%$, respectively. Especially in the category of Straight Knife, Folding Knife and Utility Knife, DOAM achieves a large improvemnt. Particularly for Straight Knife, which is the category with the worst occlusion, DOAM sees a remarkable amount over SSD \cite{liu2016ssd} baseline by $6.48\%$ and Non-local \cite{wang2018non} by $5.12\%$. The performance of DOAM in Scissor, the category with the lightest occlusion, is merely improved by $1.71\%$ compared with SSD \cite{liu2016ssd} and similar to Non-local \cite{wang2018non}. Obviously, DOAM surpasses these current widely-used attention mechanisms over different categories in the field of prohibited item detection based on X-ray images. \begin{table}[h] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{5pt} \caption{Performance comparison between DOAM-O and other different attention mechanisms on object categories.} \label{Classification-result} \small \begin{tabular}{lcccccc} \hline \multicolumn{1}{l}{\multirow{2}{*}{Method}} & \multicolumn{1}{c}{\multirow{2}{*}{mAP}} & \multicolumn{5}{c}{Categories} \\ \cline{3-7} \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{FO} & \multicolumn{1}{c}{ST} & \multicolumn{1}{c}{SC} & \multicolumn{1}{c}{UT} & \multicolumn{1}{c}{MU} \\ \hline SSD \cite{liu2016ssd} & 70.89 & 76.91 & 35.02 & 93.41 & 65.87& 83.27\\ \hline +SE \cite{hu2018squeeze} & 71.85 & 77.17 & 38.29 & 92.03 & 66.10 & \textbf{85.67} \\ +Non-local \cite{wang2018non} & 71.41 & 77.55 & 36.38 & 95.26 & 64.86 & 82.98 \\ +DA \cite{fu2019dual} & 71.96 & 79.68 & 37.69 & 93.38 & 64.14 & 84.90\\ +DOAM-T &74.01&81.37&41.50&95.12&68.21&83.83\\ \textbf{+DOAM-O} &\textbf{74.57} & \textbf{81.67} & \textbf{41.90} & \textbf{95.37} & \textbf{68.41} & 84.23\\ \hline \end{tabular}\label{table_attention_catagories} \end{table} \subsubsection{Object Occlusion Levels} Tab. \ref{level-table} shows different attention mechanisms with SSD on each object occlusion level. And Fig. \ref{improve} is drawning from Tab. \ref{level-table}, which illustrates the improvement of DOAM compared with other methods in terms of different occlusion levels. In other words, with the increasing of occlusion level, the model gets improved more obviously. It implies that DOAM is able to tackle the occlusion issue especially in heavily occluded cases. (Note that in OL3, the performance of ``SSD+Non-local" is lower than ``SSD". We conjecture the decrease is closely related to the attention mechanism of Non-local, which captures spatial information by building the relations between regions. Unfortunately, this type of relation reduces effect with the increasingly disorder of image.) \begin{table}[h] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{12pt} \caption{Performance comparison between DOAM and other different attention mechanisms on object occlusion levels.} \label{Occlusion-result} \small \begin{tabular}{lccc} \hline Method & OL 1 & OL 2 & OL 3\\ \hline SSD \cite{liu2016ssd} & 75.45 & 69.54 & 66.30\\ \hline +SE \cite{hu2018squeeze} & 76.02 & 70.11 & 67.53\\ +Non-local \cite{wang2018non} & 75.99 & 70.17 & 65.87\\ +DA \cite{fu2019dual} & 77.41 & 69.68 & 66.93\\ \textbf{+DOAM} & \textbf{77.87} & \textbf{72.45} & \textbf{70.78}\\ \hline \end{tabular} \label{level-table} \end{table} \begin{figure}[!t] \centering \includegraphics[width=0.9\linewidth]{images/improvement.pdf} \caption{The amount changes of performance improvement of DOAM over different models with occlusion level increasing.} \label{improve} \end{figure} \subsection{Comparing with Different Detection Approaches} To further evaluate the effectiveness and applicability of DOAM, we conduct experiments on three popular detection approaches, SSD \cite{liu2016ssd}, YOLOv3 \cite{redmon2018yolov3} and FCOS \cite{tian2019fcos}. The results are shown in Tab. \ref{detection_approach}. \begin{table}[!h] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{5pt} \caption{Performance comparison between DOAM-integrated network and baselines for three famous detection approaches.} \small \begin{tabular}{lcccccc} \hline \multicolumn{1}{l}{\multirow{2}{*}{Method}} & \multicolumn{1}{c}{\multirow{2}{*}{mAP}} & \multicolumn{5}{c}{Category} \\ \cline{3-7} \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{FO} & \multicolumn{1}{c}{ST} & \multicolumn{1}{c}{SC} & \multicolumn{1}{c}{UT} & \multicolumn{1}{c}{MU} \\ \hline SSD \cite{liu2016ssd} & 70.89 & 76.91 & 35.02 & 93.41 & 65.87& 83.27 \\ +DOAM-T & 74.01 & \textbf{81.37} & 41.50 & 95.12 & 68.21 & 83.83 \\ \textbf{+DOAM-O} &\textbf{74.57} & \textbf{81.67} & \textbf{41.90} & \textbf{95.37} & \textbf{68.41} & \textbf{84.23}\\ \hline YOLOv3 \cite{redmon2018yolov3} & 78.21 & \textbf{92.53} & 36.02 & \textbf{97.34} & 70.81 & 94.37 \\ +DOAM-T & 79.25 & 90.23 & 41.73 & 96.96 & 72.12 & 95.23\\ \textbf{+DOAM-O} & \textbf{80.40} & \textbf{92.00} & \textbf{43.57} & \textbf{97.70} & \textbf{73.05} & \textbf{95.65}\\ \hline FCOS \cite{tian2019fcos} & 82.02 & 86.41 & 68.47 & 90.22 & 78.39 & 86.60\\ +DOAM-T & 82.41 & 86.71 & 68.58 & \textbf{90.23} & 78.84 & 87.67\\ \textbf{+DOAM-O} & \textbf{83.80} & \textbf{87.61} & \textbf{72.73} & 90.03 & \textbf{80.77} & \textbf{87.80}\\ \hline \end{tabular}\label{detection_approach} \end{table} As shown in Tab. \ref{detection_approach}, the performance of DOAM-integrated networks are promoted by 3.12\%, 1.04\% and 0.39\% in comparison with SSD \cite{liu2016ssd}, YOLOv3 \cite{redmon2018yolov3} and FCOS \cite{tian2019fcos} respectively, which means that our module can be utilized as a plug-and-play module into most detection networks. Note that the performances on Folding Knife and Scissor are slightly diminished after DOAM-integrated. We speculate it is because our attention mechanism focuses more on the parts occludede heavily, so it is better at dealing with high-level occlusions. While samples in these two categories above are slightly occluded, which may lead to the little fall in performance for Folding Knife and Scissor. \subsection{Comparing with Other Over-sampling Mechanisms} In this section, to verify the effectiveness of our over-sampling strategy, we compare three variants of retraining strategies, including easy-sample retraining strategy (+Easy), random-sample retraining strategy (+Random) and Focal Loss strategy (+Focal). Tab. \ref{oversamplingstr} reports the performances of all models. Note that in Tab. \ref{oversamplingstr}, the four training strategies are implemented in the base model (SSD+DOAM) for fair comparison. Besides, ``DOAM-T'' refers to the model ``SSD+DOAM'' with traditioanl training strategy and ``DOAM-O'' refers to the model ``SSD+DOAM'' with the over-sampling training strategy. \begin{table}[h] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{5pt} \caption{Performance comparison between the over-sampling strategy and other congeneric mechanisms on object categories.$^*$} \label{Classification-result} \small \begin{threeparttable} \begin{tabular}{lcccccc} \hline \multicolumn{1}{l}{\multirow{2}{*}{Method}} & \multicolumn{1}{c}{\multirow{2}{*}{mAP}} & \multicolumn{5}{c}{Categories} \\ \cline{3-7} \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{FO} & \multicolumn{1}{c}{ST} & \multicolumn{1}{c}{SC} & \multicolumn{1}{c}{UT} & \multicolumn{1}{c}{MU} \\ \hline SSD \cite{liu2016ssd} & 70.89 & 76.91 & 35.02 & 93.41 & 65.87& 83.27\\ +DOAM-T &74.01&81.37&41.50&95.12&68.21&83.83\\ \hline +DOAM-E & 73.22 & 81.36 & 36.67 & 92.38 & 71.21 & \textbf{84.47} \\ +DOAM-R & 73.42 & 81.28 & 35.57 & 95.21 & 71.89 & 83.12 \\ +DOAM-F \cite{fu2019dual} & 73.25 & 80.69 & 34.59 & 95.25 & \textbf{73.34} & 82.40\\ \textbf{+DOAM-O} &\textbf{74.57} & \textbf{81.67} & \textbf{41.90} & \textbf{95.37} & 68.41 & 84.23\\ \hline \end{tabular}\label{oversamplingstr} \begin{tablenotes} \footnotesize \item $^*$Note that ``E'' refers to the easy sample over-sampling training strategy, ``R'' refers to the random sample over-sampling training strategy and ``F'' refers to the focal loss strategy. \end{tablenotes} \end{threeparttable} \end{table} \subsection{Module Complexity Analysis} In this section, we analyze the model complexity including the total number of parameters, model size and computation cost of different attention mechanisms in the case of that the base model is SSD. The three attention mechanisms that we make comparison with DOAM are SE \cite{hu2018squeeze}, Non-local \cite{wang2018non} and DA \cite{fu2019dual}, which focus on channel information, spatial information, and combination of the two kinds of information above respectively. All of them lead to quite a bit of extra parameter size and model size, while DOAM only brings negligible addition (about 0.41\% and 0.11\% in parameter size and model size) to SSD without attention mechanisms. As for computational cost, our DOAM slightly brings extra computational cost compared with SSD without any attention mechanisms (about 7.14\% in GFLOPs), meanwhile SE, Non-local, and DA bring 3.15\%, 6.54\% and 23.07\% respectively. \begin{table}[h] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{8pt} \caption{Complexity comparison of different models.$^*$} \label{param-result} \small \begin{threeparttable} \begin{tabular}{lccc} \hline Method & PARAMs & SIZE(MB) & GFLOPs\\ \hline SSD \cite{liu2016ssd} & $24.2\times10^6$ & 92.6 & 30.6522\\ \hline +SE \cite{hu2018squeeze} & $32.0\times10^6$ & 122.1 & \textbf{31.6169}\\ +Non-local \cite{wang2018non} & $30.9\times10^6$ & 117.9 & 32.6577\\ +DA \cite{fu2019dual} & $45.6\times10^6$ & 174.1 & 37.7231\\ \textbf{+DOAM} & $\bm{24.3\times10^6}$ & \textbf{92.7} & 32.8435\\ \hline \end{tabular} \label{param-table} \begin{tablenotes} \footnotesize \item $^*$Note that ``PARAMs'', ``SIZE'' and ``GFLOPs'' refer to ``the total number of parameters'', ``the Model Size'' and ``the Giga Floating Point operations'', respectively. \end{tablenotes} \end{threeparttable} \end{table} DOAM prevails the other three attention mechanisms in parameter size and model size, while is slightly more expensive in computational cost. It might result from different values of parameter which causes repetitive computation in RIA. \subsection{Ablation Study}\label{sec:ablation-study} In this section, we report the results of ablation studies in Tab. \ref{Ablation-result} to verify the effectiveness of each part of our model. Note that in Tab. \ref{Ablation-result}, "C" refers simply concatenating the edge image and original RGB image. "DOAM/MA" refers DOAM without Material Awareness module and "DOAM/$\mathcal{G}$" refers DOAM without the Gate Convolutional Neural Network. "DOAM" refers to the DOAM with traditional training strategy and ``DOAM-O'' refers to the model we proposed in this paper. \begin{table}[!h] \centering \renewcommand\arraystretch{1.3} \setlength{\tabcolsep}{5pt} \caption{Ablation studies of Over-sampling De-occlusion Attention Network.$^*$} \small \begin{threeparttable} \begin{tabular}{lcccccc} \hline \multicolumn{1}{l}{\multirow{2}{*}{Method}} & \multicolumn{1}{c}{\multirow{2}{*}{mAP}} & \multicolumn{5}{c}{Category} \\ \cline{3-7} \multicolumn{1}{l}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{FO} & \multicolumn{1}{c}{ST} & \multicolumn{1}{c}{SC} & \multicolumn{1}{c}{UT} & \multicolumn{1}{c}{MU} \\ \hline SSD \cite{liu2016ssd} & 70.89 & 76.91 & 35.02 & 93.41 & 65.87& 83.27 \\ \hline +Concate &72.32 & 79.00 & 36.46 & 94.13 & 68.85 & 83.18 \\ +EG & 72.75 & 80.26 & 35.54 & 94.81 & 67.96 & \textbf{85.19} \\ +EG+MA/$\mathcal{G}$ & 73.12 & 79.94 & 38.58 & 93.39 & 69.40 & 84.28 \\ +EG+MA & 74.01 & \textbf{81.37} & 41.50 & 95.12 & 68.21 & 83.83 \\ \textbf{+EG+MA+O} & \textbf{74.27} & 81.06 & \textbf{42.45} & \textbf{95.37} & \textbf{70.31} & 82.16 \\ \hline \end{tabular}\label{Ablation-result} \begin{tablenotes} \footnotesize \item $^*$Note that ``Concate'' refers to simply concating the edge image and the corresponding original image, ``MA/$\mathcal{G}$'' refers to the Material Awareness module without the Gated convolutional network and ``+EG+MA+O" refers to the model ``DOAM-O''. \end{tablenotes} \end{threeparttable} \end{table} Tab. \ref{Ablation-result} shows that EG improved the performance by $0.43\%$, compared with the strategy of simply concatenating the input image and the corresponding edge image without any other operations of EG. We speculate that it is mainly because the EG gives specific emphasis to edge information via optimizing the loss function, which contributes to the ability to focus more on prohibited items. While simply concatenating operates all the objects for feature fusion in the image equally, even the object is not what we desire to detect. Meanwhile, model integrating both EG and MA performs better than integrating EG alone by $0.37\%$, which unequivocally proves the effectiveness of MA. Note that the size of prohibited item is about 10$\times$10 in average, so we pick 10$\times$10 as the region scale for each position of the feature map. \begin{figure*}[!t] \centering \includegraphics[width=\linewidth]{images/HOTMAP.png} \caption{The whole network with Grad-CAM visualization results. We compare the visualization results of the DOAM-integrated network (VGG16+DOAM) with baseline (VGG16). The grad-CAM visualization is calculated for the last convolutional outputs. The ground-truth label is shown on the top of each input image.} \label{hotmap} \end{figure*} We choose three different scales of the regions, ($5\times5$, $10\times10$, $15\times15$ respectively), and introduce the gated convolutional neural organization \cite{yu2019free} $\mathcal{G}$ into MA to adaptively select the best feature map, which is generated by average pooling operation with proper pooling size. As a result, the performance improves by $0.9\%$ after drawing $\mathcal{G}$. \subsection{Visualization Experiment} \begin{figure}[!b] \centering \includegraphics[width=\linewidth]{images/att_vis.pdf} \caption{Attention distribution map visualization.} \label{att_vis} \end{figure} \subsubsection{Attention Distribution Map Visualization} In this section, we visualize the attention map in Fig. \ref{att_vis}, which generated by DOAM to have an intuitive observation of how DOAM works. We showcase 10 input Xray images (two for each class) in rows 1 and 3, as well as the corresponding attention visualizations in rows 2 and 4. It is notable that edge and region information are captured by DOAM accurately. For example, in column 4, a red box is marked on a utility knife in the X-ray image (in row 1), and the boundaries of the utility knife are clearly portrayed in the attention visualization (in row 2). Moreover, in the first column, a red box is marked on a folding knife and the corresponding attention map (in row 2) highlights most of the areas where the folding knife lies on. To sum up, visualizations further demonstrate that our proposed model is effective in collecting edge and region information to improve feature representation in occluded prohibited items detection. \subsubsection{The Whole Network Visualization with Grad-CAM} In this section, we apply the Grad-CAM \cite{selvaraju2017grad} to the two networks, VGG16 network \cite{simonyan2014very} and DOAM-integrated network (VGG16+DOAM), and using images from the OPIXray dataset. As we can clearly observe from Fig. \ref{hotmap} that the Grad-CAM masks of DOAM-integrated network cover the target object regions more widely than the single VGG16 network \cite{simonyan2014very}, which verifies the effectiveness of Material Awareness (MA) for utilizing the RIA to aggregate the region information. And also the Grad-CAM masks circle the shape and cover the target more evenly, due to the emphasized edge information by Edge Guidance (EG). From these observations, we verify that the feature refinement process of our model eventually leads the networks to perform better. % % % % % \section{Conclusion}\label{Section:conclusion} In this paper, we contribute the first high-quality dataset named OPIXray, which aims at occluded prohibited items detection in security inspection. The images of the dataset are gathered from an airport and these prohibited items are annotated manually by professional inspectors to simulate the real security inspection scene to the greatest extent. Besides, we introduce abundant analysis and discussions for the OPIXray dataset, including construction principles, quality control procedures and potential tasks. To address the occlusion problem in noisy X-ray images, we propose an over-sampling de-occlusion attention network (DOAM-O). To the best of our knowledge, this is the first work that explicitly exploits attention mechanisms and over-sample technologies to prohibited items detection, which might provide a new way to solve the occlusion problem under noisy data scenarios. We hope that our contributions can be beneficial to promote the development of prohibited items detection in noisy X-ray images. \section{Acknowledgement} This work was supported by National Natural Science Foundation of China (62022009, 61872021), Beijing Nova Program of Science and Technology (Z191100001119050), and State Key Lab of Software Development Environment (SKLSDE-2020ZX-06). % \begin{comment} \appendices \section{Proof of the First Zonklar Equation} Appendix one text goes here. \section{} Appendix two text goes here. \section*{Acknowledgment} This work was supported by National Natural Science Foundation of China (61872021), Beijing Nova Program of Science and Technology (Z191100001119050), State Key Lab of Software Development Environment (SKLSDE-2020ZX-06) and Fundamental Research Funds for Central Universities (YWF-20-BJ-J-646). \begin{IEEEbiography}{Michael Shell} Biography text here. \end{IEEEbiography} \begin{IEEEbiographynophoto}{John Doe} Biography text here. \end{IEEEbiographynophoto} \begin{IEEEbiographynophoto}{Jane Doe} Biography text here. \end{IEEEbiographynophoto} \end{comment} \ifCLASSOPTIONcaptionsoff \newpage \fi \bibliographystyle{IEEEtran}

\section{Introduction} Text-to-Image synthesis has become an important research topic in the area of artificial intelligence \cite{pan2019gansurvey}. The introduction of Generative Adversarial networks (GANs) by Goodfellow et al. \cite{goodfellow2014gan} in 2014 has enabled rapid improvement in state-of-the-art methods on text-to-image synthesis. Built upon GANs, the Stacked Generative Adversarial Network (StackGAN) proposed by Zhang et al. in 2016 \cite{zhang2016stackgan} has enhanced the stability of GAN’s training process and the output resolution of the samples generated. The authors of the StackGAN have proposed the improved StackGAN-v2 \cite{zhang2018stackgan++} in the following year, further improving the training model with a series of multi-scale image distributions. In most existing image synthesis research works, the training images used are of the best quality with a clear, high resolution photo of a target, which might be hard to collect and produce in practical scenarios. In most of the use cases, the images used for machine learning may have deformation. Noise distortion can arise with multiple objects in the image and watermark added. The issue of motion blur can also occur due to a moving camera or moving target objects. In certain scenarios, the image may not fully capture the target objects but is represented as a cropped part of target. Often, the performance of the generative models is susceptible to small image transformations \cite{samuel2016imagequality} in contrast to robust human visual systems \cite{samuel2017humancomparison}. Therefore, in this paper, we focus on robustness evaluation of StackGAN-v2 specifically on image transformation, by introducing foreign objects into the training image dataset and investigate the quality of the image generated from the altered images. These images are generated based on several metamorphic relations that we defined in this study and Inception Score (IS) will be used to assess and compare the results. With these in mind, the following research questions need to be addressed: \begin{itemize} \item How can metamorphic relations between the characteristics of the model and its robustness be identified? \item How do variations in training data affect the IS of the synthesised images? \end{itemize} This paper aims to develop a means of generating metamorphic relations to aid the evaluation and interpretation of image synthesis models. Specifically, the objectives of this paper includes: \begin{itemize} \item To propose metamorphic relations that are capable of examining the relations between the characteristics of the StackGAN-v2 model and its robustness; \item To identify the effects of variations in training data on the generated model. \end{itemize} We use Inception Score (IS) \cite{barratt2018note} as the metric to evaluate the quality of the images generated from StackGAN-v2 model. IS is a widely used metric to evaluate the quality of an image generative model. \[ \mathrm{IS}(G) = \exp (\mathbb{E}_{\mathrm{x}\sim p_{g}} D_{KL}(p(y|\mathrm{x})\|p(y))) \] where \(\mathrm{x}\sim p_{g}\) denotes a generated sample \(\mathrm{x}\) from distribution of images \(p_{g}\), and \(y\) is the label classified by inception model. \(D_{KL}(p(y|\mathrm{x})\|p(y))\) indicates $KL$ divergence between the conditional class distribution \(p(y|\mathrm{x})\) and the marginal class distribution \(p(y)\). Therefore, if a generative model is capable of generating a high diversity of images, large $KL$ divergence is expected, leading to a high IS. While IS is able to evaluate the model quantitatively, it is difficult to capture the correctness of the generated images with respect to the intended output from a given input text, given the black-box nature of the StackGAN-v2 model. As such, we propose a metamorphic testing approach to better understand the synthesized image, given a set of input data. Instead of identifying every metamorphic relation (MR) prior to conducting the experiments, we employ investigative steps to generate MR based on the analysis of test case results from the previous metamorphic relation consecutively, to verify and emphasise the findings from the previous test case. Using our motivating example that it is extremely difficult to collectivise perfect training images without any noise, we first propose a metamorphic relation to evaluate the robustness of StackGAN-v2 algorithm when there are presences of a foreign object in the training data. As experiments are conducted, the unknown behaviour of StackGAN-v2 is revealed, from which new metamorphic relations are further identified. Details of the experiments on these metamorphic relations are given in Section \ref{experiments}. The main discovery and contribution of this study are that metamorphic testing works effectively in an iterative process model. While most of the metamorphic testing project had defined the relations prior based on the expected outcome, like the testing done in the work by \cite{wang2019metamorphic, zhou2019driverlesscar, zhang2018metamorphicautonomousdriving}, This paper shows that metamorphic testing can follow an iterative or cyclic process flow, similarly to how we have conducted our study as summarised in Figure \ref{fig:methodology_flowchart}. \begin{figure}[t] \includegraphics[width=0.25\textwidth]{images/methodology_flowchart_new_black.PNG} \centering \caption{Overall workflow for this study.} \label{fig:methodology_flowchart} \end{figure} This is a significant finding in the field of software testing, especially in applications where the algorithm is too complex for researchers to predict the outcome of a given input. \section{Background} \subsection{Stacked Generative Adversarial Networks} Stacked Generative Adversarial Network (StackGAN) \cite{zhang2016stackgan} is a framework designed based on Generative Adversarial Network (GAN) \cite{goodfellow2014gan} to synthesise high-quality images with photo-realistic details reflecting the context of the given text description. StackGAN is designed as a two-stage GAN to address the instability issue of GANs’ training process upon high resolution image generation. Stage-I GAN first generates low resolution images by sketching the primitive shape and basic colours of the target described in the given text input. Then, Stage-II GAN improves the semantics of the Stage-I result images by conditioning with the text input again to capture the omitted details. An improved version of StackGAN, called the StackGAN-v2 \cite{zhang2018stackgan++}, was proposed to further refine and stabilise the earlier model architecture by arranging multi-stages of GAN in a tree-like structure. The root of the tree represents the input dataset, and the deepest level of the branches provides the final output of generated samples, where the most semantic and photo-realistic images are drawn. The intermediate GANs gradually improve resolution of the sample to be generated at each branch level. In this manner, the generative model of GAN at each branch learns to capture image distributions, and the feedback from one model can improve the learning process of other generative models at different branches. The discriminative models take the original datasets and samples from generative models with respective conditioning variables to train for classification. \subsection{Metamorphic Testing} Metamorphic testing \cite{10.1145/3143561} is an approach widely used for testing machine learning systems to handle the oracle problem as test oracle may not always exist \cite{wang2019metamorphic}. In metamorphic testing, new test cases are generated based on the set of input-output pairs from the previous test cases to better understand the relation among the series of inputs and their corresponding outputs, with what is known as metamorphic relation (MR). The implementation of metamorphic testing can be broken down into three steps \cite{segura2016surveyonmt, Zhou2019Meta}. \begin{enumerate} \item Identify the characteristics of the program under test and represent them in the form of relations among a set of test inputs and their corresponding expected outputs. \item Generate source test cases using traditional test case selection techniques. The source test cases are used to serve as seeds to generate the follow-up test cases using identified metamorphic relation. \item Execute the test cases to verify whether the outputs of the source test case and follow-up test cases have satisfied or violated the metamorphic relation they were built on. \end{enumerate} In this paper, we apply metamorphic testing to show how the unexpected image dataset can impact the performance of the StackGAN-v2 model. \section{Related Work} Most of the GAN models are susceptible to errors caused by the discrepancies in training data. Not only is there a lack of test oracles to ensure efficiency and correctness of machine learning models, but there is also a lack of metrics to evaluate model correctness and image quality. Metamorphic testing \cite{zhou2019driverlesscar} is chosen in this study to augment existing testing strategies and to identify unknown limitations of the current StackGAN-v2 model. Metamorphic testing has shown tremendous success in evaluating the robustness of self-driving or autonomous driving algorithms because the testing of these algorithms lacks diversity in test cases. This is largely due to the complexity involved in generating driving scenes which are known to impair driving abilities \cite{zhang2018metamorphicautonomousdriving}. The problem was resolved with the employment of a GAN-based technique to synthesise driving scenes with varying weather conditions, and the expected steering angle signals were then predicted based on the metamorphic relations \cite{zhang2018metamorphicautonomousdriving}. Metamorphic testing was used to detect violations of the metamorphic relations through prompt analysis of the outputs to evaluate robustness of the autonomous driving system. Furthermore, existing studies found that testing on object detection algorithms experienced poor test coverage due to lack of diversity and realness in test inputs \cite{wang2019metamorphic}. MetaOD, a combination of metamorphic testing with image processing to add different objects into background images was introduced to generate modified test inputs that look realistic \cite{wang2019metamorphic}. MetaOD does so by checking the consistency of object detection results between synthetic images and the corresponding backgrounds by employing metamorphic relations and evaluating the extent to which they are upheld \cite{wang2019metamorphic}. However, the success of metamorphic testing heavily relies on the quality of metamorphic relations, which are not straightforward and difficult to identify. Metamorphic relation pattern (MRP) is one of the methods that can be utilised to identify metamorphic relations \cite{Zhou2020MetamorphicRF}. Examples of metamorphic relation patterns include input equivalence and shuffling \cite{Segura2019MetamorphicRP}. Apart from this, existing studies have also utilised symmetry as a metamorphic relation pattern which assumes that the system would appear the same under different viewpoints \cite{Zhou2020MetamorphicRF}. This study is motivated by the ability of metamorphic testing to exploit metamorphic relations in evaluating the robustness of an algorithm \cite{wang2019metamorphic} \cite{zhou2019driverlesscar} \cite{zhang2018metamorphicautonomousdriving}. Thus, in the context of our work, the first step is to identify the potential metamorphic relations between the training dataset and the photo-realistic images generated by the model. Test data is then altered and manipulated correspond to the metamorphic relations, in order to identify the unknown limitations or faults of the StackGAN-v2 model, that are otherwise hidden from the researchers and users. Ultimately, the findings of our research can be used to evaluate the robustness of StackGAN algorithm. \section{Methods} \label{methods} We setup an experimental test bed using \textit{GeForce RTX 2080 Ti} with \textit{CUDA 10.0} to allow the execution of the study objectives. Further details of our implementation can be found at our GitHub page\footnote{https://github.com/FIT4003StackGAN/FIT4003-StackGAN-v2-Metamorphic-Testing}. In general, we employ investigative steps to identify the metamorphic relations in the StackGAN-v2 model by manipulating the training image dataset. This is necessary and extremely crucial to our study as we cannot predict the output of the model when given a set of unexpected inputs, and thus are unable to identify more metamorphic relations without performing some form of initial analysis. We introduce noise to the entirety of the training image dataset by adding foreign object bodies that tend to be absent from the original dataset, such as an addition of bird or trees. The modified images are inspected to ensure minimal obstruction of the focal objects, which in this study are birds. Training would then be performed on the modified images, which requires approximately 33 hours to run on 600 epochs (as per recommended in StackGAN-v2). Following this, we generate the images for the model which are then analysed using two different evaluation metrics, namely the IS and the Likert Scale \cite{joshi2015likert}. Finally, we decipher the results of the test case, and from there, design and identify new metamorphic relations and the cycle continues. The detailed implementation of our experiments are discussed in Section \ref{experiments}. \section{Experiments} \label{experiments} \subsection{Data Collection} We evaluate StackGAN-v2 on the CUB \cite{WahCUB_200_2011} datasets which contain 200 bird species with 11,788 images, as was utilised in the original StackGAN-v2 paper. We utilised the image list for class-disjoint training and test sets as specified by the original StackGAN-v2 paper without any modifications, to reproduce the results as closely as possible. This is done so that the results from the unmodified datasets may act as the ground truth for our experiments. \subsection{Evaluation Metrics} It is important that we apply the same metrics used to get an unbiased comparison, and so we utilise the IS model fine-tuned by StackGAN-v2 as our quantitative evaluation metric. The higher the IS for a set of images, the better the quality and diversity of the images. All reported ISs in this study will be utilising the best result of each test case. \begin{table}[t] \caption{Comparison of ISs and Likert Scale Evaluation between Test Cases} \label{table:1} \tiny \centering \sisetup{ table-number-alignment = center, table-figures-integer = 1 } \begin{tabular}{ |S| *{3}{@{\hspace{0.5\tabcolsep}}S[ separate-uncertainty, table-figures-uncertainty = 1 ]@{\hspace{0.5\tabcolsep}}|} } \hline {Test Cases} & {IS} & {Semantic Likert Scale} & {Realistic Likert Scale} \\ \hline {StackGAN-v2 paper} & 4.05(5) & {-} & {-} \\ {Pre-Trained} & 4.08(5) & {-} & {-} \\ {Self-Trained} & 4.16(3) & 2.37(88) & 2.59(111) \\ {\textbf{TC\textsubscript{01}}} & 3.50(4) & 1.60(85) & 1.73(100) \\ {\textbf{TC\textsubscript{02}}} & 3.84(4) & 1.48(62) & 1.76(99) \\ {\textbf{TC\textsubscript{03}}} & 3.99(5) & 1.75(88) & 1.84(110) \\ {\textbf{TC\textsubscript{04}}} & 3.96(6) & 1.18(38) & 1.15(42) \\ {\textbf{TC\textsubscript{05}}} & 3.92(4) & 1.98(80) & 2.25(101) \\ {\textbf{TC\textsubscript{06}}} & 4.06(2) & 1.89(97) & 1.90(98) \\ {\textbf{TC\textsubscript{07}}} & 3.88(5) & 1.91(91) & 1.77(107) \\ {\textbf{TC\textsubscript{08}}} & 3.83(3) & 1.68(70) & 1.68(92) \\ \hline \end{tabular} \end{table} While studies have shown that the IS correlates well with human judgement, it is our concern that the usage of a fine-tuned model provided by StackGAN-v2 may introduce some form of partiality in our experiments. Thus, we perform an additional analysis on the generated images with the highest IS for each test case. The set of highest quality (IS) images would be evaluated using the Likert Scale -- a form of evaluation metric to quantify human thoughts and perception \cite{joshi2015likert}. This quantitative analysis is conducted amongst the co-authors, using a five-point scale with 1 representing very poor, and 5 representing excellent. For all test cases, we compare the generated images with the training image datasets in terms of semantics and photo-realism. A total of 40 generated images are chosen for the Likert Scale evaluation of the White breasted Kingfisher species. The compilation of the details of the Likert Scale evaluation will be elaborated towards the end of the Section \ref{results}. Finally, much of this study involves qualitative analysis, where we manually review the images by eye and noting interesting details for each result. Explanations of our observations are elaborated in Section \ref{results}. \subsection{Preliminaries} \label{preliminaries} Metamorphic relations are discovered and revised as we progress through our experiments and reveal more behaviour of the StackGAN-v2 model. We were able to identify a metamorphic relation prior to conducting the experiments which are built on sound reasoning. The following metamorphic relation continue to be relevant throughout the course of this study and is proposed as follows: \begin{displayquote} \textbf{MR\textsubscript{01}} Introduction of a minimally obtrusive object consistently in all training image datasets should not drastically affect the IS. \end{displayquote} \textbf{MR\textsubscript{01}} assumes that the secondary object introduced into the training image should not severely affect the results of the IS, especially when the object has minimal obscuration on the main focus of the image, which is a single bird. This metamorphic relation was built on observations of the unmodified training datasets, which were found to contain watermarks in more than one instance. Thus, we find it reasonable to assume the added object should not affect the results by a significant factor. Prior to executing any test cases, we tested the pre-trained model provided by StackGAN-v2 and performed training on the original dataset to establish a stable ground truth. In the following sections, the term \textit{self-trained} will be used to denote the model trained from unmodified dataset. We found that we were able to closely replicate the IS provided in the StackGAN-v2 paper, and the generated images were relatively realistic in nature. The compilations of the ISs are presented in Table \ref{table:1}. In addition to the birds model, we have also initialised the church model for investigative purposes. Having established the fundamental metamorphic relation, we began exploring objects of interest to insert into the training images. These objects need to be causing minimal obstruction to the principal object, the bird, and still be of theoretical interest to our study. The introduced objects were also scaled to a manageable size such that it would be smaller in size than the principal object. Manual validation was conducted to ensure the introduced objects result in minimal obstruction of the view of the principal object. Samples of the modified images are presented in Figure \ref{fig:sample_modified_img}, where the labels \textit{TC01-TC08} correspond to the test cases generated from different \textbf{MRs}, which will be discussed in the subsequent sections. Furthermore, a summary that presents the associated description and test cases of all the MRs that were covered in this paper are available on Table \ref{table:2}. \begin{figure}[t] \includegraphics[width=0.48\textwidth]{images/sample_modified_images_new.PNG} \centering \caption{Sample of the modified training images in different test cases.} \label{fig:sample_modified_img} \end{figure} \subsection{Test Cases and Results} \label{results} \begin{figure}[t] \includegraphics[width=0.35\textwidth]{images/bounding_box_issue_2x4.jpg} \centering \caption{The real samples images for self-trained model. Real samples are the processed training image through usage of bounding boxes to identify the focal birds. } \label{fig:bounding_box_issue} \end{figure} \begin{table*}[h!] \caption{Summarised Table of MRs} \label{table:2} \centering \sisetup{ table-number-alignment = center, table-figures-integer = 1 } \begin{tabular}{|m{0.04\linewidth}| @{}c@{}|} \hline MRs & Information \\ \hline {MR\textsubscript{01}} & \begin{tabular}{m{0.1\linewidth}| @{}c@{}} {Description} & \begin{tabular}{m{0.8\linewidth}} {Introduction of a minimally obtrusive object consistently in all training image dataset should not drastically affect the IS.} \end{tabular} \\\hline {Causal Relation} & \begin{tabular}{m{0.8\linewidth}} {-} \end{tabular} \\\hline {Test Cases} & \begin{tabular}{m{0.03\linewidth}|@{}c@{}} {TC\textsubscript{01}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one bird to 100\% of the training images.} \\\hline {Variation} & {10 different birds chosen at random for each image.}\\ \end{tabular} \\ \end{tabular} \\ \end{tabular} \\\hline {MR\textsubscript{02}} & \begin{tabular}{m{0.1\linewidth}| @{}c@{}} {Description} & \begin{tabular}{m{0.8\linewidth}} {Introduction of a minimally obtrusive object consistently in all training image dataset should result in a grey-tinted effect and impact the IS in a similar manner, regardless of the type of object introduced.} \end{tabular} \\\hline {Causal Relation} & \begin{tabular}{m{0.8\linewidth}} {MR\textsubscript{01}} \end{tabular} \\\hline {Test Cases} & \begin{tabular}{m{0.03\linewidth}|@{}c@{}} {TC\textsubscript{02}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one tree to 100\% of the training images.} \\\hline {Variation} & {10 different trees chosen at random for each image.}\\ \end{tabular} \\ \end{tabular} \\ \end{tabular} \\\hline {MR\textsubscript{03}} & \begin{tabular}{m{0.1\linewidth}| @{}c@{}} {Description} & \begin{tabular}{m{0.8\linewidth}} {Introduction of a minimally obtrusive object in only a selected portion of the training image dataset should result in a diminished effect of the grey-tinted effect and a higher IS than the model in which all training image dataset were modified.} \end{tabular} \\\hline {Causal Relation} & \begin{tabular}{m{0.8\linewidth}} {MR\textsubscript{01}} \end{tabular} \\\hline {Test Cases} & \begin{tabular}{m{0.03\linewidth}|@{}c@{}} {TC\textsubscript{03}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one bird to 30\% of the training images.} \\\hline {Variation} & {10 different birds chosen at random for each image.}\\ \end{tabular} \\\hline {TC\textsubscript{04}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one tree to 30\% of the training images.} \\\hline {Variation} & {10 different trees chosen at random for each image.}\\ \end{tabular} \\ \end{tabular} \\ \end{tabular} \\\hline {MR\textsubscript{04}} & \begin{tabular}{m{0.1\linewidth}| @{}c@{}} {Description} & \begin{tabular}{m{0.8\linewidth}} {Introduction of a minimally obtrusive object consistently in all training images should result in a grey-tinted effect and impact the IS in a similar manner, regardless of the colour of the object introduced.} \end{tabular} \\\hline {Causal Relation} & \begin{tabular}{m{0.8\linewidth}} {MR\textsubscript{02}, MR\textsubscript{03}} \end{tabular} \\\hline {Test Cases} & \begin{tabular}{m{0.03\linewidth}|@{}c@{}} {TC\textsubscript{05}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one bird to 100\% of the training images.} \\\hline {Variation} & {1 green bird for all images.}\\ \end{tabular} \\\hline {TC\textsubscript{06}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one bird to 100\% of the training images.} \\\hline {Variation} & {1 blue bird for all images.}\\ \end{tabular} \\\hline {TC\textsubscript{07}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one bird to 100\% of the training images.} \\\hline {Variation} & {1 red bird for all images.}\\ \end{tabular} \\ \end{tabular} \\ \end{tabular} \\\hline {MR\textsubscript{05}} & \begin{tabular}{m{0.1\linewidth}| @{}c@{}} {Description} & \begin{tabular}{m{0.8\linewidth}} {Introduction of an unobtrusive object consistently in all training images should not result in a grey-tinted effect and not impact the IS by a huge factor.} \end{tabular} \\\hline {Causal Relation} & \begin{tabular}{m{0.8\linewidth}} {MR\textsubscript{02}, MR\textsubscript{03}, MR\textsubscript{04}} \end{tabular} \\\hline {Test Cases} & \begin{tabular}{m{0.03\linewidth}|@{}c@{}} {TC\textsubscript{08}} & \begin{tabular}{m{0.08\linewidth}|m{0.64\linewidth}} {Modification} & {Introduction of one bird to 100\% of the training images, without obstructing the actual bird-of-interest.} \\\hline {Variation} & {10 different birds chosen at random for each image.}\\ \end{tabular} \\ \end{tabular} \\ \end{tabular} \\\hline \end{tabular} \end{table*} In order to validate \textbf{MR\textsubscript{01}}, we propose the first test case, \textbf{TC\textsubscript{01}} to look into the ability of the StackGAN-v2 model to manage the quantification of the focal object (the birds). We have found that the model utilises bounding boxes when searching for birds in an image and are prone to poor performance where multiple birds were not properly identified and only parts of their bodies were included in the bounding box (see Figure \ref{fig:bounding_box_issue}). Thus, we are interested to observe the ability of the model to correctly identify the actual bird within the image, instead of the added bird we introduced. It was our expectation that the resulting synthesised images would not be affected adversely as was stated in \textbf{MR\textsubscript{01}}. Visually, we expect to see an additional bird in all generated images, as it was consistently present in all training images. \begin{figure*}[b] \includegraphics[width=0.9\textwidth]{images/table_generated_images_new.png} \centering \caption{Sample of the generated images for each test case.} \label{fig:table_generated_images} \end{figure*} However, the results deviated from our expectations. Upon visual inspection, it was apparent that the resulting images were all tinted in grey colour as can be noted in Figure \ref{fig:table_generated_images}, with the colours being largely washed-out. In this regard, the quantitative analysis in Table \ref{table:1} agreed with user judgement and showed a significant decrease in both IS and Likert scores. Further examination had revealed the issue may lay solely in the third and final generator, which was the only generator to yield fake samples with a grey-tinted effect. To elaborate, when performing training on a new model instance, multiple fake samples are generated as the output. Generally, a fake sample is the output from the generator of the StackGAN-v2 model, of which there are a total of 3. Thus, we are able to observe 3 different fake samples for a particular iteration of the training process. We observed in multiple instances that fake samples 2 were consistently tinted grey compared to the fake samples 0 and 1, which were vibrant in colour. Efforts were made to uncover the cause for such an effect, where both the StackGAN-v2 paper and source code were studied in detailed. We concluded that the colour consistency regularisation was not responsible for this effect, as was clarified in the original paper that it was not required in the text-to-image synthesis task, and thus irrelevant to the current study; the source code also revealed no distinctive difference in behaviour for the third generator compared to the other two generators. Finally, visual inspection also revealed that an assortment of behaviour was observed with the added bird in the generated images, where they may exhibit one of the following properties: \begin{itemize} \item Without any of the introduced birds; \item With one single introduced bird; \item With multiple introduced birds. \end{itemize} This reveals an inconsistency in the StackGAN-v2 model, which may be attributed to instability or poor bounding box behaviour. The poor result generated for \textbf{MR\textsubscript{01}} indicates that the model may have mistakenly identified the added bird as the actual bird of the image. However, the model may have also been confused by the 10 different types of birds added and will thus be further investigated in other test cases. The results from \textbf{TC\textsubscript{01}} lead us to define two new metamorphic relations, \textbf{MR\textsubscript{02}} and \textbf{MR\textsubscript{03}}, with regards to the effect of the introduced objects on the model. \textbf{MR\textsubscript{02}} was introduced to help identify if the bounding box issue was truly prevalent in \textbf{TC\textsubscript{01}}, and if so, a different object, such as a tree should reduce the confusion of the bounding box in identifying the focal birds in the image, which implies a higher IS and better quality of the generated images. \textbf{MR\textsubscript{03}} was also introduced for a similar reason, where we assume that if a lesser extent of images is modified with extra birds, then the level of confusion introduced to the model should diminish, thus resulting in a higher IS and a less severe grey-tinted effect. \begin{figure}[!b] \includegraphics[width=0.45\textwidth]{images/chart_TC1_2_3_4.png} \centering \caption{IS for \textbf{TC\textsubscript{01}}, \textbf{TC\textsubscript{02}}, \textbf{TC\textsubscript{03}} and \textbf{TC\textsubscript{04}}. The difference between \textbf{TC\textsubscript{02}} and \textbf{TC\textsubscript{04}} is less than the other two test cases, indicating the impact of modifying a proportion of training image dataset with the bird object having a larger effect as compared to using the tree object.} \label{fig:chart_TC1_2_3_4} \end{figure} The second test case, \textbf{TC\textsubscript{02}}, was built on the \textbf{MR\textsubscript{02}} , we expect there to be a similar effect as observed in \textbf{TC\textsubscript{01}}, especially so if the bounding box was not misguided by the newly introduced objects. While the result was consistent with \textbf{MR\textsubscript{02}}, we observed that the grey-tinted effect was less prominent in this particular test case, and the IS also exhibited a higher value than \textbf{TC\textsubscript{01}} (see Table \ref{table:1}). This observation, where the result of \textbf{TC\textsubscript{02}} was significantly better than the \textbf{TC\textsubscript{01}}, implies that the type of object introduced does impact the model differently. This effect can be attributed to the different shape, colour, position, or other variables of the two test cases and will be further explored in the subsequent test cases. Following this, the third test case \textbf{TC\textsubscript{03}} building on the \textbf{MR\textsubscript{03}}, was conducted to observe the stability of the StackGAN-v2 model with respect to the proportion of the modified training image dataset. Consistent with \textbf{MR\textsubscript{03}}, it was observed that the grey-tinted effect was much less noticeable in \textbf{TC\textsubscript{03}} compared to its predecessor. The IS also reveals the highest value amongst all the test cases, substantiating the claim from \textbf{MR\textsubscript{03}}. At this stage of our experiments, we are interested to verify that the \textbf{MR\textsubscript{03}} is also applicable to the test case when the foreign object is a tree. The results of the \textbf{TC\textsubscript{04}} were found to be faithful to \textbf{MR\textsubscript{03}}, similarly to its predecessor, \textbf{TC\textsubscript{03}}. Despite this, we were able to interpret a few irregularities in the results of Table \ref{table:1}. It had been earlier remarked that the \textbf{TC\textsubscript{02}} had a significantly higher IS when compared to \textbf{TC\textsubscript{01}}. Thus, we have assumed that when the added object is a tree, we may expect it to have a higher IS than a similar scenario where the added object is a bird. In contradiction to this assumption, the results from \textbf{TC\textsubscript{03}} were found to produce a higher IS than \textbf{TC\textsubscript{04}}. Interested to pursue this irregularity further, we began checking if this was just an isolated case and not a consistent behaviour. This lead to the finding that the effect of modifying only 30\%, rather than 100\% of the training image dataset for trees is minute compared to the effect in birds as can be observed from Figure \ref{fig:chart_TC1_2_3_4}. The observation was also in contradiction with \textbf{TC\textsubscript{02}}, where it was assumed that the impact of the grey-tinted effect and IS would be consistent for all test cases regardless of the type of object. This presents multiple possible scenarios as the cause for such an irregularity, of which we have established two: \begin{itemize} \item One such theory hypothesise that colour may be a factor in attributing the grey-tinted effect; \item The second theory is built on the idea of shape of the added object being a contributing factor to the grey-tinted effect in the generated images. Theoretically speaking, the effect would be especially prominent if the object was a bird, since it adds more confusion to the bounding box in trying to identify the focal bird in the image. \end{itemize} \begin{figure}[!b] \includegraphics[width=0.45\textwidth]{images/chart_TC5_6_7.png} \centering \caption{IS for \textbf{TC\textsubscript{06}}, \textbf{TC\textsubscript{07}}, and \textbf{TC\textsubscript{08}}. Noted that \textbf{TC\textsubscript{08}} in particular achieved the worst IS while \textbf{TC\textsubscript{07}} achieved the best IS. This may be an indication that the colours of the introduced object have a significant effect on the StackGAN-v2 model. } \label{fig:chart_TC5_6_7} \end{figure} We then proposed our fourth metamorphic relation \textbf{MR\textsubscript{04}} and its corresponding test cases \textbf{TC\textsubscript{05}}, \textbf{TC\textsubscript{06}} and \textbf{TC\textsubscript{07}}, based on the first theory identified from the result of \textbf{TC\textsubscript{04}} where the colour of the added object may or may not be an attributing factor in the grey-tinted effect. The test cases \textbf{TC\textsubscript{05}}, \textbf{TC\textsubscript{06}} and \textbf{TC\textsubscript{07}} were designed such that we may review if specific colours in the RGB scheme may prove a difficult concept for the StackGAN-v2 model to grasp and visualise. In particular, we have reason to believe that the model may struggle explicitly with the colour green and/or brown, based on our observations with the previous test cases \textbf{TC\textsubscript{03}} and \textbf{TC\textsubscript{04}} concerning trees. In order to ensure the consistency and reliability of our results, the test cases \textbf{TC\textsubscript{06}}, \textbf{TC\textsubscript{06}}, and \textbf{TC\textsubscript{07}} were implemented using the same bird image but represented in the respective colours specified by the test cases, and are all positioned identically for a particular image in each test case. The results of \textbf{TC\textsubscript{05}} conformed to our expectations, where the grey-tinted effect was still visibly present. The IS for \textbf{TC\textsubscript{06}}, \textbf{TC\textsubscript{06}}, and \textbf{TC\textsubscript{07}} was placed between \textbf{TC\textsubscript{01}} and \textbf{TC\textsubscript{03}}, which was expected, considering the fact that \textbf{TC\textsubscript{05}} had introduced more noise to the sample compared to \textbf{TC\textsubscript{03}}, where only 30\% of the images were modified and thus introduce less confusion to the StackGAN-v2 model (see Table \ref{table:1}). Both \textbf{TC\textsubscript{06}} and \textbf{TC\textsubscript{07}} displayed consistent attributes to what had been described in \textbf{TC\textsubscript{05}}, with the exception of \textbf{TC\textsubscript{06}}, where we notice that it has a higher IS as compared to \textbf{TC\textsubscript{03}}. This is unexpected since \textbf{TC\textsubscript{03}} had a much smaller proportion of images (30\% as compared to 100\%) with added objects, which in accordance to \textbf{MR\textsubscript{03}}, was anticipated to result in a higher IS. In order to identify the cause for this irregularity, we began looking into the differences between \textbf{TC\textsubscript{03}} and \textbf{TC\textsubscript{06}} in closer detail: \begin{itemize} \item Proportion of modified images where \textbf{TC\textsubscript{03}} had 30\% and \textbf{TC\textsubscript{06}} had 100\%; \item \textbf{TC\textsubscript{03}} randomly chooses 1 of 10 different birds to add to a training image, while \textbf{TC\textsubscript{06}} only introduces a single type of bird in blue colour. While it may appear that both colour and shape could be a contributing factor here, it seems likely that shape could be overruled as a contributing factor given that the results from \textbf{TC\textsubscript{05}} and \textbf{TC\textsubscript{07}} had a lower IS than \textbf{TC\textsubscript{03}}; \item Positions in which birds were added to the images were not consistent. Further adjustments were necessary to ensure minimal obstruction of the focal bird in the image. \end{itemize} From our results, the most feasible reason for the observed irregularity appears to be either the colour or the position of the birds. Upon closer inspection, we believe colour has little to no effect on the model as the 3 test cases \textbf{TC\textsubscript{05}}, \textbf{TC\textsubscript{06}}, and \textbf{TC\textsubscript{07}} had relatively similar results. Having now established that both shape and colour are unlikely to be the contributing factor, we began evaluating how the different positions of the birds could be affecting the results in such a manner. As was emphasised in all our test cases, we make an effort to \textit{minimise} the obstruction of the added object on the focal birds. In hindsight, while we are able to guarantee that the focal birds are not being absurdly overshadowed by the newly added objects, there may exist instances where the added object lightly brushes the focal bird's wings or torso. This became an even more relevant topic when we consider the previously examined bounding box issue. We now hypothesise that if a certain proportion of the birds in the training images is obscured by the object in some form, it would affect the ability of the bounding box to correctly identify the focal bird within the image, which is applicable to the irregularity in \textbf{TC\textsubscript{06}} if it had only a very small proportion of mistakenly identified bounding boxes when compared to \textbf{TC\textsubscript{03}}. To simplify, while \textbf{TC\textsubscript{06}} had a much larger proportion of modified images, it is possible that the bounding box was able to identify the focal birds more accurately in this particular test case if the introduced birds did not obscure the focal bird in any way. We now define a new metamorphic relation \textbf{MR\textsubscript{05}} and its resulting test case \textbf{TC\textsubscript{08}} to further investigate this hypothesis. In \textbf{TC\textsubscript{08}}, we hope to compare its results with \textbf{TC\textsubscript{01}} to attest our claim that the obstruction of the view of the focal bird would impede the ability of the bounding box to correctly identify it. The resulting model substantiates our claim with \textbf{MR\textsubscript{05}}, exhibiting a much higher IS than \textbf{TC\textsubscript{01}}, as can be noted from Table \ref{table:1}. Another surprising detail was observed, where we found the generated images to be completely free of any grey-tinted effect (see Figure \ref{fig:table_generated_images}). This further supports our claim that the bounding box is easily confused if an object obscures the focal body of the image, and is likely the contributing factor to the grey-tinted effect. Finally, we review the results of our Likert Scale analysis for all test cases (see Table \ref{table:1}). Generally, we found that the IS model corresponded well with human judgement in our analysis, with the exception of \textbf{TC\textsubscript{04}}, where it was found to be represented most poorly in both the semantic and realistic senses. This is in contradiction with the results from the IS, and may act as a point of reference in future investigations. When simply observed in a qualitative manner, we believe some anomalies appeared consistently in all test cases: \begin{itemize} \item The ability of the model to differentiate between species of birds are relatively weak. Different bird species were sparsely distributed in each other's categories, even in the self-trained model instance (see Figure \ref{fig:inconsistent_species}); \item The model has a tendency to generate multiple similar deformed images in a single test case (see Figure \ref{fig:deformed_birds}). \end{itemize} \begin{figure}[t] \includegraphics[width=0.4\textwidth]{images/inconsistent_species_new.png} \centering \caption{Sample of generated images from the self-trained model. } \label{fig:inconsistent_species} \end{figure} \begin{figure}[t] \includegraphics[width=0.3\textwidth]{images/deformed_birds.jpg} \centering \caption{Sample of generated images from \textbf{TC\textsubscript{05}}. The deformed images seem to follow some form of pattern, thus appearing repetitive in nature.} \label{fig:deformed_birds} \end{figure} \section{Conclusion} In this study, we proposed a metamorphic testing approach to identifying the robustness of text-to-image synthesis models, specifically the StackGAN-v2 model. The first contribution of this study is the identification of five metamorphic relations that may be applied to any text-to-image synthesis model to verify the robustness of the algorithm in the presence of unexpected input training images. The metamorphic relations were derived successively based on one naive metamorphic relation as a starting point, which is applicable to any text-to-image synthesis model : \emph{``How would the text-to-image synthesis model behave when noises were introduced to the training images?''}. Our findings suggest that by using the prior test case results, it is possible to derive more complicated metamorphic relations and test cases to evaluate the robustness of text-to-image synthesis models. This can then follow an iterative or cyclic process model, instead of traditional metamorphic testing methods which tend to follow a waterfall process model. Our subsequent contribution is the identification of the robustness issues in the StackGAN-v2 model, which we have found to be subjected to undocumented behaviour, particularly the grey-tinted effect when unexpected input training images are utilised. With the aid of metamorphic testing, we have successfully revealed the aforementioned behaviour of the StackGAN-v2 model, which are otherwise hidden from the researchers. Ultimately, our findings can help researchers to better understand the behaviour of the StackGAN-v2 model and further develop a robust image synthesis model. There are some potential directions to approach for future research. The naturalness of the modified images from the dataset can be enhanced since we think that realism and naturalness will not significantly affect the result from metamorphic testing. Improve the naturalness of the modified training images may benefit this research by revealing the errors which are more likely found in real life. Other than that, another limitation of this work is that we believe it is limited in its characteristic of requiring intensive labour for qualitative analysis. As such, we intend to automate our metamorphic testing process in future research. In addition to that, we are also interested in looking into the following topics: 1.) to perform further examination of the impact of colour in the StackGAN-v2 model; 2.) to investigate the effects of shape of the added object on the StackGAN-v2 model; 3.) to incorporate text input modifications in metamorphic testing. \section*{Acknowledgment} We would like to thank all the authors of the StackGAN-v2 paper \cite{zhang2018stackgan++} for responding to our queries. This work was partly supported by the Advanced Engineering Platform's CyberSecurity \& AI $\varphi^{2}$ Cluster Funding (AEP-2020-Cluster-04), Monash University, Malaysia campus. \printbibliography \end{document}

\section{} \section{ Introduction} Multiferroics are materials in which ferroic properties, e.g., magnetism and polar order coexist. Magnetic and ferroelectric ordering couple microscopically or macroscopically to form the magnetic ferroelectrics. The coupling of the two ordering leads to the so-called magnetoelectric effect in which the magnetization can be tuned by the external electric field, and vice versa\cite{1,2,3}. Magnetic ferroelectrics have potential applications in information storage, actuators, sensors, and functional devices. Perovskite BiFeO$_3$ exhibits both weak ferromagnetism and ferroelectric characteristics, and has been studied extensively in recent days.\cite{4,5,6,7,8}. The G-type antiferromagnetic (AFM) order of Fe magnetic moments exhibits a canting caused by the antisymmetric Dzyaloshinskii-Moriya interaction(DMI) under rhombohedral $R3c$ space group. However, a spiral spin structure of AFM Fe sublattice rotates through the crystal with a long-wavelength period of 620{\AA}, and decrease the weak ferromagnetism further. In our previous work, we suggested that the effect of decreasing magnetism caused by the spiral spin structure can be suppressed by doping magnetic transitional metal ions in perovskite B sites\cite{9}. We suggest perovskite Bi$_2$FeMnO$_6$ is a good candidate to fulfill this requirement in that Mn and Fe have different magnetic moments and hence lower DMI to produce the rotation of AFM vectors. Comparing with BiFeO$_3$, the new system should have an observed magnetization and reversal of magnetization macroscopically. Except the ferroelectric displacement of Bi site driven by Bi-6s stereochemically active lone pair induced by the mixing between the $(ns)^2$ ground state and a low-lying $(ns)^1(np)^1$ excited state, there exists another structural distortion, the alternating sense of rotation of the oxygen octahedra along [1 1 1] direction, which is known as antiferrodistortive(AFD) distortion \cite{10,11}. Also in our former paper, it is shown that AFD distortion couples with the weak ferromagnetism due to the DMI which is decreasing with increasing on-site Coulomb iteraction($U$) \cite{12}. We aim to extend our study to perovskite Bi$_2$FeMnO$_6$ to investigate the coupling between AFD distortion and weak ferromagnetism under DMI using $ab$ $initio$ calculations with considering the spin-orbital(SO) coupling effect and the noncollinear spin configuration, while Mn were doped periodically along [1 1 1] direction in Fe site of BiFeO$_3$ .Does AFD distortion couple with magnetism in Bi$_2$FeMnO$_6$ when Fe and Mn are arranged in G-type AFM ordering? What is the origin of coupling between the AFD distortion and the magnetism under DMI in Bi$_2$FeMnO$_6$? What is the role of DMI in the coupling between AFD distortion and magnetization? How does Coulomb iteraction($U$) take effect on the DMI? In this paper we have proposed the mechanism of DMI in Bi$_2$FeMnO$_6$,using first-principles calculations based on density functional theory(DFT).This work can shed light on discovering the magnetoelectric coupling process in perovskite multiferroics. The remainder of this paper is organized as follows: In section 2, we presented the computational details of our calculations. We reported the calculated results and discussions in section 3. In section 4, we drew conclusion based on our calculation. \section{ Computational details} Our calculations were performed within the local spin density approximation(LSDA) to DFT using the ABINIT package\cite{13,14}. The ion-electron interaction was modeled by the projector augmented wave (PAW) potentials \cite{15,16} with a uniform energy cutoff of 500 eV. Bi 5d, 6s, and 6p electrons, Fe 4s, 4p,and 3d electrons, and O 2s and 2p electrons were considered as valence states. Two partial waves per $l$ quantum number were used. The cutoff radii for the partial waves for Bi, Fe, and O were 2.5, 2.3, 1.1 a.u., respectively. $6\times6\times6$ Monkhorst-Pack sampling of the Brillouin zone were used for all calculations. We calculated the net magnetization per unit cell and the electronic properties within the LSDA+U method where the strong Coulomb repulsion between localized $d$ states has been considered by adding a Hubbard-like term to the effective potential\cite{17,18,19}. The effective Hubbard parameter, the difference between the Hubbard parameter $U$ and the exchange interaction $J$ ($U-J$), was changing in the range between 0 and 5 eV for the Fe and Mn $d$ states. For 0 eV of $(U-J)$, $J$ was varying as 0,0.5, 0.8,and 1 eV, respectively. $J$ remained 1 eV for other effective Hubbard values. Taking into account the SO interaction, we introduced the noncollinear spin configuration to construct the G-type AFM magnetic order with the AFM axis being along the $x$ axis in Cartesian coordinates in our $ab$ $initio$ calculation. \section{ Results and discussion} The initial lattice parameters are introduced as same as BiFeO$_3$ in Ref.[4]. Then cell shape and the atomic positions are fully relaxed, and the relaxed parameters are given in table 1. One can see that the lattice constant is decreased with doping Mn in BiFeO$_3$. The rhombohedral angle is also decreased comparing with BiFeO$_3$. This shows that the lattice cell is compressed with substitution of Mn in Fe site while lattice shape is also changed with decreasing rhombohedral angle. For the AFD distortion, a rotational vector $\mathbf{R}$ has been introduced to describe the direction of the rotation of the oxygen octahedra\cite{12}. $\mathbf{R}_{out}$ corresponds to the state in which the rotational vectors of two neighboring oxygen octahedra are deviating away, while $\mathbf{R}_{in}$ is pointing inward. The rotational angle is 10\textordmasculine in the Cartesian coordinates\cite{12}. $U$ and $J$ are the Coulomb interaction and superexchange interaction parameter which were implemented in LSDA+U calculation as in ref.[12]. $U$ is the amount of energy required to transfer one electron from one site to its nearest neighbor, while $J$ indicates the strength of the magnetic superexchange interaction. In table 2, we present the net magnetization per unit cell with respect to $\mathbf{R}_{in}$ and $\mathbf{R}_{out}$ in Cartesian coordinates for different $U$ and $J$. For the sake of clarity, we only set varying value of $J$ for $U$=0 eV. Again, one can see that $J$ value have no effect on the resulting magnetization when $U$ remains constant. The initial magnetic moment of Fe are arranged in ferrimagnetic order along $x$ direction which is known as G-type AFM order. The resultant magnetic moment in $x$ direction is due to the ferrimagnetic arrangement of Fe and Mn magnetization. The neighboring moment will interact under the DMI and deviate away from the original $x$ direction to produce a resultant magnetic moment in $y$ direction. When the rotational vector $\mathbf{R}$ is reversed, the resultant magnetic moment will be reversed too until the $U$ is smaller than 2.7 eV. This indicates the AFD and magnetic ordering is coupled and DMI is still active in this case. As $U$ is greater than 2.7 eV, the coupling between these two ordering is prohibited. That is to say, the DMI is precluded in this case. It is worth mentioning that the component of magnetic moment in $y$ direction for U =2.7 eV becomes zero. This indicates the critical value of $U$ for depressing DMI is 2.7 eV in which the antisymmetric Fe and Mn magneitc moments stop interacting with each other. This critical value is smaller than that for BiFeO$_3$ where the DMI is turned off as $U$ is approaching 2.9 eV\cite{12}. This indicates that Fe-Mn antisymmetric interaction where interacting spin moment are $(5/2)S-(4/2)S$ is weaker than Fe-Fe antisymmetric interaction where interacting spin moment are $(5/2)S-(5/2)S$. We believe the weaker DMI and alternating spin moments in the new system would destroy the spiral spin structure through the crystal in comparison with the case in BiFeO$_3$. Hereby the magnetization and reversal of magnetization can be improved in the new system. In order to make the reversal of magnetization clear, the resultant AFM vectors are illustrated in Fig.1 with respect to $U$. AFM vectors have three components corresponding to $x$, $y$, and $z$ direction, respectively. It is apparent that there is an anomaly as $U$ is approaching 2.7 eV which corresponds to the critical value of eliminating DMI. The magnetization is inverted as $U$ is under the critical value which indicates the DMI is still active in this $U$ range. Meanwhile the DMI is inversely proportional to the $U$ value. The deviation of original magnetic moment away from $x$ axis only occurs in $xoy$ plane. That is why the reversal of magnetic moment can not be observed in $z$ direction. The relationship between DMI and Coulomb on-site interaction can be understood theoretically as follows. The Hamiltonian of the new perovskite system reads, \begin{equation} H_{Bi_2FeMnO_6}=-2\sum_{<Fei,Mnj>}\textbf{J}_{Fei,Mnj}\mathbf{S}_{Fei}\cdot\mathbf{S}_{Mnj}+\sum_{<Fei,Mnj>}\textbf{D}_{Fei,Mnj}\cdot\mathbf{S}_{Fei}\times\mathbf{S}_{Mnj}. \end{equation} The first term is attributed to the symmetric superexchange, and the second term is the antisymmetric DMI contribution which only occurs when the inversion symmetry of the cations is broken. The constant \textbf{D} is related to the AFD displacement and the rotational vector. It functions like a 'DMI constant' which indicates the strength of the DMI. It reads by the second order perturbation in the case of one electron per ion \begin{equation} \textbf{D}_{Fe,Mn}^{(2)}=(4i/U)[b_{nn'}(Fe-Mn)C_{n'n}(Mn-Fe)-C_{nn'}(Fe-Mn)b_{n'n}(Mn-Fe)], \end{equation} From this relation, it can be seen that the DMI constant is inversely proportional to the on-site Coulomb interaction and has nothing to do with the symmetric superexchange parameter. The new system should possess a good insulating property in terms of the band gap calculated, and hence a reversal states of ferroelectricity. The band gap for metal oxides can be opened with adopting an appropriate $U$ value. The corresponding value for BiFeO$_3$ is about 4.3 eV. We take 5 eV $U$ value for Fe and Mn in Bi$_2$FeMnO$_6$ to analyze the total density of states(DOS) in that the band gap can be increased slightly with increasing $U$ value. The band gap corresponding to two rotational vectors are demonstrated in Fig.2. Generally, the band gap corresponding to the two rotational vectors are increasing linearly with respect to $U$ value, while there exist two anomalies to the two AFD states as $U$ approaching 2.7 eV where the band gap are disappeared. These anomalies in the curve are consistent with the calculating results for AFM vectors shown in table 2 and Fig. 1, respectively. The anomalies show the disappearance of DMI occurs at the critical $U$ value and is caused by the hopping of electrons through Fermi energy. Moreover, the band gap to $\mathbf{R}_{out}$ is more narrower than to $\mathbf{R}_{in}$ which is probably caused by the different ligand field related with structure. In comparison with the total DOS without applying the on-site Coulomb interaction, the total DOS with $U$ value of 5 eV is shown in Fig. 3. The finite DOS in the vicinity of Fermi level is pushed away to form the band gap as $U$ is equal to 5 eV. A 1.26 eV and 0.75 eV band gap is obtained for $\mathbf{R}_{out}$ and $\mathbf{R}_{in}$ rotational states, respectively. This indicates that these two AFD states both possess a semiconducting property. A ferroelectric reversal would be expected to these two reversal AFD states. However, Bi$_2$FeMnO$_6$ remains metal property without applying $U$. Again this shows that the on-site Coulomb interaction is significant in opening the band gap for perovskite metal oxides. In order to shed light on the role of DMI between neighboring Fe and Mn ions, we report the orbital resolved density of states(ODOS) of the magnetic ions as $U$ is equal to 2.6 eV which is exactly under the critical value for precluding the antisymmetric DMI. The ODOS of Fe and Mn corresponding to $\mathbf{R}_{out}$ and $\mathbf{R}_{in}$ states are shown in Fig. 4. It is apparent that there are large amount of doubly degenerate $e_g$ states for Fe and Mn in the occupied bands around the vicinity of the Fermi level. The charge hybridization between Fe-3d and Mn-3d electrons is realized mainly by the $e_g$-$e_g$ states interaction. It suggests that the $e_g$-$e_g$ AFM interaction play an important role in the antisymmetric DMI between Fe and Mn ions. $e_g$ state is composed by $d_{z^2-r^2}$ and $d_{x^2-y^2}$ orbitals. It is worth mentioning that the finite ODOS in the vicinity of Fermi level is mainly attributed to the $d_{z^2-r^2}$ orbital. We suggest that $e_g$-$e_g$ AFM interaction is exactly mediated by neighboring $d_{z^2-r^2}$ orbitals of Fe and Mn ions. The charge transformation between Fe and Mn ions, and hence the inversion of magnetization can be seen from Fig. 5. The hopping of electrons between neighboring $e_g$-$e_g$ pairs occurs due to the DMI as the inversion symmetry of cations is broken. Moreover the hopping of electrons couples with the rotation of the neighboring oxygen octahedra to produce the resultant magnetization. The charge distribution in neighboring $e_g$-$e_g$ states is polarized to a certain rotational vector, and the charge distribution is inversely polarized as the rotational vector is changing direction, leading to the reversal of resultant magnetization eventually. The charge distribution is supposed to distribute uniformly without applying the DMI. However, the homogenous distribution is destroyed by the DMI between neighboring magnetic ions to form a polarized status. We suggest this constitutes the microscopic mechanism of DMI in perovskite multiferroics. On the other hand, the AFM interaction in triply degenerate $t_{2g}$-$t_{2g}$ composed by $d_{xy},d_{yz},$ and $d_{xz}$ orbitals is relatively weak comparing with $e_g$-$e_g$ AFM interaction. Again this shows that the latter is deeply involved in the antisymmetric DMI. \section{ Conclusion} The critical value of $U$ for depressing DMI in Bi$_2$FeMnO$_6$ is found to be 2.7 eV which is smaller than that in BiFeO$_3$. This indicates that the DMI in the new system functions weaker in comparison with that in BiFeO$_3$ . ODOS of Fe and Mn corresponding to different AFD displacements show that the microscopic mechanism of the DMI is originating from the hopping of electrons between neighboring $e_g$-$e_g$ states, and the $e_g$-$e_g$ AFM interaction couples with the rotation of the neighboring oxygen octahedra.

\section{A short introduction to the Bohm Picture of Quantum Mechanics (BP)} Quantum Theory, as it is explained in textbooks, comes in two fundamental versions, classified by Dirac with the word \emph{Pictures} {[}1]: the Schroedinger Picture (SP), where time developments are entirely accounted for by the wavefunctions (which can be described in many different \emph{representations}) and the Heisenberg Picture (HP), where time-developments are entirely encoded in the operators. \subsection{Starting from the SP} The differential form of the Schroedinger equation of motion is (we set $\hbar=1,$ unless otherwise stated): \begin{equation} i\frac{d}{dt}|\Psi(t)>=H|\Psi(t)>\label{eq:(1)}\end{equation} An integral representation of the Schroedinger ket vector $|\Psi(t)>$ is \begin{equation} |\Psi(t)>=e^{-iH(t-t')}|\Psi(t')>\label{eq:(2)}\end{equation} The Hamiltonian H in this context has a double role: it is the operator generating time-translations, assumed to be itself time-independent (for isolated systems), and it is the operator whose eigenvectors are stationary states $\Phi_{\nu}$ of the system: \begin{equation} H(N)\Phi_{\nu}(\{\vec{x}\};E_{\nu})=E_{\nu}\Phi_{\nu}(\{\vec{x}\};E_{\nu})\label{eq:(3)}\end{equation} We shall use the simplifying short-hand notation: \begin{equation} \{\vec{x}\}\equiv\vec{x}_{1},\vec{x}_{2},...,\vec{x}_{N}\label{eq:(4)}\end{equation} \begin{equation} \{\vec{p}\}\equiv-i\vec{\nabla}_{1},-i\vec{\nabla}_{2},...,-i\vec{\nabla}_{N}\label{eq:(5)}\end{equation} In general, let us assume some complete set of state vectors in Hilbert space: \begin{equation} \sum_{\alpha}|\alpha><\alpha|=1\label{eq:(6)}\end{equation} \begin{equation} <\beta|\alpha>=\delta(\beta,\alpha)\label{eq:(7)}\end{equation} Then the wavefunctions (or their complex conjugates) are just components (complex numbers) of the ket vector $|\Psi(t)>($or the bra vector $<\Psi(t)|$) in the Hilbert space in whatever the chosen basis vectors is {[}1] : \begin{equation} \Psi_{\beta}(t)=<\beta|\Psi(t)>\label{eq:(8)}\end{equation} \begin{equation} \Psi_{\alpha}^{*}(t)=<\Psi(t)|\alpha>\label{eq:(9)}\end{equation} We work in this paper with an entirely different Picture, unfortunately not even mentioned in textbooks. To give it a name, we shall refer to it as the Bohm Picture (BP){[}2,3,4] and shall be discussed presently. Before we proceed, let us ask the question : would it make any real difference, had Bohm started instead from the HP, instead of the SP? From one of Dirac's last publications {[}1] we can infer that for non-relativistic systems the answer is NO; for relativistic Quantum Field Theory (QFT), the case seems to remain open. Practical approximation methods for solving the time-dependent and time-independent Schroedinger equation, for reallistic systems, both in the textbook formulation and in published BP papers {[}2,3], are generally based on some standard combinations of Perturbation - Variation theory with computer power. As we shall see, wavefunctions associated with specific physical systems are an essential part of input information, but only a part. This statement sharply disagrees not only with textbook quantum theory but also with all (as far as I know) alternative ontological formulations {[}2]. Let us attempt to look more closely into the question of time development in Quantum Mechanics, as accounted for in the BP. As it is perhaps widely agreed, an universal time seems to be required by the very definition of the Schroedinger equation. We follow Bohm by choosing the basis set of wavefunctions to be the eigenstates of the Schroedinger (hermitean) {}``position operators'' $\hat{x}_{k}$: \begin{equation} \hat{x}_{k}|\vec{x}_{1},\vec{x}_{2},...,\vec{x}_{N}>=\vec{x}_{k}|\vec{x}_{1},\vec{x}_{2},...,\vec{x}_{N}>\label{eq:(10)}\end{equation} \begin{equation} [\hat{x}_{k},\hat{x}_{j}]=0\label{eq:(11)}\end{equation} \begin{equation} [\hat{p_{k}},\hat{p_{j}}]=0\label{eq:(12)}\end{equation} \begin{equation} [\hat{x}_{k},\hat{p_{j}}]=i\delta(k,j)\label{eq:(13)}\end{equation} \begin{equation} \Psi(\{\vec{x}\};t)\equiv<\{\vec{x}\}|\Psi(t)>\label{eq:(14)}\end{equation} The time-dependent Schroedinger equation for describing a system with countable N=1,2,..., degrees of freedom can always be rewritten in the form\begin{equation} H(N)\Psi(\{\vec{x}\};t)=i\frac{\partial}{\partial t}\Psi(\{\vec{x}\};t)\label{eq:(15)}\end{equation} where H(N) is the (supposed hermitean) Hamiltonian:\begin{equation} H(N)=U_{0}-\sum_{i=1}^{N}\frac{1}{2m_{i}}\nabla_{i}^{2}+V(\{\vec{x}\})\label{eq:(16)}\end{equation} $U_{0}$ is an arbitrary constant. The function V represents a $preassigned$ classically definable $local$ function: \begin{equation} <\{\vec{x}'\}|V|\{\vec{x}\}>\equiv\delta(\{\vec{x}'\}-\{\vec{x}\})V(\{\vec{x}\})\label{eq:(17)}\end{equation} So a general time-dependent wavefunction is obviously \begin{equation} \Psi(\{\vec{x}\};t)=\sum_{\nu,E_{\nu}}C_{\nu}(E_{\nu})\Phi(\{\vec{x}\};E_{\nu})\exp(-iE_{\nu}t)\label{eq:(18)}\end{equation} and is completely defined, if the values of the coefficients $C_{\nu}$ be given at some initial time t=0:\begin{equation} \Psi(\{\vec{x}\};0)=\sum_{\nu,E_{\nu}}C_{\nu}(E_{\nu})\Phi(\{\vec{x});E_{\nu})\label{eq:(19)}\end{equation} Thus assuming that an initial wave-packet is given, we immediately deduce an explicit expression for the coefficients\begin{equation} \Psi(\{\vec{x}\};0)=\sum_{i}a_{i}\varphi_{i}(\{\vec{x}\})\label{eq:(20)}\end{equation} \begin{equation} \Phi_{\nu}(\{\vec{x}\},E_{\nu})=\sum_{i}b_{\nu i}(E_{\nu})\varphi_{i}(\{\vec{x}\})\label{eq:(21)}\end{equation} \begin{equation} C_{\nu}(E_{\nu})=\sum_{i}b_{\nu i}^{*}(E_{\nu})a_{i}\label{eq:(22)}\end{equation} where $\Psi$and $\Phi$ are expanded over some convenient complete set of functions $\left\{ \varphi_{i}\right\} $. We shall simply assume familiarity with this standard material and carry on from there. Let us next introduce the BP on the basis of the following postulates: POSTULATE 1 As the wavefunction is in general a complex number, let us write\begin{equation} \Psi(\{\vec{x}\};t)=R(\{\vec{x}\};t)e^{iS(\{\vec{x}\};t)}\label{eq:(23)}\end{equation} where R and S are both real. This by definition a $beable$ {[}2,3,4] oficially known {}``wavefunction'' . It obeys its own equation of motion, which is none other than the Schroedinger equation. The latter can be rewritten as a system of time-dependent coupled differential equations involving two real functions R and S: \[ \frac{\partial S(\{\vec{x}\};t)}{\partial t}+\] \begin{equation} +\sum_{k=1}^{N}\frac{1}{2m_{k}}|\vec{\nabla}_{k}S(\{\vec{x}\};t)|^{2}+V(\{\vec{x}\})+Q(\{\vec{x}\};t)\label{eq:(24)}\end{equation} \begin{equation} \frac{\partial P(\{\vec{x}\};t)}{\partial t}+\vec{\nabla}.{\displaystyle {\displaystyle \sum_{k}}}\vec{J}_{k}(\{\vec{x}\};t)=0\label{eq:(25)}\end{equation} with the definitions \begin{equation} \vec{J}_{k}(\{\vec{x}\};t)=\vec{\nabla}_{k}S(\{\vec{x}\};t)\label{eq:(26)}\end{equation} \begin{equation} P\times\vec{J}_{k}=Im[\Psi^{*}\vec{\nabla}_{k}\Psi]\label{eq:(27)}\end{equation} \begin{equation} P=R^{2}\label{eq:(28)}\end{equation} \begin{equation} Q(\{\vec{x}\};t)\equiv-\sum_{k}\frac{1}{2m_{k}}\frac{\nabla_{k}^{2}R(\{\vec{x}\};t)}{R(\{\vec{x}\};t)}\label{eq:(29)}\end{equation} The last quantity introduced is known as the '' Quantum Potential'' or rather, {}``Quantum Information Potential'' (QIP) {[}2] and it is, together with definition (\ref{eq:(26)}), the core of the alternative formulation of non-relativistic quantum mechanics proposed by Bohm and his co-workers {[}2,3,4]. As shown by these authors, it can be naturally generalised to Quantum Field Theory (QFT). In any case, we emphasize that all the above definitions are derived from the time-dependent Schroedinger equation, and that their fundamental status stands or falls with the validity of this equation. Let us first introduce the other set of further relevant $beables$ {[}2,4] in BP, i.e. particles and fields: this $completes$ the formulation of Quantum Theory in the BP. One begins by defining dynamical quantities such as particle momenta, in complete agreement - in the appropriate limit of course, i.e. when Q is negligible - with the classical definition of momenta that is part of the Hamilton-Jacobi formulation of classical mechanics {[}2,3]. We thus get Bohm's equations of motions for particles \[ \vec{p}_{k}(\{\vec{X}(t)\};t)=m_{k}\frac{d\vec{X}_{k}(\{\vec{X}(t)\};t)}{dt}=\] \begin{equation} =\{\vec{\nabla}_{k}S(\{\vec{x}\};t)\}_{\{\vec{x}\}=\{\vec{X}(t)\}}\label{eq:(30)}\end{equation} or (equivalently) the velocity of a particle k at its position $\vec{X}_{k}(t)$: \[ \frac{d\vec{X}_{k}(\{\vec{X}(t)\};t)}{dt}=\] \begin{equation} =\frac{1}{m_{k}}[\frac{1}{P}Im(\Psi^{*}\vec{\nabla}_{k}\Psi)]_{\{\vec{x}\}=\{\vec{X}(t)\}}\label{eq:(31)}\end{equation} This non-linear set of ordinary differential equations is the fundamental set of Bohm equation of motion for $beable$ particles {[}2]: it is the essential part of the theoretical framework that , according to Bohm and co-workers, is missing from standard textbook formulations and alternative ontological interpretations {[}2]. It is then demonstrated that a complete removal of all the dead wood of ambiguities and {}``mysteries'' of textbook QM can thereby be elegantly achieved. Eq.(\ref{eq:(28)}) can be rearranged to give an alternative version: \[ m_{k}\frac{d^{2}\vec{X}_{k}(t)}{dt^{2}}=\] \begin{equation} =-[\vec{\nabla}_{k}(V(\{\vec{x}\})+Q(\{\vec{x}\};t))]_{\{\vec{x}\}=\{\vec{X}(t)\}}\label{eq:(32)}\end{equation} One can thus clearly see that any beable particle is subject to $quantum$ $accelerations$, regardless of the existence of classical (inertial) accelerations. This sharply illuminates the profound difference between what a {}``particle'' means in classical physics and what a {}``beable particle'' means in Bohm's quantum physics {[}2]. Unawereness of, or misunderstandings about this fundamental difference can only lead to unnecessary confusions and endless irrelevant objections and production of paradoxes. The individual Bohm beable particle history (i.e. existing regardless of whether the particle is, or is not, a member of an ensemble) is then formally defined to be a solution of these non-linear ordinary differential equations (31),(32). Key features of quantum theory, such as quantum non-locality and quantum wholeness {[}2], are thus precisely expressed by these equations for particle (and field) motions as they are ultimately interpreted in spacetime: (i) single particle histories are primarily completely entangled (i.e. $in$ $principle$ defined by, unseparable from ) the entire environment (literally, the entire universe), because they are $\textit{actively }$$quantum$ $in-formed$ by the $phase$ of the wavefunction {[}2]. This has of course nothing to do with $classically$ defined {}``medium entanglement'' (many-body effects) through {}``potentials, gravitation, thermal influences'' and so forth; (ii) an essential feature of this quantum information is the de Broglie $guidance$ condition on the particles eq (\ref{eq:(30)}), which puts the $potential$ beable information encoded in the common wave function into the $actual$ beable motion of every $individual$ beable particle (as distinct from statistical ensembles of such particles), participating in the potentially present common pool of information provided by the wavefunction {[}2]. So, the physics at the quantum level of accuracy differs totally from that assumed and expressed by classical mechanics, first and foremost because it is this $quantum$ $wholeness$ that appears to be the primary reality {[}2]. This is encoded $ab$$initio$ in the origin, structure and every feature of the QIP {[}2]. It should be quite clear that one is indeed introducing here crucial novel concepts that altogether single out the BP. The beable Bohm histories of particles (and fields) then go $continuously$ over to $observable$ newtonian (or Hamilton-Jacobi) trajectories. But one should be in no doubt {[}2] that the presence of the QIP profoundly changes the entire situation from the bottom up, in a way that can be neither understood nor interpreted in classical mechanics terms. Thus the emergence of familiar classical spacetime descriptions is indeed an immediate consequence of certain especial features of the QIP. Two extreme situations can be immediately envisaged: (i) The QIP is numerically completely negligible relative to the classical potential (or, in more physical terms, compared to unavoidable thermal fluctuations). In this case, the quantum forces derived from it are completely overrun by ordinary classical mechanical and thermodynamical forces. A $\textit{continuous transition}$ to a classical regime (i.e. pure spacetime descriptions, with use of coordinates) then occurs. Numerical quantum corrections are then too small to play any important role. (ii) The QIP is separable, i.e. \[ Q(\{\vec{x}\};t)=\] \begin{equation} =-\sum_{k=1}^{N}\frac{1}{2m_{k}}\frac{1}{R_{k}(\{\vec{x}\};t)}\nabla_{k}^{2}R(\{\vec{x}\};t)\equiv\sum_{k=1}^{N}Q_{k}(\{\vec{x}\};t)\label{eq:(33)}\end{equation} In this case, the specific quantum mechanical non-locality dissolves and each particle behaves classically as if completly unaware of its surroundings, except that it may interact with these surroundings through eventual classically definable preassigned potentials, or more physically speaking, through the omnipresent thermal fluctuations. POSTULATE 2: In order to confront the BP (one should never forget that BP deals primarily with individual beables {[}2,3,4]) with any existing experimental material, one needs first to analyse in detail the anatomy of actual measurement processes. This was thoroughly done by de Broglie, Bohm and their co-workers {[}2,3,4] in general, at least for non-relativistic systems. The situation when relativivistic requirements are essential is still far less digested, both in textbook QFT and in the relativistic BP {[}2,3,4]. Let us then simply summarize one of the main results : one has to apply some definite stochastic averaging procedures to the above theory of individual beables, very much the same way as it is and was always done in a classical context, e.g. in classical Statistical Mechanics {[}2,3,4]. Given these individual Bohm histories of beables, it is straightforward to compute statistical averages (i.e. averaging all initial positions and velocities) over pure ensembles of such histories {[}2,3] . For example, assuming for simplicity just one spatial dimension, \begin{equation} <X(t)>_{EA}\equiv\int dX(t)R^{2}(X(t),t)X(t)\label{eq:(34)}\end{equation} \begin{equation} <P(t)>_{EA}\equiv\int dX(t)R^{2}(X(t),t)[\frac{\partial S(x,t)}{\partial x}]_{x=X(t)}\label{eq:(35)}\end{equation} \begin{equation} <E(t)>_{EA}\equiv\int dX(t)R^{2}(X(t),t)[-\frac{\partial S(x,t)}{\partial t}]_{x=X(t)}\label{eq:(36)}\end{equation} where the history $X(t)$ is given by (\ref{eq:(30)}). This turns out to be - of course! - exactly the same as ordinary quantum mechanical expectation values, taken with the $same$ wavefunction. So, statistics (a practical necessity, simply because $we$ cannot know, nor is it necessary $for$ $us$ to know, everything that happened, happens and shall happen in the world!) does not have in the context of the present formulation the same fundamental status given to it by textbook formulations of quantum theory, which deal solely with {}``observables'', i.e. with the actual experimental data resulting from actually performed experiments. Most of these observables are context dependent, because of the participatory nature of {}``observations'' in Quantum Mechanics {[}2]. Perhaps the one important exception to this fact of life is the observable {}``position of things'' {[}4], i.e the eigenvalues of the spatial position operator. This observable can be directly linked to the beable {}``position of (discrete) particles in space'' or {}``continuous fields at specified points in space'', introduced in the previous postulate. Space is, after all, where we exist and do experiments! It is the playground of classical physics (and of common sense). As both de Broglie and Bohm demonstrated, both mathematically and using physics, the probability distributions to do this necessary stochastic averaging, whatever they might originally be, rapidly relax to none other than the familiar Born probabilities {[}2]. Because of the local conservation of probability (\ref{eq:(25)}), if this holds at some particular time, it will hold for all times. Hence, within the context of any given real experiment, the results predicted in principle by textbook and the ensemble averaged BP over all initial positions $and$ velocities must be exactly the same. A misunderstanding (or unawareness) of this most important POSTULATE 2 has caused a great deal of unnecessary confusion in the published literature {[}5] about the meaning and status of the BP. Real differences would be seen only for histories of individual systems, as distinct from ensembles of such systems. Only if, and when, someone some day has gathered sufficient such data can one make meaningful judgements on the status of BP. \section{Case studies from contemporary Nuclear and Particle physics} We consider simple, but illustrative, additional examples inspired by contemporary practices in Nuclear and Particle Physics. These fields of research offer excellent new oppurtunities for testing specific interpretations and predictions based on the BP of Quantum Theory. A good reason is, they often draw attention to the understanding of two key features of all (quantum) processes, viz. the question of transitions in time and the closely related feature of quantum non-locality. \subsection{EPR-like correlations in pure $\alpha-$decay channels of very light nuclei} Among the very light nuclei, the $^{8}Be$ and $^{12}C$ isotopes can decay into pure two or three $\alpha-$channels with appreciable branching ratios. These channels always played an important role in the theories and experiments on ordinary stellar evolution. Thus the momenta and energy distributions of the final state particles have been thoroughly investigated {[}6,7,8], but not the predictable EPR-like quantum correlations - the main issue here. The point of interest is that the outgoing $\alpha's$ in these decays are bosons, even if the parent nuclei are bags of nucleons, which are fermions. So the wavefunctions in the outgoing decay channels must be fully symmetric under exchanges of arbitrary labels attached to them (in this case, just their position coordinates). So, any model candidate to a beable wavefunction in the Center of Mass (CM) reference frame of this nucleus must be a fully symmetric function of the internal coordinates. We shall ignore for simplicity the final state Coulomb interaction {[}7] that in principle operates among these outgoing particles and accelerates them away from each other. We could assume, as it is done in actual numerical fits {[}7], that this Coulomb interaction is screened; in any case, it plays only a secondary role here, at it leads basically to what could be called $classical$ $entanglement$. However, the outgoing $\alpha-$channels, even if not interacting through pre-assigned classical potentials, are EPR-correlated through the Bose symmetry requirement on the wavefunction {[}2]. As is well-known these EPR-correlations are essentially non-local {[}2], and thus beyond any classical accountability. We shall show from the viewpoint of the BP of Quantum Mechanics how these correlations ( that must be there) can be traced back to the symmetry requirement on the wavefunction and, as a by-product, attempt to give a precise meaning to words such as {}``identical, equivalent, distinguishable, undistinguishable,...'' particles. In order to make this discussion as simple as possible but without losing generality, let us consider only one spatial dimension and imagine a $^{8}Be$ nucleus (at rest when t=$0$) fissioning into two $\alpha$-particles. Extensions to heavier nuclei decays involving more $\alpha-$ channels {[}6] do not add essentially new physics, but do complicate the detailed mathematics considerably. So, let the Hamiltonian for t>0 be \begin{equation} H=-\frac{1}{2M_{\alpha}}{\displaystyle \sum_{k=1}^{2}}\frac{\partial^{2}}{\partial x_{k}^{2}}\;\;\;\;\; t>0\label{eq:(37)}\end{equation} A central feature of this Hamiltonian is its invariance under the two permutations of the labels 1,2 as a result of a common mass for the particles. True stationary states must then share this symmetry with the Hamiltonian. Quantum Mechanics establishes that these must be fully symmetric functions of the coordinates, as one is dealing with bosons {[}10]. The Schroedinger equation is \begin{equation} H\Psi(x_{1},x_{2};t)=i\frac{\partial}{\partial t}\Psi(x_{1},x_{2};t)\;\;\; t>0\label{eq:(38)}\end{equation} It must be complemented with the initial condition \begin{equation} \Psi(x_{1},x_{2};0+)=F(x_{1},x_{2})\label{eq:(39)}\end{equation} where F is a given (fully symmetric) function of spatial coordinates. It is assumed of course that the wavefunction is continuous at t=0. The consequences of the symmetry requirement are obvious, if we transform to Jacobi coordinates: \begin{equation} x=x_{1}-x_{2}\label{eq:(40)}\end{equation} \begin{equation} X_{CM}=\frac{x_{1}+x_{2}}{2}\label{eq:(41)}\end{equation} The coordinate $x$ is antisymmetric under the exchange $1\leftrightarrow2$ whereas $X_{CM}$ is symmetric. Hence we must have, \begin{equation} \Psi(x_{1},x_{2};t)=g(x;t)\times e^{iPX_{CM}}\label{eq:(42)}\end{equation} where \begin{equation} g(x;t)=g(-x;t)\label{eq:(43)}\end{equation} Let us assume that we are in the CM coordinate system First consider a single particle guided by the wave-packet \[ \Psi(x,t;\{A\})=\int_{-\infty}^{+\infty}exp(-i\frac{p^{2}}{2M_{\alpha}}t)*e^{-(p-p_{A})^{2}/\sigma_{A}^{2}}*\] \begin{equation} *exp(ip(x-x_{A})dp\;\;\;\; t>0\label{eq:(44)}\end{equation} Putting \begin{equation} \Psi(x,t;\{A\})=R(x,t;\{A\})exp(iS(x,t;\{A\}))\label{eq:(45)}\end{equation} we find that the particle velocity at time t is given by the guidance condition (\ref{eq:(30)}): \[ \frac{dX(t;\{A\}))}{dt}=\frac{1}{M_{\alpha}}\{\frac{\partial}{\partial x}S(x,t;\{A\}))\}_{x=X(t)}=\] \begin{equation} =\frac{1}{1+(\frac{\sigma_{A}^{2}}{2M_{\alpha}}t)^{2}}[\frac{p_{A}}{M_{\alpha}}+\frac{\sigma_{A}^{2}}{M_{\alpha}}(X(t;\{A\})-x_{A})t]\label{eq:(46)}\end{equation} A simple quadrature then gives the particle position at time t. This shows that the particle is accelerated from its initial velocity $p_{A}/M_{\alpha}$ at time t=0 to its final velocity at t=+$\infty$. The fundamental reason for this pure quantum effect is, as discussed in Section 1, the action of the QIP( \ref{eq:(29)} ) that is directly linked to the Schroedinger wavefunction. Next, consider two $\alpha's$ but ignore the symmetry requirement. Then the wavefunction is simply \begin{equation} \Psi(x_{1},x_{2},t;\{A,B\})=\Psi(x_{1},t;\{A\})\times\Psi(x_{2},t;\{B\})\label{eq:(47)}\end{equation} hence \begin{equation} R(x_{1},x_{2},t;\{A,B\})=R(x_{1},t;\{A\})\times R(x_{2},t;\{B\})\label{eq:(48)}\end{equation} \begin{equation} S(x_{1},x_{2},t;\{A,B\})=S(x_{1},t;\{A\})+S(x_{2},t;\{B\})\label{eq:(49)}\end{equation} Using the guidance condition (\ref{eq:(30)}) the velocities are \begin{equation} \frac{d}{dt}X_{1}(t;A)=\{\frac{\partial}{\partial x_{1}}S(x_{1},t;\{A\})\}_{x_{1}=X_{1}(t)}\label{eq:(50)}\end{equation} \begin{equation} \frac{d}{dt}X_{2}(t;B)=\{\frac{\partial}{\partial x_{2}}S(x_{2},t;\{B\})\label{eq:(51)}\end{equation} We conclude that the $\alpha-$particle that was labelled 1 at time t=0 follows the beable trajectory A, i.e. $X_{1}(t;A)$, and is therefore still the same particle at any later time t, exactly as one classically would expect ; the same applies to particle 2, moving along the beable spacetime trajectory $X_{2}(t;B)$. As expected classically, 1 and 2 are completely {}``unaware'' of each other, in spite of the fact that they have a common origin, that is, the fissioning of $^{8}Be$. There is no question about which is 1 and which is 2, as one can always in principle follow their trajectories at any time. This applies, even if 1 and 2 are {}``indistinguishable, equivalent, ...''. We are so far in complete harmony with classical physics thinking. Next, consider again two $\alpha's$ but insist that the symmetry requirement must be obeyed by the wavefunction, as it should. Then the correct wavefunction is \[ \Psi(x_{1},x_{2},t;\{A,B\})=\frac{1}{\sqrt{2}}\{\Psi(x_{1},t;\{A\})\times\Psi(x_{2},t;\{B\})+\] \[ +\Psi(x_{2},t;\{A\})\times\Psi(x_{1},t;\{B\})\}\] \begin{equation} =\Psi(x_{2},x_{1},t;\{A,B\})\label{eq:(52)}\end{equation} Now the separability expressed by the definitions (\ref{eq:(48)}) and (\ref{eq:(49)}) is lost, and the results (\ref{eq:(50)}) and (\ref{eq:(51)}) are no longer valid. Then the more general condition (\ref{eq:(30)}) prevails: \[ \frac{d}{dt}X_{1}(t;A,B)=\] \begin{equation} =\{\frac{\partial}{\partial x_{1}}S(x_{1},x_{2},t;\{A,B\})\}_{x_{1},x_{2}=X_{1}(t),X_{2}(t)}\label{eq:(53)}\end{equation} \[ \frac{d}{dt}X_{2}(t;A,B)=\] \begin{equation} =\{\frac{\partial}{\partial x_{2}}S(x_{1},x_{2},t;\{A,B\})\}_{x_{1},x_{2}=X_{1}(t),X_{2}(t)}\label{eq:(54)}\end{equation} This is tantamount to quantum non-locality, here meaning dependence on $x=x_{1}-x_{2}$ that persists no matter how large $x$ is, even in absence of any classically describable forces. So, a natural (and classically meaningful) question would be: suppose one places an $\alpha-$detector at a position D$_{A}$ and simultaneously another identical $\alpha$-detector at a position D$_{B}$. These detectors duly detect at some time t the two $\alpha-$particles upon their arrival there. But is the particle that is detected at D$_{A}$ the {}``same'' particle called 1 at time t=0, and the particle detected at D$_{B}$the {}``same'' called 2 at time t=0? The exact and real answer is given by (\ref{eq:(53)}) and (\ref{eq:(54)}) which, within the classical physics philosophy (and within the philosophy of the usual common language as well), are meaningless! Another feature of results (\ref{eq:(53)}) and (\ref{eq:(54)}) that are closer to the original EPR-correlation problem, is: instead of position detectors we could try momentum detectors. Eqs.(\ref{eq:(53)})and (\ref{eq:(54)}) show that the $momentum$ of say particle 1 at time t must depend on the $position$ at the $same\; time\; t$ of particle 2, independently of the distance separating the two particles and vice-versa. Note that this conclusion applies to individual beables and thus has nothing to do with the usual probabilities and statistics. However, upon taking appropriate ensemble averages (Section 1), i.e. doing the usual statistics of measurements {[}2,3,4], we would end up exactly with the standard textbook predictions, which apply of course only to some actual experimental result and refer only to that particular experimental set up that produces that result. \subsection{Neutron interferometry under gravitational fields} The subject of specific quantum features seen in accelerated reference systems is a promising field for future experimental research. Generally speaking, ultracold neutron physics {[}9] is especially interesting, due to high-quality experimental work now available. Essentially quantum effects are deeply linked with changes in the phase of the wavefunction {[}10]. So, the most revealing tell-tale of the underlying quantum reality is most simply and directly provided by interference experiments. Classical examples are provided by the Aarhanov-Bohm and the Aarhanov-Casher effects {[}2,3]. More recently, very precise experimental results on motions of neutrons in the earth's gravitational field became available {[}9]. Consider a free neutron in motion under the gravitational field of the earth {[}9,10]. Classically, any massive particle would move along the newtonian trajectory given by a solution of the equation of motion \begin{equation} m_{in}\frac{d^{2}z}{dt^{2}}=-m_{gr}g\label{eq:(55)}\end{equation} where the z-axis is chosen to be parallel to the gravitational lines of force. As well-known {[}11] \begin{equation} m_{in}/m_{gr}=universal\;\; constant\label{eq:(56)}\end{equation} So a choice of mass units allows this universal constant to be set to unity, and we get a simple $geometrical$ law of motion under gravity \begin{equation} \frac{d^{2}z(t)}{dt^{2}}=-g\label{eq:(57)}\end{equation} \begin{equation} z(t)=z(0)-\frac{1}{2}gt^{2}\label{eq:(58)}\end{equation} quite independent of the mass parameter. When using QM to reexamine this phenomenon, this most remarkable independence on mass seems to disappear. However, this holds only if $\hbar\neq0$, which of course is the case: the Schroedinger equation of motion is then \begin{equation} [-\frac{d^{2}}{dz^{2}}-2(\frac{M_{N}}{\hbar})^{2}gz]\Psi(z,t)=2i(\frac{M_{N}}{\hbar})\frac{\partial}{\partial t}\Psi(z,t)\label{eq:(59)}\end{equation} where $M_{N}$is the neutron mass. So the mass does contribute, but only in the combination $M_{N}/\hbar$. Note that there is no contradiction at all with the experimental fact {[}11], just as there isn't any in classical mechanics. Thus the Bohm equation of motion for any individual neutron that replaces (\ref{eq:(57)}) can be written down. The neutron velocity is given by eq \begin{equation} v(Z(t),t)=\frac{1}{M_{N}}[\frac{d}{dz}S(z,t)]_{z=Z(t)}\label{eq:(60)}\end{equation} where $S(z,t)/\hbar$ is the phase of the wavefunction $\Psi(z,t)$. From this the neutron history can be predicted, given the initial conditions. Consider now the following experiment {[}10]: a neutron beam with energy E is split into two channels, a channel ABD and another channel ACD. The channels are made to meet again at the neutron detector D. For example, in the table-top experiment discussed in {[}10] the neutron following the path ABD first climbs from the ground level up an inclined plane (say, at angle $\delta$ with the ground) from A to B along a path of length $l_{1}$then at B follows the path BD (length $l_{2}$) parallel to the ground (thus normal to the gravitational lines of force) until it reaches the detector at D. If a neutron is instead guided along the path ACD, the opposite sequence is followed, i.e. first AC (corresponding to BD) then CD (corresponding to AB) until the neutron reaches the detector at D. The detectors at D record the arrival at D of any neutron member of a statistical ensemble, sharing the same wavefunction, as a function of the inclination $\delta$ for a given beam energy E. This is of course just another example of a {}``which-path'' experiment, this time using neutron interferometry {[}10]. Let us then assume that \[ \Psi(Z(t),t)=R_{1}(Z(t),t)exp(iS_{1}(Z(t),t)/\hbar)+\] \begin{equation} +R_{2}(Z(t),t)exp(iS_{2}(Z(t),t)/\hbar)\label{eq:(61)}\end{equation} The probability pattern at the detector is then \begin{equation} P(Z(t),t)=R_{1}^{2}(Z(t),t)+R_{2}^{2}(Z(t),t)+2R_{1}R_{2}cos\Phi_{12}\label{eq:(62)}\end{equation} where $\Phi_{12}$ is the phase difference between the two paths : \begin{equation} \Phi_{12}=(S_{1}(ABD)-S_{2}(ACD))/\hbar\label{eq:(63)}\end{equation} It can be quickly estimated in the WKB approximation (which, remember, neglects the QIP). Phase change along ABD: \begin{equation} \frac{S_{ABD}}{\hbar}=\int_{ABD}pdl/\hbar=\int_{ABD}\sqrt{2M_{N}(E-M_{N}gz)}dl/\hbar\label{eq:(64)}\end{equation} and along ACD: \begin{equation} \frac{S_{ACD}}{\hbar}=\int_{ACD}pdl/\hbar=\int_{ACD}\sqrt{2M_{N}(E-M_{N}gz)}dl/\hbar\label{eq:(65)}\end{equation} hence \begin{equation} \Phi_{12}=-l_{2}\sqrt{2M_{N}E}/(2E\hbar)*M_{N}gl_{1}sin\delta\label{eq:(66)}\end{equation} Recalling that \begin{equation} \sqrt{2M_{N}E}/E=2M_{N}\lambda/h\label{eq:(67)}\end{equation} where $\lambda$ is the de Broglie wavelength of the incoming wavefunction. We thus get our final result {[}10] \begin{equation} \Phi_{12}=-(M_{N}/\hbar)^{2}gl_{1}l_{2}sin\delta*\frac{\lambda}{2\pi}\label{eq:(68)}\end{equation} This prediction appears to be in excellent agreement with the experimental results {[}9,10]. Let us now consider the realistic case that any incoming neutron with given velocity belongs in a statistical ensemble of such particles. Upon arrival at time t=0 at A they are all redirected (i.e. the trajectories bifurcate). Depending on the initial position and velocity of a particle, either the path ABD or the path ACD will be followed, but every single body is forced to meet at D. The histories Z(t) of individual neutrons are given by formula (30).The actually observed pattern of arrivals of individuals at D is fixed by (62). The $probability$ that at time t=0 any particular neutron $is$ at Z=0 with velocity V(0) is $assumed$ (Postulate 2) to be given by \begin{equation} P(Z(0)=0,t=0)=|\Psi(Z(0)=0,t=0)|^{2}\label{eq:(69)}\end{equation} We emphasize that in BP there is nothing fundamental about this assumption. As a matter of fact, this is the only place where probabilities enter the game at all, just as they do in Classical Statistical Mechanics. Everything else for t>0 is completely determined by the initial conditions. Thus the probability that a neutron $will$ $be$ at the position Z(t) (with velocity V(t) given by (30 )) is \begin{equation} P(Z(t),t)=|\Psi(Z(t),t)|^{2}\label{eq:(70)}\end{equation} in full formal agreement with textbooks. \subsection{Quantum transitions in nuclear $\beta$-decay processes} The ongoing experimental and theoretical work in Nuclear and Particle Physics has by now initiated an exciting new era of fundamental physics. Random samples of the enormous output of publications in later years that are relevant to this paper are {[}9],{[}12],{[}13 -17],{[}18]). An especially interesting and relevant feature of recent developments is the strong light they cast on the nature of quantum transitions in electroweak and QCD physics. \subsubsection{The time-dependent wavefunction} Good examples of contemporary interest is K-capture in highly ionized H-like atoms {[}18] and the physics of ultracold neutrons in traps {[}9]. We shall $reinterpret$ this according to the philosophy sketched in this paper. Let us consider the decays \begin{equation} _{Z}A_{N}+e^{-}\rightarrow_{Z-1}A_{N+1}+\nu_{e}\label{eq:(71)}\end{equation} This is a K-capture process, meaning capture of an 1S atomic electron by a proton in a parent nucleus $_{Z}A_{N}$ with Z protons and N neutrons through electroweak interactions that instantaneously replace a proton with a neutron in the daughter nucleus $_{Z-1}A_{N+1}$ , together with an outgoing electron neutrino $\nu_{e}$. This is accompanied by a sudden, but in this case rather small, change of the Coulomb field of the nucleus. An interesting feature of this kind of experiments is that the parent atom is supposed to be H-like, i.e. to be completely ionized except for a single bound electron, say in the 1S atomic state {[}18]. We shall consider this without any explicit or implicit reference to {}``meaurements'', {}``probabilities'' , ''observations'' or any other related standard concepts and terminology used in textbooks {[}2,3,4], where they are considered to be essential for self-consistency. As already discussed, this is not the case with the BP {[}2,3,4], where {}``measurements'' of {}``observables'' (implying actually no more than a permanent record of events stored somewhere and referring exclusively to a specific experimental set-up) are just special cases of quantum transition processes for beables {[}2,3,4]. Postulate 2 would link this to any specific available or predictable experimental results and, of course, to any textbook postdictions or predictions of such results. Again, this is not the purpose of the present paper. We imagine a movie showing the time development of a K-capture process ocurring say at time t=0. The Hamiltonian driving the evolution of the wave function is assumed to be \begin{equation} H=U_{0}+H_{0}+V_{eff}\label{eq:(72)}\end{equation} $U_{0}$ represents the (infinite) Dirac vacuum energy. The next term is defined as \[ H_{0}=H_{nuc}-\frac{\hbar^{2}}{2m_{e}}\nabla_{y}^{2}(e)-Z\alpha\int d^{3}\vec{x}\frac{\varrho_{el}(\vec{x})}{|\vec{x}-\vec{y}|}+\] \begin{equation} {\displaystyle +\sum_{n=1}^{3}}[-i\hbar\vec{\alpha}(n).\vec{\nabla}_{y}(n)+\beta(n)m_{n}c^{2}]\label{eq:(73)}\end{equation} We shall reserve the notation $\{\vec{x}\}$ for space coordinates of the nucleons. The symbol $\vec{y}$ refers exclusively to the spatial coordinates of leptons. The first term on the rhs of (\ref{eq:(72)}) represents the nuclear Hamiltonian: \[ H_{nuc}\Phi_{\alpha}(\{\vec{x},\vec{x}_{CM}\};Z,N)=\] \begin{equation} =E_{\alpha}(Z,N)\Phi_{\alpha}(\{\vec{x},\vec{x}_{CM}\};Z,N)\label{eq:(74)}\end{equation} where \begin{equation} \vec{x}_{CM}\equiv\frac{{\displaystyle {\displaystyle M_{P}\sum_{i=1}^{Z}\vec{x}_{iP}+M_{N}\sum_{i=1}^{N}\vec{x}_{iN}}}}{ZM_{P}+NM_{N}}\label{eq:(75)}\end{equation} are the CM coordinates of the nucleus defined in the atomic CM reference system.The next two terms represent the Hamiltonian of the single atomic 1S electron with $\varrho_{el}(\vec{x})$ as the electric charge density profile of the nucleus. In the non-relativistic approximation \[ [-\frac{\hbar^{2}}{2m_{e}}\nabla_{y}^{2}(e)-Z\alpha\int d^{3}\vec{x}\frac{\varrho_{el}(\vec{x})}{|\vec{x}-\vec{y}|}]\varphi_{1S\sigma}(\vec{y})=\] \begin{equation} =-|\varepsilon_{1S}|\varphi_{1S\sigma}(\vec{y})\label{eq:(76)}\end{equation} The last term in definition (73) stands for the Dirac Hamiltonian of massive neutrinos, with the Bjorken-Drell conventions and definitions {[}20]. We need actually only the positive energy solutions $u_{n\vec{p}\sigma}(\vec{y})$: \[ [-i\hbar\vec{\alpha}(n).\vec{\nabla}_{y}(n)+\beta(n)m_{n}c^{2}]u_{n\vec{p}\sigma}(\vec{y})=\] \begin{equation} =\sqrt{|\vec{p}|^{2}c^{2}+m_{n}^{2}c^{4}}u_{n\vec{p}\sigma}(\vec{y})\label{eq:(77)}\end{equation} The term $V_{eff}$ in definition (72) is the electroweak interaction as originally suggested by the Standard Model {[}19] but adapted here (as an effective interaction) to the physical conditions inside nuclei {[}21]. As such it must be used with careful insight and according to standard ideas on modern renormalization techniques and effective field theories {[}21]. Assume that at time $t\leq0$ the initial wavefunction of an individual H-like ion at rest in its CM system is \[ \Psi_{I,1S\sigma}(\{\vec{x},\vec{x}_{M}\},\vec{y};t)\equiv<\{\vec{x},\vec{x}_{M}\},\vec{y}|I,1S\sigma;t>=\] \begin{equation} =\Phi_{I}(\{\vec{x},\vec{x}_{M}\};Z,N)\varphi_{1S\sigma}(\vec{y})e^{-iE_{I1S}t}\;\;\;\; t\leq0\label{eq:(78)}\end{equation} where $\Phi_{I}(\{\vec{x},\vec{x}_{M}\};Z,N)$ is the internal wavefunction of the parent nucleus. The total available energy is $E_{I1S}$: \begin{equation} E_{I1S}=U_{0}+{\displaystyle (}M_{I}(Z,N)+m_{e})c^{2}-|\varepsilon(1S)|\label{eq:(79)}\end{equation} where M$_{I}$(Z,N) is the mass of the parent nucleus. However, this state is not really a stationary state, because of the background weak couplings of quarks inside the nucleons to virtual flavour changing heavy electroweak bosons $W^{\pm},Z^{0}$ {[}21]. This is represented by the (hermitean) last term $V_{eff}$ on the rhs of the effective Hamiltonian (\ref{eq:(72)}) for allowed transitions {[}21]: \begin{equation} V_{eff}(\vec{x},\vec{y})=-\frac{G_{F}}{\sqrt{2}}\delta(\vec{x}-\vec{y})(V_{F}(n)+V_{GT}(n))\label{eq:(80)}\end{equation} where $G_{F}=8.7\times10^{-5}MeV.fm^{2}$ is the Fermi constant , $V_{F}(V_{GT})$ are related to the standard Fermi (Gamov-Teller) interaction operators. Let us make the following ansatz for the time-dependent quantum state that should be good enough for sufficiently $weak$ $interactions$ {[}10]: \[ |I,1S\sigma;t>=\] \begin{equation} =exp(-i(E_{I1S}-i\frac{\Gamma_{I1S}}{2})t)|I,1S\sigma;t=0>\;\;\;\Gamma_{I1S}>0\label{eq:(81)}\end{equation} It can be easily shown that {[}10] \[ \Gamma_{I1S}=2\pi{\displaystyle \sum_{F}{\displaystyle \sum_{n,\sigma_{n}}\int d^{3}\vec{p}_{n}}}*\delta(E_{I1S}-E_{nF}(p_{n}))*\] \begin{equation} *|<F(-\vec{p}_{n}),n\vec{p}_{n}\sigma_{n}|V_{eff}|I,1S\sigma>|^{2}\label{eq:(82)}\end{equation} in agreement with Fermi's Golden Rule {[}10]. Furthermore \[ E_{nF}(p_{n})=\] \[ =U_{0}+{\displaystyle (}M_{F}(Z-1,N+1)c^{2}+\frac{|\vec{p}_{n}|^{2}}{2M_{F}}+\] \begin{equation} +\sqrt{|\vec{p}|_{n}^{2}c^{2}+m_{n}^{2}c^{4}}\label{eq:(83)}\end{equation} in obvious notation. The leptonic coordinates in the atomic CM system are related to $\vec{x}_{CM}$ by the definition \begin{equation} m_{e}\vec{y}+M_{I}\vec{x}_{CM}=0\label{eq:(84)}\end{equation} in the initial state and \begin{equation} m_{n}\vec{y}+M_{F}\vec{x}_{CM}=0\label{eq:(85)}\end{equation} in the final state. We have used momentum conservation when writing down the expression for the recoil energy of the final nucleus. The Bohm survival probability for an individual parent at any time t>0 is defined to be \begin{equation} P_{I1S}(t)=<I,1S\sigma;t|I,1S\sigma;t>=exp(-\Gamma_{I1S}t)P_{I1S}(0)\label{eq:(86)}\end{equation} This must satisfy the probability conservation law (25). So it does, to first order, since $V_{eff}$ is by construction hermitean. Let us define as usual \[ \Psi_{I1S\sigma}(\{\vec{x},\vec{x}_{CM}\},\vec{y};t)=R_{I1S\sigma}(\{\vec{x},\vec{x}_{CM}\},\vec{y};t)*\] \begin{equation} =*exp(iS_{I1S\sigma}(\{\vec{x},\vec{x}_{CM}\},\vec{y};t)\label{eq:(87)}\end{equation} where \[ R_{I1S\sigma}(\{\vec{x},\vec{x}_{CM}\},\vec{y};t)=\] \begin{equation} =exp(-\frac{\Gamma_{I1S}}{2}t)|\Psi_{I,1S\sigma}(\{\vec{x},\vec{x}_{CM}\},\vec{y},0)|\label{eq:(88)}\end{equation} \begin{equation} P_{I1S}(t)=R_{I1S\sigma}^{2}(\{\vec{x},\vec{x}_{CM}\},\vec{y};t)\label{eq:(89)}\end{equation} Any built-in degeneracies of $\Psi_{I1S\sigma}(\{\vec{x},\vec{x}_{CM}\},\vec{y};0)$ would imply that its phase may not be zero. \subsubsection{Particle histories} In the context of the ontological BP under discussion a good question (though one that is never asked by any textbook!) could be: what are precisely the beable particle histories, i.e. the positions $and$ velocities of those inside our decaying initial ion and the outgoing neutrino, at any time after t=0, until eventually some entirely new conditions - such as an observation! - occur {[}2]? In the context of the BP of Quantum Theory, not only the question does make perfect sense, but can also be easily answered in quantitative detail, at least in perturbation theory as sketched above (even if, for the present, we are quite unable to plan, let alone carry out, any experiments based on such ideas !). Especially interesting is any possibly {}``measurable quantity'', e.g. the momentum of the outgoing beable neutrino n or rather, the recoil energy of the daughter nucleus. The precise answer given by the BP is (section 1): \[ \vec{p}_{n}(\{\vec{X}(t),\vec{X}_{CM}(t)\},\vec{Y}(t);t)=\] \begin{equation} =\{\vec{\nabla}_{\vec{y}}S_{I1S\sigma}(\{\vec{x},\vec{x}_{CM}(t)\},\vec{y};t)\}_{\{\vec{x}\}\vec{x}_{CM},\vec{y}=\{\vec{X}(t)\},\vec{X}_{CM}(t),\vec{Y}(t)}\label{eq:(90)}\end{equation} Immediate confrontation with existing experimental material on this question is however possible ( Postulate 2 ,section 1) if one first carries out an appropriate stochastic averaging over all initial positions and velocities of the (entangled) beable nucleons and leptons initially involved : \[ <\vec{p}_{n}(\{\vec{X}(t),\vec{X}_{CM}(t)\},\vec{Y}(t);t)>_{EA}=\int\int\int[R_{I1S\sigma}^{2}(\{\vec{x}\}\vec{x}_{CM},\vec{y};t))*\] \begin{equation} *\vec{\nabla}_{y}(n)S_{I1S\sigma}(\{\vec{x}\}\vec{x}_{CM},\vec{y};t)\}_{\{\vec{x}\}\vec{x}_{CM},\vec{y}=\{\vec{X}(t)\}\vec{X}_{CM}(t),\vec{Y}(t)}\label{eq:(91)}\end{equation} The integrations are to be carried out over all histories of all participating beable particles. As mentioned in section 2, this is guaranteed to be in numerical agreement with the textbook answer at t=0. As also mentioned in section 2, it will then be true at any time : \[ <(-i\vec{\nabla}_{y}(n)>=\] \[ =\frac{1}{2}\int\int[\Psi_{I1S\sigma}^{*}(-i\vec{\nabla}_{y}(n)\Psi_{I1S\sigma})+\Psi_{I1S\sigma}(-i\vec{\nabla}_{y}(n)\Psi_{I1S\sigma})^{*}]\equiv\] \begin{equation} \equiv<\vec{p}_{n}(\{\vec{X}(t)\}\vec{X}_{CM}(t),\vec{Y}(t);t)>_{EA}\label{eq:(92)}\end{equation} We emphasize that the word {}``time'' means here literally {}``time taken by a particle, say the neutrino n, to be guided from its position $\vec{X}(t)$ (and velocity $\vec{V}(t)$) to position $\vec{X}'(t')$ (and velocity $\vec{V}'(t')$)''. If a {}``distance'' L is {}``measured'' from the neutrino position to the CM of the parent atom , then at time t>0 the distance L can be traded with time t elapsed because to sufficient accuracy $t=L/c.$ Hence, the time label t in (\ref{eq:(91)}) can be replaced by the space label L, and so all talk about a possibly vague {}``time'' can be avoided {[}22]. The point being, {}``time'' here is directly linked to the physical motions of material particles. \section{Summary and conclusions} The ontological formulation of non-relativistic Quantum Mechanics, originally initiated by Louis de Broglie in the twenties but independently recreated and reformulated in the fifties and sixties by D. Bohm and his co-workers, is briefly presented and its explanatory powers are further illustrated with examples from new fields of research in contemporary Nuclear and Particle Physics, fields that were not mature during the lifetime of those pioneers. It is argued that the reformulation in question has indeed the capability of not only by-passing all the so-called paradoxes and ambiguities characterizing the official (textbook) epistemological formulations of Quantum Theory, but more importantly, lead to genuine testable $physical\;\; explanations$ (and not only to highly abstract mathematical representations) of actual physical phenomena, at the quantum level of accuracy. This is achieved without any need of arbitrary breaks of continuity between the so-called {}``classical'' and {}``quantal'' modes of description that is so vital for all textbook formulations. Furthermore, this novel ontological formulation can lead naturally to proposals for new type of experimentation, thus in principle greatly extending the predictive powers of present day Quantum Theory. One can thus open up for new ways of thinking and trying fresh approaches to the age-old fundamental problems of Relativistic Quantum Field Theory and Theoretical Cosmology.

Introduction ============ The T-Line TL-200 (Tensys Medical Inc., San Diego, CA, USA) is a noninvasive arterial blood pressure (BP) monitoring system allowing continuous beat-to-beat monitoring of systolic arterial pressure (SAP), mean arterial pressure (MAP), and diastolic arterial pressure (DAP). It provides a real-time BP waveform like that obtained using an arterial catheter for BP monitoring. The aim of this study was to compare BP measurements obtained using the T-Line TL-200 system with simultaneous invasive BP measurements using a femoral arterial catheter in unselected critically ill medical patients. Methods ======= In 28 patients treated in a medical ICU of a German university hospital, BP values were simultaneously obtained using a femoral arterial catheter and the T-Line TL-200 device. All recorded data were included in the final analysis. For comparison of BP measurements, Bland-Altman analysis accounting for repeated measurements was performed. Results ======= A total of 76,826 pairs of BP measurements (each consisting of SAP, MAP, and DAP) were analyzed. For MAP, Bland-Altman analysis revealed a mean difference of +0.47 mmHg (95% limits of agreement: -16.53 to +17.46 mmHg). For SAP and DAP, the bias and 95% limits of agreement were -9.01 mmHg (-37.47 to +19.45 mmHg) and +5.22 mmHg (-13.50 to +23.94 mmHg), respectively. Conclusion ========== The T-Line TL-200 system allows determination of MAP with a satisfactory agreement when compared to invasive assessment of MAP using a femoral arterial catheter in unselected critically ill medical patients. Higher mean differences and 95% limits of agreement for SAP and DAP measurements might be explainable by limited comparability of central (femoral) and peripheral (radial) SAP and DAP measurements.

\section{Introduction} Understanding how people's opinions are received and evaluated is a fundamental problem that arises in many domains, such as in marketing studies of the impact of reviews on product sales, or in political science models of how support for a candidate depends on the views he or she expresses on different topics. This issue is also increasingly important in the user interaction dynamics of large participatory Web sites. Here we develop a framework for understanding and modeling how opinions are evaluated within on-line communities. The problem is related to the lines of computer-science research on opinion, sentiment, and subjective content \citep{Pang+Lee:08a}, but with a crucial twist in its formulation that makes it fundamentally distinct from that body of work. Rather than asking questions of the form ``What did Y think of X?'', we are asking, ``What did Z think of Y's opinion of X?'' Crucially, there are now three entities in the process rather than two. Such three-level concerns are widespread in everyday life, and integral to any study of opinion dynamics in a community. For example, political polls will more typically ask, ``How do you feel about Barack Obama's position on taxes?'' than ``How do you feel about taxes?'' or ``What is Barack Obama's position on taxes?'' (though all of these are useful questions in different contexts). Also, Heider's theory of {\em structural balance} in social psychology seeks to understand subjective relationships by considering sets of three entities at a time as the basic unit of analysis. But there has been relatively little investigation of how these three-way effects shape the dynamics of on-line interaction, and this is the topic we consider here. \vspace*{-.1in} \xhdr{The Helpfulness of Reviews} The evaluation of opinions takes place at very large scales every day at a number of widely-used Web sites. Perhaps most prominently it is exemplified by one of the largest online e-commerce providers, Amazon.com, whose website includes not just product reviews contributed by users, but also evaluations of the helpfulness of these reviews. (These consist of annotations that say things like, ``26 of 32 people found the following review helpful'', with the corresponding data-gathering question, ``Was this review helpful to you?'') Note that each review on Amazon thus comes with both a {\em star rating\xspace} --- the number of number of stars it assigns to the product --- and a {\em \helpfulness vote\xspace } --- the information that $a$ out of $b$ people found the review itself helpful. (See Figure \ref{fig:plagiarism_example} for two examples.) This distinction reflects precisely the kind of opinion evaluation we are considering: in addition to the question ``what do you think of book X?'', users are also being asked ``what do you think of user Y's review of book X?'' A large-scale snapshot of Amazon reviews and {\hvote}s\xspace will form the central dataset in our study, as detailed below. The factors affecting human helpfulness\xspace evaluations are not well understood. There has been a small amount of work on automatic determination of helpfulness\xspace, treating it as a classification or regression problem with Amazon {\hvote}s\xspace providing labeled data \citep{Kim+al:06a,Ghose+Ipeirotis:07a,Liu+al:07a}. Some of this research has indicated that the {\hvote}s\xspace of reviews are not necessarily strongly correlated with certain measures of review quality; for example, Liu et al. found that when they provided independent human annotators with Amazon review text and a precise specification of helpfulness in terms of the thoroughness of the review, the annotators' evaluations differed significantly from the {\hvote}s\xspace observed on Amazon. All of this suggests that there is in fact a subtle relationship between two different meanings of ``helpfulness'': helpfulness in the narrow sense --- does this review help you in making a purchase decision? --- and helpfulness ``in the wild,'' as defined by the way in which Amazon users evaluate each others' reviews in practice. It is a kind of dichotomy familiar from the design of participatory Web sites, in which a presumed design goal --- that of highlighting reviews that are helpful in the purchase process --- becomes intertwined with complex social feedback mechanisms. If we want to understand how these definitions interact with each other, so as to assist users in interpreting helpfulness evaluations, we need to elucidate what these feedback mechanisms are and how they affect the observed outcomes. \xhdr{The present work: Social mechanisms underlying helpfulness evaluation} In this paper, we formulate and assess a set of theories that govern the evaluation of opinions, and apply these to a dataset consisting of over four million reviews of roughly 675,000 books on Amazon's U.S. site, as well as smaller but comparably-sized corpora from Amazon's U.K., Germany, and Japan sites. The resulting analysis provides a way to distinguish among competing hypotheses for the social feedback mechanisms at work in the evaluation of Amazon reviews: we offer evidence against certain of these mechanisms, and show how a simple model can directly account for a relatively complex dependence of helpfulness on review and group characteristics. We also use a novel experimental methodology that takes advantage of the phenomenon of review ``plagiarism''\xspace to control for the text content of the reviews, enabling us to focus exclusively on factors outside the text that affect helpfulness\xspace evaluation. In our initial exploration of non-textual factors that are correlated with helpfulness evaluation on Amazon, we found a broad collection of effects at varying levels of strength.\footnote{For example, on the U.S. Amazon site, we find that reviews from authors with addresses in U.S. territories outside the 50 states get consistently lower {\hvote}s\xspace. This is a persistent effect whose possible bases lie outside the scope of the present paper, but it illustrates the ways in which non-textual factors can be correlated with helpfulness evaluations. Previous work has also noted that {\em longer} reviews tend to be viewed as more helpful; ultimately it is a definitional question whether review length is a textual or non-textual feature of the review.} A significant and particularly wide-ranging set of effects is based on the relationship of a review's star rating\xspace to the {\smark}s\xspace of other reviews for the same product. We view these as fundamentally {\em social} effects, given that they are based on the relationship of one user's opinion to the opinions expressed by others in the same setting. \footnote{A contrarian might put forth the following non-socially-based alternative hypothesis: the people evaluating review helpfulness\xspace are considering actual product quality rather than other reviews, but aggregate opinion happens to coincide with objective product quality. This hypothesis is not consistent with our experimental results. However, in future work it might be interesting to directly control for product quality.} Research in the social sciences provides a range of well-studied hypotheses for how social effects influence a group's reaction to an opinion, and these provide a valuable starting point for our analysis of the Amazon data. In particular, we consider the following three broad classes of theories, as well as a fourth straw-man hypothesis that must be taken into account. \begin{itemize} \item[(i)] {\em The \conformityhyp.} One hypothesis, with roots in the social psychology of conformity \cite{Bond+Smith:96a}, holds that a review is evaluated as more helpful when its star rating\xspace is closer to the consensus star rating\xspace for the product --- for example, when the number of stars it assigns is close to the average number of stars over all reviews. \item[(ii)] {\em The \biashyp.} Alternately, one could hypothesize that when a user considers a review, he or she will rate it more highly if it expresses an opinion that he or she agrees with.\footnote{Such a principle is also supported by structural balance considerations from social psychology; due to the space limitations, we omit a discussion of this here.} Note the contrasts and similarities with the previous hypothesis: rather than evaluating whether a review is close to the mean opinion, a user evaluates whether it is close to their own opinion. At the same time, one might expect that if a diverse range of individuals apply this rule, then the overall helpfulness evaluation could be hard to distinguish from one based on conformity; this issue turns out to be crucial, and we explore it further below. \item[(iii)] {\em The brilliant-but-cruel hypothesis.} The name of this hypothesis comes from studies performed by \citet{Amabile:83a} that support the argument that ``negative reviewers [are] perceived as more intelligent, competent, and expert than positive reviewers.'' One can recognize everyday analogues of this phenomenon; for example, in a research seminar, a dynamic may arise in which the nastiest question is consistently viewed as the most insightful. \item[(iv)] {\em The \contenthypprefix straw-man hypothesis.} Finally, there is a challenging methodological complication in all these styles of analysis: without specific evidence, one cannot dismiss out of hand the possibility that helpfulness is being evaluated purely based on the textual content of the reviews, and that these non-textual factors are simply correlates of textual quality. In other words, it could be that people who write long reviews, people who assign particular {\smark}s\xspace in particular situations, and people from Massachusetts all simply write reviews that are textually more helpful --- and that users performing helpfulness evaluations are simply reacting to the text in ways that are indirectly reflected in these other features. Ruling out this hypothesis requires some means of controlling for the text of reviews while allowing other features to vary, a problem that we also address below. \end{itemize} We now consider how data on {\smark}s\xspace and {\hvote}s\xspace can support or contradict these hypotheses, and what it says about possible underlying social mechanisms. \xhdr{Deviation from the mean} A natural first measure to investigate is the relationship of a review's star rating\xspace to the mean star rating\xspace of all reviews for the product; this, for example, is the underpinning of the conformity hypothesis. With this in mind, let us define the {\em helpfulness ratio\xspace} of a review to be the fraction of evaluators who found it to be helpful (in other words, it is the fraction $a/b$ when $a$ out of $b$ people found the review helpful), and let us define the {\em product average} for a review of a given product\xspace to be the average star rating\xspace given by {\em all} reviews of that product\xspace. We find (Figure~\ref{fig:unsigned-diff-novar}) that the median helpfulness ratio\xspace of reviews decreases monotonically as a function the absolute difference between their star rating\xspace and the product average. (The same trend holds for other quantiles.) In fact the dependence is surprisingly smooth, with even seemingly subtle changes in the differences from the average having noticeable effects. This finding on its own is consistent with the conformity hypothesis: reviews in aggregate are deemed more helpful when they are close to the product average. However, a closer look at the data raises complications, as we now see. First, to assess the brilliant-but-cruel hypothesis, it is natural to look not at the absolute difference between a review's star rating\xspace and its product average, but at the {\em signed difference}, which is positive or negative depending on whether the star rating\xspace is above or below the average. Here we find something a bit surprising (Figure~\ref{fig:signed-diff-novar}). Not only does the median helpfulness as a function of signed difference fall away on both sides of $0$; it does so {\em asymmetrically}: slightly negative reviews are punished more strongly, with respect to helpfulness\xspace evaluation, than slightly positive reviews. In addition to being at odds with the brilliant-but-cruel hypothesis for Amazon reviews, this observation poses problems for the conformity hypothesis in its pure form. It is not simply that closeness to the average is rewarded; among reviews that are slightly away from the mean, there is a bias toward overly positive ones. \xhdr{Variance and individual bias} One could, of course, amend the conformity hypothesis so that it becomes a ``conformity with a tendency toward positivity'' hypothesis. But this would beg the question; it wouldn't suggest any underlying mechanism for where the favorable evaluation of positive reviews is coming from. Instead, to look for such a mechanism, we consider versions of the individual-bias hypothesis. Now, recall that it can be difficult to distinguish conformity effects from individual-bias effects in a domain such as ours: if people's opinions (i.e., {\smark}s\xspace) for a product come from a single-peaked\xspace distribution with a maximum near the average, then the composite of their individual biases can produce overall {\hvote}s\xspace that look very much like the results of conformity. We therefore seek out subsets of the products on which the two effects might be distinguishable, and the argument above suggests starting with products that exhibit high levels of individual variation in {\smark}s\xspace. In particular, we associate with each product the {\em variance} of the {\smark}s\xspace assigned to it by all its reviews. We then group {\product}s\xspace by variance, and perform the signed-difference analysis above on sets of {\product}s\xspace having fixed levels of variance. We find (Figure~\ref{fig:signed-diff-var}) that the effect of signed difference to the average changes smoothly but in a complex fashion as the variance increases. The role of variance can be summarized as follows. \begin{itemize} \item When the variance is very low, the reviews with the highest helpfulness ratios are those with the average star rating\xspace. \item With moderate values of the variance, the reviews evaluated as most helpful are those that are slightly above the average star rating\xspace. \item As the variance becomes large, reviews with {\smark}s\xspace both above {\em and} below the average are evaluated as more helpful than those that have the average star rating\xspace (with the positive reviews still deemed somewhat more helpful). \end{itemize} These principles suggest some qualitative ``rules'' for how --- all other things being equal --- one can seek good helpfulness evaluations in our setting: With low variance go with the average; with moderate variance be slightly above average; and with high variance avoid the average. This qualitative enumeration of principles initially seems to be fairly elaborate; but as we show in Section~\ref{sec:gaussians}, all these principles are consistent with a simple model of individual bias in the presence of controversy. Specifically, suppose that opinions are drawn from a mixture of two single-peaked\xspace distributions --- one with larger mixing weight whose mean is above the overall mean of the mixture, and one with smaller mixing weight whose mean is below it. Now suppose that each user has an opinion from this mixture, corresponding to their own personal score for the product, and they evaluate reviews as helpful if the review's star rating\xspace is within some fixed tolerance of their own. We can show that in this model, as variance increases from $0$, the reviews evaluated as most helpful are initially slightly above the overall mean, and eventually a ``dip'' in helpfulness\xspace appears around the mean. Thus, a simple model can in principle account for the fairly complex series of effects illustrated in Figure~\ref{fig:signed-diff-var}, and provide a hypothesis for an underlying mechanism. Moreover, the effects we see are surprisingly robust as we look at different national Amazon sites for the U.K., Germany, and Japan. Each of these communities has evolved independently, but each exhibits the same set of patterns. The one non-trivial and systematic deviation from the pattern among these four countries is in the analogue of Figure~\ref{fig:signed-diff-var} for Japan: as with the other countries, a ``dip'' appears at the average in the high-variance case, but in Japan the portion of the curve {\em below} the average is higher. This would be consistent with a version of our two-distribution individual-bias model in which the distribution below the average has higher mixing weight --- representing an aspect of the brilliant-but-cruel hypothesis in this individual-bias framework, and only for this one national version of the site. \xhdr{Controlling for text: Taking advantage of ``plagiarism''\xspace} Finally, we return to one further issue discussed earlier: how can we offer evidence that these non-textual features aren't simply serving as correlates of review-quality features that are intrinsic to the text itself? In other words, are there experiments that can address the \contenthypprefix straw man hypothesis above? To deal with this, we make use of rampant ``plagiarism''\xspace and duplication of reviews on Amazon.com (the causes and implications of this phenomenon are beyond the scope of this paper). This is a fact that has been noted and studied by earlier researchers \cite{David+Pinch:06a}, and for most applications it is viewed as a pathology to be remedied. But for our purposes, it makes possible a remarkably effective way to control for the effect of review text. Specifically, we define a {\em ``plagiarized''\xspace pair} of reviews to be two reviews of different products with near-complete textual overlap, and we enumerate the several thousand instances of plagiarized pairs on Amazon. (We distinguish these from reviews that have been cross-posted by Amazon itself to different versions of the same product.) Not only are the two members of a ``plagiarized''\xspace pair associated with different products; very often they also have significantly different {\smark}s\xspace and are being used on products with different averages and variances. (For example, one copy of the review may be used to praise a book about the dangers of global warming while the other copy is used to criticize a book that is favorable toward the oil industry). We find significant differences in the {\hratio}s\xspace within plagiarized pairs, and these differences confirm many of the the effects we observe on the full dataset. Specifically, within a ``plagiarized''\xspace pair, the copy of the review that is closer to the average gets the higher helpfulness ratio\xspace in aggregate. Thus the widespread copying of reviews provides us with a way to see that a number of social feedback effects --- based on the score of a review and its relation to other scores --- lead to different outcomes even for reviews that are textually close to identical. \xhdr{Further related work} We also mention some relevant prior literature that has not already been discussed above. The role of social and cognitive factors in purchasing decision-making has been extensively studied in psychology and marketing \citep{Cialdini+Goldstein:04a,Escalas+Bettman:05a,Fitzsimons+al:02a,Underhill:99a}, recently making use of brain imaging methodology \citep{Knutson+al:07}. Characteristics of the distribution of review {\smark}s\xspace (which differ from {\hvote}s\xspace) on Amazon and related sites have been studied previously \citep{Chevalier+Mayzlin:06a,Hu+Pavlou+Zhang:06a,Wu+Huberman:08c}. Categorizing text by quality has been proposed for a number of applications \citep{Jindal+Liu:07a,Sen+al:07a,Agichtein+al:08a,Harper+al:08a}. Additionally, our notion of variance is potentially related to the idea that people play different roles in on-line discussion \citep{Welser+al:07a}. \section{Data} \label{sec:data} Our experiments employed a dataset of over 4 million Amazon.com book reviews (corresponding to roughly 675,000 books), of which more than 1 million received at least 10 {\hvote}s\xspace each. We made extensive use of the Amazon Associates Webservice (AWS) API to collect this data.\footnote{We used the AWS API version 2008-04-07. Documentation is available at \url{http://docs.amazonwebservices.com/AWSECommerceService/2008-04-07/DG/} .} We describe the process in this section, with particular attention to measures we took to avoid sample bias. We would ideally have liked to work with {\em all} book reviews posted to Amazon. However, one can only access reviews via queries specifying particular books by their Amazon product ID, or ASIN (which is the same as ISBN for most books), and we are not aware of any publicly available list of all Amazon book ASINs. However, the API allows one to query for books in a specific category (called a {\em browse-node} in AWS parlance and corresponding to a section on the Amazon.com website), and the best-selling titles up to a limit of 4000 in each browse-node can be obtained in this way. To create our initial list of books, therefore, we performed queries for all 3855 categories three levels deep in the Amazon browse-node hierarchy (actually a directed acyclic graph) rooted at ``Books$\rightarrow\,$Subjects''. An example category is {\newcommand{$\rightarrow\,$}{$\rightarrow\,$} \newcommand{\abn}[3]{\emph{#1$\rightarrow\,$#2$\rightarrow\,$#3}\xspace} \abn{Children's Books}{Animals}{Lions, Tigers \& Leopards}. } These queries resulted in the initial set of 3,301,940 books, where we count books listed in multiple categories only once. We then performed a book-filtering step to deal with ``cross-posting'' of reviews across versions. When Amazon carries different versions of the same item --- for example, different editions of the same book, including hardcover and softcover editions and audio-books --- the reviews written for all versions are merged and displayed together on each version's product page and likewise returned by the API upon queries for any individual version.% \footnote{At the time of data collection, the API did not provide an option to trace a review to a particular edition for which it was originally posted, despite the fact that the Web store front-end has included such links for quite some time.} This means that multiple copies of the same review exist for ``mechanical'', as opposed to user-driven, reasons.\footnote{We make use of human-instigated review copying later in this study.} To avoid including mechanically-duplicated reviews, we retained only one of the set of alternate versions for each book (the one with the most complete metadata). The above process gave us a list of 674,018 books for which we retrieved reviews by querying AWS. Although AWS restricts the number of reviews returned for any given product query to a maximum of 100, it turned out that 99.3\% of our books had 100 or fewer reviews. In the case of the remaining 4664 books, we chose to retrieve the 100 {\em earliest} reviews for each product to be able to reconstruct the information available to the authors and readers of those reviews to the extent possible. (Using the earliest reviews ensures the reproducibility of our results, since the 100 earliest reviews comprise a static set, unlike the 100 most helpful or recent reviews.) As a result, we ended up with 4,043,103 reviews; although some reviews were not retrieved due to the 100-reviews-per-book API cap, the number of missing reviews averages out to roughly just one per ASIN queried. Finally, we focused on the 1,008,466 reviews that had at least 10 {\hvote}s\xspace each. The size of our dataset compares favorably to that of collections used in other studies looking at {\hvote}s\xspace: \citet{Liu+al:07a} used about 23,000 digital camera reviews (of which a subset of around 4900 were subsequently given new {\hvote}s\xspace and studied more carefully); \citet{Zhang+Varadarajan:06a} used about 2500 reviews of electronics, engineering books, and PG-13 movies after filtering out duplicate reviews and reviews with no more than 10 {\hvote}s\xspace; \citet{Kim+al:06a} used about 26,000 MP3 and digital-camera reviews after filtering of duplicate versions and duplicate reviews and reviews with fewer than 5 {\hvote}s\xspace; and \citet{Ghose+Ipeirotis:08a} considered ``all reviews since the product was released into the market'' (no specfic number is given) for about 400 popular audio and video players, digital cameras, and DVDs. \section{Effects of \headerformat{deviation\xspace} from average and \headerformat{variance\xspace}} \label{sec:variance} Several of the \hypotheses that we have described concern the relative position of an opinion about an entity vis-\`a-vis the average opinion about that entity. We now turn, therefore, to the question of how the helpfulness ratio\xspace of a review depends on its star rating\xspace's deviation\xspace from the average star rating\xspace for all reviews of the same book. According to the \conformityhyp, the helpfulness ratio\xspace should be {\em lower} for reviews with {\smark}s\xspace \emph{ either above or below} the product\xspace average, whereas the \brillianthyp translates to the ``asymmetric'' prediction that the helpfulness ratio\xspace should be {\em higher} for reviews with {\smark}s\xspace \emph{below} the product\xspace average than for overly positive reviews. (No specific predictions for helpfulness ratio\xspace vis-\`a-vis product\xspace average is made by either the \biasprefix or \contenthyp without further assumptions about the distribution of individual opinions or text quality.) \xhdr{Defining the average} For a given review, let the {\em computed \product-average {\smark}\xspace} (abbreviation: {\em computed star average\xspace}) be the average star rating\xspace as computed over all reviews of that product\xspace in our dataset. {This differs in principle from the {\em Amazon-displayed \product-average {\smark}\xspace} (abbreviation: {\em displayed star average\xspace}), the ``Average Customer Review'' score that Amazon itself displayed for the book at the time we downloaded the data. One reason for the difference is that Amazon rounds the displayed star average\xspace to the nearest half-star (e.g., 3.5 or 4.0) --- but for our experiments it is preferable to have a greater degree of resolution. Another possible source of difference is the very small (0.7\%) fraction of books, mentioned in Section \ref{sec:data}, for which the entire set of reviews could not be obtained via AWS: the displayed star average\xspace would be partially based on reviews that came later than the first 100 and which would thus not be in our dataset. However, the mean absolute difference between the computed star average\xspace when rounded to the nearest half-star (0.5 increment) and the displayed star average\xspace is only 0.02. Note that both scores can differ from the ``Average Customer Review'' score that Amazon displayed {\em at the time a helpfulness evaluator\xspace provided their \helpfulness vote\xspace}, since this time might pre-date some of the reviews for the book that are in our dataset (and hence that Amazon based its displayed star average\xspace on). In the absence of timestamps on {\hvote}s\xspace, this is not a factor that can be controlled for. \xhdr{Deviation\xspace experiments} \begin{figure}[t] \begin{center} \includegraphics[width=3in,viewport= 50 200 550 600,clip]{unsigned-difference-step05_median} \caption{ \label{fig:unsigned-diff-novar} Helpfulness ratio\xspace declines with the absolute value of a review's deviation\xspace from the computed star average\xspace; this behavior is predicted by the \conformityhyp but not ruled out by the other hypotheses. \newline \hspace*{.1in} The line segments within the bars (connected by the descending line) indicate the median helpfulness ratio\xspace; the bars depict the helpfulness ratio\xspace's second and third quantiles. \newline \hspace*{.1in}Throughout, grey bars indicate that the amount of data at that $x$ value represents $.1\%$ or less of the data depicted in the plot. } \end{center} \end{figure} We first check the prediction of the \conformityhyp that the helpfulness ratio\xspace of a review will vary inversely with the absolute value of the difference between the review's star rating\xspace and the computed \product-average {\smark}\xspace --- we call this difference the review's {\em deviation\xspace}. Figure \ref{fig:unsigned-diff-novar} indeed shows a very strong inverse correlation between the median helpfulness ratio\xspace and the absolute deviation\xspace , as predicted by the \conformityhyp. However, this data does not completely disprove the \brillianthyp, since for a given absolute deviation\xspace $|x|>0$, it could conceivably happen that reviews with positive deviations\xspace $|x|$ (i.e. more favorable than average) could have much worse {\hratio}s\xspace than reviews with negative deviation\xspace $-|x|$, thus dragging down the median helpfulness ratio\xspace. Rather, to directly assess the \brillianthyp, we must consider {\em signed {\deviation}\xspace}, not just absolute deviation\xspace. \begin{figure}[ht] \begin{center} \includegraphics[width=3in, viewport= 50 200 550 600,clip]{signed-difference-step05_median} \caption{\label{fig:signed-diff-novar} The dependence of helpfulness ratio\xspace on a review's signed {\deviation}\xspace from average is inconsistent with both the \brilliantprefix and, because of the asymmetry, the \conformityhyp. } \end{center} \end{figure} Surprisingly, the effect of signed {\deviation}\xspace on median helpfulness ratio\xspace, depicted in the ``Christmas-tree'' plot of Figure \ref{fig:signed-diff-novar}, turns out to be different from what {\em either} hypothesis would predict. The \brillianthyp clearly does not hold for our data: among reviews with the same absolute deviation\xspace $|x|>0$, the relatively positive ones (signed {\deviation}\xspace $|x|$) generally have a higher median helpfulness ratio\xspace than the relatively negative ones (signed {\deviation}\xspace of $-|x|$), as depicted by the positive slope of the {green dotted} lines connecting ($-|x|$,$|x|$) pairs of datapoints. But Figure \ref{fig:signed-diff-novar} {\em also} presents counter-evidence for the \conformityhyp, since that hypothesis incorrectly predicts that the connecting lines would be horizontal. \begin{figure*}[t] \begin{center} \includegraphics[height=4in, viewport= 50 200 550 600,clip]{signed-difference_median_variances} \caption{\label{fig:signed-diff-var} As the variance of the {\smark}s\xspace of reviews for a particular product increases, the median helpfulness ratio\xspace curve becomes two-humped and the helpfulness ratio\xspace at signed deviation\xspace 0 (indicated in red) no longer represents the unique global maximum. There are non-zero signed {\deviations}\xspace in the plot for variance\xspace 0 because we rounded variance\xspace values to the nearest .5 increment.} \end{center} \end{figure*} To account for Figure \ref{fig:signed-diff-novar}, one could simply impose upon the \conformityhyp an extra ``tendency towards positivity'' factor, but this would be quite unsatisfactory: it wouldn't suggest any underlying mechanism for this factor. So, we turn to the \biashyp instead. In order to distinguish between \conformity effects and \biasprefix effects, we need to examine cases in which individual people's opinions do {\em not} come from exactly the same (single-peaked\xspace, say) distribution; for otherwise, the composite of their \biases could produce {\hratio}s\xspace that look very much like the results of \conformity. One natural place to begin to seek settings in which \bias and \conformity are distinguishable, in the sense just described, is in cases in which there is at least high {\em variance\xspace} in the {\smark}s\xspace. Accordingly, Figure \ref{fig:signed-diff-var} separates {\product}s\xspace by the variance of the {\smark}s\xspace in the reviews for that product\xspace in our dataset. One can immediately observe some striking effects of variance\xspace. First, we see that as variance increases, the ``camel plots'' of Figure \ref{fig:signed-diff-var} go from a single hump to two.\footnote{This is a reversal of nature, where Bactrian (two-humped) camels are more agreeable than one-humped Dromedaries.} We also note that while in the previous figures it was the reviews with a signed {\deviation}\xspace of exactly zero that had the highest {\hratio}s\xspace, here we see that once the variance\xspace among reviews for a product\xspace is 3.0 or greater, the highest {\hratio}s\xspace are clearly achieved for {\product}s\xspace with signed {\deviations}\xspace close to but still noticeably above zero. (The beneficial effects of having a star rating\xspace slightly above the mean are already discernible, if small, at variance 1.0 or so.) % Clearly, these results indicate that variance\xspace is a key factor that any hypothesis needs to incorporate. In Section \ref{sec:gaussians}, we develop a simple \biasprefix model that does so; but first, there is one last hypothesis that we need to consider. \vspace*{.6in} \section{Controlling for text quality: experiments with \headerformat{``plagiarism''\xspace}} As we have noted, our analyses do not explicitly take into account the actual text of reviews. It is not impossible, therefore, that review text quality may be a confounding factor and that our straw-man \contenthyp might hold. Specifically, we have shown that {\hratio}s\xspace appear to be dependent on two key non-textual aspects of reviews, namely, on deviation\xspace from the computed star average\xspace and on star rating\xspace variance\xspace within reviews for a given product; but we have not shown that our results are not simply explained by review quality. Initially, it might seem that the only way to control for text quality is to read a sample of reviews and determine whether the Amazon {\hratio}s\xspace assigned to these reviews are accurate. Unfortunately, it would require a great deal of time and human effort to gather a sufficiently large set of re-evaluated reviews, and human re-evaluations can be subjective; so it would be preferable to find a more efficient and objective procedure. \footnote{ \citet{Liu+al:07a} did perform a manual re-evaluation of 4909 digital-camera reviews, finding that the original {\hratio}s\xspace did not seem well-correlated with the stand-alone comprehensiveness of the reviews. But note that this could just mean that at least some of the original {\hevaluator}s\xspace were using a different standard of text quality (Amazon does not specify any particular standard or definition of helpfulness\xspace). Indeed, the exemplary ``fair'' review quoted by Liu et al. begins, ``There is nothing wrong with the [product] except for the very noticeable delay between pics. [Description of the delay.] Otherwise, [other aspects] are fine for anything from Internet apps to ... print enlarging. It is competent, not spectacular, but it gets the job done at an agreeable price point.'' Liu et al. give this a rating of ``fair'' because it only comments on some of the product's aspects, but the Amazon {\hevaluator}s\xspace gave it a helpfulness ratio\xspace of 5/6, which seems reasonable. Also, reviews might also be evaluated vis-\`a-vis the totality of all reviews, i.e., a review might be rated helpful if it provides complementary information or ``adds value''. For instance, a one-line review that points out a serious flaw in another review could well be considered ``helpful'', but would not rate highly under Liu et al.'s scheme. It is also worth pointing out subjectiveness can remain an issue even with respect to a given text-only evaluation scheme. The two human re-evaluators who used Liu et al.'s \citeyearpar{Liu+al:07a} standard assigned different helpfulness\xspace categories (in a four-category framework) to 619=12.5\% of the reviews considered, indicating that there can be substantial subjectiveness involved in determining review quality even when a single standard is initially agreed upon.} A different potential approach would be to use machine learning to train an algorithm to automatically determine the degree of helpfulness\xspace of each review. Such an approach would indeed involve less human effort, and could thus be applied to larger numbers of reviews. However, we could not draw the conclusions we would want to: any mismatch between the predictions of a trained classifier and the {\hratio}s\xspace observed in held-out reviews could be attributable to errors by the algorithm, rather than to the actions of the Amazon {\hevaluator}s\xspace.\footnote{ \citet{Ghose+Ipeirotis:08a} observe that their trained classifier often performed poorly for reviews of products with ``widely fluctuating'' {\smark}s\xspace, and explain this with an assertion that the Amazon {\hevaluator}s\xspace are not judging text quality in such situations. But there is no evidence provided to dismiss the alternative hypothesis that the {\hevaluator}s\xspace are correct and that, rather, the algorithm makes mistakes because reviews are more complex in such situations and the classifier uses relatively shallow textual features. } We thus find ourselves in something of a quandary: we seem to lack any way to derive a sufficiently large set of objective and accurate re-evaluations of helpfulness\xspace. Fortunately, we can bring to bear on this problem two key insights: \begin{enumerate} \item Rather than try to re-evaluate all reviews for their helpfulness\xspace, we can focus on reviews that are {\em guaranteed} to have very similar levels of textual quality. \item Amazon data contains many instances of nearly-identical reviews \citep{David+Pinch:06a} --- and identical reviews must {\em necessarily} exhibit the same level of text quality. \end{enumerate} Thus, in the remainder of this section, we consider whether the effects we have analyzed above hold on pairs of {\em ``plagiarized''\xspace} reviews. \xhdr{Identifying ``plagiarism''\xspace (as distinct from ``justifiable copying'')} Our choice of the term ``plagiarism''\xspace is meant to be somewhat evocative, because we disregard several types of arguably justifiable copying or duplication in which there is no overt attempt to make the copied review seem to be a genuinely new piece of text; the reason is because this kind of copying does not suit our purposes. However, ill intent cannot and should not be ascribed to the authors of the remaining reviews; we have attempted to indicate this by the inclusion of scare quotes around the term. In brief, we only considered pairs of reviews where the two reviews were posted to different books --- this avoids various types of relatively obvious self-copying (e.g., where an author reposts a review under their user ID after initially posting it anonymously), since obvious copies might be evaluated differently. We next adapted the code of \citet{Sorokina+al:06a} to identify those pairs of reviews of different {\product}s\xspace that have highly similar text. To do so, we needed to decide on a similarity threshold that determines whether or not we deem a review pair to be ``plagiarized''\xspace. A reasonable option would have been to consider only reviews with identical text, which would ensure that the reviews in the pairs had exactly the same text quality\xspace. However, since the reviews in the analyzed pairs are posted for different products, it is normal to expect that some authors modified or added to the text of the original review to make the ``plagiarized''\xspace copy better fit its new context. For this reason, we employed a threshold of $70\%$ or more nearly-duplicate sentences, where near-duplication was measured via the code of \citet{Sorokina+al:06a}.\footnote{\citet{Kim+al:06a}, who also noticed that the phenomenon of review alteration affected their attempts to remove duplicate reviews, used a similar threshold of 80\% repeated bigrams.} This yielded $8,313$ ``plagiarized''\xspace pairs; an example is shown in Figure \ref{fig:plagiarism_example}. Manual inspection of a sample revealed that the review pairs captured by our threshold indeed seem to consist of close copies. \begin{figure}[h] \begin{center} \begin{minipage}{3.35in}{\includegraphics[width=3.35in]{rev1_s}} \end{minipage}\\ \begin{minipage}{3.35in}{\includegraphics[width=3.35in]{rev1}}\hspace*{-1.9in}$\cdots$ \end{minipage}\\ \begin{minipage}{3.35in}{\includegraphics[width=3.35in]{rev_space}} \end{minipage}\\ \begin{minipage}{3.35in}{\includegraphics[width=3.35in]{rev2}}\hspace*{-2in}$\cdots$ \end{minipage} \caption{The first paragraphs of ``plagiarized''\xspace reviews posted for the {\product}s\xspace \texttt{Sing 'n Learn Chinese} and \texttt{Sing 'n Learn Korean}. In the second review, the title is different and the word ``chinese'' has been replaced by ``korean'' throughout. Sources: {\footnotesize \protect\url{http://www.amazon.com/review/RHE2G1M8V0H9N/ref=cm_cr_rdp_perm}} and \protect\url{http://www.amazon.com/review/RQYHTSDUNM732/ref=cm_cr_rdp_perm}. \label{fig:plagiarism_example}} \end{center} \end{figure} \vspace*{0.3in} \xhdr{Confirmation that text quality\xspace is not the (only) explanatory factor} Since for a given pair of ``plagiarized''\xspace reviews the text quality\xspace of the two copies should be essentially the same, a statistically significant difference between the {\hratio}s\xspace of the members of such pairs is a strong indicator of the influence of a non-textual factor on the {\hevaluator}s\xspace. An initial test of the data reveals that the mean difference in helpfulness ratio\xspace between ``plagiarized''\xspace copies is very close to zero. However, a confounding factor is that for many of our pairs, the two copies may occur in contexts that are practically indistinguishable. Therefore, we bin the pairs by how different their absolute deviations\xspace are, and consider whether {\hratio}s\xspace differ at least for pairs with very different deviations\xspace. More formally, for $i,j \in \{0,0.5,\cdots,3.5\}$ where $i < j$, we write $i$$\succ$}%{\cellcolor[gray]{0.9} $\succ$$j$ (conversely, $i$$\prec$}%{\cellcolor[gray]{0.9} $\prec$$j$) when the helpfulness ratio\xspace of reviews with absolute deviation\xspace $i$ is significantly\xspace larger (conversely, smaller) than that for reviews with absolute deviation\xspace $j$. Here, {\em significantly\xspace larger} or {\em smaller} means that the Mantel-Haenszel test for whether the helpfulness\xspace odds ratio is equal to 1 returns a $95\%$ confidence interval that does not contain 1.\footnote{To avoid drawing conclusions based on possible numerical-precision inaccuracies, we consider any confidence interval that overlaps the interval [0.995,1.005] to contain 1. This ``overlap'' policy affects only two bins in Table \ref{tab:plagiarism_unsigned} and two bins in Table \ref{tab:plagiarism_signed}.} The Mantel-Haenszel test \cite{Agresti:96a} measures the strength of association between two groups, giving more weight to groups with more data. (Experiments with an alternate empirical sampling test were consistent.) We disallow $j= 4$ since there are only relevant 24 pairs which would have to be distributed among $8$ $(i,j)$ bins. The existence of even a {\em single} pair $(i,j)$ in which $i$$\succ$}%{\cellcolor[gray]{0.9} $\succ$$j$ or $i$$\prec$}%{\cellcolor[gray]{0.9} $\prec$$j$ would already be inconsistent with the \contenthyp. Table \ref{tab:plagiarism_unsigned} shows that in fact, there is a significant\xspace difference in a {\em large majority} of cases. Moreover, we see no ``$\prec$}%{\cellcolor[gray]{0.9} $\prec$'' symbols; this is consistent with Figure \ref{fig:unsigned-diff-novar}, which showed that the helpfulness ratio\xspace is inversely correlated with absolute deviation\xspace prior to controlling for text quality\xspace. We also binned the pairs by {\em signed} deviation\xspace. The results, shown in Table \ref{tab:plagiarism_signed}, are consistent with Figure \ref{fig:signed-diff-novar}. First, all but one of the statistically significant results indicate that ``plagiarized''\xspace reviews with star rating\xspace closer to the product\xspace average are judged to be more helpful. Second, the ($-i,i$) results are consistent with the asymmetry depicted in Figure \ref{fig:signed-diff-novar} (i.e., the ``upward slant'' of the green lines). Note that the sparsity of the ``plagiarism''\xspace data precludes an analogous investigation of variance\xspace as as a contextual factor. \newcommand{\ci}{\textcolor{white}} \begin{table}[htdp] \begin{center} {\small \begin{tabular}{|c||c|c|c|c|c|c|c|}\hline \backslashbox{$i$}{$j$} & 0.5 & 1 & 1.5 & 2 & 2.5 & 3 & 3.5 \\\hline \hline 0 & \ci$\prec$}%{\cellcolor[gray]{0.9} $\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ &$\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline 0.5 & \ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & ~}%{$\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline 1 & \ec &\ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & ~}%{$\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline 1.5 & \ec &\ec &\ec & \morehu}%{$\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline 2 & \ec & \ec &\ec & \ec & \morehu}%{$\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline 2.5 & \ec & \ec & \ec &\ec & \ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline 3 &\ec & \ec &\ec & \ec & \ec &\ec & \ci$\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\\hline \end{tabular} } \caption{\label{tab:plagiarism_unsigned} ``Plagiarized''\xspace reviews with a lower absolute deviation\xspace tend to have larger {\hratio}s\xspace than duplicates with higher absolute deviations. Depicted: whether reviews with deviation $i$ have an helpfulness ratio\xspace significantly larger ($\succ$}%{\cellcolor[gray]{0.9} $\succ$) or significantly smaller ($\prec$}%{\cellcolor[gray]{0.9} $\prec$, no such cases) than duplicates with absolute deviation\xspace $j$ (blank: no significant difference). } \end{center} \end{table} \newtheorem{theorem}{Theorem}[section] \newcommand{\xproof}[1]{ {\noindent {\it Proof.} {#1} \rule{2mm}{2mm} \vskip \belowdisplayskip} } \def{p}{{p}} \def{q}{{q}} \def{\alpha}{{\alpha}} \def{\varepsilon}{{\varepsilon}} \section{A Model Based on \headerformat{\bias} and Mixtures of Distributions} \label{sec:gaussians} We now consider how the main findings about helpfulness, variance, and divergence from the mean are consistent with a simple model based on individual bias with a mixture of opinion distributions. In particular, our model exhibits the phenomenon observed in our data that increasing the variance shifts the helpfulness distribution so it is first unimodal and subsequently (with larger variance) develops a local minimum around the mean. The model assumes that helpfulness evaluators can come from two different distributions: one consisting of evaluators who are positively disposed toward the product, and the other consisting of evaluators who are negatively disposed toward the product. We will refer to these two groups as the {\em positive} and {\em negative} evaluators respectively. We need not make specific distributional assumptions about the evaluators; rather, we simply assume that their opinions are drawn from {\em some} underlying distribution with a few basic properties. Specifically, let us say that a function $f : {\bf R} \rightarrow {\bf R}$ is {\em $\mu$-centered}, for some real number $\mu$, if it is {\em unimodal} at $\mu$, centrally symmetric, and $C^2$ (i.e. it possesses a continuous second derivative). That is, $f$ has a unique local maximum at $\mu$, $f'$ is non-zero everywhere other than $\mu$, and $f(\mu + x) = f(\mu - x)$ for all $x$. We will assume that both positive and negative evaluators have one-dimensional opinions drawn from (possibly different) distributions with density functions that are $\mu$-centered for distinct values of $\mu$. Our model will involve two parameters: the {\em balance} between positive and negative reviewers ${p}$, and a {\em controversy level} ${\alpha} > 0$. Concretely, we assume that there is a ${p}$ fraction of positive evaluators and a $1 - {p}$ fraction of negative evaluators. (For notational simplicity, we sometimes write ${q}$ for $1 - {p}$.) The controversy level controls the distance between the means of the positive and negative populations: we assume that for some number $\mu$, the density function $f$ for positive evaluators is $(\mu + {q} {\alpha})$-centered, and the density function $g$ for negative evaluators is $(\mu - {p} {\alpha})$-centered. Thus, the density function for the full population is $h(x) = {p} f(x) + {q} g(x)$, and it has mean ${p} (\mu + {q} {\alpha}) + {q} (\mu - {p} {\alpha}) = \mu$. In this way, our parametrization allows us to keep the mean and balance fixed while observing the effects as we vary the controversy level ${\alpha}$. Now, under our individual-bias assumption, we posit that each helpfulness evaluator has an opinion $x$ drawn from $h$, and each regards a review as helpful if it expresses an opinion that is within a small tolerance of $x$. For small tolerances, we expect therefore that the helpfulness ratio of reviews giving a score of $x$, as a function of $x$, can be approximated by $h(x)$. Hence, we consider the shape of $h(x)$ and ask whether it resembles the behavior of helpfulness ratios observed in the real data. Since the controversy level ${\alpha}$ in our model affects the variance in the empirical data (${\alpha}$ is the distance between the peaks of the two distributions, and is thus related to the variance, but the balance ${p}$ is also a factor), we can hope that at as ${\alpha}$ increases one obtains qualitative properties consistent with the data: first a unimodal distribution with peak between the means of $f$ and $g$, and then a local minimum near the mean of $h$. In fact, this is precisely what happens. The main result is the following. \newcommand{\textcolor{white}{-}}{\textcolor{white}{-}} \begin{table}[tdp] \begin{center} {\small \begin{tabular}{|c||c|c|c|c|c|c|c|}\hline \backslashbox{$i$}{$j$} & -0.5 & -1 & -1.5 & -2 & -2.5 & -3 & -3.5 \\\hline \hline 0 & \ci$\prec$}%{\cellcolor[gray]{0.9} $\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline -0.5 & \ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ \\ \hline -1 & \ec &\ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ \\ \hline -1.5 & \ec &\ec &\ec & \morehu}%{$\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline -2 & \ec & \ec &\ec & \ec & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline -2.5 & \ec & \ec & \ec &\ec & \ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\ \hline -3 &\ec & \ec &\ec & \ec & \ec &\ec & \ci$\succ$}%{\cellcolor[gray]{0.9} $\succ$ \\\hline \multicolumn{7}{c}{} \\ \hline \backslashbox{$i$}{$j$} & \ms0.5 & \ms1 & \ms1.54 & \ms2 & \ms2.5 & \ms3 & \ms3.5 \\\hline \hline 0 & \morehu}%{$\prec$ & \morehu}%{$\prec$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline 0.5 & \ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline 1 & \ec &\ec & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\prec$}%{\cellcolor[gray]{0.9} $\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline 1.5 & \ec &\ec &\ec & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline 2 & \ec & \ec &\ec & \ec & \morehu}%{$\prec$ & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline 2.5 & \ec & \ec & \ec &\ec & \ec & \morehu}%{$\prec$ & \morehu}%{$\prec$ \\ \hline 3 &\ec & \ec &\ec & \ec & \ec &\ec & \morehu}%{$\prec$ \\\hline \multicolumn{7}{c}{} \\ \hline \backslashbox{}{$i$} & -0.5 & -1 & -1.5 & -2 & -2.5 & -3 & -3.5 \\\hline \hline $-i$ & & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & $\succ$}%{\cellcolor[gray]{0.9} $\succ$ & & & & \\\hline \end{tabular} } \end{center} \caption{\label{tab:plagiarism_signed} The same type of analysis as Table \ref{tab:plagiarism_signed} but with {\em signed} deviation. The first (resp. second) table is consistent with the lefthand (resp. righthand) side of Figure \ref{fig:signed-diff-novar}. The third table is consistent with the ``upward slant'' of the green lines in Figure \ref{fig:signed-diff-novar}: for the same absolute deviation\xspace value, when there is a significant difference in helpfulness\xspace odds ratio, the difference is in favor of the positive deviation\xspace. \newline \hspace*{.2in}{\small (There are a noticeable number of blank cells, indicating that a statistically significant difference was not observed for the corresponding bins, due to sparse data issues: there are twice as many bins as in the absolute-deviation analysis but the same number of pairs.)} } \end{table} \begin{theorem} For any choice of $f$, $g$, and ${p}$ as defined as above, there exist positive constants ${\varepsilon}_0 < {\varepsilon}_1$ such that \begin{itemize} \item[(i)] When ${\alpha} < {\varepsilon}_0$, the combined density $h(x)$ is unimodal, with maximum strictly between the mean of $f$ and the mean of $g$. \item[(ii)] When ${\alpha} > {\varepsilon}_1$, the combined density function $h(x)$ has a local minimum between the means of $f$ and $g$. \end{itemize} \label{thm:general-mixture} \end{theorem} \xproof{ We first prove (i). Let us write $\mu_f = \mu + {q} {\alpha}$ for the mean of $f$, and $\mu_g = \mu - {p} {\alpha}$ for the mean of $g$. Since $f$ and $g$ have unique local maxima at their means, we have $f''(\mu_f) < 0$ and $g''(\mu_g) < 0$. Since these second derivatives are continuous, there exists a constant $\delta$ such that $f''(x) < 0$ for all $x$ with $|x - \mu_f| < \delta$, and $g''(x) < 0$ for all $x$ with $|x - \mu_g| < \delta$. Since $\mu_f - \mu_g = \alpha$, if we choose $\alpha < \delta$, then $f''(x)$ and $g''(x)$ are both strictly negative over the entire interval $[\mu_g, \mu_f]$. Now, $f'(x)$ and $g'(x)$ are both positive for $x < \mu_g$, and they are both negative for $x > \mu_f$. Hence $h(x) = {p} f(x) + {q} g(x)$ has the properties that (a) $h'(x) > 0$ for $x < \mu_g$; (b) $h'(x) < 0$ for $x > \mu_f$, and (c) $h''(x) < 0$ for $x \in [\mu_g, \mu_f]$. From (a) and (b) it follows that $h$ must achieve its maximum in the interval $[\mu_g, \mu_f]$, and from (c) it follows that there is a unique local maximum in this interval. Hence setting ${\varepsilon}_0 = \delta$ proves (i). For (ii), since $f$ and $g$ must both, as density functions that are both centered around their respective means, go to $0$ as $x$ increases or decreases arbitrarily, we can choose a constant $c$ large enough that $f(\mu_f - x) + g(x + \mu_g) < \min({p} f(\mu_f), {q} g(\mu_g))$ for all $x > c$. If we then choose ${\alpha} > c/\min({p},{q})$, we have $\mu_f - \mu > c$ and $\mu - \mu_g > c$, and so $h(\mu) = {p} f(\mu) + {q} g(\mu) \leq f(\mu) + g(\mu) < \min({p} f(\mu_f), {q} g(\mu_g)) \leq \min(h(\mu_f), h(\mu_g))$, where the second inequality follows from the definition of $c$ and our choice of $\alpha$. Hence, $h$ is lower at its mean $\mu$ than at either of $\mu_f$ or $\mu_g$, and hence it must have a local minimum in the interval $[\mu_g,\mu_f]$. This proves (ii) with ${\varepsilon}_1 = c / \min({p},{q})$. } \begin{figure}[h] \centering \includegraphics[width=0.45\textwidth,viewport= 70 200 550 600,clip]{gaussians_synthetic_theta0_sigma1_alpha045_step20-2} \caption{An illustration of Theorem \ref{thm:symmetric-mixture}, using translated Gaussians as an example --- the theorem itself does not require any Gaussian assumptions. } \label{fig:gaussians-synthetic} \end{figure} For density functions at this level of generality, there is not much one can say about the unimodal shape of $h$ in part (i) of Theorem~\ref{thm:general-mixture}. However, if $f$ and $g$ are translates of the same function, and their next non-zero derivative at $0$ is positive, then one can strengthen part (i) to say that the unique maximum occurs between the means of $h$ and $f$ when ${p} > \frac12$, and between the means of $h$ and $g$ when ${p} < \frac12$. In other words, with this assumption, one recovers the additional qualitative observation that for small separations between the functions, it is best to give scores that are slightly above average. We note that Gaussians are one basic example of a class of density functions satisfying this condition; there are also others. See Figure~\ref{fig:gaussians-synthetic} for an example in which we plot the mixture when $f$ and $g$ are Gaussian translates, with ${p}$ fixed but changing ${\alpha}$ and hence changing the variance. (Again, it is not necessary to make a Gaussian assumption for anything we do here; the example is purely for the sake of concreteness.) Specifically, our second result is the following. In its statement, we use $f^{(j)}(x)$ to denote the $j^{\rm th}$ derivative of a function $f$, and recall that we say a function is $C^j$ if it has at least $j$ continuous derivatives. \begin{theorem} Suppose we have the hypotheses of Theorem~\ref{thm:general-mixture}, and additionally there is a function $k$ such that $f(x) = k(x - \mu_f)$ and $g(x) = k(x - \mu_g)$. (Hence $k$ is unimodal with its unique local maximum at $x = 0$.) Further, suppose that for some $j$, the function $k$ is $C^{j+1}$ and we have $k^{(j)}(0) > 0$ and $k^{(i)}(0) = 0$ for $2 < i < j$. Then in addition to the conclusions of Theorem~\ref{thm:general-mixture}, we also have \begin{itemize} \item[(i$'$)] There exists a constant ${\varepsilon}_0'$ such that when ${\alpha} < {\varepsilon}_0'$, the combined density $h(x)$ has its unique maximum strictly between the mean of $f$ and the mean of $h$ when ${p} > \frac12$, and strictly between the mean of $g$ and the mean of $h$ when ${p} < \frac12$. \end{itemize} \label{thm:symmetric-mixture} \end{theorem} \xproof{ We omit the proof, which applies Taylor's theorem to $k'$, due to space limitations. } We are, of course, not claiming that our model is the only one that would be consistent with the data we observed; our point is simply to show that there exists at least one simple model that exhibits the desired behavior. \section{Consistency among countries} \label{sec:countries} In this section we evaluate the robustness of the observed social-effects phenomena by comparing review data from three additional different national Amazon sites: Amazon.co.uk (U.K), Amazon.de (Germany) and Amazon.co.jp (Japan), collected using the same methodology described in Section \ref{sec:data}, except that because of the particulars of the AWS API, we were unable to filter out mechanically cross-posted reviews from the Amazon.co.jp data. It is reasonable to assume that these reviews were produced independently by four separate populations of reviewers (there exist customers who post reviews to multiple Amazon sites, but such behavior is unusual). There are noticeable differences between reviews collected from different regional Amazon sites, in both average helpfulness ratio\xspace and review variance (Table \ref{countries}). The review dynamics in the U.K. and Japan communities appear to be less controversial than in the U.S. and Germany. Furthermore, repeating the analysis from Section \ref{sec:variance} for these three new datasets reveals the same qualitative patterns observed in the U.S. data and suggested by the model introduced in Section \ref{sec:gaussians}. Curiously enough, for the Japanese data, in contrast to its general reputation of a collectivist culture \cite{Bond+Smith:96a}, we observe that the left hump is higher than the right one for reviews with high variance, i.e., reviews with {\smark}s\xspace \textit{below} the mean are more favored by {\hevaluator}s\xspace than the respective reviews with positive deviations (Figure \ref{fig:countries_var3}). In the context of our model, this would correspond to a larger proportion of negative evaluators (balance $p<0.5$). \begin{table} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline & Total reviews& Avg h.ratio & Avg star rating\xspace var. \\ \hline U.S. & 1,008,466 & 0.72 & 1.34 \\ U.K. & 127,195 & 0.80 & 0.95 \\ Germany & 184,705 & 0.74 & 1.24 \\ Japan & 253,971 & 0.69 & 0.93 \\ \hline \end{tabular} \caption{\label{countries} Comparison of review data from four regional sites: number of reviews with 10 or more {\hvote}s\xspace, average helpfulness ratio\xspace, and average variance\xspace in star rating\xspace.} \end{center} \end{table} \begin{figure}[h] \begin{center} \begin{minipage}{1.5in}{\includegraphics[width=1.5in]{countries_jp_var3}} \end{minipage} \qquad \begin{minipage}{1.5in}{\includegraphics[width=1.5in]{countries_us_var3}} \end{minipage} \caption{\label{fig:countries_var3}Signed {\deviations}\xspace vs. helpfulness ratio\xspace for variance\xspace $=3$, in the Japanese (left) and U.S. (right) data. The curve for Japan has a pronounced lean towards the left.} \end{center} \end{figure} \section{Conclusion} \label{sec:conclusion} We have seen that helpfulness evaluations on a site like Amazon.com provide a way to assess how opinions are evaluated by members of an on-line community at a very large scale. A review's perceived helpfulness depends not just on its content, but also the relation of its score to other scores. This dependence on the score contrasts with a number of theories from sociology and social psychology, but is consistent with a simple and natural model of individual bias in the presence of a mixture of opinion distributions. There are a number of interesting directions for further research. First, the robustness of our results across independent populations suggests that the phenomenon may be relevant to other settings in which the evaluation of expressed opinions is a key social dynamic. Moreover, as we have seen in Section~\ref{sec:countries}, variations in the effect (such as the magnitude of deviations above or below the mean) can be used to form hypotheses about differences in the collective behaviors of the underlying populations. Finally, it would also be very interesting to consider social feedback mechanisms that might be capable of modifying the effects we observe here, and to consider the possible outcomes of such a design problem for systems enabling the expression and dissemination of opinions. \xhdr{Acknowledgments} We thank Daria Sorokina, Paul Ginsparg, and Simeon Warner for assistance with their code, and Michael Macy, Trevor Pinch, Yongren Shi, Felix Weigel, and the anonymous reviewers for helpful (!)\ comments. Portions of this work were completed while Gueorgi Kossinets was a postdoctoral researcher in the Department of Sociology at Cornell University. This paper is based upon work supported in part by a University Fellowship from Cornell, DHS grant N0014-07-1-0152, the National Science Foundation grants BCS-0537606, CCF-0325453, CNS-0403340, and CCF-0728779, a John D. and Catherine T. MacArthur Foundation Fellowship, a Google Research Grant, a Yahoo!\ Research Alliance gift, a Cornell University Provost's Award for Distinguished Scholarship, a Cornell University Institute for the Social Sciences Faculty Fellowship, and an Alfred P.\ Sloan Research Fellowship. Any opinions, findings, and conclusions or recommendations expressed are those of the authors and do not necessarily reflect the views or official policies, either expressed or implied, of any sponsoring institutions, the U.S.\ government, or any other entity. \newcommand{\vspace{-.09in}}{\vspace{-.09in}}

1. Introduction {#sec1} =============== The cornea is a hard, transparent tissue through which light enters the eyes first. It is transparent, without blood vessels, and has many sensory nerve endings. Serving as a barrier to debris and infection as well as two-thirds of the refractive power of the eye, the cornea is imperative to proper vision \[[@B1]\]. Cornea is part of the eye exposed to the outer environment and thus most likely to sustain damage due to various insults, such as trauma, inflammation, infection, and so on \[[@B2]\]. The corneal wound healing results in the formation of opaque corneal scar. In fact, millions of people around the world have corneal scars that cause vision loss \[[@B3]\]. Corneal transplantation is the traditional treatment for clinically significant corneal opacities. Regenerative medicine for the cornea represents a novel treatment strategy for patients with corneal diseases. Stem cell-based therapies represent a novel strategy that in some instances (using stromal stem cells) may substitute conventional corneal transplantation \[[@B1], [@B4]\]. The formation of corneal scar is a complex process: its mechanism and potential genetic control are not fully understood. It involves the synthesis of various growth factors, cytokines, and proteases produced by epithelial cells, stromal keratocytes, inflammatory cells, and lacrimal gland cells, \[[@B5]\]. Therefore, it is very important to understand the precise molecular mechanism of corneal scar formation and formulate effective diagnosis and treatment strategies. Gene chip or gene mapping is a gene detection technology that has been used for more than ten years. Gene chip can quickly detect all genes in the expression information of the same sample at the same time point. It is especially suitable for screening differentially expressed genes, \[[@B6]\]. We can use microarray technology and bioinformatics analysis to screen genetic alterations at the genome level \[[@B7]\]. Using microarray technology and bioinformatics analysis, we can accurately determine the differential expression level of genes, thus providing an effective method for large-scale gene expression research \[[@B8], [@B9]\]. In previous articles, some potential differentially expressed genes, hub genes, and signal pathways related to corneal scar formation have not been identified, which may help us understand the molecular mechanism of corneal scar formation and provide candidate targets for the treatment of corneal scar. In addition, there are no researches about differential gene analysis, Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment pathway analyses, and protein-protein interaction (PPI) network analyses of corneal scar using the dataset of GEO database. Therefore, gene expression profile analysis can better understand the genetic, cellular, and molecular changes that occur during the formation of corneal scar, which will provide an opportunity to design treatment options, selectively regulate the key stages of healing process, and thus produce scars closer to normal corneal structure \[[@B5]\]. In the present study, we downloaded an original microarray data set GSE6676 from the NCBI Gene Expression Comprehensive Database (NCBI-GEO) and analyzed to obtain DEGs between mice that have the overexpression of TGF beta and wildtype litter mate to the mice that have the overexpression of TGF beta. Scar formation of the cornea is critically modulated by the expression of transforming growth factor-beta (TGF-beta), TGF beta can induce the transformation of corneal keratocytes into myofibroblasts, which is the main cause of corneal fibrosis or scar formation \[[@B10]\]. Subsequently, in order to identify the related genes, pathways, and existing molecular mechanisms, we carried out Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment pathway analyses and protein-protein interaction (PPI) network analyses. In summary, the mechanism of corneal scar formation was studied by bioinformatics, which provided potential biomarker candidates for clinical treatment and drug target discovery. 2. Materials and Methods {#sec2} ======================== 2.1. Microarray Data {#sec2.1} -------------------- GEO (<http://www.ncbi.nlm.nih.gov/geo>) \[[@B11]\] is a public functional genomics data repository of high throughout gene expression data, chips, and microarrays. One gene expression dataset \[GSE6676\] was downloaded from GEO (GPL1261 \[Mouse430_2\] Affymetrix Mouse Genome 430 2.0 Array). According to the annotation information in the platform, the probes are transformed into corresponding gene symbols. The GSE6676 data set contained 8 samples, including 4 corneas from wildtype mice samples and 4 corneas under the influence of high doses of TGF-beta samples. 2.2. Identification of Differentially Expressed Genes (DEGs) {#sec2.2} ------------------------------------------------------------ The GEO2R (<http://www.ncbi.nlm.nih.gov/geo/geo2r/>) is used to identify differentially expressed genes (DEGs). GEO2R is based on *R* that comes with the GEO databases. Differentially expressed genes (DEGs) were determined by \|logFC\| of no less than 1 and *t*-tests with *P* \< 0.05. 2.3. GO Enrichment and KEGG Pathway Analysis of the DEGs {#sec2.3} -------------------------------------------------------- The Database for Annotation, Visualization, and Integrated Discovery (DAVID; david. abcc. ncifcrf. Gov) software was used to interpret functions of extensive genes obtained from previous genome studies \[[@B12]\]. The Gene Ontology database (GO; <http://www.geneontology.org>) contained structured ontologies or vocabularies that depict basic characteristics of genes and gene products \[[@B13]\]. The Kyoto Encyclopedia of Genes and Genomes database (KEGG; <http://www.genome.jp/kegg/>) synthesizes information of biological systems from genomic, chemical, and systemic functional aspects \[[@B14]\]. Using the DAVID software, functional and pathway enrichment analyses were conducted separately, for upregulated and downregulated genes. *P* \< 0.05 was considered to indicate a statistically significant difference for the screening of significant GO terms and KEGG pathways. 2.4. PPI Network Construction and Module Analysis {#sec2.4} ------------------------------------------------- PPI networks are predicted by searching tools for interacting genes (STRING; [(http://string-db.org)](http://string-db.org)) (version 10.0) \[[@B15]\] online database. Analyzing the functional interactions between proteins may provide insights into the mechanisms of generation or development of diseases. In the present study, PPI network of DEGs was constructed using STRING database, and an interaction with a combined score \>0.4 was considered statistically significant. Cytoscape (version 3.4.0) is an open source bioinformatics software platform for visualizing molecular interaction networks \[[@B16]\]. Cytoscape Plug-in Molecular Complex Detection (MCode) (Version 1.4.2) is an application for clustering a given network based on topology to discover densely connected regions \[[@B17]\]. The PPI network is drawn by Cytoscape, and the most important module is identified by MCODE. The criteria for selection were as follows: MCODE scores \>5, degree cut-off = 2, node score cut-off = 0.2, Max depth = 100, and k-score = 2. Subsequently, the KEGG and GO analyses for genes in this module were performed using DAVID. 2.5. Hub Genes Selection {#sec2.5} ------------------------ The hub genes were selected using the cytoHubba plugin, a Cytoscape plugin, was used to determine the hub proteins or genes in the PPI network. We used five methods (degree, Maximum Neighborhood Component (MNC), Radiality centrality, Stress centrality, Closeness centrality) to sequence and evaluate central genes \[[@B18]\]. 3. Results {#sec3} ========== 3.1. Identification of Differentially Expressed Genes (DEGs) {#sec3.1} ------------------------------------------------------------ We downloaded gene expression profile GSE6676 from the GEO database, and use the GEO2R method to identify DEGs in the corneas under the influence of high doses of TGF-beta samples compared with the corneas from wildtype mice samples. Based on the initial data, we included the inclusion criteria: *P* \< 0.05 and \|logFC\| ≥ 1.0. A total of 1377 differentially expressed genes were identified, of which 776 were up-regulated and 601 were down-regulated between the corneas under the influence of high doses of TGF-beta samples compared with the corneas from wildtype mice samples. At the same time, the volcano plot of all DEGs and the heat map of the top 50 up-regulated genes and the top 50 down-regulated genes are shown in Figures [1](#fig1){ref-type="fig"} and [2](#fig2){ref-type="fig"}, respectively. 3.2. GO Term Enrichment Analysis {#sec3.2} -------------------------------- Then, the online software DAVID was used to functionally categorize these 1377 significant DEGs. As shown in [Figure 3](#fig3){ref-type="fig"}, the top 5 enrichment analyses are shown for each part of gene ontology (GO) analysis. For biological process (BP) enrichment analysis, the results showed that the upregulated genes significantly took part in the ion transport (GO:0006811), synaptic transmission (GO:0007268), visual perception (GO:0007601), sensory perception of light stimulus (GO:0050953), and neurotransmitter transport (GO:0006836). The down-regulated genes were significantly involved in epidermal cell differentiation (GO:0009913), inflammatory response (GO:0006954), response to wounding (GO:0009611), epidermis development (GO:0008544), regulation of platelet activation (GO:0010543) as shown in [Figure 3(a)](#fig3){ref-type="fig"}. For cell component (CC) enrichment analysis, the present study showed that the upregulated genes were mainly involved in synapse (GO:0045202), plasma membrane (GO:0005886), synapse part (GO:0044456), and plasma membrane part (GO:0044459), cell junction (GO:0030054). The Downregulated genes mainly revolved in extracellular region (GO:0005576), extracellular region part (GO:0044421), extracellular space (GO:0005615), extracellular matrix (GO:0031012), and proteinaceous extracellular matrix (GO:0005578) as shown in [Figure 3(b)](#fig3){ref-type="fig"}. In addition, in the enrichment analysis of molecular function (MF), up-regulated genes are mainly enriched in ion channel activity (GO:0005216), substrate specific channel activity (GO:0022838), gated channel activity (GO:0022836), channel activity (GO:0015267), and passive transmembrane transporter activity (GO:0022803). For down-regulated genes, they mainly take part in cytokine activity (GO:0005125), chemokine activity (GO:0008009), chemokine receptor binding (GO:0042379), sulfotransferase activity (GO:0008146), and endopeptidase activity (GO:0004175) as shown in [Figure 3(c)](#fig3){ref-type="fig"}. 3.3. KEGG Pathway Analysis {#sec3.3} -------------------------- KEGG pathway analysis showed that the up-regulated DEGs were mainly enriched in Glycosphingolipid biosynthesis (mmu00601), Neuroactive ligand-receptor interaction (mmu04080), Melanogenesis (mmu04916), PPAR signaling pathway (mmu03320), Nitrogen metabolism (mmu00910), while the down-regulated DEGs were mainly enriched in Cytokine-cytokine receptor interaction (mmu04060), Jak-STAT signaling pathway (mmu04630), Sulfur metabolism (mmu00920), Toll-like receptor signaling pathway (mmu04620). The KEGG pathway enrichment analysis results of up-regulated and down-regulated are shown in [Figure 4](#fig4){ref-type="fig"}. 3.4. Construction of PPI Network and Module Analysis {#sec3.4} ---------------------------------------------------- Based on the information in the STRING database, we used the plug-in MCODE in Cytoscape find 893 nodes and 3041 edges in the PPI network. Nodes represent DEGs and edges represent interactions between DEGs. Then, we use MCODE to filter the modules of PPI network. Therefore, two important modules are obtained from PPI network, as shown in [Figure 5](#fig5){ref-type="fig"}. Following the GO and KEGG pathway enrichment, analysis of DEGs in module 1 and module 2 were analyzed using DAVID. However the result shows that there were not go function enrichment and specific KEGG pathways in module 1, while the DEGs in module 2 was mostly distributed in the GO function enrichment. The module 2 most significantly took part in the molecular function (MF) as shown in [Table1](#tab1){ref-type="table"}. For the KEGG pathway analysis, it was mainly distributed in Neuroactive ligand-receptor interaction and Calcium signaling pathway. 3.5. Hub Gene Selection {#sec3.5} ----------------------- In the present study, we used cytoHubba to choose hub genes. According to the five classification methods in cytoHubba, the top 15 hub genes selected by these ranked methods in cytoHubba are shown in [Table 2](#tab2){ref-type="table"}. Finally, seven central genes were identified by overlapping the first 15 genes, as shown in [Figure 6](#fig6){ref-type="fig"}. IL6 (interleukin-6) is the most excellent central genes based on five ranked methods. MMP9 (matrix metallopeptidase 9), CXCL10 (C-X-C motif chemokine ligand 10), MAPK8 (mitogen-activated protein kinase 8), TLR4 (toll-like receptor 4), HGF (hepatocyte growth factor), EDN1 (endothelin 1) were also selected as hub genes. 4. Discussion {#sec4} ============= The only widely accepted treatment for severe corneal scarring or clouding is the traditional techniques of corneal transplantation \[[@B19], [@B20]\]. The number of corneas required for these transplants, however, exceeds those available, leading to a shortage of donor tissues, especially in developing countries, which means that many patients can never undergo keratoplasty \[[@B20]\]. Therefore, it is very important to understand the precise molecular mechanism of corneal scar formation in order to formulate an effective diagnosis and look for alternative treatment strategies. GEO2R is based on R software for analysis. GEO2R performs comparisons on original submitter-supplied processed data tables using the GEOquery and limma R packages from the Bioconductor project. Thus, GEO2R provides a simple interface that allows users to perform R statistical analysis without command line expertise. Both are equally accurate and many researchers have used GEO2R in the DEG analysis before. So we used GEO2R to analyze differentially expressed genes at our study. In the present study, the gene expression profile of GSE6676 was downloaded and various bioinformatics analysis methods were performed to reveal the presence of differentially expressed molecules in an opaque cornea with blood vessels and lack of sensory nerves versus a normal cornea. Our study showed that there were 1377 differentially expressed genes in these two different clinical scenarios. The GO analyses showed that Upregulated DEGs mainly participated in cell component (CC), while down regulated DEGs mainly took part in both cell component (CC) and molecular function (MF). The upregulated genes were mainly enriched in Synapse, synapse part, plasma membrane, plasma membrane part, and ion transport. However the downregulated genes were mainly involved with extracellular region, extracellular region part, cytokine activity, chemokine activity, and chemokine receptor binding. Synapses and synapse part may be related to nerves in corneal wound healing. Corneal nerves are very important to maintain the protective reflexes (blinking and tearing) and to provide support to epithelial regeneration, through growth factors \[[@B21], [@B22]\]. Plasma membranes and ion transport may be due to a large amount of material exchange during corneal wound healing. Some study had shown that during the healing process of epithelial wounds, a large amount of extracellular matrix is needed to promote the migration of epithelial cells to cover the wounds \[[@B23]\]. More importantly, they play an important role in signaling cross-talk. Previous evidence suggests the presence and importance of intracellular signal crosstalk during epithelial healing after injury \[[@B2]\]. Previous studies have reported that extracellular region and extracellular region part play important roles in the cornea wound repair. For example, extracellular matrix (ECM) provides chemical and physical signals to keratocytes, thus affecting their differentiation to fibroblasts and/or myofibroblasts, and plays an important role in the formation of corneal scar. It also provides a matrix through which they migrate during wound repair \[[@B24]--[@B26]\]. In addition, recent studies have shown that cytokines and chemokines are important regulators of wound healing-related cell growth, proliferation, migration, differentiation, adhesion, ECM deposition, and protease regulation. Corneal cells express a variety of growth factors and cytokines that have special effects on epithelial cells. It is activated during corneal wound healing and plays a major role in corneal inflammation and angiogenesis \[[@B27]\]. THE KEGG pathway analysis also indicated that the DEGs were mainly enriched in Neuroactive ligand-receptor interaction and Cytokine-cytokine receptor interaction, this is consistent with our previous results of functional enrichment analysis. Ji et al. \[[@B28]\] showed that the cytokine-cytokine-receptor interaction pathway is an important way to participate in miR-296 gene-related inflammation in cornea of alkali burned mice. The PPI network of DEGS is constructed by strings and visualized by cytoscape. Five significant modules (score ≥5) were identified by MCODE, then the GO and KEGG pathway enrichment analysis of DEGs in this most significant module was analyzed using DAVID. Our study shows that no specific GO and KEGG pathways were found in module 1. Module 2 most significantly took part in G-protein coupled receptor protein signaling pathway, cell surface receptor linked signal transduction, peptide receptor activity, G-protein coupled, peptide binding, and neuropeptide receptor activity. G-protein coupled receptor protein signaling pathway, peptide receptor activity, and G-protein coupling played important roles in Lysophosphatidic Acid Promoting Corneal Epithelial Wound Healing. LPA accelerates corneal epithelial wound healing by inducing the ability of autocrine HB-EGF signal. G protein-coupled receptor ligands can be used as stimulants to activate LPA pathway and epidermal growth factor receptor to meet the pathophysiological challenges of human corneal epithelial cells \[[@B29]\]. Previous study showed that some cell surface receptor linked signal transduction plays a major role in the pathogenesis of corneal neovascularization, such as Vascular Endothelial Growth Factors (VEGFs). Vascular endothelial cell proliferation can be promoted by allogeneic T cells from corneal transplantation host by using VEGF-A signaling. Vascular endothelial growth factor is one of the important factors affecting corneal neovascularization. The role of vascular endothelial growth factor in corneal neovascularization has been confirmed. The inhibitory effect of vascular endothelial growth factor is currently being studied as a treatment for corneal neovascularization \[[@B30], [@B31]\]. KEGG pathways in module 2 were mainly involved in Neuroactive ligand-receptor interaction and Calcium signaling pathway. The Neuroactive ligand-receptor interaction was also found in KEGG pathway analysis of the up-regulated DEGs. Neuroactive ligands can transmit important signals to maintain corneal reflex and recovery of nervous function \[[@B21]\]. Some study showed that early reactions after corneal epithelial injury include the release of nucleotides, the transmission of calcium waves from the wound, and the promotion of migration to reconstruct the cytoskeleton rearrangement of the epithelial barrier. So calcium signaling pathway promotes migration to reestablish the epithelial barrier \[[@B32]\]. In addition, calcium blocker diltiazem is to be used for therapeutic control of pain following corneal injury or surgery because it can retain blink reflex and promote wound healing \[[@B33]\]. Therefore, an in-depth understanding of these pathways is helpful to elucidate the key mechanism of corneal wound healing. Moreover, we analyzed the PPI network of the DEGs and seven hub genes were identified by five sequencing methods in cytoHubba: IL6 (interleukin-6), MMP9 (matrix metallopeptidase 9), CXCL10 (C-X-C motif chemokine ligand 10), MAPK8 (mitogen-activated protein kinase 8), TLR4 (toll-like receptor 4), HGF (hepatocyte growth factor), EDN1 (endothelin 1). IL6, the most outstanding hub gene, which had been reported to play an important role in ocular surface immune defense and decreases the severity and susceptibility of contact lens--related keratitis \[[@B34]\]. However Di Girolamo illustrated IL6 that are induced by UV radiation may play a key role in initiating blood vessel formation, cellular proliferation, tissue invasion, and inflammation \[[@B35]\]. Ghasemi also found IL-6 plays and modulates many important roles in ocular inflammation and angiogenesis in the cornea and it can result in unwanted neovascularization or exacerbation of inflammation, leading to the destruction and damage of delicate eye factors \[[@B36]\]. In addition, IL-6 can stimulate neutrophil degranulation, causing localized tissue damage with loss of corneal transparency and visual impairment \[[@B37]\]. In a word, IL-6 plays a dual role in the cornea. MMP9 (matrix metallopeptidase 9), as one of imperative matrix metalloproteinase, synthesized by corneal epithelial cells and stromal cells, have long been suspected of having a significant role in keratoconus \[[@B38]\]. Matrix metalloproteinases (MMPs) affect cell migration by degradation of extracellular matrix or by altering cell adhesion properties, and increase during corneal wound healing. Previous studies have shown that Galectin-3-induced MMP-9 promotes corneal cell movement and wound healing by destroying cell contact, while knockout mice show delayed skin wound healing in vivo \[[@B39], [@B40]\]. Groblewska et al. reported that MMP-9 are able to degrade collagen type IV from the basement membrane and ECM, a process that is related to tumor progression, metastasis, growth, and angiogenesis \[[@B41]\]. Du et al. had reported that blockade of MMP -9 could inhibit corneal lymphangiogenesis \[[@B42]\]. A better understanding of this factor may provide novel therapies for corneal transplant rejection and other lymphatic disorders. CXCL10 (C-X-C motif chemokine ligand 10) as a member of the interferon-inducible tripeptide motif Glu-Leu-Arg-negative (ELR-) CXC chemokines, is a fibrotic and vascular inhibitory chemokine produced by macrophages, endothelial cells, and fibroblasts. It is reported to be involved in the natural defense of cornea against microbial infection. Interferon regulatory factors (IRF1) play a role in corneal innate immune response by regulating the expression of CXCL10. IFN*γ*-producing NK cells enhance the epithelial expression of IRF1 and CXCL10, thus contributing to the corneal natural defense against Pseudomonas aeruginosa infection \[[@B43]\]. Gao et al. reported that CXCL10 in inflammatory cornea may be an endogenous factor, which can prevent angiogenesis and control vascular regression. In addition, CXCL10 can be locally applied to the eye surface by derivative peptide or AAV-CXCL10 to reduce or prevent corneal angiogenesis associated with chemical damage and graft rejection, which is related to corneal lymphangiogenesis \[[@B44]\]. MAPK8, also known as JNK1, is a member family of mitogen-activated protein kinases (MAPK), belonging to conserved serine and threonine protein kinases. Some studies have shown that c-Jun Nterminal kinase (JNK) 1 and 2 play an important role in the formation of corneal scar. The increase of JNK 1/2 level can block the expression of connective tissue growth factor gene induced by TGF-beta 1 and subsequent cell proliferation, migration, and collagen synthesis \[[@B45], [@B46]\]. Previous studies have shown that JNK1/2 participates in the expression of CTGF induced by TGF-beta 1. JNK1/2 pathway inhibitor SP600125 plays a role in the inhibition of CTGF expression. Therefore, the study and identification of JNK1/2 pathway may help to provide a new strategy to reduce corneal scar formation \[[@B46]\]. TLR4 (Toll-like receptor 4) plays an important role in corneal epithelial wound healing. Eslani et al. show that Functional TLR 4 expression is induced after corneal epithelial injury, then Toll-like receptor 4 signal promotes epithelial repair by inducing cell migration, proliferation, and regulates the production of inflammatory cytokines in damaged corneal epithelium \[[@B47]\]. The wound closure rate of human corneal epithelial cells (HCEC) treated with lipopolysaccharide (LPS) increased significantly. In these cells, TLR4 mediates the increase of cytokines such as IL-6, TNF-a, and CXCL8/IL-8, and the phosphorylation of ERK1/2 and p38 MAP kinases \[[@B2]\]. Therefore, TLR4 contributes to inflammatory response and epithelial cell migration and proliferation. Therefore, the regulation of TLR4 signal may provide a new treatment for corneal epithelial wound healing disorder. Hepatocyte growth factor (HGF) and its receptor protein was found in cornea \[[@B48]\]. Gupta et al. show that HGF plays an important role in corneal repair through c-met receptor tyrosine kinase \[[@B49]\]. HGF exists in three main cell types of human cornea, including epithelial cells, stromal cells, and endothelial cells. Corneal epithelial injury can up-regulate the expression of HGF in corneal cells, and then regulate corneal wound healing through secretion, autocrine, and paracrine ways \[[@B50]\]. Moreover Gupta et al. reported that Localized topical BMP7 + HGF gene therapy can treat improving transparency and corneal fibrosis in vivo, and reduce overhealing and selective apoptosis of myofibroblasts \[[@B49]\]. EDN1 is a key protein of non-neuronal autoregulation of retinal blood flow, which can cause vasoconstriction and vascular tension, and enhance endothelial production of vasodilators such as nitric oxide and prostacyclin \[[@B51]\]. The interaction between vascular endothelial growth factor (VEGF) and EDN1 may play an important role in the simultaneous proliferation of vascular wall endothelial cells and smooth muscle cells. Jinzi Zhou had reported elevated EDN1 in vitreous of patients with proliferative diabetic retinopathy (PDR) \[[@B52]\]. However, there was no report shown that whether EDN1 plays a role in corneal wound healing or not. However, the limitation of our study is that there is no dataset of humans related to corneal scar in the GEO database. We only found the dataset of mouse. Although it can not be directly adapted to human beings, it will be of great help to our next research. In addition, based on our analysis, we can infer the potential mechanism of Hub gene in corneal scar, but it needs further molecular biological experiments to verify and confirm the potential mechanism of hub genes in corneal scar formation. 5. Conclusion {#sec5} ============= In conclusion, 1 377 DEGs, 7 hub genes, and 2 important network modules were identified. Seven hub genes IL6, MMP9, CXCL10, MAPK8, TLR4, HGF, and EDN1 and their signaling pathways, identified in this research, warrant further studies in order to explore their application in the clinical setting. However, molecular biological experiments are required to confirm the function of the identified genes in the formation of corneal scar. This work was supported by Henan Provincial Innovation and Outstanding Talent Program (154200510027), 2016 Henan provincial science and technology research project: Construction and clinical application of digital precision spine surgery technology system (162102310018) and 2018 Henan provincial Medical Science and Technology Tackling Program Provincial-ministerial Co-construction Project (SBGJ2018076). Data Availability ================= All data generated or analyzed during this study are included in this published article. Conflicts of Interest ===================== The authors declare that they have no conflicts of interest. Authors\' Contributions ======================= Binfeng Liu, Ang Li, and Hongbo Wang contributed equally to this work. {#fig1} {#fig2} {#fig3} {#fig4} {#fig5} {#fig6} ###### The top 10 Gene Ontology (GO) functions and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways enriched for the genes involved in module 2. Description *P* Value Count Gene symbol ---------------------------------------------------------------------- ----------- ------- ----------------------------------------------------------------------------- GO:0007186 G-protein coupled receptor protein signaling pathway (BP) 1.25E-07 11 GRM5, LTB4R1, CYSLTR1, TACR3, EDN1, OXTR, PROKR1, GPR143, NMBR, HTR2B, OPN4 GO:0007166 cell surface receptor linked signal transduction (BP) 1.97E-06 11 GRM5, LTB4R1, CYSLTR1, TACR3, EDN1, OXTR, PROKR1, GPR143, NMBR, HTR2B, OPN4 GO:0008528 peptide receptor activity, G-protein coupled (MF) 6.35E-05 4 TACR3, OXTR, PROKR1, NMBR GO:0042277 peptide binding (MF) 1.93E-04 4 TACR3, OXTR, PROKR1, NMBR GO:0008188 neuropeptide receptor activity (MF) 3.53E-04 3 TACR3, PROKR1, NMBR GO:0042165 neurotransmitter binding (MF) 0.001843 3 TACR3, PROKR1, NMBR GO:0004974 leukotriene receptor activity (MF) 0.003007 2 LTB4R1, CYSLTR1 GO:0003013 circulatory system process (BP) 0.004138 3 EDN3, EDN1, OXTR GO:0005886 plasma membrane (CC) 0.009375 8 GRM5, LTB4R1, CYSLTR1, TACR3, OXTR, PROKR1, NMBR, HTR2B GO:0042310 vasoconstriction (BP) 0.011425 2 EDN3, EDN1 mmu04080: Neuroactive ligand-receptor interaction 5.51E-08 7 GRM5, LTB4R1, CYSLTR1, TACR3, OXTR, NMBR, HTR2B mmu04020: Calcium signaling pathway 3.85E-05 5 GRM5, CYSLTR1, TACR3, OXTR, HTR2B ###### The top 15 hub genes rank in cytoHubba. Degree Stress MNC Radiality Closeness -------- -------- -------- ----------- ----------- IL6 IL6 IL6 IL6 IL6 GNGT1 UBC CXCL1 MAPK8 MAPK8 CXCL10 MAPK8 CXCL10 TLR4 TLR4 TLR4 TLR4 GNGT1 INS2 MMP9 CXCL1 GNGT1 MMP9 HGF CXCL10 MMP9 POMC CXCL2 MMP9 INS2 MAPK8 INS2 TLR4 POMC POMC CXCL2 SLCLA2 CCR5 CXCL10 HGF CCR5 HGF MAPK8 EDN1 CXCL1 UBC SYP CCL20 SLCLA2 EDN1 POMC CXCL10 CXCL5 CXCL1 CXCL2 CCL20 PVALB EDN1 UBC CCR5 CXCL5 EDN1 HGF CCR5 UBC HGF RAB3A CXCL3 CCL20 CCL20 EDN1 MMP9 CCR3 SYP GNGT1 [^1]: Academic Editor: Ernesto Picardi

Introduction ============ Transmigration is a dental pathology in which a migrating tooth germ passes across the median plane. This problem is very rare and is estimated to affect 0.1% of the general population. However, this disorder is being detected more frequently due to routine panoramic radiographs taken before patients undergo orthodontic treatment. Transmigration was first described by Thoma in 1952, and the term itself was originally used by Ando et al. \[[@b1-med-2015-0096],[@b2-med-2015-0096]\]. However, Tarsitano and collegues were the first to define transmigration as the passing of an unerupted tooth across the median line \[[@b3-med-2015-0096]\]. Javid claims that the term transmigration can be applied when at least half the length of a tooth has passed the median line \[[@b4-med-2015-0096]\]. Joshi and others consider this view to be incorrect. In their opinion, what is more important is the tendency of a tooth to migrate and pass the median plane rather than the distance a tooth has passed \[[@b5-med-2015-0096]\]. Howard observed a link between the migration of a tooth and the angle between the longitudinal axis of the impacted canine and the median line. Transmigration does not occur if the angle is 25--30°. An angle of 30--95° promotes transmigration \[[@b6-med-2015-0096]\]. Transmigration of a tooth germ most commonly occurs during maturation, when the alveolar processes grow intensely. Transmigrating teeth are frequently impacted in the bone, usually without causing any clinical symptoms \[[@b4-med-2015-0096]\]. However, some authors have described rare cases of displaced teeth erupting in places atypical for canine teeth. According to Joshi, in 20% of all cases canines erupt in the median line area, or on the opposite side of the dental arch, usually in the vestibular aspect. In the case of the longest recorded displacement, a canine migrated to the mesial root of the first molar. The rate of migration varies. Joshi observed a displacement of up to 3 mm during the course of one year \[[@b7-med-2015-0096]\]. Other authors describe cases of a canine remaining in a stable position over a period of up to five years \[[@b5-med-2015-0096],[@b8-med-2015-0096]\]. In percentage terms, transmigration occurs more frequently in women than in men. This condition most commonly affects the lower canines, although the literature also includes cases of transmigration in the upper arch. More cases of unilateral transmigration have been described than bilateral cases \[[@b3-med-2015-0096]\]. According to Costello, transmigration occurs more frequently in the right canines \[[@b9-med-2015-0096]\]. Mupparapu distinguished between five different types of transmigration. Type I is diagnosed when the coronal part of the tooth passes the median line and the canine is in a mesial or distal position in relation to the incisor and its long axis is inclined mesially. In type II, the canine is positioned horizontally, close to the lower border of the body of mandible and below the roots of the incisors. In type III, the canine erupts mesially or distally in relation to the canine on the opposite side. In type IV, the canine is in a horizontal position close to the lower border of the body of the mandible and below the root apices of premolars or molars. Finally, in type V the canine is in a vertical position, and its axis passes beyond the median line \[[@b10-med-2015-0096]\]. The etiology of transmigration is not fully understood. Javid and Joshi believe it is caused by an excessively high eruption force, which displaces the tooth germ beyond the median plane \[[@b4-med-2015-0096],[@b5-med-2015-0096]\]. However, Broadway claims that the process is due to the incorrect position of the tooth germ in the mandible \[[@b11-med-2015-0096]\]. Other etiological factors that need to be considered include odontogenic tumors, cysts \[[@b12-med-2015-0096],[@b13-med-2015-0096]\], premature loss of deciduous teeth, tooth crowding, supernumerary teeth, hypodontia, and an abnormal canine anatomy where the crown is excessively long \[[@b14-med-2015-0096],[@b15-med-2015-0096]\]. These can impede proper canine eruption and potentially promote transmigration \[[@b7-med-2015-0096],[@b16-med-2015-0096]\]. Some authors have highlighted cases of inflammation surrounding a deciduous tooth, which ruptures the band of cells connecting the tooth germ with the oral cavity \[[@b9-med-2015-0096],[@b14-med-2015-0096]\]. Mitchell has also described cases of canine migration caused by mandibular fractures in the area of growing tooth germs \[[@b17-med-2015-0096]\]. Holla, Saity, andParashar consider that Dental Class II malocclusion with deep bite, deep curve of Spee, excess tooth material, and increased lower anterior facial height indicating an increase in space available apical to the mandibular incisors may be predisposing factors in the etiology of the transmigrating canine \[[@b18-med-2015-0096]\]. Five procedures for treating canine transmigration have been described in the literature and the choice of method depends on the severity of the transmigration. The procedures are as follows: 1\) surgical extraction of a displaced canine with retarded eruption, 2) autotransplantation, 3) surgical exposure of a transmigrated tooth, 4) the alignment of the tooth in the dental arch using a fixed orthodontic appliance, and 5) observation of a displaced tooth without therapeutic intervention. Autotransplantation is recommended in cases where the lower incisors are in the correct position and when there is sufficient space in the dental arch to accommodate the transplanted canine. There have been no case reports of canine transmigration in patients with mandibular hypoplasia secondary to neonatal and infantile osteitis. The purpose of this paper is to present a series of three cases in which the transmigration of four lower canines was diagnosed. Case I ====== Female patient K.U. was treated for bilateral suppurative otitis during the neonatal period when she was eight days old, which resulted in sepsis. At the age of five, the patient was diagnosed with the following disorders: facial asymmetry and mandibular retrognathism secondary to right-sided temporo-mandibular joint ankylosis. We detected impaired mandibular kinetics in the form of limited opening and excursion to the left. During protrusion, the mandible deviated to the right. The patient was fitted with a functional orthodontic appliance type Klammt 1 with a reduced palatal plate to compensate for mandibular retrognathism and facial asymmetry. The mandible opening was improved to 37 mm. A panoramic radiograph taken when the patient was aged eight years and nine months revealed significant mesial inclination of both lower permanent canines in relation to the median plane. The lower right canine (tooth 43) was angulated at 86°, while the lower left canine (tooth 33) was angulated at 80°. In addition, there was no space for tooth 43. Using a fixed appliance, space was created for tooth 43. When the patient was aged 11 years and nine months, part of crown 43 appeared in the oral cavity, which was brought into the arch using elastics. Over a period of three years, tooth 33 migrated from a position distal to tooth 72 to a position mesial to the root of tooth 44. This tooth was inclined at 80° in relation to the median plane, which had not changed over the three-year period. A panoramic radiograph taken when the patient was aged 12 years and two months revealed a short and deformed ramus of the mandible on the right side. Horizontally positioned impacted tooth 33, which had transmigrated from the opposite side, was visible close to the border of the body of the mandible in the area of teeth 43--41. Its crown was oriented distally (towards the angle of mandible on the ankylosis side), while its root was positioned towards the median line. The crown of transmigrated tooth 33 was located between the root apex of tooth 44 and the lower border of the body of the mandible and was more radiopaque than the root, which indicated that the tooth was inclined in the vestibular-lingual plane. The angulation of the inclination was 94 degrees. The image of tooth 33 was blurred at mid height of the body of the mandible below the root apices of teeth 41 and 31. A CT scan showed that the lateral surface of crown 33 and part of its incisive cusp were covered only by a thin layer of compact bone, whereas the apex was oriented towards the lingual cortical plate. An odontogenic cyst was detected. The 1--2-mm thick lower border of the body of the mandible below crown 33 remained undamaged by the dentigerous cyst ([Fig. 1](#f1-med-2015-0096){ref-type="fig"},[2](#f2-med-2015-0096){ref-type="fig"}). The transmigration observed in this case corresponded to Class IV on the Mupparapu scale. Because of the cyst and the possibility of injuring the lower incisors when aligning tooth 33 in the arch, we decided to extract the tooth with the cyst capsule. Case II ======= Female Patient R.A., age 11, contracted a staphylococcal infection during the perinatal period, which developed into omarthritis on the right side as well as inflammation of the right temporo-mandibular joint. As a result, ankylosis of the temporo-mandibular joint occurred together while the right upper extremity was underdeveloped. At the age of 11, the patient sought orthodontic treatment. An examination revealed facial asymmetry, protrusion of the upper lip and significant recession of the chin. Scars caused by a mandibular distraction procedure were visible on the skin of right cheek. The distractor was founded at the age of 8 years. The following conditions were diagnosed: mandibular retrognathism, left-side displacement of the mandible, impaired kinetics of the mandible with an opening limited to 30 mm, a limited left-sided excursion, and an asymmetric protrusion. There was insufficient space for the lower left canine (tooth 33) ([Fig. 3](#f3-med-2015-0096){ref-type="fig"},[4](#f4-med-2015-0096){ref-type="fig"}). The diagnostics were confirmed with a CT scan. This made it possible to establish the presence of tooth 33, which was impacted and displaced beyond the median line. It was positioned vertically in the projection of the lower central incisors (teeth 41--31). Crown 33 was not covered from the vestibular aspect by the cortical plate, and its cusp was located at the height of the root apices of teeth 41 and 31, which were positioned lingually. The root apex of tooth 33 was oriented towards the lower border of the body of mandible and positioned midway along the body, between the vestibular and lingual cortical plate. Tooth 33 was rotated. This type of canine transmigration corresponded to Class V on the Mupparapu scale. The angulation was 62 degrees. It was decided to surgically expose tooth 33 in order to align it in the dental arch using a fixed appliance ( [Fig. 5](#f5-med-2015-0096){ref-type="fig"}). Case III ======== Female patient Ć.D. had been receiving orthodontic treatment since she was one year old due to left temporo-mandibular joint ankylosis (a complication of inflammation: local infections in the neonatal period can lead to generalized, dangerous infections called sepsis) and mandibular retrognathism. The child received an oral screen in order to train the orbicularis oris muscle and duce the distal position of the mandible. Symptoms of ankylosis were diagnosted when the child was 4 years old and could not open her mouth wide during spoon-feeding. Because of the recurring stiffness of the left temporo-mandibular joint, surgical release of this joint was performed three times when the patient was six, nine and 12 years old. Lateral and anterior-posterior cephalograms revealed that the ramus of the mandible on the side of the defect was shortened by five mm, and the body of mandible was shortened by six mm and recessed by 18 mm. The overjet was increased by 11 mm. The patient was fitted with upper and lower fixed light-archwire appliances. The treatment plan was to even out skeletal discrepancies in the form of mandibular retrognathism and open bite. A radiograph taken when the patient was aged 13 years revealed the absence of the coronoid and condylar processes in the mandible on the left side and the presence of an impacted upper left canine (tooth 23), lower left first premolar (tooth 34), lower left canine (tooth 33), and lower right canine (tooth 43). In the anterior segment of the mandible, which was shown in a number of projections, horizontally positioned teeth 33 and 43 were identified, which were displaced parallel to one another and below the incisor roots. Increased translucency around the displaced and impacted canines was evident, which suggested a dentigerous cyst. Tooth 43 was positioned transversally with the cusp between teeth 32 and 34, and the root apex was near to the lower border of the mandible body in the projection of tooth 42. The crown of tooth 43 was in a vestibular position and was partially covered by the cortical plate, while the apex of its root was close to the lingual cortical plate of the body of the mandible. Tooth 33 was positioned below tooth 43, horizontally along the lower border of the body of the mandible. The incisor cusp of tooth 33 was located in the projection of the root apex of tooth 42 and the root apex of tooth 34. The crown of the tooth was in a vestibular position, adjacent to the lower margin of the body of the mandible and was not covered by the cortical lamella in the cusp region. The root apex of tooth 33 was located midway along the body of the mandible in the region of its lower margin. An osteolytic defect indicating a forming dentigerous cyst was only evident around the crown of tooth 33 ([Fig. 6](#f6-med-2015-0096){ref-type="fig"},[7](#f7-med-2015-0096){ref-type="fig"}). Teeth 43 and 33 were in class IV transmigration according to Mupparapu. The angulation of the impacted canine 33 to the midline was 75 degrees and of the second impacted canine 43 -- 94 degrees. The impacted tooth 34 was exposed, and then orthodontic brackets were bonded to their surfaces. Elastics were engaged to bring these teeth into the arch ([Fig. 8](#f8-med-2015-0096){ref-type="fig"},[9](#f9-med-2015-0096){ref-type="fig"}). The transmigrated tooth 33 and tooth 43 were removed with a developing dentigerous cyst. The impacted tooth 23 was exposed as well and bring these teeth into the arch. Ethical approval ---------------- The research related to human use has been complied with all the relevant national regulations, institutional policies and in accordance the tenets of the Helsinki Declaration, and has been approved by the authors' institutional review board or equivalent committee. Informed consent ---------------- Informed consent has been obtained from all individuals included in this study. Conclusions =========== In the cases described in the present study, canine transmigration occurred in parallel with mandible hypoplasia, which was an osteitic complication that developed when the patients contracted general infections in the neonatal stage and in infancy. The available literature includes no discussion of the etiology of the disorder discussed in the present paper. In the present article, transmigration was diagnosed in adolescent girls. Other authors have also noted that the complication occurs more frequently in female patients and usually develops during puberty \[[@b4-med-2015-0096],[@b7-med-2015-0096],[@b19-med-2015-0096],[@b20-med-2015-0096]\]. One of the cases described above involved a very rare bilateral displacement of canines in the mandible - three of the four transmigrating canines had been displaced from the left to the right side. Costello claims, however, that this anomaly is more common in teeth on the right side \[[@b9-med-2015-0096]\]. The transmigrated teeth showed a tendency to erupt on the vestibular surface of the body of the mandible in the region of the median line or in the projection of the incisors. This could be confirmed by the absence of cortical bone covering the crowns of the transmigrated teeth. Such a tendency has also been observed by other authors \[[@b14-med-2015-0096],[@b16-med-2015-0096],[@b21-med-2015-0096],[@b22-med-2015-0096]\]. Due to the significant retrusion of the anterior segment of the mandible, the three cases presented above were difficult to diagnose based on standard radiological diagnostics. Panoramic radiographs proved to be useless because they represent flat (2D) pictures. Because of mandible hypoplasia, the chin area fell outside the imaging layer, which made X-ray images also useless. Therefore, it was decided to extend the imaging diagnostics through the use of CT. A reconstruction of the picture was obtained with transverse cross-sections of the mandible. In addition, a three-dimensional (3D) reconstruction was made, which allowed us to assess the position of the transmigrated canines with regard both to the vestibular and lingual cortical plate and to the roots of other teeth. This information ensured a better evaluation of the existing conditions, which proved helpful when selecting therapeutic management. There is no data in the available literature regarding the use of computer tomography in imaging transmigrations. Currently, the preferred diagnostic approach in such cases should include cone beam CT (CBCT) imaging as it reduces the radiation dose. The practice of surgical exposure and aligning a transmigrated canine into the dental arch was first described in 1994 by Wertz \[[@b23-med-2015-0096]\]. However, this approach is difficult and depends greatly on the maturity as well as the position and degree of tooth transmigration. Many authors have opted to observe this disorder, because the process is asymptomatic in most patients \[[@b3-med-2015-0096],[@b7-med-2015-0096],[@b15-med-2015-0096],[@b17-med-2015-0096],[@b19-med-2015-0096]\]. However, it is absolutely essential that a follow-up X-ray be taken, in order to assess any possible development of dentigerous cysts, root resorption of adjacent teeth or potential infection foci. These signs are indications for extracting a tooth \[[@b14-med-2015-0096]\]. It is also important to be aware of any possible neurological symptoms caused by the compression of a transmigrated tooth on nerve endings \[[@b3-med-2015-0096],[@b4-med-2015-0096],[@b14-med-2015-0096]\]. In the case of our study, we decided to extract transmigrated canines if they were horizontally positioned along the lower border of the mandible body (class II according to Mupparapu) and if we observed signs of a dentigerous cyst. Two-thirds of the length of one of the teeth was displaced beyond the median line, and up to half of the length of the other was similarly displaced. The remaining two canines were surgically exposed in an attempt to align them in the dental arch. Our observations indicate that patients with mandibular hypoplasia secondary to osteitis in the growth period should undergo screening with panoramic radiographs at the age of eight to nine years (when canine roots achieve a third of their maturity) to ensure early detection of a canine's eventual tendency to transmigrate. Extending imaging diagnostics to include CT, most preferably CBCT, makes it possible to precisely evaluate a transmigrated canine's position and to plan a course of treatment. Treatment planning in such cases should be performed by a team comprising an orthodontist and a surgeon. Such a management approach makes it possible to determine the optimal time for commencing treatment and to choose the optimal procedure for teeth at risk of transmigration. **Conflict of interest statement:** Authors state no conflict of interest {#f1-med-2015-0096} {#f2-med-2015-0096} {#f3-med-2015-0096} {#f4-med-2015-0096} {#f5-med-2015-0096} {#f6-med-2015-0096} {#f7-med-2015-0096} {#f8-med-2015-0096} {#f9-med-2015-0096}

Introduction {#S1} ============ From recent employers' surveys, behavioral or emotional intelligence competencies, also defined as 'soft skills' in the labor market's jargon ([@B91]), are becoming the most critical capabilities for organizations to acquire at all levels ([@B70]; [@B85]). These competencies are assuming a growing relevance for the workforce of the future, especially in the age of automation and artificial intelligence which are progressively replacing routine tasks and jobs ([@B74]; [@B27]; [@B113]). The literature that has extensively contributed to the debate on emotional, social, and cognitive intelligence has primarily identified three main theoretical streams and related measurement models ([@B7]; [@B28]; [@B110]; [@B20]). The first stream conceives emotional intelligence as a mental ability which encompasses four branches, namely the ability to perceive emotions, the ability to use emotions to facilitate thought, the ability to understand emotions, and the ability to manage emotions ([@B72]). The second stream considers emotional intelligence as a personality trait and the major contributions have been provided by the [@B11] model which contemplates five main components (intrapersonal skills, interpersonal skills, adaptability, stress management, and general mood) ([@B11]) and by the [@B83] model which consists of four components (well-being, sociability, self-control, and emotionality). The third stream focuses on competencies as behavioral manifestations of emotional intelligence ([@B46]; [@B20]). Specific behavioral competencies are grouped into five main clusters: (1) the self-awareness cluster, which concerns knowing one's internal states, preferences, resources, and intuitions, consisting of emotional self-awareness; (2) the self-management cluster, which enables the management of one's internal states, impulses, and resources; (3) the social awareness cluster, which refers to how people are aware of others' feelings, needs, and concerns; (4) the relationship management cluster, which includes abilities that induce desirable responses in others; and (5) the cognitive cluster, namely the ability to analyze information and situations. Notwithstanding the ongoing debate on the different definitions and measurement models of emotional intelligence - which does not represent the topic of this article - empirical research has provided extensive evidence that the behavioral manifestation of emotional intelligence plays a crucial role in preparing individuals for contemporary careers, and subsequently in favoring their employability and entry-level success in the job market ([@B44]; [@B57]; [@B31]). This confirms that acquiring a proper set of emotional, social, and cognitive competencies during the education experience or in the preliminary stages of the career path can greatly enhance the chances of people's success during the recruiting and selection process. Moreover, the aforementioned competencies are also labeled life skills ([@B47]; [@B33]), since they are required to deal with the demands and challenges of everyday life, and thus necessary along one's career path to maintain employability over time and to attain career success ([@B1]). However, academic studies and employers' surveys have still revealed, on the one hand, a mismatch between the expectations of companies and the behavioral competency profile of candidates ([@B8]; [@B85]), and on the other hand, a misalignment between the types of competencies learned by students attending higher education programs and those required by the labor market ([@B94]). This skills gap poses a challenge to organizations and to educational institutions, whose responsibility is to formally support the development of emotional and social competencies (ESCs) along with professional/technical skills. When in their book "Competence at Work," [@B100]: 12) stated that "you can teach a turkey to climb a tree, but it is easier to hire a squirrel," they meant that behavioral competencies are more complex and difficult to develop than professional or hard skills; thus pursuing their development can imply challenges, such as the different role instructors have to assume, the active engagement of the learners, and the more experiential pedagogical approach teachers need to adopt ([@B13]). Since behavioral competencies are currently the most in-demand skills in the labor market, and a wide body of research since the 1980s has demonstrated their positive impact on individual performance, career success, and wellbeing across sectors and professional roles ([@B18]; [@B100]; [@B58]; [@B67]; [@B110]), a large number of theoretical studies on how these ESCs can be effectively developed has emerged over time. The purpose of this paper is to deliver a systematic review of the literature on this topic in order to make the following contributions: (i) provide the first comprehensive critical analysis of the distinctive features of the theoretical and methodological frameworks adopted for designing ESCs learning experiences; (ii) review the contexts in which these training initiatives have been delivered and the categories of learners involved; (iii) discuss the learning outcomes of the educational programs and how they have been assessed; (iv) based on the identification of knowledge gaps and inconsistencies in the current state of the literature, suggest promising paths for future research; and (v) stimulate valuable insights for educators, human resource managers, executives, and policymakers, by organizing and critically analyzing the extant contributions on competency development. The paper is organized as follows. In the next section, the methods and the data set are introduced. Specifically, we describe the criteria adopted for selecting the articles that contributed to the debate on competency development. Subsequently, we describe the characteristics of the participants who attended the educational programs on ESCs. Furthermore, we analyze the different pedagogical frameworks and the methodological approaches implemented in training behavioral competencies, as well as the ESCs most frequently targeted by these programs. Then, we critically examine how the effectiveness of these educational initiatives has been measured. We conclude by discussing the gaps and the inconsistencies in the current literature, suggesting how future research should overcome them. Overview of the Method {#S2} ====================== The data collection process was preceded by a thorough preparatory stage, during which the authors conducted an exploration of the extant literature ([@B106]). This activity allowed the identification of the subject areas in which the debate has been developed and the keywords to use for the searching query. We decided to focus our attention on the developmental and learning processes of emotional, social, and cognitive competencies that occur after formal training or educational courses provided by universities or certified organizations and directed to students or practitioners. Our intention, in fact, is to disentangle the development of behavioral competencies as a result of on the job learning ([@B37]) from the effects that training programs have on the same skills, by focusing on the latter. We conducted an exploratory analysis using the following search query: \[("develop^∗^" OR "learn^∗^") AND ("emotional intelligenc^∗^" OR "social intelligenc^∗^" OR "cognitive intelligenc^∗^" OR "social competenc^∗^" OR "emotional competenc^∗^" OR "cognitive competenc^∗^" OR "behavioral competenc^∗^" OR "behavioral competenc^∗^" OR "behavioral skill^∗^" OR "behavioral skill^∗^" OR "emotional capabilit^∗^" OR "social capabilit^∗^" OR "cognitive capabilit^∗^" OR "behavioral capabilit^∗^" OR "behavioral capabilit^∗^" OR "emotional skill^∗^" OR "social skill^∗^" OR "cognitive skill^∗^") AND ("education")\]. We searched the terms in the keywords, titles, and abstracts of documents contained in two databases, Scopus and Web of Science, recognized by the scientific community as having wide coverage ([@B115]). Moreover, we limited our research to articles published in English and to the three subject areas that were shown to welcome the largest part of the debate, namely business and management, psychology, and social sciences. The business and management field was included to capture development programs aimed at students and practitioners which have an impact on organizations through their actual or future human capital. The psychology area was considered with the aim of acquiring a deep understanding of the psychological dimensions, processes, and outcomes that characterize behavioral competencies training and development programs. Finally, we considered the social sciences area in general in order to include those contributions neglected by the other categories, such as studies in learning and education. Since the aim of our review is to contribute to the existing academic literature while relying on well-established and impactful knowledge, we decided to limit the analysis to peer-reviewed academic articles. Indeed, the search query was defined to exclude conference papers, books, and book chapters. Applying the aforementioned keywords, Scopus and the Web of Science databases retrieved 6,106 unique articles, after checking for duplications. Starting from the consideration that medium to top journals publish works that have an impact on academic research, we refined our review by considering only studies published within the first 300 journals in the SCImago list in the areas of business and management, psychology, and social sciences. The titles and abstracts of these articles were checked to assess their fit with our research purposes. In so doing, we also applied some exclusion criteria and boundary conditions. Firstly, we excluded from the data set those articles focused on children and elderly people as target subjects, since our interest is oriented toward research dealing with the development of ESC competencies in people of working age or approaching the labor market. As a result, we focused on studies centered on adolescents/students and adults. We also excluded studies dealing with physical and behavioral disorders, such as autism, anxiety, and depression. The reason lies in the consideration that a focus on these pathologies can alter the design of the program, whose aim is to cope with these disturbances, and consequently the effectiveness of the learning outcomes. Finally, we excluded research conducted at the group or organization level of analysis, while considering those studies centered on individuals' competency development. We did not insert any time limit. This selection process allowed us to collect a final sample of 71 articles published within the period 1982--2020. This selection process further allowed the identification of the scholars who contributed the most to the debate under investigation. [Figure 1](#F1){ref-type="fig"} provides a visual representation of the data collection flow. {#F1} Data Set Description and Data Analysis {#S3} ====================================== The final sample consists of 71 peer-reviewed articles, 38 of which present or test the effectiveness of programs that develop behavioral competencies. The remaining works encompass both theoretical and empirical studies aimed at (i) reviewing different learning approaches for enhancing ESCs (e.g., [@B56]; [@B75]; [@B81]); (ii) testing the impact of competency development on variables such as life satisfaction, stress reduction, performance (see, for instance, [@B26]; [@B1]; [@B34]); and (iii) valuing or discussing the impact and use of different tools and methods for the aforementioned purposes (e.g., [@B68]; [@B112]; [@B17]; [@B77]). For each of the 71 articles, we examined the theoretical frameworks adopted, coherently with the different streams of research illustrated in the introduction section. From this analysis emerged that more than half of the works (54%) explicitly referred to the behavioral approach to emotional and social intelligence competencies, whereas, respectively, 7% and 6% of the articles pertain to the mental ability and the trait models. A few articles combine the different streams of research (2%), or adopt other theoretical framework such as the cultural intelligence approach (4%) or the cognitive skills (6%). The remaining 21% of the articles do not explicitly refer to a theoretical model, focusing instead on the impact of the training programs on few specific skills measured through disparate self-rating scales. This evidence shows the emphasis ascribed to the behavioral stream and to the other self-report approaches in addressing the issue of emotional and social intelligence development. As previously discussed in the literature ([@B110]), this can be due to the challenges posit by the other approaches and associated measures used for the assessment of emotional and social intelligence. Focusing on the contributions illustrating competency development programs, extant research suggests that developing behavioral competencies requires pedagogical approaches and related didactic techniques which actively engage the learner ([@B57]). People, and adults in particular, learn what they want to learn and, as a consequence, individuals who want to develop their ESC competencies should be conscious of what, how, and for what purposes they want to change or improve their set of behaviors ([@B19]). We treated learners' awareness as a discriminant factor when we analyzed the contributions presented by the educational programs. The results of the analysis suggest that literature splits into two broad categories. The articles included in the first category describe and test the effectiveness of courses and programs designed and developed with the clear purpose of improving behavioral competencies. This is the case of the Leadership Assessment and Development (LEAD) course implemented in an MBA curriculum at Case Western Reserve University ([@B22]; [@B88]; [@B82]) or the Emotional Intelligence Training Program (EITP) ([@B45]), to cite just two. [Table 1](#T1){ref-type="table"} summarizes these works and their main findings. ###### Courses designed for ESCs development. Study Main findings --------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- [@B9] The Professional Fellows Program for midcareer professionals is a learner-centered, non-graded program that provides participants with the tools for understanding and facing current leadership challenges, while encouraging the creation of a personalized learning agenda and a network of peers that promote lifelong learning. An assessment of the program is provided along with its description. [@B10] The application of a whole person learning approach to family business education proves its effectiveness for developing the ESC competencies of the next generation of entrepreneurs. [@B23] The article combines cross-sectional and longitudinal data on multiple cohorts of MBA attendees. Results show that emotional and social competencies can be improved in MBA programs that are not only based on traditional knowledge acquisition methods, but that adopt a more holistic approach to students' development. [@B24] The Ecotonos cross-cultural simulation game helps increase the metacognitive, motivational, and behavioral components of participants' cultural intelligence. [@B26] The article examines the learning that occurred during an MBA, measured in terms of competency development and its impact on career advancement and compensation. [@B36] The MindOut social and emotional learning program targeting adolescents shows its effectiveness for improving students' emotional and social skills and wellbeing. [@B41] The Happy emotional education program is a gamified tool that supports adolescents' emotional competency development. [@B45] The Emotional Intelligence Training Program, in its online, in-class, and coaching versions, can effectively improve students' emotional intelligence competencies. [@B50] The reflecting team case model is presented and described as an alternative didactic method for improving participants' ESC intelligence competencies. [@B53] The Community Approach to Learning Mindfully (CALM) program for educators proves its effectiveness in promoting teachers' mindfulness and emotional functioning, with effects in terms of individuals' wellbeing. [@B54] The Tuning into Teens (TINT) parents' emotion socialization program reveals an effective course for improving parents' emotional responses to their adolescent children's behaviors, while reducing family conflicts. [@B57] The study presents, discusses, and assesses the effectiveness of the whole person learning approach, confirming its soundness for developing individuals' ESC intelligence competencies. [@B60] The Cultivating Awareness and Resilience in Education (CARE) mindfulness-based program for teachers enhances teachers' mindfulness and emotion regulation, and reduces the level of their distress. [@B69] The Girls First Resilience Curriculum that targets girls in high poverty conditions displays its effectiveness for improving participants' emotional resilience, social and emotional assets, self-efficacy, and psychological and social wellbeing. [@B78] The paper describes the features and use of consciousness-enhancing experiences for executives. [@B82] A longitudinal assessment of the Leadership Assessment and Development (LEAD) course implemented in an MBA program has been used by the authors to understand how patterns of competency development change over time and what factors intervene in this change. The authors identify five factors that affect competency development. [@B87] Cross-cultural management education increases students' cultural intelligence, along with their commitment and satisfaction with the course. [@B88] This longitudinal study is based on the Leadership Assessment and Development (LEAD) course implemented in an MBA program for understanding the dynamics of ESC competency development. Results indicate that different ESC competencies are developed at different periods throughout the course. [@B89] The article presents the features of the MS program in Executive Leadership and Organization Change (ELOC), which is based on the action learning, competency development, and teamwork pedagogical framework. The program aims to develop leaders who provide outstanding performance. [@B96] The PREDEMA 2.0 program targeting university students improves students' emotional competencies and wellbeing in the short run. [@B97] The PREDEMA program aims to boost adolescents' emotional competencies in order to promote their wellbeing and improve the quality of relationships they build with their peers. The program is effective for reducing cyberbullying behaviors. [@B98] Adopting a classroom-as-organization (CAO) approach significantly contributes to the development of students' emotional intelligence competencies. [@B101] The Cleveland Clinic Academy leadership development course is used as a point of reference in providing guidelines for leadership development in healthcare. [@B107] The authors test and confirm that different types of mental training intervention (targeting attention, socio-affective skills, and socio-cognitive competencies) intervene in different areas of the human brain, thus affecting people's behaviors. [@B109] Simulations are effective methods for developing teamworking and team coordination capabilities in medical emergency teams. In the second category, we considered those programs for which the improvement of individuals' behavioral skills is a collateral, although welcome, byproduct of courses designed with different learning objectives. A synthesis of these studies is provided in [Table 2](#T2){ref-type="table"}. For instance, in a recent longitudinal study, [@B48] show that attending pre-school has long-term effects on the development of Chinese adolescents' social skills. Similar findings are discussed by [@B80], who discovered that adolescents advancing in formal education improve their non-cognitive skills, an umbrella term that indicates competencies such as conscientiousness, self-efficacy, and future orientation. More generally, and moving to graduate students, it has been found that MBA programs support the development of attendees' emotional intelligence ([@B105]). ###### Studies designed with different purposes that develop ESCs as a byproduct. Study Main findings --------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ [@B3] The peer-tutoring program (PTP) positively influences tutors' social skills. [@B15] Academic courses relying on traditional learning methods should complement these methods with individual experiential learning and social experiential learning to develop students' ESC skills. [@B25] In courses aimed at enhancing declarative and procedural knowledge, students' attending face-to-face, instead of web-based, classes report higher levels of cognitive and social competencies. [@B39] Students assess simulation as the most effective method for developing interpersonal skills and self-awareness. Simulation and case study are believed to improve students' problem-solving competencies more than lectures. [@B48] Pre-school attendance has a long-term impact on adolescents' social skills. [@B61] Academic entrepreneurial programs should not only focus on teaching the 'know what' (content knowledge), but also work to develop the 'know why' (attitude and motivational skills) and the 'know how' (social skills). [@B76] A cooperative team learning approach positively affects the development of students' social competencies. [@B79] The incorporation of elements of the liberal arts, especially visual arts and poetry, can enhance students' self-awareness and recognition of their emotions. [@B80] Continuing formal education after adolescence and young adulthood improves persons' non-cognitive skills such as conscientiousness and future orientation. [@B90] Youths participation in after-school programs positively impacts on their concentration and emotion regulation skills. [@B95] Adolescents who participate in an 8-week training using Pokémon-Go display improvement in selective attention, concentration, and sociability levels. [@B105] MBA programs support participants in developing their emotional intelligence competencies. [@B108] Enterprise education equips students with behavioral competencies such as self-confidence, communication capabilities, creative thinking, and problem solving. [@B114] Teaching strategies relying on inquiry-oriented class discussions, *ad hoc* examination practices, and evaluations by both peers and graduate students are effective in developing students' higher-order cognitive skills (HOCS) (i.e., decision-making, problem solving, critical thinking). The following sections will present the results of the analyses we conducted on the selected articles. We will focus our attention on the characteristics of participants in educational programs, describe the methodological approaches that constitute the theoretical foundations of each program, evaluate the nature of the developed competencies, and conclude with remarks on the effectiveness of the considered programs in light of their purposes. Participants Involved in Competency Development Programs {#S3.SS1} -------------------------------------------------------- Our analysis of the articles in our data set, considering both programs specifically designed to develop ESC competencies and those designed with different purposes, allows us to identify three groups of participants considered by the extant research on behavioral competency development: (i) youths and adolescents, mostly high school students (the sampled individuals in this category are 12 to 18 years old); (ii) undergraduate and graduate students, also encompassing full-time MBA attendees; and (iii) adults and professionals, namely people who are actively employed in a working position and for which universities and/or organizations offer training programs. Research targeting youths and adolescents is focused on two fundamental aspects: individuals' well-being (see, for instance, [@B69]; [@B36]) and promotion of the coexistence of youths with their peers (e.g., [@B97]). The development of behavioral competencies is believed to support adolescents in coping with the transitions that occur during this age and that are characterized by many possible stressors ([@B36]). The process of maturation in fact brings both biological and psychosocial changes that affect the lives of youths and adolescents. It is during this period that individuals start to construct their identity, experiencing conflicts with their family, seeking acceptance among their peers, and laying the foundations of their future lives and careers ([@B36]). "Programs aim\[ing\] to enhance adolescents' social and emotional skills \[...\] act as protective factors for promoting positive wellbeing and reducing the onset of mental health difficulties and health risk behaviors" ([@B36]: 1258). Furthermore, encouraging the development of adolescents' behavioral competencies is a strategic action that can improve the co-existence of youths with their peers and prevent the occurrence of bullying behaviors ([@B41]; [@B95]; [@B97]). Emotional and social competency development programs are meant to provide individuals with the tools to better interact and relate with their peers and teachers, while improving the recognition and management of their emotions to reduce conflicts and aggressiveness ([@B97]). The second identified category considered students who attend competency development programs in the higher education setting. The motivations behind the design of the development programs presented by these works lead us to distinguish between two subgroups of target subjects: undergraduate/graduate students and MBA attendees. As suggested by [@B96], individuals leaving the high school world and entering a university career can go through a period of stress due to the increased amount of responsibilities and expectations coming from the environment. They become independent, moving from their hometown or home country; they are required to build social relationships and to face the academic challenges while coping with the competition of the university environment. They also bear the burden of imagining and projecting their lives after graduation. Emotional training programs aimed at developing students' behavioral competencies can support individuals going through these unavoidable steps, with a positive impact on their wellbeing ([@B96]). In addition, university programs have to prepare students for the job market. The current business landscape asks new job market entrants to be effective in translating the concepts and behaviors learned during the academic path into an organizational reality characterized by complexity and dynamicity ([@B45]). Furthermore, globalization asks workers to interact and collaborate with people from different countries and cultures ([@B24]). It is not surprising that the development of cultural intelligence has entered educators' learning agendas during the last decades, and it has become the focus of a number of educational programs ([@B87]). Cultural intelligence is defined as "a person's capability for successful adaptation to new cultural settings, that is, for unfamiliar settings attributable to cultural context" ([@B38]: 9). Furthermore, organizations seek people with sound social skills who are able to enrich and enhance their own and the organization's knowledge and capabilities through lifelong learning ([@B76]). In this scenario, the development of emotional and social competencies becomes of paramount importance ([@B24]; [@B45]; [@B76]). This explains the presence of competency development courses within MBA programs whose aim is to prepare the future managers and leaders ([@B23]; [@B82]; [@B105]), who should impact and transform their organizations and communities for the better ([@B89]). It is well established in the academic literature that advanced professionals, and adult learners in general, have different learning needs compared with young people or university students ([@B9]). The articles that belong to the third category highlight the role that ESC competencies play in particular groups of subjects: educators and managers. Parents, as educators, play a key role in influencing their children's emotional development ([@B54]). For this reason, it is of paramount importance that parents learn to recognize and manage not only their own and but also their children's emotions. In youths' development, another key role is played by teachers. They are responsible not only for the quality of education, but also for providing stimuli and learning opportunities that boost students' curiosity and performance ([@B53]; [@B60]). However, according to [@B60], about 50% of K-12 teachers in the US attest to being under pressure and experiencing a lot of stress in relation to their job. The causes are attributable to factors such as being constantly under the judgment of students, parents, and school principals; being asked to manage students' misbehaviors, emotional reactions, and scarce interest in school topics; and feeling unable to perform a job in which there is no working time and for which they need to take the work home ([@B75]; [@B53]; [@B60]). Programs that aim to reduce teachers' stress have been designed and implemented to support teachers in coping with these possible critical situations. Reducing teachers' stress is a key step in the creation of a more relaxed and enjoyable climate in class, which has an impact on students' emotional responses to in-class activities and on their performance ([@B53]; [@B60]). In the organizational setting, considerable efforts and resources have been spent on development programs that target managers' competency portfolio. Universities and business organizations acknowledge that having great leaders contributes to organizational success and is instrumental in the achievement of a sustainable competitive advantage. In this scenario, the development of ESC competencies sustains managers' outstanding performance and supports them in facing the challenges of the current business environment ([@B78]; [@B63]). Surprisingly, in the articles analyzed, we found that some studies have primarily focused only on educational programs delivered to professionals operating in the healthcare sector. In this particular setting, most managerial and leading positions are assigned in response to academic performance or cognitive ability. However, being effective doctors, medical researchers, or nurses does not necessarily reflect an effective manifestation of social and emotional competencies ([@B101]). Successfully managing, coordinating, inspiring, and working in teams are crucial behaviors both for facing emergencies and coping with the everyday challenges that hospitals and clinics face. In such complex organizations, where there are a lot of stakeholders with different roles and requirements, leaders are asked to take forward a culture of effective coordination and to manage the trade-off between costs, access, and quality of health services ([@B109]; [@B101]). These skills are the products of learning and practicing efforts trained by *ad hoc* courses such as team coordination programs ([@B109]). Pedagogical Frameworks {#S3.SS2} ---------------------- This section presents the results of the analysis we conducted on the theoretical foundations and methodological approaches of the didactic programs purposely designed to enhance behavioral competencies ([Table 1](#T1){ref-type="table"}). Thus, we did not consider those studies in which the competency development was an unintentional outcome of the learning experience ([Table 2](#T2){ref-type="table"}). A common thread characterizes these development programs: despite relying on different frameworks, these studies agree on the importance of experiential learning methodologies to improve behavioral skills. The review shows that two main theoretical approaches are extensively adopted in designing these programs. The first is Kolb's experiential learning theory ([@B66]), furtherly advanced by programs like the Leadership Assessment and Development (LEAD) course ([@B23]; [@B88]; [@B82]), or the Executive Leadership and Organization Change (ELOC) program ([@B89]). According to Kolb, the process of learning from experience encompasses an understanding and re-elaboration of new experiences that adds to previous ones, modifying and enriching them. The experiential learning cycle explains this process and depicts the four sequential phases through which individuals elaborate and internalize experiences: (i) concrete experience of a new behavior, thought or action; (ii) reflective observation of how the individual or other people behaved or acted in the same situation or under similar conditions; (iii) abstract conceptualization of the outputs of these actions or behaviors, in terms of both individuals' feelings and cause-effect relationships; and (iv) active experimentation of the experience enriched by a new conceptualization of the same ([@B65]). The second approach is the whole person learning pedagogy ([@B92]), which has been applied in programs illustrated, for instance, by [@B57] and [@B10]. Specifically, [@B57] describe the approach they implemented in an MBA course to respond to the needs of enterprises that seek employees who possess teamworking, communication, conflict resolution, and problem-solving skills. When accompanied by an assessment of the ESC competencies, the whole person learning pedagogy is revealed to be an effective method for acquiring and retaining skills that allow executives to be successful in their roles. This approach is an extension of the concept of whole person learning presented by [@B92]. In particular, it is a learning pedagogy that impacts not only on people's cognitive capabilities, typically involved in a learning process, but also on their emotions and behaviors: hence the term 'whole person.' Involvement of the cognitive, emotional, and behavioral dimensions supports engagement and motivation toward the learning process ([@B57]). This pedagogy has proved to be effective both when implemented into MBA courses and for developing the skills of undergraduate students involved in their family businesses ([@B57]; [@B10]). Further support for these results is provided by scholars who focus on programs that strongly rely on experiential methods for developing teachers' competencies such as self-awareness, self-control, resilience, empathy, and compassion ([@B75], [@B60]). Finally, [@B24], and [@B81] contend that experiential training is of paramount importance when developing managers and, more generally, practitioners' cultural intelligence competencies. Literature introducing and testing the effectiveness of programs for ESCs development draws on a number of different experiential and person-centered methodologies that constitute the application of different learning techniques. Based on the methodology implemented in the educational program, we group these studies as follows: programs based on the intentional change theory (ICT; [@B23]; [@B88]; [@B82]); on the social and emotional learning (SEL) approach ([@B32]; [@B103]; [@B41]; [@B97], [@B96]; [@B36]); on mindfulness-based programs ([@B75]; [@B35]; [@B53]; [@B60]; [@B107]); on cultural intelligence development courses ([@B24]; [@B87]; [@B81]), and on other pedagogies which rely on experiential learning such as consciousness-raising experiences and the classroom-as-organization approach ([@B78]; [@B98]). The rest of this section is devoted to describing in more detail the differences among these methodological approaches and the behavioral competencies nurtured by these programs. Originally defined as self-directed learning theory, the ICT describes the essential components that lead an individual toward a desirable and sustainable change in his behaviors, thoughts, feelings, and perceptions ([@B19]). The LEAD program ([@B23]; [@B22]; [@B88]; [@B82]), implemented in an MBA curriculum since 1987 at Weatherhead School of Management, Case Western Reserve University, follows the ICT pedagogy to support students in the development of their emotional and social competencies. In harmony with Boyatzis' ICT, the course is structured to guide participants throughout five phases, also defined "discoveries" or "epiphanies": (1) identification of the ideal self and development of a personal vision; (2) discovery of the real self; (3) creation of a personal learning path; (4) the practice of desired behaviors, which includes reflections on actions and feelings; and (5) the creation of relationships that support individuals during the training phase, such as the involvement of participants in coaching sessions with trained professional coaches ([@B82]). Inspired by the LEAD course, the Professional Fellows Program is aimed at midcareer professionals who are entering managerial positions. The program guides participants onto a path where they are introduced to leadership notions and challenges, and they have also the opportunity to assess their competencies, develop their learning agenda, and construct a network with other students, with the aim of sustaining each other's lifelong learning commitment ([@B9]). Both LEAD and the Professional Fellows Program integrate experiential learning theory with the whole person learning approach ([@B9]; [@B22]). Social and emotional learning (SEL) is a concept born in 1994, when a group of researchers and professionals met at the Fetzer Institute (Kalamazoo, MI, United States) to discuss the importance and opportunity of inserting the development of social and emotional competencies into school curricula. In the same year, Daniel Goleman and Eileen Growald, who participated in this meeting, founded the Collaborative for Academic, Social, and Emotional Learning (CASEL), whose aim is to support the inclusion of SEL for scholars and teachers ([@B32]). SEL is defined as "a strategy to nurture students' social and emotional competences by way of explicit teaching. SEL uses a student-centered approach that encourages student participation in the learning process and in the development of analytical communication and collaborative behaviors" ([@B32]: 2). The SEL approach contends that schools should explicitly teach emotional and social competencies, along with the cognitive ones already taught, applying a more holistic approach through which the learning of these skills occurs not only in class but also through the involvement of families and communities. In this scenario, a key role is played by teachers, who become the promotors and the cornerstone in the implementation of SEL's purposes ([@B32]). The SEL approach targets five ESC competencies: self-awareness, self-management, social awareness, relational competencies, and responsible decision-making, integrating the teaching of these skills into school curricula and practices ([@B32]; [@B103]). An example is provided by the MindOut program, developed to support older adolescents' emotional and social development, and designed according to CASEL's theoretical framework. This program is based on a number of experiential learning practices such as collaborative learning, creation of scenarios, games, and simulations, and it adopts a holistic approach to continuing the training outside the classroom and at home ([@B36]). Another SEL program that targets adolescents is presented and tested by [@B97]. This program, named PREDEMA, aims to enhance the understanding and management of students' emotions, contributing to build a positive climate in the classroom. Indeed, PREDEMA's ultimate goal is to improve students' co-existence and to facilitate the relationships between students and teachers, while reducing cyberbullying behaviors and enhancing students' wellbeing in the long run. The program is based on the four-branch model of emotional intelligence ([@B71]) describing and categorizing the emotional intelligence skills into: (1) the ability to perceive emotions; (2) the ability to use emotions to sustain the cognitive process; (3) the ability to understand or 'label' emotions; and (4) the ability to manage emotions. The first group of skills (perceiving emotions) contains competencies that reflect a lower level of complexity. When moving to the other groups, however, the tasks become more challenging and the competencies more complex. The PREDEMA program, as well as its 2.0 version targeting university students ([@B96]), has its roots in the dialogical paradigm of learning, wherein the learning process is the result of communicative actions between learners and the subjects with which they interact ([@B86]). Moreover, responding to a need to boost emotional education, [@B41] present and evaluate the effectiveness of Happy Software 12--16. This gamified program was developed to support adolescents in the management of their emotions. It is based on the model of emotional competencies presented by the Group of Psychopedagogical Orientation, which defines five categories of emotional skills: emotional awareness, emotion regulation, autonomy (e.g., self-esteem, motivation, resilience), social skills, and life competencies (e.g., adaptivity, seeking help, personal wellbeing, ability to contribute to community, society's wellbeing) ([@B41]). Mindfulness-based interventions (MBIs) represent another set of practices adopted to nurture behavioral competencies. Mindfulness is defined as the awareness that derives from purposefully and non-judgmentally paying attention to current experiences while maintaining "openness, curiosity, and acceptance toward \[these\] experiences" ([@B75]; [@B53]: 144). It has found a plethora of applications and related exercises, such as breathing meditation, body scan, loving kindness meditation, and yoga, among others, to cultivate in participants present moment attention and self-awareness, but also affective qualities of care, gratitude, and loving kindness, as well as social awareness through perspective taking ([@B75]; [@B35]). The strength of mindfulness-based programs is that they create the mental conditions under which behavioral competencies can be manifested and enhanced. These programs support trainees' resilience and wellbeing, reduce stress, improve the way people relate and interact with others, and promote people's commitment to change ([@B75]; [@B35]). Mindfulness-based programs intervene in the brain's plasticity---specifically, in the areas of the brain devoted to emotion regulation ([@B75])---and contribute to enhancing competencies such as empathy, creative thinking, teamworking and collaboration, social awareness, and self-awareness, to name a few ([@B75]; [@B35]; [@B60]). Besides the well-known mindfulness-based stress reduction program (MBSR) ([@B62]), a number of programs are mentioned, tested, and analyzed in the literature we reviewed. [@B53], for instance, test the efficacy of a yoga-based intervention, the Community Approach to Learning Mindfully (CALM), to support educators' wellbeing. According to the authors, mindfulness, along with other contemplative practices, is at the base of yoga exercises. In a literature review published in 2012, Meiklejohn and other scholars compare the effectiveness and features of three mindfulness-based experiential training programs aimed at improving teachers' resilience and wellbeing: the Mindfulness-Based Wellness Education (MBWE) program; Stress Management and Relaxation Techniques (SMART) in education; and the Cultivating Awareness and Resilience in Education (CARE) program. In this last case, [@B60] empirically validate the positive effects of the program on teachers' mindfulness, distress, and emotion regulation. Finally, a recent study by [@B107] has tested and confirmed that different types of contemplative or meditation-based practices induce plasticity in different areas of the human brain, suggesting to researchers and practitioners the need for more exploration of targeted interventions devoted to developing different types of skill. A group of articles included in our analysis focus on the development of cultural intelligence competencies, confirming that, as happens for emotional intelligence skills, cross-cultural cognitive and behavioral capabilities can be improved through dedicated training and development programs ([@B87]). Scholars contend that courses that embrace experiential learning principles show their effectiveness in developing cultural intelligence, especially when combined with experiences abroad and interactions with people from different countries ([@B81]). This is confirmed by [@B24], who tested the effectiveness of a cross-cultural simulation-based game, Ecotonos. Moving from contributions that support the need to complement knowledge acquisition with experiential learning practices, this study supports the effectiveness of games as a dimension for training new cultural behaviors in a safe environment. Games, in fact, allow evaluation of the consequences of practiced behaviors and the implementation of corrections if such behaviors do not provide the expected results ([@B24]). Consciousness-raising experiences are practices that have their roots "in sociopolitical discourse and empowerment movements ([@B43]), feminist theory and study groups ([@B99]), myriad forms of psychotherapy ([@B84]), and many spiritual disciplines ([@B59])" ([@B78]: 174). However, they prove their effectiveness even when applied to the business context ([@B78]). Consciousness-raising experiences in organizational environments refer to a set of activities, usually part of leadership development programs, that aim at cultivating executives' self-and social awareness. They can be in the form of sessions/meetings in which people speak about their lives, work, aspirations, what motivates and guides them, and the challenges they face. They can also translate into shared experiences, such as outdoor activities or team challenges that promote participants' self-reflection, stimulating their cognitive and emotional intelligence competencies. An example of a consciousness-raising experience is service learning ([@B78]). In service learning programs, students or course participants take part in community service for learning and development purposes. Answering the question of what it takes "to develop enlightened leaders who can transform their organizations and communities," [@B89]: 163) describe a leadership development program based on service-learning: the ELOC course. The ELOC program, like other service-learning based courses, naturally integrates emotional and social competencies development to sustain leaders in their effectiveness. It is based on the action learning, competency development, teamwork (ACT) philosophy. The program is designed to actively involve participants and guide them in putting their competencies into practice. At the same time, they are stimulated to reflect on their actions and make corrections when necessary, in view of continuous personal improvement. This first stage is inspired by Kolb's experiential learning cycle ([@B64]) and in line with the double-loop learning theory ([@B5]; [@B4]). As for the competency development, the ELOC program provides students with the tools to assess their emotional and social competencies while they are guided to develop a customized and personalized learning plan. Moreover, students are divided into groups according to their learning styles (through the Kolb Learning Style Inventory), and they are challenged to practice their teamworking abilities and provide feedback in order to contribute to each other's development ([@B89]). A similar approach is employed in the health care leadership development program offered by the Cleveland Clinic Academy. The course articulates the development of leadership competencies through three complementary experiences: (i) lectures, in which participants learn the characteristics of an effective and outstanding leader; (ii) mentoring and coaching sessions, in which they receive and ask for feedback; and (iii) opportunities to practice the competencies that participants want to develop ([@B101]). Another experiential learning methodology is the classroom-as-organization approach (CAO), wherein [@B98] tested the positive effects on improving behavioral competencies on a sport management event course. The CAO approach is an experiential learning pedagogy that stimulates students through activities such as simulations, role play, teamwork, and group case studies. Mimicking a real working environment, participants are exposed to the acquisition of conceptual knowledge on the topic of their academic course and are provided with the opportunity to develop their behavioral skills (classified into self-awareness, self-management, social awareness, and relationship management). [@B98] first explore the connection between the CAO approach and the development of ESC competencies, treating the CAO methodology as an opportunity to test whether students' emotional intelligence competencies can be improved "without formally instructing them on emotional intelligence theory in the classroom" ([@B98]: 79). Students are also stimulated in reflective practices (keeping a personal journal, through meetings, and writing a final essay on their experience) in order to evaluate their personal experiences. Scholars cite cooperative learning as an effective methodology for collectively learning new concepts. In a recent work, [@B76] contend that when cooperating, students can also practice their social skills. For this reason, the authors suggest that a cooperative learning methodology can constitute an effective training stage for the development of social competencies in students once the trainees have been informed of the characteristics of 'proper' social behaviors during lectures and seminars ([@B76]). A similar approach is described for the training of trauma teams in hospitals. The skill of teamwork, necessary for managing emergencies, is developed through experiential learning methods, in particular constructing scenarios. First, participants attend lectures to learn the characteristics of effective technical and teamworking behaviors; they are then challenged to put these behaviors into practice during the aforementioned simulations. Finally, feedback is provided during the simulations and at the end of the training experience ([@B109]). Behavioral Competencies as Learning Objectives {#S3.SS3} ---------------------------------------------- The aforementioned programs have been also analyzed in consideration of the behavioral competencies that they aimed at enhancing, regardless of their effectiveness in terms of learning outcomes. For each educational program, [Table 3](#T3){ref-type="table"} reports a short description of the learning objectives stated in the study. ###### Programs' learning outcomes. Educational Program Study ESCs -------------------------------------------------------------------------------- -------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- CALM [@B53] Positive emotions (compassion, acceptance); self-awareness; self-control and emotion regulation; sense of community CARE for Teachers [@B60] Emotional support; self-awareness; self-control; and emotion regulation Classroom-as-organization approach [@B98] Self-awareness competencies, self-management competencies, social awareness competencies and relationship management competencies Consciousness-raising experiences [@B78] Self-awareness; social awareness; dealing with diverse people; promoting inclusiveness; relating to society Cross-cultural management course [@B87] Cultural intelligence (cognitive, metacognitive, motivational, and behavioral components) Ecotonos cultural simulation game [@B24] Cultural intelligence (cognitive, metacognitive, motivational, and behavioral components) Emergency medicine crisis resource management (EMCRM) [@B109] Team coordination skills (anticipation of potential problems, asking for help, attention allocation, communication, delegation, knowledge of the environment, leadership, relational capabilities, teamworking and team leadership, use of information and resources) Emotional Intelligence Training Program [@B45] Adaptability and decision-making (impulse control, problem solving); emotion management; interpersonal skills (empathy, relationship management, social responsibility); mood and self-expression (emotional expression, assertiveness, independence); self-perception (self-awareness, self-esteem, self-realization); stress management (flexibility, stress tolerance, optimism) Executive Leadership and Organization Change (ELOC) program [@B89] Change catalyst; conflict management; leadership; system thinking; service orientation; team management Girls First Resilience Curriculum [@B69] Adaptability; coping capabilities; emotional resilience; persistence; self-efficacy; social skills Happy 12--16 [@B41] Conflict resolution; self-control and emotion regulation Leadership Assessment and Development (LEAD) course [@B23]; [@B88]; [@B26]; [@B82] Achievement orientation; adaptability; coaching and mentoring; conflict management; empathy; goal setting; influence; information analysis; information gathering; initiative; inspirational leadership; organizational awareness; pattern recognition; positive outlook; self-awareness; self-control; system thinking; teamwork MindOut social and emotional learning program [@B36] Relationship management; responsible decision-making; self-awareness; self-management; social awareness PREDEMA [@B97] Expression of emotions; self-awareness; self-control and emotion regulation; social awareness PREDEMA 2.0 [@B96] Assertiveness; conflict resolution; empathy; positive mood; self-awareness; self-control and emotion regulation; self-esteem; social awareness Professional Fellows program [@B9] Attention to detail; communication; developing others; efficiency orientation; empathy; flexibility; group management; initiative; negotiation; networking; pattern recognition; persuasiveness; planning; quantitative analysis; self-confidence; self-control; social objectivity; systems thinking; use of concepts ReSource Project [@B107] Attention; awareness; socio-affective skills (i.e., compassion, prosocial motivation); socio-cognitive skills (i.e., perspective taking, understanding others) Reflective team case teaching model [@B50] Cognitive intelligence skills (critical thinking); emotional intelligence skills (communication, perspective taking, teamworking) The Cleveland Clinic Academy curriculum for health care leadership development [@B101] Leadership competencies Tuning into Teens [@B54] Emotion coaching (conflict resolution, empathy, management of others' emotions); self-awareness; self-control and emotion regulation Whole person learning approach [@B57]; [@B10] Communication; decision-making (negotiation, conflict management, and resolution); leadership; planning; self-awareness; social awareness; teamwork When considering the most targeted competencies, literature reveals a predominant focus on emotional self-awareness ([@B23]; [@B78]; [@B88]; [@B57]; [@B10]; [@B53]; [@B45]; [@B82]; [@B97], [@B96]; [@B36]; [@B60]; [@B107]), namely the ability to understanding one's own emotions and their effects. The LEAD course, the whole person learning approach, and other programs that aim to promote learners' wellbeing, such as the MindOut social and emotional learning program, CARE for teachers, or the CALM course, are structured in ways that allow a deeper reflection on learners' emotions, reactions, desires, and values ([@B23]; [@B88]; [@B53]; [@B82]; [@B36]; [@B60]). The second most developed behavioral competency is self-control ([@B9]; [@B23]; [@B88]; [@B104]; [@B54]; [@B53]; [@B41]; [@B45]; [@B82]; [@B97], [@B96]; [@B36]; [@B60]), followed by social awareness competencies like empathy ([@B9]; [@B23]; [@B109]; [@B78]; [@B88]; [@B57]; [@B10]; [@B54]; [@B45]; [@B82]; [@B97], [@B96]; [@B36]; [@B107]). Being able to manage ourselves even during the most stressful events has important implications for the quality of our social relationships, and it has an impact on our performance. When we lose control of our emotions, we are subjected more to feelings and thoughts that nourish our negative states, impeding a lucid evaluation of situations and negatively affecting our ability to manage and solve problems. Moreover, people who demonstrate the ability to understand other people's emotions deeply and 'step into others' shoes' are more inclined to collaborate with others, be good team workers and leaders and, more generally, they can reach a superior performance thanks to their ability to 'harmonize' with people with which they interact ([@B46]). Taken together, these results suggest that knowledge of our emotions, along with a deep understanding of others' feelings and the ability to act and react in accordance with our own and others' emotional states, is the foundation of most of the development programs considered in this review. Indeed, if we consider the educational programs based on MBIs, they are often devoted to nurturing self-awareness in combination with emotion regulation/stress reduction, empathy, and compassion ([@B107]). Another example where these competencies are jointly developed is offered by the Tuning into Teens program ([@B54]). This is a modified version of the Tuning into Kids program, adapted to address adolescents' needs. Both programs aim at training parents to recognize and respond to their children's emotions in order to promote and support emotional awareness and regulation in youths. To accomplish this goal, the Tuning into Teens and Tuning into Kids programs also promote parents' emotional self-awareness and their ability to manage their emotions. These programs are designed according to the model of emotion coaching proposed by [@B49] that condenses into five suggestions: (1) be aware of your children's emotions, especially when they are of low intensity; (2) recognize the importance of emotions for establishing deep contact with your children; (3) be empathetic with your children and communicate to them that they are understood; (4) support children in communicating, describing, and labeling their feelings; and (5) support and help children in solving problems in relation to their emotions, if needed, but leave room for self-improvement and autonomy ([@B55]). Social and emotional self-awareness represent the premise for successful development of other behavioral competencies ([@B46]), like those most frequently addressed in other studies, such as conflict management ([@B88]; [@B89]; [@B54]; [@B41]; [@B82]; [@B96]) and leadership ([@B23]; [@B88]; [@B57]; [@B89]; [@B101]; [@B10]; [@B82]). The ability to manage difficult situations, encouraging an open dialogue about problems and finding creative solutions to possible frictions, is at the heart of success in many roles and positions. It is fundamental especially for activities that require many interactions, and for people working under pressure ([@B46]). It is also a fundamental skill that outstanding leaders should manifest. Effective leaders are able to be at the forefront of any project/activity with enthusiasm and energy. They are able to motivate their followers, being a model of conduct, but also deeply understanding others' emotions and the reasons behind their behaviors. The Effectiveness of the Competency Development Programs {#S3.SS4} -------------------------------------------------------- This section aims to shed light on the effectiveness of those educational programs that are designed for both the direct and the indirect development of behavioral competencies. We present the results of our analysis as follow: the first section focuses on the methodology adopted by the authors to test the effectiveness of the analyzed educational programs. The second paragraph explores the determinants of the development of emotional and social intelligence competencies, specifically considering both the learners' educational experience and their personal characteristics. Finally, we present the effects of these educational programs on other variables, particularly on participants' entrepreneurial intention, wellbeing, academic performance, and career-related variables. ### Methodology Adopted for Assessing Behavioral Competency Development {#S3.SS4.SSS1} An analysis of the methodologies employed in the empirical studies (48 out of 71 articles) that assess the impact of the aforementioned programs allows us to highlight some peculiarities. Most of the studies directly targeting competency development adopt an experimental design (42%). In these works, participants are randomly divided into two groups, the experimental and the control group. The first receives the treatment, participating in a competency development program, while the second does not attend any course. Competencies are measured in both groups before (at t~1~) and after (at t~2~) the experiment. This procedure is adopted in studies such as those by [@B24], [@B45], [@B36], and other scholars ([@B54]; [@B69]; [@B53]; [@B60]). A quarter of the empirical articles assessing programs specifically targeting ESC competency development are further characterized by a quasi-experimental design. The experimental and quasi-experimental procedures differ in the assignment of participants to experimental and control groups, which is random in the first case. Ultimately, the procedure for assessing the improvement of behavioral competencies in the intervention group after the course is similar to that described when presenting the experimental approach (see, for instance, [@B57]; [@B87]; [@B41]; [@B96]). The work by [@B97], however, is noteworthy. The authors collected measures about the effects of the intervention program at three different points in time: before the experiment, just at the end of the protocol, and 6 months post-test, in order to measure the long-term effect of the program. Ten studies in our dataset adopt a longitudinal design ([@B23]; [@B68]; [@B112]; [@B26]; [@B1]; [@B48]; [@B73]; [@B82]; [@B34]; [@B80]). In these works, authors assess the long-term impact of each development program and discuss its effectiveness in light of a sustainable development. More details are provided in [Table 4](#T4){ref-type="table"} which displays the characteristics of the research conducted in the quantitative studies in our data set. ###### Research design of the quantitative studies on competency development. Study Study design Sample size Target audience Target behavioral competencies Developed behavioral competencies Pre--post-test Long-term follow-up Self/other assessment --------- --------------------------- ------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------ --------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------- [@B1] Longitudinal cohort study *N* = 266 Graduate students of an MBA program at the Weatherhead School of Management (WSOM), Case Western Reserve University (Cleveland, OH, United States). Attention to detail, developing others, efficiency orientation, emotional self-control, empathy, flexibility, group management, negotiating, networking, oral communication, pattern recognition, persuasiveness, planning, quantitative analysis, self-confidence, social objectivity, system thinking, use-of-concepts, use- of-technology, written communication. Attention to detail, developing others, efficiency orientation, emotional self-control, empathy, flexibility, group management, negotiating, networking, oral communication, pattern recognition, persuasiveness, planning, quantitative analysis, self-confidence, social objectivity, system thinking, use-of-concepts, use- of-technology, written communication. Pre--post-test assessment. -- Self-assessment, external raters and third-party evaluation. [@B3] Experiment *N* = 141 Freshmen and tutors from the University of Granada (Spain). Social skills \[ability of saying 'no,' make requests, initiate positive interactions with the opposite sex, defense of own interests, expression of discomfort, self-expression\]. Social skills \[ability of saying 'no,' make requests, initiate positive interactions with the opposite sex, defense of own interests, expression of discomfort, self-expression\]. Pre--post-test assessment. -- Self-assessment. [@B9] Mixed *N* = 53 People enrolled in the Professional Fellows Program. Attention to detail, developing others, efficiency orientation, empathy, flexibility, group management, initiative, negotiating, networking, oral communication, pattern recognition, persuasiveness, planning, quantitative analysis, self-confidence, self-control, social objectivity, systems thinking, theory building, use of concepts, using technology, written communication. Developing others, efficiency orientation, empathy, flexibility, group management, initiative, networking, oral communication, pattern recognition, persuasiveness, planning, self-confidence, self-control, systems thinking, use of concepts. Pre--post-test assessment. -- Self-assessment, external raters and third-party evaluation. [@B15] Correlational study *N* = 95 First year university students enrolled in master's degree at an Italian university. Emotional competencies \[accurate self-assessment, achievement orientation, adaptability, conflict management, conscientiousness, emotional awareness, emotional self-control, initiative, optimism, self-confidence, trustworthiness\], Social competencies \[building bonds, change catalyst, communication, cultural awareness, developing others, empathy, influence, inspirational leadership, organizational awareness, service orientation, teamwork\], Cognitive competencies \[pattern recognition, systems thinking\]. IEL methods are particularly effective for developing self-awareness and self-management. SEL and TL methods are particularly effective for developing Social competencies; IEL and SEL techniques are particularly effective for developing Cognitive competencies. Post-test assessment -- Self-assessment and external raters. [@B16] Causal comparative study *N* = 142 Students from an Italian university who were close to graduate or recently graduated. Emotional competencies \[achievement orientation, adaptability, conscientiousness, emotional self-awareness, emotional self-control, positive outlook\], Social competencies \[change catalyst, coach and mentor, conflict management, empathy, influence, inspirational leadership, organizational awareness, teamwork\], Cognitive competencies \[pattern recognition, systems thinking\]. International Experience positively influence all group of competencies. Cultural Experience positively influences Social and Cognitive competencies. Post-test assessment. -- External raters. [@B23] Longitudinal cohort study *N* = 1241 MBA students at the Weatherhead School of Management, Case Western Reserve University (Cleveland, OH, United States). Relationship-management \[developing others, empathy, group management, helping, leadership, negotiating, networking, oral communication, persuasiveness, relationship, social objectivity\]. Cognitive competencies \[information analysis, information-gathering, pattern recognition, quantitative analysis, sense-making, systems-thinking, theory-building, use-of-concepts, use-of-technology, written communication\]. Self-management competencies \[action, attention-to-detail, efficiency orientation, flexibility, goal setting, initiative, planning, self-confidence, self-control\]. Relationship-management \[developing others, empathy, group management, helping, leadership, negotiating, networking, oral communication, persuasiveness, relationship, social objectivity\]. Cognitive competencies \[information analysis, information-gathering, pattern recognition, quantitative analysis, sense-making, systems-thinking, theory-building, use-of-concepts, use-of-technology, written communication\]. Self-management competencies \[action, attention-to-detail, efficiency orientation, flexibility, goal setting, initiative, planning, self-confidence, self-control\]. Pre--post-test assessment. -- Self-assessment, external raters and third-party evaluation. [@B24] Quasi-experiment *N* = 81. University students studying International Management. Cultural intelligence (behavioral, cognitive, metacognitive, motivational). Cultural intelligence (behavioral, metacognitive, motivational). Pre--post-test assessment. -- Self-assessment. [@B25] Causal comparative study *N* = 180 Undergraduate business students at a Southern U.S. university. Cognitive competencies \[analytical thinking, knowledge application, reflective thinking\], Social competencies \[social awareness, relationship management\]. Students in face-to-face classes develop their Cognitive competencies \[analytical thinking, knowledge application, reflective thinking\] and Social competencies \[social awareness, relationship management\] better than students in web-based classes. Post-test assessment. -- Self-assessment. [@B26] Longitudinal cohort study *N* = 44 Graduates of an MBA program held at the Fondazione CUOA (Italy). Attention to detail, business bargaining, customer orientation, developing others, directing others, efficiency orientation, empathy, flexibility, group management, information gathering, initiative, leadership, negotiating, networking, oral communication, organizational awareness, organizational commitment, pattern recognition, persuasiveness, planning, result orientation, quantitative analysis, self-confidence, self-control, social objectivity, system thinking, theory building, use of concepts, use of technology, visioning, written communication. Directing others, efficiency orientation, networking, written communication. Pre--post-test assessment. -- Self-assessment, external raters and third-party evaluation. [@B33] Cross-sectional study *N* = 407 Physical Education students from five schools in England and one school in Ireland. Life skills \[decision making, emotional skills, goal setting, interpersonal communication, leadership, problem solving, social skills, teamwork, time management\]. Life skills \[decision making, emotional skills, goal setting, interpersonal communication, leadership, problem solving, social skills, teamwork, time management\]. Students completed the survey during physical education lessons or at the beginning of the school day. -- Self-assessment. [@B34] Longitudinal study *N* = 1.400 Canadian children and youths. Emotional Intelligence related competencies \[interpersonal skills, intrapersonal skills, stress management, and adaptability\]. Emotional Intelligence related competencies \[interpersonal skills, intrapersonal skills, stress management, and adaptability\]. The study uses data collected by the National Longitudinal Survey of Children and Youth starting from 2004. Follow-ups were conducted as part of the National Longitudinal Survey in 2006 and 2008. Self-assessment. [@B36] Experiment *N* = 497 Students in schools for youth from disadvantaged communities. Relationship management, responsible decision-making, self-awareness, self-management, social awareness. Self-management. Pre--post-test assessment. -- Self-assessment. [@B39] Correlational study *N* = 194 Undergraduate and MBA students from a Canadian business school. Interpersonal skills, problem solving skills, self-awareness. Simulation are effective for developing interpersonal competencies and self-awareness, followed by case study and lecture, respectively. Simulations and case studies are particularly effective for developing problem solving skills. Post-test assessment. -- Self-assessment. [@B40] Quasi-experiment *N* = 862 University students at the University of Granada (Spain). Empathy, motivation, self-awareness, self-regulation, social skills-leadership. Empathy, motivation, self-awareness, self-regulation, social skills-leadership. Pre--post-test assessment. -- Self-assessment. [@B41] Quasi-experiment *N* = 903 Spanish adolescents enrolled in secondary education programs. Conflict resolution, emotion regulation, self-control. Emotion regulation, self-control. Pre--post-test assessment. -- Self-assessment. [@B45] Experiment *N* = 192 University students. Adaptability and decision-making \[impulse control, problem solving\], emotion management, interpersonal skills \[empathy, relationship management, social responsibility\], mood and self-expression \[emotional expression, assertiveness, independence\], self-perception \[self-awareness, self-esteem, self-realization\], stress management \[flexibility, stress tolerance, optimism\]. Adaptability and decision-making \[impulse control, problem solving\], emotion management, interpersonal skills \[empathy, relationship management, social responsibility\], mood and self-expression \[emotional expression, assertiveness, independence\], self-perception \[self-awareness, self-esteem, self-realization\], stress management \[flexibility, stress tolerance, optimism\]. Pre--post-test assessment. -- Self-assessment. [@B48] Longitudinal study *N* = 1.795 Chinese rural children. Attention skills \[cognitive self-control, concentration, discipline, organization\], Social skills \[sociability/number of friends, leadership at school\]. Social skills \[sociability/number of friends, leadership at school\]. The study uses data from the China Family Panel Studies. Data were collected starting from 2010. This study uses the follow-up conducted in 2012. Self-assessment. [@B53] Experiment *N* = 64 Educators in two middle schools. Positive emotions (compassion, acceptance), self-awareness, self-control and emotion regulation, sense of community. Self-control and emotion regulation. Pre--post-test assessment. -- Self-assessment. [@B54] Experiment *N* = 449 Parents and their adolescent offspring. Emotion coaching (conflict resolution, empathy, management of others' emotions), self-awareness, self-control and emotion regulation. Emotion coaching (conflict resolution, empathy, management of others' emotions), self-awareness, self-control and emotion regulation. Pre--post-test assessment. -- Self-assessment and external raters. [@B57] Quasi-experiment *N* = 483 MBA students who completed a course in executive skills. Communication, decision-making \[negotiation, conflict management, conflict resolution\], leadership, planning, self-awareness, social awareness, teamwork. Communication, decision-making \[negotiation, conflict management, conflict resolution\], leadership, planning, self-awareness, social awareness, Pre--post-test assessment. -- Third-party evaluation. [@B60] Experiment *N* = 224 Elementary school teachers in high poverty areas of New York. Emotional support, self-awareness, self-control and emotion regulation. Emotional support, self-awareness, self-control and emotion regulation. Pre--post-test assessment. -- Self-assessment and external raters. [@B68] Longitudinal study *N* = 100 Graduates of an MBA program. Attention to detail, developing others, efficiency orientation, emotional self-control, empathy, flexibility, group management, initiative, negotiating, networking, pattern recognition, persuasiveness, planning, self-confidence, social objectivity, system thinking. Attention to detail, developing others, efficiency orientation, emotional self-control, empathy, flexibility, group management, initiative, negotiating, networking, pattern recognition, persuasiveness, planning, self-confidence, social objectivity, system thinking. Pre--post-test assessment. -- Self-assessment, and third-party evaluation. [@B69] Experiment *N* = 2.387 Rural adolescent girls in Bihar (India). Adaptability, coping capabilities, emotional resilience, persistence, self-efficacy, social skills. Social skills. Pre--post-test assessment. -- Self-assessment. [@B73] Longitudinal study Two studies: the Minnesota Twin Family Study: *N* = 1.382; the Sibling Interaction and Behavior Study: *N* = 617. The Minnesota Twin Family Study (MTFS): pairs of twins and their parents; The Sibling Interaction and Behavior Study (SIBS): adoptive and nonadoptive families with two offspring and their rearing parents. Non-Cognitive skills \[academic effort, academic problems, aggression, alienation, control, externalizing, harm avoidance, stress reaction\]. -- MTFS: twins are assessed at 11 or 17 years old; SIBS: adolescents were assessed at the mean age of 15 years old. MTFS: follow-ups were conducted at the age of 14, 17, 20, 24, 29; SIBS: follow-ups were conducted after 3, 7 years. Self-assessment. [@B76] Quasi-experiment *N* = 346 University students from the Degrees in Infant Education and Primary Education at the University of Extremadura (Spain). Social skills: (1) self-assertion \[asking for changes in behavior, taking criticism on board, stopping interaction\], (2) receiving information \[actively listening, empathizing, summarizing, asking for help, asking questions\], (3) imparting information \[motivating, imparting information, convincing others, explaining oneself, giving help\]. Imparting information \[motivating, imparting information, convincing the others, explaining oneself or helping\], self-assertion \[asking for changes in behavior, receiving criticism). Pre--post-test assessment. -- Self-assessment. [@B80] Longitudinal study *N* = 9.291 U.S. individuals from adolescence to adulthood. Non-cognitive skills \[risk aversion, future-orientation, and conscientiousness\]. Non-cognitive skills \[risk aversion, future-orientation, and conscientiousness\]. The study uses data collected by the National Longitudinal Study of Adolescent to Adult Health during 1994. Follow-ups were conducted in 1995, 2001, 2008. Self-assessment. [@B82] Longitudinal cohort study *N* = 531 Full-time graduate MBA students. Achievement orientation, adaptability, coach and mentor, conflict management, emotional self-awareness, emotional self-control, empathy, influence, inspirational leadership, organizational awareness, pattern recognition, positive outlook, teamwork, system thinking. Achievement orientation, adaptability, coach and mentor, emotional self-awareness, empathy, influence, inspirational leadership, pattern recognition, system thinking. Pre--post-test assessment. -- Self-assessment and external raters. [@B87] Quasi-experiment *N* = 281 U.S graduate students in business subjects. Cultural intelligence (cognitive, metacognitive, motivational, and behavioral). Cultural intelligence (cognitive, metacognitive, motivational, and behavioral). Pre--post-test assessment. -- Self-assessment. [@B88] Mixed *N* = 25 MBA students who attended the LEAD course in an MBA program at the Weatherhead School of Management, Case Western Reserve University (Cleveland, OH, United States). Action, goal setting, help, information analysis, information gathering, initiative, leadership, quantitative analysis, relationship building, sense making, theory building, use of technology. Action, goal setting, help, information analysis, information gathering, initiative, leadership, quantitative analysis, sense making, theory building, use of technology. Pre--post-test assessment. -- Self-assessment and third-party evaluation. [@B90] Correlational study Study 1: *N* = 46; Study 2: *N* = 118 Study 1: Adolescents in low-income status; Study 2: Latino students with poor academic performance, poor classroom conduct and low parent involvement in school. Study 1: self-worth; Study 2: concentration and emotional regulation skills. Study 1: self-worth; Study 2: concentration and emotional regulation skills, only for individuals with preexisting concentration and emotional regulation problems. Pre--post-test assessment. -- Self-assessment and external raters. [@B95] Experiment *N* = 190 Adolescents enrolled at two summer schools in Andalusia (Spain). Emotional intelligence \[self-control, emotionality, sociability\]. Emotional intelligence \[sociability\]. Pre--post-test assessment. -- Self-assessment. [@B97] Quasi-experiment *N* = 148 Adolescents enrolled in first and second years of secondary school. Expression of emotions, self-awareness, self-control and emotion regulation, social awareness. Expression of emotions, self-awareness, self-control and emotion regulation, social awareness, not immediately after the intervention but after the 6 months follow-up. Pre--post-test assessment. A follow-up was conducted 6 months after the end of the intervention. Self-assessment. [@B96] Quasi-experiment *N* = 250 Undergraduate students enrolled at a Spanish university. Assertiveness, conflict resolution, empathy, positive mood, self-awareness, self-control and emotion regulation, self-esteem, social awareness. -- Pre--post-test assessment. A follow-up was conducted 3 months after the end of the intervention. Self-assessment. [@B98] Quasi-experiment *N* = 49 Undergraduate and graduate students. Emotional and Social competencies \[achievement orientation, accurate self-assessment, adaptability, building bonds, change catalyst, communication, conflict management, cultural awareness, developing others, emotional self-awareness, emotional self-control, empathy, influence, initiative, inspirational leadership, optimism, organizational awareness, self-confidence, service orientation, teamwork and collaboration, transparency, trustworthiness\]. Emotional and Social competencies: \[accurate self-assessment, achievement orientation, trustworthiness, adaptability, optimism, service orientation, organizational awareness, cultural awareness, communication, building bonds, and teamwork and collaboration\]. Post-test assessment. No follow-ups. External raters and third-party evaluation. [@B105] Cross-comparative study *N* = 113 MBA students, psychology students, nonstudents \[control groups\]. Emotional Intelligence. Emotional Intelligence only for MBA students. Pre--post-test assessment. -- Self-assessment. [@B107] Experiment *N* = 332 Healthy adults. Attention, awareness, socio-affective skills (compassion, prosocial motivation), socio-cognitive skills (perspective taking, understanding others). Attention, compassion. Pre--post-test assessment. -- Third-party evaluation. [@B108] Correlational study *N* = 117 Business Studies students from secondary schools and one university in Dundee (Scotland, United Kingdom). Entrepreneurial skills \[collaboration, communication, confidence, creativity and innovation, decision-making, leadership, organization and management, problem-solving, resolving conflict, reflection\]. Entrepreneurial skills \[communication, confidence, creativity and innovation, organization and management, reflection\]. Post-test assessment. -- Self-assessment. [@B109] Experiment *N* = 15 Medical students at the beginning of their clinical career. Team coordination skills \[anticipation of potential problems, asking for help, attention allocation, communication, delegation, knowledge of the environment, leadership, relational capabilities, teamworking and team leadership, use of information and resources\]. Team coordination skills \[anticipation of potential problems, asking for help, attention allocation, communication, delegation, knowledge of the environment, leadership, relational capabilities, teamworking, use of information and resources\]. Pre--post-test assessment. -- Self-assessment and third-party evaluation. [@B112] Longitudinal study *N* = 30 Part-time MBA students. Emotional competencies \[attention to detail, efficiency orientation, flexibility, initiative, planning, self-control\], Social competencies \[developing others, empathy, group management, leadership, negotiating, networking, persuasiveness, self-confidence\], Cognitive competencies \[pattern thinking, social objectivity, systems thinking\]. Emotional competencies \[efficiency orientation, flexibility, initiative, self-control\], Social competencies \[leadership, self-confidence\], Cognitive competencies \[social objectivity, systems thinking\]. Pre--post-test assessment. A follow-up was conducted between 18 and 30 months after the graduation. Self-assessment and third-party evaluation. This comparative analysis report information about the research design, the sample, the assessment procedure (pre-posttest assessment, follow-ups, actors involved), along with the targeted and developed behavioral competencies. As for this last aspect, a number of studies reveals their effectiveness for the development of a wide range of skills, such as the emotional, social and cognitive intelligence competencies (see for instance [@B9]; [@B88]; [@B98]; [@B15], [@B16]; [@B45]; [@B82]), while other research narrows the focus to more specific clusters related, for instance, to individuals' social sphere and relationship management (e.g., [@B3]; [@B76]). Some differences in the methods adopted for measuring the effects of the learning experience on behavioral competencies were found in those programs not explicitly designed for developing ESCs. The majority (53%) of articles in this last group lean toward a quantitative design, collecting cross-sectional data. This is the case, for instance, of [@B25] article, in which the authors rely on a survey design to compare the unintended outcomes of web-based and face-to-face classes. Questionnaires were also administrated by [@B39] to compare the effects of three teaching procedures, and by [@B15], who assess the joint impact of different learning methods on students' ESC competencies. Moreover, in this last group of works, only two articles ([@B3]; [@B76]) adopt, respectively, an experimental and a quasi-experimental design, further highlighting the differences between the two aforementioned groups of programs. ### Factors Influencing the Development of Behavioral Competencies {#S3.SS4.SSS2} Along with the courses that proved their effectiveness for developing ESCs, literature has identified other intervening factors that may enhance or hinder the effectiveness of the learning process. We categorized these variables into two broad categories: the one referring to characteristics of the learning experience, while the other to learners' personal characteristics. For what concerns the former category, literature has acknowledged the importance of complementing traditional learning methods, such as lectures, with other, more holistic and experience-based, approaches. In this regard, [@B15] showed that emotional and social competencies can be effectively developed in students only when individual and social experiential learning methods (i.e., simulations, individual/cooperative assignments, business games) are integrated into the more traditional educational activities. Similar results are provided by [@B39] who tested the effectiveness of three different teaching methods aimed at enhancing learners' self-awareness, interpersonal and problem-solving skills. The authors confirm that simulations are perceived as the most effective teaching method for developing relational capabilities and self-awareness. Case study approach is perceived as effective as simulations for stimulating problem solving skills, while it is the second-best method for developing interpersonal competencies and self-awareness. More interestingly, students classify lectures as the less effective approach for their behavioral competency development ([@B39]). To further corroborate these results and moving from the in-class environment, [@B16] suggest that living an international experience and taking part to cultural activities favor the development of behavioral competencies. When abroad, students get in touch with cultures, uses and norms that detach them from the comforting familiarity of their habits. They experience aloneness and need to construct new relationships to cope with it and to integrate in their new social environment. Similarly, when taking part to cultural activities (e.g., political parties, artistic groups, volunteerism) individuals interact and cooperate with others, often facing situations that ask them to reflect on, and sometimes revise, their values and ethical beliefs. These experiences proved to be of stimulus for the development of students' emotional, social and cognitive competencies. As for the category grouping learners' personal features, the work by [@B33] confirms that students' basic needs satisfaction acts as a mediator between the teaching style of their coach and the development of their life skills (i.e., interpersonal communication, emotional capabilities, leadership skills, problem solving competencies). The authors suggest that a teaching style oriented to encouraging students' proactivity and initiative and that promotes their problem-solving and interpersonal skills has the effect of fulfilling learners' need for autonomy and relatedness while enhancing their behavioral competency development. In studying the process of change that underlies the development of behavioral competencies, [@B88] discovered the key role played by students' career aspirations, opportunity mode and desire for continuous learning. The author suggests that individuals who possess a higher propensity to continuous learning, who have planned a clear career path and who try to exploit their learning opportunity in order to maximize their learning outcomes obtain better results in term of competency development compared to students who do not show these characteristics. Finally, in a work reaffirming the importance of having a clear and structured learning plan that guide the acquisition of behavioral competencies, [@B68] provides support to the hypothesis that individuals who have a clear image of who they want to become and of the competencies they need to improve are more successful in their process of change than those who do not establish clear learning goals. ### The Effectiveness of Competency Development Programs on Other Variables {#S3.SS4.SSS3} Among the reviewed programs, we identified a group of studies that not only address the enhancement of behavioral skills *per se*, but analyze the effect of competency development on other variables. [Table 5](#T5){ref-type="table"} reports the description of the dependent variables that have been affected by the behavioral competencies developed by the educational programs. ###### The impact of competency development on other variables. Study Dependent variables -------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ [@B1] Life satisfaction, career satisfaction, and career success: self-informant based assessment of validated measures. [@B16] Entrepreneurial orientation measured as the expression of a conscious plan to set up a business. [@B26] Career advancement. Two measures: (i) position held by each subject (classified according to an adapted version of the Watson Wyatt Global Grading System and weighted considering the information related to the subject's position in the company hierarchy and the number of collaborators that he/she manages); (ii) the characteristics of the company which employed the subject at the moments under examination (in terms of turnover, number of employees, organizational complexity). Career compensation: total annual wage (fixed and variable) of each subject. [@B40] Entrepreneurial intention: self-informant based assessment of a validated measure. [@B53] Positive affect, distress tolerance, teaching efficacy, professional, burnout, physical symptoms: self-informant based assessment of validated measures. Blood pressure readings. Diurnal assays of cortisol. [@B60] Psychological distress and physical distress: self-report and report on student assessments of validated measures. Emotional support: classroom observation on teacher--student interactions. [@B69] Positive psychological wellbeing: self-informant based assessment with the KIDSCREEN-Psychological Wellbeing Subscale. Social wellbeing: with the KIDSCREEN-Social Support and Peers Subscale [@B73] College outcome: measured as having completed or currently attending college at the second follow-up. [@B97] Cyberbullying, life satisfaction: self-informant based assessment of a validated measures of before (T1), immediately after (T2), and 6 months after the intervention (T3). [@B96] Subjective well-being: self-informant assessment of a validated scale of life satisfaction. [@B16] and [@B40] test and confirm the effectiveness of ESC competency development on students' entrepreneurial intention. Entrepreneurial intention is defined as a planned desire to start a new business in the future ([@B16]). Persons who possess or develop their entrepreneurial intention are more likely to show entrepreneurial behaviors that concretize into the creation of new businesses ([@B40]). In particular, it has been discovered that students who receive entrepreneurial training and develop their emotional competencies during these courses show higher entrepreneurial intention, since they develop their entrepreneurial attitude and self-efficacy. Emotional competencies in fact increase students' self-identification as future entrepreneurs and strengthen confidence in their ability as future business owners ([@B40]). Moreover, learning how better to understand and manage themselves and others' emotions has a direct impact on learners' entrepreneurial intention ([@B16]). Some studies have considered the impact of competency development on improving people's wellbeing. This is the case of the PREDEMA and PREDEMA 2.0 programs ([@B97], [@B96]). The authors assert that intervening in students' behavioral competencies helps them cope with the pressures associated with academic life, especially during the transition from high school to the university ([@B96]), while it also shows positive effects in improving the coexistence between adolescents in classroom environments, limiting the incidence of cyberbullying behaviors ([@B97]). Similar findings are reported in a study on the efficacy of the Girls First Resilience Curriculum, designed and implemented for girls in high-poverty conditions ([@B69]). The program reveals its effectiveness in enhancing girls' psychological and social wellbeing, meaning that people enrolled in this curriculum show, at the end of the course, higher levels of life satisfaction and greater capacity to build relationships with peers. These results are obtained through the improvement of behavioral competencies such as self-efficacy, tolerance of negative situations, empathy, and teamworking ([@B69]). As mentioned earlier, teachers are responsible not only for the quality of education they provide to their students, but also for students' development as human beings. Promoting teachers' wellbeing and their capacity to manage stressful situations is at the heart of CARE for teachers and CALM programs. These MBIs reveal their effectiveness in improving educators' ESC competencies and in promoting participants' wellbeing and distress, with positive results also in terms of educational outcomes ([@B53]; [@B60]). Some studies have tested the impact of the development of behavioral competencies on learners' educational attainments and career-related variables such as career satisfaction and career success. According to [@B73], cognitive and non-cognitive competencies, especially self-control, emotion control, and motivation, are independently associated with students' academic achievements. Thus, students who are able to manage their negative emotions even in the toughest periods are more inclined to achieve challenging objectives in the long run instead of pursuing immediate reward. Another explanation comes from a literature review conducted by [@B111], in which the authors contend that MBIs increase learners' attention, memory, and emotion regulation. Meditation is able to change learners' brains and to foster learners' cognitive functions and self-control, supporting their motivation and the creation of a personalized learning plan, thus leading them toward successful academic outcomes, greater wellbeing, and better social relationships ([@B111]). As for career-related variables, the literature suggests that the development of emotional intelligence competencies positively impacts on career satisfaction and career success, used as umbrella terms for capturing one's own idea of satisfaction with one's position, income level, advancement chances, and development opportunities inside the organization ([@B1]). More specifically, a study by [@B26] reveals that students who participated in an MBA program and who reported higher improvements (measured by self-, peer, and third party evaluations) in their ESC competencies during the course, have better career paths (measured by a composite index considering the subjects' position, the number of collaborators, and the characteristics of the organization) at least 1 year after the end of the MBA program. Literature Gaps and Directions for Future Research {#S4} ================================================== There is little doubt that the debate on how emotional, social, and cognitive competencies should be learnt has flourished during the last decade, with contributions in the literature that document the educational programs designed to nurture different behaviors in diverse learning contexts. The results of our analysis also allow the identification of major gaps in the current debate and promising areas for future research. Categories of Learner: Who Is Missing? {#S4.SS1} -------------------------------------- The review reveals that competency development programs have been designed and delivered in different settings, from high school to higher education and organizational contexts. The studies claim a positive influence of ESC programs in supporting individuals toward transition stages in their educational and career paths. However, where the higher education setting is concerned, the ESC learning experience offered in Bachelor, graduate, or Master's programs are often attended only by students enrolled in management curricula. Research has shown that behavioral competencies are crucial in every job and every sector; therefore, their development does not pertain only to professional roles operating in the management field. Consequently, further studies should consider the design of ESC courses in different disciplinary areas, focusing on those competencies most required in the labor market and distinctive for the specific professional roles prepared by the academic curricula. Moreover, among the educational initiatives described in the selected literature, the programs targeting adults primarily focus on the health care sector. Notwithstanding the relevance of nurturing emotional and social intelligence competencies in this specific organizational context, the studies analyzed do not consider the training initiatives implemented by many organizations in both the public and private sectors, aiming at providing executives with the ESCs to exercise their leadership successfully ([@B63]). Considering the huge amount of effort and resources invested by companies in various forms of training and development program to nurture emotional intelligence in their employees ([@B30]), scholars may further investigate how these initiatives vary in terms of objectives, content, methodologies, and behavioral change according to the different professional roles to whom the courses are addressed and the organizational contexts in which they are delivered. An interesting area of enquiry, that is also indicated by some scholars as one of the critique to the competency model, is represented by the analysis of how competency development programs should be tailored according to the specificity of the firm and job requirements as well as the industrial setting, since organizational context and group relationships assume a relevant role in influencing the learning process ([@B2]) and in supporting and encouraging individuals to undertake the learnt behaviors ([@B14]). Methodological Insights for Behavioral Competency Development {#S4.SS2} ------------------------------------------------------------- The review we conducted reveals that emotional and social competencies development should promote the adoption of a holistic approach to education that requires educators to rethink and innovate their own teaching methods and styles, shifting from a passive lecture approach to one that actively engages learners in activities designed for practice and experience. Indeed, in line with the experiential and whole person learning approach, a variety of teaching techniques have been proven to be effective in emotional and social education programs. These encompass self-reflection activities and meditation practice, but also interactive techniques such as coaching, role play, simulations, or the adoption of gamification. Despite the contributions provided by these studies, further investigation is necessary to provide an in-depth analysis and systematization of the different learning techniques that can be adopted to develop ESCs. Future studies should identify and describe the conditions for their effective use and their association with specific competencies. For instance, if the techniques adopted during mindfulness-based interventions can be successfully proposed to enhance emotional self-awareness, self-control, and empathy, they may not be appropriate for other types of competency. Furthermore, even though the literature seems to converge on experiential learning theory ([@B66]) as the pedagogical framework for designing competency development programs, studies do not address the issue of how each learning style adopted by individuals to practice the new behaviors calls for different techniques. Another limitation we found in the analysis of the literature is related to the fact that many educational programs explicitly adopting the whole person learning approach limit the learning experience to just a set of competencies defined at institutional level. As suggested by other contributions, the effectiveness of a learning program is based on the intention of the individuals to undertake the development of specific competencies in line with their desired career goal ([@B29]; [@B22]; [@B21]). Specifically, this different approach has been described in ICT, which advances self-directed learning by providing a theoretical framework for describing the essential components of desirable, sustainable change in one's behavior, thoughts, feelings, and perceptions ([@B19]). The facts that the change is desired and that it is sustainable represent two fundamental aspects of effective application of the ICT framework to emotional and social competencies. The change is desired by the person who wishes it, since learning goals are set intentionally by individuals to be coherent with their future personal and professional lives. The fact that, during the training experience, individuals intentionally identify the competencies they wish to learn represents an essential motivational driver to change their behavior successfully. Moreover, in contrast to the majority of training experiences, intentional change is also sustainable, in that it endures for a relatively long time. In line with this methodology, scholars should devote more attention to the development of didactic techniques which enable learners to become more aware of their overall portfolio of competencies at the beginning of the program and support them in identifying the ESCs needed to attain their desired future. Another feature of the analyzed training interventions is that the learning programs engage students with experiential techniques for promoting self-reflection and active learning in the classroom environment. Only a few initiatives provide the participants with a structured plan to practice the competencies in real life contexts in order to pursue a long-lasting change in their behavior. Adopting a more experimental approach to competency development offers individual more opportunity for self-discovery and the exploration of effective behavior in life and the organizational settings ([@B14]), and consequently it may favor a more sustainable change. The intentional change approach seems to respond to the critique moved by [@B2] to competency development models. Indeed, the authors call for a more holistic, self-directed and experimental approach to competency development in line with the literature on adult education. Moreover, they pointed out that the competency learning initiatives mainly adopt a short-term logic in order to satisfy the current organizational needs. Instead, a future-oriented perspective is more in line with the concept of organization as dynamic and learning entities. In order to support organizations in their continuous improvement, educational institutions may play a pivotal role in identifying and developing those emotional and social intelligence competencies that will be central in the future jobs and are necessary to enable individuals to learn and adapt during periods of change. Another limitation highlighted by the literature review is related to the types of learning objective proposed by these programs. As emerged from our analysis, social and emotional learning initiatives have primarily been designed for the development of a few types of emotional and social competency, specifically self-awareness, empathy, and some relationship management competencies such as teamwork, leadership, and conflict management. Apart some studies ([@B9]; [@B23]; [@B68]; [@B88]; [@B112]; [@B26]; [@B15], [@B16]; [@B25]; [@B45]; [@B82]; [@B33]; [@B36]; [@B40]; [@B80]) that demonstrated a positive impact of the learning programs on a broad set of emotional, social and cognitive competencies, more efforts need to be devoted to the cognitive competencies and the self-management cluster, with the exception of the competencies of emotion regulation/self control frequently considered in the articles analyze. Future studies should advance the debate on the development of these skills that are assuming increasing relevance in the labor market, such as resilience, adaptability, problem-solving, and creative thinking, as highlighted by recent employer surveys ([@B70]; [@B85]). Finally, there is also a need for more robust studies which assess the effectiveness of ESC programs delivered in different cultural environments, since the majority of the studies have been conducted in the European and US contexts. Learning Outcomes and Assessment Methods {#S4.SS3} ---------------------------------------- A limitation that almost all the studies analyzed have in common is that the outcomes in terms of individual behavioral change were measured immediately after the intervention had ended. Prior research suggests that the effects of many interventions diminish over time---the so-called honeymoon effect---and therefore future studies should devote more effort to investigating if the effects will endure post-intervention. Follow-up studies conducted several months after the end of the intervention could investigate the longer-term effects of the program on learners' outcomes, as indicated in the SPR's Standards Committee ([@B42]). The review shows that even though the majority of the empirical works relies on the behavioral approach or third emotional intelligence stream in defining the competencies considered as learning outcomes, the same competency in different studies has been measured adopting diverse psychometric measures. This may limit the comparability of the learning experiences examined in the studies analyzed. This review was not intended to add to the debate on the different approaches to competency measurement, however, scholars should consider this issue as critical in the future research if they aim to compare how different learning methodologies affect the development of a specific competency. Scholars should motivate the decision to rely on a specific psychometric scale for measuring the behavioral competency that is the target of the training programs, reporting satisfying criteria for discriminant and convergent validity. Related to the issue of competency measurement, another critical issue that should be considered in designing competency development studies concerns the data collection technique. Indeed, self-report assessment in the pre- and post-intervention stages could be biased by the honeymoon effect or by a higher awareness in that specific competency matured during the learning program that allows the participant to assess the competency more objectively. As proposed by [@B20] in the Stream 4 measures of emotional intelligence, the person's behavior should be assessed by external raters who have the opportunity to observe the activation of the competencies in the work and life contexts pre and post the learning experience and during a long-term follow up session. Moreover, future research may measure the behaviors that the programs aim to develop relying not only on both self- and others' evaluation through 360 assessments but also on direct observations, limiting the potential bias and the competing dynamics often ascribed to 360-degree appraisal tools used both for assessment and development purposes ([@B14]). Only a few studies included in this review devote attention to other variables influenced by ESC development, focusing on entrepreneurial intention, individual wellbeing, career success, and satisfaction, and concentrating the attention primarily on students and educators. We suggest further consideration of other outcomes of the learning programs. For instance, an interesting line of enquiry is represented by the impact of the ESC learning programs on generating differences in individual performance in workplace settings, but also on employability and perceived employability, namely "the perceived ability to attain sustainable employment appropriate to one's qualification level" ([@B93]: 2). Since employability depends on the optimal use of personal competences ([@B102]), and only anecdotal evidence has been provided on this relationship, an interesting area of enquiry is represented by investigating the effect of competency development on increasing individuals' employability. Conclusion {#S5} ========== Much has been written about the relevance of emotional and social competencies across professional roles and industrial settings, and our article systematically reviews the literature which has contributed so far to the debate on how behavioral competencies can be developed in the higher education and organizational environments. The contribution of our work is, on the one hand, to document the diversities and the commonalities among the theoretical and methodological frameworks adopted in several educational programs and, on the other hand, to identify major opportunities for their further advancement. Education institutions and organizations should formally recognize the critical role of ESCs in contributing to individuals' employability, job performance, and career success, investing appropriately in their development. This review represents the first attempt to systematize the methodologies of the educational programs for competency development and to assess their effectiveness in order to assist educators and executives in their ongoing efforts to equip students and employees with the relevant skills needed to achieve superior performance in the workplace. At the institutional level, policymakers should promote a dedicated agenda with concrete actions to equip people with emotional and social intelligence competencies---for instance, sponsoring and funding career pathway projects. Finally, the modern concept of the employment psychological contract, in which careers are conceived as more self-directed, value-driven, and boundaryless ([@B6]; [@B51]; [@B52]), spurs individuals to be proactive in undertaking lifelong learning through the identification of training opportunities which enable them to broaden their repertoire of emotional and social behaviors in line with their desired career path. Author Contributions {#S6} ==================== SB, FG, RZ, and RB contributed to the design of the study, organized and analyzed the database. SB, FG, RZ, and RB wrote the different sections of the manuscript. All authors contributed to manuscript revision, read and approved the submitted version. Conflict of Interest {#conf1} ==================== The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. This research has been conducted at the Ca' Foscari Competency Centre of the Ca' Foscari University of Venice. We thank all the colleagues of the Ca' Foscari Competency Centre with whom we had the opportunity to share our research. We express our gratitude to the reviewers and the editor, whose support was fundamental to direct the manuscript. [^1]: Edited by: Jon Gruda, Maynooth University, Ireland [^2]: Reviewed by: Jim Allen McCleskey, Western Governors University, United States; Raul Antonio Berrios, Universidad de Santiago de Chile, Chile [^3]: This article was submitted to Organizational Psychology, a section of the journal Frontiers in Psychology

\section{Introduction} \label{intro} In the last years we have witnessed a vast increase in the amount of data that is being stored and processed by organizations of all types. As stated by Xindong et al.~\cite{XindongWu2014}, the so-called big data revolution has come to us not only with many challenges but also with plenty of opportunities that these organizations are willing to embrace. According to Rajamaran and Ulman~\cite{rajaraman2014mining}, the main challenge is to extract useful information or knowledge from these huge volumes of data that enable us to predict or to better understand the phenomena involved in its generation. These tasks are commonly tackled using data mining. As part of the data mining process, feature selection or feature subset selection is a crucial preprocessing step for identifying the most relevant attributes from a dataset. Removing irrelevant and redundant attributes not only can generate less complex and more accurate models but also a reduced dataset allows us to enhance the performance of many data mining schemes. Feature selection algorithms are usually classified into three types: \emph{wrappers}, \emph{filters} and \emph{embedded methods}. Wrappers refer to methods that require the learning of a classifier based on a single or on a subset of the original features, they are usually more computationally expensive. Filter methods rely only on the characteristics of data and are independent of any learning scheme, thereby requiring less computational effort. Finally, embedded methods refer to those techniques where the selection of features is carried out as part of the classification process. Another important feature selection classification, as stated by Garc\'ia et al.~\cite{García2015}, is the one that comes from viewing the feature selection as a search problem with the aim of finding a high quality feature subset, and thereby, they classify the methods according to the type of search performed: \emph{exhaustive}, \emph{heuristic} and \emph{stochastic}. However, as not all feature selection techniques can be viewed as these types of search, a fourth category needed to be included, \emph{feature weighting}. In this context, ReliefF~\cite{Kononenko1994} is a widely applied feature weighting technique that estimates the quality of the features from a given dataset by assigning weights to each of them. It can be used as filter feature selection method by defining a significance threshold and selecting features with quality above it. As a result of its advantages such as being able to work with nominal or continuous features, handling multi-class problems, detecting features interactions, handling missing data and noisy tolerance, it is considered one of the most convenient filter-based feature selection methods available \cite{Bolón-Canedo2012}. On the other hand, after its 2004 seminal paper, Google's MapReduce~\cite{Dean2004a,Dean2008} emerged as a programing model that simplified the development of scalable applications that process and generate large scale datasets. These applications are defined in terms of map and reduce tasks that are automatically parallelized by a programming framework and executed in clusters of tens or even thousands of machines, handling failures and scheduling resources. One important fact is that MapReduce was designed to run on commodity hardware, because it leads to lower costs per processor and per unit of memory of the cluster. The standard MapReduce implementation to date is Hadoop MapReduce~\cite{hadoop}, an open source implementation mainly developed at Yahoo Labs. However, the framework uses disk writing between every MapReduce job with the objective of recovering from failures, and this becomes a bottleneck for iterative nature algorithms like the ones used in machine learning and data mining, including the original ReliefF algorithm. For that reason, Spark~\cite{Zaharia2010} has gained much attention in the last couple of years, since it presents an improved model that is capable of handling most of its operations in-memory while maintaining the fault tolerance and scalability of MapReduce. In this paper we present DiReliefF\footnote{\url{https://github.com/rauljosepalma/DiReliefF}}, a distributed and scalable redesign of the original ReliefF algorithm based on the Spark computing model, enabling it to deal with much larger datasets in terms of both instances and features than the traditional version would be able to handle. In addition, after presenting and discussing the design of the algorithm we compare several runs of it with the traditional version implemented on the WEKA \cite{Hall2009} platform, using 4 open datasets with numbers of instances in the order of $10^7$ and a number of features that ranges in the order of $10^1$ until $10^3$. The main conclusion obtained is that is practically unfeasible to deal with such high amounts of data using the traditional version on standard hardware, while the distributed version is able to handle them smoothly. \section{Related Work} \label{sec:relatedWork} ReliefF was first described by Kononenko in 1994 \cite{Kononenko1994} and since then, many applications, extensions and improvements have been published. Moreover, ReliefF is itself an extension of the original Relief algorithm developed by Kira and Rendell \cite{Kira1992}, the latter was initially limited to binary class problems while the former can handle multi-class problems. As recent examples of those extensions, Reyes et al.~\cite{Reyes2015} presented three of them for multi-label problems and compared them with previous extensions for the same purpose. Zafra et al. \cite{Zafra2012} extended ReliefF to the problem of multi-instance learning. Greene et al. \cite{Greene2009} proposed an adaptation to enhance ReliefF's ability to detect feature interactions called SURF (Spatially Uniform ReliefF) and applied it in the genetic analysis of human diseases. One of ReliefF's mayor flaws is its incapacity to detect redundant features, and hereby some attempts have been made to overcome this flaw. For example, Li and He~\cite{Li2016} used a forward selection algorithm to select non redundant critical quality characteristics of complex industrial products. Zhang et al.~\cite{Zhang2008} combined ReliefF with the mRMR (minimal-redundancy-maximal-relevancy) algorithm \cite{Peng2005} to select non redundant gene subsets that best discriminate biological samples of different types. As it might be expected, most feature selection algorithms have asymptotic complexities that depend on the number of features or instances, and in our case, ReliefF depends linearly on both of them~\cite{Robnik2003}. Thereby, its performance gets compromised when faced with datasets with high dimensionality and/or high number of instances. For this reason, many attempts have been made to make feature selection methods, including ReliefF more scalable. Recently, Canedo et al.~\cite{Bolon-Canedo2015} proposed a framework to deal with high dimensional data, distributing the dataset by features, processing in parallel the pieces, and then performing a merging procedure to obtain a single selection of features, however this method is oriented for high dimensional datasets and no tests were made with datasets with high amounts of instances. A some way similar approach was followed by Peralta et al.~\cite{Peralta2015}, who used the MapReduce model to implement a wrapper-based evolutionary search feature selection method. In this case, the data is split by instances, and an evolutionary feature selection is performed over each of these pieces. Furthermore, the reduce step basically uses a simple majority voting of the selected features with a user-defined threshold to select the definitive subset of features. All the tests were carried out with the EPSILON dataset, that we also use here (see Section~\ref{sec:experiments}). Zhao et al.~\cite{Zhao2012} presented a distributed parallel feature selection method based on variance preservation using the SAS High-Performance Analytics \footnote{\url{http://www.sas.com/en_us/software/high-performance-analytics.html}} proprietary software. Experimental work was carried out with both high dimensional and high number of instances datasets. Kubica et al.~\cite{Kubica2011} developed parallel versions of three forward search based feature selection algorithms, a wrapper with a logistic regression classifier is used to guide the search that is parallelized using the MapReduce model. It is also worth mentioning the recent work of Wang et al.~\cite{Wang2016} that uses the Spark computing model to implement feature selection strategy for classification of network traffic. Their approach basically consists in two phases, first it generates a feature ranking based on the Fisher score, and then a forward search is done using a wrapper approach. Nevertheless, this method can only be applied to continuous data. As it can be seen, none of the above contributions implements a purely filter approach, even in the case of the framework proposed by Bol\'on-Canedo et al.~\cite{Bolon-Canedo2015} that allows us to apply filters to parts of the dataset, the results are then merged using a wrapper. However, in a recent publication, Ram\'irez-Gallego et al.~\cite{Ramirez-Gallego2016} presented three implementations of an extended version of the popular mRMR feature selection filter, including a distributed version under Spark. Moreover, for the specific case of ReliefF, Huang et al.~\cite{Huang2009} proposed an optimization to improve the computation efficiency on large datasets, but this improvement is only useful when no random sampling of the instances is performed for the weights approximation. In other words, it only works when the $m$ parameter (see Section~\ref{sec:reliefF}) is set equal to the number of instances. Furthermore, the experimental work they performed was carried out with datasets with a number of instances in the order of $10^4$, which is still manageable by a single machine. Finally, to the best of our knowledge, no published contribution has ever attempted to redesign the ReliefF filter method to a distributed environment as is proposed here. \section{ReliefF} \label{sec:reliefF} We briefly describe the original ReliefF algorithm by Kononenko \cite{Kononenko1994} that will serve as a conceptual basis for the redesign presented in Section~\ref{sec:direliefF}. ReliefF's central idea consists in evaluating the quality of the features by their ability to distinguish instances from one class to another in a local neighborhood, i.e. the best features are those that contribute more to increase distance between different class instances while contribute less to increase distance between same class instances. ReliefF, as mentioned above, is an extension of the original Relief method, but it is capable of working with multi-class, noisy and incomplete datasets. \begin{algorithm} \caption{ReliefF~\cite{Robnik2003,Kononenko1994}} \label{alg:reliefF} \begin{algorithmic}[1] \STATE calculate prior probabilities $P(C)$ for all classes \STATE set all weights $W[A] := 0.0$ \FOR{$i = 1$ \TO $m$} \STATE randomly select an instance $R_i$ \STATE find $k$ nearest hits $H_j$ \FORALL{classes $C \neq cl(R_i)$} \STATE from class $C$ find $k$ nearest misses $M_j(C)$ \ENDFOR \FOR{$A := 1$ \TO $a$} \STATE $AH := - \sum_{j = 1}^{k} \mathit{diff}(A,R_i,H_j) / (m \cdot k)$ \label{alg:diff1} \STATE $AM := \sum_{C \neq cl(R_i)} \left [ \left ( \frac{P(C)}{1 - P(cl(R_i))} \right )\sum_{j = 1}^{k} \mathit{diff}(A,R_i,M_j(C)) \right ] / (m \cdot k)$ \label{alg:diff2} \STATE $W[A] := W[A] + (AH + AM)$ \ENDFOR \ENDFOR \end{algorithmic} \end{algorithm} Algorithm \ref{alg:reliefF} displays ReliefF's pseudo-code, mostly preserving the original notation used in \cite{Robnik2003}. As we can observe, it consists of a main loop that iterates $m$ times, where $m$ corresponds to the number of samples from data to perform the quality estimation. Each selected sample $R_i$ equally contributes to the $a$-size weights vector $W$, where $a$ is the number of features in the dataset. The contribution for the $A$-esim feature is calculated by first finding $k$ nearest neighbors of the actual instance for each class in the dataset. The $k$ neighbors that belong to the same class as the actual instance are called hits($H$), and the other $k \cdot (c - 1)$ neighbors are called misses ($M$), where $c$ is the total number of classes, and $cl(R_i)$, represents the class of the $i-$esim sample. Once neighbors are found, their respective contributions to $A$-esim feature are calculated. The contribution of the hits collection $AH$ is equal to the negative of the average of the differences between the actual instance and each hit, note that this is a negative contribution because only non desirable features should contribute to create differences between neighbor instances of the same class. Analogously, the contribution of the misses collection $AM$ is equal to the weighted average of the differences between the actual instance and each miss, this is of course a positive contribution because good features should help to differentiate between instances of a different class. The weights for this summation are defined according to the prior probability of each class, calculated from the dataset. Finally, it is worth mentioning that dividing both $AH$ and $AS$ by $m$ simply indicates another average between the contributions of all $m$ samples. Since the $\mathit{diff}$ function returns values between $0$ and $1$, the ReliefF's weights will lie in the range $[-1, 1]$, and must be interpreted in the positive direction: the higher the weight, the higher the corresponding feature's relevance. The $\mathit{diff}$ function is used in two cases in the ReliefF algorithm, the obvious one is between lines \ref{alg:diff1} and \ref{alg:diff2} to calculate the weight. It is also used to find distances between instances, defined as the sum of the differences over every feature (Manhattan distance). The original $\mathit{diff}$ function used to calculate the difference between two instances $I_1$ and $I_2$ for a specific feature $A$ is defined in~(\ref{eq:diffNom}) for nominal features, and as in~(\ref{eq:diffNumOrig}) for numeric features. However, the latter has been proved to cause an underestimation of numeric features with respect to nominal ones in datasets with both types of features. Thereby, a so-called ramp function, depicted in~(\ref{eq:diffNumRamp}), has been proposed by Hong~\cite{Hong1997} to deal with these problem. The idea behind is to relax the equality comparison on~(\ref{eq:diffNumOrig}) by using two thresholds: $t_{eq}$ is the maximum distance between two features to still consider them equal, and analogously, $t_{diff}$ is the minimum distance between two features to still consider them different. Their default values are set to $5\%$ and $10\%$ of the feature's value interval respectively. In addition, there are also versions of the $\mathit{diff}$ function to deal with incomplete data, however, since the datasets chosen for the experiments in this work do not have missing values, we do not consider them here. \begin{equation} \label{eq:diffNom} \mathit{diff}(A,I_1,I_2) = \begin{cases} 0 & \text{if } value(A, I_1) = value(A, I_2),\\ 1 & \text{otherwise } \end{cases} \end{equation} \begin{equation} \label{eq:diffNumOrig} \mathit{diff}(A,I_1,I_2) = \frac{\left | value(A, I_1) - value(A, I_2) \right |}{ max(A) - min(A)} \end{equation} \begin{equation} \label{eq:diffNumRamp} \mathit{diff}(A,I_1,I_2) = \begin{cases} 0 & \text{if } d \leq t_{eq}, \\ 1 & \text{if } d > t_{\mathit{diff}}, \\ \frac{d-t_{eq}}{t_{\mathit{diff}}-t_{eq}} & \text{if } t_eq < d \leq t_{\mathit{diff}} \end{cases} \end{equation} \section{Spark Cluster Computing Model} We now briefly describe the main concepts behind the Spark computing model, focusing on those that will complete the conceptual basis for the description of our proposal in Section~\ref{sec:direliefF}. We also provide a short comparison with other existent computing models such as MapReduce, with the aim of justifying our selection of Spark. The main concept behind the Spark model is the Resilient Distributed Dataset or in short RDD. Zaharia et al.~\cite{Zaharia2010,Zaharia2012} defined an RDD as a read-only collection of objects, that is partitioned and distributed across the nodes of a cluster. It has the ability to automatically recover lost partitions through the record of lineage that knows the origin of the data and optionally, the calculations that went through it. Even more relevant is the fact that the operations run for an RDD are automatically parallelized by the Spark engine, this abstraction frees the programmer of having to deal with threads, locks and all the complexities involved in the traditional parallel programming. There are two types of operations that can be executed on an RDD: (i) actions and (ii) transformations. On the one hand, \emph{actions} are the mechanism that permit to obtain results from a Spark cluster; five commonly used actions are: $reduce$, $sum$, $aggregate$, $sample$ and $collect$. The action $reduce$ is used to aggregate the elements of an RDD, by applying a commutative and associative function that receives as arguments two elements of the RDD an returns one element of the same type. Action $sum$ is simply a shorthand for a reduce action that sums all the elements on the RDD. Next, action $aggregate$ has a similar behavior to $reduce$, but its return type can be different from the type of the elements of the RDD. It works in two steps: the first one aggregates the elements of each partition and returns an aggregated value for each of them, the second one, merges these values between all partitions to a single one, that becomes the definitive result of the action. Lastly, actions $sample$ and $collect$ are also similar, the former takes an amount of elements and returns a random sample of this size from the RDD, and the latter, simply returns an array with all the elements in the RDD, this of course has to be done with care to prevent exceeding the maximum memory available at the driver (as later explained). On the other hand, \emph{transformations} are the mechanism for creating an RDD from another one. Since RDDs are read-only, a transformation does not affect the original RDD but creates a new one. Three of the most important transformations are: $map$, $flatMap$ and $filter$. The first two: $map$ and $flatMap$ are similar, they return a new RDD that is the result of applying a function to each element of the original one. In the case of $map$, the function applied takes a single argument and returns a single element, thus the new RDD has the same number of elements that the original one. In the case of $flatMap$, the applied function takes a single element but it can return zero or more elements, therefore the resulting RDD is not required to have the same number of elements as the original one. Finally, $filter$ is straightforward, it receives a boolean function to discriminate dataset elements to return a subset of it. In respect to the cluster's architecture, Spark follows the master-slave model. There is a cluster manager (master) through which the so-called driver program can access the cluster. The driver coordinates the execution of a user application by assigning tasks to the executors, which are programs that run in worker nodes (slaves). By default only one executor is run per worker. With regard to the data, RDD partitions are distributed across the worker nodes, and the number of tasks launched by the driver for each executor will be according to the number of partitions of the RDD residing in the worker. A detailed view of the discussed architecture with respect to the physical nodes can be seen in Figure~\ref{fig:sparkArch}. \begin{figure*} \includegraphics[width=1\textwidth]{fig1.eps} \caption{Spark Cluster Architecture} \label{fig:sparkArch} \end{figure*} \subsection{Other Cluster Computing Models} Spark has quickly emerged from other cluster computing models (e.g. MapReduce as the most relevant) as the prefered platform due to its advantages. First of all, Spark was designed from the beginning to efficiently handle iterative jobs in memory, such as the ones used by many data mining schemes. This has lead to the quick development of a machine learning library~\cite{Meng2015} that contains redesigned distributed algorithms such as the one in this work. Moreover, besides the Spark author's own comparison \cite{Zaharia2010,Zaharia2012}, others \cite{Shi2015} have shown that Spark is faster than MapReduce in most of the data analysis algorithms tested. Second, any MapReduce program can be directly translated to Spark, i.e. the MapReduce model can be completely expressed using the $flatMap$ and $groupByKey$ operations in Spark. Finally, as previously stated, Spark provides a wider range of operations other than $flatMap$ and $groupByKey$. This being said, its worth mentioning that other models have already tried to fulfill the lack of efficient iterative job handling in MapReduce. Two of them are HaLoop \cite{Bu2010} and Twister \cite{Ekanayake2010}. However, even though they have support for executing iterative MapReduce jobs, automatic data partitioning, and Twister has the ability to keep it in-memory, they both prevented an interactive data mining and can indeed be considered subsets of Spark functionality. By the time, both projects have become outdated with newest versions dating from 2012 and 2011 respectively. A recently work by Liu et al~\cite{Liu2016} compared parallelized versions of a neural network algorithm over Hadoop, HaLoop and Spark, concluding that Spark was the most efficient in all cases. \section{DiReliefF} \label{sec:direliefF} In this section, we describe our proposed algorithm. The first design decision is where to concentrate the parallelization effort as the ReliefF algorithm could be described as an embarrassingly parallel algorithm since its outermost loop goes through completely independent iterations, each of these can be directly executed in different threads as stated by Robnik and Kononenko~\cite{Robnik2003}. However, parallelizing the algorithm in such way ties the parallelization to the number of samples $m$, and prevents Spark from doing optimizations based on the resources available, the size of the dataset, the number of partitions and the data locality. This would also require that every thread would read through the whole dataset, while as we show below, there is only one pass needed to process the distances and calculate the feature weights. Furthermore, continuing with Robnik and Kononenko discussion, ReliefF algorithm's complexity is $\mathcal{O}(m \cdot n \cdot a)$, where $n$ is the number of instances in the dataset, $m$ is the number of samples taken from the $n$ instances and $a$ is the number of features. Moreover, the most complex operation is the selection of the $k$ nearest neighbors for two reasons: first, the distance from the current sample to each of the instances must be calculated with $\mathcal{O}(n \cdot a)$ steps; and second, the selection must be carried out in $\mathcal{O}(k \cdot log(n))$ steps. As a result, the parallelization is focused on these stages rather than on the $m$ independent iterations. We need to bear in mind that the ReliefF algorithm can be considered as a function applied to a dataset $DS$, having as input parameters the number of samples $m$ and the number of neighbors $k$, and returning an $a-$size vector of weights $W$, as shown in~(\ref{eq:reliefFFunction}). Thus, the ReliefF algorithm can be interpreted as the calculation of each individual weight $W[A]$, using~(\ref{eq:weightA}), where $\mathit{sdiffs}$~(\ref{eq:sumOfDiffs}) represents a function that returns the total sum of the differences in the $A-$esim feature between a given instance $R_i$, and a set $NN_{C,i}$ of $k$ neighbors of this instance where all belong to the same class $C$. \begin{equation} \label{eq:reliefFFunction} \mathit{reliefF}(DS,s,w) = W \end{equation} \begin{equation} \label{eq:weightA} W[A] = \frac{1}{m} \cdot \sum_{i = 1}^{m} \left [ -\mathit{sdiffs}(A, R_i, cl(R_i)) + \sum_{C \neq cl(R_i)} \left [ \left ( \frac{P(C)}{1 - P(cl(R_i))} \right ) \mathit{sdiffs}(A,R_i,C) \right ] \right ] \end{equation} \begin{equation} \label{eq:sumOfDiffs} \mathit{sdiffs}(A,R_i,C) = \frac{1}{k} \cdot \sum_{j = 1}^{k} \mathit{diff}(A,R_i,NN_{C,i,j}) \end{equation} The dataset $DS$ can be defined (see~(\ref{eq:dataset})) as a set of $n$ instances each represented as a pair $I_i = (F_i,C_i)$ of a features vector $F_i$ and a class $C_i$. \begin{equation} \label{eq:dataset} DS = \left \{ (F_1,C_1),(F_2,C_2),\cdots,(F_n,C_n) \right \} \end{equation} Given the initial definitions and assuming that the features types (nominal or numeric) are stored as metadata, we first calculate the maximum and minimum values for all continuous features in the dataset. These values are needed by the $\mathit{diff}$ function (see~(\ref{eq:diffNumRamp}) and~(\ref{eq:diffNumOrig})). Our implementation, as the original version, uses~(\ref{eq:diffNom}) for nominal attributes and selects between~(\ref{eq:diffNumRamp}) and~(\ref{eq:diffNumOrig}) for numeric attributes via an initialization parameter. The task of finding maximum and minimum values is efficiently achieved applying a $reduce$ action with a function $fmax$ ($fmin$) that given two instances returns a third one containing the maximum (minimum) values for each continuous feature. This is shown for maximum values on~(\ref{eq:max}). \begin{align} \label{eq:max} MAX & = DS.reduce(fmax) \nonumber \\ MAX & = (max(F_1[1],\cdots,F_n[1]), \cdots , max(F_1[a],\cdots,F_n[a])) \end{align} The next step consists in calculating the prior probabilities of all classes in $DS$, this is also a requirement of the $\mathit{diff}$ function. This values can be essentially obtained by the means of a $map$ and a $reduceByKey$ transformation, the former simply returns a dataset with all instance classes paired with a value of one, and the latter sums these ones using the class as a key, thereby obtaining a set of pairs $(C_i, count(C_i))$ containing the classes and the number of instances in $DS$ belonging to that class, which can simply be divided by $n$ to obtain the priors. Equations~(\ref{eq:priors}) depict the previous discussion, note that with the use of $collect$, $P$ is turned into a local array rather than an RDD. \begin{align} \label{eq:priors} f(I) & = (cl(I),1) \nonumber \\ g(a,b) & = a + b \nonumber \\ h((a,b)) & = (a, b / n) \nonumber \\ P & = DS.map(f).reduceByKey(g).map(h).collect() \nonumber \\ P & = \left \{ (C_1, prior(C_1)), \cdots , (C_c, prior(C_c)) \right \} \end{align} A rather short step, is the selection of the $m$ samples, this is accomplished with the use of the $takeSample$ action, as shown in~(\ref{eq:samples}). \begin{equation} \label{eq:samples} R = ( R_1,\cdots,R_m) = DS.takeSample(m) \end{equation} Now we can proceed to the computationally intensive step of finding the $k$ nearest neighbors of each sample for each class. We first find the distances from every sample $m$ to each of the $n$ instances. This can be directly accomplished by means of a $map$ transformation applied to $DS$ where for every instance $I$, a vector of distances from it to the $m$ samples is returned (as shown in~(\ref{eq:distancesDataset})). \begin{align} \label{eq:distancesDataset} distances(I) & = (distance(I,R_1), \cdots ,distance(I,R_m)) \nonumber \\ f(I) & = (I, distances(I)) \nonumber \\ DD & = DS.map(f) \nonumber \\ DD & = \left \{ (I_1, distances(I_1)),\cdots (I_n, distances(I_n))) \right \} \end{align} Next, as shown in~(\ref{eq:nearestNeighbors}), we are able to obtain the nearest neighbors $NN$ matrix by using an $aggregate$ action, where each element $NN_{C,i}$ of this matrix is a vector with the $k$ nearest neighbors of a sample $R_i$ belonging to a class $C$. As stated before, the $aggregate$ action has two steps. The first step is defined in the function $localNN$ that returns a local neighbors matrix $LNN$ for each partition of the RDD. This matrix, has a similar structure as the $NN$ matrix but each vector element is treated instead as a $k$-sized binary heap that is used to incrementally store the nearest neighbors found during the traverse of the local partition. The second step of the $aggregate$ action is the merging of the local matrices, the defined function $mergeNN$ combines two local matrices by merging its individual binary heaps, keeping only the elements with shorter distances. Once both functions have been defined, a call to the $aggregate$ action can be performed, providing also an empty matrix $LNN$ structure (with empty heaps) so the $localNN$ can start aggregating neighbors to it. Lastly note that since the $NN$ matrix is obtained via an action, it is a local object and not an RDD. Calculating it concludes the complex step in the algorithm, step that has been fully implemented using parallel operations. \begin{align} \label{eq:nearestNeighbors} NN & = \begin{bmatrix} NN_{1,1} & \cdots & NN_{1,m}\\ \vdots & \ddots & \vdots \\ NN_{c,1} & \cdots & NN_{c,m} \end{bmatrix} \nonumber \\ NN_{C,i} & = (N_1,\cdots,N_k \mid \forall N \ cl(N) = C) \nonumber \\ localNN(I, LNN_{in}) & = LNN_{out} \nonumber \\ mergeNN(LNN_a, LNN_b) & = LNN_{merged} \nonumber \\ NN & = DS.aggregate(emptyNN, localNN, mergeNN) \end{align} Once the $NN$ matrix is stored on the driver program, operations are not distributed in the cluster anymore. However, this is not a problem but a requirement, because $NN$ matrix is small: $c \times m$. The last remarkable step, consists in obtaining the matrix $SDIF$. Each element $SDIF_{C,i}$ represents an $a$-size vector, and each element of this vector stores the sum of the differences for the $A$-esim feature between the $NN_{C,i}$ group of neighbors and the $R_i$ sample. Moreover, each element of the vector $SDIF_{C,i}$ can be calculated by mapping the $\mathit{diff}$ function over the $A$-esim feature of all the instances in the $NN_{C,i}$ vector and then summing this differences. Observe that this $map$ and $sum$ functions are not RDD-related anymore, but local equivalents that can be parallelized only on the driver's local threads. Finally, note that each element of each vector of the matrix $SDIF$ effectively represents the value shown in~(\ref{eq:sumOfDiffs}), thus, the final vector of weights, $W$, can be easily calculated by applying the formula given on~(\ref{eq:weightA}) using the already obtained $P$ set with the prior probabilities. Equations \ref{eq:sumOfDiffsMatrix} depict the above discussion and a resume of the algorithm's main pipeline discussed in this section is shown in Figure \ref{fig:relieffMain}. \begin{align} \label{eq:sumOfDiffsMatrix} SDIF_{C,i} & = (\mathit{sdiffs}(1,R_i,C),\cdots,\mathit{sdiffs}(a,R_i,C)) \nonumber \\ f(N) & = diff(A,N,R_i) \nonumber \\ SDIF_{C,i,A} & = \mathit{sdiffs}(A,R_i,C) = NN_{C,i}.map(f).sum / k \nonumber \\ SDIF & = \begin{bmatrix} SDIF_{1,1} & \cdots & SDIF_{1,m}\\ \vdots & \ddots & \vdots \\ SDIF_{c,1} & \cdots & SDIF_{c,m} \end{bmatrix} \nonumber \\ \end{align} \begin{figure} \includegraphics[width=1\textwidth]{fig2.eps} \caption{DiReliefF's Main Pipeline} \label{fig:relieffMain} \end{figure} Finally, there is one implementation issue worth mentioning. As previously stated, an RDD is designed to be stored in memory, but this does not happen automatically, so, if the RDD is going to be used in future operations it must be explicitly cached. In our case, the most complex part of the algorithm is the calculation of the distances $DD$ and the $NN$ matrices. This calculation is effectively performed in a single pass through the dataset initiated by the $aggregate$ action, and therefore, caching of any intermediate result would indeed cause a waste of resources. However, the initial part of the algorithm that requires the calculations of the maximum, minimum and priors, each require a pass through the dataset $DS$ and therefore can take some advantage of caching it but only when it fits in the distributed memory of the cluster. When it does not, caching does not help because its benefits are overshadowed due to the time needed for writing the dataset to disk. As a result, in our implementation caching is disabled by default but can be enabled with a parameter (it should be enabled only when we can assure that the dataset fits in the distributed memory). \section{Experimental Evaluation} \label{sec:experiments} In this section, experimental results obtained from different executions of the proposed algorithm are presented. The experiments were performed with the aim of testing the algorithm scalability and its time and memory consumption with respect with the traditional version. Tests were also performed in order to observe sample sizes were the algorithm's weights become stable. For the realization of the tests, an 8-node cluster of virtual machines was used, one node is used as a master and the rest are slaves, the cluster runs over the following hardware-software configuration: \begin{itemize} \item Host Processors: 16 X Dual-Core AMD Opteron 8216 \item Host Memory: 128 GB DDR2 \item Host Storage: 7 TB SATA 2 Hard Disk \item Host Operating System: CentoOS 7 \item Hypervisor: KVM (qemu-kvm 1.5.3) \item Guests Processors: 2 \item Guests Memory: 16 GB \item Guests Storage: 500 GB \item Java version: OpenJDK 1.8 \item Spark version: 1.6.1 \item HDFS version: 2.6.4 \item WEKA version: 3.8 \end{itemize} During the first part of the tests, the ECBDL14~\cite{Bacardit2012} dataset was used, this dataset comes from the Protein Structure Prediction field, and it was used during the ECBLD14 Big Data Competition of the GECCO'2014 international conference. The dataset has approximately 33.6 million instances, 631 attributes, 2 classes, 98\% of negative examples and occupies about 56GB of disk space. In a second part of the experiments, we use other three datasets briefly described in Table~\ref{tb:datasets}. HIGGS~\cite{Baldi2014}, from the UCI Machine Learning Repository~\cite{Lichman2013}, is a recent dataset that represents a classification problem that distinguishes between a signal process which produces Higgs bosons and a background process which does not. KDDCUP99~\cite{Ma2009} represents data from network connections and classifies them between normal connections and different types of attacks (a multi class problem). Finally, EPSILON is an artificial data set built for the Pascal Large Scale Learning Challenge in 2008\footnote{\url{http://largescale.ml.tu-berlin.de/about/}}. \begin{table}[tbp] \caption{Datasets used in the experiments} \label{tb:datasets} \begin{tabular}{lp{0.7in}p{0.5in}p{1.0in}l} \hline\noalign{\smallskip} Dataset & No. of Inst. & No. of Feats. & Features Types & Problem Type \\ \noalign{\smallskip}\hline\noalign{\smallskip} ECBDL14 & $\sim$33.6 million & 632 & Numerical and Categorical & Binary \\ HIGGS & 11 million & 28 & Numerical & Binary \\ KDDCUP99 & $\sim$5 million & 42 & Numerical and Categorical & Multiclass \\ EPSILON & 1/2 million & 2,000 & Numerical & Binary \\ \noalign{\smallskip}\hline \end{tabular} \end{table} Initially, all tests are run with a number of neighbors $k=10$ which is a typical choice \cite{Kononenko1994}, and a relatively low number of samples $m=10$ to keep execution times reasonable. However, during the stability tests larger number of samples are used. In addition, HDFS is used to store all the datasets or samples of datasets in the experiments related to the distributed version. Conversely, the local file system is used for tests of the traditional version of ReliefF. Regarding the rest of the implementation parameters, two of them were left fixed for all the experiments. The first one selects the original $\mathit{diff}$ function in~(\ref{eq:diffNumOrig}) for the distance evaluation, as this is the one used in the ReliefF implementation in WEKA. The second parameter refers to whether or not apply caching in Spark and as stated before, caching provides benefits only when the entire datasets fully fits in memory. In order to prevent making the results obscure, it was decided that caching was disabled for all the experiments. An important configuration issue refers to driver memory consumption. In Spark computation model, there is no communication between tasks, so all the task's results will be sent to the driver, this is especially important for the $aggregate$ action, because every task performing the $localNN$ operation will return a $LNN$ matrix to the driver that then is going to be merged. The Spark configuration parameter $spark.driver.maxResultSize$ has a default value of 1GB but it was set to 6 for all the experiments performed. This is specifically important for tests involving larger matrix sizes, i.e., those with higher values of $m$ or $c$. \subsection{Empirical Complexity} \label{subsec:complexity} Figure \ref{fig:timeAndMemory} shows time and memory consumption behavior of the distributed version of ReliefF version versus the one implemented in the WEKA platform~\cite{Hall2009} for incrementally sized samples of the ECBDL14 dataset. To make the comparison possible, the WEKA version was executed under the host environment with no virtualization. It is worth noting that for the WEKA version a 30\% sample was the largest that could be tested because larger samples would need more memory than is available in the system (showing the lack of scalability). The distributed version, in addition to being able to handle the whole dataset, preserves a linear behavior in relation to the number of instances, and it is also capable of processing data in less time by leveraging the cluster nodes. Another fact to observe is the change in the slope of the linear behavior observed by the Spark version between the 40\% and 50\% samples of the dataset. This is due to the fact that during this interval the dataset overflows the available memory in the cluster and the Spark engine starts using disk storage. In other words, the algorithm is capable of maintaining a linear complexity even when the dataset does not fit into memory. The mentioned overflow can be observed on the right graph in Figure~\ref{fig:timeAndMemory}. This graph shows a previously mentioned advantage of Spark, i.e, it is designed to run in commodity hardware. A simple look to the memory consumption of the traditional version shows that the required amount of memory quickly overgrows beyond the limits of an ordinary computer. However, in our distributed version, we can observe that using nodes, in our case with 16GB of memory, is enough to handle the task. \begin{figure*} \includegraphics[width=1\textwidth]{fig3.eps} \caption{Execution time and memory consumption of Spark DiRelieF and WEKA ReliefF versions} \label{fig:timeAndMemory} \end{figure*} Analogous run time results are obtained by varying the number of features and the number of samples (see Figure~\ref{fig:featsAndSampleSize}) confirming an empirical complexity equivalent to the original one, i.e., $\mathcal{O}(m \cdot n \cdot a)$. \begin{figure*} \includegraphics[width=1\textwidth]{fig4.eps} \caption{Execution time of Spark DiReliefF and WEKA ReliefF with respect to parameters $a$ and $m$} \label{fig:featsAndSampleSize} \end{figure*} \subsection{Scalability} \label{subsec:scalability} Now, we delve into the topic of scalability. To keep execution times within manageable limits, the largest test was performed with a 30\% sample size of the ECBDL14 dataset, following this, 10\% and 1\% sample sizes were used. Also, in oder to examine a smaller dataset, tests with 50\% and 10\% samples of the HIGGS dataset were run. Figure~\ref{fig:scalability} shows the behavior of the distributed algorithm with respect to the number of cores used. As it can be observed, adding cores at the beginning greatly contributes to reducing the execution time, but once a threshold is reached, the contribution is rather subtle or even null. Such threshold depends on the size of the dataset, so, larger datasets can take benefit more of larger number of cores. On the other hand, smaller datasets will face the case were they do not have enough partitions to be distributed over all the cluster nodes. In this latter case, adding more nodes will not provide any performance improvement. As an example to quantify this fact, it can be observed in Figure~\ref{fig:scalability} that the 30\% sized sample of the ECBDL14 dataset can take advantage of 7 cores, while the 1\% sample size only can take advantage of 4 cores. Similarly, there is no practical advantage of using more than 1 core for the 10\% sample size of the HIGGS dataset. \begin{figure*} \includegraphics[width=1\textwidth]{fig5.eps} \caption{Execution time of Spark DiReliefF and WEKA ReliefF with respect to the number of cores involved} \label{fig:scalability} \end{figure*} Figure~\ref{fig:scalability} also shows with an horizontal line, the WEKA version time of the 30\% sample of the ECBDL14 dataset, since it can only take advantage of one core, the execution time is constant. However, the time is better than the distributed version in the case where 4 or less cores are involved, clearly because it does not need to deal with the driver scheduling, the selection of an executor and communication between both of them over the network, as the distributed version does. For this reason, the distributed version of ReliefF is only useful for large datasets, and in fact, regarding execution time, the 30\% sample of ECBDL14 is the only dataset in our tests that can take real advantage of the distributed version. \subsection{Stability} \label{subsec:stability} The following set of tests was made in order to check the stability of the algorithm. In this case, with stability we refer to similarity of the algorithm rankings for different executions in the same dataset. Many stability measures have been defined, a commonly used is the Consistency Index presented by Kuncheva~\cite{Kuncheva2007}, but it cannot be directly applied because it requires to define of a threshold for the selection of a subset. In addition, Kalousis et al.~\cite{Kalousis2006} proposed the use of the Pearson correlation coefficient to measure the similarity between features rankings, but previous tests showed that paradoxically returns unstable results for the EPSILON and HIGGS (pure numerical) datasets. For these reasons, we opted here for a simpler stability indicator, a mere average of the absolute differences between every feature weights of two rankings. More formally, if we have two feature ranking vectors $W1$ and $W2$, the average difference between them is calculated as shown in~(\ref{eq:stability}). \begin{align} \label{eq:stability} W1 & = (w1_1,w1_2,\cdots,w1_a) \nonumber \\ W2 & = (w2_1,w2_2,\cdots,w2_a) \nonumber \\ \mathit{AvgDiff} & = \frac{1}{a}\sum_{A=1}^{a} \left | w1_A - w2_A \right | \end{align} As Robnik and Kononenko~\cite{Robnik2003} stated, the correct sample size $m$ is problem dependent, so for the tests, we used all of the datasets described in Table~\ref{tb:datasets}. Moreover, their experiments show that as the number of examples increases, the required sample size diminishes. Figure~\ref{fig:stability} shows the results obtained and evidences that the biggest gain in stability is obtained with a sample size between 50 and 100 instances, thereby confirming that relative small sample sizes are enough to obtain stable results in the tested large scale datasets. \begin{figure*} \includegraphics[width=1\textwidth]{fig6.eps} \caption{DiReliefF's average difference in weight ranks for increasing values of $m$ in different datasets} \label{fig:stability} \end{figure*} \section{Conclusions} In this work, we have presented DiRelieF, a distributed version of the well-known ReliefF feature selection algorithm. This version was implemented using the emerging Apache Spark programming model to deal with current Big Data requirements such as such as failure recovery and scalability over a cluster of commodity computers. Even when the ReliefF algorithm is easily parallelizable by associating jobs to each independent iteration and then merging this results, this method ties the number of jobs to the sample size (number of iterations) and requires and equal number of passes through the dataset. For this reason we designed an alternative version whose core is based on the Spark $map$ and $aggregate$ operations and on the use of binary heaps. This method does not suffer from these problems and requires a single pass through the whole dataset to perform the main operations of the algorithm compared to the $m$ passes required by the original version. As part of the experimental work, we have also compared our proposal with a non distributed version of the algorithm implemented on the WEKA platform. Our results showed that the non distributed version is poorly scalable, i.e., it is unable to handle large datasets due to memory requirements. Conversely, our version is fully scalable and provides better execution times and memory usage when dealing with very large datasets. Our experiments also showed that the algorithm is capable of returning stable results with sample sizes that are much smaller than the size of the complete dataset. \bibliographystyle{spmpsci}

\section{Introduction} The relay channel was introduced in 1971 when van der Meulen \cite {Meu71} proposed a channel model consisting of one source, one relay, and one destination. The relay aids the communication between the source and the destination so that increased robustness, higher transmission efficiency, and/or larger coverage range can be achieved. As smallest but fundamental unit of large network topologies, the relay channel has been extensively studied focusing on both theoretical and implementation aspects. Decode-and-forward (DF) relaying is the most researched protocol for relay channels. In particular, the design of distributed channel codes has attracted considerable attention. The concept of distributed Turbo coding (DTC) was proposed in \cite{ZV03}, which offered a new fashion of distributed code design. Low-density parity-check (LDPC) codes were considered for distributed coding for example in \cite{CBSA07}, \cite{HD07} and \cite{RY07}. Different approaches were presented to optimize LDPC codes for given channel conditions. For LDPC block codes, an irregular degree distribution needs to be derived to match a given channel. For a variety of channel conditions, extensive re-optimization is required. This leads to a high complexity for code adaptation and may not be feasible in practice. In this paper we propose to use LDPC convolutional codes for distributed channel coding in relay networks. LDPC convolutional codes were first proposed in \cite{FZ99} as a time-varying periodic LDPC code variation. Then the idea was further developed in, e.g., \cite{PFSLZC08}, \cite{LSCZ10}. Recently, it has been proven analytically in \cite{KRU10} that the belief-propagation (BP) decoding threshold of an LDPC convolutional code achieves the optimal maximum \emph{a posteriori} probability (MAP) threshold of the corresponding LDPC block code with the same variable and check degrees. This code in turn approaches the capacity as the node degrees increase. Furthermore, regular LDPC convolutional codes allow us to avoid complicated re-optimization of the degree distributions for varying channel conditions. Meanwhile, LDPC convolutional codes enable recursive encoding and sliding-window decoding \cite{LSCZ10}, which dispels the concerns over complexity and delay. Motivated by the good properties of LDPC convolutional codes, we consider in this paper the design of bilayer LDPC convolutional codes for the relay channel. A similar code construction was proposed in \cite{RUAS10} for the wiretap channel. A protograph-based bilayer code was proposed in \cite{NND10} which applies the concept of bilayer-lengthened codes. In contrast to \cite{NND10} we present bilayer expurgated codes \cite{RY07} in this paper. In the following, we will discuss the construction of bilayer LDPC convolutional codes for given relay channels. We will prove analytically that the proposed bilayer code is capable of achieving the highest rate with DF relaying in binary erasure channels (BEC). Moreover, the regularity of degree distributions significantly simplifies the code optimization. Numerical results are provided to verify the theoretical analysis. \section{Preliminaries} \label{sec:Preliminary} In this section, firstly we introduce the transmission model we use throughout the paper. Then we briefly review the coding strategy which leads to the highest achievable rate \cite{KSA03} with DF relaying. The construction of bilayer codes \cite{RY07} is described as a practical realization of the coding strategy. \subsection{System Model} In this paper, we restrict ourself to the three-node relay channel which is composed of one source, one relay, and one destination. The source ($S$) intends to transmit its information to the destination ($D$) while the relay ($R$) provides assistance. The system model is shown in Figure~\ref{fig:Transmod}. Due to practical constraints the relay works in a half-duplex mode, which means it cannot transmit and receive at the same time or the same frequency. This implies that the transmission from the source to the destination is carried out in two phases. In the first phase, the source broadcasts while the relay and the destination listen. In the second phase, the relay transmits to the destination while the source keeps silent. We assume the transmissions on the three links to be orthogonal. \begin{figure}[htb] \psfrag{B}[][]{\scriptsize $\word{B}$} \psfrag{D}[][]{\scriptsize $\hat{\word{B}}$} \psfrag{Xr}[][]{\scriptsize $\word{X}\!_R$} \psfrag{Xs}[][]{\scriptsize $\word{X}\!_S$} \psfrag{Y1}[][]{\scriptsize $\word{Y}\!_{SR}$} \psfrag{Y2}[][]{\scriptsize $\word{Y}\!_{SD}$} \psfrag{Y3}[][]{\scriptsize $\word{Y}\!_{RD}$} \centering \includegraphics[width=.45\textwidth]{Fig/GerenalModel.eps} \vspace{-3mm} \caption{Transmission model.} \vspace{-2mm} \label{fig:Transmod} \end{figure} In the following we use $X_i$, $i\in\{S,R\}$, to denote the BPSK modulated signals which are transmitted from the source and the relay, and we use $Y_{i,j}$, $i\in\{S,R\}$, $j\in\{R,D\}$, for the channel observations of the three links. We use $C_{ij}=\emph{I}(X_i;Y_{ij})$, $i\in\{S,R\}$, $j\in\{R,D\}$ to denote the capacity of each link constrained to the BPSK modulation. In this paper we assume that perfect channel-state information (CSI) is available for code construction. \subsection{Achievable Rate} The highest transmission rate using decode-and-forward protocol for the half-duplex relay channel with orthogonal receive components is given as \cite{KSA03} \begin{equation} \label{AchiRate} \begin{split} R_{DF}= \sup_{\alpha,0\leq \alpha\leq 1} \min (\alpha \emph{I}(X_S;Y_{SR}&),\\ \alpha \emph{I}(X_S;Y_{SD})+(1-&\alpha) \emph{I}(X_R;Y_{RD})) \end{split} \end{equation} where $\alpha$ is the fraction of channel uses in the first phase, and $(1-\alpha)$ is the fraction of channel uses in the second phase. To achieve $R_{DF}$, in the first phase the source employs a capacity-achieving code for the source-relay link which guarantees successful decoding at the relay. This code may not be decodable at the destination due to the poorer channel condition on the source-destination link. Therefore, in the second phase the relay forwards additional bits to the destination in order to construct an overall lower rate code which is capacity-achieving for the source-destination link. A practical implementation of the idea is presented in the following. \subsection{Bilayer LDPC Block Codes for Relay Channels} \label{sec:BlockBiLDPC} The construction of bilayer LDPC block codes \cite{RY07} is realized in two steps corresponding to the two transmission phases. In the first phase, $K_1$ information bits $\word{B}$ are encoded by a length-$N_1$ codeword $\word{X}_S$ through a rate-$R_1$ LDPC code $\mathcal{C}_1$ (i.e., $K_1=N_1 \cdot R_1$) with the check matrix $\word{H}_S$ and transmitted. At the end of the first phase, the relay decodes $\mathcal{C}_1$, using the check matrix $\word{H}_S$, and recovers $\word{X}_S$. At the destination, additional $K_2$ bits are needed for successfully decoding $\word{X}_S$: \begin{equation*} K_2=N_1(\emph{I}(X_S;Y_{SR})-\emph{I}(X_S;Y_{SD})). \end{equation*} Therefore, in the second phase the relay generates $K_2$ new bits (syndrome, $\word{S}$) using the check matrix $\word{H}_R$. These $K_2$ syndrome bits are transmitted to the destination via a channel encoder $\mathcal{C}_2$ of rate $R_2$ using $N_2$ channel uses, i.e., $K_2=N_2 \cdot R_2$. To simplify the discussion, we assume these syndrome bits are perfectly known at the destination after decoding $\mathcal{C}_2$. Then the overall code $\mathcal{C}$ is described by the stacked check matrix $\word{H}$, and we have \begin{equation*} \word{H}\word{X}_S=\left[\begin{array}{c} \word{H}_S\\\word{H}_R\end{array}\right]\word{X}_S=\left[\begin{array}{c} \word{0}\\\word{S}\end{array}\right]. \end{equation*} That is, at the destination ($N_1-K_1$) zero check equations and $K_2$ non-zero check equations need to be satisfied in the decoding. The Tanner graph of a bilayer code example is plotted in Figure~\ref{fig:BilayerLDPC}. \begin{figure}[htb] \vspace{-2mm} \centering \includegraphics[width=.25\textwidth]{Fig/BilayerLDPC.eps} \vspace{-2mm} \caption{The Tanner graph of a bilayer LDPC code. Circles represent variable nodes, and squares indicate check nodes. The solid lines correspond to the edges in $\word{H}_S$, and the dashed lines show the connections determined by $\word{H}_R$.} \vspace{-2mm} \label{fig:BilayerLDPC} \end{figure} To achieve the optimal performance, the design of $\word{H}_S$ and $\word{H}$ needs to guarantee that $\mathcal{C}_1$ and $\mathcal{C}$ are simultaneously capacity-achieving for the source-relay link and the source-destination link respectively. The authors of \cite{RY07} approached this target by applying irregular LDPC block codes. Consequently, re-optimization is required for every given channel, which results in high complexity and infeasibility. In the next section we will show how this goal can be achieved by using regular LDPC convolutional codes leading to significantly reduced optimization overhead. If the channel codes $\mathcal{C}_1$ and $\mathcal{C}_2$ are both capacity-achieving, i.e., $R_1=C_{SR}$ and $R_2=C_{RD}$, then the achievable rate in (\ref{AchiRate}) is maximized by \vspace{-1mm} \begin{equation} \label{fraction} \alpha = \frac{N_1}{N_1+N_2} = \frac{C_{RD}}{C_{RD}+C_{SR}-C_{SD}}. \end{equation} Later in this paper we will prove that in BECs $R_{DF}$ can be achieved by applying bilayer LDPC convolutional codes. \section{Bilayer LDPC Convolutional Codes} \label{sec:BilayerLDPCCC} \subsection{LDPC Convolutional Codes} A regular $(l,r)$ time-varying binary LDPC convolutional code can be defined by a syndrome former matrix \cite{LSCZ10} \begin{equation*} \word{H}^T=\left[\begin{array} {c c c c c} \ddots & & \ddots & &\\ \word{H}_0^T(1) & \ldots & \word{H}_w^T(1+w) & & \\ & \ddots & & \ddots & \\ & & \word{H}_0^T(t) & \ldots & \word{H}_w^T(t+w)\\ & & \ddots & & \ddots \\ \end{array}\right], \end{equation*} where $l$ is the variable degree and $r$ is the check degree. We assume that at each time instant $t$ ($t=1,2,...,L$) the number of variable nodes is $M$. Then each submatrix $\word{H}_i^T(t+i)$ is a $M\times(Ml/r)$ binary matrix. The largest $i$ such that $\word{H}_i^T(t+i)$ is nonzero for some $t$ is called the syndrome former memory $w$. The matrix $\word{H}^T$ is sparse. There are many variations of LDPC convolutional codes in the literature. In this paper, we denote an LDPC convolutional code by four parameters $\{l,r,L,w\}$. The memory constraint $w$ can be any non-negative integer. We assume that each of the $l$ edges of a variable node at time $t$ uniformly and independently connects to the check nodes in the time range $[t,...,t+w]$. More precisely, for each variable node at time $t$, one can define a \emph{type} $\mathcal{M}_t$\footnote{Index of the variable node is omitted for the ease of notation.} \cite{KRU10} which is a $w$-tuple of non-negative integers, $\mathcal{M}_t=(m_{t,t},...,m_{t,t+j},...,m_{t,t+w})$, $j\in[0,w]$, and $\sum_{j} m_{t,t+j}=l$. The element $m_{t,t+j}$ indicates that there are $m_{t,t+j}$ edges connecting the designated variable node at time $t$ and the check nodes at time $t+j$. For each variable node, $\mathcal{M}_t$ is uniformly and independently chosen from all possible types. It has been stated in \cite{KRU10} that the $\{l,r,L,w\}$ code ensemble is capacity achieving and easier to analyze. However, experimentally it shows a worse trade-off between rate, threshold and block length. Another variant, the $\{l,r,L\}$ ensemble, can be considered as a special case of the more general code ensemble mentioned above. For this ensemble, the memory length $w$ always equals $l-1$. Exactly one of the $l$ outgoing edges of each variable node at time $t$ is connected to one check node at position $[t,...,t+(l-1)]$, i.e., $m_{t,t+j}\!=\!1$ for all $j\in[0,l-1]$. We observe through experiments that this type of ensemble provides good performance with moderate $M$ and $L$ when $l\geq 3$. In this paper, we use the $\{l,r,L,w\}$ ensemble for theoretical analysis while employing the $\{l,r,L\}$\! ensemble in simulations. \subsection{Bilayer LDPC Convolutional Codes for Relay Channels} Firstly, we define the structure of a bilayer LDPC convolutional code. We assume the number of variable nodes to be $N=M \cdot L$. The connections between the $N$ variable nodes and the check nodes in the first (second) layer are determined by the ensemble $\{l_1,r_1,L,w_1\}$ ($\{l_2,r_2,L,w_2\}$). If $w_1=w_2$, we denote the bilayer code by $\{l_1,l_2,r_1,r_2,L,w\}$. Note that only the edges belonging to the same layer are connected to one check node. The structure of the overall check matrix is illustrated in Figure~\ref{fig:LDPCCBilayer}. \begin{figure}[htb] \vspace{-3mm} \psfrag{D}[][]{$\word{H}^T$} \psfrag{A}[][]{ $\word{H}_{S}^T$} \psfrag{B}[][]{$\word{H}_{R}^T$} \psfrag{E}[][]{ $\ddots$} \centering \includegraphics[width=.35\textwidth]{Fig/BilayerMatrixT.eps} \vspace{-1mm} \caption{Overall check matrix of a bilayer LDPC convolutional code. The white blocks correspond to the non-zero submatrices in the first layer, and the grey blocks are for those submatrices in the second layer.} \vspace{-3mm} \label{fig:LDPCCBilayer} \end{figure} The protocol for transmitting a bilayer LDPC convolutional code for the relay channel is similar to the strategy we explained in Section~\ref{sec:BlockBiLDPC}. The information bits from the source are encoded by the single-layer code $\{l_1,r_1,L,w_1\}$ and broadcasted in the first phase. After successful decoding, the relay generates the syndrome bits using $\{l_2,r_2,L,w_2\}$. These syndrome bits are transmitted to the destination under perfect protection by another channel code in the second phase. The destination decodes the overall code by considering the zero check equations in the first layer and the non-zero check equations in the second layer. \subsection{Analysis for Binary Erasure Channels} It has been shown in \cite{KRU10} that the $\{l,r,L,w\}$ ensemble with infinite $M$ has the following properties in a binary erasure channel: for the rate of the code $R$ \vspace{-1mm} \begin{equation} \lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}R(l,r,L,w)=1-\frac{l}{r}, \label{Rate} \end{equation} and for the decoding threshold \begin{equation*} \lim_{w\rightarrow\infty}\!\lim_{L\rightarrow\infty}\!\epsilon^{B\!P}(l,r,L,w)\!=\!\!\lim_{L\rightarrow\infty}\!\epsilon^{M\!A\!P}(l,r,L,w)\!=\!\epsilon^{M\!A\!P}(l,r), \end{equation*} where $\epsilon^{BP}$ and $\epsilon^{MAP}$ are respectively the BP threshold and the MAP threshold for decoding. If we increase the degrees of the nodes, its decoding threshold approaches the Shannon limit $\epsilon_{Sh}=1-R$, \vspace{-1mm} \begin{equation} \lim_{r\rightarrow\infty}\lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}\epsilon^{BP}(l,r,L,w)=\epsilon_{Sh}. \label{Capacity} \end{equation} In the following, we will show in Theorem 2 that the bilayer LDPC convolutional code $\{l_1,l_2,r,r,L,w\}$ achieves the same Shannon limit as the standard single-layer ensemble $\{l_1+l_2,r,L,w\}$ \cite{RUAS10}. As a preparation for the theorem, we introduce the following lemma. \begin{lemma} If $M$, $L$ and $w$ go to infinity in this order, the density evolution of a single-layer LDPC convolutional code $\{l,r,L,w\}$ in a binary erasure channel can be written as \begin{equation*} p^{(i)}=\epsilon (q^{(i-1)})^{l-1} \hspace{3mm} \mathrm{and} \hspace{3mm} q^{(i)}=1-(1-p^{(i)})^{r-1}, \end{equation*} where $p^{(i)}$ ($q^{(i)}$) is the erasure probability from a variable (check) node to a check (variable) node in the $i$-th iteration, and $\epsilon$ is the erasure probability of the channel. \end{lemma} \begin{proof} In the following we refer to the check nodes connected to a given variable node as the active check nodes for that variable node. We use $p_{t,t+j}^{m,(i)}$ to denote the probability that the message from a given variable node at time $t$ to the $m$-th active check node at time $t+j$ in decoding iteration $i$ is erased. In the first iteration, $p_{t,t+j}^{m,(1)}=\epsilon$ for all $m$ and $j$. We use $q_{t+j,t}^{n,(i)}$ to represent the probability that the message from the $n$-th active check node at $t+j$ to the given variable node at $t$ is erased. Then we have \begin{equation} \label{Lem1} p_{t,t+j}^{m,(i)}=\epsilon\prod_{k=0,k\neq j}^{w} \left(\prod_{n=1}^{m_{t,t+k}}q_{t+k,t}^{n,(i-1)}\right) \cdot \prod_{v=1,v\neq m}^{m_{t,t+j}}q_{t+j,t}^{v,(i-1)}. \end{equation} If $M\rightarrow\infty$, the messages from different nodes at the same time instant behave identically \cite{LSCZ10}. Then (\ref{Lem1}) reduces to \begin{equation} \label{Lem2} p_{t,t+j}^{(i)}=\epsilon\prod_{k=0,k\neq j}^{w} (q_{t+k,t}^{(i-1)})^{m_{t,t+k}} \cdot (q_{t+j,t}^{(i-1)})^{m_{t,t+j}-1}. \end{equation} The messages from nodes at different time instants can behave differently and are usually tracked separately. However, if we have $L\rightarrow\infty$, the effect of boundaries caused by the initialization and the termination of the code vanishes. We can then consider the code asymptotically regular \cite{LFZC09}. The message updating is averaged over $w+1$ time instants. Therefore, if $w\rightarrow\infty$, the messages from the nodes at different time instants have asymptotically identical distribution. Eventually, (\ref{Lem2}) is simplified to \vspace{-0.5mm} \begin{equation*} p^{(i)}=\epsilon (q^{(i-1)})^{l-1}. \end{equation*} \vspace{-5mm} Similarly, we also obtain \vspace{-1mm} \begin{equation*} q^{(i)}=1-(1-p^{(i)})^{r-1}. \vspace{-5mm} \end{equation*} \end{proof} Now we look at the relation between the bilayer LDPC convolutional code ensemble $\{l_1,l_2,r,r,L,w\}$ and the standard single-layer ensemble $\{l_1+l_2,r,L,w\}$. \begin{theorem}{\cite{RUAS10}} We denote a bilayer LDPC convolutional code of length $N=M \cdot L$ by $\{l_1,l_2,r_1,r_2,L,w\}$, where $l_1$ and $l_2$ are respectively the variable degrees of the two layers, $r_1$, $r_2$ are the check degrees of the two layers, and $w$ is the common memory constraint. If we assume the two layers take the same check degree, i.e., $r_1=r_2=r$, then the bilayer LDPC convolutional code $\{l_1,l_2,r,r,L,w\}$ approaches the same Shannon limit as the single-layer LDPC convolutional code $\{l_1+l_2,r,L,w\}$. \end{theorem} \begin{proof} For the completeness of the proof, we repeat the derivation of the BP decoding threshold which was previously given in \cite{RUAS10}. According to Lemma 1, we write for the first layer of the bilayer LDPC convolutional code, \begin{equation*} p_1^{(i)}=\epsilon (q_1^{(i-1)})^{l_1-1} (q_2^{(i-1)})^{l_2} \hspace{2mm} \mathrm{and} \hspace{2mm} q_1^{(i)}=1-(1-p_1^{(i)})^{r_1-1}. \end{equation*} For the second layer of the code, we have \begin{equation*} p_2^{(i)}=\epsilon (q_1^{(i-1)})^{l_1} (q_2^{(i-1)})^{l_2-1} \hspace{2mm} \mathrm{and} \hspace{2mm} q_2^{(i)}=1-(1-p_2^{(i)})^{r_2-1}. \end{equation*} Since $r_1=r_2=r$ and $p_1^{(1)}=p_2^{(1)}=\epsilon$, we obtain from the iterations $p_1^{(i)}=p_2^{(i)}$. Then the recursion can be written as \begin{equation*} p^{(i)}=\epsilon(1-(1-p^{(i-1)})^{r-1})^{l_1+l_2-1}. \end{equation*} This indicates that the bilayer LDPC convolutional code has the same BP threshold as the $\{l_1+l_2,r,L,w\}$ ensemble. The rate of the bilayer LDPC convolutional code satisfies \vspace{-1mm} \begin{equation} \label{BLDPCCRate} \lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}R(l_1,l_2,r,r,L,w)=1-\frac{l_1}{r}-\frac{l_2}{r}, \vspace{-1mm} \end{equation} and the decoding threshold achieves the Shannon limit, \begin{equation} \label{BLDPCCLimit} \lim_{r\rightarrow\infty}\lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}\epsilon^{BP}(l_1,l_2,r,r,L,w)=\frac{l_1}{r}+\frac{l_2}{r}. \end{equation} According to (\ref{Rate}) and (\ref{Capacity}), obviously the single-layer code $\{l_1+l_2,r,L,w\}$ has the same rate as in (\ref{BLDPCCRate}) and achieves the same limit as in (\ref{BLDPCCLimit}). Therefore, the theorem is proven. \end{proof} For the design of bilayer LDPC convolutional codes in relay channels, firstly we choose an $\{l_1,r,L,w\}$ ensemble which is capacity achieving for the source-relay link. Afterwards the relay generates the syndrome bits according to $\{l_2,r,L,w\}$ and forwards them to the destination. The overall code structure is consequently $\{l_1,l_2,r,r,L,w\}$. In the following we will show this overall code is capacity achieving for the source-destination link. In addition, it enables the highest achievable rate $R_{DF}$ of the relay channel. \begin{theorem} For a binary erasure relay channel, we can find an LDPC convolutional code $\mathcal{C}_1=\{l_1,r,L,w\}$ achieving the capacity for the source-relay link and simultaneously its bilayer extension $\mathcal{C}=\{l_1,l_2,r,r,L,w\}$ achieving the capacity for the source-destination link. Meanwhile, the above code construction provides the highest achievable rate with decode-and-forward relaying as in (\ref{AchiRate}). \end{theorem} \begin{proof} We assume that the erasure probability for the source-relay link and the source-destination link are $\epsilon_{SR}$ and $\epsilon_{SD}$, respectively, and $\epsilon_{SR}<\epsilon_{SD}$. The corresponding channel capacities for these two links are \begin{equation*} C_{SR}=1-\epsilon_{SR},\ C_{SD}=1-\epsilon_{SD}. \end{equation*} We use a regular LDPC convolutional code $\{l_1,r,L,w\}$ with $l_1/r=\epsilon_{SR}$ for the transmission in the first phase. According to (\ref{Rate}) and (\ref{Capacity}), we have \vspace{-1mm} \begin{equation*} \lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}R(l_1,r,L,w)=1-\frac{l_1}{r}=1-\epsilon_{SR} \end{equation*}\vspace{-2mm} \begin{equation*} \mathrm{and}\hspace{3mm}\lim_{r\rightarrow\infty}\lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}\epsilon^{BP}(l_1,r,L,w)=\epsilon_{SR}. \end{equation*} Hence, $\mathcal{C}_1$ is capacity achieving, and error-free decoding can be guaranteed at the relay. We assume that the number of variable nodes of $\mathcal{C}_1$ is $N_V$ and the number of check nodes of $\mathcal{C}_1$ is $N_{C1}$, then \begin{equation*} N_{C1}=l_1 N_V/r. \end{equation*} The number of additional bits needed at the destination is \begin{equation*} N_{C2}=N_V(C_{SR}-C_{SD})=N_V(\epsilon_{SD}-\epsilon_{SR}), \end{equation*} and these bits are provided by the syndrome generated at the relay. At the destination, the total number of check nodes is \begin{equation*} N_C=N_{C1}+N_{C2}=N_V(\frac{l_1}{r}+\epsilon_{SD}-\epsilon_{SR})=\epsilon_{SD}N_V. \end{equation*} The additional $N_{C2}$ check equations bring in $r N_{C2}$ edges, and the corresponding variable degree $l_2$ follows as \begin{equation*} l_2=r N_{C2}/N_V=r(\epsilon_{SD}-\epsilon_{SR}). \end{equation*} From Theorem 2, we have for the source-destination link \begin{equation*} \lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}R(l_1,l_2,r,r,L,w)=1-\frac{l_1+l_2}{r}=1-\epsilon_{SD}, \end{equation*}\vspace{-2mm} \begin{equation*} \mathrm{and}\hspace{3mm}\lim_{r\rightarrow\infty}\lim_{w\rightarrow\infty}\lim_{L\rightarrow\infty}\epsilon^{BP}(l_1,l_2,r,r,L,w)=\epsilon_{SD}. \end{equation*} Therefore, the overall code $\mathcal{C}$ achieves the capacity of the source-destination link. The number of channel uses in the first phase is $N_1=N_V$. In the second phase, we can use another capacity-achieving LDPC convolutional code to transmit the $N_{C2}$ syndrome bits to the destination. Therefore, $N_2=N_{C2}/C_{RD}$ channel uses are needed. The fraction \begin{equation*} \alpha'=\frac{N_1}{N_1+N_2} = \frac{C_{RD}}{C_{RD}+C_{SR}-C_{SD}} \end{equation*} equals the one in (\ref{fraction}), which maximizes the achievable rate. \end{proof} From Theorem 3, we can conclude that the proposed regular bilayer LDPC convolutional codes significantly simplify the code optimization. Appropriate variable and check node degrees can easily be computed from the parameters of the channels, and a complicated optimization of irregular degree distributions as for example in \cite{RY07} can be avoided. \section{Numerical Results} \label{sec:Results} In this section, we firstly give numerical results for bilayer LDPC convolutional code ensembles $\{l,r,L\}$ in binary erasure relay channels. The source broadcasts its information bits with an $\{l_1\!\!=\!\!3,r\!\!=\!\!10,L\!\!=\!\!100\}$ LDPC convolutional code. At each time instant, the number of variable nodes is set to be $M\!=\!2000$. At the relay, different values of $l_2$ ($l_2\in\{2,3,4,5\}$) are chosen. Consequently, bilayer LDPC convolutional codes of different rates are constructed. Note that rate loss is inevitable for finite $L$ \cite{LSCZ10}. We compare the decoding thresholds of both the single-layer code and the bilayer codes with the corresponding Shannon limits. It can be seen from Figure~\ref{fig:BilayerBECBER} that the gaps in between are impressively small. \begin{figure}[htb] \vspace{-3.5mm} \centering \includegraphics[width=0.83\columnwidth]{Fig/BilayerBECBER.eps} \vspace{-3mm} \caption{Bit erasure rate of bilayer LDPC convolutional codes with different overall rates in BECs. The solid curves show the simulation results, and the dashed lines indicate the Shannon limits.} \vspace{-2.5mm} \label{fig:BilayerBECBER} \end{figure} To evaluate the proposed bilayer LDPC convolutional codes under more practical conditions, Figure~\ref{fig:BilayerAWGNBER} shows the bit-error-rate performance for the AWGN channel. For comparison purpose, we also include a regular bilayer LDPC block code. For both types of the codes, we set $l_1=3$, $l_2=2$, and $r=10$, which leads to approximately $R_{SR}=0.7$ and $R_{SD}=0.5$. In addition, the lengths of both codes are chosen in the way that the same hardware complexity \cite{PFSLZC08} is needed. It can be observed that the bilayer LDPC convolutional code clearly outperforms its block code counterpart. Signal-to-noise ratio (SNR) gains of $0.5$~dB and $1.3$~dB are obtained at the relay and at the destination, respectively. \begin{figure}[htb] \vspace{-2mm} \centering \includegraphics[width=0.83\columnwidth]{Fig/BilayerAWGNBER.eps} \vspace{-3mm} \caption{Comparison of bit error rate between a bilayer LDPC convolutional code and a regular bilayer LDPC block code in AWGN channels.} \vspace{-3mm} \label{fig:BilayerAWGNBER} \end{figure} \section{Conclusions} \label{sec:Conclusion} In this paper bilayer LDPC convolutional codes were proposed for three-node relay channels. For a binary erasure relay channel, we can find a bilayer LDPC convolutional code which is able to simultaneously achieve the capacities of the source-relay link and the source-destination link. Meanwhile, this code provides the highest possible transmission rate with decode-and-forward relaying. Moreover, the regular code structure significantly reduces the complexity by avoiding the optimization of irregular degree distributions. Numerical results were provided in both binary erasure channels and AWGN channels. In binary erasure channels, we can observe that the decoding thresholds are very close to the Shannon limits. In AWGN channels, a significant gain in terms of SNR is achieved compared with its block code counterpart. \small \bibliographystyle{IEEEbib}

Introduction {#s0005} ============ Allele-specific gene expression refers to the differential expression of the two alleles in a hybrid genome, such as gene imprinting phenomena in endosperm tissues in flowering plants and placental tissues in animals. The allele-specific expression of a gene is attributed to a distinct epigenetic status of its two parental alleles [@b0005; @b0010; @b0015]. Two types of epigenetically repressive marks were found to label the silenced allele of an allele-specifically expressed locus. The first is DNA methylation, which plays an important role in establishing the imprinting status prior to fertilization both in plants and animals. Many allele-specifically expressed genes were located in the differentially methylated regions (DMRs) in neighboring upstream and downstream cytosine-rich sequences of the two alleles [@b0020]. Histone H3 lysine 27 tri-methylation (H3K27me3) is another silencing mark that is associated with allele-specific expression, whose formation is mediated by Polycomb-group (PcG) proteins [@b0005; @b0025]. In *Arabidopsis*, members of the PcG complex contribute to allele-specific expression. The maternal PcG is able to autonomously silence the paternal alleles by adding H3K27me3 to the regulatory sequences of the unwanted paternal alleles, resulting in imprinting phenomena [@b0005; @b0025]. However, how PcG specifically recognizes one of the two alleles and whether other PcG-targeted genes are also subjected to allele-specific H3K27me3 is not known yet. Allele-specific expression may also involve activation of transposable elements (TEs) and uniparental expression of non-coding RNAs (ncRNAs) resulting from allele-specific hypo-methylation in endosperm [@b0030; @b0035; @b0040; @b0045; @b0050; @b0055]. Nonetheless, the exact underlying mechanisms remain elusive. Identification of allele-specifically expressed genes has relied on RNA-based technology to infer differentially expressed alleles based on single nucleotide polymorphisms (SNPs). However, this type of analysis has been hindered due to the lack of SNPs in many genes from parental genomes. The genome-wide discovery of imprinted genes has been carried out by performing array- and sequencing-based RNA profiling using SNPs to distinguish paternally or maternally derived transcripts [@b0060]. The array method is very imprecise, since one nucleotide difference on an array probe containing the SNP site is not sensitive enough to differentiate the probe hybridization signals to confirm allelic expression. Sequencing methods are more accurate, as they rely on directly calling the mismatches from reads. However, technical issues do exist. For example, reads are usually non-uniformly distributed in different portions of a gene, due to experimental bias in sample preparation and sequencing. Thus, multiple SNP points on the same gene often have varied coverage, which makes statistical evaluation quite inaccurate. Other disadvantages of RNA-based methods include: (1) because high-quality SNPs are usually located in coding regions, other types of genomic variation such as insertion and deletions (INDELs), intronic and UTR-related polymorphisms are usually not used for calling allelic expression; (2) for genes without SNPs, it is impossible to measure any allele-specific expression and (3) the intergenic regions where potential allele-specific epigenetic modification may exist will be missed by RNA analysis. To overcome these limitations in RNA-based methods, epigenetic modifications associated with allele-specifically expressed genes can be profiled to infer the molecular mechanisms underlying the allele-specific expression and may facilitate identification of more allele-specifically expressed genes. The feasibility of this type of analysis has been demonstrated in studying the association between allele-specific DNA methylation and imprinted gene expression by mapping the DNA methylome in maize along with mRNA expression levels [@b0065]. We propose a similar computational framework to conduct a combinatorial analysis of mRNA and epigenetic marks to improve the accuracy of identifying allele-specifically expressed genes in the seedling tissues of hybrid maize. This strategy identified allele-specifically expressed genes by simultaneously tracking allele-specific gene expression and epigenetic modification in maize seeding tissues in reciprocal F1 hybrids crossed from the B73 and Mo17 inbred lines. After differentiating the sequencing reads derived from B73 and Mo17 genomes, we determined the allele-specifically expressed genes by combining allele-specific epigenetic marks and RNA expression. The RNA-seq and ChIP-seq data used for developing the pipeline were obtained from an unpublished project that profiled DNA methylation, H3K27me3 and mRNA transcriptome in the seedling tissues of reciprocal F1 hybrids of maize. Methods {#s0010} ======= Macro-syntenic map construction {#s0015} ------------------------------- The Mo17 contigs were first partitioned into 2-kbp fragments, so that the large chunks of genomic DNA can be aligned by the BLAST-like alignment tool (BLAT) with small gaps allowed [@b0070]. The adjacently aligned fragments within 1 kbp were then merged as one syntenic block. Finally, various types of genomic variation were indexed, including SNPs and INDELs. Gaps longer than 100 bp were excluded, which may possibly result from the poor coverage of 454 sequencing of Mo17 genome. Read mapping, genomic position conversion and data visualization {#s0020} ---------------------------------------------------------------- The chromosome sequences of the B73 genome and the contig sequences of the Mo17 genome were first indexed by Bowtie 2 software. These sequences were prepared as two index databases in order to map the reads to the genomes [@b0075]. The mapping results in the SAM format were converted to BAM format using the SAMtools software [@b0080]. We then converted the chromosomal coordinates for the ChIP-seq and RNA-seq reads in hybrids mapped on the Mo17 contigs to the corresponding positions on the B73 genome using a customized Perl Script. To visualize the RNA-seq and ChIP-seq data in the UCSC genome browser, three output files containing the B73-unique reads, Mo17-unique reads and commonly mapped reads, respectively, were first converted into wiggle format files to represent base coverage in each 100 bp bin. Then, the generated wiggle files were uploaded to the UCSC genome browser installed on the local server for visualization (<http://genome.ucsc.edu/>). Chi-square test for detecting allele-specific gene expression and epigenetic modification {#s0025} ----------------------------------------------------------------------------------------- We used the Chi-square test to detect allele-specific gene expression patterns and epigenetic modification, with a slight modification of the regular Poisson distribution that introduced another parameter *λ* to represent average bias during sample preparation and sequencing. Let *X*′ represent the reads mapped on a gene territory with length *L*, the read counts in genes encoded by different loci are $x_{1}^{\prime}\text{,}x_{2}^{\prime}\text{,}x_{3}^{\prime}\text{,}\mspace{2mu}\ldots\text{,}\mspace{2mu} x_{L}^{\prime}$. Then, the overall *X*′ followed a generalized Poisson distribution:$$\mathit{\Pr}(X^{\prime} = x^{\prime}) = \left\{ \begin{array}{l} {\theta\left( {\theta + x^{\prime}\lambda} \right)^{x^{\prime} - 1}e^{- \theta - x^{\prime}\lambda}/x^{\prime}!} \\ {0\quad\text{for}\mspace{6mu} x^{\prime} > q\mspace{6mu}\text{if}\mspace{6mu}\lambda < 0} \\ \end{array} \right)$$ The optimal parameter *λ* was calculated by the maximum likelihood estimation using the Newton--Raphson equation $\sum_{i = 1}^{L}\frac{x_{i}^{\prime}(1 - x_{i}^{\prime})}{\overline{x^{\prime}} + (x_{i}^{\prime} - \overline{x^{\prime}})} = L\overline{x^{\prime}}$, where $\overline{x^{\prime}}$ was the average read count at different positions of a gene. *q* is an empirical parameter representing the minimum count of reads used in the test with the default value of *q* set as 20. The optimal parameter *θ* was calculated by $\widehat{\theta} = \overline{x^{\prime}}(1 - \widehat{\lambda})$. Results {#s0030} ======= Construct a macro-syntenic map of the B73 and Mo17 genomes {#s0035} ---------------------------------------------------------- The genome of maize inbred line B73 has been recently released with well-assembled chromosome sequences and annotated gene sets [@b0085]. The Mo17 genome was sequenced independently by Joint Genome Institute using 454 shotgun sequencing technology and 25 Gbps of sequencing reads were generated. These sequencing reads for Mo17 were *de novo* assembled to 528 Mbps non-redundant contig sequences (<http://www.phytozome.net>), which cover most of gene-encoding regions and sequences upstream or downstream of protein-coding genes. Thus, the sequence availability of both genomes facilitated identification of allele-specifically expressed genes using the proposed strategy. Based on the 528 Mbps Mo17 contigs and 2.5 Gbps B73 genome sequences, we developed a pipeline to build a macro-syntenic map for these two inbred lines (see Methods). The syntenic map was built to identify usable natural genomic variation and to convert the position of any read mapped on the Mo17 contigs to the corresponding position in the B73 genome. Shown in [Figure 1](#f0005){ref-type="fig"} is one exemplary syntenic block, fragments mapped in small pieces are usually located at the flanking regions or untranslated regions (UTRs) of genes, indicating that more genomic variations occur in regulatory regions or UTRs of genes than in coding regions. Our pipeline mapped 315,224 (63% of total 496,300) Mo17 contigs to B73 genome, and generated 46,099 syntenic blocks, which cover 18% (451 Mbps) of the B73 genome. We found that 35,244 genes (90.1%) in the filtered gene set (39,097 high-confidence genes in total) and 80,660 genes (73.5%) in the working gene set (109,626 genes in total without excluding pseudo-genes and putative TEs) partially overlapped with or were located entirely inside these 46,099 syntenic blocks. These data indicated that Mo17 genome is not completely covered; however, most high-confidence coding genes and non-repetitive regions can be covered for identification of imprinted genes. Details of all the identified genomic variations in B73 and Mo17 syntenic blocks were stored in plain text file. Based on the Bowtie 2 mapping score, the SNP information is used to differentiate the reads uniquely/preferentially mapped onto each genome, or commonly mapped onto both genomes. Correlating the mapped locations of reads with the syntenic blocks between B73 and Mo17 genomes {#s0040} ----------------------------------------------------------------------------------------------- After mapping short reads to the B73 and Mo17 genomes, we developed a script to correlate the mapped positions of reads in the B73 genome with their corresponding positions in the syntenic blocks. Such correlation information would be used to infer the allele-specifically mapped reads in the homologous regions of the two genomes. Subsequently, we developed another script to load the information from the syntenic database and the sequencing read databases in order to differentiate the reads in the two reciprocal F1 hybrids. Reads were classified into three groups: (1) reads that were only or preferentially mapped to the B73 genome (B73-unique reads); (2) reads that were only or preferentially mapped to Mo17 (Mo17-unique reads); and (3) reads mapped to both genomes with equal bowtie mapping score (common reads). Visualization of allele-specific gene expression and epigenetic modification patterns {#s0045} ------------------------------------------------------------------------------------- To facilitate the analysis of allele-specific epigenetic modification and gene expression patterns in hybrid maize, the counts of reads mapped on the two parental genomes were visualized in the UCSC browser (<http://genome.ucsc.edu/>). In the browser, we illustrated densities of three types of reads aforementioned for a specified genomic locus using B73 as the reference genome. While most RNA-seq reads were commonly mapped to the coding regions where nucleotide polymorphisms are fewer, the untranslated regions (5′ & 3′ UTRs) usually contained more B73-unique or Mo17-unique reads as expected ([Figure 2](#f0010){ref-type="fig"}). In the first case in [Figure 2](#f0010){ref-type="fig"}, there are more B73-unique RNA reads at the 5′ end of gene while there are more Mo17-unique reads at the 3′ end of gene in the two reciprocal crosses ([Figure 2](#f0010){ref-type="fig"}A). Thus, allele-specific expression of the B73 and Mo17 alleles can be inferred from the ratio of the amount of unique reads and common reads. A statistical test then can be used to estimate significance. In the second case in [Figure 2](#f0010){ref-type="fig"}, a miRNA gene, zma-MIR171n, contains more Mo17-unique H3K27me3 reads at the 5′ end, indicating it is unevenly modified on two alleles ([Figure 2](#f0010){ref-type="fig"}B). [Figure 2](#f0010){ref-type="fig"}C shows that two alleles may be equally modified by H3K27me3 since there were no significantly different inbred-unique reads detected on these two genes. [Figure 2](#f0010){ref-type="fig"}D shows an allele-specific DNA methylation at the 5′ end of another gene. However, these genes should not be regarded as imprinted, since no parent-of-origin effect was observed. The allele-specific pattern is due to the occurrence of natural sequence variations in these two genomes. In the case of endosperm, consistent detection of the inbred-unique reads, regardless of the direction of the cross, will facilitate the identification of imprinted genes. Moreover, these examples indicated that genomic variations in the intergenic regions can better differentiate the maternal and paternal reads, and they are more suitable for statistical evaluation of the differences in the reads mapped on the two parental alleles than SNPs in the coding regions alone. Identification of allele-specific expression and epigenetic modification based on Chi-square test {#s0050} ------------------------------------------------------------------------------------------------- To assess allele-specific gene expression and epigenetics, we first analyzed the RNA-seq data in two reciprocal F1 hybrids using the Chi-square test. The Chi-square test is a relatively simple method, which does not require distribution modeling or parameter estimation. In Chi-square test, ratio of the expression levels from the two alleles was assumed to be 1:1. This null hypothesis was also applied to histone modification and DNA methylation, since there was one copy of maternal DNA and one copy of paternal DNA in the genome of seedling tissues. However, uneven read distribution within a gene, caused by experimental bias in sample preparation and sequencing, would greatly affect the estimation of allele-specific patterns ([Figure 3](#f0015){ref-type="fig"}). Thus, systematic errors need to be minimized before the statistical test is performed. The Poisson distribution has been widely used to model the read count feature of RNA-seq and ChIP-seq data. Therefore, we performed a Poisson regression to estimate read coverage for the B73 and Mo17 alleles based on generalized Poisson distribution by considering the amount of B73- and Mo17-unique reads as well as common reads within an allele ([Figure 3](#f0015){ref-type="fig"}). We used the Poisson test to evaluate differences of read coverage between the B73 and Mo17 alleles, based on the null hypothesis that the ratio was 1:1 (maternal expression & modification versus paternal expression & modification). To correct the type I error derived from multiple testing, we employed the Sidák correction method to calculate the adjusted *P* value, *adjusted P* = 1 − (1 − *P~i~*)^1/^*^n^*, where *n* is the number of *P* values derived from multiple testing. Finally, genes and genomic regions with significant allele-specific patterns in terms of gene expression and epigenetic modification were selected using a *P* value of 0.01. Conclusion {#s0055} ========== Allele-specific expression is found ubiquitously in both plant and animal kingdoms, which has been thought largely attributable to associated *cis*-element variations and allele-specific epigenetic modification patterns. This interesting expression pattern was first observed in maize endosperm and then was expanded to maize meristem evaluated by massively parallel signature sequencing (MPSS). Allele-specific expression phenomenon was also reported in Arabidopsis, suggesting that this phenomenon is widely spread in plants. Allele-specific expression could result in allelic gene expression variation in hybrid and thus may have an impact on heterosis performance of F1 hyrbid. Therefore, profiling of transcriptome and epigenomic data from the hybrid plants that were generated from crosses of the subspecies or different inbred lines has been carried out in many plant species, in order to investigate the correlation between allele-specific gene expression and epigenetic modification and heterosis in crop plants. However, no mature bioinformatic pipelines and computational framework have been developed for this type of study. Using the Illumina high-throughput sequencing data of the mRNA transcriptome, DNA methylome and H3K27me3 modification in hybrid maize seedling tissue, we developed a computational procedure for detecting allele-specific gene expression and epigenetic modification. The first step was to generate a file storing a variety of natural genomic variations within the syntenic blocks identified in the two parental maize genomes. The second step was to generate a file storing the mapping information from the sequenced RNAs and epigenetic marks in F1 hybrids. The third step was to differentiate the reads mapped only/preferentially to B73 and Mo17 genomes, and commonly to both genomes, finally convert positions of reads on Mo17 to the reference B73 genome. The fourth step generated outcomes in the wiggle format, ready to be visualized in the UCSC genome browser. We also introduced a statistical method to measure the significance of allele-specific gene expression and epigenetic modification using the Chi-square test to accommodate the bias of the inbred-unique reads inherited from the two parental genomes. This computational framework can be applied to both plants and animal research. Authors' contributions {#s0060} ====================== XFW conceived the project. XXW performed the analysis. XFW and XXW wrote the manuscript. Both authors read and approved the final manuscript. Competing interests {#s0065} =================== The authors declare no conflict of interests. This work was supported by the National Natural Science Foundation of China (Grant No. 31260015), Natural Science Foundation of Qinghai Province (Grant No. 2012-Z-919Q), the Research Fund for Youth Project of Qinghai University (Grant No. 2011-QYY-1) and the State Key Laboratory for Agrobiotechnology of China Agricultural University (Grant No. 2012SKLAB06-5). Peer review under responsibility of Beijing Institute of Genomics, Chinese Academy of Sciences and Genetics Society of China. {#f0005} {#f0010} {#f0015}

\section{Introduction \label{sec:1}} Black holes are currently accepted as members of the bestiary of astronomical objects. Their success is explained on the one hand by the simplicity and elegance of their mathematical properties, and on the other hand by the (indirect) observation of the existence of very dense and intrinsically very dark distributions of matter in our universe. For instance, observations of the center of our own galaxy point to the existence of an object with such characteristics that the only known theoretical structure in general relativity (GR) that can be identified with it is a supermassive black hole \cite{Schoedel2002,Schoedel2003,Ghez2005}. However, it is noteworthy that we lack direct evidence of the most characteristic property of a black hole, namely its horizon (this may actually be impossible \cite{Abramowicz2002}; for a short discussion of this issue see Sec.~\ref{sec:6a}), or even of the very dynamical formation of black holes. The latter deficiency could eventually be alleviated in the future with the advent of gravitational wave astronomy \cite{Fryer2011}. Moreover, black holes contain odd features that suggest that their theoretical portrait is far from completely understood. The most prominent of these features is the unavoidable presence of a singularity, a region in spacetime in which the known laws of physics break down \cite{Hawking1976,Hawking1973}. It is expected that the successful combination of quantum mechanics and gravity will lead to the regularization of these singularities. In the absence of a full theory of quantum gravity or, in a wider sense, of an ultraviolet completion of GR, there has nonetheless emerged, especially after the work of Hawking \cite{Hawking1975,Hawking1976}, a consensus picture about the kind of ultraviolet modifications of the classical behavior one should expect. Naturally, the overall picture is far from being completely self-consistent (see for instance the information loss problem \cite{Hawking2005} or the recent firewall controversy \cite{Almheiri2013}), and surprises may arise as our knowledge about the high-energy properties of the gravitational interaction improve. This is the vision we align with in this paper. Guided by the discomfort with a ubiquitous property of the models that are currently in the market, namely the absurdly large lifetime of evaporating black holes, we elaborate on a completely different (and in some sense to be detailed, opposite) view on the possible ultraviolet effects on gravitational collapse processes, that was first discussed in \cite{Barcelo2011,Barcelo2014e,Barcelo2015}. Quite concisely, our geometric construction represents the transition from a black-hole geometry to a white-hole geometry in a short characteristic time scale. It is not our aim to provide here a first-principle justification of the overall proposal, though we shall include some thoughts on promising developments to consider in this direction. Our approach is heuristic, trying to elucidate what kind of effects have the potential of being eventually measurable, even if the evaluation of this potential could be at the moment rather theoretical. In a genuine bottom-up approach, we shall focus in the self-consistency of the picture and the implications that derive from it, for both the high-energy theoretical aspects and the low-energy observational features. This paper complements our previous discussions, presenting a more detailed account of some of the motivations behind this model and an in-depth exploration of its main geometric and physical features. The body of the paper is divided in five main sections. In Secs. \ref{sec:2} and \ref{sec:3} we review, respectively, some of the basic features of gravitational collapse of massive stars and properties of black holes in GR, and the quantum-mechanical modifications of some of these properties that are nowadays contemplated from a theoretical perspective. In Sec. \ref{sec:4} we look at the different possible ultraviolet modifications from a phenomenological perspective. In Sec. \ref{sec:5} we discuss our proposal in detail, presenting a step-by-step construction of a family of geometries describing the process we want to describe. In Sec. \ref{sec:6} we go through the possible observational implications of the proposal, that may be of relevance if these processes are indeed realized in nature. We finish with a brief conclusions section. \section{Black holes and spacetime singularities: the standard GR picture \label{sec:2}} \subsection{Event horizons \label{sec:2a}} We shall concentrate on the simplest models of gravitational collapse, which are spherically symmetric. Eventually we will consider the effect of introducing small, non-spherical perturbations in this simple picture. As is well known, in the spherically symmetric case, and for a star with mass $M$ whose radius is larger than its Schwarzschild radius, \begin{equation} r_{\rm s}=\frac{2GM}{c^2}, \end{equation} the external metric is unique and is given precisely by the Schwarzschild solution. The Schwarzschild metric has to be glued with the internal metric of the matter distribution along the surface of the star. A stable stellar structure has to be in hydrostatic equilibrium described by means of the Tolman-Oppenheimer-Volkoff equation \cite{Oppenheimer1939}. On the basis of this equation and known equations of state for the matter inside the star, one concludes that there exists an upper bound to the mass of stable configurations \cite{Oppenheimer1939,Rhoades1974,Bombaci1996}. Accordingly, any star in equilibrium will become unstable if accreting more than a certain amount of matter. In those extreme situations, the strength of the gravitational interaction surpasses that of any other possible known force in the system, provoking an indefinite contraction of the stellar structure. The resulting trajectory of the surface of the collapsing star can be described by means of a function $R(\tau)$, where $\tau$ is the proper time of an observer attached to the surface. If the regime in which the strength of the gravitational interaction surpasses that of any other relevant force in the system is reached, this surface will inevitably cross the Schwarzschild radius in finite time. In terms of equations, for a given initial stellar radius at $\tau=\tau_{\rm i}$ there always exists a finite proper time interval $\Delta\tau$ so that \begin{equation} R(\tau_{\rm i}+\Delta\tau)=r_{\rm s}. \end{equation} A prime example of this behavior is the Oppenheimer-Snyder model of gravitational collapse \cite{Oppenheimer1939b}, in which the perfect fluid representing the matter content is pressureless, so that gravity is strictly the only force present in the model. The Oppenheimer-Snyder model is thus considered to provide a reliable description of the late stages of the gravitational collapse process for massive enough distributions of matter with small asphericities. Upon crossing, a trapping horizon is formed in the position that corresponds to the Schwarzschild radius. While we leave to the next section the rigorous definition of such a notion, for the moment it is enough to have in mind the intuitive picture that this trapping horizon \emph{locally} forbids that signals originated at the surface of the star reach external observers. This can be seen explicitly if we take the black-hole patch of the Kruskal manifold and use suitable coordinates, for instance Painlev\'e-Gullstrand coordinates (see \cite{Martel2001} and references therein). It is important to keep in mind that this observation by itself does not imply that lumps of matter will never be able to cross outwards the radial position at which the trapping horizon was first formed. The overall dynamical evolution of spacetime around the outgoing distribution of matter should be taken into account in order to consistently conclude so. As we will see, restrictions to the geometry of spacetime in the form of energy conditions are useful to tackle this issue. We leave this discussion for the following section, assuming for the moment the standard picture that horizons are inviolable in classical GR. The resulting object, a so-called black hole, would then possess as defining characteristic an event horizon. They would be absolutely inert objects, the ultimate end point of gravitational collapse, the dead state of stellar physics. No matter content can ever escape from the inside, as the region behind the Schwarzschild radius is causally disconnected from the external observers. Black holes are characterized by just three numbers, mass $M$, electric charge $Q$ and angular momentum $J$, for generic initial conditions (even if not spherically symmetric) for the matter that went to form the black hole; see \cite{Chrusciel2012} for a thorough technical discussion. This simplicity is an appealing feature for a large number of theorists. In our idealized setting we are demanding spherical symmetry ($J=0$) and electrical neutrality ($Q=0$), leaving us with just one relevant parameter. \subsection{Singularities \label{sec:2b}} According to classical GR, the fate of matter after crossing the trapping horizon is ominous. While the horizon is characterized by a pronounced deformation of the light cones so that (locally) the light that is emitted at the horizon cannot escape, light cones just behind the horizon are deformed even more drastically. The resulting deformation ultimately implies that every light-like trajectory starting inside the horizon must (locally) go inwards, i.e., with the value of the radial coordinate decreasing, until ``hitting'' the singularity. Let us be more precise on these notions, which not only will permit us to make sharp statements, but also understand the generality of this picture. In principle one could think (and this was so historically \cite{Penrose2010}) that the occurrence of the singularity sketched above may be an artifact of the high degree of symmetry of the solution being used. Quite the opposite, these features are completely generic, as was first understood by means of the so-called Penrose singularity theorem. This was the first of a sequence of results, usually known as Hawking-Penrose theorems (see \cite{Senovilla2014} for a recent account of the history of the subject and later developments up to date). These theorems are formulated on spacetime manifolds that possess a non-compact Cauchy hypersurface, i.e., the topology of the manifold is $\mathbb{R}\times\Sigma$ with $\Sigma$ being a non-compact spacelike tridimensional manifold. Roughly speaking, this means that the geometry does not incorporate new regions that require additional data for their description, and that there are well-defined notions of future and past. The first notion we need concerns the ``surface'' of the black hole, that is, the two-dimensional manifold defined by the equation $r=r_{\rm s}$ at a given moment of time (being the Schwarzschild metric independent of time, this surface is stationary under its flow). We shall use the conventions and notation of \cite{Senovilla2014}. Let us imagine that we are emitting a spherical wavefront of light rays from a given position $r> r_{\rm s}$. Then light rays pointing outwards will move outwards, meaning that the radius of the corresponding sphere of light will grow in time, while light rays pointing inwards will move inwards, with decreasing radius in time. This is no longer true when we enter the Schwarzschild radius: for $r< r_{\rm s}$ both spheres of light move inwards. These surfaces are thus a particular case of the general concept of a closed future-trapped surface: a closed surface whose \emph{future-directed} null geodesics do not flow outwards. A closed past-trapped surface would correspond to the situation in which \emph{past-directed} null geodesics do not flow outwards. In the spherically symmetric case one can consider these surfaces as spherical, filling up the black hole as an onion. In this case the $r=r_{\rm s}$ surface marks the boundary between this trapping behavior and the normal behavior for $r>r_{\rm s}$, and represents the so-called trapping horizon, or marginally future-trapped surface.\footnote{In the Schwarzschild solution the event and trapping horizons coincide. This is not generally true for dynamical spacetimes, being the simplest example a small, non-spherical perturbation of the Schwarzschild solution \cite{Hawking1973}.} The second notion is that of a singularity. This is a slippery concept as, strictly speaking, singularities do not belong to spacetime. We can however characterize the singular behavior by means of particles traveling to or from the singularity or, in geometric terms, of geodesics which finish or start at the singularity. In general, curves are maps from a subset of the real line $\mathbb{R}$ onto the spacetime manifold. When all the geodesics in a manifold are defined for the entire real line $\mathbb{R}$ when parametrized in terms of their affine parameter, the manifold is said to be geodesically complete, being incomplete in the opposite. The usual definition is that a spacetime manifold is singular if and only if it is geodesically incomplete.\footnote{Most definitions also include the condition of being non-extendable \cite{Hawking1973}, but this technical point is not really relevant for our purposes.} From a physical perspective this would imply the existence of particles or observers that disappear (or materialize) abruptly, which clearly corresponds to an abhorrent behavior. It is important to keep in mind that, for a manifold to be geodesically incomplete, it is enough that there exists just one incomplete geodesic. The last ingredient is a condition on the curvature of the manifold. This crucial geometrical condition for the theorem is motivated by its translation, through the use of the Einstein field equations, into a condition upon the matter content. The so-called null convergence condition holds by definition if, for every null vector field $u^a$ in the manifold, \begin{equation} \mathscr{R}_{ab}u^au^b\geq0\label{eq:null1} \end{equation} with $\mathscr{R}_{ab}$ is the Ricci curvature tensor. If we apply the Einstein field equations, \begin{equation} G_{ab}=\mathscr{R}_{ab}-\frac{1}{2}\mathscr{R}g_{ab}=\frac{8\pi G}{c^4}T_{ab},\label{eq:einsteineqs} \end{equation} then Eq. \eqref{eq:null1} implies the null energy condition for the matter stress-energy tensor $T_{ab}$: \begin{equation} T_{ab}u^au^b\geq0.\label{eq:null2} \end{equation} With these ingredients, the Penrose singularity theorem is formulated as follows \cite{Penrose1965}: given a spacetime containing a non-compact Cauchy hypersurface $\Sigma$ and a closed future-trapped surface, the convergence condition \eqref{eq:null1} being valid for all the null vectors $u^a$ implies that there exist null geodesics that are incomplete in the future. In all its generality, this theorem does not provide us with any clue about the behavior of the spacetime near the singularity. To put an example, it does not assert that there is some curvature invariant (a scalar made up from the metric and its derivatives only) which blows up in the singularity, or that an entire set of observers in a given region (e.g., the interior of the future-trapped surface) will certainly experience the singular behavior in the future. Developing new techniques is necessary in order to tackle this thorny problem and understand in fine detail the physical properties of singularities. Among the attempts of doing so in the framework of the classical theory, the so-called BKL conjecture \cite{Khalatnikov1970,Belinskii1971}, contemporaneous to the singularity theorem stated above, occupies a prominent place. While in ordinary situations gravity is the weakest of all the forces in nature, singularities are the place in which gravity becomes the most prominent actor. Following this idea, in this conjecture it is assumed that matter fields do not play any essential role near the singularity. Moreover, it is postulated that the temporal derivatives of the metric field are the dominant ones, so that spatial derivatives can be ignored near the singularity (which is assumed to be spacelike). Every spatial point of spacetime will then become disconnected from the others, and will evolve following ordinary differential equations that present chaotic behavior. Numerical studies have shown evidence in favor of this behavior; see \cite{Berger2002,Ashtekar2011} for modern discussions. \section{Black holes and spacetime singularities: ultraviolet effects beyond GR\label{sec:3}} \subsection{Trapping horizons \label{sec:3a}} When trying to build coherent scenarios in which the collapse of matter is affected by the very quantum nature of matter, the classical picture significantly changes. Robust semiclassical calculations tell us that black holes cannot be absolutely stationary, as they evaporate through the emission of Hawking quanta~\cite{Hawking1975,Unruh1976}. This is the last piece composing the so-called black-hole thermodynamics: a black hole of mass $M$ has a temperature \begin{equation} T_{\rm H}=\frac{\hbar c^3}{8\pi G M k_{\rm B}}. \end{equation} Given this result, it is assumed that the classical spacetime representing the collapse of a star to form a black hole should be modified so that the event horizon, instead of setting at a stationary position (once the absorption of surrounding matter and its stabilization has ended up), would shrink up to eventually disappear in a final explosion~\cite{Hawking1974}. It is also clear that the quantum effects responsible of the evaporation of the horizon would entail even more drastic modifications in the region surrounding the singularity. These realizations inevitably trigger the questions: Are these evaporating black holes really black holes in the sense of having an event horizon? Is there some information loss in a complete evaporation process? These questions have been a matter of controversy, and a strong driving force for theoretical development, since the publication of Ref. \cite{Hawking1976}. However, nowadays even Hawking concedes \cite{Hawking2014} that the most reasonable solution is that no event horizon would ever form, but only a structure which is locally similar but that lives for a finite, albeit extremely long, amount of time; that is, a trapping horizon as defined in Sec. \ref{sec:2b}. Under this view black holes, understood as causally disconnected regions of spacetime, would not strictly exist in nature. However, owing to the similarity of these quantum-corrected objects with classical black holes, they usually keep their name. Here we will generically call them regularly evaporating black holes (REBHs) to distinguish them from their classical cousins. \subsection{Preventing singularities \label{sec:3b}} As happens with the event horizon the introduction of quantum mechanics, or any other kind of ultraviolet modifications, is likely to bring important changes to the classical picture concerning the singular behavior of GR. The first indication of these changes is that all kinds of energy conditions, that play a central role in the singularity theorems, are violated by quantum effects \cite{Barcelo2002,Ford2003}. Indeed, from a qualitative perspective it is widely expected that any consistent ultraviolet completion of GR will regularize the singular classical behavior. One of the ways to do so is the occurrence of repulsive forces preventing the gravitational collapse process to proceed indefinitely.\footnote{It would not be the first known example of this phenomenon: for example, the electron degeneracy pressure is a genuinely quantum effect that stabilizes white dwarfs \cite{Fowler1926,Chandrasekhar1931}.} Given their repulsive character, these effective forces will generally violate the energy conditions and, eventually, would unlock new options for the fate of gravitational collapse. These qualitative expectations have been embedded in mathematical frameworks of different nature, essentially since the first explorations in semiclassical gravity \cite{Parker1973}, and most frequently in the field of cosmology \cite{Novello2008}. One of the most popular implementations, probably because of its fundamental flavor, is the one that results from the application of the loop quantum gravity techniques (see \cite{Ashtekar2005b} for an introduction to the subject). Loop quantum cosmology (see, e.g., \cite{Banerjee2012}) is the result of applying non-perturbative quantization techniques borrowed from the general theory and applied to some highly symmetric cosmological spacetimes. Although the results of this procedure have to be taken with a grain of salt \cite{Kuchar1989,Barbero2010}, these models are regarded as valuable tools in understanding the implications of the wider quantization program of the canonical structure of GR. One of the robust results of this approximation is that near the cosmological singularities, there appear effective forces with a net repulsive effect. These forces are strong enough to overcome the fatal attraction of gravity which would otherwise engender a singularity, provoking the \emph{bounce} of the matter distribution and connecting a classical contracting cosmology with a classical expanding cosmology \cite{Bojowald2001}; the Big Bang event is identified with the moment of the bounce, so that the expanding branch corresponds to our present universe. The bounce generally takes place when the density of matter is of the order of the Planck density, \begin{equation} \rho_{\rm c}\sim\rho_{\rm P}=\frac{m_{\rm P}}{\ell_{\rm P}^3}=\frac{c^5}{\hbar G^2}\simeq 5\times 10^{96}\mbox{ kg m}^{-3}.\label{eq:planckden} \end{equation} Following this result, quantum effects would be triggered when the radius of stellar structure is several orders of magnitude greater than the Planck length; for a star of mass $M$, its minimum radius $R_0$ will be approximately given by \begin{equation} R_0\sim\ell_{\rm P}\left(\frac{M}{m_{\rm P}}\right)^{1/3}.\label{eq:minrad} \end{equation} Although this may seem surprising at first, given that quantum gravity effects are usually bound to the Planck length, this very association is only reasonable in the absence of matter, that is, in the pure gravity case. Let us stress that the order of magnitude \eqref{eq:minrad} is the natural one on the basis of the Einstein field equations and the expectation that quantum gravity effects should appear when the curvature is of the order of $\ell_{\rm P}^{-2}$ \emph{in the presence of matter}, being independent of the specific ultraviolet completion one pursues for the gravitational interaction. Actually, for dust matter content of Planck density the trace of the Einstein field equations \eqref{eq:einsteineqs} leads to the very same order the magnitude: \begin{equation} \mathscr{R}=g^{ab}\mathscr{R}_{ab}=-\frac{8\pi G}{c^4}g^{ab}T_{ab}=\frac{8\pi G\rho_{\rm P}}{c^2}=\frac{8\pi}{\ell_{\rm P}^2}.\label{eq:planckcur} \end{equation} Thus in the presence of matter the association between Planckian densities and ultraviolet effects is completely straightforward, independently of the specific ultraviolet completion one considers. In loop quantum cosmology, the bouncing behavior is successfully captured by an affective Friedmann equation \cite{Taveras2008} which, in the flat ($k=0$) case, takes the form \begin{equation} \left(\frac{\dot{a}}{a}\right)^2=\frac{8\pi G}{3}\rho\left(1-\frac{\rho}{\rho_{\rm c}}\right).\label{eq:frimod} \end{equation} In this equation it is easy to see that reaching the critical density corresponds to a turning point, $\dot{a}=0$. The second term in the right-hand side of this equation corresponds to an effective stress-energy tensor that strongly modifies the classical behavior. It will be instructive to have in mind this feature in our later discussion of gravitational collapse. The occurrence of effective repulsive forces will probably be a robust ultraviolet effect, independently of the specific quantization program or ultraviolet completion one choses (see, e.g., \cite{Barcelo2011,Parker1973,Broda2011,Abedi2015} for different suggestions), leading eventually to a resolution of the singular behavior of the classical theory. However, while the cosmological situation is quite well understood nowadays, how these considerations would affect the issue of gravitational collapse is a more subtle issue, as we discuss in the following. \section{Ultraviolet effects: phenomenological considerations\label{sec:4}} \subsection{Horizon predominance \label{sec:4a}} The necessity of finding an ultraviolet completion of GR has been in the air at least since the first successes of quantum field theory, in particular when applied to electrodynamics. The generalizations of these field-theoretical developments to GR were plagued with difficulties, thus forcing to use partial, and therefore necessarily incomplete, approaches. These have been used since the sixties to get insights into the ultraviolet behavior of gravitational collapse processes. Before Hawking's groundbreaking developments, there already were some interesting attempts to see what could result from the incorporation of quantum-mechanical effects to a black hole spacetime. As early as 1966, Sakharov~\cite{Sakharov1966} and Gliner~\cite{Gliner1966} suggested that at the high densities close to singularity formation the effective matter content might develop a vacuum-like equation of state $\rho=-p$, providing a repulsive force. Soon after, Bardeen showed that it is possible to construct black hole geometries (here meaning having an event horizon) satisfying the null energy condition but having no singularities~(\cite{Bardeen1968}, see also~\cite{Borde1994}). Singularity avoidance (and hence a way out of the singularity theorems) is permitted due the non-existence of an open Cauchy surface: the spacetime develops a topology change from open hypersurfaces to closed ones~\cite{Borde1997}. Matter crossing the black-hole horizon and falling towards the apparent singularity would end up reappearing from a white-hole horizon into a \emph{different universe}, i.e., not in the same asymptotic region. Bardeen's spacetime and others with similar structure (e.g.~\cite{Dymnikova1992,Mars1996,Ayon-Beato1998,Bronnikov2001,Bronnikov2001b,Hayward2006,Olmo2014,Olmo2015}) are examples of regularizations of the black hole singularity that do not affect the event horizon. Hawking's idea that event horizons should evaporate, leading eventually to some fundamental loss of information, introduced a new ingredient into the singularity problem. Many different evaporation scenarios have been put forward since then to accommodate solutions to the information and singularity problems. These proposals have generated a large amount of work; see for instance the sample of recent papers \cite{Ashtekar2006,Ashtekar2008,Hossenfelder2010,Kawai2013,Kawai2014,Torres2014,Frolov2014,Bojowald2014,Gambini2014,Torres2015,Rovelli2014,Barrau2014b,Bambi2013,Liu2014,Zhang2015,Modak2015,Modak2015b,Hawking2015,Hooft2015}. What we want to emphasize here is that, at least qualitatively, the resulting effective geometries advocated by most of these scenarios (what we have called REBHs) are essentially equivalent. Diagrams describing these geometries can be found for example in~\cite{Frolov1979,Frolov1981,Roman1983,Ashtekar2005,Hayward2006,Rovelli2014,Bambi2014,Bardeen2014,Kawai2015}. Aside from their fine theoretical details, all of them contemplate REBHs as essentially hollow and almost stationary regions of spacetime; all of them share the substitution of event horizons by \emph{extremely slowly} evaporating trapping horizons as seen by external observers, even if the specific time scales could vary between different models. In this sense these proposals may be regarded as ``conservative''. Thus, the prevalent view is that REBHs indeed form in astrophysical scenarios and, ignoring their gravitational interaction with the surrounding matter, they would remain almost inert for very many Hubble times except for a tiny evaporative effect that would eventually make them disappear. The evaporation rate of stellar-mass objects would be so slow, $10^{67}$ years to halve the size of a Solar mass object, that for all practical purposes they could be considered stationary. REBHs have a ridiculously large lifetime, in whatever measure: the estimated age of the universe is of the order of $10^{10}\mbox{ years}$. Whereas the distinction between a classical BH and a REBH is of fundamental interest on purely theoretical grounds, it is almost certainly irrelevant for all astrophysical purposes, and arguably for any meaningful operational perspective \cite{Visser2014}. This becomes even worse when considering that astrophysical black holes are actually on average growing and not yet evaporating because their Hawking temperature is smaller than the approximately 3K of the cosmic microwave background \cite{Samtleben2007}. This situation could be partially alleviated if primordial black holes were generated in our universe, as it was already proposed by Hawking~\cite{Hawking1974}; see~\cite{Carr2010} for some recent constraints on their existence. In this paper we shall focus on contemporary gravitational collapse processes, which are certainly the most interesting ones from an experimental perspective. Given the irrelevance of Hawking radiation for astrophysical purposes, this implies that any experimental test, whose result is available to observers outside black holes, of the precise way in which the evaporation proceeds would be almost certainly beyond the reach of humankind. Most importantly, the very nature of any kind of process taking place in the interior of the trapping horizon, and in particular the fate of the matter inside it and the ultraviolet regularization of the singularity, will be inevitably hidden for us unless something radical changes this picture. While observers that fall into the black hole and cross the trapping horizon will be able to catch a glimpse of what is going on inside, they will not be able to communicate their experience to the exterior, enforcing the ignorance of external observers.\footnote{We encourage the reader to note the amusing analogy with the existence of life after death.} For the standard REBH perspective the only possible, albeit partial, experimental corroboration of the overall scenario would be the real detection of Hawking radiation.\footnote{Even this assertion is debatable, as stellar objects hovering near its Schwarzschild radius could produce a radiation signal very close to the predicted Hawking radiation \cite{Stephens1994,Barcelo2006,Barbado2011}.} In the real world the systems hosting black holes are so complex that there exist multitude of forms of radiation that would eventually frustrate any attempt to measure such a tiny effect.\footnote{A great deal of effort has been put on the detection of the analogue of the Hawking effect in the analogue gravity framework \cite{Barcelo2005,Steinhauer2014,Steinhauer2015}. It is, however, unclear to what extent the analogy is deep enough to consider the measure of these effects on fluids a detection of the `genuine' gravitational Hawking radiation; some arguments in favor can be read in \cite{Unruh2014}.} On the other hand, given the ridiculously long times associated with the evaporation process, any event happening inside the trapping horizons formed in gravitational collapse processes is certainly irrelevant for practical purposes. Any attempt to understand the physics behind the horizon, for instance on the lines sketched in Sec. \ref{sec:3b}, would be unquestionably pointless: while the fate of matter behind the trapping horizon and the horizon itself could prove an interesting intellectual exercise, it would be impossible to find any experimental corroboration of these developments. The black hole trapping horizon will effectively act as an event horizon, forcing our ignorance about what is behind. This observation is not a matter of improving the sensitivity of experiments, but the issue is a completely different (and for us, quite unpleasant) one. One could argue that, even accepting that the gravitational collapse process itself could be useless to boost our understanding of the physics near singularities, the Big Bang singularity could be used as a substitute from which to obtain this knowledge. The first observation that comes to mind concerning this assertion is the issue of repeatability: while we expect that numerous processes of gravitational collapse are taking place now around us in the universe, and will be quite surely taking place in the future, in the currently accepted model of the universe there is only one Big Bang event. Also the Big Bang singularity lies in the past so that, while we can hopefully describe its properties in a simple way, it is questionable to what extent we are able to perform experiments in the usual sense of the word, or if the most we can hope to do is constructing some sort of cosmological ``archeology''. Leaving aside these issues, which are indeed ubiquitous to the field of cosmology (see \cite{Goenner2010} for instance), there are good physical reasons to believe that the singularities associated with black holes should be different from the initial cosmological singularity. These arguments are based on the second law of thermodynamics and have been repeatedly exposed by Penrose \cite{Penrose2010}. While one may think that these singularities correspond from a mechanical point of view as the time-reversal of each other, thermodynamical considerations break this apparent time reversal. On the one hand, if the second law of thermodynamics is to be applicable to the universe, the behavior near the initial singularity is to be associated with a low amount of entropy. On the other hand, the process of gravitational collapse to black holes is generally expected to lead to high entropies. In particular, this should indeed be the case in the standard scenario, in order to fulfill the thermodynamic description of black holes and match the Bekenstein-Hawking entropy formula \begin{equation} S=\frac{c^3 A}{4 G\hbar}, \end{equation} where $A$ is the area of the horizon. This points to a deep physical distinction between these situations that supersedes the apparent time reversal symmetry. To end this section let us say a few words about other popular and somewhat more exotic approaches to the information and singularity problems. Our comments here will be restricted to their relation with the lifetime problem which is central for this work. The complementarity approach of Susskind~\cite{Susskind1993} would share the REBH geometric description but maintaining that it would be only relevant for observers actually crossing the horizon. A complete experimental proof of the overall complementarity scenario would therefore face the same practical problems than more standard REBHs scenarios. The fuzzball proposal of Mathur~\cite{Mathur2005,Mathur2009} takes the view that the evaporating horizon marks a real border for the spacetime. Information is accumulated at the gravitational radius being retrieved only after enormous amounts of time. The same situation occurs in the condensed state scenario of Dvali~\cite{Dvali2014,Dvali2015}. Let us recall that the Page time (half-mass evaporation, the assumed time of retrieval of the information in these scenarios) for an astrophysical black hole is as extremely long as the time scale associated with Hawking evaporation. The Giddings remnant scenario~\cite{Giddings1992,Giddings1994} adds to the REBHs the presence of massive remnants. Only the last stages of the REBHs picture are expected to be modified so that this scenario also shares the lifetime problem. Let us note for completeness that there is a recent construction by Haggard and Rovelli~\cite{Haggard2014} that shares some geometric properties with our proposal, as it also describes the black to white hole transition with only one asymptotic region. However, it is completely different in essence: by construction, the duration of the total transition as seen from external observers is extremely long. Therefore this scenario is indistinguishable for an extremely long time from any REBH model, being therefore not better regarding the lifetime problem. \subsection{Preponderance of the singularity regularization\label{sec:4b}} The possibility of measuring the ultraviolet effects that are responsible for singularity avoidance in reasonable time scales would imply that Hawking radiation would not be able to carry away from the object a significant amount of energy in the meanwhile. Indeed, the horizon barrier should be broken in some way so that some signals from the inside may reach an external observer and, as the evaporation process is extremely slow, the net evaporative effect would be negligible. Even if we wait the entire age of the universe (specifically, the Hubble time $t_{\rm H}$), the energy loss of a stellar-mass black hole would be of the order of \begin{equation} 10^{-4}\,\frac{\hbar c^6}{G^2}\frac{t_{\rm H}}{M^2_\odot}\sim10^{-11}\mbox{ J}. \end{equation} This represents a $10^{-49}$ part of its original rest energy, which is absurdly negligible. Again, from an operational perspective only part of these theoretical developments associated with gravitational collapse could be relevant. We end up with a dilemma: if the physics of horizons as long-lived thermodynamical entities is to be realized in nature, there is very little hope that we will be able to observe any trace of the physics associated with the corresponding (regulated) singularities while, if the physics of these spacetime singularities is to be observable in reasonable time scales, any dissipation through the emission of Hawking radiation would be physically irrelevant in energetic terms. An additional tension, of a different kind but related to the previous one, appears when ultraviolet effects inside the collapsing distribution of matter are taken into account. As a first approximation let us consider the homogeneous and isotropic situation inside the star, so that we can use a patch of a cosmological solution to describe the internal geometry. In the pressureless case the surface of the star (as well as any point in the interior) will follow a geodesic in the external metric, i.e., the Schwarzschild metric. Any nonzero pressure will lead to deviations from this behavior. As an extreme example one could imagine a strong enough pressure to halt the collapse before the event horizon is formed; the radius of the star will then reach a fixed value. While this is what happens for structures below the Tolman-Oppenheimer-Volkoff limit of neutron stars \cite{Oppenheimer1939,Rhoades1974,Bombaci1996}, for larger masses there is no known force which can compete with the gravitational contraction \cite{Woosley2002}. It seems unavoidable that, in these situations, the matter distribution crosses its Schwarzschild radius, so that we can even take the idealized Oppenheimer-Snyder model, neglecting any known form of pressure with respect to the gravitational force. Deviations from GR would not appear until high enough densities are reached, and certainly \emph{not} at the moment when the star generates its trapping surface (at least if the collapse proceeds in a free-fall manner; departures from this behavior could lead to strong deviations even before of horizon crossing \cite{Barcelo2008}). As we discussed above, if singularities are to be avoided in a suitable ultraviolet completion, a universal effective pressure with a net repulsive effect should appear at some point. This pressure would prevent the formation of singularities, leading to bouncing solutions of similar nature as the ones that were previously described in a cosmological framework. How is this compatible with the geometry outside the star? The bouncing process of the matter distribution implies that the radius of the star should grow in proper time for an observer in the surface. But this is in contradiction with the external metric: there are no timelike geodesics of non-decreasing radius inside the Schwarzschild radius. In order to avoid a causality violation we should accept a modification of the external metric around the bouncing star. Given the spherical symmetry we are assuming, these geometry deviations should take the form of an effective matter content violating energy conditions. In other words, these general arguments point to the necessary existence of modifications of the geometry even outside the star. These modifications are generally expected to be localized in a small region around the minimum radius of the star $R_0$ in which the curvature invariants associated with the classical geometry are still large enough. The situation is, in our opinion, more subtle. Once we have accepted that a bounce of the star should occur near the singularity, both geometric pictures inside and outside the radius of the star seem to be hardly reconcilable. More specifically, gluing these geometries along the wordline (for fixed angular variables) of the surface of the star no longer appears as a viable option, as these present what we can regard as competitive trends. While the external geometry will try to stop the matter distribution to escape, the internal geometry of matter will try to overcome the overwhelming horizon influence. It seems that one should rather face the problem of how these effects compete in order to have a glance to the final outcome. In the most general case the resulting geometry will present a transition layer between these two regimes. How to describe the properties of this layer seems to be currently unknown. In the standard view of REBHs, the internal geometry is completely dominated by the external geometry, so that the transition layer affects the behavior of the overall internal metric. This would correspond, e.g., to the effective stabilization of the bouncing star at some asymptotic radius \emph{inside} the horizon. Such a situation corresponds to one of the extremes of the entire family of possible interpolating geometries. The situation we want to analyze here is the complementary extreme, with the internal geometry unchanged and the external metric modified in order to guarantee a smooth matching. As we will discuss, in this case the transient may be analogous to a shock wave phenomenon as those occurring in normal fluids. From a theoretical perspective, the fine knowledge of the resulting geometry will probably have to wait until our understanding of the ultraviolet properties of the gravitational interaction is deep and firm enough to tackle these complex questions. Although ultraviolet effects are expected to be triggered only when curvatures are high enough (i.e., Planckian), we have virtually no knowledge of how these regions of high-curvature, that violate energy conditions, evolve once they are generated, how they propagate and decay in time. While a first-principles study of this issue is highly interesting, we can always revert the logic and try a phenomenological approach. We can study the possible effective geometries and try to extract physical and observational consequences, information that could be eventually used in order to progress in our theoretical understanding. \section{Effective bounces, black to white hole transitions and shock waves \label{sec:5}} In plain terms, what we want to analyze is what would happen if the spherically-symmetric collapsing matter, upon reaching Planck density, slowed down and bounced back. In geometric terms we expect the description near the matter crossing the horizon outwards to correspond with the time-reversed geometry to that of a black hole, namely the one associated with a white hole. Birkhoff's theorem is certainly a uniqueness result for the vacuum geometry outside a spherical distribution of matter with mass parameter $M$ and radius greater than its Schwarzschild radius. However, for vacuum geometries inside the Schwarzschild radius this is not true. Birkhoff's theorem asserts that any vacuum patch must be locally isometric to a patch of the maximally extended Kruskal manifold~\cite{Hawking1973}. The vacuum geometry outside a spherical distribution of matter but still inside the Schwarzschild radius could therefore either correspond to the black-hole or the white-hole patch. From this perspective, the bounce can be represented by a transition between these two patches which will necessarily contain features that go beyond GR. It is an essential feature that the flat asymptotic regions corresponding to these geometries are one and the same. The bounce is a physical event taking place entirely in our universe; it does not describe the escape of matter to remote regions that are causally disconnected from the static observers standing outside the stellar structure (as discussed in Sec. \ref{sec:4a}, proposals of this kind fall again within the REBH paradigm). As we will see later, the most general metric describing this bounce contains a number of unknown parameters, so that its fine details certainly depend on the underlying ultraviolet completion of GR. However, it also possesses some robust characteristics that clearly distinguishes it from REBH proposals even from a physical perspective. Let us proceed in a constructive way, and build first a simple toy-model geometry of a time-symmetric bouncing regularization of a collapsing star, that however encapsulates the most relevant features of the process we want to describe. While this geometry will present some singular properties, it can be considered as the distributional limit of well-behaved geometries, thus providing indeed a very good analytical approximation. We shall describe the bouncing geometry from an explicitly time-symmetric point of view. As stated above, for simplicity we are concentrating on the spherically symmetric collapse, with the collapsing matter being characterized as a pressureless dust cloud. Also we will ignore dissipation effects in a first stance: classically, a spherically-symmetric collapse does not produce any dissipation in the form of gravitational radiation. If we took into account quantum corrections, there would indeed exist some (rather small) dissipation. In the limit of very large masses, this quantum radiation could in principle be made as small as desired. The time-symmetric geometry we are going to present is a reasonable dissipationless approximation to more realistic, dissipative situations, that we explore in Sec. \ref{sec:6}. It will prove useful to describe the metric using generalized Painlev\'e-Gullstrand coordinates~\cite{Kanai2011}. These coordinates are adapted to observers attached to the stellar body. We shall write the metric as \begin{eqnarray} ds^2=-A^2(t,r) dt^2 + {1 \over B^2(t,r)}[dr - v(t,r) dt]^2+r^2 d\Omega^{\,2}_2, \label{generalized-acoustic-metric} \end{eqnarray} with $d\Omega^{\,2}_2$ being the line element of the unit 2-sphere. The three functions $A(t,r)$, $B(t,r)$ and $v(t,r)$ will in the following be given patch by patch. Let us begin considering that the collapse process starts at rest at infinity, so that the line element \eqref{generalized-acoustic-metric} reduces to the the standard Painlev\'e-Gullstrand one, with $A^2(t,r)=c^2$ and $B^2(t,r)=1$. When familiarized with this situation, we will move on to consider that the more general case in which the collapse starts from an initial radius $r_{\rm i}$ in Sec. \ref{sec:5c}. \subsection{Collapse from infinity and homogenous thin-layer transition \label{sec:5a}} To describe the bounce, we shall glue two geometries, one corresponding to the Oppenheimer-Snyder collapse of a homogeneous ball of dust, the other one being its time-reversal. Let us consider for the moment the metric outside the star, corresponding to the Schwarzschild solution. The velocity profile $v(t,r)$ presents a flip of sign between the black-hole and white-hole patches of the Kruskal manifold. Thus the metric we are seeking for will be characterized by a velocity profile \begin{equation} v(t,r) =v_{\rm s}(r)[1-2\theta(t)]=-c\,[1-2\theta(t)]\sqrt{{r_{\rm s}\over r}},\qquad R(t)\leq r\leq+\infty. \label{velocity-profile_1} \end{equation} Here $v_{\rm s}(r)=c\sqrt{r_{\rm s}/r}$ is the absolute value of the standard velocity profile of the Schwarzschild solution, and the Heaviside function $\theta(t)$ is used to perform the transition between the different patches. Concerning the function $R(t)$ describing the trajectory of the surface of the star, for the moment the only condition we demand, following our previous discussion, is that it is bounded from below by $R(0)=R_0$ where $t=0$ corresponds to the moment of the bounce in the coordinates we are using. This gluing procedure by itself is nothing but a brute force exercise, and the result presents unpleasant features. However, we want to present a constructive procedure in which the details are progressively added on demand, eventually arriving to the complete picture in all its generality. For instance, using the Heaviside function to construct the geometry makes all the $t=0$ hyperplane singular for $r>R_0$, in the sense that the metric is discontinuous there. Of course, this does not necessarily reflect any physical reality, as we can always introduce a regulator to make the geometry finite. Let us introduce the continuous, differentiable but non-analytic function \begin{equation} f(t)=\left\{\begin{array}{cr}\exp(-1/t)\qquad&0\leq t,\\ 0\qquad&t\leq0,\\\end{array} \right. \end{equation} that permits us to define for $t_{\rm R}\in\mathbb{R}$ the function \begin{equation} g_{t_{\rm R}}(t)=\frac{f(1/2+t/t_{\rm R})}{f(1/2+t/t_{\rm R})+f(1/2-t/t_{\rm R})}.\label{eq:interfun} \end{equation} This function represents a regulated version of $\theta(t)$ that interpolates between the values $g_{t_{\rm R}}=0$ (for $t\leq -t_{\rm R}/2$) and $g_{t_{\rm R}}=1$ (for $t\geq t_{\rm R}/2$). The value of $t_{\rm R}$ thus controls the duration of the interpolation. We can actually define the pointwise limit \begin{equation} \theta(t)=\lim_{t_{\rm R}\rightarrow 0}g_{t_{\rm R}}(t), \end{equation} thus fixing the convention $\theta(0)=1/2$. \begin{figure}[h]% \vbox{ \hfil \includegraphics[width=0.55\textwidth]{figure.jpg}\hfil} \bigskip% \caption{Transition from the black-hole patch to the white-hole patch by using a smooth interpolation with characteristic time scale $t_{\rm R}$. The qualitative behavior of light cones for different values of the radial coordinate is shown.} \label{Fig:transition}% \end{figure}% With respect to the field equations, the strict limit $t_{\rm R}\rightarrow 0$ may not well defined as a solution even in the distributional sense, as its curvature may contain terms that are quadratic in the Dirac delta function, which are ill-defined (in general, the product of two distributions cannot be defined \cite{Grubb2009}). Let us emphasize again that we always keep in the back of our minds that the relevant physical situation corresponds to $t_{\rm R}$ small, but nonzero. For $t_{\rm R}>0$ we can safely insert the metric in the Einstein field equations \eqref{eq:einsteineqs} to obtain (using, e.g., Eqs. 3.2 in \cite{Kanai2011}) the following nonzero components of the effective stress-energy tensor supporting this geometry: \begin{align} \mathscr{T}^1_{\,\,\,\,\,0}=-\frac{c}{2\pi}(1-2g_{t_{\rm R}})\dot{g}_{t_{\rm R}}\frac{M}{r^2},\nonumber\\ \mathscr{T}^1_{\,\,\,\,\,1}=-\frac{c}{2\pi}\dot{g}_{t_{\rm R}}\frac{M}{r^2}\left(\frac{r_{\rm s}}{r}\right)^{-1/2},\nonumber\\ \mathscr{T}^2_{\,\,\,\,\,2}=-\frac{c}{8\pi}\dot{g}_{t_{\rm R}}\frac{M}{r^2}\left(\frac{r_{\rm s}}{r}\right)^{-1/2}. \label{eq:effset} \end{align} We use the notation $\mathscr{T}_{ab}$ in order to make a clear distinction between this effective source originated in the bounce, and the stress-energy tensor $T_{ab}$ of the perfect fluid that forms the stellar structure. It is interesting to note that the effective stress-energy tensor \eqref{eq:effset} only involves first derivatives of the interpolating function $g_{t_{\rm R}}$, so that the distributional limit is indeed well-defined, corresponding to a $\delta-$function source located on $t=0$; for $t_{\rm R}>0$ this source presents certain width. The effective source \eqref{eq:effset} violates the null energy condition \eqref{eq:null2} in all its radial extension; let us consider for example the hyperplane $t=0$ and the null vector $u^a$ on it, with components $u^0=u^1=1$, $u^2=u^3=0$. Then \begin{equation} \mathscr{T}_{ab}u^au^b=-\frac{c}{\pi}\frac{M}{r^2}\left(\frac{r_{\rm s}}{r}\right)^{-1/2}\left.\dot{g}_{t_{\rm R}}\right|_{t=0}<0. \end{equation} Taking into account the metric inside the collapsing star, the complete geometry is given by: \begin{equation} A^2(t,r)= c^2 B^2(t,r)=c^2, \end{equation} and \begin{eqnarray} v(t,r) =-c\,[1-2g_{t_{\rm R}}(t)]\times\left\{ \begin{array}{cr} \displaystyle\sqrt{{r_{\rm s} \over r}}\qquad &R(t)\leq r, \\ \displaystyle{r \over R(t)}\sqrt{{r_{\rm s} \over R(t)}}\qquad &r\leq R(t). \end{array} \right. \label{velocity-profile1} \end{eqnarray} The function $R(t)$ is constructed from the calculation in GR of the Oppenheimer-Snyder collapse starting from infinity \cite{Kanai2011}: \begin{eqnarray} R(t)=\left(\frac{9r_{\rm s}}{2}\right)^{1/3}|c\,t|^{2/3}. \end{eqnarray} The point $t=0$ in this trajectory would correspond to the classical singularity. For $t<0$ the trajectory goes inwards, while for $t>0$ it goes outwards. To make the geometry smooth, we shall consider that the function $R(t)$ is modified by means of a smooth interpolation in the interval $t\in[-t_{\rm R}/2,t_{\rm R}/2]$ so that the classical singularity is never formed; as said before, $R(t)$ should be bounded from below by $R(0)=R_0>0$. The order of magnitude of $R_0$ is fixed through model-independent considerations as given by Eq. \eqref{eq:minrad}. To sum up, the two main features of the geometry presented in this section are the regularization of the singularity by the bounce of the collapsing matter at $R(0)=R_0$ and the existence of a thin-layer transition region between the black-hole and white-hole patches that violate the null convergence condition \eqref{eq:null1}. This thin layer is homogeneous in the sense that the interpolation time $t_{\rm R}$ is independent of the radial coordinate; this dependence will be included in Sec. \ref{sec:5c}. Let us discuss for the moment some additional relevant properties of this region of spacetime. \subsection{Non-perturbative ultraviolet effects \label{sec:5b}} The curvature scalar corresponding to the effective source \eqref{eq:effset} is given by \begin{equation} \mathscr{R}=6\frac{\dot{g}_{t_{\rm R}}}{c}\sqrt{\frac{r_{\rm s}}{r^3}}.\label{eq:infcur} \end{equation} This expression permits us to make an order of magnitude estimate for $t_{\rm R}$, by the following argument. Let us assume that, when the stellar radius reaches its minimum value, $R_0$, its density is of the order of the Planck density, corresponding to Planckian curvatures by virtue of \eqref{eq:planckcur}. These Planckian curvatures generate non-perturbative contributions that prevent the structure from collapsing, triggering the bounce process [recall the form of the modified Friedmann equation \eqref{eq:frimod}]. For continuity reasons, it is reasonable to expect that the modifications of the classical geometry outside, but very close to the stellar radius, are at this stage also Planckian. This is guaranteed if, as using Eqs. \eqref{eq:minrad} and \eqref{eq:infcur}, and the estimation $\dot{g}_{t_{\rm R}}\sim 1/t_{\rm R}$ reveals, \begin{equation} t_{\rm R}\sim \frac{\ell_{\rm P}}{c}=t_{\rm P},\label{eq:treg} \end{equation} with $t_P$ the Planck time. These extremely short times, in conjunction with the mechanism acting as the trigger and the information about the curvature \eqref{eq:infcur}, suggest the following interpretation of the effective source \eqref{eq:effset} by analogy: it may be understood as a shock wave, originated by the violent bounce of the stellar structure at $r=R_0$, that propagates outwards. The decay rate of this ``curvature wave'' with the radius goes as $r^{-3/2}$, which is to be compared with, e.g., the decay of the classical value of the Kretschmann scalar $\mathscr{K}=\mathscr{R}_{abcd}\mathscr{R}^{abcd}$ of the Schwarzschild geometry, valid for $|t|>t_{\rm R}/2$, \begin{equation} \sqrt{\left.\mathscr{K}\right|_{|t|\geq t_{\rm R}/2}}=2\sqrt{3}\,\frac{r_{\rm s}}{r^3}.\label{eq:kre1} \end{equation} Indeed, to be more consistent we may compare this expression with the ultraviolet corrections to the Kretschmann scalar at $t=0$ for instance. The evaluation of the leading order of this quantity for short transients shows that it nevertheless behaves as Eq. \eqref{eq:infcur}: \begin{equation} \sqrt{\left. \Delta \mathscr{K}\right|_{t=0}}\simeq 6\sqrt{2}\,\frac{\dot{g}_{t_{\rm R}}}{c}\sqrt{\frac{r_{\rm s}}{r^{3}}}.\label{eq:kremod} \end{equation} While on $R_0$ the classical and ``ultraviolet'' parts of the Kretschmann scalar are of the same order by construction, their decay rates for $r>R_0$ are rather different. A look at Eqs. \eqref{eq:kre1} and \eqref{eq:kremod} suffices to check that the effects associated with the shock wave decrease more slowly than the curvature of the classical geometry. We could say that the wave preserves its non-perturbative character, with respect to the background, in its way out. As we will discuss in detail in Sec. \ref{sec:5f}, this is the mechanism that permits to understand how ultraviolet modifications are able to alter dramatically the geometry on the horizon, even if the curvatures might no longer be Planckian there. Notice that the infinite radial extension of this shock wave is an unphysical feature of this simplified model that will not survive in more elaborated models such as we explain below. \subsection{Collapse from a finite radius and triangular-shaped transition \label{sec:5c}} Up to now, we have considered that the initial radius of the stellar structure is infinite. Let us look for the description of the more realistic case of gravitational collapse from a finite radius $r_{\rm i}$. The resulting metric will be written again in the form of Eq. \eqref{generalized-acoustic-metric}. The collapse of the star begins at rest at the initial radius $r_{\rm i}$ and is represented by the trajectory of the star's surface $R(t)$. The arbitrary zero of time is chosen such that the collapse starts at $t=-t_{\rm B}/2$, where \begin{equation} t_{\rm B}=\pi\, \frac{r_{\rm i}}{c}\sqrt{\frac{r_{\rm i}}{r_{\rm s}}}\label{eq:tbounce} \end{equation} is twice the classical collapsing time from $r=r_{\rm i}$ to $r=0$ in the Oppenheimer-Snyder model. For $t\leq-t_{\rm B}/2$, i.e., before the collapse has started, the metric is exactly Schwarzschild outside $r_{\rm i}$, while for $r\leq r_{\rm i}$, \begin{equation} A^2(t,r)=c^2,\qquad B^2(t,r)=1-\frac{r_{\rm s}r^2}{r_{\rm i}^3},\qquad v(t,r)=0. \end{equation} This metric corresponds to a star of homogeneous density, maintained static by an appropriate internal pressure. Once the collapse has started, the three coordinate patches we use are $0\leq r\leq R(t)$, $R(t)\leq r\leq r_{\rm i}$, and $r_{\rm i}\leq r\leq +\infty$. The functions describing the metric are then: \begin{eqnarray} A^2(t,r)=c^2\times\left\{ \begin{array}{cr} \displaystyle \frac{1-r_{\rm s}/ r}{1-r_{\rm s}/r_{\rm i}}\qquad &r_{\rm i}\leq r, \\ 1\qquad &R(t)\leq r\leq r_{\rm i}, \\ 1\qquad &r\leq R(t); \end{array} \right. \label{c-profile} \end{eqnarray} \begin{eqnarray} B^2(t,r)= \left\{ \begin{array}{cr} \displaystyle1-r_{\rm s} /r\qquad &r_{\rm i}\leq r, \\ \displaystyle1-r_{\rm s}/r_{\rm i}\qquad &R(t)\leq r\leq r_{\rm i}, \\ \displaystyle1-{r_{\rm s} \over r_{\rm i}}\left({r \over R(t)}\right)^2\qquad &r\leq R(t); \end{array} \right. \label{epsilon-profile} \end{eqnarray} \begin{eqnarray} v(t,r)=-c\,[1-2g_{t_{\rm R}(r)}(t)]\times\left\{ \begin{array}{cr} 0\qquad &r_{\rm i}\leq r, \\ \displaystyle\sqrt{{r_{\rm s} \over r}-{r_{\rm s} \over r_{\rm i}}}\qquad &R(t)\leq r\leq r_{\rm i}, \\ \displaystyle{r \over R(t)}\sqrt{{r_{\rm s} \over R(t)}-{r_{\rm s} \over r_{\rm i}}}\qquad &r\leq R(t). \end{array} \right. \label{velocity-profile} \end{eqnarray} In this case, the function $R(t)$ is constructed from the calculation in GR for the Oppenheimer-Snyder collapse from a finite radius $r_{\rm i}$: \begin{eqnarray} R(t)= {r_{\rm i} \over 2}(1 +\cos \eta),\qquad t= \frac{t_{\rm B}}{2\pi}(\eta +\sin \eta-\pi),\label{eq:oppsny} \end{eqnarray} with $\eta \in [0,2\pi]$. The point $\eta=\pi$ ($t=0$) in this trajectory would correspond to the classical singularity; for $\eta<\pi$ the trajectory goes inwards, while for $\eta>\pi$ it goes outwards. Again, we shall modify the function $R(t)$ so that it is bounded from below by $R(0)=R_0>0$. A new feature in this situation is the necessary introduction of a radial dependence on the parameter $t_{\rm R}$ that controls the interpolation time, which then becomes a function $t_{\rm R}(r)$. This quantity enters through the function $g_{t_{\rm R}(r)}(t)$ in Eq. \eqref{velocity-profile}. This is necessary to guarantee that curvature invariants are kept finite in the surroundings of the $r=r_{\rm i}$ hypersurface. Had we taken this parameter as being constant, the leading order of the Ricci scalar in the limit $r\rightarrow r_{\rm i}$ from below would be given, at $t=0$ for instance, by \begin{equation} \left.\mathscr{R}\right|_{t=0}\simeq \frac{4r_{\rm s}(3r_{\rm i}-4r)}{r_{\rm i}\,r^2\sqrt{\displaystyle\frac{r_{\rm s}}{r}-\frac{r_{\rm s}}{r_{\rm i}}}}\frac{1}{c\,t_{\rm R}}.\label{eq:domricc} \end{equation} On the other hand, the consideration of \emph{decreasing} functions $t_{\rm R}(r)$ with the radius is clearly motivated by the understanding of the modifications of the near-horizon geometry as the result of the propagation of non-perturbative ultraviolet effects from $r=R_0$ up to $r=r_{\rm i}$. It is therefore quite remarkable that avoiding that the scalar curvature blows up at $r=r_{\rm i}$ at $t=0$ demands that $t_{\rm R}(r)$ verifies the differential equation \begin{equation} (3r_{\rm i}-4r)\frac{1}{t_{\rm R}(r)}+2r(r_{\rm i}-r)\frac{d}{dr}\left(\frac{1}{t_{\rm R}(r)}\right)=3C\sqrt{r_{\rm i}-r}, \end{equation} where $C$ is an arbitrary constant. In this equation, the first term on the left-hand side corresponds to the contribution reflected in Eq. \eqref{eq:domricc}, while the second term on the left-hand side corresponds to additional contributions to the Ricci scalar that show up when the radial dependence of the parameter $t_{\rm R}(r)$ is included. The right-hand side of this equation is the minimal expression which cancels the term that goes to zero in the denominator of Eq. \eqref{eq:domricc}, thus ensuring a good behavior of the Ricci scalar in the limit $r\rightarrow r_{\rm i}$. The inhomogeneous solution to this differential equation is given by the decreasing function \begin{equation} t_{\rm R}(r)=\frac{1}{C}\sqrt{r_{\rm i}-r}.\label{eq:dectr0} \end{equation} The simplest ansatz for our geometry is assuming that Eq. \eqref{eq:dectr0} holds for the entire interval $R_0\leq r\leq r_{\rm i}$. Demanding $t_{\rm R}(R_0)=t_{\rm P}$ fixes the constant $C$, leading to \begin{equation} t_{\rm R}(r)=t_{\rm P}\sqrt{\frac{r_{\rm i}-r}{r_{\rm i}-R_0}}.\label{eq:dectr} \end{equation} One can show by direct evaluation that this choice regularizes at the same time other curvature invariants such as, e.g., the Kretschmann scalar. Fixing $t_{\rm R}(r)$ as in Eq. \eqref{eq:dectr} specifies completely the geometry, in which the thin-layer region that encloses the non-standard ultraviolet effects that is depicted in Fig. \ref{Fig:transition} will be generally transformed into a (smoothed) triangular-shaped region, with one of the vertices located at $(t,r)=(0,r_{\rm i})$. The inevitable introduction of the function $t_{\rm R}(r)$ is intimately related to the fact that non-perturbative ultraviolet effects are naturally confined in this case into a compact ball of radius $r_{\rm i}$, so that the form of the Schwarzschild metric is preserved for $r>r_{\rm i}$. In the most general case the maximum radius reached by ultraviolet effects could be an \emph{independent} parameter $r_{\rm m}$ such that $r_{\rm m}>r_{\rm s}$. The explicit construction of the corresponding geometries is more involved; we refer the reader to \cite{Barcelo2015} for a discussion of this issue. The geometries we have constructed explicitly have the necessary properties to reflect appropriately the main implications of the black to white hole transition in short characteristic time scales. We shall therefore take them as the starting point for our following discussions. \subsection{Short-lived trapping horizons \label{sec:5d}} The first of the quantities of interest we can evaluate shows one of the nontrivial features that characterizes the entire family of bouncing geometries we are considering in this paper. In the first model we have discussed with a thin-layer transition region, calculations performed on the limit $t_{\rm R}\rightarrow 0$ will be exact up to $\mathscr{O}(t_{\rm R}/t_{\rm B})$ terms, thus implying this distributional limit will prove a very good analytical approximation on the basis of Eqs. \eqref{eq:treg} and \eqref{eq:tbounce}. This is also true in the second model with a triangular-shaped transition region, but now with $t_{\rm R}(R_0)\rightarrow 0$. Let us therefore start by considering the distributional limit $t_{\rm R}(R_0)\rightarrow 0$ for mathematical simplicity. In this case, the proper time for observers situated at $r=r_{\rm i}$ for the bounce to take place is exactly the same as for observers attached to the surface of the star. Indeed, from the perspective of an observer collapsing with the star, the entire process of collapse and bounce with respect to some reference initial position takes a time $t_{\rm B}$ as defined in (\ref{eq:tbounce}), that is, twice the free-fall collapsing time. By construction of the coordinate system, the proper time of the observer attached to the stellar structure is simply the Painlev\'e-Gullstrand time, $d\tau=dt$, as for that observer $dr/dt=v(t,r)$. This same process seen by an observer always at rest at the initial position $r_{\rm i}$ takes {\em this very same time}. For this observer $dr=0$ and $v(t,r_{\rm i})=0$, implying $d\tau= dt$, so that its proper time is given the same temporal coordinate we are using to write down the metric. Seen by observers far outside the collapsing star ($r\gg r_{\rm s}$) the entire process would take this time multiplied by the standard general-relativistic redshift factor, $(1-r_{\rm s}/ r_{\rm i})^{-1/2}$. This redshift factor will be of order unity for stars initially larger than a few times their Schwarzschild radius. The fact that the proper time separation between these two events, the start of the collapse and its return to the initial position, is the same for the surface-attached observer and the one standing still at the initial radial position, is a general property of the geometries we are considering: a finite value of $t_{\rm R}(R_0)$ of the order of \eqref{eq:treg} leads, taking into account Eq. \eqref{eq:tbounce}, to extremely small $\mathscr{O}[t_{\rm R}(R_0)/t_{\rm B}]$ corrections coming from the finite transient zones. Therefore, these geometries do not exhibit long-lived trapped regions of any sort, but only short-lived trapped regions. From a different point of view, imagine that one were to monitor with high time resolution this time-symmetric collapse from the reference initial position, that is assumed to be sufficiently far outside the Schwarzschild radius. One sets up two synchronized clocks at this initial position before the collapse. Then one clock is left to follow the collapsing structure and the other is kept at rest in the reference position. By observing with a telescope the ticks of the clock falling with the star, one would see that in the collapsing phase the ticks slow down progressively. However, at some point they start to speed up in such a way that when the two clocks are finally back together they show precisely the same time. This is easily understood if we think in terms of the analogue metric in fluids \cite{Barcelo2005}. When the star is collapsing, the velocity profile is that of a sink, so that signals originated on the star's surface will be emitted at different positions, picking up a delay that depends on the time that light needs to cover up that distance. On the contrary, in the white-hole patch the velocity profile is reminiscent of a source, which effectively shortens the time between different signals for the external observer. Overall, for an external observer, that uses essentially the Schwarzschild time coordinate as its proper time (but for an irrelevant redshift factor depending on his position), the collapsing phase will last longer than the expanding phase, so that he might not realize the time-symmetric character of the process. In summary, one of the essential features of the process we are considering is that the time lapse associated with the collapse is short, of the order of tenths of a millisecond for neutron-star-like initial configurations. This is equally true both for observers attached to the structure as well as for external stationary observers. Within this quite general family of geometries (arguably a complete set of geometries interpolating between a black-hole geometry with the white-hole geometry), the only way to prescribe geometries that allow for extremely long times of the bounce process as seen by external observers is to introduce extremely slow transients. By using the adjective \emph{extremely slow} we understand a regularization that takes a very long time, measured by observers attached to the surface of the star, in order to overcome the gravitational attraction and start a noticeable expansion of the stellar structure; in mathematical terms this corresponds to very large $t_{\rm R}(R_0)/t_{\rm B}$ quotients. From the perspective of the regularization of the singularity as well as observers \emph{inside} the stellar structure, this quasi-static possibility is arguably quite unnatural. In the framework of our discussion in Sec. \ref{sec:4}, these solutions would correspond to prioritizing the role of the external metric in the entire process, subjugating the behavior of matter as well as the possible ultraviolet effects to the prevalence of the long-lived horizon. \subsection{Short transients and the propagation of non-perturbative ultraviolet effects \label{sec:5f}} All the transients lead to a characteristic imprint as we have already discussed: the deviation of the near-horizon outer geometry, that is, the Schwarzschild geometry beyond $r=r_{\rm s}$. Such a feature is anathema in the orthodox view on the possible relevance of ultraviolet effects on classical geometries. It belongs to the conventional wisdom that appreciable deviations from the classical behavior are to be expected in regions of large curvature in, e.g., Planck units. The argument is that only then the possible corrections to the Einstein field equations are expected to be non-perturbative. In the Schwarzschild solution the Ricci curvature tensor is zero but the Weyl part of the Riemann curvature tensor leads to the Kretschmann scalar \begin{equation} \mathscr{K}=\mathscr{R}_{abcd}\mathscr{R}^{abcd}=\frac{12 r_{\rm s}^2}{r^6}.\label{eq:kre} \end{equation} This implies that the geometry around $r=r_{\rm s}$ of stellar-size black holes is to be regarded as a robust feature, not to be affected by any kind of ultraviolet modifications. While this view certainly makes sense in the study of the static case (i.e., an eternal black hole), the role of these considerations is far from clear for us when talking about dynamical situations. For instance, in our proposal there is no modification at all of the behavior of the geometry near the moment of formation of the trapping horizon in the black-hole patch. An observer who suddenly left the surface of the star to pend close, but outside the horizon, will not notice any deviations from the expected behavior in GR in the few initial instants of the process. It is only when matter reaches a dense enough situation that relevant ultraviolet effects altering the geometry are originated. While at first these effects are confined to regions of high curvature, there is nothing that prevents them to propagate outwards, modifying the geometry in their wake. In doing so these effects can reach regions that were characterized by low curvatures, such as the near-horizon region. However, this effect should not be seen as a modification of the behavior of regions with low curvature by unnaturally large ultraviolet effects, but rather as the propagation of a non-perturbative wave of high-curvature through regions of low curvature. Interestingly, this picture is self-consistent only for short transients, with characteristic time scale $t_{\rm R}(R_0)$ of the order of the Planck time. This is neatly illustrated by taking as working example the geometry considered in Sec. \ref{sec:5a}, though the situation is generically the same for all the geometries we have considered. The Kretschmann scalar of the transient in this case is given in \eqref{eq:kremod}. As argued above, the short (Planckian) time $t_{\rm R}$ associated with the transient implies that the modifications of the spacetime curvature in the surroundings of the stellar structure, when the latter reaches its minimum radius $R_0$, are also of the order of the Planck curvature. Thus in geometries with short transients, and only in these cases, the trigger of non-perturbative ultraviolet effects on the metric outside the star is inextricably tied up to spacetime regions with high curvature. This region of high curvature will propagate outwards, getting diluted in this process: already when reaching the horizon, the magnitude of this curvature several orders of magnitude lower, roughly by \begin{equation} \left(\frac{R_0}{r_{\rm s}}\right)^{3/2}\sim\frac{m_{\rm P}}{M}. \end{equation} One might be tempted to argue on this basis that ultraviolet effects should not be able to significantly alter the near-horizon geometry, being the corresponding curvature far smaller than Planckian.\footnote{It is interesting to note that it is in principle possible to construct models in which the consideration of triangular-shaped transition regions could even lead to Planckian curvatures near the horizon, by just making an appropriate ansatz for $t_{\rm R}(r)$ from $r=R_0$ up to the near-horizon region.} The situation is, indeed, the contrary. While the classical distribution of the curvature (measured by the Kretschmann scalar) decays with the radius as $r^{-3}$ [Eq. \eqref{eq:kre1}], the curvature of the non-standard region associated with ultraviolet effects decays as $r^{-3/2}$ [Eq. \eqref{eq:kremod}]. This guarantees that modifications that are non-perturbative at $R_0$, remain non-perturbative near the horizon. Thus modifications of the near-horizon geometry do not appear because of unnaturally large ultraviolet effects originated there, but rather as a result of the propagation of sudden ultraviolet effects that are originated when the stellar structure undergoes a violent bounce at Planckian densities. It is certainly not possible, given our present knowledge of the gravitational interaction, to back up the picture with short characteristic time scales from a first-principle, complete perspective. The work of {H{\'a}j{\'{\i}}cek} and collaborators \cite{Hajicek2001,Hajicek2003,Ambrus2004,Ambrus2005}, represents nevertheless a first step in this direction. In their study of the quantization of collapsing shells, they have shown that the transition from a collapsing branch to an expanding branch occurs, and that a given definition of the crossing time between an external observer and the dynamical shell is short, being essentially of the same order as the time $t_{\rm B}$ we are considering, given by Eq. \eqref{eq:tbounce}. It is equally worth remarking their skepticism about this result, as the same authors were expecting very long times to occur, in order for their model to be accommodated within the REBH family. The definition of observables concerning time intervals in these bouncing processes is however far from straightforward \cite{Ambrus2004,Ambrus2005}. At the light of the developments presented here, this issue certainly deserves further study, both in this particular model and in frameworks that use different quantization techniques, such as \cite{Hartle2015} for instance. In any case, lacking for the moment a fundamental justification for the short transients does not prevent our discussion to be self-consistent at low energies, while presenting new intuitions about the way in which ultraviolet modifications to the behavior of GR would behave. On top of this, what we want to stress is the genuine opportunity that these hypothetical processes offer: in contrast with REBH proposals, which are experimentally inert for extremely long times and thus for practical purposes, a process with a short characteristic time scale should lead to clear imprints that could be hopefully detectable. Thus this proposal is audacious, but not without consequences, as it offers prospects of being falsifiable much more easily than any other models that are nowadays present the literature. \section{Physical and observational consequences \label{sec:6}} If there exists a regularization of the classical behavior of the form we have described, the collapse process itself would not constitute the final stage of collapse in stellar physics. One would immediately be impelled to wonder about what would happen after the bounce. The search for new states of equilibrium on the one hand, and the understanding of the transient collapse process itself on the other, become entirely distinct issues. If our idea is at work in nature it should have many observable effects, though additional work is needed in order to determine the possible signatures of both phases. In the collapse of, say, a neutron star, matter would remain apparently frozen at the Schwarzschild radius for just a few tenths of a millisecond before being expelled again. Even neglecting dissipation, the metric is non-Schwarzschild during an extremely short time interval in a region extending outside the gravitational radius. In realistic situations the bounce will not be completely time-symmetric: part of the matter will go through towards infinity in the form of dissipative winds while the remaining mass will tend to recollapse. In this way one would have a brief and violent transient phase, composed of several bounces, followed by the formation of a new (stable or metastable) equilibrium object with the resulting mass. \subsection{Towards new figures of equilibrium \label{sec:6a}} In an ideal situation, perfectly spherically symmetric and without dissipation, the collapsing body would enter into a never-ending cycle of contracting and expanding phases. In a realistic situation, though, one expects that the system will dissipate at least quantum mechanically while searching for new equilibrium configurations. Here the panorama of possibilities is almost unexplored. Let us discuss the possibility we think more plausible. Notice that any development on this direction should properly acknowledge and deal with, for instance, the observations of the existence of extremely dense objects in our own galaxy. In~\cite{Barcelo2008} it was shown that if the velocity of trapping-horizon crossing by the collapsing matter distribution were rather small, then quantum effects of vacuum polarization would become so powerful that they might even stop the collapse. It is very unlikely that the classically expected almost-free-falling collapse of stellar structures like neutron stars would lead directly to strong vacuum polarization effects. However, in our scenario, when taking into account dissipation, one would expect that each new recollapsing phase would start from a position closer to its Schwarzschild radius than the previous one. In this way, at some stage vacuum polarization effects could start to be relevant and even stop further collapses. This might lead for the time being to hypothetical almost-stationary structures hovering extremely close to their gravitational radius. There might be other mechanisms underlying the stabilization of stellar structures close to their gravitational radius. What is relevant here is that the final metastable object could be small, dark, and heavy, but without black- or white-hole districts (see~\cite{Visser2008} and references therein). These black stars will not be voids in space, but they will be filled with matter. Since they have no horizons, they will in principle be completely open to astrophysical exploration. Let us stress that black holes as described by classical GR might still continue to be very good models for the external gravitational behavior of these black stars. A natural question in this regard is whether Hawking-like evaporation, being a paradigmatic theoretical feature of black holes, would be also a characteristic of these new objects. When the system stabilizes close to its Schwarzschild radius, it might emit or not and with different spectral properties depending on the specifics of the structure, which at this stage are difficult to envisage. In other words, the Hawking effect \emph{does not} need to be preserved in the black-star scenario. However, we have already mentioned in our previous discussion that there are ways in which these objects could acquire emission and evaporation properties resembling Hawking's scenario. At least two different mechanisms are known. One such structure could emit a Hawking-like flux if it were continuously and asymptotically approaching its Schwarzschild radius (without crossing it)~\cite{Barcelo2006}, or if it were pulsating in a close to free-fall manner~\cite{Barbado2011}. Should this radiation exist, it would in both cases be Planckian but not strictly thermal, as correlations are maintained by both mechanisms, and the total amount of energy radiated would be in principle negligible given the extremely short time associated with the transients. In relation with Hawking radiation, notice that these scenarios do not invite us to wonder whether information is lost or not, as no singularities and no long-lived trapping horizons are formed in the first place. On the other hand, during the transient phases one would expect quantum dissipation in the form of particle production. This particle production will be in general non-thermal, though at trapping-horizon crossings it would have the form of short bursts of thermal radiation~\cite{Stephens1994,Barcelo2006,Barbado2011}. Once the object has settled down, it would probably be extremely difficult to distinguish it from a standard black hole through astrophysical observations. There have been some proposals to discern whether or not there exists an explorable surface in objects associated with black holes \cite{Narayan2002,Narayan2008}. The absence of Type I X-ray emission in binaries containing a black-hole mimicker have been argued to imply that the candidate did not have an external surface but an event horizon~\cite{Narayan2002}. However, things are clearly not that simple (for some specific criticism see~\cite{Abramowicz2002}). The reaction of the black hole mimicker when absorbing some matter from its companion would strongly depend on its specific heat capacity. If the black-hole candidate has a heat capacity similar to those of a black hole, which can be expected due the high redshift of its surface, its behavior would be difficult to distinguish in this regard from that of a proper black hole. This kind of observations can play a significant role, though, in putting constraints to specific models of black hole mimickers (see, e.g.,~\cite{BroderickNarayan2007} for constraints on gravastars). This task could even be more difficult for proposals in which the compact object lacks a hard surface \cite{Vincent2015}. A different issue is the possibility, in principle, of the hypothetical use of a radar to check the presence or not of a surface. As discussed in~\cite{Barcelo2015}, a \emph{elastic} scattering of a wave in the supposed surface would distinguish in no time whether a surface exists or not. \subsection{Energetics of the transient phase \label{sec:6b}} When considering realistic situations with dissipation, the transient phase might leave some traces, for instance in the physics of gamma-ray bursts (see, e.g., \cite{Piran2004}). There is experimental evidence that a subset of these events, the so-called \emph{long} gamma-ray bursts, are associated with the final stages in the life of very massive stars. The most widely accepted theoretical picture is known as the collapsar model \cite{MacFadyen1998}. It is natural to expect that a modification of the standard gravitational collapse process to a black hole that is considered here could leave clear imprints associated with a reverberant collapse. However, in the collapsar model of GRBs the emission zone is supposed to be very far from the collapsed core \cite{Piran2004}. This means that the connection between the processes at the core and those at the external wind shells could be very far from direct. However, the general features of the model are enough in order to roughly compare its energetics to those of GRBs. This comparison may be used in order to understand whether or not the bounce process is a reasonable candidate for the mechanism behind these bursts. So let us assume that the picture discussed above is realized in nature: the occurrence of violent bouncing processes, dissipation and final stabilization in the form of a black star. We can estimate the effect of the energy loss in the entire process by using the following argument. Recall that for dust matter content, and in the absence of rotation, the differential equation for the trajectory of the surface of the star is mathematically equivalent to that of a test particle with the overall mass of the star following a radial geodesic in Schwarzschild spacetime. We can then use the conserved quantities associated with the geodesic equations in this spacetime. In particular, we shall use the conserved quantity $E$ that is associated with energy. So let $r_{\rm i}$ be the initial radius and $r_{\rm s}$ the Schwarzschild radius of the star, and consider the Schwarzschild effective potential for radial motion. If the structure was originally at rest, its energy is given by \begin{equation} \left(\frac{E}{M c^2}\right)^2=1-\frac{r_{\rm s}}{r_{\rm i}}.\label{eq:schen} \end{equation} This equation has a clear interpretation: the positive term on the right-hand side is the rest energy of the star, while the second term corresponds to the negative gravitational energy of the structure. Then the energy of the resulting compact body, as defined in \eqref{eq:schen}, is essentially zero. Energy balance implies that the energy that has to be released in the entire process, for example by means of the emission of a shell of matter that escapes to spatial infinity, is given by: \begin{equation} \Delta E=Mc^2\sqrt{1-\frac{r_{\rm s}}{r_{\rm i}}}.\label{eq:energetics} \end{equation} This is a model-independent estimation that tell us that the object has to get rid of a significant portion of its original rest energy in order to reach stabilization. If we take for example a neutron star with $r_{\rm i}=2\,r_{\rm s}$, then the emitted energy is $Mc^2\sqrt{1/2}\simeq 0.71\, Mc^2$. How is this compared with the GRBs energetics? Interestingly, \eqref{eq:energetics} is of the same order of magnitude as the energy emission in those events if the emission is considered isotropic (see, e.g., \cite{Piran2004,Mobberley2009}). Indeed, it is of the same order of magnitude as other theoretical estimates that are based, for instance, on the Penrose process \cite{Frolov2011} or similar mechanisms for energy extraction from charged black holes \cite{Christodoulou1971,Damour1975}. A comment that applies to all these estimations is that there is experimental evidence that the emission is collimated so that the real energy that is emitted is smaller than \eqref{eq:energetics}, roughly by a factor of $10^{-2}$ . Of course, \eqref{eq:energetics} is a crude estimate that does not take into account other effects that would take place near the collapsing object, besides being evaluated in an isotropic model that does not take into account rotation. That with simple, model-independent ingredients we are able to get that close to the observed energetics of GRBs is a strong incentive to consider further developments of the picture. This is a hint in favor of the possibility that the bounce process we are describing might be behind of some GRB events. As we cover in the next section, gravitational waves produced deep inside these violent events may provide a much better observational opportunity, eventually permitting to elucidate the kind of mechanism that is behind them. \subsection{Ripples from the transient phase \label{sec:6c}} One of the primary predictions of GR that still awaits experimental corroboration is the generation and propagation of local disturbances of spacetime, or gravitational waves. Gravitational-wave astronomy is nowadays a mature branch from a theoretical perspective, while there has been a great deal of experimental effort in order to overcome the difficulties in the detection of these tiny ripples in spacetime. The scientific potential that this new observational window promises is huge; see \cite{Fryer2011,Buonanno2014} for instance. The fine theoretical knowledge of the gravitational wave patterns associated with different gravitational phenomena would make possible to unveil information about astrophysical processes that is definitely not possible to obtain from their electromagnetic counterparts. As a prime example, this observational technique is arguably the best tool to finally determine the physical mechanisms that are behind both short and long GRBs. If the process that we are describing in this paper is realized in nature, the information encoded in gravitational waves may be significantly different, and therefore more surprising than initially expected. The observation of the gravitational wave pattern associated with the gravitational collapse of a massive star into a black hole, if properly correlated with the electromagnetic counterpart of a long GRB, would be the smoking gun of the collapsar model. The generation and form of these wave patterns are well understood nowadays. In the spherically symmetric case, non-spherical inhomogeneities that are present in the initial stellar structure will generate a gravitational wave signal, that terminates with the relaxation of the perturbed horizon to its stable Schwarzschild form (in the presence of rotation it will be described by the Kerr solution instead \cite{Hawking1973}). After that, there is complete silence in the gravitational-wave channel. This is a definite characteristic of the classical gravitational collapse process as described in GR, that will be shared by \emph{all} the REBH models. The robustness and generality of this result is what makes any departures from it highly interesting. We shall describe in the following why do we expect distinctive departures of this behavior to occur in our model, and how can these be computed. It is noteworthy that this is only proposal in the literature to our knowledge that advocates for these kind of modifications on the gravitational wave patterns of collapsing stars. Let us start with a very simple analogue, given by the electromagnetic radiation of a pulsating, or bouncing, charged sphere. In the spherically symmetric case there will no emission of radiation. As in the gravitational case, one has to consider non-spherical distributions of charge in order to trigger the emission of electromagnetic radiation. Now the electromagnetic radiation measured far from the sphere will depend on the given trajectory of the radius of the sphere, $R(t)$, bounded both from below and above. We take this trajectory as the analogue of the trajectory of the star surface in our bouncing geometry. The overall emission of radiation can be obtained in the framework of standard electrodynamics. The emission of electromagnetic radiation in the collapsing branch will be followed by the emission of new radiation pulses coming from the expanding branch, perhaps with a burst corresponding to the bounce event. Notice that the emission of radiation breaks the time symmetry of the trajectory $R(t)$, if present. Additionally, in several bounces we shall get several repetitions of this wave pattern. This is what we expect to occur in the gravitational case, thus leading to a very different gravitational signal when compared with the one obtained in GR. Now when we turn back to gravity, there is a crucial technical difference to take into account when considering this analogy. While the equations of electrodynamics enforce the conservation of charge, the equations of GR do the equivalent with energy. In the electromagnetic case the energy that is emitted in the form of radiation is ultimately provided by the mechanism behind the kinematic evolution of the charge distribution. In other words, the non-spherical distributions of charge that cause the emission of radiation do not decay over time. On the contrary, the equations of GR automatically take into account the decay of non-spherical perturbations due to the emission of gravitational waves. This makes the problem more involved mathematically but, as we discuss in the following, having at hand the metric describing the bounce of the stellar structure one should be able to obtain a definite answer for the spectrum of gravitational waves that is produced in the process for a given perturbation of the initial configuration. Let us sketch the necessary steps in order to do so, taking for instance the metric described in Sec. \ref{sec:5c} as a specific representative of the bounce process. This metric is a solution of the field equations \begin{equation} G_{ab}=\frac{8\pi G}{c^4}\left(T_{ab}+\mathscr{T}_{ab}\right),\label{eq:eins} \end{equation} where $G_{ab}$ is the Einstein tensor, $T_{ab}$ is the dust stress-energy tensor, and $\mathscr{T}_{ab}$ is the non-standard stress-energy tensor, the components of which would be similar to \eqref{eq:effset}, that describes the shock wave produced in the violent bounce. Our previous discussion corresponds to the spherically symmetric situation; in order to produce gravitational waves we need to introduce non-spherical perturbations. These perturbations are introduced in the initial state of the system, namely the star at rest with radius $r_{\rm i}$, with asymptotic conditions ensuring the absence of radiation at spatial infinity. We can perform a perturbative expansion around the spherically symmetric situation of the form \begin{equation} G_{ab}=G_{ab}^{(0)}+G_{ab}^{(1)}+..., \end{equation} where $G_{ab}^{(0)}$ is the Einstein tensor of the spherically symmetric geometry describing the bounce, and $G_{ab}^{(1)}$ will contain the information about the non-spherical perturbations and, in particular, the gravitational wave signal. At first order in the deviations from absolute spherical symmetry, the evolution of these perturbations are then determined by the equations \begin{equation} G^{(1)}_{ab}=\frac{8\pi G}{c^4}T_{ab}^{(1)}.\label{eq:einspert} \end{equation} Notice that $\mathscr{T}_{ab}$ drops off from this equation, as it is constructed to identically cancel the zeroth-order terms. This last equation sets the basis for the study of gravitational wave emission of this process. The study of its implications is currently being carried out, the results of which will be reported elsewhere. It is probably not necessary to stress the appeal of the possibility, even if it could appear remote at present, of detecting a characteristic gravitational wave signal that deviates from the expected classical template. Such an observation would be the smoking gun of the bounce process as described here, providing a low-energy observational window to genuine ultraviolet effects acting on the gravitational collapse of massive stars. \section{Conclusions \label{sec:7}} The outcome of extreme gravitational collapse processes is one of the great theoretical open problems in gravitational physics. The precise determination of the ultraviolet modifications to the classical behavior encapsulated in GR is of course the essential key to unveil its solution. It is not so often stressed that the present understanding of this problem is facing an important dilemma: most of the models in the market largely preserve the semiclassical picture of long-lived trapping horizons, thus obstructing their very experimental verification due to the ridiculous large lifetime of any REBH model. While we are far from denying the theoretical value of these models in understanding the gravitational interaction, in our opinion a serious effort should be made in analyzing the theoretical and observational characteristics of models that exhibit clear testable signatures in the near future. In this paper we have reviewed and worked out new properties of one such alternative model to REBHs. While the perturbative view of the classical picture that is encoded in the REBH paradigm certainly represents the consensus of the community, we have tried to transmit that there exists an interesting alternative. In this alternative model, ultraviolet effects are no longer perturbative, so that the regularization of the singular behavior of GR opens new unexplored avenues for the evolution of the system. The resulting geometries cannot be described as representing essentially the semiclassical perturbation of the trapping horizons that are formed in the collapse, but give preeminence to the regularization of the singularity. In plain terms, these geometries represent the bounce of the matter distribution when reaching Planckian densities, originating a shock wave that propagates outwards, modifying the near-horizon Schwarzschild geometry. All these features have been neatly justified through the careful construction of the explicit form of the corresponding geometries. One of the new general conclusions reached in this paper, by means of the study of curvature invariants of these geometries, is that only for extremely short transients (of the order of the Planck time) the non-perturbative modification of the near-horizon Schwarzschild geometry is justified by the propagation of non-perturbative ultraviolet effects originated at the moment of the bounce. Given the singular character of this model, it provides a unique opportunity for further studies of unexplored issues. For instance, the suggestions it presents about the possible low-energy effects coming from the ultraviolet gravitational regime, and the characteristic time scales for the black to white hole transition, are worth contemplating from a first-principles perspective. On the other hand this model has the potential, if getting mature enough, to lead to surprises for our conception of astrophysical black holes. As we have tried to remark, the most relevant property that singles out this proposal within the vast literature in the subject is its overall short characteristic time scale, which makes possible to think about the (always exciting) possibility of making contact with the empirical reality. \acknowledgments Financial support was provided by the Spanish MICINN through the projects FIS2011-30145-C03-01 and FIS2011-30145-C03-02 (with FEDER contribution), and by the Junta de Andaluc\'{\i}a through the project FQM219. R.C-R. acknowledges support from CSIC through the JAE-predoc program, co-funded by FSE. \bibliographystyle{unsrt}

\section{Introduction} A lattice path in the plane-$(x,y)$ is a set of edges $\lbrace{p_0},p_{1},\ldots,p_{n}\rbrace$ in ${\mathbb Z}^{2}$, such that two edges are related by one vertex, the set of vertices connecting $p_{0}$ to $p_{n}$ is called a lattice path. Several authors studied and enumerated the lattice path, Mohanty and Handa \cite{MH} enumerate the unrestricted lattice paths from $(0,0)$ to $(n,k)$ where $u$ diagonal steps are allowed at each position, Dziemianczuk \cite{D} count the lattice path with four steps horizontal $H=(1,0)$, vertical $V=(0,1)$, diagonal $D=(1,1)$, and sloping $L=(-1,1)$, Fray and Roselle \cite{FR} determine the number of unrestricted weighted lattice paths from $(0,0)$ to $(n,k)$ with horizontal, vertical, and diagonal steps, Rohatgi \cite{R} enumerate the paths which must remain below the line $y = ax + b$ where the diagonal steps are allowed in addition to the usual horizontal and vertical steps. In a Pascal triangle, the binomial coefficients $\binom {n}{k}$ count the number of lattice paths from $(0,0)$ to $(n,k)$ using the steps $\lbrace H=(1,0)\rightarrow, D=(1,1) \nearrow \rbrace$, see {\scriptsize{SLOANE}} \cite{S1} $A007318$.\\ It is well known that the terms of Fibonacci sequence $(F_{n})_{n}$ are obtained by summing the elements crossing the principal diagonal rays in the Pascal triangle, \begin{align} F_{n+1}=\sum_{k}\binom{n-k}{k}, \end{align} where $\binom{n}{k}=\frac{n!}{k!(n-k)!}$ for $n\geq k \geq 0$ and $\binom{n}{k}=0$ otherwise.\\ The generating function of the binomial coefficients is given by \begin{align} \sum_{n\geq0}\binom{n}{k}x^{n}=\frac{x^{k}}{(1-x)^{k+1}}. \end{align} Alladi and Hoggat \cite{tribo1} extended the Pascal triangle. They established the Tribonacci triangle and proved that the sum of elements lying over the principal diagonal rays in the Tribonacci triangle gives the Tribonacci sequence \begin{align*} T_{n+1}=T_{n}+T_{n-1}+T_{n-2}, \end{align*} with $T_{0}=0, T_{1}=1, T_{2}=1$.\\ Denote by $\binom{n}{k}_{[2]}$ the element in the $n^{th}$ row and $k^{th}$ column of the Tribonacci triangle, the triangle is produced by the recurrence relation \begin{align*} \binom{n}{k}_{[2]}=\binom{n-1}{k}_{[2]}+\binom{n-1}{k-1}_{[2]}+\binom{n-2}{k-1}_{[2]}, \end{align*} where $\binom{n}{0}_{[2]}$= $\binom{n}{n}_{[2]}=1.$ We use the convention $\binom{n}{k}_{[2]}=0$ for $k\notin \left \{ 0,\ldots ,n\right \}.$\\ Moreover, Barry \cite{PB} has shown that for $0\leq k\leq n$ these coefficients satisfy the relation \begin{align} \binom{n}{k}_{[2]}=\sum_{j=0}^{k}\binom kj \binom{n-j}{k}. \end{align} \[ \begin{tabular}{|c|ccccccccccc} \hline $_{\mathbf{n}\backslash \mathbf{k}}$ & \textbf{0} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} & \textbf{6} & \textbf{7} & \textbf{8} & \textbf{9} \\ \hline \textbf{0} & 1\ & & & & & & & & & &\\ \textbf{1} & 1 & 1 & & & & & & & & & \\ \textbf{2} & 1 & 3 & 1 & & & & & & & &\\ \textbf{3} & 1 & 5 & 5 & 1 & & & & & & &\\ \textbf{4} & 1 & 7 & 13 & 7 & 1 & & & & & & \\ \textbf{5} & 1 & 9& 25 & 25 & 9 & 1 & & & & & \\ \textbf{6} & 1 & 11 & 41 & 63 & 41 & 11 & 1 & & & &\\ \textbf{7} & 1 & 13 & 61 & 129 &129 & 61 & 13 & 1 & & &\\ \textbf{8} & 1 & 15 & 85 & 231 & 321 & 231 & 85 & 15 & 1 & &\\ \textbf{9} & 1 & 17 & 113 & 377 & 681 & 681 & 377 & 113 & 17 & 1 & \end{tabular \] \begin{center} \text{Table 1. Tribonacci triangle.\smallskip} \end{center} We find in {\scriptsize{SLOANE}} \cite{S1} A008288 that $\binom{n}{k}_{[2]}$ counts the number of lattice paths from $(0,0)$ to $(n,k)$ using the steps $\lbrace H=(1,0), D=(1,1), L=(2,1) \rbrace$.\\ In what follows $s$ is a positive integer. \subsection{ The s-Pascal triangle} Let $ k \in \lbrace 0,1,\ldots, sn \rbrace$, the bi$^{s}$nomial coefficient $\binom{n}{k}_{s}$ is defined as the $k^{th}$ coefficient in the expansion\\ $$(1+x+x^{2}+\cdots+x^{s})^{n}=\sum_{k \in {\mathbb Z}}\binom{n}{k}_{s}x^{k},$$ with $\binom{n}{k}_{s}=0$ for $k >sn$ or $k < 0$.\\ Using the classical binomial coefficient, see \cite{A,HBB,B}, one has\\ \begin{align} \binom{n}{k}_{s}=\sum_{j_{1}+j_{2}+\cdots+j_{s}=k}\binom{n}{j_{1}}\binom{j_{1}}{j_{2}}\cdots\binom{j_{s-1}}{j_{s}}. \end{align} Some other readily well known established properties are:\\ the symmetry relation\\ \begin{align} \binom{n}{k}_{s}=\binom{n}{sn-k}_{s}, \end{align} the longitudinal recurrence relation \begin{align} \binom{n}{k}_{s}=\sum_{j=0}^{s}\binom{n-1}{k-j}_{s}, \end{align} the diagonal recurrence relation \begin{align} \binom{n}{k}_{s}=\sum_{j=0}^{n}\binom{n}{j}\binom{j}{k-j}_{s-1}, \end{align} and de Moivre expression see \cite{AD,AD1} \begin{align}\label{js} \binom{n}{k}_{s}=\sum_{j=0}(-1)^{j}\binom{n}{j}\binom{k-j(s+1)+n-1}{n-1}. \end{align} Belbachir and Benmezai \cite{bb} gave the following relation \begin{align}\label{bb1} \binom{n}{k}_{s}=(-1)^{k}\sum_{j_{1}+j_{2}+\cdots+j_{s}=k}\binom{n}{j_{1}}\binom{n}{j_{2}}\cdots\binom{n}{j_{s}}a^{-\sum_{r=1}^{s}rj_{r}}, \end{align} where $a=\exp{(2i\pi/(s+1))}$.\\ These coefficients, as for the usual binomial coefficients, are built, as for the Pascal triangle, the called s-Pascal triangle. One can find the first values of the s-Pascal triangle in {\scriptsize{SLOANE}} $A027907$ for $s = 2$, $A008287$ for $s = 3$ and $A035343$ for $s = 4$.\\ \ \[ \begin{tabular}{|c|ccccccccccccccc} \hline $_{\mathbf{n}\backslash \mathbf{k}}$ & \textbf{0} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} & \textbf{6} & \textbf{7} & \textbf{8} & \textbf{9} & \textbf{10}& \textbf{11}& \textbf{12}& \textbf{13} \\ \hline \textbf{0} & 1\ & & & & & & & & & & & & & \\ \textbf{1} & 1 & 1 & 1 & 1 & & & & & & & & & & \\ \textbf{2} & 1 & 2 & 3 & 4 & 3 & 2 & 1 & & & & & & \\ \textbf{3} & 1 & 3 & 6 & 10 & 12 & 12& 10 & 6 & 3 & 1 & & & & \\ \textbf{4} & 1 & 4 & 10 & 20 &31 & 40 & 44 & 40 &31 & 20 & 10& 4 & 1 & \\ \textbf{5} & 1 & 5 & 15 & 35 &65 & 101 & 135 & 155 &155 & 135 & 101& 65 & 35 & 15 \\ \end{tabular \] \begin{align*} \text{Table 2. Biquadranomial triangle ($s=3$).\smallskip} \end{align*} \subsection{The $q$-binomial coefficient} The $q$-analogue of binomial coefficient or the $q$-binomial coefficient $[^{n} _{k}]$ generalizes the binomial coefficient $\binom{n}{k}$ \cite{ht,JC}. It is defined as follows \begin{align*} {n \brack k}=\frac{[n]_{q}!}{[k]_{q}![n-k]_{q}!}q^{\binom k2}, \end{align*} with $ [n]_{q}= 1+q+q^{2}+\cdots+q^{n-1}$ and $[n]_{q}! =[1]_{q}[2]_{q}\cdots[n]_{q}$, we use the convention ${n \brack k}=0$ for $k \notin \lbrace 0,\ldots, n \rbrace$.\\ The $q$-binomial coefficient satisfies the following recurrence relations \begin{align}\label{03} {n \brack k}={n-1 \brack k}+q^{n-1}{n-1 \brack k-1}, \end{align} and \begin{align}\label{04} {n \brack k}=q^{k}{n-1 \brack k}+q^{k-1}{n-1 \brack k-1}. \end{align} And the generating functions are \begin{align}\label{1} \sum_{k=0}^{n}{n \brack k}x^{k}=(1+x)(1+qx)(1+q^{2}x)\cdots(1+q^{n-1}x), \end{align} and \begin{align}\label{2t} \sum_{n\geq0}{n \brack k}x^{n}=\frac{x^{k}q^{\binom{k}{2}}}{(1-x)(1-qx)\cdots(1-q^{k}x)}. \end{align} Belbachir and Benmezai \cite{bb} proposed the $q$-bi$^{s}$nomial coefficient, denoted by ${n \brack k}^{(s)}$, as follows \begin{align}\label{10} {n \brack k}^{(s)}:=(-1)^{k}\sum_{j_{1}+j_{2}+\cdots+j_{s}=k}{n \brack j_{1}}{n \brack j_{2}}\cdots{n \brack j_{s}}a^{-\sum_{r=1}^{s}rj_{r}}. \end{align} The $q$-bi$^{s}$nomial coefficient satisfies the following recurrence relations \begin{align} {n \brack k}^{(s)}=\sum_{j=0}^{s}q^{k-j}{n-1 \brack k-j}^{(s)}, \end{align} \begin{align} {n \brack k}^{(s)}=\sum_{j=0}^{s}q^{(n-1)j}{n-1 \brack k-j}^{(s)}. \end{align} According to (\ref{1}) the generating function is given by \begin{align}\label{gh} \sum_{k=0}^{ns}{n \brack k}^{(s)}x^{k}=\prod_{j=0}^{n-1}(1+q^{j}x+(q^{j}x)^{2}+\cdots+(q^{j}x)^{s}). \end{align} In the first section we introduce the quasi $s$-Pascal triangle by using a family of lattice paths; we establish an explicit formula for the elements of the quasi $s$-Pascal triangle, and we prove that the sums of the elements crossing the diagonal rays yield the terms of $s$-bonacci sequence; we close this section by giving a relation between $s$-Pascal triangle and quasi $s$-Pascal triangle. The second section is devoted to the linear recurrence relation obtained by summing the elements lying over any finite rays of the quasi $s$-Pascal triangle and we give the corresponding generating function. In the third section we give the de Moivre like summation with some other identities. In section four, we establish the $q$-analogue of the elements of the quasi $s$-Pascal triangle. \section{The quasi s-Pascal triangle } In this section we define the quasi $s$-Pascal triangle, we denote by $\binom{n}{k}_{[s]}$ the coefficient in the $n^{th}$ row and $k^{th}$ column of the quasi $s$-Pascal triangle. \begin{definition}\label{fd} The quasi-bi$^{s}$nomial coefficient $\binom{n}{k}_{[s]}$ is defined by the number of lattice path from $(0,0)$ to $(n,k)$ with steps in $\lbrace L=(1,0), L_{1}=(1,1), L_{2}=(2,1),\ldots, L_{s}=(s,1)\rbrace$. With $\binom{n}{0}_{[s]}=\binom{n}{n}_{[s]}=1 \ \text{and the convention} \ \binom{n}{k}_{[s]}=0 \ \text{for} \ k > n $ or $k<0.$ \end{definition} \begin{center} \definecolor{ududff}{rgb}{0.30196078431372547,0.30196078431372547,1.} \begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm] \clip(-3.5,-3) rectangle (17,6); \draw [->,line width=0.4pt] (0.52,5.) -- (1.52,5.6); \draw [->,line width=0.4pt] (2.52,5.) -- (3.98,5.58); \draw [->,line width=0.4pt] (6.,5.) -- (8.52,5.56); \draw [->,line width=0.4pt] (-1.,5.) -- (-0.34,5.58); \draw [->,line width=0.4pt] (-2.64,5.) -- (-2.,5.); \begin{scriptsize} \draw [fill=ududff] (-2.64,5.) circle (1.5pt); \draw [fill=ududff] (-1.,5.) circle (1.5pt); \draw [fill=ududff] (-0.32,4.98) circle (1.5pt); \draw [fill=ududff] (-0.98,5.58) circle (1.5pt); \draw [fill=ududff] (0.52,5.) circle (1.5pt); \draw [fill=ududff] (1.,5.) circle (1.5pt); \draw [fill=ududff] (1.5,4.98) circle (1.5pt); \draw [fill=ududff] (1.52,5.6) circle (1.5pt); \draw [fill=ududff] (1.04,5.58) circle (1.5pt); \draw [fill=ududff] (0.52,5.56) circle (1.5pt); \draw [fill=ududff] (2.52,5.) circle (1.5pt); \draw [fill=ududff] (2.54,5.56) circle (1.5pt); \draw [fill=ududff] (3.04,5.56) circle (1.5pt); \draw [fill=ududff] (3.,5.) circle (1.5pt); \draw [fill=ududff] (3.48,5.) circle (1.5pt); \draw [fill=ududff] (3.5,5.56) circle (1.5pt); \draw [fill=ududff] (3.98,5.58) circle (1.5pt); \draw [fill=ududff] (4.,5.) circle (1.5pt); \draw [fill=ududff] (6.,5.) circle (1.5pt); \draw [fill=ududff] (6.52,4.98) circle (1.5pt); \draw [fill=ududff] (8.,5.) circle (1.5pt); \draw [fill=ududff] (8.52,5.) circle (1.5pt); \draw [fill=ududff] (8.52,5.56) circle (1.5pt); \draw [fill=ududff] (8.04,5.56) circle (1.5pt); \draw [fill=ududff] (6.5,5.56) circle (1.5pt); \draw [fill=ududff] (5.98,5.54) circle (1.5pt); \draw [fill=ududff] (7.34,5.54) circle (0.5pt); \draw [fill=ududff] (7.,5.54) circle (0.5pt); \draw [fill=ududff] (7.64,5.54) circle (0.5pt); \draw [fill=ududff] (7.62,5.) circle (0.5pt); \draw [fill=ududff] (7.36,5.) circle (0.5pt); \draw [fill=ududff] (7.,5.) circle (0.5pt); \draw [fill=ududff] (4.6,4.98) circle (0.5pt); \draw [fill=ududff] (5.,5.) circle (0.5pt); \draw [fill=ududff] (5.4,5.) circle (0.5pt); \draw [fill=ududff] (-0.34,5.58) circle (1.5pt); \draw [fill=ududff] (-2.,5.) circle (1.5pt); \large\draw (-2.7,4.5) node[anchor=north west] { $\text{L}$}; \large\draw (-1,4.5) node[anchor=north west] { $\text{$L_{1}$}$}; \large\draw (.8,4.5) node[anchor=north west] { $\text{$L_{2}$}$}; \large\draw (3,4.5) node[anchor=north west] { $\text{$L_{3}$}$}; \large\draw (7,4.5) node[anchor=north west] { $\text{$L_{s}$}$}; \end{scriptsize} \end{tikzpicture} \vspace*{-7cm} \end{center} \begin{center} Ilustration of possible steps for the coefficient $\binom{n}{k}_{[s]}.$ \end{center} \ \begin{lemma} The quasi-bi$^{s}$nomial coefficient $\binom{n}{k}_{[s]}$ satisfies the following recurrence relation \begin{align}\label{aabc} \binom{n}{k}_{[s]}=\binom{n-1}{k}_{[s]}+\binom{n-1}{k-1}_{[s]}+\binom{n-2}{k-1}_{[s]}+\cdots+\binom{n-s}{k-1}_{[s]}, \end{align} \end{lemma} \begin{proof} By Definition \ref{fd}, the last step of any path is one of $\lbrace L=(1,0), L_{1}=(1,1), L_{2}=(2,1),\ldots, L_{s}=(s,1)\rbrace$, then if the last step is $L=(1,0)$, it remains to enumerate the number of lattice paths from $(0,0)$ to $(n-1,k)$ which is $\binom{n-1}{k}_{[s]}$, or if the last one is $L_{1}=(1,1)$ it remains to enumerate the number of lattice paths from $(0,0)$ to $(n-1,k-1)$ which is $\binom{n-1}{k-1}_{[s]},\ldots,$ if the last step is $L_{s}=(s,1)$ it remains to enumerate the number of lattice paths from $(0,0)$ to $(n-s,k-1)$ which is $\binom{n-s}{k-1}_{[s]}$, considering all possibilities we construct our recurrence. \end{proof} \subsection{Generating function} Here is given the generating function of $\lbrace \binom{n}{k}_{[s]} \rbrace_{n}$; \begin{theorem}\label{aqw} Let $F_{k}(x):=\sum_{n\geq 0}\binom{n}{k}_{[s]} x^{n}$, then $$F_{k}(x)=(1+x+x^{2}+\cdots+x^{s-1})^{k}\frac{x^{k}}{(1-x)^{k+1}}.$$ \end{theorem} \begin{proof} It follows from Relation (\ref{aabc}) that \begin{align*} F_{k}(x)=xF_{k}(x)+xF_{k-1}(x)+x^{2}F_{k-1}(x)+\cdots+x^{s}F_{k-1}(x), \end{align*} repeated applications of this recurrence give the result. \end{proof} \subsection{Binomial coefficients explicit formula} The following result gives an explicit formula for the coefficients of the quasi $s$-Pascal triangle in terms of binomial coefficients and a variant with multinomial coefficients. \begin{theorem}\label{po} The explicit formula for the quasi-bi$^{s}$nomial coefficient is given by \begin{align}\label{T1} \binom{n}{k}_{[s]}=\sum_{j_{1}}\sum_{j_{2}}\cdots\sum_{j_{s-1}}\binom {k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\binom{n-\sum_{i=1}^{s-1}j_{i}}{k}, \end{align} the multinomial version is \begin{align}\label{T2} \binom{n}{k}_{[s]}&=\sum_{k_{1},k_{2},\ldots,k_{s}}\binom {k}{k_{1},k_{2},\ldots,k_{s}}\binom {n+k-\sum_{i=1}^{s}ik_{i}}{k}, \end{align} where $ \binom {k}{k_{1},k_{2},\ldots,k_{s}}=\frac{k!}{k_{1}!k_{2}!\cdots k_{s}!}$ for $k_{1}+k_{2}+,\cdots+k_{s}=k$ and $ \binom {k}{k_{1},k_{2},\ldots,k_{s}}=0$, else. \end{theorem} \begin{proof} For Relation $(\ref{T1})$ we need to prove that \begin{align*} &\sum_{n=0}^{\infty}\sum_{j_{1}}\sum_{j_{2}}\cdots\sum_{j_{s-1}}\binom {k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\binom{n-\sum_{i=1}^{s-1}j_{i}}{k}x^{n}=\left(\frac{x+x^{2}+\cdots+x^{s}}{1-x}\right)^{k}\frac{1}{1-x}.\\ &\text{So}\\ &\sum_{n=0}^{\infty}\sum_{j_{1}}\sum_{j_{2}}\cdots\sum_{j_{s-1}}\binom {k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\binom{n-\sum_{i=1}^{s-1}j_{i}}{k}x^{n}\\ &=\sum_{j_{1}}\binom {k}{j_{1}}x^{j_{1}}\sum_{j_{2}}\binom {j_{1}}{j_{2}}x^{j_{2}}\cdots\sum_{j_{s-1}}\binom{j_{s-2}}{j_{s-1}}x^{j_{s-1}}\sum_{n=0}^{\infty}\binom{n-\sum_{i=1}^{s-1}j_{i}}{k}x^{n-\sum_{i=1}^{s-1}j_{i}}\\ &=\frac{x^{k}}{(1-x)^{k+1}}\sum_{j_{1}}\binom {k}{j_{1}}x^{j_{1}}\sum_{j_{2}}\binom {j_{1}}{j_{2}}x^{j_{2}}\cdots\sum_{j_{s-2}}\binom{j_{s-3}}{j_{s-2}}x^{j_{s-2}}\sum_{j_{s-1}}\binom{j_{s-2}}{j_{s-1}}x^{j_{s-1}}\\ &=\frac{x^{k}}{(1-x)^{k+1}}\sum_{j_{1}}\binom {k}{j_{1}}x^{j_{1}}\sum_{j_{2}}\binom {j_{1}}{j_{2}}x^{j_{2}}\cdots\sum_{j_{s-2}}\binom{j_{s-3}}{j_{s-2}}(x+x^{2})^{j_{s-2}}\\ \vdots \hspace*{-0.3cm}\\ &=\frac{x^{k}}{(1-x)^{k+1}}\sum_{j_{1}}\binom{k}{j_{1}}(x+x^{2}+\cdots+x^{s-1})^{j_{1}}=\left(\frac{x+x^{2}+\cdots+x^{s}}{1-x}\right)^{k}\frac{1}{1-x}. \end{align*} For Relation $(\ref{T2})$ we have \begin{align*} &\sum_{n\geq0}\sum_{k_{1},k_{2},\ldots,k_{s}}\binom {k}{k_{1},k_{2},\ldots,k_{s}}\binom {n-\sum_{i=1}^{s}ik_{i}+k}{k}x^{n}\\ &=\frac{1}{x^{k}}\sum_{k_{1},k_{2},\ldots,k_{s}}\binom {k}{k_{1},k_{2},\ldots,k_{s}}x^{\sum_{i=1}^{s}ik_{i}}\sum_{n\geq0}\binom {n-\sum_{i=1}^{s}ik_{i}+k}{k}x^{n-\sum_{i=1}^{s}ik_{i}+k}\\ &=\frac{1}{(1-x)^{k+1}}\sum_{k_{1},k_{2},\ldots,k_{s}}\binom {k}{k_{1},k_{2},\ldots,k_{s}}x^{\sum_{i=1}^{s}ik_{i}}\\ &=\left(\frac{x+x^{2}+\cdots+x^{s}}{1-x}\right)^{k}\frac{1}{1-x}. \end{align*} \end{proof} \subsection{Link with generalized Dellanoy matrix} Ramirez and Sirvent \cite{RS} propose a generalization of Dellanoy and Pascal Riordan arrays, they denoted by ${\mathcal D}_{m}(n,k)$ the element in the $n^{th}$ row and $k^{th}$ column of the generalized Dellanoy matrix, such that ${\mathcal D}_{m}(n,k)$ satisfy the following recurrence relation \begin{align*} {\mathcal D}_{m}(n+1,k)=a{\mathcal D}_{m}(n+1,k-1)+\sum_{i=0}^{m-1}a_{i+1}{\mathcal D}_{m}(n,k-i), \end{align*} with $k\geq m-1, n\geq 1$ and initial conditions ${\mathcal D}_{m}(0,k)=a^{k}$ and ${\mathcal D}_{m}(n,0)=a_{1}^{n}$.\\ The coefficient ${\mathcal D}_{m}(n,k)$ is given by the following explicit formula \begin{align*} {\mathcal D}_{m}(n,k)=\sum_{j_{1},j_{2},\ldots,j_{m-1}}\binom{n}{j_{1}}\binom{n-j_{1}}{j_{2}}\cdots\binom{n-j_{1}-\cdots-j_{m-2}}{j_{m-1}}\binom{n+k-u}{n}\times\\ \times a_{1}^{j_{1}}a_{2}^{j_{2}}\cdots a_{m-1}^{j_{m-1}}a_{m}^{n-\sum_{i=1}^{m-1}j_{i}}a^{k-u}, \end{align*} where $u=(m-1)(n-j_{1})+\sum_{i=2}^{m-1}(i-m)j_{i}.$\\ The following result gives a relation between ${\mathcal D}_{m}(n,k)$ and $\binom{n}{k}_{[s]}$. \begin{theorem}For $m=s$, $a=1$ and $a_{i}=1$, $i \in \lbrace 1,\ldots,s \rbrace$ \begin{align*} {\mathcal D}_{s}(k,n-k)=\binom{n}{k}_{[s]}. \end{align*} \end{theorem} \begin{proof} For $a=1$ and $a_{i}=1$, $i \in \lbrace 1,\ldots,s \rbrace$ we have \begin{align*} {\mathcal D}_{s}(n,k)=\sum_{j_{1},j_{2},\ldots,j_{s-1}}\binom{n}{j_{1}}\binom{n-j_{1}}{j_{2}}\cdots\binom{n-j_{1}-\cdots-j_{s-2}}{j_{s-1}}\binom{n+k-u}{n},\hspace*{5cm} \end{align*} where $u=(s-1)(n-j_{1})+\sum_{i=2}^{s-1}(i-s)j_{i},$ then \begin{align*} {\mathcal D}_{s}(n,k)=\sum_{j_{1},j_{2},\ldots,j_{s-1}}\binom{n}{n-j_{1}}\binom{n-j_{1}}{n-j_{1}-j_{2}}\cdots\binom{n-j_{1}-\cdots-j_{s-2}}{n-j_{1}-\cdots-j_{s-2}-j_{s-1}}\binom{n+k-u}{n},\hspace*{1cm} \end{align*} we put $ j_{v}^{'}\rightarrow n-\sum_{l=1}^{v}j_{l}, \ v=\lbrace 1,\ldots,s-1 \rbrace $ then $ j_{1}^{'}+j_{2}^{'}+\cdots+j_{s-1}^{'}=u$, \begin{align*} {\mathcal D}_{s}(n,k)=\sum_{j_{1}^{'},j_{2}^{'},\ldots,j_{s-1}^{'}}\binom {n}{j_{1}^{'}}\binom {j_{1}^{'}}{j_{2}^{'}}\cdots \binom{j_{s-2}^{'}}{j_{s-1}^{'}}\binom{n+k-\sum_{i=1}^{s-1}j_{i}^{'}}{n}.\hspace*{6.5cm}\\ \text{Thus} \ {\mathcal D}_{s}(k,n-k)=\binom{n}{k}_{[s]}.\hspace*{13cm} \end{align*} \end{proof} \subsection{Recurrence relation or s-bonacci sequence} Now we establish the recurrence relation for the $s$-bonacci sequence, which is a generalization of Fibonacci sequence. Let $(T_{n,s})_{n}$ be the terms of the $s$-bonacci sequence obtained by summing the elements lying over the principal diagonal rays in the quasi $s$-Pascal triangle.\\ Let be the sequence \begin{align*} T_{n+1,s}:=\sum_{k}\binom{n-k}{k}_{[s]}, \end{align*} with $T_{0,s}=0$. \begin{theorem}\label{aqwd} For $ n \geq 0$, $(T_{n,s})_{n} $ satisfies the following recurrence relation \[ T_{n+1,s}=T_{n,s}+T_{n-1,s}+\cdots+T_{n-s,s}, \] with $T_{1,s}=1,T_{-i,s}=0 \ \text{for} \ i \in \lbrace 0,-1,\ldots,-(s-1) \rbrace$. \end{theorem} It is not else $s$-bonacci sequence. \begin{proof} We have $T_{n+1,s}=\sum_{k}\binom{n-k}{k}_{[s]}$ and by Relation (\ref{aabc}) we obtain\\ $ T_{n+1,s}=\sum_{k}\binom{n-k-1}{k}_{[s]}+\sum_{k}\binom{n-k-1}{k-1}_{[s]}+\cdots+\sum_{k}\binom{n-k-s}{k-1}_{[s]}\\ \hspace*{4.2cm} k^{'}\rightarrow k-1 \\ \hspace*{0.8cm}=\sum_{k}\binom{n-k-1}{k}_{[s]}+\sum_{k^{'}}\binom{n-k^{'}-2}{k^{'}}_{[s]}+\cdots+\sum_{k^{'}}\binom{n-k^{'}-s-1}{k^{'}}_{[s]}\\ \\ \hspace*{0.8cm}=T_{n,s}+T_{n-1,s}+\cdots+T_{n-s,s}. $ \end{proof} For $s=1$ and $s=2$ we obtain the terms of Fibonacci and Tribonacci sequences respectively. \begin{example} For $s=3$ we have the quadrabonacci triangle \[ \begin{tabular}{|c|cccccccccc} \hline $_{\mathbf{n}\backslash \mathbf{k}}$ & \textbf{0} & \textbf{1} & \textbf{2} & \textbf{3} & \textbf{4} & \textbf{5} & \textbf{6} & \textbf{7} & \textbf{8} & \textbf{9} \\ \hline \textbf{0} & 1\ & & & & & & & & & \\ \textbf{1} & 1 & 1 & & & & & & & & \\ \textbf{2} & 1 & 3 & 1 & & & & & & & \\ \textbf{3} & 1 & 6 & 5 & 1 & & & & & & \\ \textbf{4} & 1 & 9 & 15 & 7 & 1 & & & & & \\ \textbf{5} & 1 & 12& 33 & 28 & 9 & 1 & & & & \\ \textbf{6} & 1 & 15 & 60 & 81 & 45 & 11 & 1 & & & \\ \textbf{7} & 1 & 18 & 96 & 189 &66 & 33 & 13 & 1 & & \\ \textbf{8} & 1 & 21 & 141 & 378 & 459 & 281 & 91 & 15 & 1 & \\ \textbf{9} & 1 & 24 & 195 & 675 & 1107 & 946 & 449 & 120 & 17 & 1% \end{tabular \] \vspace*{-0.52cm} \begin{center} \text{Table 2. The quadrabonacci triangle.\smallskip} \end{center} By Relation (\ref{aabc}) the elements of the quadrabonacci triangle ($s=3$) are given by $\binom{n}{0}_{[3]}=\binom{n}{n}_{[3]}=1$, and $$\binom{n}{k}_{[3]}=\binom{n-1}{k}_{[3]}+\binom{n-1}{k-1}_{[3]}+\binom{n-2}{k-1}_{[3]}+\binom{n-3}{k-1}_{[3]}.$$ For $s=3$, $\binom{n}{k}_{[3]}$ counts the number of lattice paths with steps in $(1,0)$, $(1,1)$, $(2,1)$, $(3,1)$ from $(0,0)$ to $(n,k)$, for example for the value $6$ in the quadrabonacci triangle we have the lattice path \vspace*{-0.64cm} \begin{center} \definecolor{ududff}{rgb}{0.30196078431372547,0.30196078431372547,1.} \begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm] \clip(-7.8,-2.7) rectangle (16.24,7.8); \draw [->,line width=0.2pt] (-6.,6.) -- (-5.,7.); \draw [->,line width=0.2pt] (-5.,7.) -- (-4.,7.); \draw [->,line width=0.2pt] (-4.,7.) -- (-3.,7.); \draw [->,line width=0.2pt] (-2.,6.) -- (-1.,6.); \draw [->,line width=0.2pt] (-1.,6.) -- (0.,7.); \draw [->,line width=0.2pt] (0.,7.) -- (1.,7.); \draw [->,line width=0.2pt] (2.,6.) -- (3.,6.); \draw [->,line width=0.2pt] (3.,6.) -- (4.,6.); \draw [->,line width=0.2pt] (4.,6.) -- (5.,7.); \draw [->,line width=0.2pt] (-6.,4.) -- (-5.,4.); \draw [->,line width=0.2pt] (-5.,4.) -- (-3.,5.); \draw [->,line width=0.2pt] (-2.,4.) -- (0.,5.); \draw [->,line width=0.2pt] (0.,5.) -- (1.,5.); \draw [->,line width=0.2pt] (2.,4.) -- (5.,5.); \begin{scriptsize} \draw [fill=ududff] (-6.,7.) circle (1.5pt); \draw [fill=ududff] (-5.,7.) circle (1.5pt); \draw [fill=ududff] (-4.,7.) circle (1.5pt); \draw [fill=ududff] (-3.,7.) circle (1.5pt); \draw [fill=ududff] (-6.,6.) circle (1.5pt); \draw [fill=ududff] (-5.,6.) circle (1.5pt); \draw [fill=ududff] (-4.,6.) circle (1.5pt); \draw [fill=ududff] (-3.,6.) circle (1.5pt); \draw [fill=ududff] (-2.,7.) circle (1.5pt); \draw [fill=ududff] (-2.,6.) circle (1.5pt); \draw [fill=ududff] (-1.,7.) circle (1.5pt); \draw [fill=ududff] (-1.,6.) circle (1.5pt); \draw [fill=ududff] (0.,7.) circle (1.5pt); \draw [fill=ududff] (0.,6.) circle (1.5pt); \draw [fill=ududff] (1.,7.) circle (1.5pt); \draw [fill=ududff] (1.,6.) circle (1.5pt); \draw [fill=ududff] (2.,7.) circle (1.5pt); \draw [fill=ududff] (2.,6.) circle (1.5pt); \draw [fill=ududff] (3.,7.) circle (1.5pt); \draw [fill=ududff] (3.,6.) circle (1.5pt); \draw [fill=ududff] (4.,7.) circle (1.5pt); \draw [fill=ududff] (5.,7.) circle (1.5pt); \draw [fill=ududff] (5.,6.) circle (1.5pt); \draw [fill=ududff] (4.,6.) circle (1.5pt); \draw [fill=ududff] (-6.,5.) circle (1.5pt); \draw [fill=ududff] (-6.,4.) circle (1.5pt); \draw [fill=ududff] (-5.,4.) circle (1.5pt); \draw [fill=ududff] (-5.,5.) circle (1.5pt); \draw [fill=ududff] (-4.,4.) circle (1.5pt); \draw [fill=ududff] (-4.,5.) circle (1.5pt); \draw [fill=ududff] (-3.,4.) circle (1.5pt); \draw [fill=ududff] (-3.,5.) circle (1.5pt); \draw [fill=ududff] (-2.,4.) circle (1.5pt); \draw [fill=ududff] (-2.,5.) circle (1.5pt); \draw [fill=ududff] (-1.,4.) circle (1.5pt); \draw [fill=ududff] (-1.,5.) circle (1.5pt); \draw [fill=ududff] (0.,5.) circle (1.5pt); \draw [fill=ududff] (0.,4.) circle (1.5pt); \draw [fill=ududff] (1.,4.) circle (1.5pt); \draw [fill=ududff] (1.,5.) circle (1.5pt); \draw [fill=ududff] (2.,4.) circle (1.5pt); \draw [fill=ududff] (2.,5.) circle (1.5pt); \draw [fill=ududff] (3.,5.) circle (1.5pt); \draw [fill=ududff] (3.,4.) circle (1.5pt); \draw [fill=ududff] (4.,5.) circle (1.5pt); \draw [fill=ududff] (4.,4.) circle (1.5pt); \draw [fill=ududff] (5.,5.) circle (1.5pt); \draw [fill=ududff] (5.,4.) circle (1.5pt); \end{scriptsize} \end{tikzpicture} \end{center} \vspace*{-5.8cm} \begin{center} Title: Illustration of possible paths from $(0,0)$ to $(3,1)$ using the steps\\ $(1,0), (1,1), (2,1), (3,1).$ \end{center} \end{example} Notice that for $s=1$ and $s=2$ the obtained triangles are symmetric, unlike the cases where $s > 2$.\\ \subsection{$s$-Pascal triangle versus quasi-$s$-Pascal triangle} The following result establishes the relation between the quasi $s$-Pascal triangle and $s$-Pascal triangle \begin{theorem}\label{azer} For fixed non negative integers $n, k$ and $s$, we have \begin{align*} \binom{n}{k}_{[s]}=\sum_{i}\binom{n-i}{k}\binom{k}{i}_{s-1}. \end{align*} \end{theorem} \begin{proof} We have \begin{align*} \binom{n}{k}_{[s]}=\sum_{j_{1}}\sum_{j_{2}}\cdots\sum_{j_{s-1}}\binom {k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\binom{n-j_{1}-j_{2}-\cdots-j_{s-1}}{k},\hspace*{1.5cm}\\ \text{considering the summations by blocks} \ j_{1}+j_{2}+\cdots+j_{s-1}=i \ \text{we get}\hspace*{2cm}\\ \binom{n}{k}_{[s]}=\sum_{i}\binom{n-i}{k}\sum_{j_{1}+j_{2}+\cdots+j_{s-1}=i}\binom {k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}=\sum_{i}\binom{n-i}{k}\binom{k}{i}_{s-1}.\hspace*{0.3cm}\\ \end{align*} \end{proof} \vspace*{-1.95cm} \section{Linear recurrence relation and generating function associated to finite transversals of the quasi s-Pascal triangle} This section is devoted to establish a recurrence relation associated to the sums of the elements lying over the transversals of direction $(\alpha,r)$ in the quasi $s$-Pascal triangle. In \cite{bzk} we find the details about the concept of direction in Pascal triangle. The study was extended for the arithmetic triangle, see \cite{ABR,BS1}. We generalize the concept to our triangle as an extension of Theorem \ref{aqwd} to the case where $r\in {\mathbb Z},\ \alpha \in {\mathbb N}$, $\beta \in{\mathbb Z}^{+}$ with $0\leq \beta<\alpha$ and $r+\alpha>0$. This corresponds to the finite sequences lying over finite transversals of the quasi $s$-Pascal triangle.\\ Let be the sequence \begin{align*} T_{n+1,s}^{(\alpha,\beta,r)}:=\sum_{k}\binom{n-rk}{\beta +\alpha k}_{[s]}, \ \text{with} \ T_{0,s}^{(\alpha,\beta,r)}=0. \end{align*} The following figure illustrate the direction $(\alpha,r)=(2,1)$ and $\beta=0$ on the Tribonacci triangle. \begin{center} \begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.0cm,y=1.0cm] \clip(-6.405887445887451,6.667012987013018) rectangle (10.78891774891774,16.121558441558467); \draw (3.8971428571428497,10.545800865800894) node[anchor=north west] {$1$}; \draw (-3.1331601731601793,15.446233766233792) node[anchor=north west] {$1$}; \draw (-3.1331601731601793,14.788225108225136) node[anchor=north west] {$1$}; \draw (-0.120173160173167,13.420259740259768) node[anchor=north west] {$1$}; \draw (-3.1331601731601793,14.060952380952408) node[anchor=north west] {${1}$}; \draw (-2.111515151515158,14.095584415584444) node[anchor=north west] {${3}$}; \draw (-3.115844155844162,13.437575757575786) node[anchor=north west] {${1}$}; \draw (-2.128831168831175,14.82285714285717) node[anchor=north west] {${1}$}; \draw (-2.111515151515158,13.437575757575786) node[anchor=north west] {${5}$}; \draw (-1.0898701298701365,13.437575757575786) node[anchor=north west] {${5}$}; \draw (2.910129870129863,11.325021645021673) node[anchor=north west] {${1}$}; \draw (-1.1937662337662405,12.814199134199162) node[anchor=north west] {${13}$}; \draw (-0.120173160173167,12.848831168831197) node[anchor=north west] {${7}$}; \draw (0.8841558441558371,12.866147186147215) node[anchor=north west] {${1}$}; \draw (-2.111515151515158,12.796883116883144) node[anchor=north west] {${7}$}; \draw (-3.115844155844162,12.76225108225111) node[anchor=north west] {${1}$}; \draw (-1.2110822510822576,12.069610389610418) node[anchor=north west] {${25}$}; \draw (-2.128831168831175,12.069610389610418) node[anchor=north west] {${9}$}; \draw (0.8668398268398199,12.069610389610418) node[anchor=north west] {${9}$}; \draw (-0.25,12.069610389610418) node[anchor=north west] {${25}$}; \draw (-3.1331601731601793,12.034978354978383) node[anchor=north west] {${1}$}; \draw (1.9058008658008585,12.069610389610418) node[anchor=north west] {${1}$}; \draw (1.8365367965367894,10.545800865800894) node[anchor=north west] {${61}$}; \draw (0.7975757575757506,11.307705627705657) node[anchor=north west] {${41}$}; \draw (0.7283116883116814,10.563116883116912) node[anchor=north west] {${129}$}; \draw (-3.115844155844162,11.290389610389639) node[anchor=north west] {${1}$}; \draw (-1.21,10.58043290043293) node[anchor=north west] {${61}$}; \draw (-0.33,10.563116883116912) node[anchor=north west] {${129}$}; \draw (1.819220779220772,11.325021645021673) node[anchor=north west] {${11}$}; \draw (-0.25,11.325021645021673) node[anchor=north west] {${63}$}; \draw (-1.21,11.325021645021673) node[anchor=north west] {${41}$}; \draw (-2.1980952380952443,11.307705627705657) node[anchor=north west] {${11}$}; \draw (-3.098528138528145,10.563116883116912) node[anchor=north west] {${1}$}; \draw (-1.1071861471861537,14.095584415584444) node[anchor=north west] {${1}$}; \draw (-2.1980952380952443,10.597748917748946) node[anchor=north west] {${13}$}; \draw (2.806233766233759,10.545800865800894) node[anchor=north west] {${13}$}; \draw [->,line width=0.1pt] (-2.7522077922077974,15.22112554112557) -- (-1.1764502164502222,15.879134199134228); \draw [->,line width=0.1pt] (-2.7695238095238146,14.511168831168861) -- (-1.1071861471861526,15.221125541125572); \draw [->,line width=0.1pt] (-0.7,13.85)--(0.7695238095238146,14.511168831168861) ; \draw [->,line width=0.1pt] (1.25,12.55)--(2.5,13.1) ; \draw [->,line width=0.1pt] (-2.8387878787878833,13.80121212121212) -- (-1.141818181818187,14.511168831168861); \draw [->,line width=0.1pt] (-0.7435497835497893,13.177835497835527) -- (0.7975757575757516,13.853160173160203); \draw [->,line width=0.1pt] (-0.65,12.5) -- (0.8841558441558381,13.14); \draw [->,line width=0.1pt] (-2.734891774891781,12.5) -- (-1.1,13.1); \draw [->,line width=0.1pt] (-2.769523809523816,11.740606060606092) -- (-1.2457142857142922,12.346666666666698); \draw [->,line width=0.1pt] (-2.821471861471868,11.065281385281416) -- (-1.228398268398275,11.671341991342022); \draw [->,line width=0.1pt] (-0.65,11.85) -- (0.8495238095238024,12.415930735930766); \draw [->,line width=0.1pt] (-2.8041558441558503,13.10857142857146) -- (-1.124502164502171,13.731948051948084); \draw (-3.115844155844162,9.87047619047622) node[anchor=north west] {${1}$}; \draw (-2.180779220779227,9.905108225108254) node[anchor=north west] {${15}$}; \draw (-1.21,9.922424242424272) node[anchor=north west] {${85}$}; \draw (-0.3,9.922424242424272) node[anchor=north west] {${231}$}; \draw (0.72,9.905108225108254) node[anchor=north west] {${321}$}; \draw (1.75,9.922424242424272) node[anchor=north west] {${231}$}; \draw (3.87,9.9) node[anchor=north west] {${85}$}; \draw (2.82,9.922424242424272) node[anchor=north west] {${15}$}; \draw (4.884155844155837,9.887792207792238) node[anchor=north west] {${1}$}; \draw [->,line width=0.1pt] (-2.7522077922077983,10.268744588744617) -- (-1.2457142857142922,10.857489177489207); \draw [->,line width=0.1pt] (-0.7089177489177556,11.08259740259743) -- (0.8148917748917679,11.67134199134202); \draw [->,line width=0.1pt] (-2.7002597402597464,9.61073593073596) -- (-1.228398268398275,10.13021645021648); \draw [->,line width=0.1pt] (-0.7089177489177556,10.32069264069267) -- (0.7283116883116814,10.857489177489207); \draw [->,line width=0.1pt] (1.35,11.05) -- (2.875497835497828,11.67134199134202); \draw [->,line width=0.1pt] (1.2304761904761834,11.80987012987016) -- (2.754285714285707,12.38129870129873); \end{tikzpicture} \end{center} \vspace*{-3cm} \begin{center} Figure 1: Tribonacci triangle. \end{center} \begin{theorem}\label{s} For $n \geq \alpha s+r$, $ (T_{n+1,s}^{(\alpha,\beta,r)})_{n}$ satisfies the following linear recurrence relation \begin{align}\label{asc} \sum^{\alpha}_{i=0}(-1)^{i}\binom{\alpha}{i}T_{n-i,s}^{(\alpha,\beta,r)}=\sum^{\alpha(s-1)}_{i=0}\binom{\alpha}{i}_{s-1}T_{n-\alpha-r-i,s}^{(\alpha,\beta,r)}, \end{align} we can recover the initial conditions ${\mathcal T}_{1,s}^{(\alpha,\beta,r)},\ldots,{\mathcal T}_{\alpha s+r-1,s}^{(\alpha,\beta,r)}$ by $ \sum_{k}\binom{n-rk}{\beta +\alpha k}_{[s]} $. \end{theorem} For the proof we need the following Lemma. \begin{lemma}[\cite{bzk}] Let $a$,$b$ and $\alpha$ be non negative integers satisfying the conditions $\alpha\leq a$, then \begin{align}\label{asc1} \sum^{\alpha}_{i=0}(-1)^{i}\binom{\alpha}{i}\binom{a-i}{b}=\binom{a-\alpha}{b-\alpha}. \end{align} \end{lemma} \begin{proof} Of Theorem \ref{s}, from Relations (\ref{T1}) and (\ref{asc1}), we get \begin{align*} &\sum^{\alpha}_{i=0}(-1)^{i}\binom{\alpha}{i}T_{n-i,s}^{(\alpha,\beta,r)} \\ &=\sum^{\alpha}_{i=0}(-1)^{i}\binom{\alpha}{i}\sum_{k}\binom{n-rk-i-1}{\beta +\alpha k}_{[s]}\\ &=\sum^{\alpha}_{i=0}(-1)^{i}\binom{\alpha}{i}\sum_{k}\sum_{j_{1},j_{2},\ldots, j_{s-1}}\binom {\beta+\alpha k}{j_{1}}\binom {j_{1}}{j_{2}}\ldots \binom{j_{s-2}}{j_{s-1}}\binom{n-rk-i-\sum_{i=1}^{s-1}j_{i}-1}{\beta+\alpha k}\\ &=\sum_{k}\sum_{j_{1},j_{2},\ldots, j_{s-1}}\binom {\beta+\alpha k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\sum^{\alpha}_{i=0}(-1)^{i}\binom{\alpha}{i}\binom{n-rk-i-\sum_{i=1}^{s-1}j_{i}-1}{\beta+\alpha k}\\ &=\sum_{k}\sum_{j_{1},j_{2},\ldots, j_{s-1}}\binom {\beta+\alpha k}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\binom{n-rk-\alpha-\sum_{i=1}^{s-1}j_{i}-1}{\beta+\alpha(k-1)}\\ &\quad (k^{'}\rightarrow k-1)\\ &=\sum_{k^{'}}\sum_{j_{1},j_{2},\ldots, j_{s-1}}\binom {\beta+\alpha k^{'}+\alpha}{j_{1}}\binom {j_{1}}{j_{2}}\cdots \binom{j_{s-2}}{j_{s-1}}\binom{n-rk^{'}-r-\alpha-\sum_{i=1}^{s-1}j_{i}-1}{\beta+\alpha k^{'}}\\ &=\sum_{k^{'}}\sum_{j_{1},j_{2},\ldots, j_{s-1}}\binom {\beta+\alpha k^{'}+\alpha}{j_{1}}\binom {j_{1}-i_{1}+i_{1}}{j_{2}}\cdots \binom{j_{s-2}-i_{s-2}+i_{s-2}}{j_{s-1}}\times\\ &\times\binom{n-rk^{'}-r-\alpha-\sum_{i=1}^{s-1}j_{i}-1}{\beta+\alpha k^{'}}\\ &\text{by Vandermonde Formula}\\ &=\sum_{k^{'}}\sum_{j_{1},j_{2},\ldots, j_{s-1}}\sum_{i_{1},i_{2},\ldots, i_{s-1}}\binom {\alpha}{i_{1}}\binom {\beta+\alpha k^{'}}{j_{1}-i_{1}}\binom {i_{1}}{i_{2}}\binom{j_{1}-i_{1}}{j_{2}-i_{2}}\cdots \binom{i_{s-2}}{i_{s-1}}\binom{j_{s-2}-i_{s-2}}{j_{s-1}-i_{s-1}}\times\\ &\times\binom{n-rk^{'}-r-\alpha-\sum_{i=1}^{s-1}j_{i}-1}{\beta+\alpha k^{'}}\\ & (l_{v}\rightarrow j_{v}-i_{v})\\ &=\sum_{k^{'}}\sum_{l_{1},l_{2},\ldots, l_{s-1}}\sum_{i_{1},i_{2},\ldots, i_{s-1}}\binom {\alpha}{i_{1}}\binom {\beta+\alpha k^{'}}{l_{1}}\binom {i_{1}}{i_{2}}\binom{l_{1}}{l_{2}}\cdots \binom{i_{s-2}}{i_{s-1}}\binom{l_{s-2}}{l_{s-1}}\times\\ &\times\binom{n-rk^{'}-r-\alpha-\sum_{j=1}^{s-1}i_{j}-\sum_{j=1}^{s-1}l_{j}-1}{\beta+\alpha k^{'}}\\ &\text{we take the summation as by block} \ i_{1}+i_{2}+\cdots+i_{s-1}=i \\ \end{align*} \begin{align*} &=\sum_{k^{'}}\sum_{i}\sum_{l_{1},l_{2},\ldots, l_{s-1}}\sum_{i_{1}+i_{2}+\cdots +i_{s-1}=i}\binom {\alpha}{i_{1}}\binom {\beta+\alpha k^{'}}{l_{1}}\binom {i_{1}}{i_{2}}\binom{l_{1}}{l_{2}}\cdots \binom{i_{s-2}}{i_{s-1}}\binom{l_{s-2}}{l_{s-1}}\times\\ &\times\binom{n-rk^{'}-r-\alpha-i-\sum_{j=1}^{s-1}l_{j}-1}{\beta+\alpha k^{'}}\\ &=\sum_{k^{'}}\sum_{i}\sum_{l_{1},l_{2},\ldots, l_{s-1}}\binom {\beta+\alpha k^{'}}{l_{1}}\binom{l_{1}}{l_{2}}\cdots \binom{l_{s-2}}{l_{s-1}}\binom{n-rk^{'}-r-\alpha-i-\sum_{j=1}^{s-1}l_{i}-1}{\beta+\alpha k^{'}}\times\\ &\times\sum_{i}\binom {\alpha}{i_{1}}\binom {i_{1}}{i_{2}}\binom {i_{1}}{i_{2}}\cdots\binom{i_{s-2}}{i_{s-1}}\hspace*{7cm}\\ &=\sum^{\alpha(s-1)}_{i=0}\binom{\alpha}{i}_{s-1}{\mathcal T}_{n-\alpha-r-i,s}^{(\alpha,\beta,r)}. \end{align*} \end{proof} For $\alpha=1,r=1,\beta=0$ we obtain the terms of $s$-bonacci sequence, for $s=2$, we obtain Theorem $7$ of \cite{AB}. \\ \begin{example} For $\alpha=2,r=1,\beta=0$ and $n\geq 2s+1$ we have the following recurrence relation \begin{align*} T_{n,s}^{(2,0,1)}&=\sum_{j=1}^{2}(-1)^{j+1}\binom{2}{j}T_{n-j,s}^{(2,0,1)}+\sum_{j=0}^{2(s-1)}\binom{2}{j}_{s-1}T_{n-j-3,s}^{(2,0,1)}\\ &=2T_{n-1,s}^{(2,0,1)}-T_{n-2,s}^{(2,0,1)}+T_{n-3,s}^{(2,0,1)}+2T_{n-4,s}^{(2,0,1)}+\cdots\\ &\cdots+sT_{n-s-2,s}^{(2,0,1)}+\cdots+2T_{n-2s,s}^{(2,0,1)}+T_{n-2s-1,s}^{(2,0,1)}, \end{align*} as $\binom{2}{s-1}_{s-1}=s$. \end{example} The following result establishes the generating function for the sequence $({\mathcal T}_{n,s}^{(\alpha,\beta,r)})_{n}$ for quasi $s$-Pascal triangles. \begin{theorem} The generating function of the sequence $\lbrace{\mathcal T}_{n,s}^{(\alpha,\beta,r)}\rbrace_{n\geq0}$ is given by \begin{align*} \sum_{n\geqslant0}{\mathcal T}_{n+1,s}^{(\alpha,\beta,r)}x^{n}=\frac{(1-x)^{\alpha-\beta-1}(x+x^{2}+\cdots+x^{s})^{\beta}}{(1-x)^{\alpha}-x^{r+\alpha}(1+x+\cdots+x^{s-1})^{\alpha}}. \end{align*} \end{theorem} \begin{proof} We have \begin{align*} \sum_{n\geqslant0}{\mathcal T}_{n+1,s}^{(\alpha,\beta,r)}x^{n}&=\sum_{n\geqslant0}\sum_{k}\binom{n-rk}{\beta +\alpha k}_{[s]}x^{n}\\ &=\sum_{n\geqslant qk}\sum_{k}\binom{n-rk}{\beta +\alpha k}_{[s]}x^{n-rk}x^{rk}\\ &=\sum_{k}\frac{(x+x^{2}+\cdots+x^{s})^{\beta+\alpha k}x^{rk}}{(1-x)^{\beta+\alpha k+1}}\\ &=\frac{(x+x^{2}+\cdots+x^{s})^{\beta}}{(1-x)^{\beta+1}}\sum_{k}\left( \frac{(x+x^{2}+\cdots+x^{s})^{\alpha}x^{r}}{(1-x)^{\alpha}}\right)^{k}\\ &=\frac{(x+x^{2}+\cdots+x^{s})^{\beta}}{(1-x)^{\beta+1}}\frac{1}{1-(\frac{x^{r}(x+x^{2}+\cdots+x^{s})^{\alpha}}{(1-x)^{\alpha}})}\\ &=\frac{(1-x)^{\alpha-\beta-1}(x+x^{2}+\cdots+x^{s})^{\beta}}{(1-x)^{\alpha}-x^{r+\alpha}(1+x+\cdots+x^{s-1})^{\alpha}}. \end{align*} \end{proof} \section{The de Moivre summation and other nested sums} Butler and Karasik see \cite{BK} showed how the binomial coefficient can be written as nested sums, in this section we establish an identity for the quasi bi$^{s}$nomial coefficients $\binom{n}{k}_{[s]}$ equivalent to the de Moivre summation for bi$^{s}$nomial coefficient and we give some other nested sums for the coefficient $\binom{n}{k}_{[s]}$. The following identity is important in the sense that it gives a simple summation with a product of two binomials. \begin{theorem} The following identity holds true \begin{align} \label{917} \binom{n}{k}_{[s]}=\sum_{j}(-1)^{j}\binom kj\binom{n+k-sj}{2k}. \end{align} \end{theorem} \begin{proof} By Theorem \ref{aqw} we have \begin{align*} \sum_{n\geq0}\binom{n}{k}_{[s]}x^{n}&=\frac{x^{k}(1+x+x^{2}+\cdots+x^{s-1})^{k}}{(1-x)^{k+1}}\\ &=x^{k}(1-x^{s})^{k}\frac{1}{(1-x)^{2k+1}}\\ &=x^{k}\sum_{j}(-1)^{j}\binom kj x^{js}\sum_{i}\binom{i+2k}{2k}x^{i}\\ &=\sum_{n}\sum_{i+sj=n}(-1)^{j}\binom kj\binom{i+2k}{2k}x^{n+k}\\ &=\sum_{n}\sum_{j}(-1)^{j}\binom kj\binom{n-sj+k}{2k}x^{n}.\\ \end{align*} Identity (\ref{917}) is a dual version of Relation (\ref{js}) \end{proof} Now we give an identity for $\binom{n}{k}_{[s]}$ dual to Relation (\ref{bb1}). \begin{theorem}\label{b4} For $w=\exp{(2i\pi/s)}$ we have \begin{align*} \binom{n}{k}_{[s]}=\sum_{j}\binom{n-j}{k}(-1)^{j}\sum_{k_{1}+k_{2}+\cdots+k_{s-1}=j}\binom{k}{k_{1}}\binom{k}{k_{2}}\cdots\binom{k}{k_{s-1}}\times w^{-\sum_{r=1}^{s-1}rk_{r}}. \end{align*} \end{theorem} \begin{proof}By Theorem \ref{aqw}, we have \begin{align*} &\sum_{n\geq0}\binom{n}{k}_{[s]}x^{n}\\ &=\frac{x^{k}}{(1-x)^{k+1}}\prod_{j=1}^{s-1}(x-w^{j})^{k}\\ &=\sum_{n\geq0}\binom{n}{k}x^{n}\sum_{k_{1}}(-1)^{k-k_{1}}\binom{k}{k_{1}}w^{k-k_{1}}x^{k_{1}} \sum_{k_{2}}(-1)^{k-k_{2}}\binom{k}{k_{2}}w^{2(k-k_{2})}x^{k_{2}}\cdots\\ &\cdots\sum_{k_{s-1}}(-1)^{k-k_{s-1}}\binom{k}{k_{s-1}}w^{(s-1)(k-k_{s-1})}x^{k_{s-1}}\\ \end{align*} \begin{align*} &=\sum_{n\geq0}\binom{n}{k}x^{n}\sum_{j\geq0}\sum_{k_{1}+k_{2}+\cdots+k_{s-1}=j}\binom{k}{k_{1}}\binom{k}{k_{2}}\cdots\binom{k}{k_{s-1}}\times(-1)^{(s-1)k-j}{w^{\sum_{r=1}^{s-1}r(k-k_{r})}x^{j}}\\ &=\sum_{n\geq0}\sum_{j}\binom{n-j}{k}\sum_{k_{1}+k_{2}+\cdots+k_{s-1}=j}\binom{k}{k_{1}}\binom{k}{k_{2}}\cdots\binom{k}{k_{s-1}}\times(-1)^{(s-1)k-j}w^{\sum_{r=1}^{s-1}r(k-k_{r})}x^{n}.\\ &=\sum_{n\geq0}\sum_{j}\binom{n-j}{k}\sum_{k_{1}+k_{2}+\cdots+k_{s-1}=j}\binom{k}{k_{1}}\binom{k}{k_{2}}\cdots\binom{k}{k_{s-1}}(-1)^{j}w^{-\sum_{r=1}^{s-1}rk_{r}}x^{n},\\ &\text{This yields the result}. \end{align*} \end{proof} \begin{remark}We can also deduce Theorem \ref{b4} using Theorem \ref{azer} and Identity (\ref{bb1}). \end{remark} Using the generating function we obtain the following beautiful nested relation.\\ \begin{theorem} The terms of the $s$-Pascal triangle satisfy the following identity \begin{align*} \binom{n}{k}_{[s]}=\sum_{j_{1},j_{2},\ldots,j_{s-1}}\binom{k}{j_{1}}\binom{j_{1}}{j_{2}}\cdots\binom{j_{s-2}}{j_{s-1}}\binom{n-k-\sum_{i=1}^{s-2}j_{i}}{j_{s-1}}\times(2)^{j_{1}}(3/2)^{j_{2}}\cdots(s/s-1)^{j_{s-1}}. \end{align*} \end{theorem} \begin{proof}We have \begin{align*} &\sum_{n\geq0}\sum_{j_{1},j_{2},\ldots,j_{s-1}}\binom{k}{j_{1}}\binom{j_{1}}{j_{2}}\cdots\binom{j_{s-2}}{j_{s-1}}\binom{n-k-\sum_{i=1}^{s-2}j_{i}}{j_{s-1}}(2)^{j_{1}}(3/2)^{j_{2}}\cdots(s/s-1)^{j_{s-1}}x^{n}\\ &=x^{k}\sum_{j_{1}}\binom{k}{j_{1}}(2x)^{j_{1}}\sum_{j_{2}}\binom{j_{1}}{j_{2}}(3x/2)^{j_{2}}\cdots\sum_{j_{s-2}}\binom{j_{s-3}}{j_{s-2}}(x(s-1)/(s-2))^{j_{s-2}}\times\\ &\times\sum_{j_{s-1}}\binom{j_{s-2}}{j_{s-1}}(s/s-1)^{j_{s-1}}\sum_{n\geq0}\binom{n-k-\sum_{i=1}^{s-2}j_{i}}{j_{s-1}}x^{n-k-\sum_{i=1}^{s-2}j_{i}}\\ &=\frac{x^{k}}{1-x}\sum_{j_{1}}\binom{k}{j_{1}}(2x)^{j_{1}}\sum_{j_{2}}\binom{j_{1}}{j_{2}}(3x/2)^{j_{2}}\cdots\sum_{j_{s-2}}\binom{j_{s-3}}{j_{s-2}}(x(s-1)/(s-2))^{j_{s-2}}\times\\ &\times\sum_{j_{s-1}}\binom{j_{s-2}}{j_{s-1}}\left( \frac{(s/(s-1))x}{1-x}\right)^{j_{s-1}}\\ &=\frac{x^{k}}{1-x}\sum_{j_{1}}\binom{k}{j_{1}}(2x)^{j_{1}}\sum_{j_{2}}\binom{j_{1}}{j_{2}}(3x/2)^{j_{2}}\cdots\\ &\cdots\sum_{j_{s-2}}\binom{j_{s-3}}{j_{s-2}}\left(\frac{x(s-1)/(s-2)+x^{2}/(s-2)}{1-x}\right)^{j_{s-2}}\\ &=\frac{x^{k}}{1-x}\sum_{j_{1}}\binom{k}{j_{1}}(2x)^{j_{1}}\sum_{j_{2}}\binom{j_{1}}{j_{2}}(3x/2)^{j_{2}}\cdots\\ &\cdots\sum_{j_{s-3}}\binom{j_{s-4}}{j_{s-3}}\left(\frac{(x(s-2)+x^{2}+x^{3})/(s-3)}{1-x}\right)^{j_{s-3}}\\ \ &\vdots \\ &=\frac{x^{k}}{1-x}\sum_{j_{1}}\binom{k}{j_{1}}\left(\frac{2x+x^{2}+\cdots +x^{s-1}}{1-x}\right)^{j_{1}}\\ &=\frac{x^{k}(1+x+x^{2}+\cdots+x^{s-1})^{k}}{(1-x)^{k+1}}. \end{align*} \end{proof} \section{The $q$-analogue of the quasi s-Pascal triangle} In this section we define the $q$-analogue of the quasi $s$-Pascal triangle, we denote by ${n \brack k}_{[s]}$ these coefficients; for that, we give an explicit formula and generating function of ${n \brack k}_{[s]}$, and finally we propose a $q$-deformation for the $s$-bonacci sequence. \begin{definition}\label{02} We define the $q$-quasi-bi$^{s}$nomial coefficient, according to Relation (\ref{aabc}), as \begin{align}\label{bn1} {n \brack k}_{[s]}={n-1 \brack k}_{[s]}+\sum_{j=1}^{s}q^{n-j}{n-j \brack k-1}_{[s]},\hspace*{0.6cm} \end{align} or equivalently \begin{align}\label{bn2} {n \brack k}_{[s]}=q^{k}{n-1 \brack k}_{[s]}+\sum_{j=1}^{s}q^{(k-1)j}{n-j \brack k-1}_{[s]}. \end{align} We use the convention ${0 \brack 0}_{[s]}=1 $ and ${n \brack k}_{[s]}=0$ for $k \notin \lbrace 0,\ldots,n \rbrace$. \end{definition} \begin{remark} For $s=1$ we obtain Relations (\ref{03}) and (\ref{04}) respectively. \end{remark} The generating function of ${n \brack k}_{[s]}$ is given by \begin{theorem}\label{79} Let $\mathbb{F}_{k}(x):=\sum_{n\geq0}{n \brack k}_{[s]}x^{n}$ the generating function of the $q$-quasi bi$^{s}$nomial coefficient, then \begin{align*} \mathbb{F}_{k}(x)=\frac{x^{k}q^{\binom{k}{2}}\prod_{j=0}^{k-1}(1+q^{j}x+(q^{j}x)^{2}+\cdots+(q^{j}x)^{s-1})}{\prod_{j=0}^{k}(1-q^{j}x)}. \end{align*} \end{theorem} \begin{proof} We have $\mathbb{F}_{k}(x)=\sum_{n\geq0}{n \brack k}_{[s]}x^{n}$, then using Relation (\ref{bn1}) \begin{align*} \mathbb{F}_{k}(x)&=\sum_{n\geq0}{n-1 \brack k}_{[s]}x^{n}+\sum_{n\geq0}\sum_{j=1}^{s}q^{n-j}{n-j \brack k-1}_{[s]}x^{n}\\ &=x\sum_{n\geq0}{n \brack k}_{[s]}x^{n}+(x+x^{2}+\cdots+x^{s})\sum_{n\geq0}{n \brack k-1}_{[s]}(qx)^{n},\\ \end{align*} thus \begin{align}\label{2210} \frac{(1-x)\mathbb{F}_{k}(x)}{(x+x^{2}+\cdots+x^{s})}=\sum_{n\geq0}{n \brack k-1}_{[s]}(qx)^{n}, \end{align} on the other side, using Relation (\ref{bn1}) again \begin{align*} &\frac{(1-x)\mathbb{F}_{k}(x)}{(x+x^{2}+\cdots+x^{s})}\\ &=\sum_{n\geq0}{n-1 \brack k-1}_{[s]}(qx)^{n}+\sum_{n\geq0}q^{n-1}{n-1 \brack k-2}_{[s]}(qx)^{n}+\cdots+\sum_{n\geq0}q^{n-s}{n-s \brack k-2}_{[s]}(qx)^{n}\\ &=qx\sum_{n\geq0}{n \brack k-1}_{[s]}(qx)^{n}+(qx+(qx)^{2}+\cdots+(qx)^{s})+\sum_{n\geq0}{n \brack k-2}_{[s]}(q^{2}x)^{n}, \end{align*} and by Relation (\ref{2210}) we obtain \begin{align} \frac{(1-x)(1-qx)\mathbb{F}_{k}(x)}{(x+x^{2}+\cdots+x^{s})(qx+(qx)^{2}+\cdots+(qx)^{s})}=\sum_{n\geq0}{n \brack k-2}_{[s]}(q^{2}x)^{n}, \end{align} we repeat the process and get\\ \begin{align*} \frac{(1-x)(1-qx)\cdots(1-q^{k-1}x)\mathbb{F}_{k}(x)}{\prod_{j=0}^{k-1}(q^{j}x+(q^{j}x)^{2}+\cdots+(q^{j}x)^{s}}&=\sum_{n\geq0}{n \brack 0}_{[s]}(q^{k}x)^{n}\\ &=\frac{1}{(1-q^{k}x)}, \end{align*} and finally we conclude to the result.\\ \end{proof} \begin{remark} With the same method we can also prove Theorem \ref{79}, using Relation (\ref{bn2}). \end{remark} The following result establishes the explicit formula of the $q$-quasi-bi$^{s}$nomial coefficient \begin{theorem}\label{bn} The coefficient ${n \brack k}_{[s]}$ satisfy \begin{align} {n \brack k}_{[s]}=\sum_{j}{n-j \brack k}{k \brack j}^{(s-1)}. \end{align} \end{theorem} \begin{proof} We have \begin{align*} \sum_{n\geq0}\sum_{j}{n-j \brack k}{k \brack j}^{(s-1)}x^{n}&=\sum_{j}{k \brack j}^{(s-1)}x^{j}\sum_{n\geq0}{n-j \brack k}x^{n-j}\\ &=\frac{x^{k}q^{\binom{k}{2}}\prod_{j=0}^{k-1}(1+q^{j}x+(q^{j}x)^{2}+\cdots+(q^{j}x)^{s-1})}{\prod_{j=0}^{k}(1-q^{j}x)}, \end{align*} this last equality comes from Relation (\ref{2t}) and Relation (\ref{gh}). \end{proof} Cigler and Carlitz propose the $q$-analogue of the Fibonacci sequence see \cite{CL1,JC1}, the following result establish the recurrence relation for the $q-$analogue of the $s-$bonacci sequence, this generalize s Theorem \ref{aqwd}. \begin{theorem} Let $\mathbb{T}_{n+1}^{(s)}(x):=\sum_{k}{n-k \brack k}_{[s]}x^{k}$ for $n\geq 0$ and $\mathbb{T}_{0}^{(s)}(x)=0$ then\\ \begin{align}\label{es} \mathbb{T}_{n+1}^{(s)}(x)=\mathbb{T}_{n}^{(s)}(x)+x\sum_{j=1}^{s}q^{n-j-1}\mathbb{T}_{n-j}^{(s)}(x/q), \end{align} and \begin{align}\label{es1} \mathbb{T}_{n+1}^{(s)}(x)=\mathbb{T}_{n}^{(s)}(xq)+x\sum_{j=1}^{s}\mathbb{T}_{n-j}^{(s)}(xq^{j}). \end{align} \end{theorem} \begin{proof} For Relation (\ref{es}), we have \begin{align*} \mathbb{T}_{n+1}^{(s)}(x)=\sum_{k}{n-k \brack k}_{[s]}x^{k}, \end{align*} then by Relation (\ref{bn1}), we have \begin{align*} \mathbb{T}_{n+1}^{(s)}(x)=\sum_{k}{n-k-1 \brack k}_{[s]}x^{k}+\sum_{k}{n-k-1 \brack k-1}_{[s]}x^{k}q^{n-k-1}+\cdots\hspace*{1.5cm}\\ \cdots+\sum_{k}{n-k-s \brack k-1}_{[s]}x^{k}q^{n-k-s}\\ =\sum_{k}{n-k-1 \brack k}_{[s]}x^{k}+x\sum_{k^{'}}{n-k^{'}-2 \brack k^{'}}_{[s]}x^{k^{'}}q^{n-k^{'}-2}+\cdots \hspace*{0.9cm}\\ \hspace*{0.7cm}\cdots+x\sum_{k^{'}}{n-k^{'}-s-1 \brack k^{'}}_{[s]}x^{k^{'}}q^{n-k^{'}-s-1}\\ =\mathbb{T}_{n}^{(s)}(x)+x\sum_{j=1}^{s}q^{n-j-1}\mathbb{T}_{n-j}^{(s)}(x/q).\hspace*{5cm} \end{align*} \end{proof} The proof is the same for the Relation (\ref{es1}), we use Relation (\ref{bn2}).

\section{INTRODUCTION} Colloidal rods have received long-lasting attention as building-blocks for complex materials because of their effective gelling properties and their ability to orient upon hydrodynamic forces, allowing the manufacture of materials with tailored anisotropy.\citep{Solomon2010,Rosen2020b,Calabrese2020b} Soft materials with anisotropic orientation have shown a large number of advantages, such as their mechanical strength,\cite{Hakansson2014,Nechyporchuk2019} structural color,\citep{Liu2014} electrical and thermal conductivity,\citep{Kiriya2012,Xin2019} and directional control of cell growth.\citep{DeFrance2017} To achieve directionality in soft materials, microfluidic platforms have been extensively used to control hydrodynamic forces, mainly through shearing and extensional-dominated flows, to aid particle alignment while keeping negligible inertia effects. The relative strength between diffusion and hydrodynamic forces is typically described by the P\'{e}clet number $Pe=|E|/Dr$, where $|E|$ is the characteristic deformation rate, and the rotational diffusion coefficient for non-interacting rods, $Dr$, is described as\citep{LangStiffness} \begin{equation} Dr=\frac{3 k_b T \mathrm{ln}(l/d_{eff})}{\pi \eta_{s} l^3}~, \label{eqn:Dr} \end{equation} where $k_b$ is the Boltzmann constant, $T$ the temperature, $\eta_{s}$ the solvent shear viscosity, $l$ is the length and $d_{eff}$ the effective diameter of the rod which accounts for the thickness of the electric double layer. For $Pe<1$ the particles are dominated by Brownian fluctuations, whilst for $Pe>1$, the particles are perturbed by the flow field. It has been shown, experimentally and theoretically, that shearing flows enable a gradual alignment of anisotropic particles towards the flow direction, where particles align with a preferential angle of 45$^\circ$ at $Pe\simeq1$ and achieve orientation parallel to the flow direction, 0$^\circ$, at $Pe\gg1$.\citep{Vermant2001,Dhont2007,Reddy2018,Winkler2004} Although shear-dominated flows are relatively simple to study \textit{via} conventional rheo-optics techniques, ``pure'' extensional flows in conditions where shear-forces are negligible are experimentally difficult to assess.\citep{Haward2016,Lang2019a} Consequently, much less is known on how extensional rates affect the orientation of particles in shear-free conditions. Nonetheless, it has been shown that extensional rates enable an additional control on particle orientations, inducing, for instance, particle alignment perpendicular to the flow direction, which is not possible in shearing flows.\citep{Trebbin2013,Kiriya2012,Qazi2011,Pignon2003} Of considerable importance, is the work of Corona et al., who investigated the effects of shearing and extension-dominated flows on the alignment of a concentrated suspension of colloidal rods using a fluidic four-roll mill device.\citep{Corona2018} However, no significant differences could be discerned between shearing and extensional-dominated flows as excluded volume effects were likely dominant at the high concentration tested. Recent work from Ros\'{e}n et al.\citep{Rosen2020a} elucidated that for extension-dominated flows, generated in flow-focusing and converging channels, a reduced rotational diffusion coefficient was obtained when compared to shearing flows. Nonetheless, difficulties rising from interparticle interactions, particle flexibility, and a non uniform extensional rate within the microfluidic channel, hindered a comprehensive quantitative examination to decouple the effects of shearing and extensional dominated flows. To date, a quantitative comparison between shear and extensional rate--driven alignment of rod-like particles, in conditions where interparticle interactions are negligible, is still missing. In this article we couple quantitative flow field measurements with state-of-the-art flow-induced birefringence analysis to elucidate the role of shear and extensional rates at aligning rod-like particles. A dilute cellulose nanocrystal (CNC) suspension consisting of negatively charged rigid rod-like nanoparticles is used as a model system.\citep{Hasegawa2020a} \section{MATERIALS AND METHODS} \subsection{Test fluid} The CNC was purchased from CelluForce (Montreal, Canada) as an aqueous 5.6 wt\% stock dispersion at pH 6.3. A 0.1wt\% CNC dispersion was prepared by dilution of the stock dispersion with deionised water and used without further treatment. Extensive characterization of CNC from the same industrial producer is described by Bertsch et al.\citep{Bertsch2017,Bertsch2019} and Reid et al.\citep{Reid2017} \subsection{Shear Rheometry} Steady shear rheology of the 0.1 wt\% CNC dispersion was measured using a strain-controlled ARES-G2 rotational shear rheometer (TA Instruments Inc.) equipped with a double gap geometry (with an inner and outer gap of 0.81 and 1.00~mm, respectively) composed of a stainless steel bob and a hard-anodized aluminum cup. The dispersion was covered with a solvent trap and measurements were performed at 25$^\circ$C (controlled by an advanced Peltier system with temperature accuracy of $\pm$~0.1$^\circ$C). A microfluidic slit rheometer (m-VROC RheoSense Inc.), equipped with an A10 pressure cell, 3 mm wide and \SI{100}{\micro\meter} high, was used to access the rheological response of the test fluid at high values of shear rate (up to $\dot\gamma\approx 3 \times 10^4$ s$^{-1}$). The experiment was carried out at $25^{\circ}$C (controlled via an external circulating water bath with temperature accuracy of $\pm$~0.1$^\circ$C). \subsection{Atomic force microscopy (AFM)} A 0.01~wt\% CNC dispersion was drop-casted on a mica substrate and imaged using an atomic force microscope (Dimension ICON3, Bruker) in tapping mode. The distribution of the particle contour length, $l$, and diameter, $d$, was extrapolated by tracking 976 isolated particles using an open-source code, FiberApp.\citep{Usov2015a} The value of $d$ was obtained from the AFM height profile. \subsection{Microfluidic platforms} A shearing flow-dominated channel (SFC) and an optimized shape cross-slot extensional rheometer (OSCER)\citep{Haward2012a,Haward2016} were used to generate two-dimensional (2D) flows that, in discrete areas of the geometries, provide good approximations to purely shearing and purely extensional flows, respectively (Figure~\ref{fgr:Channels}). The SFC consists of a straight channel of fused silica glass with a rectangular cross section (length $L=25$ mm along the $x$-axis, height $H=2$~mm along the $z$-axis, and width $W=0.4$~mm along the $y$-axis, resulting in an aspect ratio $\alpha=H/W=5$), fabricated with a selective laser-induced etching (SLE) technique\citep{Burshtein2019,Haward2019} (Figure~\ref{fgr:Channels}a). The OSCER is based on a planar cross-slot geometry with two incoming and outgoing flows placed orthogonal to each other, as described by Haward et al.\citep{Haward2012a,Haward2016} and depicted in Figure~\ref{fgr:Channels}b. The device has a height of $H=2.1$~mm along the $z$-axis and a channel width of $W=0.2$ mm at the inlets and outlets, yielding $\alpha=10.5$, generating a good approximation of a 2D flow field which is extensional-dominated in a large region around the stagnation point, at $x=y=0$ (see coordinate system in Figure~\ref{fgr:Channels}b). The fluid elements are compressed along the $y$-axis and extended along the $x$-axis (referred to as the compression and elongation axes, respectively, Figure~\ref{fgr:Channels}b). The flow inside the channels is driven by Nemesys low-pressure syringe pumps (Cetoni, GmbH) and Hamilton Gastight syringes, which infuse the liquid at the inlet and withdraw it at an equal and opposite volumetric rate, $Q$ (m$^3$~s$^{-1}$), from the outlet. To ensure steady flows, the highest flow rates were set at Reynolds numbers, $Re=\rho U W/\eta$, of 15.4 and 6.4, for the SFC and OSCER, respectively, considering a fluid density, $\rho=1000$ kg~m$^{-3}$, an average flow velocity as $U=Q/HW$ (m~s$^{-1}$) and a fluid viscosity $\eta=1.3$ mPa~s as determined by rheological measurements (detailed in the following section). The flows were equilibrated for at least 5~s before all measurements and confirmed as being steady by inspection of the micro-particle image velocimetry ($\mu$-PIV, see next section) flow fields prior to their time-averaging. All the experiments were carried out at the ambient laboratory temperature ($ 25 \pm 1^{\circ}$C). \begin{figure}[h!] \includegraphics[width=8cm]{New1.jpg} \caption{Schematic diagrams of (a) the shearing flow-dominated channel (SFC) and (b) the optimized shape cross-slot extensional rheometer (OSCER) with respective coordinate system and scale bar. The dashed arrows indicate the flow direction.} \label{fgr:Channels} \end{figure} \subsection{Micro-particle image Velocimetry ($\mu$-PIV)} The flow fields in the SFC and OSCER devices were obtained using time averaged $\mu$-PIV of the test fluid seeded with \SI{1.1}{\micro\meter} fluorescent particles (Fluoro-Max\textsuperscript{TM}, Thermo Fisher), to a concentration of $\approx0.02$ wt\%. The $\mu$-PIV measurements were conducted using a volume illumination system (TSI Inc., MN) installed on an inverted microscope (Nikon Eclipse Ti). Nikon PlanFluor objective lenses of 4$\times$ and 10$\times$ with numerical apertures of NA=0.13 and 0.30 were used for the OSCER and SFC device, respectively. Each geometry was placed with $z$-axis parallel to the light source and for all the flow rates tested, a sequence of at least 100 image pairs were acquired at the midplane of the geometries ($z=1$ mm). For the SFC, images were acquired at a distance of $\approx L/2$ from the inlets to ensure fully developed velocity profiles. The average displacement of the seeded particles between the two images in each pair was kept constant at $\approx$ 4 pixels. The measurement depths, $\delta z_m$, corresponding to the depth over which the seeded particles contribute to the determination of the velocity field was $\delta z_m =150~ $\SI{}{\micro\meter} and \SI{35}{\micro\meter}, for the OSCER and SFC, respectively.\citep{Meinhart2000} Cross-correlation between image pairs provided velocity vectors on a square grid with a spatial resolution of 2.0 and 0.8 \SI{}{\micro\meter}/pixel for the OSCER and SFC, respectively. Data analysis was performed using a custom-made Matlab routine. \subsection{Flow-induced birefringence (FIB)} Flow-induced birefringence (FIB) measurements were performed using an Exicor MicroImager (Hinds Instruments, Inc., OR). Monochromatic light of wavelength $\lambda=450$~nm was shone through a linear polarizer at 0$^\circ$, a photoelastic modulator (PEM) at 45$^\circ$, the SFC or the OSCER containing the testing fluid, a PEM at 0$^\circ$ and a linear polarizer at 45$^\circ$, in the order given. The geometries were imaged using a 5$\times$ objective in the same position as described for $\mu$-PIV. The instrument performs Mueller matrix decomposition, determining the elements of 4$\times$4 Mueller matrices using a stroboscopic light source. The retardance $R$, describing the total phase shift occurring between the two orthogonally polarized light beams, and the orientation of the slow optical axis (extraordinary ray), $\theta$, were obtained from a total of 7 images acquired at 1 frame/s. The retardance, $R$ (measured in nm), was then converted to birefringence as $\Delta n =R/H$. The background value of $\Delta n$ was determined for the test fluid at rest and subtracted for all the analysis presented. The background value determined in both geometries was $\Delta n \approx 6\times10^{-7}$ and comparable to the instrument resolution of $\Delta n \approx 3\times10^{-7}$. The spatial resolution of the measurement was $\approx 2$ \SI{}{\micro\meter}/pixel and data analysis was performed using a custom-made Matlab routine. \section{RESULTS AND DISCUSSION} \subsection{Characterization of the test fluid} We begin by estimating the volume density, $\nu=N / V$, of the 0.1 wt\% CNC dispersion, where $N$ is the number of rods and $V$ is the sample volume. In suspensions of monodisperse rod-like particles, the volume density, $\nu$, is commonly used to distinguish between the dilute, semi-dilute and concentrated regimes.\citep{Solomon2010} For monodisperse rods, the calculation of $\nu$ is straightforward as a single particle length ($l$) and diameter ($d$) are sufficient to describe the whole particle population. Contrarily, for a polydisperse distribution of rods such as CNC, the large span of lengths must be considered for a more accurate estimation of $\nu$. Thus, the size distribution of the CNC was extrapolated from atomic force microscopy (AFM) images (Figure~\ref{fgr:Fig1}a,b). The average contour length was $\langle l \rangle = 260 \pm180$ nm and the average diameter $\langle d \rangle = 4.8 \pm1.8$ nm. The effective volume density, $\nu_e$, was thereafter estimated as \begin{equation} \nu_e=\frac{N_e}{V}=\frac{\sum_{i=1}^{l_{max}} \bigg(\frac{V_{CNC}\phi_i}{Vcyl_i}\bigg)}{V}, \label{eqn:nueff} \end{equation} where $N_e$ is the effective number of rods, $l_{max}$ is the longest detected contour length (700~nm), $V_{CNC}$ is the volume of the CNC in the sample (which can be estimated using a density of 1500~kg~m$^{-3}$),\citep{RyanWagnerArvindRaman2010} $\phi_i$ is the volume fraction of the rods with length $i$ and $Vcyl_i$ is the volume occupied by a single rod with length $i$, for which a cylindrical morphology can be approximated. It is noted that the distribution of the CNC diameters, $d$, is not accounted for in eqn.~\ref{eqn:nueff} since it has only a minimal effect on the estimation of $\nu_e$ of slender objects and $\langle d \rangle$ is used to obtain $Vcyl_i$. For the 0.1 wt\% CNC, the $\nu_e\approx 1/ \langle l \rangle^3$, indicating that at this concentration, the CNC dispersion is at the onset of the semi-dilute regime, where the particles rarely interact (although more frequently than in the dilute regime).\citep{LangStiffness,Lang2019} The absence of pronounced interparticle interactions was reported by Bertsch et al. for concentrations below 0.5~wt\%, as shown by small angle X-ray scattering studies of CNC from the same source as that used in the present work.\citep{Bertsch2019} In addition, from the AFM images, a persistence length, $l_p\approx30\langle l\rangle$ was extrapolated using the method of the mean-squared midpoint displacement (MSMD) within the FiberApp routine and detailed by Usov and Mezzenga,\citep{Usov2015a} indicating that CNC can be well described as rigid rods. A value of $l_p\gg l$ was also obtained when $l_p$ was calculated as\citep{Usov2015} \begin{equation} l_p=\frac{\pi \langle d \rangle^4 G}{64 k_b T}, \label{eqn:lp} \end{equation} using the reported values of the CNC elastic modulus, $G$, between 5 and 150 GPa.\citep{RyanWagnerArvindRaman2010} \begin{figure}[h!] \includegraphics[width=8cm]{Fig1.jpg} \caption{(a) Tapping-mode atomic force microscopy (AFM) image of a drop-casted 0.01wt\% CNC dispersion on mica substrate. (b) Particle contour length, $l$, and particle diameter, $d$ (inset in (b)), distributions as obtained from the AFM counting of 976 isolated particles. (c) Flow curve of the 0.1 wt\% CNC dispersion presented as shear stress, $\sigma$ \textit{vs} shear rate, $\dot\gamma$. The black circles are data obtained using a rotational strain controlled rheometer (ARES-G2). The data represented by the diamonds are obtained using a microfluidic slit rheometer (m-VROC). The dashed (black) and solid (blue) lines are the predictions of the zero shear viscosity from eqn.\ref{eqn:eta} using values of average length, $\langle l \rangle$, and maximum length, $l_{max}$, respectively, as obtained from AFM analysis. } \label{fgr:Fig1} \end{figure} The steady shear rheology of the 0.1~wt\% CNC dispersion is shown in Figure~\ref{fgr:Fig1}c. The shear stress, $\sigma$, of the 0.1 wt\% CNC suspension followed a linear relationship with the shear rate, $\dot\gamma$, as for a Newtonian fluid with shear viscosity $\eta= 1.3$~mPa~s. Anisotropic particles are expected to exhibit a viscosity plateau at low shear rate, namely the zero shear viscosity, $\eta_0$, followed by a shear thinning behavior up to a second viscosity plateau at higher shear rates, $\eta_\infty$.\citep{Lang2019,Lang2020b,Tanaka2014,Kobayashi2011,Lang2016} From eqn.~\ref{eqn:Dr} it is clear that longer rods will align at lower shear rates than shorter rods since $Dr\propto 1 / l^3$. As such, in polydisperse suspensions of non-interacting colloidal rods, the onset of shear thinning is dictated by the longest population of rods. Substituting $l$ with $l_{max}$ and accounting for the contribution of the electric double layer ($\delta{d}=22.6$~nm)\citep{Bertsch2019} to the effective diameter ($d_{eff}\approx~\delta{d} +\langle d \rangle$) in eqn.~\ref{eqn:Dr}, we estimated the expected onset of the shear thinning at $\dot\gamma\simeq 40 \ $~s$^{-1}$ ($Pe~\simeq ~1$). However, at this CNC concentration, the value of viscosity of the test fluid is close to that of the solvent shear viscosity, $\eta_{s}$, making practically impossible to capture the shear thinning region expected for anisotropic particles. For a dilute dispersion of monodisperse rods, $\eta_0$ can be estimated as\citep{LangStiffness} \begin{equation} \eta_0\simeq \eta_{s}+\nu_e k_b T \bigg(\frac{1}{30} \frac{1}{Dr} \bigg). \label{eqn:eta} \end{equation} When the value of $l$ in eqn.~\ref{eqn:Dr} is substituted with either $\langle l \rangle$ or $l_{max}$, the predicted values of zero shear viscosity well encompass the experimental data (Figure~ \ref{fgr:Fig1}c). The lack of any further contribution needed to account interparticle interactions in eqn. \ref{eqn:eta} suggests that the 0.1~wt\% CNC dispersion can be considered in the dilute regime. \subsection{Flow profiles} The bulk rheometry gives a good indication that the diluted CNC dispersion behaves as a Newtonian fluid. Consequently, no information regarding the onset of alignment, as usually associated with the onset of shear thinning, can be discerned from the flow curve. To define a deformation rate--alignment relationship of the dilute CNC dispersion we focus on the control of the shear rate, $\dot\gamma$, and extensional rate, $\dot\varepsilon$ in two separate microchannels. \begin{figure}[h!] \includegraphics[width=16cm]{Fig2.jpg} \caption{(a) Time averaged results of flow velocimetry ($\mu$-PIV) with superimposed streamlines for the SFC containing the 0.1~wt\% CNC dispersion. Two representative average flow velocities, $|U|$, are displayed. (b) Normalized spatially averaged velocity profiles taken across the $y$-axis of the channel, for values of $|U|$ as in (a). (c) Magnitude of shear rate, $|\dot\gamma|=\partial|V_x|/\partial|y|$, as obtained from the velocity profile displayed in (b). (d) Magnitude of the shear rate obtained from the average of $|\dot\gamma|$ at $y=\pm0.1$ mm, $|\dot\gamma|_{(\pm0.1)}$, as a function of the average flow velocity. In the panels (b), (c), and (d), the solid lines are the infinite series analytical solution for creeping Newtonian flow.\citep{Shah}} \label{fgr:Fig2} \end{figure} The SFC was used to provide a well approximated 2D shearing flow. Figure~\ref{fgr:Fig2}a shows two representative time averaged flow fields obtained by $\mu$-PIV at the midplane ($z=1$~mm) of the SFC. The velocity magnitude $|V|$, as measured by $\mu$-PIV, is scaled by the average flow velocity $|U|$. The flow field displays a velocity gradient across the channel width ($y$-axis) with the greater velocity along the centerline. This is evident when plotting the spatially averaged velocity, namely $\langle|V|\rangle$ (determined by averaging $|V|$ along 0.2 mm of the $x$-axis) as $\langle|V|\rangle/|U|$ \textit{vs} the channel width ($y$-axis), where a parabolic (Poiseuille) flow profile is displayed (Figure~\ref{fgr:Fig2}b). The measured flow profile is in quantitative agreement with an infinite series analytical solution for creeping Newtonian flow\citep{Shah} as depicted by the solid line in Figure~\ref{fgr:Fig2}, consistent with the Newtonian-like behavior described by the rheological measurement in Figure~\ref{fgr:Fig1}. From the velocity profile displayed in Figure~\ref{fgr:Fig2}b, we compute the shear rate profile as $|\dot\gamma|=\partial|V_x|/\partial|y|$, where $|V_x|$ is the $x$-component of velocity (Figure~\ref{fgr:Fig2}c). The magnitude of the shear rate is also in good agreement with the expectation for a Newtonian fluid (black lines). Since the shear rate varies substantially along the channel $y$-axis, we select the value of the shear rate at the location $y=\pm$~$0.1$~mm, namely $|\dot\gamma|_{(\pm 0.1)}$ (computed as the average of the $|\dot\gamma|$ at $y=0.1$ mm and $y=-0.1$~mm), which is the mid-point in $|y|$ between the minimum and the maximum value of $|\dot\gamma|$. The location $y=\pm$~$0.1$~mm is also far from the channel side walls, where the $\mu$-PIV limitations become evident, while still providing relatively high values of~shear~rate. \begin{figure}[ht!] \includegraphics[width=16.5cm]{Fig3.jpg} \caption{(a) Time averaged results of flow velocimetry ($\mu$-PIV) with superimposed streamlines for the OSCER containing the 0.1~wt\% CNC dispersion. Two representative average flow velocities, $|U|$, are displayed. (b) Velocity component along the elongation axis, $V_x$, at $y=0$~mm. (c) Velocity component along the compression axis, $V_y$, at $x=0$~mm. (d) Magnitude of the extension rate, $|\dot\varepsilon|$, as obtained along the compression and elongation axes. The solid line is the expected relationship for a Newtonian fluid.\citep{Haward2012a} } \label{fgr:Fig3} \end{figure} The relation between $|\dot\gamma|_{(\pm0.1)}$ and the average velocity $|U|$ can therefore be established as shown in Figure~\ref{fgr:Fig2}d, leading to $|\dot\gamma|_{(\pm0.1)}=176|U|$, which satisfies the expectation for a Newtonian fluid (solid line). The OSCER is used to generate a 2D flow field with an extensional-dominated flow in a large region around the stagnation point (the point of zero velocity at $x=y=0$, Figure~\ref{fgr:Channels}b).\citep{Haward2012a,Haward2016,Hawarda2016} The two incoming and outgoing flows are orthogonal to each other and indicated by the dashed arrows in Figure~\ref{fgr:Fig3}a. The $\mu$-PIV profiles for two characteristic average velocities, $|U|$, display a symmetric and Newtonian-like behavior (Figure~\ref{fgr:Fig3}a). The $x$-component of velocity, $V_x$, obtained along the elongation axis (at $y=0$) increases linearly with $x$ (Figure~\ref{fgr:Fig3}b) and, analogously, the $y$-component of velocity, $V_y$ (at $x=0$) along the compression axis, decreases linearly with $y$ (Figure~\ref{fgr:Fig3}c). This indicates uniform extensional rates along both the compression and elongation axes. As such, the values of compression and elongation rates can be directly obtained from the slope of the plots in Figure~\ref{fgr:Fig3}b,c. In Figure~\ref{fgr:Fig3}d, the magnitude of the compression and elongation rate as a function of the average flow velocity $|U|$ collapsed on a single curve, indicating that the extensional rates, $\dot\varepsilon$, along the compression and extension axes are equal and opposite, as for an ideal planar extension.\citep{Haward2012a} The $|\dot\varepsilon|$ followed a linear relationship with $|U|$ as $|\dot\varepsilon|=1070|U|$ (indicated by the solid line), in good agreement with previously reported Newtonian fluids.\citep{Haward2012a,Haward2016} This behavior suggests that the diluted CNC dispersion behaves as a Newtonian fluid also in extension, giving an extensional viscosity $\eta_{E}=4\eta$ where $\eta$ is the shear viscosity obtained from rheometry measurements displayed in Figure~\ref{fgr:Fig1}c and the proportionality factor of 4 is the Trouton ratio for Newtonian fluids in planar elongation flow.\citep{Hawarda2016} \subsection{Flow-induced alignment} Since the spatially resolved deformation rates, $|\dot\gamma|$ and $|\dot\varepsilon|$, have been established, we shift our attention to the impact that the two different deformation rates have on the structural orientation of the dilute CNC dispersion. The contourplot in Figure~\ref{fgr:Fig4}a displays the FIB fields in the SFC at two different values of $|U|$. The birefringence intensity, $\Delta n$, is displayed by the contourplot and describes the extent of anisotropy in the system and the orientation of the slow optical axis, $\theta$, is displayed by the superimposed solid segments, which directly probes the CNC orientation angle.\citep{Vermant2001,Sun2016} As a metric of comparison, it has been recently shown that both $\Delta n$ and $\theta$ yield excellent agreement with parameters describing the extent of alignment and favourable orientation extrapolated from small angle X-ray scattering (SAXS) of anisotropic particles, enabling a trustworthy comparison of the FIB with the rheo-small angle scattering (SAS) literature.\citep{Rosen2020} The FIB fields in the SFC display the minimum at $y=0$ (where $\dot\gamma=0$~s$^{-1}$) and increases towards the channel walls (Figure~\ref{fgr:Fig4}a). Note that we attribute the narrow region of low FIB signal very close to the walls to ``shadowing'' by the wall, i.e. due to a slight imprecision of the orthogonal alignment of the microfluidic device on the imaging system (see Figure~S1 for $\Delta n$ profiles across the channel width). Since $\Delta n$ scales with the volume fraction of aligned particles ($\phi_{aligned}$), it is clear that a greater number of CNC become aligned as the shear rate is increased. The CNC orientation, displayed by the solid segments in Figure~\ref{fgr:Fig4}a, shows an overall alignment of the particles in the flow direction. However, for the lowest value of $|U|$, the CNC orientation is not perfectly mirrored between the portion of the channel at $y<0$ mm and $y>0$ mm, and a degree of heterogeneity in the CNC orientation angle is observed. In the OSCER device, a large region around the stagnation point displays a strong $\Delta n$ signal, as clearly visible for the larger value of $|U|$ in Figure~\ref{fgr:Fig4}b (displayed by the light blue color in the contourplot). The direction of the CNC alignment in the compression axis is perpendicular to the flow direction ($y$-axis) due to the deceleration of the fluid element upon approaching the stagnation point and the consequent negative extensional rate along the $y$-axis. Contrarily, the CNC aligns parallel to the flow ($x$-axis) along the elongation axis due to the positive extension rate. Similar orientation trends have been reported for suspensions of anisotropic particles in extensional-dominated flows (in cross slots and fluidic four-roll mill devices), although in conditions where interparticle interactions play a crucial role on particle alignment.\citep{Pignon2003,Qazi2011} \begin{figure}[h!] \includegraphics[width=16cm]{Fig4.jpg} \caption{Time averaged FIB profiles of a 0.1~wt\% CNC dispersion, for the (a) SFC and (b) OSCER device. The birefringence, $\langle \Delta n \rangle$, displayed by the contourplot whilst the direction of the slow optical axis, $\theta$, indicated by the the solid segments (in white or black, for easy visualization). (c) Spatially averaged birefringence, $\langle \Delta n \rangle$, and (d) spatially averaged orientation angle, $\langle \theta \rangle$, as a function of the magnitude of shear rate, $|\dot\gamma|$, and extension rate $|\dot\varepsilon|$. The $\langle \Delta n \rangle$ and $\langle \theta \rangle$ are obtained in the same channel location as for the relevant deformation rate. The inset in (c) displays the plot as in (c) with a re-scaled $x$-axis as $4|\dot\varepsilon|$. The solid lines in (c) are power law fittings. The dashed (black) and solid (blue) vertical lines are the values of $Dr$ and $Dr_{E}$ (s$^{-1}$) estimated \textit{via} eqn.\ref{eqn:Dr}, using values of the solvent shear viscosity ($\eta_{s}$), and solvent extensional viscosity ($\eta_{s,E}$), respectively, and $l=l_{max}$.} \label{fgr:Fig4} \end{figure} To have a good quantitative description of the deformation rate--alignment relationship, we plot the spatially resolved values of $\Delta n$ as a function of the relevant deformation rate previously determined in the specific locations of the SFC and OSCER geometries (Figure~\ref{fgr:Fig4}c). For the SFC, the spatially averaged birefringence signal, $\langle \Delta n \rangle$, is obtained at $y=\pm0.1$ mm averaging $\Delta n$ along 1 mm in the $x$-direction, whilst for the OSCER device, $\langle \Delta n \rangle$ is obtained by averaging $\Delta n$ along 1 mm of the elongation axis (at $y=0$ over $-0.5\leq x\leq0.5$~mm, see Figure~S1, S2 for the spatially-resolved $\Delta n$ profiles). The curves showed in Figure~\ref{fgr:Fig4}c for the SFC and the OSCER device display a similar increase in $\langle \Delta n \rangle$ with the deformation rate, $|E|$, which is well captured by a power law trend as $\langle \Delta n \rangle=A |E|^{p}$, where $A$ is a proportionality factor and $p$ is the power law exponent. For the SFC, $A~=~(0.9~\pm~0.2~)~\times~10^{-7}$~s and $p=0.90 \pm 0.03$, whilst for OSCER, $A=( 1.4 \pm 0.2) \times 10^{-7} $~s and $p=0.98 \pm 0.04$. Importantly, the FIB technique enabled to capture the CNC alignment with the deformation rate, otherwise unexpected by interpretation of the rheological measurements alone (see Figure~\ref{fgr:Fig1}c). It is noted that for a given applied $|U|$ (hence $\dot\gamma$ profile, Figure~\ref{fgr:Fig2}c), this empirical power law is able to capture the $\langle \Delta n \rangle$ profile across most of the SFC width with reasonable accuracy (see Figure~S1 for spatially-resolved $\Delta n$ profiles with the predictions of the empirical power law fitting). Interestingly, from Figure~\ref{fgr:Fig4}c, it is apparent that lower extensional rates are required to induce rod alignment in the OSCER device, than the shear rates required to induce alignment in the SFC. This appears to indicate that extensional forces are more effective at inducing CNC alignment. The $\langle \Delta n \rangle$ profiles obtained for the extensional and shearing flow collapsed on a single master curve when scaling the extensional rate as $4|\dot\varepsilon|$ (inset in Figure~\ref{fgr:Fig4}c). This indicates that extensional and shear forces have a similar relationship with $\phi_{aligned}$ and that the extensional forces are 4 times more effective for the alignment of anisotropic particles when compared to shear forces. Therefore for a Newtonian fluid containing non-interacting rods, the following proportionality can be proposed $\Delta n\propto |\dot\gamma|^{0.90} \propto 4|\dot\varepsilon|^{0.98}\propto \phi_{aligned}$. It is noted that the proportionality factor of 4 corresponds exactly to the Trouton ratio for Newtonian fluids in a planar elongation flow, which set the relationship between the solvent shear viscosity, $\eta_{s}$, and the solvent extensional viscosity, $\eta_{s,E}=4\eta_{s}$.\citep{Hawarda2016} The strong dependence of the solvent viscosity on the onset of flow alignment is reflected by the $Dr\propto 1/\eta_{s}$ in eqn.\ref{eqn:Dr}. Assuming a dominant extensional viscosity $\eta_{s,E}$ along the extensional axis, it is plausible to substitute $\eta_{s}$ with $\eta_{s,E}=4\eta_{s}$ in eqn.~\ref{eqn:Dr}, yielding a $Dr$ in extension ($Dr_{E}$) 4 times smaller compared to the $Dr$. Although not validated, the consideration of $\eta_{s}$ or $\eta_{s,E}$ for a rotational diffusion coefficient that accounts for shearing and extensional-dominated flows, respectively, has been previously proposed by Qazi et al.\citep{Qazi2011} A decreased rotational diffusion coefficient in extensional-dominated flows has also been demonstrated by Ros\'{e}n et al.\citep{Rosen2020a} for interacting cellulose nanofibrils. However, the absence of interparticle interactions and the uniform, shear-less, elongational flow presented in this work enables the first quantitative elucidation of differences between the rotational dynamics of colloidal rods in shear and extensional flows. Considering that the longest CNC population ($l_{max}$) aligns at lower deformation rates, the predicted onset of alignment based on $Dr$ (dashed, black, lines in Figure~\ref{fgr:Fig4}c,d) and $Dr_{E}$ (solid, blue, lines in Figure~\ref{fgr:Fig4}c,d) do not exactly match the onset of birefringence. However, the scaling factor between the curves in Figure~\ref{fgr:Fig4}c strongly suggests that the larger value of $\eta_{s,E}$ compared to $\eta_{s}$ is the cause of the earlier onset of alignment. To better evaluate the onset of alignment in shear and extensional flow, we further evaluate quantitatively the CNC orientation as captured by the angle of orientation, $\theta$, with respect to the flow direction (Figure~\ref{fgr:Fig4}d). For the SFC, the spatially averaged angle of the slow optical axis, $\langle \theta \rangle$, was determined in the same location as for the $\langle \Delta n \rangle$, i.e. averaging $|\theta|$ along 1 mm in the $x$-direction at $y=\pm 0.1$~mm, whilst for the OSCER device, the $\langle \theta \rangle$ was obtained from averaging $\theta$ along 1 mm of the elongation axis at $y=0$ (averaging $\theta$ over $-0.5\leq x\leq0.5$~mm) and compression axis at $x=0$ (averaging $\theta$ over $-0.5\leq y\leq0.5$~mm, see Figure~S2 for $\theta$ profiles along the extensional axes). In shear, for $|\dot\gamma| < 10^2$ s$^{-1}$, within experimental error (note error bars), the CNC alignment is almost constant at $\langle \theta \rangle \simeq 28 ^\circ$ with respect to the flow direction. For $|\dot\gamma|\gtrsim 10^2$ ~s$^{-1}$, $\langle \theta \rangle$ progressively decreases to values of around $22 ^\circ$ at the highest shear rates, in good agreement with SAXS and birefringence studies of a diluted CNC suspension.\citep{Rosen2020} Along the elongation axis of the OSCER, the CNC aligns parallel to the flow, with an orientation of $\langle \theta \rangle \simeq 0 ^\circ$ to the flow direction. Contrarily, along the compression axis, the CNC align perpendicular to the flow with $\langle \theta \rangle \simeq 90 ^\circ$. Such perpendicular alignment of elongated particles has been previously reported for shear thinning fluids in extensional-dominated flows.\citep{Kiriya2012,Trebbin2013} However, it is clear from our results that shear thinning is not a requirement for such orientation to occur. As previously anticipated, the expected value for the onset of CNC alignment is $|\dot\gamma|\simeq~40$~s$^{-1}$ (depicted by $Dr$, dashed (black) lines in Figure~\ref{fgr:Fig4}c,d) when considering the longest CNC population. Although the onset of birefringence, $\langle \Delta n \rangle$, occurs at values of $|\dot\gamma|\approx 10$~s$^{-1}$, the predicted onset of alignment at $|\dot\gamma|\simeq 40$~s$^{-1}$ is in close agreement with the origin of the significant decrease in $\langle \theta \rangle$ in the SFC (Figure~\ref{fgr:Fig4}d). The coexistence of $\langle \Delta n \rangle > 0$ and the steady value of $\langle \theta \rangle$ with a relatively large error at $|\dot\gamma| \lesssim 40$ s$^{-1}$, suggests the presence of an intermediate shear rate region where Brownian diffusion and deformation rate are of comparable magnitude, leading to a time-dependent and collective particle rotation in the velocity-gradient plane.\citep{Dhont2007,Mewis1997} Contrarily, a Brownian dominated region occurs for $|\dot\gamma| < 10$ s$^{-1}$, where the system is isotropic and $\langle \Delta n \rangle\approx 0$, whilst, at $|\dot\gamma| \gtrsim 40$ s$^{-1}$, the hydrodynamic forces orient the particles towards a $\langle \theta \rangle = 0^\circ$, causing a pronounced increase in $\langle \Delta n \rangle$. A similar intermediate shear rate regime has been described for a concentrated CNC suspension\citep{Ebeling1999} as well as wormlike micelles\citep{Mu2019} and polymeric liquid crystals.\citep{Mewis1997} For the elongational flow, $\langle \theta \rangle$ is constant over the whole range of $|\dot\varepsilon|$, in agreement with the reported analytical solutions for diluted Brownian suspensions of rod-like particles.\citep{Dhont2007} Therefore, the onset of a convection-dominated particle alignment in a shear-dominated flow is more adequately described by the orientation angle $\langle \theta \rangle$, whilst, for the extensional-dominated flow, the onset of birefringence is a better parameter to consider since $\langle \theta \rangle$ is independent of the rate of deformation. When this is considered, the onset of alignment is satisfactorily captured by a diffusion coefficient that accounts for the longest particle population and the appropriate viscosity that the fluid experiences in the specific location of the channel, i.e. $Dr$ or $Dr_{E}$. \section{CONCLUSION} Combining flow visualization and flow induced birefringence, we are able to decouple the effects of shear and extensional forces on a dilute dispersion of cellulose nanocrystals, consisting of negatively charged polydisperse and rod-like rigid particles. Simultaneous analysis of the birefringence with the orientation of the slow optical axis elucidate that extensional forces are \textit{ca.} 4 times more effective at inducing particle alignment compared to shear forces. This difference is explained by the different viscosity that the fluid experiences between a shear and extensional-dominated region, namely the shear and extensional viscosity, respectively. In addition, shear and extension rates have displayed a remarkably different effect on particle orientation. In a shear-dominated flow, the particles exhibit a gradual decrease in the orientation angle with increasing shear rate. Contrarily, in the extensional-dominated flow the particle orientation is deformation-rate independent and a parallel or perpendicular orientation of the particles to the flow direction is observed along the elongation and compression axes of the geometry, respectively. To the best of our knowledge, this is the first experimental report where the effects of shear and pure extensional flows on the alignment of rigid anisotropic particles are quantitatively compared. This understanding will provide the basis to decouple the orientation dynamics in industrially relevant flows, where a combination of shear and extensional rates are usually present and in fluids where interparticle interactions play a crucial role. We envisage that the knowledge provided in this work will help to optimize engineering processes involved with controlled anisotropy. Having demonstrated the influence of the solvent shear and extensional viscosity at controlling the alignment of non-interacting rods for a Newtonian solvent, we aim in future work to broaden this understanding for the more complex non-Newtonian case, by using solvent fluids with larger and deformation rate--dependent Trouton ratios. \begin{acknowledgement} The authors gratefully acknowledge the support of Okinawa Institute of Science and Technology Graduate University with subsidy funding from the Cabinet Office, Government of Japan. S.J.H. and A.Q.S. also acknowledge financial support from the Japanese Society for the Promotion of Science (JSPS, Grant Nos. 18K03958 and 18H01135) and the Joint Research Projects (JRPs) supported by the JSPS and the Swiss National Science Foundation (SNSF). The authors thank Dr.~Riccardo Funari from the Micro/Bio/Nanofluidics unit at OIST for his assistance with AFM measurements and Dr.~Vikram Rathee from the Micro/Bio/Nanofluidics unit at OIST for comments on the manuscript. \end{acknowledgement} \providecommand{\latin}[1]{#1} \makeatletter \providecommand{\doi} {\begingroup\let\do\@makeother\dospecials \catcode`\{=1 \catcode`\}=2 \doi@aux} \providecommand{\doi@aux}[1]{\endgroup\texttt{#1}} \makeatother \providecommand*\mcitethebibliography{\thebibliography} \csname @ifundefined\endcsname{endmcitethebibliography} {\let\endmcitethebibliography\endthebibliography}{} \begin{mcitethebibliography}{45} \providecommand*\natexlab[1]{#1} \providecommand*\mciteSetBstSublistMode[1]{} \providecommand*\mciteSetBstMaxWidthForm[2]{} \providecommand*\mciteBstWouldAddEndPuncttrue {\def\unskip.}{\unskip.}} \providecommand*\mciteBstWouldAddEndPunctfalse {\let\unskip.}\relax} \providecommand*\mciteSetBstMidEndSepPunct[3]{} \providecommand*\mciteSetBstSublistLabelBeginEnd[3]{} \providecommand*\unskip.}{} \mciteSetBstSublistMode{f} \mciteSetBstMaxWidthForm{subitem}{(\alph{mcitesubitemcount})} \mciteSetBstSublistLabelBeginEnd {\mcitemaxwidthsubitemform\space} {\relax} {\relax} \bibitem[Solomon and Spicer(2010)Solomon, and Spicer]{Solomon2010} Solomon,~M.~J.; Spicer,~P.~T. {Microstructural regimes of colloidal rod suspensions, gels, and glasses}. \emph{Soft Matter} \textbf{2010}, \emph{6}, 1391--1400\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Ros{\'{e}}n \latin{et~al.}(2020)Ros{\'{e}}n, Hsiao, and S{\"{o}}derberg]{Rosen2020b} Ros{\'{e}}n,~T.; Hsiao,~B.~S.; S{\"{o}}derberg,~L.~D. {Elucidating the opportunities and challenges for nanocellulose spinning}. \emph{Advanced Materials} \textbf{2020}, 2001238\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Calabrese \latin{et~al.}(2020)Calabrese, da~Silva, Porcar, Bryant, Hossain, Scott, and Edler]{Calabrese2020b} Calabrese,~V.; da~Silva,~M.~A.; Porcar,~L.; Bryant,~S.~J.; Hossain,~K. M.~Z.; Scott,~J.~L.; Edler,~K.~J. {Filler size effect in an attractive fibrillated network: a structural and rheological perspective}. \emph{Soft Matter} \textbf{2020}, \emph{16}, 3303--3310\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[H{\aa}kansson \latin{et~al.}(2014)H{\aa}kansson, Fall, Lundell, Yu, Krywka, Roth, Santoro, Kvick, {Prahl Wittberg}, W{\aa}gberg, and S{\"{o}}derberg]{Hakansson2014} H{\aa}kansson,~K. M.~O.; Fall,~A.~B.; Lundell,~F.; Yu,~S.; Krywka,~C.; Roth,~S.~V.; Santoro,~G.; Kvick,~M.; {Prahl Wittberg},~L.; W{\aa}gberg,~L.; S{\"{o}}derberg,~L.~D. {Hydrodynamic alignment and assembly of nanofibrils resulting in strong cellulose filaments}. \emph{Nature Communications} \textbf{2014}, \emph{5}, 4018\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Nechyporchuk \latin{et~al.}(2018)Nechyporchuk, H{\aa}kansson, Gowda.V, Lundell, Hagstr{\"{o}}m, and K{\"{o}}hnke]{Nechyporchuk2019} Nechyporchuk,~O.; H{\aa}kansson,~K. M.~O.; Gowda.V,~K.; Lundell,~F.; Hagstr{\"{o}}m,~B.; K{\"{o}}hnke,~T. {Continuous assembly of cellulose nanofibrils and nanocrystals into strong macrofibers through microfluidic spinning}. \emph{Advanced Materials Technologies} \textbf{2018}, \emph{4}, 1800557\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Liu \latin{et~al.}(2014)Liu, Wang, Ma, Tian, Gu, and Lin]{Liu2014} Liu,~D.; Wang,~S.; Ma,~Z.; Tian,~D.; Gu,~M.; Lin,~F. {Structure-color mechanism of iridescent cellulose nanocrystal films}. \emph{RSC Advances} \textbf{2014}, \emph{4}, 39322--39331\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Kiriya \latin{et~al.}(2012)Kiriya, Kawano, Onoe, and Takeuchi]{Kiriya2012} Kiriya,~D.; Kawano,~R.; Onoe,~H.; Takeuchi,~S. {Microfluidic control of the internal morphology in nanofiber-based macroscopic cables}. \emph{Angewandte Chemie - International Edition} \textbf{2012}, \emph{51}, 7942--7947\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Xin \latin{et~al.}(2019)Xin, Zhu, Deng, Cheng, Zhang, Chung, De, and Lian]{Xin2019} Xin,~G.; Zhu,~W.; Deng,~Y.; Cheng,~J.; Zhang,~L.~T.; Chung,~A.~J.; De,~S.; Lian,~J. {Microfluidics-enabled orientation and microstructure control of macroscopic graphene fibres}. \emph{Nature Nanotechnology} \textbf{2019}, \emph{14}, 168--175\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[{De France} \latin{et~al.}(2017){De France}, Yager, Chan, Corbett, Cranston, and Hoare]{DeFrance2017} {De France},~K.~J.; Yager,~K.~G.; Chan,~K. J.~W.; Corbett,~B.; Cranston,~E.~D.; Hoare,~T. {Injectable anisotropic nanocomposite hydrogels direct in situ growth and alignment of myotubes}. \emph{Nano Letters} \textbf{2017}, \emph{17}, 6487--6495\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Doi and Edwards(1978)Doi, and Edwards]{LangStiffness} Doi,~M.; Edwards,~S.~F. {Dynamics of rod-like macromolecules in concentrated solution. Part 1—{B}rownian motion in the equilibrium state }. \emph{Journal of the Chemical Society, Faraday Transactions 2: Molecular and Chemical Physics} \textbf{1978}, \emph{74}, 560--570\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Vermant \latin{et~al.}(2001)Vermant, Yang, and Fuller]{Vermant2001} Vermant,~J.; Yang,~H.; Fuller,~G.~G. {Rheo-optical determination of aspect ratio and polydispersity of nonspherical particles}. \emph{AIChE Journal} \textbf{2001}, \emph{47}, 790--798\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Dhont and Briels(2007)Dhont, and Briels]{Dhont2007} Dhont,~J. K.~G.; Briels,~W.~J. \emph{Soft Matter Vol. 2}; Wiley-VCH: Weinheim, Germany, 2007; Vol.~2; pp 216--283\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Reddy \latin{et~al.}(2018)Reddy, Natale, Prud'homme, and Vermant]{Reddy2018} Reddy,~N.~K.; Natale,~G.; Prud'homme,~R.~K.; Vermant,~J. {Rheo-optical analysis of functionalized graphene suspensions}. \emph{Langmuir} \textbf{2018}, \emph{34}, 7844--7851\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Winkler \latin{et~al.}(2004)Winkler, Mussawisade, Ripoll, and Gompper]{Winkler2004} Winkler,~R.~G.; Mussawisade,~K.; Ripoll,~M.; Gompper,~G. {Rod-like colloids and polymers in shear flow: a multi-particle-collision dynamics study}. \emph{Journal of Physics: Condensed Matter} \textbf{2004}, \emph{16}, 3941--3954\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Haward \latin{et~al.}(2016)Haward, McKinley, and Shen]{Haward2016} Haward,~S.~J.; McKinley,~G.~H.; Shen,~A.~Q. {Elastic instabilities in planar elongational flow of monodisperse polymer solutions}. \emph{Scientific Reports} \textbf{2016}, \emph{6}, 33029\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Lang \latin{et~al.}(2019)Lang, Hendricks, Zhang, Reddy, Rothstein, Lettinga, Vermant, and Clasen]{Lang2019a} Lang,~C.; Hendricks,~J.; Zhang,~Z.; Reddy,~N.~K.; Rothstein,~J.~P.; Lettinga,~M.~P.; Vermant,~J.; Clasen,~C. {Effects of particle stiffness on the extensional rheology of model rod-like nanoparticle suspensions}. \emph{Soft Matter} \textbf{2019}, \emph{15}, 833--841\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Trebbin \latin{et~al.}(2013)Trebbin, Steinhauser, Perlich, Buffet, Roth, Zimmermann, Thiele, and F{\"{o}}rster]{Trebbin2013} Trebbin,~M.; Steinhauser,~D.; Perlich,~J.; Buffet,~A.; Roth,~S.~V.; Zimmermann,~W.; Thiele,~J.; F{\"{o}}rster,~S. {Anisotropic particles align perpendicular to the flow direction in narrow microchannels}. \emph{Proceedings of the National Academy of Sciences of the United States of America} \textbf{2013}, \emph{110}, 6706--6711\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Qazi \latin{et~al.}(2011)Qazi, Rennie, Tucker, Penfold, and Grillo]{Qazi2011} Qazi,~S. J.~S.; Rennie,~A.~R.; Tucker,~I.; Penfold,~J.; Grillo,~I. {Alignment of dispersions of plate-like colloidal particles of Ni(OH)$_2$ induced by elongational flow}. \emph{Journal of Physical Chemistry B} \textbf{2011}, \emph{115}, 3271--3280\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Pignon \latin{et~al.}(2003)Pignon, Magnin, Piau, and Fuller]{Pignon2003} Pignon,~F.; Magnin,~A.; Piau,~J.-M.; Fuller,~G.~G. {The orientation dynamics of rigid rod suspensions under extensional flow}. \emph{Journal of Rheology} \textbf{2003}, \emph{47}, 371--388\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Corona \latin{et~al.}(2018)Corona, Ruocco, Weigandt, Leal, and Helgeson]{Corona2018} Corona,~P.~T.; Ruocco,~N.; Weigandt,~K.~M.; Leal,~L.~G.; Helgeson,~M.~E. {Probing flow-induced nanostructure of complex fluids in arbitrary 2D flows using a fluidic four-roll mill (FFoRM)}. \emph{Scientific Reports} \textbf{2018}, \emph{8}, 15559\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Ros{\'{e}}n \latin{et~al.}(2020)Ros{\'{e}}n, Mittal, Roth, Zhang, Lundell, and S{\"{o}}derberg]{Rosen2020a} Ros{\'{e}}n,~T.; Mittal,~N.; Roth,~S.~V.; Zhang,~P.; Lundell,~F.; S{\"{o}}derberg,~L.~D. {Flow fields control nanostructural organization in semiflexible networks}. \emph{Soft Matter} \textbf{2020}, \emph{16}, 5439--5449\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Hasegawa \latin{et~al.}(2020)Hasegawa, Horikawa, and Shikata]{Hasegawa2020a} Hasegawa,~H.; Horikawa,~Y.; Shikata,~T. {Cellulose nanocrystals as a model substance for rigid rod particle suspension rheology}. \emph{Macromolecules} \textbf{2020}, \emph{53}, 2677--2685\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Bertsch \latin{et~al.}(2017)Bertsch, Isabettini, and Fischer]{Bertsch2017} Bertsch,~P.; Isabettini,~S.; Fischer,~P. {Ion-induced hydrogel formation and nematic ordering of nanocrystalline cellulose suspensions}. \emph{Biomacromolecules} \textbf{2017}, \emph{18}, 4060--4066\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Bertsch \latin{et~al.}(2019)Bertsch, S{\'{a}}nchez-Ferrer, Bagnani, Isabettini, Kohlbrecher, Mezzenga, and Fischer]{Bertsch2019} Bertsch,~P.; S{\'{a}}nchez-Ferrer,~A.; Bagnani,~M.; Isabettini,~S.; Kohlbrecher,~J.; Mezzenga,~R.; Fischer,~P. {Ion-induced formation of nanocrystalline cellulose colloidal glasses containing nematic domains}. \emph{Langmuir} \textbf{2019}, \emph{35}, 4117--4124\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Reid \latin{et~al.}(2017)Reid, Villalobos, and Cranston]{Reid2017} Reid,~M.~S.; Villalobos,~M.; Cranston,~E.~D. {Benchmarking cellulose nanocrystals: From the laboratory to industrial production}. \emph{Langmuir} \textbf{2017}, \emph{33}, 1583--1598\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Usov and Mezzenga(2015)Usov, and Mezzenga]{Usov2015a} Usov,~I.; Mezzenga,~R. {FiberApp: An open-source software for tracking and analyzing polymers, filaments, biomacromolecules, and fibrous objects}. \emph{Macromolecules} \textbf{2015}, \emph{48}, 1269--1280\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Haward \latin{et~al.}(2012)Haward, Oliveira, Alves, and McKinley]{Haward2012a} Haward,~S.~J.; Oliveira,~M. S.~N.; Alves,~M.~A.; McKinley,~G.~H. {Optimized cross-slot flow geometry for microfluidic extensional rheometry}. \emph{Physical Review Letters} \textbf{2012}, \emph{109}, 128301\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Burshtein \latin{et~al.}(2019)Burshtein, Chan, Toda-Peters, Shen, and Haward]{Burshtein2019} Burshtein,~N.; Chan,~S.~T.; Toda-Peters,~K.; Shen,~A.~Q.; Haward,~S.~J. {3D-printed glass microfluidics for fluid dynamics and rheology}. \emph{Current Opinion in Colloid and Interface Science} \textbf{2019}, \emph{43}, 1--14\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Haward \latin{et~al.}(2019)Haward, Kitajima, Toda-Peters, Takahashi, and Shen]{Haward2019} Haward,~S.~J.; Kitajima,~N.; Toda-Peters,~K.; Takahashi,~T.; Shen,~A.~Q. {Flow of wormlike micellar solutions around microfluidic cylinders with high aspect ratio and low blockage ratio}. \emph{Soft Matter} \textbf{2019}, \emph{15}, 1927--1941\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Meinhart \latin{et~al.}(2000)Meinhart, Wereley, and Gray]{Meinhart2000} Meinhart,~C.~D.; Wereley,~S.; Gray,~M. {Volume illumination for two-dimensional particle image velocimetry}. \emph{Measurement Science and Technology} \textbf{2000}, \emph{11}, 809--814\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Wagner \latin{et~al.}(2010)Wagner, Raman, and Moon]{RyanWagnerArvindRaman2010} Wagner,~R.; Raman,~A.; Moon,~R.~J. \emph{10th International Conference on Wood {\&} Biofiber Plastic Composites}; 2010; pp 309--316\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Lang \latin{et~al.}(2019)Lang, Kohlbrecher, Porcar, Radulescu, Sellinghoff, Dhont, and Lettinga]{Lang2019} Lang,~C.; Kohlbrecher,~J.; Porcar,~L.; Radulescu,~A.; Sellinghoff,~K.; Dhont,~J. K.~G.; Lettinga,~M.~P. {Microstructural understanding of the length- and stiffness-dependent shear thinning in semidilute colloidal rods}. \emph{Macromolecules} \textbf{2019}, \emph{52}, 9604--9612\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Usov \latin{et~al.}(2015)Usov, Nystr{\"{o}}m, Adamcik, Handschin, Sch{\"{u}}tz, Fall, Bergstr{\"{o}}m, and Mezzenga]{Usov2015} Usov,~I.; Nystr{\"{o}}m,~G.; Adamcik,~J.; Handschin,~S.; Sch{\"{u}}tz,~C.; Fall,~A.; Bergstr{\"{o}}m,~L.; Mezzenga,~R. {Understanding nanocellulose chirality and structure–properties relationship at the single fibril level}. \emph{Nature Communications} \textbf{2015}, \emph{6}, 7564\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Lang and Lettinga(2020)Lang, and Lettinga]{Lang2020b} Lang,~C.; Lettinga,~M.~P. {Shear flow behavior of bidisperse rodlike colloids}. \emph{Macromolecules} \textbf{2020}, \emph{53}, 2662--2668\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Tanaka \latin{et~al.}(2014)Tanaka, Saito, Ishii, and Isogai]{Tanaka2014} Tanaka,~R.; Saito,~T.; Ishii,~D.; Isogai,~A. {Determination of nanocellulose fibril length by shear viscosity measurement}. \emph{Cellulose} \textbf{2014}, \emph{21}, 1581--1589\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Kobayashi and Yamamoto(2011)Kobayashi, and Yamamoto]{Kobayashi2011} Kobayashi,~H.; Yamamoto,~R. {Reentrant transition in the shear viscosity of dilute rigid-rod dispersions}. \emph{Physical Review E} \textbf{2011}, \emph{84}, 051404\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Lang \latin{et~al.}(2016)Lang, Kohlbrecher, Porcar, and Lettinga]{Lang2016} Lang,~C.; Kohlbrecher,~J.; Porcar,~L.; Lettinga,~M. {The connection between biaxial orientation and shear thinning for quasi-ideal rods}. \emph{Polymers} \textbf{2016}, \emph{8}, 291\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Shah and London(1978)Shah, and London]{Shah} Shah,~R.; London,~A. In \emph{{Laminar Flow Forced Convection in Ducts: a Source Book for Compact Heat Exchanger Analytical Data, Academic Press, New York, 1978.}}; Hartnett,~T. I. J.~P., Ed.; Academic Press, 1978\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Haward(2016)]{Hawarda2016} Haward,~S.~J. {Microfluidic extensional rheometry using stagnation point flow}. \emph{Biomicrofluidics} \textbf{2016}, \emph{10}, 043401\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Sun and Huang(2016)Sun, and Huang]{Sun2016} Sun,~C.-L.; Huang,~H.-Y. {Measurements of flow-induced birefringence in microfluidics}. \emph{Biomicrofluidics} \textbf{2016}, \emph{10}, 011903\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Ros{\'{e}}n \latin{et~al.}(2020)Ros{\'{e}}n, Wang, Zhan, He, Chodankar, and Hsiao]{Rosen2020} Ros{\'{e}}n,~T.; Wang,~R.; Zhan,~C.; He,~H.; Chodankar,~S.; Hsiao,~B.~S. {Cellulose nanofibrils and nanocrystals in confined flow: Single-particle dynamics to collective alignment revealed through scanning small-angle X-ray scattering and numerical simulations}. \emph{Physical Review E} \textbf{2020}, \emph{101}, 032610\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Mewis \latin{et~al.}(1997)Mewis, Mortier, Vermant, and Moldenaers]{Mewis1997} Mewis,~J.; Mortier,~M.; Vermant,~J.; Moldenaers,~P. {Experimental evidence for the existence of a wagging regime in polymeric liquid crystals}. \emph{Macromolecules} \textbf{1997}, \emph{30}, 1323--1328\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Ebeling \latin{et~al.}(1999)Ebeling, Paillet, Borsali, Diat, Dufresne, Cavaill{\'{e}}, and Chanzy]{Ebeling1999} Ebeling,~T.; Paillet,~M.; Borsali,~R.; Diat,~O.; Dufresne,~A.; Cavaill{\'{e}},~J.~Y.; Chanzy,~H. {Shear-induced orientation phenomena in suspensions of cellulose microcrystals, revealed by small angle X-ray scattering}. \emph{Langmuir} \textbf{1999}, \emph{15}, 6123--6126\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \bibitem[Lerouge and Berret(2009)Lerouge, and Berret]{Mu2019} Lerouge,~S.; Berret,~J.-F. In \emph{Polymer Characterization: Rheology, Laser Interferometry, Electrooptics}; Dusek,~K., Joanny,~J.-F., Eds.; Advances in Polymer Science 12; Springer Berlin Heidelberg: Berlin, Heidelberg, 2009; Vol.~58; pp 1--71\relax \mciteBstWouldAddEndPuncttrue \mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct} {\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax \unskip.} \end{mcitethebibliography} \end{document}

1. Introduction {#sec1-cancers-11-00088} =============== Glioblastoma multiforme (GBM), the most malignant type of glioma, is the most common type of brain tumor, and it is associated with a low median survival of 9--15 months \[[@B1-cancers-11-00088]\]. The bioactive lipid molecule, 2OHOA, was shown to be safe in patients with glioma and other advanced solid tumors in a phase I/IIa clinical trial ([ClinicalTrials.gov](ClinicalTrials.gov) identifier \#NCT01792310). In a xenograft model of human GBM, 2OHOA provided a greater anti-tumor effect than temozolomide (TMZ), which is the current standard-of-care first-line chemotherapy against GBM and increases patient's survival by ca. 2.5 months \[[@B2-cancers-11-00088],[@B3-cancers-11-00088]\]. In mice, tumors relapsed when treated with TMZ, but not in mice treated with 2OHOA. 2OHOA is a hydroxylated fatty acid that produces dramatic changes in membrane lipid composition and organization \[[@B4-cancers-11-00088],[@B5-cancers-11-00088]\], activating sphingomyelin (SM) synthesis \[[@B6-cancers-11-00088]\]. Sphingomyelin synthase (SMS) is an enzyme that generates SM. The two main SMS isoforms, SMS1 and SMS2, share significant homology, except for a sterile alpha motif (SAM), present only in SMS1 ([Figure S1](#app1-cancers-11-00088){ref-type="app"}), which can interact with regulatory proteins, lipids and RNA \[[@B7-cancers-11-00088]\]. In addition, the SMS isoforms differ in their subcellular localization: SMS1 is located in the Golgi apparatus, and SMS2 in the Golgi apparatus and the plasma membrane \[[@B8-cancers-11-00088]\]. Most SM is synthetized by SMS1, which is responsible for the 60--80% of the SMS activity \[[@B9-cancers-11-00088],[@B10-cancers-11-00088]\]. SMSs can reversibly catalyze SM synthesis by transferring phosphocholine or phosphoethanolamine from phosphatidylcholine or phosphatidylethanolamine onto ceramide, yielding SM and diacylglycerol \[[@B11-cancers-11-00088],[@B12-cancers-11-00088]\]. Ceramide and diacylglycerol are widely known for their pro- and anti-apoptotic roles, respectively \[[@B13-cancers-11-00088],[@B14-cancers-11-00088]\]. In addition, SM and cholesterol are the main components of lipid rafts, specialized regions of the plasma membrane with important roles in signal transduction and membrane trafficking \[[@B15-cancers-11-00088],[@B16-cancers-11-00088]\]. Therefore, regulation of SM by 2OHOA has relevant consequences for cancer cell signaling and may explain 2OHOA antitumor pharmacological effects. Neither the role of SMS1 and SMS2 in tumor progression nor the effect of 2OHOA on each SMS isozymes are fully understood. A useful tool to identify potential pharmacological targets for brain tumors is the REpository for Molecular BRAin Neoplasia DaTa (REMBRANDT) database, which integrates information on gene expression and survival \[[@B17-cancers-11-00088]\]. Because 2OHOA has relevant antitumor activity against glioma, and given that its mechanism of action involves SMS activation and altered SMS1 expression is associated with glioma patient survival, we studied the role of these enzymes in glioma tumorigenesis, prognosis and response to 2OHOA \[[@B3-cancers-11-00088],[@B18-cancers-11-00088],[@B19-cancers-11-00088]\]. In this study, we found that SMS1 and SMS2 expression had an opposing impact on patient survival, which is in line with their divergent effects on GBM cell survival and is consistent with 2OHOA-induced activation of SMS1. Whether both SMS isozymes catalyze similar reactions or differ in their activity in resting and growing (e.g., cancer) cells requires additional investigation. The present study further demonstrates that SMS isoforms are valuable anticancer drug targets and tumor biomarkers. 2. Results {#sec2-cancers-11-00088} ========== 2.1. SMS1 and SMS2 Expression Have Opposite Effects on Glioma Patient Survival {#sec2dot1-cancers-11-00088} ------------------------------------------------------------------------------ Using the REMBRANDT database, we investigated the expression of SMS1 and SMS2 in all the types of glioma. BETASTASIS website provides gene expression and survival data from 329 glioma patients, of which 54 patients were alive, and therefore censored and only used for expression analyses, and not in survival studies. Kaplan-Meier plots were used to assess the median survival for patient groups with differential SMS1 or SMS2 gene expression and significant differences in survival (*p* \< 0.05) were determined using the log-rank test. When analyzing SMS1 and SMS2 gene expression, glioma patients were classified into 2 groups defined by high or low SMS1 or SMS2 expression with respect to the median expression for each gene. Patients with high SMS1 expression (SMS1-high) had longer median survival than patients with low SMS1 expression (SMS1-low; [Figure 1](#cancers-11-00088-f001){ref-type="fig"}A,C). Compared to normal tissue from healthy subjects, SMS1 had a significantly lower expression in GBM patients ([Figure 1](#cancers-11-00088-f001){ref-type="fig"}D), thus indicating its implication in tumorigenesis and its value as a prognostic factor. Conversely, patients with low SMS2 expression had longer median survival than patients with high SMS2 expression ([Figure 1](#cancers-11-00088-f001){ref-type="fig"}B,C). Consistent with the median survival, the two-year and five-year survival rates were higher for SMS1-high or SMS2-low groups, whereas patients in SMS1-low or SMS2-high groups had a worse prognosis ([Figure 1](#cancers-11-00088-f001){ref-type="fig"} C,E,F). The importance of SMS1 expression for patient survival was confirmed in other databases for glioma (GSE4412 database, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}A) and GBM (TCGA database, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}B) and in other cancer types analyzed: Normal acute myeloid leukemia (CN-AML, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}C), gastric cancer ([Figure 2](#cancers-11-00088-f002){ref-type="fig"}D), kidney renal clear cell carcinoma (KIRC, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}E), lung cancer, including NSCLC (non-small cell lung cancer) and SCLC (small cell lung cancer) ([Figure 2](#cancers-11-00088-f002){ref-type="fig"}F), lymphoma (B-cell lymphoma and Burkitt's Lymphoma, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}G), pancreatic ductal adenocarcinoma (PDAC, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}H), sarcoma ([Figure 2](#cancers-11-00088-f002){ref-type="fig"}I) and skin cutaneous melanoma (SKCM, [Figure 2](#cancers-11-00088-f002){ref-type="fig"}J) ([Table 1](#cancers-11-00088-t001){ref-type="table"}). Despite similar enzymatic activity, SMS1 and SMS2 have a divergent influence on glioma patients' overall survival, and this divergence is possibly associated with the different roles in cancer (and possibly normal) cell differentiation, proliferation and survival. 2.2. SMS1/SMS2 Ratio is Relevant for Glioma Patient Survival and Correlates with 2OHOA's Pharmacological Efficacy {#sec2dot2-cancers-11-00088} ----------------------------------------------------------------------------------------------------------------- To evaluate the combined influence of SMS1 plus SMS2 gene expression in glioma patients, 4 groups were evaluated (SMS1-high plus SMS2-low, SMS1-low plus SMS2-low, SMS1-high plus SMS2-high and SMS1-low plus SMS2-high), and significant differences were observed ([Figure 3](#cancers-11-00088-f003){ref-type="fig"}A). SMS1-high plus SMS2-low group had a 5-year survival of 42.7%, whereas SMS1-low plus SMS2-high group had a 5-year survival of 4.6% ([Figure 3](#cancers-11-00088-f003){ref-type="fig"}B,C). These were the best and worst median survival groups, respectively, whereas the other 2 groups displayed an intermediate median survival. To assess the functional relevance of the previous findings in patients, in vitro cultures of several glioma and other tumor cell lines were used. To investigate the divergent effects of SMS isoforms on human GBM (U118) cell proliferation, gain-of-function assays for SMS1 and SMS2 were performed. To this end, cells were transfected with the Tet-On/Off 3G system containing either SMS1 or SMS2 cDNA, which induced SMS isozyme overexpression upon doxycycline addition to the culture medium ([Figures S2 and S3](#app1-cancers-11-00088){ref-type="app"}). U118 cells overexpressing SMS1 showed a 35% lower viability than non-induced control cells. In contrast, cells overexpressing SMS2 showed higher viability (20%) than control GBM cells ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}A). Human GBM (SF295) cells exhibited a similar viability behavior after overexpression of SMS1 and SMS2 ([Figure S4](#app1-cancers-11-00088){ref-type="app"}). These results indicated that SMS1 and SMS2 fulfill different roles in glioma cell growth and possibly in tumorigenesis. Loss-of-function experiments were carried out inducing 48-h SMS1 and SMS2 silencing by means of specific siRNAs. In both cases, compromised cell viability was observed, which is in agreement with the proliferative role of SMS2 (−89.7% and −48.1%, respectively, [Figure S5](#app1-cancers-11-00088){ref-type="app"}). The effect observed for SMS1 could be due to the high level of silencing achieved as discussed below. In addition, SMS1 and SMS2 protein levels were evaluated in several cancer cell types after exposure to 2OHOA. SMS1 and SMS2 were measured after 48 h in the presence or absence (control, 100%) of 2OHOA (200 µM) in 16 cancer cell lines ([Figure 4](#cancers-11-00088-f004){ref-type="fig"} and [Figure S6](#app1-cancers-11-00088){ref-type="app"}), and the IC~50~ for 2OHOA was also determined in these cell lines ([Table S1](#app1-cancers-11-00088){ref-type="app"}). An inverse correlation was observed between the change in SMS1 protein levels after 2OHOA treatment and their IC~50~ values in 7 human GBM cell lines ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}B left panel). Similar results were obtained when this correlation was assessed in another 9 non-glioma cancer cell lines ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}B right panel). 2OHOA-mediated increase in SMS1 expression correlated with lower IC~50~ values (i.e., stronger antitumor effect). In addition, cell lines with higher increase in SMS1 expression after 2OHOA treatment displayed the lowest basal SMS1 mRNA expression levels ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}D). Moreover, SMS1/SMS2 ratio inversely correlated with IC~50~ values for 2OHOA ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}C and [Table S2](#app1-cancers-11-00088){ref-type="app"}). These results indicated that glioma cell lines responded better to drug treatment upon SMS1 increase, SMS2 decrease or both. Accordingly, the ratio of these two proteins represents a relevant predictor of drug potency and efficacy. Indeed, an inverse correlation between SMS1 and SMS2 expression was observed when their mRNA levels were analyzed in 386 glioma patients (*p* \< 0.0001; [Figure 4](#cancers-11-00088-f004){ref-type="fig"}E), thus further supporting the clinical relevance of the SMS1/SMS2 ratio in glioma progression and prognosis. 2.3. Basal SMS1 mRNA Levels Anticipate 2OHOA Anti-Cancer Activity against GBM {#sec2dot3-cancers-11-00088} ----------------------------------------------------------------------------- Basal SMS1 and SMS2 mRNA levels were evaluated in 6 human glioma cell lines, four responders and two non-responders (IC~50~ \> 600 µM) to 2OHOA treatment and in one non-tumor cell line, (MRC-5). In these cell lines, the mRNA levels of SMS1 but not of SMS2 strongly correlated with the IC~50~ values for 2OHOA ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}D left panel) and similar results were observed in other cancer cell lines ([Figure 4](#cancers-11-00088-f004){ref-type="fig"}D right panel). These results suggest that basal SMS1 mRNA levels anticipate whether 2OHOA treatment might be efficacious. Moreover, all GBM cell lines investigated had lower SMS1 expression values (0.2--4.5 a.u.) than non-cancer MRC5 cells (7.5 a.u.: [Table S3](#app1-cancers-11-00088){ref-type="app"}), similarly as observed in human brain tissue from glioma patients or healthy subjects ([Figure 1](#cancers-11-00088-f001){ref-type="fig"}D). 2.4. Opposing Influence of SMS1 and SMS2 on 2OHOA Regulation of GBM Cell Proliferation and Differentiation {#sec2dot4-cancers-11-00088} ---------------------------------------------------------------------------------------------------------- To evaluate the implication of SMS1 and SMS2 on 2OHOA's pharmacological mechanism of action, and their differential role on cell viability, overexpression and silencing experiments were carried out on U118 cells in the presence (treated) or absence (control) of 2OHOA. With this aim, we analyzed the relevance of these isoforms in known events of 2OHOA's mechanism of action. Exposure to 2OHOA induced a significant decrease of dihydrofolate reductase (DHFR) levels in several cancer cell types, including U118 glioma cells, ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}A) \[[@B19-cancers-11-00088],[@B20-cancers-11-00088]\]. The proliferation biomarker (DHFR) decreased 38% when SMS1 was overexpressed and a further decrease to 90% was evident in the presence of 2OHOA ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}A). However, siRNA-mediated SMS1 downregulation had no significant effect on DHFR levels ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}B). In contrast, SMS2 overexpression had no significant effect on DHFR in glioma (U118) cells ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}A), and it produced a significant DHFR increase in SF295 and U251 cells ([Figure S7](#app1-cancers-11-00088){ref-type="app"}). Conversely, a 70% decrease in DHFR was observed after siRNA-mediated SMS2 downregulation in U118 cells ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}B). It is possible that the effect of 2OHOA on DHFR levels masked any possible effect due to SMSs overexpression. In line with these results, SMS1 silencing prevented 2OHOA-induced DHFR reduction, while SMS2 silencing had no effect on 2OHOA-induced decrease of DHFR levels. These results indicate that 2OHOA-induced activation of SMS1 mediates the ensuing decrease in DHFR levels. The increase of N-Cadherin inversely correlates with glioma cell invasive capacity and malignant behavior \[[@B21-cancers-11-00088]\], and this increase is associated with glioma cell differentiation \[[@B22-cancers-11-00088]\]. In this context, N-Cadherin levels increased 65% and 30% upon SMS1 or SMS2 overexpression respectively, although a distinct N-Cadherin cellular distribution pattern was observed by confocal microscopy depending on the isoform overexpressed ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}C,E). A 2-fold increase in N-Cadherin was observed after 2OHOA treatment, an effect that was maintained when SMS1 was overexpressed but was abolished when SMS2 was overexpressed ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}C). Accordingly, following siRNA-mediated SMS1 silencing there was a 45% decrease in N-Cadherin, whereas SMS2 downregulation caused a 30% increase in the levels of this protein ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}D). Moreover, silencing of SMS1, but not SMS2, abolished 2OHOA's effect on N-Cadherin expression. These results further demonstrate the opposing influence of both SMS isoforms on cell proliferation, differentiation and the cellular response to 2OHOA. Numerous studies have previously shown that alterations in the expression of a receptor subtype or isozyme of a given family can be compensated by the expression of another isoform. To address this possibility, levels of endogenous SMS1 and SMS2 mRNA resulting from different silencing or overexpression technics were evaluated ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}F,G). RNA silencing of SMS1 caused an increase in the expression of SMS2 mRNA levels, but the reverse situation did not occur, silencing of SMS2 did not change SMS1 expression. Interestingly, SMS2 overexpression led to a decrease in SMS1 expression, while increasing SMS1 expression had no effect on SMS2 expression. However, siRNAs treatments did not result in a strong decrease of SMS protein levels, suggesting that small changes in protein amount were enough to produce these effects, or that the protein present was not fully active. The lack of correlation between mRNA levels ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}F,G) and protein levels ([Figure S5](#app1-cancers-11-00088){ref-type="app"}) after siRNAs treatment was previously reported for SMS1 \[[@B23-cancers-11-00088]\]. 2.5. SMS1 and 2OHOA Modulate the β-Catenin Pathway {#sec2dot5-cancers-11-00088} -------------------------------------------------- Since the interaction between N-Cadherin and β-catenin modulates the β-catenin signaling pathway \[[@B24-cancers-11-00088],[@B25-cancers-11-00088]\], we investigated β-catenin response to 2OHOA treatment in the context of altered SMS expression. Abnormal activation of the Wnt/β-catenin pathway has been associated with GBM progression \[[@B26-cancers-11-00088]\] and inhibition of the pathway by siRNA suppresses malignant glioma cell growth \[[@B27-cancers-11-00088]\]. Upregulation of N-Cadherin is associated with inhibition of the β-catenin pathway, and when N-Cadherin/β-catenin complexes are cleaved at the plasma membrane, β-catenin is intensely phosphorylated at Thr41 and subsequently degraded \[[@B28-cancers-11-00088]\]. β-catenin may be localized at the plasma membrane in association with N-Cadherin, in the cytoplasm where it is usually degraded, or in the nucleus, where it acts as a transcription factor to activate the expression of genes related to cell proliferation, such as AXIN2, Cyclin D1 or c-myc \[[@B29-cancers-11-00088],[@B30-cancers-11-00088],[@B31-cancers-11-00088]\]. Here, we used AXIN2 mRNA expression as an indicator of β-catenin activity. Again, SMS1 and SMS2 displayed an inverse influence on this pathway ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}H). Thus, activation of the β-catenin pathway was enhanced 82% when SMS1 was silenced, whereas either SMS1 overexpression or treatment with 2OHOA induced inhibition of this pathway (34% and 51%, respectively; [Figure 5](#cancers-11-00088-f005){ref-type="fig"}H). Conversely, SMS2 silencing or overexpression did not alter the activation of the β-catenin pathway ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}H). In addition, the distribution of β-catenin protein was altered when SMS1 was overexpressed ([Figure 5](#cancers-11-00088-f005){ref-type="fig"}I), resulting in a clear cytoplasmic distribution and a strongly decreased presence in the nucleus. 2.6. SMS1 and SMS2 Modulate ER Stress but Only SMS1 Affects Autophagy {#sec2dot6-cancers-11-00088} --------------------------------------------------------------------- 2OHOA activates ER stress and macroautophagy in glioma cells \[[@B32-cancers-11-00088],[@B33-cancers-11-00088]\]. An immediate response under ER stress is a general protein synthesis inhibition mechanism due to increased phosphorylation of the α subunit of the translation initiation factor eIF2α at Ser51 \[[@B34-cancers-11-00088]\]. In turn, this effect mediates conversion of light chain 3 (LC3) from the soluble LC3-I to the membrane-bound LC3-II form, a key step in autophagy induction in mammalian cells \[[@B35-cancers-11-00088]\]. These two ER stress markers (LC3 and eIF2α) were assessed and the levels of phosphorylated eIF2α increased after exposure to 2OHOA \[[@B32-cancers-11-00088]\]. Overexpression of either SMS1 ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}A, left) or SMS2 ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}A, right) mimicked the increase in phospho-eIF2α induced by 2OHOA, whereas downregulation of either SMS1 or SMS2 decreased phosphorylation of eIF2α by 35% and 55%, respectively ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}B). Furthermore, phosphorylation of eIF2α induced by 2OHOA was maintained following SMS1 or SMS2 overexpression ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}A), although it was abolished when either SMS1 or SMS2 were silenced ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}B). SMS1, but not SMS2, overexpression and 2OHOA provoked increase in the LC3B II-to-LC3B I ratio, ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}C). Confocal microscopy revealed the cytoplasmic aggregation of LC3 upon SMS1 overexpression ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}E). In contrast, LC3B II-to-LC3B I ratio decreased by 30% in response to SMS1 downregulation ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}D). Furthermore, the 2-fold increase in LC3B II-to-LC3B I ratio induced by 2OHOA was maintained when SMS1 and SMS2 were overexpressed, and it was abolished when SMSs were silenced upon siRNA treatment. 2.7. 2OHOA Increases SMS Activity through SMS1 Exclusively {#sec2dot7-cancers-11-00088} ---------------------------------------------------------- To determine if the overexpressed SMS1 and SMS2 proteins were enzymatically active, conversion of NBD-C6-Ceramide into NBD-C6-SM was measured. A clear increase in protein activity was observed when SMS1 or SMS2 were overexpressed ([Figure S3](#app1-cancers-11-00088){ref-type="app"}), thus demonstrating that both overexpressed proteins were active. The same activity assay was investigated when SMS1, SMS2 or both were silenced using specific siRNAs. Only when both SMSs were silenced total SMS activity was blocked, which suggested a compensatory mechanism whereby only one SMS is needed to maintain SM production. Interestingly, this compensatory activation would only rely on enzymatic activity since their expression was not significantly affected ([Figure S5](#app1-cancers-11-00088){ref-type="app"}). In the presence of 2OHOA, the protein activity assay showed a clear increase of SMS activity, except when SMS1 or both SMSs were silenced. These results demonstrate that 2OHOA induction of SM synthesis requires SMS1, and further suggests that 2OHOA primarily interacts with SMS1 to increase SMS enzymatic activity ([Figure 6](#cancers-11-00088-f006){ref-type="fig"}F). 3. Discussion {#sec3-cancers-11-00088} ============= In the present study, we assessed the influence of SMS1 and SMS2 on glioma patients' survival, cancer cell viability and responses to the antitumor drug, 2OHOA. We showed that either high SMS1 or low SMS2 mRNA levels represented a good prognostic factor in glioma. Moreover, the combination of SMS1-high plus SMS2-low expression resulted in a significant and marked increase in median survival compared to the other expression profiles. By contrast, SMS1-low and SMS2-high expression, independently or combined, was associated with poor prognosis and significantly worse survival parameters. In addition, we reported that the pharmacological effect of 2OHOA is due to the increase of SMS1 expression and enzymatic activity, and simultaneous reduction of SMS2 in glioma cells, which regulate cell growth and patient prognostic. SMS1 expression showed robust association with glioma patients' survival as similar results were observed in 2 different databases, REMBRANDT ([Figure 1](#cancers-11-00088-f001){ref-type="fig"}) and GSE4412 ([Figure 2](#cancers-11-00088-f002){ref-type="fig"}). The effect of SMSs overexpression on U118 GBM cells in vitro was in line with the clinical data that showed a decreased or increased in cell proliferation following SMS1 or SMS2 overexpression, respectively. For example, a reduced proliferation capacity in glioma cells overexpressing SMS1 would agree with longer survival in glioma patients with high SMS1 expression. However, when SMS1 and SMS2 were silenced using specific siRNAs, cell viability decreased in both cases, and this is consistent with the clinical response to low SMS2 levels, but not with the response to low SMS1 levels of expression. Nevertheless, this discrepancy may reflect the extremely efficient in vitro silencing of SMS1, which achieved an inhibition of expression stronger than any differences found in patients. Thus, although the lowest SMS1 expression was about 57%, the highest expression measured in glioma patients ([Figure S8](#app1-cancers-11-00088){ref-type="app"}), siSMS1 caused a reduction to 10% of the control SMS1 expression in vitro, potentially compromising GBM cell survival. This premise is supported by the fact that SMS1 KO mice show deteriorated growth and male infertility, and they are born in a lower Mendelian ratio \[[@B36-cancers-11-00088],[@B37-cancers-11-00088]\], which indicates the relevance of SMS1 expression during development. These novel findings show the opposing roles of these two SMS isoforms in GBM cells and glioma patients, and that the antitumor drug 2OHOA, which has demonstrated safety and efficacy in phase I/IIA clinical trial, can differentially modulate SMS1, a potential new anticancer drug target. We also found a strong positive correlation between basal SMS1 mRNA levels and the pharmacological potency of 2OHOA based on IC~50~ values. These results suggest a direct interaction between 2OHOA and SMS1 and support the use of SMS1 basal expression as a prognostic biomarker to identify potential responders to 2OHOA treatment. This hypothesis is further supported by the facts that (1) only GBM cell lines responding to 2OHOA treatment increase SMS1 expression, and (2) the effect of increasing SMS activity by either SMS1 overexpression or 2OHOA treatment produced similar glioma cell growth inhibition. In fact, the lower SMS1 mRNA levels in human glioma cells relative to non-glioma cancer cells (both in vitro and in vivo) can explain the enhanced efficacy of 2OHOA against this type of cancer. Finally, the fact that induction of SM synthesis by 2OHOA was observed only in cells expressing SMS1 (independently of SMS2 expression) suggests a specific and direct interaction of 2OHOA with SMS1, particularly considering the short times of treatment needed to produce this increase in SMS activity. A recently published study questions the activation of SMS activity in response to treatment with 2OHOA \[[@B38-cancers-11-00088]\]. However, substantial differences in the protocol used in this publication compared to previous studies \[[@B19-cancers-11-00088]\] might well explain the divergent results obtained, such as the method of dissolution of the drug or the direct addition of the substrate NBD-Cer instead of adding it in the form of LUVs, among others. In this work \[[@B19-cancers-11-00088]\], 2OHOA treatments, performed between 5 min and 24 h, indeed increased SMS activity, and now we provide further detail on this effect showing that SMS activity increase is exclusively due to the activation of SMS1 and that this effect fades away for longer incubation times. GBM cell differentiation was assessed using N-Cadherin regulation as indicator, which was enhanced by SMS1, 2OHOA or both. SMS2 overexpression dampened SMS1 expression, which could in part explain the lack of effects of 2OHOA on N-Cadherin when SMS2 was overexpressed. Not only were the levels of N-Cadherin induced by 2OHOA relevant to cell differentiation, but its distribution was also relevant to the membrane recruitment of β-Catenin, thus driving its exclusion from the nucleus and its ensuing degradation \[[@B28-cancers-11-00088]\]. The cellular localization of β-Catenin has a great influence on its activity: it is readily degraded in the cytoplasm, activation by Wnt induces its translocation to the nucleus and its subsequent participation in the expression of genes related to cell growth, or it may bind to transmembrane cadherins to participate in their association with actin cytoskeleton and cell differentiation. Here, it was shown that SMS1 overexpression was associated with both altered distribution of N-Cadherin and β-Catenin exclusion from the nucleus, all them molecular events involved in 2OHOA's mechanism of action, that causes glioma cell cycle arrest, differentiation and death \[[@B3-cancers-11-00088]\]. Overexpression of SMS1 but not SMS2 significantly activated macroautophagy via LC3 cleavage, which may explain the opposing roles of SMS1 and SMS2 in GBM. While autophagy is a cellular recycling mechanism, macroautophagy is a pivotal means to induce cancer cell death \[[@B30-cancers-11-00088]\]. Most importantly, activation of autophagy can induce either cancer cell chemoresistance (when initiated via EGFR signaling for example) or autophagic cell death (e.g., when initiated by β-catenin, AKT or Ras signaling), depending on the tumor type and treatment. In fact, the activation of apoptosis and autophagy following knockdown of β-catenin was recently reported in head and neck squamous cancer cells \[[@B39-cancers-11-00088],[@B40-cancers-11-00088]\]. In our in vitro GBM cell model, the mechanism of action by 2OHOA was associated with mislocalization of K-Ras \[[@B3-cancers-11-00088]\]. Consistently with these results, similar K-Ras mislocalization was seen when sphingomyelinase was inhibited \[[@B41-cancers-11-00088]\], which also increases SM levels. In addition, K-Ras mislocalization is also related to phosphatidylserine decrease in the plasma membrane \[[@B42-cancers-11-00088]\] and 2OHOA produces this effect on U118 cells \[[@B19-cancers-11-00088]\]. Here, it was shown that both 2OHOA treatment and overexpression of SMS1 dampened the activity of the β-catenin pathway, evident through the decrease in AXIN2 mRNA and compensatory increase by SMS1 silencing. Hence, 2OHOA-induced macroautophagy \[[@B32-cancers-11-00088],[@B33-cancers-11-00088]\] appears to be mediated by β-catenin. In general, GBM cells in which the expression of SMS1 increased showed lower proliferative capacity (accompanied by reduced levels of DHFR), increased differentiation and induction of autophagic cell death. This effect was replicated or enhanced by 2OHOA, which in part explains its molecular mechanism of action. Finally, we showed that only U118 cells expressing SMS1 underwent increased SM synthesis upon 2OHOA treatment. Our results demonstrate a direct modulation of SMS1 by 2OHOA, most likely due to direct interactions between these two molecular entities. The rapid response to treatment \[[@B19-cancers-11-00088]\] and the ability of the SAM domain present only on SMS1 to bind to lipids \[[@B7-cancers-11-00088]\] further support the hypothesis of a direct 2OHOA-SMS1 interaction. Thus, SMS1 activation induces glioma cell cycle arrest, differentiation and death. Because the main structural and functional difference between SMS1 and SMS2 is the SAM domain in SMS1, we hypothesize that this region could be involved in the divergent effects of 2OHOA on these isozymes. In addition to the SAM domain, the different cellular localization of SMS1 and SMS2, could explain their divergent effects on cell proliferation and glioma patient prognosis, as well as their differential response to 2OHOA treatment. Notably, SMS1 silencing caused SMS2 mRNA increases, which might be associated with some of the effects observed after siSMS1 treatment. This situation reflects the importance of the balance between both isoforms and replicates the worst scenario for patients (Low-SMS1 in combination with High-SMS2). Nevertheless, SMS1 increase due to SMS2 silencing did not significantly increase the expression of the corresponding protein. This behavior has previously been reported for SMS1 \[[@B23-cancers-11-00088]\], reflecting the high post-transcriptional regulation these enzymes undergo. SMS1 is a complex gene with several splicing forms \[[@B43-cancers-11-00088]\] capable of generating circular and linear RNAs \[[@B44-cancers-11-00088],[@B45-cancers-11-00088]\]. Furthermore, it contains recursive exons and participates in a multi-step splicing process involved in the generation of a wide array of RNAs, some of them with unknown roles \[[@B46-cancers-11-00088]\]. Although neither SMS1 protein nor mRNA levels increased when SMS2 was silenced, its enzymatic activity was boosted, which suggests functional regulatory post-translational processes of this key enzyme. Alternatively, another enzymatically active SMS1 splicing isoform may be responsible for the increase in the enzymatic activity. The fact that only SMS1 silencing induced SMS2 mRNA increase, while SMS2 silencing did not affect SMS1 mRNA levels, reflected that most SM is synthesized by SMS1, so that its lack would induce a compensatory increase of SMS2 expression, whereas lack of SMS2 would not have a high impact in the cell. In favor of this, while SMS1 KO mice show male infertility and are born in a lower Mendelian ratio \[[@B36-cancers-11-00088],[@B37-cancers-11-00088]\], SMS2 KO mice are seemingly healthy and only display reduced SM levels by about 20% \[[@B47-cancers-11-00088]\]. The great complexity of the SMS1 gene and its strong post-transcriptional/translational regulation, suggest an important role in cells beyond its catalytic activity. Previous studies suggested that the effect of 2OHOA to increase SM levels was exerted on both SMS1 and SMS2 in the context of the state-of-the-art at that point in time \[[@B3-cancers-11-00088],[@B19-cancers-11-00088]\]. However, more recent discoveries showing that both enzymes have ceramide phosphoethanolamine synthase activity \[[@B47-cancers-11-00088]\] opened the possibility of a differential action of 2OHOA on these isozymes. The present work demonstrated that 2OHOA only regulated SMS1. SMS1 and SMS2 are important enzymes in glioma tumorigenesis, with value as prognostic biomarkers to patient survival and potential response to pharmacological treatment with 2OHOA. SMS1 and SMS2 appear to have opposite effects on GBM cell differentiation, proliferation and survival. Our results suggest that 2OHOA would be more efficacious in glioma/GBM (and possibly other cancer) patients with low SMS1 expression that augments upon treatment initiation. The present study elucidates unknown functions about membrane lipids in general and lipid enzymes in cancer cell proliferation and survival. It also sheds light on the mechanism of action of 2OHOA and, opens new avenues for the rational design of specific SMS2 inhibitors to treat GBM and potentially other cancers. 4. Materials and Methods {#sec4-cancers-11-00088} ======================== 4.1. Cancer Patient Data Collection and Processing {#sec4dot1-cancers-11-00088} -------------------------------------------------- The potential clinical value of SMS1 and SMS2 gene expression was assessed in different datasets to investigate their correlation with the overall survival of patients with GBM and other cancer types. The "REpository for Molecular BRAin Neoplasia DaTa" \[[@B17-cancers-11-00088]\] dataset was downloaded from NCI CA Integrator to evaluate the relationship between SMS1 and SMS2 expression and glioma patient survival. These data, which were used to produce Kaplan-Meier plots from the REMBRANDT dataset, were available at the Betastasis website ([www.betastasis.com](www.betastasis.com)). The REMBRANDT database provides gene expression and survival data from 524 glioma patients, mainly grade III and IV (including 329 GBM patients). The relationship between SMS1 gene expression and survival probability after diagnosis was also investigated in other cancer datasets using the PROGgeneV2 website tool <http://watson.compbio.iupui.edu/chirayu/proggene/database/?url=proggene>) \[[@B48-cancers-11-00088]\]. The databases used to study this relationship were: GSE4412 for patients with grade III and IV glioma; TCGA-GBM for glioblastoma; GSE12417 for cytogenetically normal acute myeloid leukemia (CN-AML); GSE62254 for gastric cancer; TCGA-KIRK for kidney renal clear cell carcinoma (KIRC); GSE30219 for lung cancer; GSE4475 for lymphoma; GSE21501 for pancreatic ductal adenocarcinoma (PDAC); TCGA-SARC for sarcoma; and TCGA-SKCM for skin cutaneous melanoma (SKCM). For each dataset, cancer patient samples were classified into groups of high and low SMS1 expression with respect to the median expression. The median survival and p-values from the log-rank tests were compared in Kaplan-Meier plots for low and high SMS expressing patient groups using the GraphPad Prism 6 or the PROGgeneV2 web tool, setting statistically significant difference between curves at *p* \< 0.05. Overall survival at 2 and 5 years after diagnosis was calculated as indicated above, and presented as the mean ± standard error, showing the 95% upper and lower Confidence Intervals (CI). The correlation between SMS1 and SMS2 expression and lifespan is represented as boxes and whiskers, showing the 2.5--97.5 percentile data. The results were analyzed using a non-parametric Mann-Whitney test and were considered significant at *p* \< 0.05 values (two tailed, unpaired data). 4.2. Cell Culture {#sec4dot2-cancers-11-00088} ----------------- GBM and other cancer cell lines were obtained from Apointech SI (Salamanca, Spain) and cultured in RPMI 1640 medium (Sigma-Aldrich, St Louis, MO, USA) supplemented with 10% Fetal Bovine Serum (FBS, Sigma-Aldrich), 100 units/mL penicillin and 100 μg/mL streptomycin (Labclinics, Barcelona, Spain). The cells were cultured in 2 mL of medium in 6-well plates for RNA extraction or in 4 mL of medium in P60 dishes for immunoblotting and incubated at 37 °C and in 5% CO~2~. 4.3. Cell Transfection and Protein Overexpression {#sec4dot3-cancers-11-00088} ------------------------------------------------- Stably transfected cells were generated using the Tet-On^®^ 3G Inducible Expression System (Clontech, Madrid, Spain), according to the manufacturer's instructions. G418 (200 ng/μL) and hygromycin (100 ng/μL) were used for maintenance of stably transfected cells. Doxycycline (1 µg/mL) was used to induce SMS1 or SMS2 overexpression. Lipofectamine 2000 (ThermoFisher, Barcelona, Spain) was used to achieve transient transfection following the manufacturer's instructions. 4.4. Gene Expression Silencing {#sec4dot4-cancers-11-00088} ------------------------------ SMS1 and SMS2 gene expression silencing was performed by transient transfection of commercial siRNAs from Dharmacon, which were introduced in cells using RNAiMax (Life Technologies, Barcelona, Spain) followed by 48-h incubations. For siSMS1 silencing, the sense and antisense sequences used were GCCCAACUGCGAAGAAUAAUU and UUAUUCUUCGCAGUUGGGCUU, respectively (target sequence: GCCCAACTGCGAAGAATAA). For siSMS2 silencing, the sense and antisense sequences used were ACCGUCAUGAUCACAGUUGUAUU and UACAACUGUGAUCAUGACGGUUU, respectively (target sequence: ACCGTCATGATCACAGTTGTA). A negative siRNA control (siNeg) from Life Technologies was used (unknown mixed sequences). 4.5. Antibody Generation {#sec4dot5-cancers-11-00088} ------------------------ Two newly generated tailor-made isoform-specific antibodies anti-SMS1 and anti-SMS2 were produced in rabbit (Proteogenix, Schiltigheim, France) using the following peptides, respectively: DIPTPDGSFSIKIKPNGMPNGY and PSDPTNTYARPAEPVEEENKNG. A complete characterization of these antibodies is showed in [Figure S1](#app1-cancers-11-00088){ref-type="app"}. 4.6. Cell Lysis, Electrophoresis, Immunoblotting and Protein Quantification {#sec4dot6-cancers-11-00088} --------------------------------------------------------------------------- Cells were washed with phosphate-buffered saline (PBS, pH 7.6) and scraped in 400 μL of protein extraction buffer (10 mM Tris-HCl, 2 mM EDTA, 1% SDS, 5 mM Protease Inhibitor Complete \[ROCHE\] and 1 mM Sodium Orthovanadate). Protein extract quantification and immunoblotting were performed as described in \[[@B32-cancers-11-00088]\] with minimum variations. For standard curves (basal protein) increasing amounts of a standard protein extract (5, 15, 30 and 45 µg) were loaded on the same gel along with the experimental samples. The membranes were incubated overnight with primary antibodies diluted in fresh blocking solution: rabbit anti-SMS1 1:5000 or rabbit anti-SMS2 1:5000 from the above-described sera (ProteoGenix, Schiltigheim, France); mouse anti-N-Cadherin, anti-β-Catenin and anti-DHFR diluted 1:1000 (*BD-*Bioscience, Madrid, Spain); rabbit anti-LC3 and mouse anti-eIF2α diluted 1:1000 (Cell Signalling, Leiden, The Netherlands); mouse anti-V5 tag was diluted 1:5000 (Thermo Scientific, Barcelona, Spain). Antibody binding was detected with IRDye (800 CW) conjugated anti-mouse or anti-rabbit IgG (Li-cor, Biosciences, Lincoln, NE, USA) in blocking solution, incubating the membranes for 1 h at room temperature (1:10,000). Finally, protein content was quantified by integrated photodensitometry after near infrared scanning at 700 nm (Odissey, Li-cor Biosciences, Lincoln, NE, USA). α-Tubulin (1:10,000, Sigma) was assessed as a loading control. Bars in figures are representative of the average quantification from several experiments, each one with various replicates. The values in every lane were normalized to tubulin content and normalized values in treated cells were referred to those of control (untreated and not overexpressing) cells. 4.7. Immunofluorescence and Confocal Microscopy {#sec4dot7-cancers-11-00088} ----------------------------------------------- Cells were seeded onto 25 mm circular coverslips in 24-well plates at a density of ca. 25 × 10^3^ cells/well in 1 mL of RPMI 1640 medium supplemented with 10% FBS. After incubation overnight, SMS1 or SMS2 cloned in pTRE3G (Clontech, Saint-Germain-en-Laye, France) was transiently expressed and overexpression was induced with doxycycline treatment for 48 h. 2OHOA treatment was maintained for 24 h. Cells were then fixed with 4% paraformaldehyde in phosphate buffer for 30 min at room temperature or with 100% methanol for 1 h at −20 °C, and they were subsequently washed with PBS. Cells were incubated overnight at 4 °C with antibodies against LC3 (Cell Signalling, Leiden, The Netherlands), N-Cadherin or β-Catenin (BD-Bioscience, then washed 5 times with TBS and then incubated for 1 h at room temperature with the appropriate Alexa Fluor^®^ 488- or Alexa Fluor^®^ 598-conjugated IgGs (Invitrogen, Carlsbad, CA, USA). Finally, cells were washed 5 times with PBS, incubated with DAPI (Sigma, Darmstadt, Germany) for 5 min at room temperature and then washed twice with PBS. Coverslides were then mounted in Vectashield (Hardse: VECTOR) medium and examined under a confocal fluorescence microscope (Leica Microsystem TCS SPE) using a 63× objective lens. 4.8. RNA Isolation {#sec4dot8-cancers-11-00088} ------------------ Cells were seeded in 6-cm^2^ plates with RPMI 1640 medium supplemented with 10% FBS, 100 U/mL penicillin and 100 µg/mL streptomycin (PAA Laboratories GmbH, Austria). After 24 h, cells were exposed to 200 μM 2OHOA for 4, 8, 24 or 48 h. After treatment, the culture medium was removed, and the cells were washed with PBS. RNA was isolated using TriPure Isolation Reagent (a monophasic solution of phenol and guanidine thiocyanate at pH 4, Roche) according to the manufacturer's instructions and it was quantified using a Nanodrop2000 Spectophotometer at 260 nm (Thermo Scientific). 4.9. Quantitative RT-PCR {#sec4dot9-cancers-11-00088} ------------------------ The isolated RNA was reverse-transcribed using the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Barcelona, Spain). The target cDNA was amplified by Real-Time quantitative polymerase chain reaction (RTqPCR) using SYBR^®^ Premix Ex Taq™ II (Tli RNaseH Plus, Clontech, Saint-Germain-en-Laye, France) and ROX Reference Dye (50×) in a Step One Plus Real Time PCR thermal cycler (Applied Biosystems). The primers used were: SMS1 Forward (Fw): TGACTCCAGTGCAACGTGACSMS1 Reverse (Rv): GTCCACACTCCTTCAGTCGCSMS2 Fw: TTAATCTGCTGGCTGCTGAGSMS2 Rv: ACCAATCTTCTGAACCCGTGAXIN2 Fw: GAGTGGACTTGTGCCGACTTCAAXIN2 Rv: GGTGGCGTTGCAAAGACATAGHPRT Fw: TGACCTTGATTTATTTTGCATACCHPRT Rv: CGAGCAAGACGTTCAGTCCT A standard curve of known cDNA concentrations was used to quantify the mRNA levels using duplicate serial dilutions starting from 50 ng of cDNA. A Step One Plus Real Time PCR thermal cycler (Applied Biosystem) was used and the relative mRNA content was calculated using StepOne Software v2.3, expressed in arbitrary units (a.u.) using standard curves with different amounts of cDNA. 4.10. Enzymatic Assay {#sec4dot10-cancers-11-00088} --------------------- After overexpression or silencing and treatment in the presence or absence of 2OHOA, 3 µM of NBD-Ceramide was added to cells in culture and incubated for 3 h. Then, cells were scraped in a proper volume of extraction buffer (10 mM Tris-HCl, 2 mM EDTA and Protease Inhibitor Complete, ROCHE). After homogenization with ultrasounds, 200 µg of total protein diluted in a final volume of 250 µL of extraction buffer were mixed with 1.25 mL of chloroform:methanol (2:1) to extract lipids. The organic phase was collected after centrifugation for 10 min at 1000× *g* and solvents were evaporated under N~2~ flux. 4.11. Thin Layer Chromatography (TLC) {#sec4dot11-cancers-11-00088} ------------------------------------- Lipids extracted as indicated above were dissolved in chloroform and loaded on 20 × 20 cm silica plates. Lipid species were resolved using a mobile phase of Chloroform:Methanol:Water:Acetic acid (60:50:1:4 v:v:v:v). Plates were scanned on a fluorescent Bio-Rad (FX) scanner. NBD-SM synthesis was quantified by integrated optical density analysis. 4.12. Data Analysis {#sec4dot12-cancers-11-00088} ------------------- For Kaplan-Meier plots, the log-rank test was used to define differences between high and low SMS expression populations. For in vitro studies, data were expressed as the mean ± SEM or the % relative to untreated controls from the number of independent experiments indicated, and the differences were analyzed using a Student's *t*-test or ANOVA. The correlations between IC~50~ values and protein levels were analyzed using a Pearson's correlation test. Differences were considered significant at *p* \< 0.05. 5. Conclusions {#sec5-cancers-11-00088} ============== The present results suggest that SMS1 and SMS2 had opposite effects on glioma patients' survival, glioma cell growth and response to 2OHOA treatment. SMSs signature could constitute a valuable prognostic biomarker, high SMS1 and low SMS2 indicating a better disease prognosis. Additionally, low basal SMS1 mRNA levels predict positive response to 2OHOA in GBM cell lines. Regarding 2OHOA's mechanism of action, we demonstrated that 2OHOA induction of SM synthesis is due exclusively to SMS1 activation, and not to SMS2 activation. We thank Guillem Ramis for his help the confocal image analysis and Daniel Wansley for style review. The following are available online at <http://www.mdpi.com/2072-6694/11/1/88/s1>, Figure S1: SMS1 and SMS2 protein sequences alignment, Figure S2: Characterization of newly generated SMS1 and SMS2 antibodies, Figure S3: Characterization of overexpressed SMS1 and SMS2 proteins, Figure S4: Human SF295 GBM cell viability, Figure S5: Cell viability after incubation with siRNA against SMS1 or SMS2, Figure S6: Representative immunoblots showing changes in SMS1 or SMS2 protein levels, Figure S7: DHFR protein levels in SF295 and U251 cells overexpressing SMS2, Figure S8: SMS1 and SMS2 gene expression in GBM patient samples, Table S1: IC~50~ values of different cancer cell lines after a 72 h exposure to 2OHOA, Table S2: Changes in SMS1-to-SMS2 protein ratio in response to 2OHOA treatment and IC~50~ for 2OHOA, Table S3: Basal mRNA levels (arbitrary units, a.u.) in several GBM cell lines and non-tumor (MRC5) cells. ###### Click here for additional data file. P.F.-G., C.A.R., V.L., P.V.E. and X.B. conceived the experiments. P.F.-G., R.R.-L., R.B.-G., C.A.R., V.L. and J.F.-D. carried out the experiments. P.F.-G., C.A.R. and V.L. analyzed data. P.F.-G., C.A.R., V.L., P.V.E. wrote the manuscript. All authors critically revised and approved the final version of the manuscript. All authors were accountable for all aspects of the work in ensuring that questions are related to the accuracy and integrity of any part of the work. This work was supported by the Spanish Ministerio de Economía y Competitividad (Grants RTC-2015-3542 and RTC-2015-4094) and co-financed by FEDER funds. The present study was also in part supported by the Marathon Foundation. P.F.-G. was supported by a Torres-Quevedo Research contract from the Spanish Ministerio de Economía y Competitividad (PTQ-12-05133), co-funded by the FSE. The European Commission also supported in part this work (H2020 Framework Programme Project CLINGLIO 755179). C.A.R. received an EU PIEF-GA-2013-627441 Grant, and she was supported by a Felip Bauçà Postdoctoral Research contract from the Govern de les Illes Balears and European Social Funds. R.R.-L. was supported by Grant BES-2014-067883 from the Spanish Ministerio de Economía y Competitividad. R.B.-G. was supported by Grant DI-14-06701 from the Spanish Ministerio de Economía y Competitividad. J.F.-D. was supported by a Research Contract from the Govern de les Illes Balears and EU Funds. V.L. was supported by a Research Contract from the Spanish Association against Cancer (AECC). The authors declare no conflict of interest. {#cancers-11-00088-f001} {#cancers-11-00088-f002} {#cancers-11-00088-f003} {#cancers-11-00088-f004} {#cancers-11-00088-f005} ![Effects of SMSs and 2OHOA on ER stress and Autophagy in U118 cells. Phospho-eIF2α \[eIF2α(P): S51\] levels in U118 cells, determined by immunoblotting, (**A**) after 72-h treatment with doxycycline (1 µg/mL) to induce SMS1 or SMS2 overexpression, 2OHOA (200 µM), or both, or (**B**) after treatment with siRNA targeting SMS1 or SMS2 for 48 h, with 2OHOA (200 µM) for 24 h, or both. Levels of LC3B II-to-LC3B I determined by immunoblotting in U118 cells (**C**) after 72-h treatment with doxycycline (1 µg/mL) to induce SMS1 or SMS2 overexpression, 2OHOA (200 µM), or both, or (**D**) after treatment with siRNA targeting SMS1 or SMS2 for 48 h, with 2OHOA (200 µM) for 24 h, or both. Untreated cells or cells not overexpressing SMSs were used as controls (100%), and the bars represent mean ± s.e.m values from 4--8 independent experiments; The differences were analyzed using a Student's *t*-test or ANOVA \*, *p* \< 0.05; \*\*, *p* \< 0.01; \*\*\*, *p* \< 0.001; \# *p* \< 0.05; between groups. SMS1 and SMS2 overexpression was measured using an anti-V5 antibody. (**E**) LC3B immunofluorescence (green) in U118 cells overexpressing SMS1, SMS2 or after 2OHOA treatment (200 µM) for 72 h. Nuclei were labeled with DAPI. Representative micrographies (single confocal plane) are shown. Scale bar, 5 µm. (**F**) NBD-SM synthesis after 3-h incubation with NBD-C6-Ceramide in U118 cells previously treated for 48 h with siRNAs (siNeg, siSMS1, siSMS2 or siSMS1+siSMS2) and in the presence or absence of 200 µM of 2OHOA (12 h).](cancers-11-00088-g006){#cancers-11-00088-f006} cancers-11-00088-t001_Table 1 ###### Patients at risk from Kaplan-Meier plots in [Figure 2](#cancers-11-00088-f002){ref-type="fig"}. Patients were censored when they were alive at the time of study; mo.: months; n: number of patients; CI: confidence interval. Cancer Type Data Set SMS1 Express n Total/Censored Median Survival (mo.) 2-Year 5-Year ------------- ----------- -------------- ------------------ ----------------------- -------- -------- ------ ------ ------- ------ ------- ------ Glioma GSE4412 High 42/18 23.87 22 49.8 36.7 67.5 N/A N/A N/A N/A Low 41/8 9.93 6 15.4 7.0 33.8 N/A N/A N/A N/A GMB TCGA-GMB High 271/60 14.83 58 27.9 22.6 34.6 13 9.9 6.3 15.6 Low 271/56 12.95 36 19.4 14.7 25.7 5 2.3 0.9 6.0 CN-AML GSE12417 High 82/30 11.18 26 37.5 28 50.1 N/A N/A N/A N/A Low 81/30 8.42 26 35.9 26.4 48.8 N/A N/A N/A N/A Gastric GSE62254 High 150/90 N/A 116 76.7 70.2 83.7 97 64.0 56.8 72.2 Low 149/58 38.99 91 60.4 53 68.8 59 40.9 33.7 49.6 KIRC TCGA-KIRC High 117/93 93.04 71 84.9 77.8 92.6 41 76.8 67.87 87.0 Low 117/74 56.35 66 80.9 73.3 89.4 24 48.4 37.8 62.0 Lung GSE30219 High 141/60 83.84 111 77.0 70.3 84.3 69 58.1 50.3 67.1 Low 141/37 28.60 77 53.8 46.2 62.7 49 37.8 30.52 46.8 Lymphoma GSE4475 High 79/46 115.00 44 66.0 56.0 77.9 23 53.4 42.5 67.3 Low 79/37 33.44 35 53.4 43.0 66.3 19 41.4 30.7 55.9 PDAC GSE21501 High 51/22 21.70 17 45.3 32.3 63.5 N/A N/A N/A N/A Low 51/14 14.79 11 22.0 12.4 38.9 N/A N/A N/A N/A Sarcoma TCGA-SACR High 117/86 81.01 59 77.7 69.2 87.2 16 57.5 45.2 73.1 Low 117/76 53.49 50 75.0 66.2 84.9 18 43.9 32.3 59.7 SKCM TCGA-SKMC High 82/46 167.87 67 88.3 81.4 95.8 41 69.0 58.9 80.8 Low 81/41 48.59 47 72.9 63 84.5 21 47.1 35.7 62.1 [^1]: These authors contributed equally to this work.

In the last decade, great efforts have been made towards seeking new sources of human cardiomyocytes for various applications, especially for drug cardiotoxicity screening and myocardial repair that require large numbers of cells. Among the candidates, human embryonic stem cells (hESCs) have attracted significant attention, because of their potential to proliferate indefinitely and to differentiate into beating cardiomyocytes (hESC-CMs) *in vitro*. Methods for cardiac differentiation have been evolving rapidly since the work of Kehat *et al*. in 2001, which was based on spontaneous formation of human embryoid bodies (hEBs)[@b1]. Greatly improved protocols have since been reported, involving the culturing of hEBs in tightly controlled manners (such as forced aggregation, in microwells, etc.), co-culture of hESC colonies with visceral-endoderm-like cells (END-2 line) or use of hESC monolayers[@b2]. These advances make hESC-CMs a possible candidate for cardiac regeneration[@b3] as well as an alternate experimental model for rarely available native human cardiomyocytes. However the interpretation of responses of hESC-CMs to physiological or pharmacological stimuli is still confounded by their immature and heterogeneous state. Attention has been given primarily to their electrophysiological function, using patch clamp, intracellular microelectrodes, and microelectrode arrays as measurement tools. However, currently available characterization is still far from complete (for review, see Blazeski *et al*.[@b4]). The action potential (AP) is central to cardiomyocyte function because it not only initiates but also regulates and coordinates tissue contraction. The morphology of APs reflects the net balance among ionic currents across the cell membrane, and is a key signal in excitation-contraction coupling. APs recorded from either single isolated hESC-CMs or cell clusters containing hESC-CMs have widely varying AP morphologies[@b4]. AP variability can be the result of different culture conditions and cell lines, but independent of that, is reflective of the subtype and maturity of these *in vitro* generated cardiomyocytes[@b5][@b6][@b7]. Among different laboratories, APs recorded from hESC-CMs have generally been classified as one of three subtypes: nodal-like, atrial-like or ventricular-like[@b8][@b9][@b10][@b11][@b12][@b13][@b14][@b15][@b16][@b17][@b18] corresponding to the major CM phenotypes in adult myocardium. However, the invasiveness and time-consuming nature of direct electrophysiological recordings substantially limit the sample sizes of these studies (ranging from 15--125 in the cited studies, with an average of 50 samples) making it unclear whether predominant phenotypes are still present in larger, more representative cell populations. Previously, we[@b19][@b20] and others[@b21][@b22][@b23] showed that optical mapping can be used to investigate the electrophysiology of confluent populations of hESC-CM. Combined with a high resolution imaging system, it is practical to study cells in large populations all at once. Following our previous observation that APs recorded from beating areas of hEBs (which are dissected out and which we will refer to as cardiac cell clusters) from the same differentiation batch had a broad variation in morphology across clusters[@b4], we obtained a large dataset of APs of hESC-CM populations within cardiac cell clusters in this study, and focused on characterizing the variability and identifying the presence of predominant phenotypes. We used well-established parameters such as spontaneous activity and AP duration (APD), as well as novel waveform-based analysis methods to characterize the variability among and within cardiac cell clusters. These measurements represent the first systematic analysis of the variability and presence of phenotypes within a large cell population. We anticipate that this approach can also be used to evaluate new strategies designed to reduce the phenotypic variation within hESC-CM populations and improve their quality for use in diagnostic and therapeutic applications and in drug screening. Results ======= Spontaneous and electrically stimulated activity of cardiac cell clusters ------------------------------------------------------------------------- We started to see spontaneously beating hEBs around day 10 of differentiation. The number of beating hEBs varied as differentiation proceeded and also varied among differentiation batches. The clusters used for this study were obtained from a single batch of differentiation where more than 90% of hEBs were beating by day 15 (day of mechanical dissection). Although similar numbers of undifferentiated hESCs were seeded for hEB formation (5000 cells/hEB), clear differences in size and shape of hEBs and their beating areas were observed ([Fig. 1A](#f1){ref-type="fig"}, left column). After mechanical dissection, all cardiac cell clusters (beating areas of hEBs) attached to the coverslip and recovered spontaneous beating within 5 days, prior to being optically mapped. From the 55 clusters obtained from the batch, spontaneous APs were recorded using optical mapping. Both continuous (35 clusters) and episodic (20 clusters) patterns of beating were observed, the latter being identified by the existence of at least 4 seconds of quiescence between APs during the recording. Among continuously beating clusters, beating rate was unstable in 6 clusters. Action potentials recorded from different clusters exhibited different spontaneous rates and had clearly different morphologies ([Fig. 1A](#f1){ref-type="fig"}, middle column). The average beating rate of stable, continuously beating clusters was 62 ± 21 bpm (mean ± SD), and their average APD~80~ (action potential duration at 80% repolarization) was 165 ± 49 ms (n = 29). Because action potential characteristics and their underlying ionic currents are known to be rate-dependent, we tested whether some of the variability in APD~80~ of the clusters could be secondary to the differences in the beat rates. Thus, electrical pacing was applied with a pair of platinum electrodes. In the examples shown, a fixed rate of 90 bpm was able to pace-capture, and reduced, but did not eliminate the differences among APs from different clusters ([Fig. 1A](#f1){ref-type="fig"}, right column). Overall, 14 of the 29 continuously beating clusters with stable beating rate could be entrained by electrical stimulation at 90 bpm. For these 14 clusters, individual APD~80~s measured during spontaneous beating did not correlate with the beat rate ([Fig. 1B](#f1){ref-type="fig"}, open circles), and different APD~80~s were observed even among clusters with similar beat rate. When 90 bpm pacing was applied, there was no significant change in the mean or SD of APD~80~ (147 ± 36 ms without pacing compared with 143 ± 33 ms with pacing, [Fig. 1B](#f1){ref-type="fig"}). In most cases, the increase in beating rate from spontaneous rate to 90 bpm produced a drop in APD~80~, consistent with the negative rate-dependence observed in natural human cardiomyocytes/tissue, although it was the opposite case for 4 of the 14 clusters where APD~80~ increased instead. The large variability in APD~80~ at 90 bpm and the distinctly different rate-dependences among clusters indicate that both rate-independent and rate-dependent effects contribute to the variable AP durations observed at spontaneous beat rates. Variability of APs within cardiac cell clusters ----------------------------------------------- We first studied rate-independent variability of APs within cardiac cell clusters by pacing the clusters at a common rate of 90 bpm, and then recording the spatial distribution of action potentials via high-resolution optical mapping. A representative example is shown in [Fig. 2A](#f2){ref-type="fig"}, where APD~80~ of 981 recordings obtained from a single cluster are shown as a histogram. The average APD~80~ was 119 ± 19 ms, and two prominent distribution peaks were observed in the histogram, suggesting the existence of two subpopulations within the cluster. The advantage of optical mapping is its ability to discriminate whether the two subpopulations are interspersed or whether they cluster together spatially. In the spatial map of this cluster ([Fig. 2B](#f2){ref-type="fig"}), two distinct regions having either long or short APD~80~ were apparent, indicating that on a local scale of the order of a hundred microns, cells have similar AP duration, but on a larger scale of the order of a mm, different AP durations can emerge. We then compared AP variability within each cardiac cell cluster to that among the entire cluster population. Fixed 90 bpm recordings were obtained from 32 cardiac cell clusters with an average of 722 ± 252 recording sites in each cluster. The average and variation in AP parameters are summarized in [Table 1](#t1){ref-type="table"}. Even though some of these clusters had spontaneous activities (including the 14 continuously beating clusters analyzed in [Fig. 1](#f1){ref-type="fig"}) while others were quiescent, we did not observe systematic differences in APs when they were paced at 90 bmp that can be correlated to their spontaneous activities. Single AP parameter distributions for individual clusters were clearly different ([Fig. 2C--F](#f2){ref-type="fig"}). In some clusters (e.g, clusters 4, 24, 25, 28), the distributions of all AP parameters were relatively narrow, indicating a relatively homogeneous population of cells within the cluster, whereas in other clusters (e.g. clusters 8, 21, 27, 30) AP parameters were distributed over a wide range and had multiple peaks, revealing multiple sub-populations or a spectrum of cells. In other clusters (e.g. cluster 7), while distributions for some AP parameters (APD~80~, APD~30~ and fractional repolarization) were relatively narrow, much wider distributions were found in triangulation. Thus, AP variability presents differently for different AP parameters. Therefore differences in subpopulation identification may be expected when a single AP parameter (such as APD~80~) is used compared to when multiple AP parameters, or the entire AP waveform is used. To further investigate the potential subpopulations among these cardiac cell clusters, we applied and compared different grouping strategies. Starting with APD~80~, we fitted two Gaussian distributions to the overall distribution of the entire population, and used the crossover value as a threshold that divided APs into two groups, namely groups of long or short AP, respectively ([Fig. 3A](#f3){ref-type="fig"}). The average action potentials from the two groups are shown in [Fig. 3D](#f3){ref-type="fig"}. We then added another action potential parameter, APD~30~ to our analysis. Using principal component (PC) analysis, we identified the linear combination of APD~80~ and APD~30~ that accounted for most of the variability in the population, and its distribution is shown in [Fig. 3B](#f3){ref-type="fig"}. By fitting two Gaussian distributions and using the crossover value, we again divided APs into two groups, and the average AP from each group is shown in [Fig. 3E](#f3){ref-type="fig"}. In a third approach, we used an automated unsupervised algorithm to provide an unbiased grouping of APs based on their wave shapes across the entire cluster population. [Figure 3C](#f3){ref-type="fig"} shows the grouping results displayed in terms of APD~80~ and color-coded by the group assignment. Only a small fraction of the action potentials (in blue) were labeled as different from the bulk of the population, leading to a very different result than those using AP parameters alone ([Fig. 3A,B](#f3){ref-type="fig"}). The average APs from the two groups identified by waveform analysis are shown in [Fig. 3F](#f3){ref-type="fig"}. Compared with the average APs using AP parameter analysis ([Fig. 3D,E](#f3){ref-type="fig"}), these had greater differences in morphology ([Fig. 3F](#f3){ref-type="fig"}), with one group having a shorter AP with more triangular profile, and the other group having a longer AP and a longer plateau phase. Differences among the three grouping methods can be visualized in the joint distribution of measured APD~30~ and APD~80~ ([Fig. 3G](#f3){ref-type="fig"}). Here, dashed lines have been drawn corresponding to decision boundaries between the two groups determined by the analysis of: (1) APD~80~ only, (2) APD~80~ + APD~30~ PC, and (3) entire waveform. We found that while the APD~80~-only method always grouped cells with same APD~80~ values together, the variability in their APD~30~ (as indicated by the vertical spread in the joint distribution) resulted in APs groups that could include very different AP shapes. The inclusion of APD~30~ introduced a tilt to the threshold line (compare lines 1 and 2), indicating the contribution of both parameters to the subtype identification. To quantify the distinction between groups, we used the Davies-Bouldin Index (DBI, [Fig. 3H](#f3){ref-type="fig"}), which compares within-group tightness against between-group separations, such that lower DBI indicates better clustering (tighter groups and/or larger separations). DBI was lower for groups identified by waveform analysis (0.75) than when groups were identified by APD~80~-threshold (0.93) or PCA (0.93) ([Fig. 3H](#f3){ref-type="fig"}). Differences in the outcomes of the parameter and waveform analyses are especially apparent in comparing APs within individual clusters (e.g, clusters 7 and 27 in [Fig. 4](#f4){ref-type="fig"}). APD~80~ maps ([Fig. 4A](#f4){ref-type="fig"}) showed considerable variability between the two clusters, reflecting differences among the AP waveforms ([Fig. 4B](#f4){ref-type="fig"}). Among the 4 APs shown, one had short duration with rapid repolarization (a), one had a short duration with a slow repolarization tail at the end of the AP (b), one was very triangular (c) and the last had a long duration with rapid repolarization (d). Different groupings were obtained with the single parameter analysis (here, being APD~80~) compared with the waveform analysis. APD~80~ analysis grouped APs b and c with AP d, whereas waveform analysis grouped these APs with AP a. The limitations of single parameter analysis are also illustrated for APs having the same APD~80~ ([Fig. 4C](#f4){ref-type="fig"}) but drastically different morphology. Analysis in terms of other parameters (such as APD~30~ in [Fig. 4D](#f4){ref-type="fig"}) also encounter morphological differences. Only when both APD~80~ and APD~30~ were used together, were the morphological differences greatly reduced ([Fig. 4E](#f4){ref-type="fig"}). Waveform analysis was successful in separating these APs into different groups. Furthermore, the slower repolarization and more triangulated shapes in APs b-c could be captured by analyzing all of the APs as 3 groups, whereupon APs b-c were identified as belonging to the third group. We then expanded our analysis of spatial heterogeneity to two other metrics of AP shape: triangulation and fractional repolarization. Five representative clusters are shown in [Fig. 5](#f5){ref-type="fig"} (one cluster per column). In some cases (cluster 4), APD~30~, APD~80~, fractional repolarization and triangulation were fairly uniform across the cluster, indicating a single population of cells. In other cases (clusters 9, 27) the 4 AP parameters had similar patterns showing two distinct regions (e.g, on the left and right sides of cluster 9, and lower left and upper right sides of cluster 27) indicative of two subpopulations, although for cluster 27 the regions were not entirely concordant across parameters. In still other cases (cluster 7, 24) the patterns for the AP parameters were dissimilar. By using the automated algorithm with entire AP waveform, cells among the clusters were separated into 2 groups, with spatial patterns as shown ([Fig. 5](#f5){ref-type="fig"}, second to last row). The 2-group analysis revealed that 14 of 32 clusters contained cells from different groups. This coexistence of cells from different groups is more notable with automated 3-group analysis ([Fig. 5](#f5){ref-type="fig"}, last row), where 17 clusters had cells from all 3 groups, 12 clusters had cells from 2 different groups, and only 3 clusters had cells from a single (and different) group. Furthermore, cells belonging to different groups were not intermingled and tended to occupy separate areas within a cluster ([Fig. 5](#f5){ref-type="fig"}, last two rows) even though spatial information was not incorporated into the grouping algorithm. Separate groups could be identified even in clusters where the patterns of APD parameters were discordant (e.g. cluster 24). Additionally, our analysis resulted in the identification of individual AP phenotypes that were common to multiple clusters (e.g., group 3 in clusters 9, 24, and 27). Note, however, that the increased number of groups was accompanied by a higher DBI value (0.91 for 3 groups vs. 0.75 for 2 groups), suggesting that 2 groups better describe this particular population of action potentials. For visualization and discussion purposes, we plotted all APs as scatter plots of joint distributions of pairs of AP parameters, along with their group labels by waveform analysis in colors ([Fig. 6](#f6){ref-type="fig"}). Although paced at the same rate, the entire set of APs encompassed a large area in all 2D spaces of AP parameter pairs, indicating a high degree of variability. When APs were divided into two groups, they occupied relatively distinct regions in all 2D spaces. Linear boundaries were determined by support vector machines (SVMs) within each 2D space, and shown as dashed lines. The degree of overlap of the two groups was defined as the total percentage of APs of each group located outside the boundary of their respective groups, and ranged from 0.9% to 8.1% across all of the 2D spaces, indicating the influence of features of AP shape not captured by just a pair of AP parameters. In general, the overlap in the 2D spaces with AP duration parameters ([Fig. 6A](#f6){ref-type="fig"}) was smaller than that in 2D spaces with triangulation and fractional repolarization parameters ([Fig. 6F](#f6){ref-type="fig"}), suggesting that AP duration parameters have a stronger contribution in grouping than triangulation or fractional repolarization. Rate dependency --------------- Depending on the cluster, electrical pacing could stimulate APs up to 330 bpm while maintaining 1:1 capture, allowing us to study the rate-dependence of paced APs ([Fig. 1B](#f1){ref-type="fig"}). We hypothesized that rate-dependency could be an additional means by which different sub-populations (phenotypes) of cells could be discriminated. In a cluster having two regions with different AP durations (cluster 26), electrical pacing was applied from 90 to 180 bpm in 30 bpm increments. While all recording sites (n = 585) in this cluster had rate-dependent shortening of APD~80~, the rate dependencies of fractional repolarization varied greatly, being positive (increased fractional repolarization with increased rate) in some regions (n = 180) and negative in others (n = 405). Two representative recordings from this cluster and their corresponding APD~80~ and fractional repolarization rate-dependencies are shown in [Fig. 7](#f7){ref-type="fig"}. The differences in rate-dependencies suggest that the use of multiple pacing rates may further assist in identifying phenotypic differences. Discussion ========== Previous studies have shown that APs generated by hESC-CMs are heterogeneous in their electrophysiological properties and not fully like adult phenotypes[@b4][@b6][@b24]. A fundamental question is: down to what structural level does this phenotypic heterogeneity extend? Prior studies using microelectrode impalements of hEBs showed that APs differ substantially among different hEBs but were mostly similar within individual hEBs[@b9] or within cell clusters[@b12]. Our results show that APs vary over a wide range across hEBs as well as over a smaller, but still substantial range within hEBs. Although numerous research groups have reported percentages of nodal-like, atrial-like and ventricular-like cells at a stage of differentiation similar to ours (20--24 days)[@b10][@b14][@b15][@b25], our results demonstrate that different phenotypes are difficult to distinguish in at least two ways: 1) individual parameters are insufficient to divide populations of APs into different groups, since APs with similar parameter values (such as APD~80~) could have very different morphology indicative of different phenotypes, and 2) the wide range and continuous variability of APs make it extremely difficult to subjectively determine cut-offs between groups. By using an automated grouping algorithm that analyzed entire AP waveform shapes, we able to objectively sort the data into 2 or 3 maximally separated groups. Cardiac cell clusters derived from hEBs are self-assembled, three-dimensional aggregates of cells. In contrast to flat cultures on two-dimensional substrates, the three-dimensional organization of the cells provides a natural microenvironment for the cells, allowing them to form and remodel their surrounding extracellular matrix, to undergo paracrine and autocrine signaling, and to contract against a compliant mechanical load, while maintaining independence among clusters. To reduce variability due to culture conditions, we used a highly efficient differentiation protocol[@b20][@b26] to obtain large numbers of beating hEBs sufficient for variability studies from the same differentiation batch. Beating cell clusters were mapped within a narrow time window, 21--24 days after initiation of differentiation. Fixed-rate pacing was used so that APs could be compared at the same beat rate. As a caveat, there still remained several uncontrolled experimental variables. Although similar numbers of hESCs were seeded to form hEBs, the ensuing hEBs still had different sizes, shapes and beat rates, which can influence the differentiation process. In addition, the mechanically dissected cardiac cell clusters may only represent a subset of all cardiomyocytes. Also, the uncontrolled presence of nonmyocytes in the cell aggregates can influence the maturation of the embedded hESC-CMs[@b27], although methods to eliminate nonmyocytes have been reported[@b18][@b28]. Our differentiation method also did not target a specific cardiomyocyte phenotype, although such strategies have been reported for the ventricular phenotype, where relatively uniform, ventricular-like action potentials have been observed[@b29]. Sharp microelectrodes and multiple impalements have been used to study the variability in electrophysiology of hEBs[@b9][@b12], but in practice are limited to around 10 or less per hEB. In contrast, high-resolution optical mapping provides hundreds of recordings from precise locations within single cardiac clusters without mechanical disruption or cellular injury due to impalements. However, there are limits on true spatial resolution, owing to diffuse light scattering of the fluorescent signal and volume integration of signals on the recording path[@b30][@b31]. Electrical coupling among neighboring cells can also electrically average APs within close proximity. The boxcar filter applied in our data analysis, which was necessary to improve signal quality, also limits the spatial resolution in our method. Nonetheless, significant heterogeneity in AP parameters was always observed, indicative of a genuine variation in electrophysiological phenotypes. This finding is not unexpected in hEBs during early stage differentiation, as they are thought to mimic human embryo development[@b9][@b32] and have the potential to develop into all major phenotypes of adult cardiomyocytes. Another limitation of optical measurements is that they measure only relative changes in transmembrane voltage and not absolute values (unless advanced methods such as dual wavelength measurements and microelectrode calibration are used)[@b33][@b34]. Nevertheless, they have fast responses that can faithfully track the transmembrane voltage and preserve AP shape[@b35][@b36]. One possible confounding effect in our experiments is that electrotonic influences among cells can produce gradients in action potential duration even in a homogenous cell population. This is particularly evident when action potentials propagate away from a stimulus site, in which case those nearer the site have longer durations than those farther away[@b37]. These effects are also modulated by tissue dimension, tissue boundaries and structural heterogeneity[@b38][@b39][@b40]. We do not believe that our results are greatly influenced by such effects. The spatial dimensions of our EBs are much smaller than the hearts in which such effects have been observed, restricting the amount of repolarization gradients that can develop. The differences in APDs observed in previous studies of electrotonic effects are of the order of 10--20 ms, much smaller than the differences we observed within and among clusters (\~100 ms). The activation times across our clusters were typically less than 15 ms, and APD shortening did not correlate with the longer activation times. The short activation times, and smooth activation sequences across clusters, also ruled out conduction delays secondary to structural heterogeneities that could produce APD gradients. In addition, variation in APD were observed across the entire cluster, without sharp gradients along the edges where the boundary effects might occur, and significant AP variability occurred among clusters, which is independent of electrotonic effects. Our high-resolution measurements of early differentiated clusters suggest that a relatively broad distribution of APs is present within the clusters ([Fig. 2](#f2){ref-type="fig"}), although not as broad as across the entire set of 32 clusters that was studied. In many, but not all clusters, a histogram of APD~80~ that was either narrow or else wide with multiple peaks was observed, suggestive of one or multiple predominant phenotypes. APD~30~, triangulation and fractional repolarization are other ways to quantify the plateau and repolarization phases of the AP, which in turn are manifestations of the underlying ionic currents which have differing temporal and voltage-dependent kinetics[@b41]. APD~30~ is a rough measure of plateau duration, while APD~80~ is a measure of the combined plateau and repolarization durations. Triangulation[@b42] is roughly the duration of phase 3 repolarization, and is considered to be proarrhythmic when prolonged. Fractional repolarization is the fraction of the total APD involved with late rapid repolarization. This parameter is also akin to the comparison between plateau (duration between APD~30~ and APD~40~) and repolarization (duration between APD~70~ and APD~80~) phases defined by others to distinguish between ventricular-like cells and atrial-like cells[@b43]. Our results showed that the spatial patterns of groupings based on each of the four parameters were usually but not always concordant with one another ([Fig. 5](#f5){ref-type="fig"}), raising the question of which parameter(s) is most important. Ideally, molecular markers would be used to identify subtypes of hESC-CMs, but at present, these are limited primarily to MLC2v or IRX-4 for the ventricular subtype[@b29][@b44][@b45] and cannot be used on live cells. Classification by AP morphology is still the most widely adopted method by which cardiomyocyte phenotypes have been assessed by microelectrode or optical recordings[@b14][@b15][@b16][@b17][@b18][@b22][@b23], Major subtypes of hESC-CMs (nodal-like, atrial-like, ventricular-like) have most often been identified subjectively by the similarity of their AP shapes to those of the adult phenotypes in terms of AP parameters. Only a few studies specified quantitative criteria for classification, including APD~90~[@b13], combination of APD~90~ and beat rate[@b12], ratio of APDs[@b43][@b46], and a combination of AP amplitude, upstroke velocity, APD~50~ and APD~90~[@b14][@b25]. In this study, we showed that grouping by histogram analysis of single parameters does not assure consistent AP morphology; action potentials having the same APD~80~ or APD~30~ can have differing AP shapes reflecting different phenotypes ([Fig. 4C,D](#f4){ref-type="fig"}). Increased numbers of parameters result in improved grouping, with the ultimate representation being the entire waveshape. Analysis of the entire data set of APs (recorded across 32 hEBs) by an automated machine learning algorithm[@b47] divided the APs into different groups based on their shape and similarity to one another. This method also produced the most distinctive differences in AP waveforms between groups ([Fig. 3F](#f3){ref-type="fig"}). Within individual hEBs the groups occupied separate spatial regions rather than being intermixed, even though the automated algorithm did not incorporate spatial information. This finding suggests that physical location or cell-cell interactions can influence phenotypic specification during hESC-CM differentiation, which is manifested in the patterns of individual parameters. Our use of the entire set of APs across all clusters also enabled cells that belong to same group to be identified in multiple clusters ([Fig. 5](#f5){ref-type="fig"}), which would not be possible by separate analyses of individual clusters. It should be noted that although APs were divided into 3 groups, those groups do not necessarily correspond to nodal-, atrial- and ventricular-like cells, but perhaps, to precursor or intermediate stage cells. Further studies are warranted to develop more advanced methods for AP shape analysis and ultimately, phenotypic classification. A further improvement for the discrimination of different phenotypes, which has not been employed so far by other laboratories, is the use of different pacing rates, which takes advantage of the different rate-dependent kinetic properties of the ionic currents underlying the AP. As we showed in one example, with increasing rate APD~80~ decreases all across the cluster, whereas fractional repolarization may change in opposite directions in different regions of the cluster ([Fig. 7](#f7){ref-type="fig"}). In terms of future applications, we have presented a strategy to obtain large datasets of APs via optical mapping and to analyze them for their electrophysiological variability and grouping. We envision that this approach will have broad applications as a tool in the study of stem cell-derived cardiomyocytes in the context of differentiation[@b2], maturation[@b24], production[@b48], drug screening[@b49], disease modeling[@b50] and myocardial regeneration[@b51]. Some examples of questions that could be addressed include: does the monolayer method of differentiation, where paracrine and juxtacrine interactions among the cells is less than in the hEB, result in a more diverse or more homogeneous distribution of cell phenotypes? When do and how many predominant phenotypes emerge during time in culture? Do certain cell lines, culture conditions or bioengineering approaches to improve maturation favor the rapid emergence of a common phenotype? How varied are responses of the cell population to drugs that affect repolarization? What is the electrophysiological consistency within and among large batches of manufactured cells? In conclusion, we observed substantial variability in AP electrophysiology of hESC-CMs contained among and even within individual cardiac cell clusters, despite measures designed to reduce differences in experimental test conditions. The variability could not be correlated with beat rate, and therefore reflects genuine phenotypic differences. Our results suggest that in early stage differentiated cell clusters, a wide range of electrophysiological phenotypes is present, and in some clusters, one or two phenotypes can dominant. Finally, the use of high resolution optical mapping and AP analysis may be useful in the development of hESC-CMs for drug cardiotoxicity screening, disease modeling and myocardial repair. Methods ======= Cardiac differentiation of human embryonic stem cells ----------------------------------------------------- The H9 line of human embryonic stem cells was used for this study. Spontaneous beating hEBs were obtained using a forced aggregation method as previously described[@b20][@b26]. Beating areas of hEBs from the same differentiation batch were mechanically dissected on day 15 of differentiation, and then transferred to plastic coverslips coated with 0.1% gelatin (Sigma-Aldrich, St. Louis, MO). These cardiac cell clusters were maintained in culture for at least 5 days to attach and recover spontaneous beating. Optical mapping --------------- On day 21--24 of differentiation, cardiac cell clusters were stained with 10 μM voltage-sensitive dye di-4-ANEPPS (Invitrogen, Grand Island, NY) for 10 min, and then transferred to a custom-made mapping chamber with Tyrode's solution containing 50 μM blebbistatin (Sigma-Aldrich, St. Louis, MO) to inhibit motion. Temperature was controlled at 37 °C throughout all experiments. For externally paced recordings, a pair of platinum field electrodes was used to deliver a 5--10 ms rectangular electrical stimulus. Pacing rate started at 60 beats per minute (bpm), and was incremented in steps of 30 bpm until the cluster failed to maintain 1:1 capture by the pacing stimulus. Optical action potentials were recorded using a MiCAM Ultima-L CMOS camera (SciMedia, Costa Mesa, CA) with 100 × 100 pixels (16 μm/pixel) at 500 frames per second (fps). Sixteen to 32 seconds of recordings were taken for each cluster at each pacing rate. Signal processing and data analysis ----------------------------------- To improve signal quality, optical recordings were convolved with a 5 × 5 boxcar filter. The border of each cluster was then traced manually to identify the recording pixels contained within that cluster. For paced activities, multiple APs from each recording site were temporally aligned by their activation times (calculated as the time of fastest AP upstroke), after which an average of the APs was calculated for that recording site. The averaged APs were quantified by 4 parameters: AP duration at 30% and 80% repolarization (APD~30~ and APD~80~, respectively), triangulation (time from APD~30~ to APD~80~) and fractional repolarization time (time from APD~30~ to APD~80~ divided by APD~80~). For triangulation, we used APD~80~, rather than APD~90~ as originally defined[@b42], because in some cells the AP baseline had a slow repolarizing tail following the repolarization phase. These four parameters were used to quantify AP variability within and among clusters. Principal component analysis[@b52] was used to determine the linear combination of AP parameters that accounted for the most variability within the dataset by orthogonal transformation. Levene's test of equal variances was used to compare the variability in APD~80~ at spontaneous or paced conditions. Automated grouping ------------------ Action potentials were recorded from clusters paced at 90 bpm and separated into groups based on their action potential morphology, using a recently published automated algorithm[@b47]. This grouping algorithm had been developed using data from 9 of the 32 clusters presented here. Briefly, APs were divided into 2 or 3 groups based on an unsupervised, spectral clustering technique that amasses APs with similar shapes together and separates APs having different shapes. The similarity between pairs of APs is calculated mathematically as a weighted function based on the cumulative squared errors between the two temporal waveforms. Davies-Bouldin index (DBI)[@b53] was also calculated using the cumulative squared errors, to compare cluster separation using different numbers of groups. Linear decision boundaries in pairs of AP parameter spaces were derived by support vector machines (SVMs) using labels determined by the spectral clustering. Additional Information ====================== **How to cite this article**: Zhu, R. *et al.* Variability of Action Potentials Within and Among Cardiac Cell Clusters Derived from Human Embryonic Stem Cells. *Sci. Rep.* **6**, 18544; doi: 10.1038/srep18544 (2016). This work was supported by NIH grants S10 RR025544 (LT), R21 HL108210 (LT), U01HL099775 (ETZ), and Maryland Stem Cell Research Fund grants 2011-MSCRF-II-0008 (ETZ) and 2013-MSCRF-II-0045 (LT). We thank Dr. Nathan Crone for allowing us to use the computational resources in his lab. **Author Contributions** R.Z. designed and performed the electrophysiology experiments, analyzed the data. L.T. supervised the study. R.Z. and L.T. wrote the manuscript. M.A.M. performed hESC culture and differentiation. E.T.Z. supervised hESC culture and differentiation. All authors reviewed the manuscript. {#f1} {#f2} {#f3} {#f4} {#f5} {#f6} {#f7} ###### Variability of AP parameters within individual and among all 32 clusters paced at 90 bpm.   APD~30~ APD~80~ Triangulation Fractional repolarization ----------------- -------------- -------------- --------------- --------------------------- Individual Mean 46 to 188 ms 85 to 257 ms 45 to 117 ms 0.27 to 0.51 Overall Mean 85 ms 137 ms 70 ms 0.41 Individual SD 4.3 to 30 ms 6.5 to 34 ms 4.8 to 58 ms 0.03 to 0.11 Overall SD 33 ms 41 ms 24 ms 0.09

\section{Introduction and Main Results} Persistent Homology is a circle of ideas \cite{frosini1990distance,frosini1992measuring,delfinado1995incremental,robins1999towards,edelsbrunner2000topological,zomorodian2005computing} related to studying the homology of diagrams of simplicial complexes or topological spaces. Often these diagrams are parametrized by a scale parameter which has some geometric meaning. One fundamental example is that given by an increasing sequence of subspaces of a given topological space $X$: $\emptyset=X_0\subset X_1\subset \cdots\subset X_n=X$. In this case, upon passing to homology (with coefficients in a field), one obtains a similar diagram of vector spaces and linear maps $V_0\rightarrow V_1\rightarrow \cdots \rightarrow V_n$. These diagrams are referred to as \emph{persistence modules} and, under mild tameness assumptions, their structure up to isomorphism can be summarized by a multiset of pairs $(i,j)$ with $i\leq j$. The intuition is that these multisets subsume the lifetime of homological features as these are born and are eventually annihilated. The poset underlying the diagram of topological spaces $\emptyset=X_0\subset X_1\subset \cdots \subset X_n=X$ from above is simply the poset on $n$ points (generated by) $\bullet \rightarrow \bullet \rightarrow \cdots \rightarrow \bullet$. This setting was generalized by Carlsson and de Silva \cite{zigzag} to allow for any diagram of topological spaces (or simplicial complexes) whose underlying poset is of the form $\bullet \leftrightarrow \bullet \leftrightarrow \cdots \leftrightarrow \bullet$, where at each occurrence of $\leftrightarrow$ exactly one choice for the direction of the arrow is made; the finite sequence of all such choices is called the \emph{type} of $\mathbb{V}$. This generalized setting is called \emph{zigzag persistence}. It provides a complete algebraic invariant for sequences $\mathbb{V} = (V_i,p_i)$ of vector spaces and linear maps of the form \begin{equation*} \begin{tikzcd}[row sep=2em,column sep=3.5em] V_1\arrow[r,"p_1"]& V_2 \arrow[l]\arrow[r,"p_2"]&\cdots\arrow[l]\arrow[r,"p_{n-2}"]& V_{n-1}\arrow[l]\arrow[r,"p_{n-1}"]& V_n\arrow[l]. \end{tikzcd} \end{equation*} As it was noted in \cite{zigzag} these complete algebraic invariants that zigzag persistence associates to $\mathbb{V}$ also take the form of persistence diagrams, but these can now be enriched with the type of the zigzag persistence module from which they arise. Zigzag persistence has found applications in neuroscience \cite{babichev2017robust,chowdhury2017importance}, and in the analysis of dynamic data \cite{corcoran2016spatio,kim2017stable,chowdhury2017importance}. See \cite{tausz2011applications} for a general description of many possible applications of zigzag persistence. In practical applications one typically wishes to use the persistence diagram of a zigzag module to gain insights about the underlying data from which the zigzag module was extracted. With applications in mind, it is important to be able to guarantee stability of zigzag persistence. Informally, a process which takes data as input and provides some invariant as output is \emph{stable} if whenever the input data is perturbed slightly, the resulting invariant changes only slightly. Since data is usually acquired with some inherent noise, stability is a very desirable property. Depending on the application, a metric $\rho$ is defined on the space of input data and stability results take the form $d_b\leq C \rho$ for some constant $C>0$. A standard metric for measuring the closeness of two persistence diagrams is the bottleneck distance $d_b$ \cite{edelsbrunner2010computational}. In the context of standard persistence, it has been proven that persistence diagrams are stable in different degrees of generality \cite{classical_stability,cohen2006vines,chazal_cohen-steiner_glisse_guibas_oudot_2009,bauer2014induced,lesnick2015theory,cohen2010lipschitz,bubenik2014categorification,bubenik2015metrics}. An alternative approach is to define a metric \textit{at the algebraic level}, measuring the distance between persistence modules directly. Stability of this form is referred to as \textit{algebraic stability} and it is a notion of algebraic stability that we study in this paper. The algebraic stability of standard persistent homology was studied in \cite{chazal_cohen-steiner_glisse_guibas_oudot_2009} (see also \cite{chazal2016structure,lesnick2015theory,bauer2014induced}), whereas the algebraic stability of zigzag persistence was approached by Botnan and Lesnick through a method different from ours in \cite{algstabzz}. In essence, the distance between zigzag modules constructed by Botnan and Lesnick first suitably extrapolates two given zigzag modules into persistence modules over $\mathbb{R}^2$ and then computes an interleaving type distance between these extrapolated modules. They were able to prove that the bottleneck distance between the persistence diagrams of the original persistence modules is bounded above by a constant times the value of the distance between them. A recent refinement by Bjerkevik \cite{bjerkevik} has found the optimal constant for this inequality. We now describe the structure of our distance and state our main stability result. \subsection{Statement of the Main Result} We introduce here a family of pseudometrics $d_{\altmathcal{R}}^p$, parametrized by $p\in[1,\infty)$, on the space of zigzag modules of length $n$ and then show that in the special case $p = 1$, the inequality $d_b^1\leq d_{\altmathcal{R}}^1$ holds, where $d_b^p$ denotes the \textit{$\ell^p$-bottleneck distance}. The $\ell^p$-bottleneck distance arises from considering a definition analogous to the standard bottleneck distance \cite{edelsbrunner2010computational} with the provision that the ground metric between points is chosen to be the $\ell^p$ norm in $\mathbb{R}^2$ (see details in Section \ref{sec:d-bottleneck}). For a given $p\in[1,\infty)$, $d_{\altmathcal{R}}^p$ is called the \emph{$p$-reflection distance.} The idea behind the definition of the reflection distance is the following: we consider some collection of transformations of zigzag modules which we will model as a collection $\altmathcal{S}$ of endofunctors on the category $n\textup{-Mod}$ of zigzag modules of a fixed length $n$. For each $p\in[1,\infty)$, we associate a cost to each functor $\altmathcal{F}\in\altmathcal{S}$ by means of a cost function \[ C_p:\altmathcal{S}\to \mathbb{R}^+.\] We then define a function $d^p_\altmathcal{R}:n\textup{-Mod}\timesn\textup{-Mod}\to \mathbb{R}$ by setting \[ d^p_\altmathcal{R}(\mathbb{V},\mathbb{W}):= \min_{(\altmathcal{F}_1,\altmathcal{F}_2)}\big\{\max\{C_p(\altmathcal{F}_1),C_p(\altmathcal{F}_2)\} \ | \ \altmathcal{F}_1(\mathbb{V})\precsim \mathbb{W} \text{ and } \altmathcal{F}_2(\mathbb{W})\precsim \mathbb{V}\big\},\] where the minimum is taken over pairs $(\altmathcal{F}_1,\altmathcal{F}_2)\in \altmathcal{S}\times \altmathcal{S}$ of functors satisfying the conditions that \mbox{$\altmathcal{F}_1(\mathbb{V})\precsim \mathbb{W}$} and $\altmathcal{F}_2(\mathbb{W})\precsim \mathbb{V}$. Here, $\mathbb{V}_1\precsim\mathbb{V}_2$ if and only if $\mathbb{V}_1$ is equivalent to a summand of $\mathbb{V}_2$, where ``equivalent" refers to equivalence of zigzag modules which differ only in the direction of linear maps representing isomorphisms. If the cost function $C_p$ satisfies the subadditivity condition \[ C_p(\altmathcal{F}_2\circ \altmathcal{F}_2)\leq C_p(\altmathcal{F}_1) + C_p(\altmathcal{F}_2)\] for all functors $\altmathcal{F}_1, \altmathcal{F}_2\in \altmathcal{S}$ then $d^p_\altmathcal{R}$ turns out to be a pseudometric on $n\textup{-Mod}$. Of course, $d^p_\altmathcal{R}$ depends both on the collection of functors which we restrict ourselves to and the cost function $C_p$ used. The functors which we will restrict ourselves to in this paper are defined by replacing certain subdiagrams of a given zigzag module by a diagram formed from its limit or colimit. Such functors are closely related to the reflection functors of Bernstein, Gelfand, and Ponomarev \cite{bgp}, hence the name \textit{the reflection distance}. The cost function chosen simply counts the number of transformations needed to transform a pair of zigzag modules into each other, weighted by the parameter $p$. Our main result is then the following \begin{restatable}[Main Theorem]{theorem}{thmmain} \label{thm:main}For all zigzag modules $\mathbb{V},\mathbb{W} \in n\textup{-Mod}$ we have \[d_{b}^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})) \leq d_{\altmathcal{R}}^1(\mathbb{V},\mathbb{W}).\] \end{restatable} \noindent Here, $d_b^1$ denotes the \textit{$\ell^1$-bottleneck distance} as described above. \subsection{Organization of the paper} Section \ref{sec:zigzag-mods} recalls the main facts about zigzag modules that we will need in this paper; Section \ref{sec:decompositions} recalls elements regarding the decomposition of zigzag modules as direct sums of interval modules; Section \ref{sec:type-transformations} sets terminology which will be used in later sections to describe transformations between different types of zigzag modules; Section \ref{sec:ref-functors} describes reflection functors and their effect on interval modules; Section \ref{sec:ref-distance} provides the precise description of our reflection distance. In Section \ref{sec:d-bottleneck} we describe the $\ell^p$-bottleneck distance and give the proof of our main theorem, Theorem \ref{thm:main}; Section \ref{sec:discussion} provides an overview of some lines of related research which may be of interest. Finally, Appendices \ref{app:cats} and \ref{app:matchings} contain background material on Category Theory and Matchings. \section{Zigzag Modules} \label{sec:zigzag-mods} Fix a field $\mathbb{F}$. All vector spaces throughout will be finite dimensional over $\mathbb{F}$. A \textit{zigzag module} $\mathbb{V} = (V_i,p_i)$ is a finite sequence \begin{equation}\label{zzmod} \begin{tikzcd}[row sep=2em,column sep=3.5em] V_1\arrow[r,"p_1"]& V_2 \arrow[l]\arrow[r,"p_2"]&\cdots\arrow[l]\arrow[r,"p_{n-2}"]& V_{n-1}\arrow[l]\arrow[r,"p_{n-1}"]& V_n\arrow[l] \end{tikzcd} \end{equation} of vector spaces and linear transformations between them. An arrow $V_i \stackrel{p_i}{\longleftrightarrow} V_{i+1}$ represents either a forward linear map $V_i \stackrel{p_i}{\longrightarrow} V_{i+1}$ or a backward linear map $V_{i} \stackrel{p_i}{\longleftarrow} V_{i+1}$, but never both. The $p_i$ are referred to as \textit{structure maps}. Note that the $i^{\textup{th}}$ structure map $p_i$ either has domain $V_i$ or $V_{i+1}$ and codomain $V_{i+1}$ or $V_{i}$ respectively. We will use the notation $p_i:V_{i_1}\to V_{i_2}$ when the direction of $p_i$ has not been specified. In other words, $i_1,i_2\in \{i,i+1\}$ with $i_1\neq i_2$, $\textup{dom}(p_i) = V_{i_1}$, and $\textup{cod}(p_i) = V_{i_2}$. The \textit{length} of a zigzag module is the length of the sequence (\ref{zzmod}) above. We will denote the collection of all zigzag modules of length $n$ by $n\textup{-Mod}$. A finite sequence $\tau$ of the symbols $\rightarrow$ and $\leftarrow$, indicating the directions of the linear maps in (\ref{zzmod}) as read from left to right, is called the \textit{type} of the zigzag module. Formally, the type of a zigzag module of length $n$ is a sequence $\tau\in\{\rightarrow,\leftarrow\}^{n-1}$. We will use the notation $\altmathcal{T}_n := \{\rightarrow,\leftarrow\}^{n-1}$. \begin{remark} Sequences in $\altmathcal{T}_n$ have length $n-1$, not length $n$. This is so that zigzag modules in $n\textup{-Mod}$ have types in $\altmathcal{T}_n$. \end{remark} We define a map \[\textup{type} = \textup{type}_n:n\textup{-Mod}\to \altmathcal{T}_n\] where $\textup{type}(\mathbb{V})$ is the type of $\mathbb{V}$. We denote the collection of all zigzag modules of type $\tau$ by $\taumod{\tau}$, that is\footnote{Note that we are slightly abusing notation since $\taumod{\tau}$ may not be a set.} \[\taumod{\tau}:= \{\mathbb{V}\in n\textup{-Mod} \ | \ \textup{type}(\mathbb{V}) = \tau\}.\] Note that $\taumod{\tau} = \textup{type}^{-1}(\tau)$ and $n\textup{-Mod} = \bigcup_{\tau\in\altmathcal{T}_n}\taumod{\tau}$. A zigzag module of type $\tau$ is also called a \textit{$\tau$-module}. \begin{example} Let $n = 3$. Then \[ \altmathcal{T}_n = \{\underbrace{(\rightarrow,\rightarrow)}_{\tau_1},\underbrace{(\rightarrow,\leftarrow)}_{\tau_2},\underbrace{(\leftarrow,\rightarrow)}_{\tau_3},\underbrace{(\leftarrow,\leftarrow)}_{\tau_4}\}.\] Consider the zigzag modules \begin{equation*} \begin{aligned}[c] \mathbb{V}_1 &= \begin{tikzcd}[column sep=2em] \mathbb{F}\arrow[r,"\textup{id}"]& \mathbb{F} \arrow[r,"\textup{id}"]& \mathbb{F} \end{tikzcd}\\ \mathbb{V}_3 &= \begin{tikzcd}[column sep=2em] \mathbb{F}&\arrow[l,swap,"0"]0\arrow[r,"0"] & \mathbb{F} \end{tikzcd}\\ \end{aligned}\quad \begin{aligned}[c] \mathbb{V}_2 &= \begin{tikzcd}[column sep=2em] \mathbb{F}\arrow[r,"\textup{id}"]& \mathbb{F} & 0 \arrow[l,swap,"0"] \end{tikzcd}\\ \mathbb{V}_4 &= \begin{tikzcd}[column sep=2em] 0\arrow[r,leftarrow,"0"]& \mathbb{F}& \mathbb{F} \arrow[l,swap,"\textup{id}"] \end{tikzcd}\\ \end{aligned} \end{equation*} where $\mathbb{F}$ is viewed as a $1$-dimensional vector space over itself and $0$ denotes the trivial vector space. Each of these zigzag modules is an element of $n\textup{-Mod}$ and we have $\textup{type}(\mathbb{V}_j) = \tau_j$ for $j = 1, 2,3 ,4$. \end{example} \subsection{Morphisms Between Zigzag Modules} Fix zigzag modules $\mathbb{V} = (V_i,p_i)$ and $\mathbb{W} = (W_i,q_i)$ in $\taumod{\tau}$. A \textit{morphism} from $\mathbb{V}$ to $\mathbb{W}$ is a collection $\phi = \{\phi_i:V_i\to W_i\}_{i = 1}^n$ of linear transformations such that the diagram \[ \begin{tikzcd}[row sep=3em,column sep=4em] V_1\arrow[r,"p_1"]\arrow[d,"\phi_1"]& V_2 \arrow[l]\arrow[r,"p_2"]\arrow[d,"\phi_2"]&\cdots\arrow[l]\arrow[r,"p_{n-2}"]& V_{n-1}\arrow[l]\arrow[r,"p_{n-1}"]\arrow[d,"\phi_{n-1}"]& V_n\arrow[l]\arrow[d,"\phi_{n}"]\\ W_1\arrow[r,"q_1"]& W_2 \arrow[l]\arrow[r,"q_2"]&\cdots\arrow[l]\arrow[r,"q_{n-2}"]& W_{n-1}\arrow[l]\arrow[r,"q_{n-1}"]& V_n\arrow[l] \end{tikzcd} \] commutes. We denote a morphism from $\mathbb{V}$ to $\mathbb{W}$ by $\phi:\mathbb{V}\to\mathbb{W}$. The linear maps $\phi_i$ comprising the morphism $\phi$ are called the \textit{components} of $\phi$. Composition of morphisms is defined component-wise and the identity morphism $\textup{id}_\mathbb{V}:\mathbb{V}\to\mathbb{V}$ is the morphism all of whose components are identity maps. With these definitions in place, for every $n\in\mathbb{N}$ and for each $\tau\in\altmathcal{T}_n$, the collection $\taumod{\tau}$ of $\tau$-modules together with the collection of all morphisms between them forms a category denoted $\taumodcat{\tau}$. We call a morphism $\phi$ an \textit{isomorphism}, \textit{monomorphism}, or \textit{epimorphism} if all of the $\phi_i$ are either bijective, injective, or surjective, respectively. If there exists an isomorphism between $\mathbb{V}$ and $\mathbb{W}$ we say that $\mathbb{V}$ and $\mathbb{W}$ are \textit{isomorphic} and write $\mathbb{V}\cong \mathbb{W}$. Given a $\tau$-module $\mathbb{V} = (V_i,p_i)$, a \textit{submodule} $\mathbb{W} = (W_i,q_i)$ of $\mathbb{V}$ is a $\tau$-module such that for all $i$ either $p_i(W_i)\subseteq W_{i+1}$ or $p_i(W_{i+1})\subseteq W_i$, depending on whether $p_i:W_i\to W_{i+1}$ or $p_i:W_{i+1}\to W_{i}$, respectively. In this case, the linear map $q_{i}$ is taken to be the restrictions of the map $p_{i}$ to the subspace $W_i$ or $W_{i+1}$, again depending on the direction of $p_i$. A $\tau$-module $\mathbb{W}$ is isomorphic to a submodule of $\mathbb{V}$ if and only if there exists a monomorphism $f:\mathbb{W}\to\mathbb{V}$. This justifies working with monomorphisms between zigzag modules instead of working with submodules directly. We write $\mathbb{W}\leq \mathbb{V}$ whenever there exists a monomorphism $f:\mathbb{W}\to \mathbb{V}$. Note that $\mathbb{W}\cong \mathbb{V}$ if and only if $\mathbb{W}\leq \mathbb{V}$ and $\mathbb{V}\leq \mathbb{W}$. \subsection{Interval Modules and the Zero Module} Fix $n\in\mathbb{N}$ and $\tau\in\altmathcal{T}_n$. For each pair $b,d\in\{1,\hdots,n\}$ with $b\leq d$ we define a zigzag module $\mathbb{I}_{\tau}([b,d]) = (I_i,p_i)\in\taumod{\tau}$, called the \textit{interval $\tau$-module} on $[b,d]$, by setting \[ I_i:= \begin{cases} \mathbb{F} & b\leq i\leq d\\ 0 & \textup{otherwise}\end{cases}\quad \text{ and }\quad p_i := \begin{cases} \textup{id}_{\mathbb{F}} & b\leq i <d\\ 0 & \textup{otherwise}.\end{cases}\] When the type is fixed we will drop the subscript $\tau$ and just write $\mathbb{I}([b,d])$. Interval $\tau$-modules of the form $\mathbb{I}_\tau([k,k])$ are called \textit{simple interval $\tau$-modules}. We also define the zero module $\scalebox{.7}[1]{$\mathbb{O}$}_\tau = (Z_i,z_i)$ of type $\tau$ to be the $\tau$-module with $Z_i = 0$ and $z_i = 0$ for all $i$. \begin{example} Let $n = 3$ and let $\tau = (\rightarrow,\leftarrow)\in\altmathcal{T}_n$. There are $6 = \binom{n+1}{2}$ nonzero interval $\tau$-modules given by \begin{equation*} \begin{aligned}[c] \mathbb{I}_\tau([1,1]) &= \begin{tikzcd}[column sep=2em] \mathbb{F}\arrow[r,"0"]& 0 & 0 \arrow[l,swap,"0"] \end{tikzcd}\\ \mathbb{I}_\tau([1,3]) &= \begin{tikzcd}[column sep=2em] \mathbb{F}\arrow[r,"\textup{id}"]& \mathbb{F} & \mathbb{F} \arrow[l,swap,"\textup{id}"] \end{tikzcd}\\ \mathbb{I}_\tau([2,3]) &= \begin{tikzcd}[column sep=2em] 0\arrow[r,"0"]& \mathbb{F} & \mathbb{F} \arrow[l,swap,"\textup{id}"] \end{tikzcd}\\ \end{aligned}\quad \begin{aligned}[c] \mathbb{I}_\tau([1,2]) &= \begin{tikzcd}[column sep=2em] \mathbb{F}\arrow[r,"\textup{id}"]& \mathbb{F} & 0 \arrow[l,swap,"0"] \end{tikzcd}\\ \mathbb{I}_\tau([2,2]) &= \begin{tikzcd}[column sep=2em] 0\arrow[r,"0"]& \mathbb{F} & 0 \arrow[l,swap,"0"] \end{tikzcd}\\ \mathbb{I}_\tau([3,3]) &= \begin{tikzcd}[column sep=2em] 0\arrow[r,"0"]& 0 & \mathbb{F} \arrow[l,swap,"0"] \end{tikzcd}\\ \end{aligned} \end{equation*} \end{example} \section{Decompositions of Zigzag Modules} \label{sec:decompositions} In this section we define the direct sum of zigzag modules of the same type and state a standard unique decomposition theorem, an adaptation of the Krull-Remak-Schmidt theorem to the context of zigzag persistence. Combined with Gabriel's Theorem which characterizes the indecomposable zigzag modules as precisely the interval zigzag modules, we are able to define \textit{persistence diagrams}, an object of fundamental importance in persistence theory. \subsection{Indecomposables and Summands} The \textit{direct sum} of two $\tau$-modules $\mathbb{X} = (X_i,\alpha_i)$ and $\mathbb{Y} = (Y_i,\beta_i)$ is a $\tau$-module $\mathbb{X}\oplus\mathbb{Y} = (Z_i,\gamma_i)$ where $Z_i = X_i\oplus Y_i$ and where $\gamma_i = \alpha_i\oplus \beta_i$ for all $i$. We say that $\mathbb{W}$ is a \textit{summand} of $\mathbb{V}$ whenever there exists a $\tau$-module $\mathbb{U}$ such that $\mathbb{V} \cong \mathbb{W}\oplus\mathbb{U}$ and we write $\mathbb{W}\preceq \mathbb{V}$. The relation $\preceq$ defines a partial order on the isomorphism classes of $\taumod{\tau}$. For later reference, we record the following easy proposition describing the structure maps of a summand: \begin{proposition} \label{types_of_summand_maps}If $(W_i,q_i) = \mathbb{W}\preceq\mathbb{V} = (V_i,p_i)$ and if the $k^\textup{th}$ structure map $p_k$ of $\,\mathbb{V}$ is injective, surjective, or bijective then the $k^\textup{th}$ structure map $q_k$ of $\,\mathbb{W}$ is also injective, surjective, or bijective, respectively. \end{proposition} A $\tau$-module $\mathbb{V}$ is said to be \textit{decomposable} if there exists nonzero $\tau$-modules $\mathbb{W}$ and $\mathbb{U}$ such that $\mathbb{V} \cong \mathbb{W}\oplus \mathbb{U}$ and is said to be \textit{indecomposable} otherwise. The following important theorem says that every zigzag module decomposes as a sum of indecomposable modules and characterizes the indecomposables as the interval modules: \begin{theorem}[Krull-Remak-Schmidt, Gabriel \cite{zigzag},\cite{gabe}]\label{krull} For each $n\in\mathbb{N}$ and for every $\tau\in\altmathcal{T}_n$, the indecomposable $\tau$-modules are precisely the interval $\tau$-modules. Moreover, every $\mathbb{V}\in\taumod{\tau}$ decomposes as a direct sum of interval $\tau$-modules. This decomposition is unique up to the order in which the summands appear. \end{theorem} \subsection{Persistence Diagrams} Fix an $n\in \mathbb{N}$ and a type $\tau\in \altmathcal{T}_n$. By Theorem \ref{krull}, every $\tau$-module $\mathbb{V}\in\taumod{\tau}$ has a decomposition of the form \begin{equation}\label{decomp} \mathbb{V} \cong \mathbb{I}_\tau([b_1,d_1])\oplus\cdots\oplus\mathbb{I}_\tau([b_N,d_N]),\end{equation} this decomposition being unique up to the ordering of the summands. We define the \textit{persistence diagram} of $\mathbb{V}$ to be the multiset \[\mathrm{Dgm}(\mathbb{V}) : = \{\!\!\{(b_i,d_i)\in\mathbb{N}\times \mathbb{N} \ | \ 1\leq i\leq N\}\!\!\},\] whose elements are ordered pairs of endpoints defining the interval modules in the decomposition (\ref{decomp}). In particular, we always have the decomposition \begin{equation} \label{persdecomp} \mathbb{V} \cong \bigoplus_{(b,d)\in \mathrm{Dgm}(\mathbb{V})}\mathbb{I}_\tau([b,d]).\end{equation} Persistence diagrams thus characterize zigzag modules. That is, for fixed type $\tau$, a $\tau$-module determines and is determined up to isomorphism by its persistence diagram. There is a simple relationship between the persistence diagram of a zigzag module and the persistence diagram of any of its summands: \begin{proposition}\label{perssum} Fix $n\in \mathbb{N}$ and $\tau\in \altmathcal{T}_n$. If $\,\mathbb{W},\mathbb{V}\in\taumod{\tau}$ with $\mathbb{W}\preceq \mathbb{V}$ then $\mathrm{Dgm}(\mathbb{W})\subseteq\mathrm{Dgm}(\mathbb{V})$. \end{proposition} \begin{proof} Since $\mathbb{W}\preceq \mathbb{V}$ there exists $\mathbb{U}\in\taumod{\tau}$ such that $\mathbb{V}\cong \mathbb{W}\oplus \mathbb{U}$. Using the decomposition (\ref{persdecomp}), we have \[ \mathbb{V} \cong \left( \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{W})}\mathbb{I}([b,d])\right)\oplus \left( \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{U})}\mathbb{I}([b,d])\right).\] By the uniqueness statement of Theorem \ref{krull}, $\mathrm{Dgm}(\mathbb{V}) = \mathrm{Dgm}(\mathbb{W})\sqcup\mathrm{Dgm}(\mathbb{U})$ so that $\mathrm{Dgm}(\mathbb{W})\subseteq\mathrm{Dgm}(\mathbb{V})$. \end{proof} \section{Type Transformations and Arrow Reversals} \label{sec:type-transformations} In this section we define several transformations of types, i.e. maps $\altmathcal{T}_n\to\altmathcal{T}_n$, whose purposes are mainly formal; they will serve to define the appropriate domains and codomains for the transformations of zigzag modules defined in Section \ref{sect_ref_func}. \subsection{Sinks and Sources} Fix $n\in \mathbb{N}$ and let $\tau\in \altmathcal{T}_n$. A zigzag module $\mathds{V} = (V_i,p_i)\in\taumod{\tau}$ has a \textit{sink} at index $k\in\{2,\hdots,n-1\}$ if it has the form \[ \mathds{V} = \begin{tikzcd}[row sep=2em,column sep=3.5em] \cdots \arrow[r,"p_{k-2}"]& V_{k-1} \arrow[l]\arrow[r,"p_{k-1}"]& V_{k}& V_{k+1}\arrow[r,"p_{k+1}"]\arrow[l,swap,"p_k"]& \cdots\arrow[l]. \end{tikzcd}\] In addition, we say that $\mathds{V}$ has a \textit{sink} at index 1 or index $n$ if the maps $p_1$ or $p_{n-1}$ are of the form $V_{1} \stackrel{p_1}{\longleftarrow} V_{2}$ or $V_{n-1} \stackrel{p_{n-1}}{\longrightarrow} V_{n}$, respectively. Similarly, $\mathds{V} = (V_i,p_i)\in\taumod{\tau}$ has a \textit{source} at index $k\in\{2,\hdots,n-1\}$ if it has the form \[ \mathds{V} = \begin{tikzcd}[row sep=2em,column sep=3.5em] \cdots \arrow[r,"p_{k-2}"]& V_{k-1} \arrow[l]&\arrow[l,swap,"p_{k-1}"] V_{k}\arrow[r,"p_{k}"]& V_{k+1}\arrow[r,"p_{k+1}"]& \cdots\arrow[l], \end{tikzcd}\] and has a \textit{source} at index $1$ or $n$ if the maps $p_1$ or $p_n$ are of the form $V_{1} \stackrel{p_1}{\longrightarrow} V_{2}$ or $V_{n-1} \stackrel{p_{n-1}}{\longleftarrow} V_{n}$, respectively. Equivalently, a $\tau$-module $\mathds{V}$ has a sink at index $k\in\{1,\hdots,n\}$ if none of the linear maps $p_i$ have domain $V_k$, and $\mathds{V}$ has a source at index $k$ if none of the linear maps $p_i$ have codomain $V_k$. Note that the property of having a sink or source at a given index depends only on the type of the zigzag module in question; that is, if a $\tau$-module $\mathds{V}$ has a sink or source at index $k\in\{1,\hdots,n\}$ then any other $\tau$-module will also have, respectively, a sink or source at index $k$. This leads to the following definitions: \begin{definition}\label{sink_source_flow}We say a type $\tau\in\altmathcal{T}_n$ has a \textit{sink} or \textit{source} at index $k$ if any (and hence every) $\tau$-module has, respectively, a sink or source at index $k$. Otherwise, $\tau$ is said to have a \textit{flow} at index $k$. If both the $(k-1)^\textup{st}$ and $k^\textup{th}$ entries of $\tau$ are $\rightarrow$, then $\tau$ is said to have a $\textit{forward flow}$, and a flow which is not a forward flow is called a \textit{backwards flow}. \end{definition} \begin{caution} When we make mention of a type $\tau\in\altmathcal{T}_n$ as having a sink or source at index $k\in\{1,\hdots,n\}$, we mean that any $\tau$-module has a sink or source, respectively, at index $k$. The word ``index" here \textbf{does not} refer to the $k^{\textup{th}}$ component of the sequence of arrows defining $\tau$. \end{caution} \subsection{Type Transformations} \begin{definition} \label{type_reversal}Fix $n\in \mathbb{N}$. For each $k\in\{1,\hdots,n-1\}$, we define the \textit{$k^\textup{th}$ reversal map} \[ r_k:\altmathcal{T}_n\to \altmathcal{T}_n\] which maps a type $\tau$ to a type $r_k\tau$ whose $k^{th}$ entry is obtained by reversing the $k^\textup{th}$ entry of $\tau$. \end{definition} \begin{example} Let $n = 4$ and $\tau = (\rightarrow,\leftarrow,\rightarrow)\in\altmathcal{T}_n$. We have \[r_1\tau = (\leftarrow,\leftarrow,\rightarrow), \quad r_2\tau = (\rightarrow,\rightarrow,\rightarrow), \quad r_3\tau = (\rightarrow,\leftarrow,\leftarrow).\] \end{example} \begin{definition}\label{type_extroversion} For fixed $n$ and for each $k\in\{1,\hdots,n\}$, we define the \textit{$k^\textup{th}$ extroversion map} \[ \sigma_k:\altmathcal{T}_n\to\altmathcal{T}_n\] which maps a type $\tau$ to a type $\sigma_k\tau$ which is obtained by placing a \textit{source} at index $k$ in $\tau$. \end{definition} More precisely, for $k\in\{2,\hdots,n-1\}$, $\sigma_k\tau$ is obtained from $\tau$ by replacing the $(k-1)^\textup{st}$ and $k^\textup{th}$ entries of $\tau$ by $\leftarrow$ and $\rightarrow$, respectively. The type $\sigma_1\tau$ is obtained by replacing the $1^\textup{st}$ entry of $\tau$ by $\rightarrow$ and the type $\sigma_n\tau$ is obtained by replacing the $(n-1)^\textup{st}$ entry of $\tau$ by $\leftarrow$. Similarly, we make the following: \begin{definition} \label{type_introversion} For fixed $n$ and for each $k\in\{1,\hdots,n\}$, we define the \textit{$k^\textup{th}$ introversion map} \[ \zeta_k:\altmathcal{T}_n\to\altmathcal{T}_n\] which maps a type $\tau$ to the type $\zeta_k\tau$ which is obtained by placing a \textit{sink} at index $k$ in $\tau$. \end{definition} \begin{example}\label{type_intro_extro_example} Let $n = 4$ and $\tau = (\rightarrow,\rightarrow,\leftarrow)\in\altmathcal{T}_n$. Then we have \begin{equation*} \begin{aligned}[c] \sigma_1\tau = \tau & = (\rightarrow,\rightarrow,\leftarrow) \\ \sigma_3\tau & = (\rightarrow,\leftarrow,\rightarrow) \end{aligned}\quad \begin{aligned}[c] \sigma_2\tau & = (\leftarrow,\rightarrow,\leftarrow)\\ \sigma_4\tau = \tau& = (\rightarrow,\rightarrow,\leftarrow). \end{aligned} \end{equation*} and \begin{equation*} \begin{aligned}[c] \zeta_1\tau & = (\leftarrow,\rightarrow,\leftarrow) \\ \zeta_3\tau = \tau & = (\rightarrow,\rightarrow,\leftarrow) \end{aligned}\quad \begin{aligned}[c] \zeta_2\tau & = (\rightarrow,\leftarrow,\leftarrow)\\ \zeta_4\tau & = (\rightarrow,\rightarrow,\rightarrow). \end{aligned} \end{equation*} \end{example} \subsection{Arrow Reversals} We wish to identify zigzag modules $\mathbb{V},\mathbb{W}\inn\textup{-Mod}$ which differ only in the direction of arrows representing isomorphisms. For example, the zigzag modules \[ \begin{tikzcd}[column sep=2em] 0\arrow[r,"0"]&\mathbb{F}\arrow[r,"\textup{id}"]& \mathbb{F}\arrow[r,"0"] & 0 \end{tikzcd}\quad \textup{ and } \quad \begin{tikzcd}[column sep=2em] 0\arrow[r,"0"]&\mathbb{F}\arrow[r,leftarrow,"\textup{id}"]& \mathbb{F}\arrow[r,"0"] & 0 \end{tikzcd}\] contain the same information and we wish to regard them as equivalent. The goal of this section is to establish notation for dealing with zigzag modules which are to be regarded as equivalent in this way. \begin{definition} Fix $\tau\in\altmathcal{T}_n$. For each $k\in\{1,\hdots,n-1\}$ we define $\taumod{\tau}^{\textup{iso},k}\subset\taumod{\tau}$ by setting \[\taumod{\tau}^{\textup{iso},k} : = \{\mathbb{V} = (V_i,p_i) \ | \ p_k \textup{ is an isomorphism}\}.\] Recall the type reversal map $r_k:\altmathcal{T}_n\to\altmathcal{T}_n$ of Definition \ref{type_reversal} which reverses the $k^\textup{th}$ arrow of a given $\tau\in \altmathcal{T}_n$. We define a map \[\altmathcal{A}_k:\taumod{\tau}^{\textup{iso},k}\to \taumod{r_k\tau}^{\textup{iso},k}\] by setting $\altmathcal{A}_k(\mathbb{V}) = (V_i,q_i)\in\taumod{r_k\tau}$, where $q_i = p_i$ for $i\neq k$ and $q_k = p_k^{-1}$. \end{definition} \begin{remarks}\label{arr_rev_remarks} \hfill \begin{enumerate} \item If $V_{i} \stackrel{p}{\longrightarrow} V_{j}$ is an isomorphism appearing in $\mathbb{V}$, and $V_j\stackrel{q}{\longleftrightarrow} V_k$ is an adjacent arrow, then we have the following isomorphisms of zigzag modules: \[ \begin{tikzcd}[row sep=3em,column sep=4em] \cdots \arrow[r]& V_i\arrow[l]\arrow[r,"p"]\arrow[d,"\textup{id}"]&V_j\arrow[r,"q"]\arrow[d,"p^{-1}"]& V_k \arrow[l]\arrow[r]\arrow[d,"\textup{id}"]&\cdots\arrow[l]\\ \cdots \arrow[r]& V_i\arrow[l]\arrow[r,"\textup{id}"]&V_i\arrow[r,"\alpha"]& V_k \arrow[l]\arrow[r]&\cdots\arrow[l] \end{tikzcd} \] where $\alpha = qp$ if $V_j\stackrel{q}{\longrightarrow}V_k$ and $\alpha = p^{-1}q$ if $V_j\stackrel{q}{\longleftarrow}V_k$ (and all vertical maps not appearing are identities). Thus it is safe to assume without loss of generality that all structure maps which are isomorphisms are in fact identities, in which case $\altmathcal{A}_k(\mathbb{V})$ is obtained by changing the type of $\,\mathbb{V}$ from $\tau$ to $r_k\tau$ but leaving all of the vector spaces $V_i$ and linear maps $p_i$ unchanged. \item Note that $\altmathcal{A}_k\circ\altmathcal{A}_k = \textup{id}_{\taumod{\tau}^{\textup{iso},k}}$ since reversing the direction of an isomorphism twice leaves the zigzag module unchanged. \end{enumerate} \end{remarks} We make the following observations about arrow reversals: \begin{proposition}\label{arrow_reversal_properties} Let $\tau\in\altmathcal{T}_n$ and suppose that $\mathbb{V}\in\taumod{\tau}^{\textup{iso},i}$ for some $i\in\{1,\dots,n-1\}$. \begin{enumerate}[(1)] \item If $\,\mathbb{V}\cong \mathbb{W}$ then $\mathbb{W}\in\taumod{\tau}^{\textup{iso},i}$ and $\altmathcal{A}_i(\mathbb{V})\cong \altmathcal{A}_i(\mathbb{W})$, \item If $\,\mathbb{W}\in\taumod{\tau}^{\textup{iso},i}$ then $\mathbb{V}\oplus\mathbb{W}\in\taumod{\tau}^{\textup{iso},i}$ and $\altmathcal{A}_i(\mathbb{V}\oplus\mathbb{W}) = \altmathcal{A}_i(\mathbb{V})\oplus\altmathcal{A}_i(\mathbb{W})$, \item If $\,\mathbb{W}\preceq \mathbb{V}$ then $\mathbb{W}\in\taumod{\tau}^{\textup{iso},i}$ and $\altmathcal{A}_i(\mathbb{W})\preceq\altmathcal{A}_i(\mathbb{V})$, \item If $\,\mathbb{V}\in\taumod{\tau}^{\textup{iso},i}\cap\taumod{\tau}^{\textup{iso},j}$ for some $j\in\{1,\dots,n-1\}$ then $\altmathcal{A}_i\altmathcal{A}_j(\mathbb{V}) = \altmathcal{A}_j\altmathcal{A}_i(\mathbb{V})$. \end{enumerate} \begin{proof} (1) That $\mathbb{W}\in\taumod{\tau}^{\textup{iso},i}$ is immediate from the commutativity relations imposed on the components of a $\tau$-module isomorphism. Moreover, it is easily verified that the components of any isomorphism $\phi:\mathbb{V}\to\mathbb{W}$ of $\tau$-modules will also serve as the components of an isomorphism of $r_i\tau$-modules between $\altmathcal{A}_i(\mathbb{V})$ and $\altmathcal{A}_i(\mathbb{W})$. (2) This follows immediately after noticing that $(p_i\oplus q_i)^{-1} = p_i^{-1}\oplus q_i^{-1}$, where $p_i$ and $q_i$ denote the $i^\textup{th}$ structure maps of $\mathbb{V}$ and $\mathbb{W}$, respectively. (3) If $\mathbb{W}\preceq\mathbb{V}\in\taumod{\tau}^{\textup{iso},i}$ then the $i^\textup{th}$ structure map of $\mathbb{V}$ is an isomorphism so that, by Proposition \ref{types_of_summand_maps}, the $i^\textup{th}$ structure map of $\mathbb{W}$ is an isomorphism as well and hence $\mathbb{W}\in\taumod{\tau}^{\textup{iso},i}$. Now if $\mathbb{U}\in \taumod{\tau}$ is such that $\mathbb{V} \cong \mathbb{W}\oplus\mathbb{U}$ then by parts (1) and (2), we have \[\altmathcal{A}_i(\mathbb{V})\cong \altmathcal{A}_i(\mathbb{W}\oplus\mathbb{U}) = \altmathcal{A}_i(\mathbb{W})\oplus\altmathcal{A}_i(\mathbb{U}),\] and thus $\altmathcal{A}_i(\mathbb{W})\preceq\altmathcal{A}_i(\mathbb{V})$. (4) If $i = j$ then the result follows immediately. If $i\neq j$, then the result follows by noting that $\altmathcal{A}_i$ and $\altmathcal{A}_j$ operate on different linear maps $p_i$ and $p_j$, so that the order in which they are applied does not matter. \end{proof} \end{proposition} \subsection{Equivalence of Zigzag Modules} We now define an equivalence relation on $n\textup{-Mod}$, formalizing our discussion at the beginning of the previous section. For $\mathbb{V},\mathbb{W}\inn\textup{-Mod}$ we write $\mathbb{V}\sim\mathbb{W}$ if and only if either $\mathbb{V}\cong \mathbb{W}$ or there is a finite sequence $k_1,\hdots,k_j$ of indices in $\{1,\hdots,n-1\}$ such that \[\mathbb{W}\cong \altmathcal{A}_{k_j}\altmathcal{A}_{k_{j-1}}\cdots\altmathcal{A}_{k_1}(\mathbb{V}).\] In words, $\mathbb{V}\sim\mathbb{W}$ if $\mathbb{W}$ can be obtained, up to isomorphism, from $\mathbb{V}$ by reversing some (possibly empty) set of arrows representing isomorphisms. Reflexivity of the relation $\sim$ is clear, while symmetry follows from Remark \ref{arr_rev_remarks} (2) and transitivity from Proposition \ref{arrow_reversal_properties} (1). Thus $\sim$ does indeed define an equivalence relation on the isomorphism classes of $n\textup{-Mod}$. \begin{remark} Note that $\scalebox{.7}[1]{$\mathbb{O}$}_\tau\sim\scalebox{.7}[1]{$\mathbb{O}$}_{\tau'}$ for any types $\tau,\tau'\in\altmathcal{T}_n$. Hence, in what follows we drop the subscript indicating type and denote the zero module of any type by $\scalebox{.7}[1]{$\mathbb{O}$}$. \end{remark} The proof of the next proposition is sketched out by Oudot in \cite{oudotbook}; we give the full details here: \begin{proposition}[\cite{oudotbook}]\label{simpers} Let $\mathbb{V},\mathbb{W}\inn\textup{-Mod}$. If $\,\mathbb{V}\sim\mathbb{W}$ then \[\mathrm{Dgm}(\mathbb{V}) = \mathrm{Dgm}(\mathbb{W}).\] \begin{proof} Let $\tau = \textup{type}(\mathbb{V})$ and write $\mathbb{V} \cong \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\mathbb{I}_\tau([b,d])$. Since $\mathbb{V}\sim\mathbb{W}$, there is a sequence of indices $k_1,k_2,\hdots,k_j\in\{1,\hdots,n-1\}$ such that \[\mathbb{W}\cong \altmathcal{A}_{k_j}\altmathcal{A}_{k_{j-1}}\cdots\altmathcal{A}_{k_1}(\mathbb{V}).\] Let $\tau' = r_{k_j}r_{k_{j-1}}\cdots r_{k_1}\tau$ so that $\textup{type}(\mathbb{W}) = \tau'$ and consider the zigzag module $\mathbb{U}\in\taumod{\tau'}$ defined by \[\mathbb{U}:= \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\mathbb{I}_{\tau'}([b,d]).\] By definition of $\mathbb{U}$, we have $\mathrm{Dgm}(\mathbb{U}) = \mathrm{Dgm}(\mathbb{V})$. We claim that $\mathbb{U}\cong\mathbb{W}$. To see this, notice that if $V_{i} \stackrel{p}{\longrightarrow} V_{j}$ is a structure map of $\mathbb{V}$ with $p$ being an isomorphism and $i,j$ being consecutive integers in $\{1,\hdots,n\}$, and if $d :=\textup{dim}(V_i) = \textup{dim}(V_{j})$, then there is an isomorphism $\psi:V_i\to \bigoplus_{m=1}^d\mathbb{F}$ so that the diagram \[\begin{tikzcd}[row sep=2.5em,column sep=3.5em] V_i\arrow[r,"p"]\arrow[d,"\psi"]& V_j \arrow[d,"\psi\circ p^{-1}"]\\ \bigoplus_{m = 1}^{d}\mathbb{F}\arrow[r,"\textup{id}"]& \bigoplus_{m = 1}^{d}\mathbb{F} \end{tikzcd}\] commutes. The above diagram commutes if and only if the diagram \[\begin{tikzcd}[row sep=2.5em,column sep=3.5em] V_i\arrow[d,"\psi"]& V_j \arrow[l,swap,"p^{-1}"]\arrow[d,"\psi\circ p^{-1}"]\\ \bigoplus_{m = 1}^{d}\mathbb{F}&\arrow[l,swap,"\textup{id}"] \bigoplus_{m = 1}^{d}\mathbb{F} \end{tikzcd}\] commutes. Applying this principle to every square at which an arrow reversal is applied, we see that $\mathbb{U}\cong \mathbb{W}$. Hence $\mathrm{Dgm}(\mathbb{V}) = \mathrm{Dgm}(\mathbb{U}) = \mathrm{Dgm}(\mathbb{W})$ by Theorem \ref{krull}. \end{proof} \end{proposition} \begin{definition} We define a relation $\precsim$ on $n\textup{-Mod}$ by declaring $\mathbb{W}\precsim \mathbb{V}$ if and only if there exists a zigzag module $\mathbb{W}'\inn\textup{-Mod}$ with $\mathbb{W}\sim \mathbb{W}'$ and $\mathbb{W}'\preceq\mathbb{V}$. \end{definition} In words, $\mathbb{W}\precsim\mathbb{V}$ if we can obtain a summand of $\mathbb{V}$ by reversing any number of the arrows of $\mathbb{W}$ representing isomorphisms. \begin{proposition}\label{precsim_preorder} $\precsim$ is a preorder on $n\textup{-Mod}$. Moreover, $\mathbb{V}\sim \mathbb{W}$ if and only if $\mathbb{W}\precsim\mathbb{V}$ and $\mathbb{V}\precsim\mathbb{W}$ so that $\precsim$ induces a partial order on $n\textup{-Mod}/\sim$. \begin{proof} Since $\mathbb{V}\sim\mathbb{V}$ and $\mathbb{V}\preceq\mathbb{V}$, we have $\mathbb{V}\precsim\mathbb{V}$. If $\mathbb{V}_1\precsim\mathbb{V}_2$ and $\mathbb{V}_2\precsim\mathbb{V}_3$ then there are zigzag modules $\mathbb{W}_1$ and $\mathbb{W}_2$ such that $\mathbb{V}_1\sim\mathbb{W}_1\preceq\mathbb{V}_2$ and $\mathbb{V}_2\sim\mathbb{W}_2\preceq\mathbb{V}_3$. That is, there are compositions of arrow reversals $\altmathcal{A}$ and $\altmathcal{B}$ such that $\altmathcal{A}(\mathbb{V}_1) \cong\mathbb{W}_1\preceq \mathbb{V}_2$ and $\altmathcal{B}(\mathbb{V}_2)\cong\mathbb{W}_2\preceq \mathbb{V}_3$. Then by Proposition \ref{arrow_reversal_properties} parts (1) and (3), we have $\altmathcal{B}\altmathcal{A}(\mathbb{V}_1) \preceq\altmathcal{B}(\mathbb{V}_2)\cong\mathbb{W}_2\preceq \mathbb{V}_3$ so that $\mathbb{V}_1\precsim\mathbb{V}_3$. This shows that $\precsim$ is a preorder. Now if $\mathbb{V}\sim\mathbb{W}$ then $\mathbb{V}\sim\mathbb{W}\preceq\mathbb{W}$ and $\mathbb{W}\sim\mathbb{V}\preceq\mathbb{V}$ so that $\mathbb{W}\precsim\mathbb{V}$ and $\mathbb{V}\precsim\mathbb{W}$. Conversely, if $\mathbb{W}\precsim\mathbb{V}$ and $\mathbb{V}\precsim\mathbb{W}$ then there exists zigzag modules $\mathbb{W}'\preceq\mathbb{W}$ and $\mathbb{V}'\preceq\mathbb{V}$ such that $\mathbb{W}\sim\mathbb{V}'$ and $\mathbb{V}\sim\mathbb{W}'$. Let $\mathbb{U}_1,\mathbb{U}_2\inn\textup{-Mod}$ be such that \[ \mathbb{V}\cong \mathbb{V}'\oplus\mathbb{U}_1 \quad \textup{ and } \quad \mathbb{W}\cong \mathbb{W}'\oplus\mathbb{U}_2.\] Since $\mathbb{W}\sim\mathbb{V}'$, there is a composition of arrow reversals $\altmathcal{A}$ such that $\altmathcal{A}(\mathbb{W})\cong \mathbb{V}'$, and similarly there is a composition of arrow reversals $\altmathcal{B}$ such that $\altmathcal{B}(\mathbb{V})\cong \mathbb{W}'$. Hence $\mathbb{W}\cong \mathbb{W}'\oplus\mathbb{U}_2 \cong \altmathcal{B}(\mathbb{V})\oplus \mathbb{U}_2$. Then we have \begin{align*}\mathbb{V}' & \cong \altmathcal{A}(\mathbb{W}) \cong \altmathcal{A}(\altmathcal{B}(\mathbb{V})\oplus\mathbb{U}_2)\\ & = \altmathcal{A}(\altmathcal{B}(\mathbb{V}))\oplus\altmathcal{A}(\mathbb{U}_2) \cong \altmathcal{A}(\altmathcal{B}(\mathbb{V}'\oplus\mathbb{U}_1))\oplus\altmathcal{A}(\mathbb{U}_2)\\ & = \altmathcal{A}(\altmathcal{B}(\mathbb{V}'))\oplus \altmathcal{A}(\altmathcal{B}(\mathbb{U}_1))\oplus \altmathcal{A}(\mathbb{U}_2). \end{align*} Using the above isomorphisms together with Propositions \ref{perssum} and \ref{simpers}, we have \begin{align*} \mathrm{Dgm}(\mathbb{V}') & = \mathrm{Dgm}(\altmathcal{A}(\altmathcal{B}(\mathbb{V}'))\oplus \altmathcal{A}(\altmathcal{B}(\mathbb{U}_1))\oplus \altmathcal{A}(\mathbb{U}_2))\\ & = \mathrm{Dgm}(\altmathcal{A}(\altmathcal{B}(\mathbb{V}')))\sqcup\mathrm{Dgm}(\altmathcal{A}(\altmathcal{B}(\mathbb{U}_1)))\sqcup\mathrm{Dgm}(\altmathcal{A}(\mathbb{U}_2)))\\ & = \mathrm{Dgm}(\mathbb{V}')\sqcup\mathrm{Dgm}(\mathbb{U}_1)\sqcup\mathrm{Dgm}(\mathbb{U}_2) \end{align*} so that $\mathrm{Dgm}(\mathbb{U}_1) = \emptyset = \mathrm{Dgm}(\mathbb{U}_2)$. Hence $\mathbb{U}_1\sim\scalebox{.7}[1]{$\mathbb{O}$}\sim\mathbb{U}_2$ so that in fact \[ \mathbb{V}\cong \mathbb{V}'\quad \textup{ and } \quad \mathbb{W}\cong \mathbb{W}'.\] Thus $\mathbb{V}\sim \mathbb{W}$, completing the proof. \end{proof} \end{proposition} \section{Reflection Functors}\label{sect_ref_func} \label{sec:ref-functors} In this section, we define \textit{reflection functors}, first introduced by Bernstein, Gelfand, and Ponomarev in \cite{bgp} as a tool for proving Gabriel's Theorem. Reflection functors were more recently applied to zigzag persistence by Kalisnik in \cite{kalisnik_thesis} to give an alternative proof of Carlsson and de Silva's Diamond Principle \cite{zigzag}. In the language of category theory, reflection functors send a diagram in the category of vector spaces to a new diagram in that same category, obtained by replacing particular subdiagrams by \textit{universal cones} or \textit{cocones}. We refer to the reader to Appendix \ref{app:cats} for a review of the necessary categorical notions. \subsection{Reflections on Zigzags}\label{sec:reflections} Suppose that $\mathbb{V} = (V_i,p_i)\in\taumod{\tau}$ for some $\tau\in \altmathcal{T}_n$. For $k\in \{2,\hdots,n-1\}$, we isolate the subdiagram \begin{equation}\label{subdiagram} \begin{tikzcd}[row sep=2em,column sep=3.5em] V_{k-1}\arrow[r,"p_{k-1}"]& V_k \arrow[l]\arrow[r,"p_k"] &V_{k+1}\arrow[l]. \end{tikzcd} \end{equation} We compute the limit $(L_k,\lambda_j)$ of the extracted diagram and then consider the new diagram \begin{equation} \label{extro_zigzag} \begin{tikzcd}[row sep=2em,column sep=3.5em] V_{k-1}&\arrow[l,swap,"\lambda_{k-1}"] L_k \arrow[r,"\lambda_k"] &V_{k+1}, \end{tikzcd} \end{equation} and then define $\altmathcal{L}_k(\mathbb{V})\in\taumod{\sigma_k\tau}$ to be the zigzag module obtained by replacing the appearance of the subdiagram (\ref{subdiagram}) in $\mathbb{V}$ by diagram (\ref{extro_zigzag}). That is, a map \[\altmathcal{L}_k:\taumod{\tau}\to \taumod{\sigma_k\tau}\] is specified by the following diamond diagram relating $\mathbb{V}$ and $\altmathcal{L}_k(\mathbb{V})$: \[\begin{tikzcd}[row sep=1.5em,column sep=2em] \mathbb{V} = \\ \\ \altmathcal{L}_k(\mathbb{V}) := \end{tikzcd} \begin{tikzcd}[row sep=2em,column sep=3em] &&&\arrow[dl] V_k&&&\\ V_1\arrow[r,"p_1"]& \cdots \arrow[r,"p_{k-2}"]\arrow[l]& V_{k-1} \arrow[l]\arrow[ur,"p_{k-1}"]& & V_{k+1}\arrow[r,"p_{k+1}"]\arrow[ul]\arrow[ul,swap,leftarrow,"p_k"]& \cdots\arrow[l]\arrow[r,"p_{n-1}"]&V_n\arrow[l]\\ &&& L_k\arrow[ul,"\lambda_{k-1}"]\arrow[ur,swap,"\lambda_k"]&&& \end{tikzcd}\quad\quad \] Similarly, if $(C_k,\gamma_j)$ is the colimit of diagram (\ref{subdiagram}), then we consider the new diagram \begin{equation} \label{intro_zigzag} \begin{tikzcd}[row sep=2em,column sep=3.5em] V_{k-1}\arrow[r,"\gamma_{k-1}"] &C_k &\arrow[l,swap,"\gamma_k"] V_{k+1}. \end{tikzcd} \end{equation} and define $\altmathcal{C}_k(\mathbb{V})\in\taumod{\zeta_k\tau}$ to be the zigzag module obtained by replacing the appearance of the subdiagram (\ref{subdiagram}) in $\mathbb{V}$ by diagram (\ref{intro_zigzag}). Thus, we have a map \[\altmathcal{C}_k:\taumod{\tau}\to \taumod{\zeta_k\tau}\] for which $\mathbb{V}$ and $\altmathcal{C}_k(\mathbb{V})$ are related by the diamond diagram \[\begin{tikzcd}[row sep=1.5em,column sep=2em] \mathbb{V} = \\ \\ \altmathcal{C}_k(\mathbb{V}) := \end{tikzcd} \begin{tikzcd}[row sep=2em,column sep=3em] &&&\arrow[dl] V_k&&&\\ V_1\arrow[r,"p_1"]& \cdots \arrow[r,"p_{k-2}"]\arrow[l]& V_{k-1} \arrow[l]\arrow[ur,"p_{k-1}"]& & V_{k+1}\arrow[r,"p_{k+1}"]\arrow[ul]\arrow[ul,swap,leftarrow,"p_k"]& \cdots\arrow[l]\arrow[r,"p_{n-1}"]&V_n\arrow[l]\\ &&& C_k\arrow[ul,leftarrow,"\gamma_{k-1}"]\arrow[ur,swap,leftarrow,"\gamma_k"]&&& \end{tikzcd}\quad\quad\] In order to define reflections at indices $1$ and $n$, we consider the diagrams \[ \begin{tikzcd}[row sep=2em,column sep=3.5em] 0\arrow[r,"0"]&V_{1}\arrow[l]\arrow[r,"p_{1}"]& V_2\arrow[l] \end{tikzcd} \qquad \textup{ and } \qquad \begin{tikzcd}[row sep=2em,column sep=3.5em] V_{n-1}\arrow[r,"p_{n-1}"]& V_n\arrow[l]\arrow[r,"0"]& 0 \arrow[l] \end{tikzcd} \] and their limits or colimits. These reflections thus depend on the choice of direction for the zero map. We wish to allow the flexibility of choosing the direction of this map each time a reflection is applied at index $1$ or $n$. Thus, we define reflections $\altmathcal{L}_1^\rightarrow$ and $\altmathcal{L}_1^\leftarrow$ obtained by replacing $V_1$ and $p_1$ with the limit of the diagram \[ \begin{tikzcd}[row sep=2em,column sep=3.5em] 0\arrow[r,"0"]&V_{1}\arrow[r,"p_{1}"]& V_2\arrow[l] \end{tikzcd} \qquad \textup{ or } \qquad \begin{tikzcd}[row sep=2em,column sep=3.5em] 0&\arrow[l,swap,"0"]V_{1}\arrow[r,"p_{1}"]& V_2\arrow[l] \end{tikzcd} \] respectively. The maps $\altmathcal{L}_n^\rightarrow$, $\altmathcal{L}_n^\leftarrow$, $\altmathcal{C}_1^\rightarrow$, $\altmathcal{C}_1^\leftarrow$, $\altmathcal{C}_n^\rightarrow$, and $\altmathcal{C}_n^\leftarrow$ are all defined analogously. For each $k\in \{2,\dots,n-1\}$, we let $\altmathcal{R}_k$ denote an unspecified choice of $\altmathcal{L}_k$ or $\altmathcal{C}_k$. Similarly $\altmathcal{R}_1$ denotes an unspecified choice of $\altmathcal{L}_1^\rightarrow$, $\altmathcal{L}_1^\leftarrow$, $\altmathcal{C}_1^\rightarrow$, or $\altmathcal{C}_1^\leftarrow$ and $\altmathcal{R}_n$ denotes an unspecified choice of $\altmathcal{L}_n^\rightarrow$, $\altmathcal{L}_n^\leftarrow$, $\altmathcal{C}_n^\rightarrow$, or $\altmathcal{C}_n^\leftarrow$. \subsection{Functoriality of Reflections} Let $\mathbb{V},\mathbb{W}\in\taumod{\tau}$ and let $\phi:\mathbb{V}\to\mathbb{W}$ be a morphism of $\tau$-modules. Denote the limits of the diagrams $V_{k-1}\stackrel{p_{k-1}}{\longleftrightarrow}V_k\stackrel{p_k}{\longleftrightarrow} V_{k+1}$ and $W_{k-1}\stackrel{q_{k-1}}{\longleftrightarrow}W_k\stackrel{q_k}{\longleftrightarrow} W_{k+1}$ by $(L_k^V,\lambda_j^V)$ and $(L_k^W,\lambda_j^W)$, respectively. Then we have the following commutative diagram: \[\footnotesize \begin{tikzcd}[row sep=2.25em,column sep=4.25em] & L_k^V \arrow[ddl,swap,thick,black,"\lambda_1^V"] \arrow[ddr,thick,black,"\lambda_2^V"] \arrow[dd,gray,"\lambda_3^V"] \arrow[drr,dashed,thick,black,"\exists!\mu"] \\ &&& L_W \arrow[ddr,thick,black,"\lambda_3^W"] \\ V_{k-1} \arrow[drr,thick,black,swap,"\phi_{k-1}"] \arrow[r,gray,"p_{k-1}"] &{\color{gray}V_k}\arrow[l,gray] \arrow[r,gray,"p_k"]\arrow[drr,gray,"\phi_k"]& V_{k+1}\arrow[l,gray] \arrow[drr,thick,black,"\phi_{k+1}"] \\ && W_{k-1} \arrow[r,swap,gray,"q_{k-1}"] \arrow[uur,leftarrow,crossing over,thick,black,"\lambda_1^W"] & {\color{gray}W_k}\arrow[r,swap,gray,"q_{k}"]\arrow[l,gray]& W_{k+1}\arrow[l,gray] \end{tikzcd}\] From this diagram, we see that $(L_k^V,\gamma_j^V)$ is a cone over $W_{k-1}\stackrel{q_{k-1}}{\longleftrightarrow}W_k\stackrel{q_k}{\longleftrightarrow} W_{k+1}$ so that by the universality of $L_k^W$, there exists a unique linear transformation $\mu:L_k^V\to L_k^W$ making the diagram commute. We then define $\psi = \altmathcal{L}_k(\phi):\altmathcal{L}_k(\mathbb{V})\to\altmathcal{L}_k(\mathbb{W})$ by setting $\psi_j = \phi_j$ for all $j\neq k$ and $\phi_k := \mu$. The fact that $\psi$ is a well-defined morphism follows from the commutativity of the boldened portion of the diagram above. Appealing to duality, we similarly obtain a morphism $\altmathcal{C}_k(\phi):\altmathcal{C}_k(\mathbb{V})\to\altmathcal{C}_k(\mathbb{W})$. \begin{definition} For each $k\in\{1,\hdots,n\}$ and for each $\tau\in\altmathcal{T}_n$, the above definitions make $\altmathcal{L}_k$ into a functor from $\taumodcat{\tau}$ to $\taumodcat{\sigma_k\tau}$, which we call the \textit{extroversion reflection functor} at index $k$. Similarly, $\altmathcal{C}_k$ is a functor from $\taumodcat{\tau}$ to $\taumodcat{\zeta_k\tau}$, which we call the \textit{introversion reflection functor} at index $k$. \end{definition} \begin{remark} When a zigzag module $\mathbb{V}$ has a \textit{sink} at index $k$, i.e., when $\mathbb{V}$ has the form \[ \mathbb{V} = \begin{tikzcd}[row sep=2em,column sep=3.5em] \cdots \arrow[r,"p_{k-2}"]& V_{k-1} \arrow[l]\arrow[r,"p_{k-1}"]& V_{k}& V_{k+1}\arrow[r,"p_{k+1}"]\arrow[l,swap,"p_k"]& \cdots\arrow[l]. \end{tikzcd}\] then, up to isomorphism, $\altmathcal{L}_k(\mathbb{V})$ is the same as the zigzag module obtained by applying the sink reflection functor of \cite{bgp} to $\mathbb{V}$. Dually, when $\mathbb{V}$ has as source at index $k$ then the zigzag module obtained by applying $\,\altmathcal{C}_k$ to $\mathbb{V}$ is isomorphic to that obtained by applying the source reflection functor. \end{remark} \subsection{Properties of Reflection Functors} A number of the results below hold for categorical reasons, and we will rely on several high-level results for their proofs. Statements and proofs of these general categorical results are contained in Appendix \ref{app:cats}. We will denote by $\mathbf{vect}_{\mathbb{F}}$ the category of finite-dimensional vector spaces over the field $\mathbb{F}$. \begin{proposition}\label{refsub} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V},\mathbb{W}\in\taumod{\tau}$. If $\,\mathbb{W}\leq \mathbb{V}$ then $\altmathcal{R}_k(\mathbb{W})\leq \altmathcal{R}_k(\mathbb{V})$. Furthermore, if $\,\mathbb{V}\cong\mathbb{W}$ then $\altmathcal{R}_k(\mathbb{V})\cong\altmathcal{R}_k(\mathbb{W})$. \begin{proof} If $\mathbb{W}\leq \mathbb{V}$ then there exists a monomorphism $j:\mathbb{W}\hookrightarrow \mathbb{V}$. By functoriality of $\altmathcal{R}_k$ we obtain a morphism $\altmathcal{R}_k(j):\altmathcal{R}_k(\mathbb{W})\rightarrow \altmathcal{R}_k(\mathbb{V})$. Theorem \ref{mono_diagrams_mono_lim} of the appendix implies that all of the components of $\altmathcal{R}_k(j)$ are monomorphisms in $\mathbf{vect}_{\mathbb{F}}$, i.e., are injective. Thus $\altmathcal{R}_k(j)$ is a monomorphism from $\altmathcal{R}_k(\mathbb{W})$ to $\altmathcal{R}_k(\mathbb{V})$ so that $\altmathcal{R}_k(\mathbb{W})\leq\altmathcal{R}_k(\mathbb{V})$. The second statement follows from the fact that $\mathbb{V}\cong\mathbb{W}$ if and only if $\mathbb{W}\leq\mathbb{V}$ and $\mathbb{V}\leq\mathbb{W}$. \end{proof} \end{proposition} \begin{proposition}\label{refsum} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V},\mathbb{W}\in\taumod{\tau}$. Then $\altmathcal{R}_k(\mathbb{V}\oplus \mathbb{W}) \cong \altmathcal{R}_k(\mathbb{V})\oplus\altmathcal{R}_k(\mathbb{W})$. This statement generalizes to the sum of any finite number of $\tau$-modules. \begin{proof} Viewing zigzag modules as diagrams in $\mathbf{vect}_{\mathbb{F}}$, the result follows from Theorem \ref{lim_prod_is_prod_lim}. The result extends to arbitrary finite sums by induction. \end{proof} \end{proposition} \begin{corollary}\label{refsummand} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V}$, $\mathbb{W}\in\taumod{\tau}$. If $\,\mathbb{W}\preceq \mathbb{V}$ then $\altmathcal{R}_k(\mathbb{W})\preceq\altmathcal{R}_k(\mathbb{V})$. \begin{proof}If $\mathbb{W}\preceq\mathbb{V}$ then $\mathbb{V} \cong \mathbb{W}\oplus\mathbb{U}$ for some $\mathbb{U}\in\taumod{\tau}$ so that by Propositions \ref{refsub} and \ref{refsum}, $$\altmathcal{R}_{k}(\mathbb{V})\cong \altmathcal{R}_k(\mathbb{W}\oplus\mathbb{U}) \cong \altmathcal{R}_k(\mathbb{W})\oplus\altmathcal{R}_k(\mathbb{U})$$ and hence $\altmathcal{R}_k(\mathbb{W})\preceq\altmathcal{R}_k(\mathbb{V})$. \end{proof} \end{corollary} \begin{corollary}\label{refprecsim} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V}$, $\mathbb{W}\inn\textup{-Mod}$. If $\,\mathbb{W}\precsim \mathbb{V}$ then $\altmathcal{R}_k(\mathbb{W})\precsim\altmathcal{R}_k(\mathbb{V})$. \begin{proof} If $\mathbb{W}\precsim \mathbb{V}$ then $\mathbb{W}\sim\mathbb{W}'\preceq \mathbb{V}$ for some $\mathbb{W}'$. Now note that that if $\mathbb{W}\sim \mathbb{W}'$ then $\altmathcal{R}_k(\mathbb{W})\sim \altmathcal{R}_k(\mathbb{W}')$. This follows from the fact that limits and colimits of diagrams are unaltered if any number of arrows representing isomorphisms are reversed. From this fact and Corollary \ref{refsummand}, we have \[\altmathcal{R}_k(\mathbb{W})\sim\altmathcal{R}_k(\mathbb{W}')\preceq \altmathcal{R}_k(\mathbb{V}),\] and hence $\altmathcal{R}_k(\mathbb{W})\precsim\altmathcal{R}_k(\mathbb{V})$. \end{proof} \end{corollary} \begin{theorem}\label{bigsumreflection} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V}\in\taumod{\tau}$. Then we have \[\altmathcal{R}_k(\mathbb{V}) \cong \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\altmathcal{R}_k(\mathbb{I}_\tau([b,d])).\] \begin{proof} Write $\mathbb{V} \cong \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\mathbb{I}_\tau([b,d])$. Using Propositions \ref{refsub} and \ref{refsum}, we have \[\altmathcal{R}_k(\mathbb{V}) \cong \altmathcal{R}_k\left(\bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\mathbb{I}_\tau([b,d])\right) \cong \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\altmathcal{R}_k(\mathbb{I}_\tau([b,d])).\qedhere\] \end{proof} \end{theorem} Theorem \ref{bigsumreflection} together with the following theorems describe exactly how reflections of zigzag modules effect their persistence diagrams: \begin{theorem}[The Diamond Principle Part I, \cite{bgp},\cite{zigzag},\cite{kalisnik_thesis}]\label{intervalreflections} Let $k\in\{2,\hdots,n-1\}$ and let $\tau\in\altmathcal{T}_n$ have a sink at index $k$ so that $\sigma_k\tau$ has a source at index $k$. Then the reflection functors $\altmathcal{L}_k$ and $\altmathcal{C}_k$ induce mutually inverse bijections between the isomorphism classes of interval $\tau$-modules and the isomorphism classes of interval $\sigma_k\tau$-modules, with the exception of the simple interval modules $\mathbb{I}_{\tau}([k,k])$ and $\mathbb{I}_{\sigma_k\tau}([k,k])$ which are annihilated by these functors.\footnote{We warn the reader that while we are referring to isomorphism classes of interval $\tau$ modules, in order to keep our notation simple, we will use the notation $\mathbb{I}_\tau([b,d])$ instead of the more correct notation $[\mathbb{I}_\tau([b,d])]$. } These functors act on the (isomorphism classes of) interval modules as follows: \begin{equation*} \begin{aligned} \mathbb{I}_\tau([k,k])\, &&\longrightarrow & \quad \scalebox{.7}[1]{$\mathbb{O}$} &&\\ \scalebox{.7}[1]{$\mathbb{O}$} && \longleftarrow \,\;& \quad \mathbb{I}_{\sigma_k\tau}([k,k])&&\\ \mathbb{I}_\tau([b,k-1])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([b,k]) && \quad \text{for }b\leq k-1\\ \mathbb{I}_\tau([b,k])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([b,k-1]) && \quad \text{for }b\leq k-1\\ \mathbb{I}_\tau([k+1,d])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([k,d]) && \quad \text{for }d\geq k+1\\ \mathbb{I}_\tau([k,d])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([k+1,d]) && \quad \text{for }d\geq k+1\\ \mathbb{I}_\tau([b,d])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([b,d]) && \quad \text{otherwise}.\\ \end{aligned} \end{equation*} For $k = 1$, we have that $\altmathcal{L}_1^\rightarrow$ and $\altmathcal{C}_1^\leftarrow$ are mutually inverse except on the simple interval modules $\mathbb{I}_\tau([1,1])$ and $\mathbb{I}_{\sigma_1\tau}([1,1])$: \begin{equation*} \begin{aligned} \mathbb{I}_\tau([1,1])\, &&{\longrightarrow} & \quad \scalebox{.7}[1]{$\mathbb{O}$} &&\\ \scalebox{.7}[1]{$\mathbb{O}$} && \longleftarrow \,\; & \quad \mathbb{I}_{\sigma_1\tau}([1,1])&&\\ \mathbb{I}_\tau([1,d])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_1\tau}([2,d]) && \quad \text{for }d\geq 2\\ \mathbb{I}_\tau([2,d])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_1\tau}([1,d]) && \quad \text{for }d\geq 2\\ \end{aligned} \end{equation*} and for $k = n$, we have that $\altmathcal{L}_n^\leftarrow$ and $\altmathcal{C}_n^\rightarrow$ are mutually inverse except on the simple interval modules $\mathbb{I}_\tau([n,n])$ and $\mathbb{I}_{\sigma_1\tau}([n,n])$: \begin{equation*} \begin{aligned} \mathbb{I}_\tau([n,n]) &&{\longrightarrow} & \quad \scalebox{.7}[1]{$\mathbb{O}$} &&\\ \scalebox{.7}[1]{$\mathbb{O}$} && \longleftarrow \,\; & \quad \mathbb{I}_{\sigma_n\tau}([n,n])&&\\ \mathbb{I}_\tau([b,n])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_n\tau}([b,n-1]) && \quad \text{for }b\leq n-1.\\ \mathbb{I}_\tau([b,n-1])\, &&\longleftrightarrow & \quad\mathbb{I}_{\sigma_n\tau}([b,n]) && \quad \text{for }b\leq n-1.\\ \end{aligned} \end{equation*} In the above, arrows pointing from left to right denote the effect of applying $\altmathcal{L}_k$ (or $\altmathcal{L}_k^{\leftarrow}, \altmathcal{L}_k^{\rightarrow}$) while arrow pointing from right to left denote the effect of applying $\altmathcal{C}_k$ (or $\altmathcal{C}_k^{\leftarrow}, \altmathcal{C}_k^{\rightarrow}$). \end{theorem} A statement completely analogous to Theorem \ref{intervalreflections} holds in the case that $\tau$ has a source at index $k$, with the roles of $\altmathcal{L}_k$ and $\altmathcal{C}_k$ reversed, $\altmathcal{L}_1^\rightarrow$, $\altmathcal{C}_1^\leftarrow$, $\altmathcal{L}_n^\leftarrow$, $\altmathcal{C}_n^\rightarrow$ replaced with $\altmathcal{C}_1^\leftarrow$, $\altmathcal{L}_1^\rightarrow$, $\altmathcal{C}_n^\rightarrow$, $\altmathcal{L}_n^\leftarrow$ respectively, and with $\sigma$ begin replaced with $\zeta$, but with the actions on interval modules being otherwise identical. \begin{figure} \centering \scalebox{1.5}{ \begin{tikzpicture} \draw[-latex,very thick] (1,0.5)--(3.5,0.5);\draw[-latex,very thick] (0.5,1)--(0.5,3.5); \draw[dashed,very thick] (0.5,0.5)--(1,0.5); \draw[dashed,very thick] (0.5,0.5)--(0.5,1); \draw[very thick] (0.479,0.5)--(0.5,0.5); \fill[gray] (1.5,2) circle [radius=1.75pt]; \fill[gray] (1,2.5) circle [radius=1.75pt]; \fill[gray] (1.5,2) circle [radius=1.75pt];\fill[gray] (2,2) circle [radius=1.75pt]; \fill[gray] (1,1.5) circle [radius=1.75pt];\fill[gray] (1,2) circle [radius=1.75pt]; \fill[gray] (1.5,1.5) circle [radius=1.75pt];\fill[gray] (1,1) circle [radius=1.75pt]; \fill[gray] (2.5,3) circle [radius=1.75pt];\fill[gray] (2,3) circle [radius=1.75pt]; \fill[gray] (1.5,3) circle [radius=1.75pt];\fill[gray] (1,3) circle [radius=1.75pt]; \fill[gray] (2.5,2.5) circle [radius=1.75pt];\fill[gray] (2,2.5) circle [radius=1.75pt]; \fill[gray] (1.5,2.5) circle [radius=1.75pt];\fill[gray] (3,3) circle [radius=1.75pt]; \tikzset{cross/.style={cross out, draw=red, fill=none, minimum size=3, inner sep=0pt, outer sep=0pt}, cross/.default={2pt}} \node[draw,cross out,red,inner sep = 1pt,minimum size = 4.5pt] at (2,2){}; \draw[-latex,semithick] (1.5,1.8)--(1.5,1.5); \draw[-latex,semithick] (1.5,1.7)--(1.5,2); \draw[-latex,semithick] (1.0,1.8)--(1.0,1.5); \draw[-latex,semithick] (1.0,1.7)--(1.0,2); \draw[-latex,semithick] (2.2,2.5)--(2.5,2.5); \draw[-latex,semithick] (2.3,2.5)--(2,2.5); \draw[-latex,semithick] (2.2,3.0)--(2.5,3.0); \draw[-latex,semithick] (2.3,3.0)--(2,3.0); \footnotesize \draw[very thin] (2,0.55) -- (2,0.45)node[anchor=north ] {$k$}; \draw[very thin] (2.5,0.55) -- (2.5,0.45) \draw[very thin] (3,0.45) -- (3,0.55); \draw[very thin] (1,0.55) -- (1,0.45); \draw[very thin] (1.5,0.55) -- (1.5,0.45) \footnotesize \draw[very thin] (.55,2) -- (.45,2)node[anchor=east ] {$k$}; \draw[very thin] (.55,2.5) -- (.45,2.5) \draw[very thin] (.55,3) -- (0.45,3); \draw[very thin] (.55,1) -- (0.45,1); \draw[very thin] (.55,1.5) -- (.45,1.5) \end{tikzpicture}\quad \begin{tikzpicture} \draw[-latex,very thick] (1,0.5)--(3.5,0.5);\draw[very thick] (0.5,0.5)--(0.5,1.5); \draw[dashed,very thick] (0.5,0.5)--(1,0.5); \draw[very thick,black] (1,1.25)--(3,1.25); \draw[very thick,orange] (1,1.0)--(2,1.0); \draw[very thick,red] (2,.75)--(3,.75); \footnotesize \draw[very thin] (2,0.55) -- (2,0.45)node[anchor=north ] {$k$}; \draw[very thin] (2.5,0.55) -- (2.5,0.45) \draw[very thin] (3,0.45) -- (3,0.55); \draw[very thin] (1,0.55) -- (1,0.45); \draw[very thin] (1.5,0.55) -- (1.5,0.45) \draw[-latex,very thick] (1,-1.5)--(3.5,-1.5);\draw[very thick] (0.5,-1.5)--(0.5,-0.5); \draw[dashed,very thick] (0.5,-1.5)--(1,-1.5); \draw[very thick,black] (1,-0.75)--(3,-0.75); \draw[very thick,orange] (1,-1)--(1.5,-1); \draw[very thick,red] (2.5,-1.25)--(3,-1.25); \footnotesize \draw[very thin] (2,-1.45) -- (2,-1.55)node[anchor=north ] {$k$}; \draw[very thin] (2.5,-1.45) -- (2.5,-1.55) \draw[very thin] (3,-1.55) -- (3,-1.45); \draw[very thin] (1,-1.45) -- (1,-1.55); \draw[very thin] (1.5,-1.45) -- (1.5,-1.55) \normalsize \node[anchor = west] at (1.73,-.25) {$\Updownarrow$ \scriptsize$\altmathcal{R}_k$}; \end{tikzpicture}\quad} \caption{Left: Points in the diagram of a zigzag module $\mathbb{V}$ move according to the arrows when the reflection functors $\altmathcal{L}_k$ or $\altmathcal{C}_k$ are applied to diagrams with sinks or sources, respectively, at index $k$. The point $(k,k)$ corresponding to the simple summand $\mathbb{I}([k,k])$ is killed. Right: An analogous picture for barcodes.} \label{reflection_diagram} \end{figure} \begin{example} Let $n = 4$, let $\tau = (\rightarrow,\leftarrow,\rightarrow)\in\altmathcal{T}_n$, and let $\mathbb{V}$ be the $\tau$-module \[\mathbb{V} = \mathbb{I}_\tau([1,4])\oplus\mathbb{I}_\tau([1,2])\oplus\mathbb{I}_\tau([2,3])\oplus\mathbb{I}_\tau([2,3])\oplus\mathbb{I}_\tau([3,3])\] Evidently, $\tau$ has a sink at index $2$. Applying $\altmathcal{L}_2$ and using Theorems \ref{bigsumreflection} and \ref{intervalreflections}, we have \[\altmathcal{L}_2(\mathbb{V}) \cong \mathbb{I}_{\sigma_2\tau}([1,4])\oplus\mathbb{I}_{\sigma_2\tau}([1,1])\oplus \mathbb{I}_{\sigma_2\tau}([3,3])^3.\] Similarly, we see that $\tau$ has a source at index $3$, and \[\altmathcal{C}_3(\mathbb{V}) \cong \mathbb{I}_{\zeta_3\tau}([1,4])\oplus\mathbb{I}_{\zeta_3\tau}([1,2])\oplus \mathbb{I}_{\zeta_3\tau}([2,2])^2.\] \end{example} The following theorem is the analogue of Theorem \ref{intervalreflections} for interval modules having a flow, i.e. neither a sink or a source, at the index to which a reflection functor is applied: \begin{theorem}[The Diamond Principle Part II]\label{flow_reflection} Let $k\in\{2,\dots,n-1\}$ and let $\tau\in\altmathcal{T}_n$ have a forward flow at index $k$. \begin{enumerate}[(1.)] \item The extroversion functor $\altmathcal{L}_k$ acts as follows on the interval $\tau$-modules: \begin{equation*} \begin{aligned} \mathbb{I}_\tau([k,k])\, &&\stackrel{\altmathcal{L}_k }{\longrightarrow} & \quad \scalebox{.7}[1]{$\mathbb{O}$} &\\ \mathbb{I}_\tau([b,k-1])\, &&\longrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([b,k]) && \quad \text{for }b\leq k-1\\ \mathbb{I}_\tau([k,d])\, &&\longrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([k+1,d]) && \quad \text{for }d\geq k+1\\ \mathbb{I}_\tau([b,d])\, &&\longrightarrow & \quad\mathbb{I}_{\sigma_k\tau}([b,d]) && \quad \text{otherwise}.\\ \end{aligned} \end{equation*} \item The introversion functor $\altmathcal{C}_k$ acts as follows on the interval $\tau$-modules: \begin{equation*} \begin{aligned} \mathbb{I}_\tau([k,k])\, &&\stackrel{\altmathcal{C}_k}{\longrightarrow} & \quad \scalebox{.7}[1]{$\mathbb{O}$} &\\ \mathbb{I}_\tau([b,k])\, &&\longrightarrow & \quad\mathbb{I}_{\xi_k\tau}([b,k-1]) && \quad \text{for }b\leq k-1\\ \mathbb{I}_\tau([k+1,d])\, &&\longrightarrow & \quad\mathbb{I}_{\xi_k\tau}([k,d]) && \quad \text{for }d\geq k+1\\ \mathbb{I}_\tau([b,d])\, &&\longrightarrow & \quad\mathbb{I}_{\xi_k\tau}([b,d]) && \quad \text{otherwise}.\\ \end{aligned} \end{equation*} \end{enumerate} \begin{proof} The proof is just a straightforward verification of the universal properties of limits and colimits. We omit the details. \end{proof} \end{theorem} The same computations can be made for reflections applied to the indices with backwards flows or to the indices $1$ or $n$, though these details are not so important for us. The upshot of Theorems \ref{intervalreflections} and \ref{flow_reflection} is that points in the diagram of zigzag module move at furthest horizontally or vertically to adjacent nodes on the integer lattice (see Figures \ref{reflection_diagram} and \ref{inex_diagram}). \begin{figure} \centering \scalebox{1.44}{ \begin{tikzpicture} \draw[-latex,very thick] (1,0.5)--(3.5,0.5);\draw[-latex,very thick] (0.5,1)--(0.5,3.5); \draw[dashed,very thick] (0.5,0.5)--(1,0.5); \draw[dashed,very thick] (0.5,0.5)--(0.5,1); \draw[very thick] (0.479,0.5)--(0.5,0.5); \fill[gray] (1.5,2) circle [radius=1.75pt]; \fill[gray] (1,2.5) circle [radius=1.75pt]; \fill[gray] (1.5,2) circle [radius=1.75pt];\fill[gray] (2,2) circle [radius=1.75pt]; \fill[gray] (1,1.5) circle [radius=1.75pt];\fill[gray] (1,2) circle [radius=1.75pt]; \fill[gray] (1.5,1.5) circle [radius=1.75pt];\fill[gray] (1,1) circle [radius=1.75pt]; \fill[gray] (2.5,3) circle [radius=1.75pt];\fill[gray] (2,3) circle [radius=1.75pt]; \fill[gray] (1.5,3) circle [radius=1.75pt];\fill[gray] (1,3) circle [radius=1.75pt]; \fill[gray] (2.5,2.5) circle [radius=1.75pt];\fill[gray] (2,2.5) circle [radius=1.75pt]; \fill[gray] (1.5,2.5) circle [radius=1.75pt];\fill[gray] (3,3) circle [radius=1.75pt]; \tikzset{cross/.style={cross out, draw=red, fill=none, minimum size=3, inner sep=0pt, outer sep=0pt}, cross/.default={2pt}} \node[draw,cross out,red,inner sep = 1pt,minimum size = 4.5pt] at (2,2){}; \draw[-latex,semithick] (1.5,1.5)--(1.5,2); \draw[-latex,semithick] (1.0,1.5)--(1.0,2); \draw[-latex,semithick] (2.0,2.5)--(2.5,2.5); \draw[-latex,semithick] (2.0,3.0)--(2.5,3.0); \footnotesize \draw[very thin] (2,0.55) -- (2,0.45)node[anchor=north ] {$k$}; \draw[very thin] (2.5,0.55) -- (2.5,0.45) \draw[very thin] (3,0.45) -- (3,0.55); \draw[very thin] (1,0.55) -- (1,0.45); \draw[very thin] (1.5,0.55) -- (1.5,0.45) \footnotesize \draw[very thin] (.55,2) -- (.45,2)node[anchor=east ] {$k$}; \draw[very thin] (.55,2.5) -- (.45,2.5) \draw[very thin] (.55,3) -- (0.45,3); \draw[very thin] (.55,1) -- (0.45,1); \draw[very thin] (.55,1.5) -- (.45,1.5) \node [anchor=east] at (2.75,1.25) {$\altmathcal{L}_k$}; \end{tikzpicture}\quad \begin{tikzpicture} \draw[-latex,very thick] (1,0.5)--(3.5,0.5);\draw[-latex,very thick] (0.5,1)--(0.5,3.5); \draw[dashed,very thick] (0.5,0.5)--(1,0.5); \draw[dashed,very thick] (0.5,0.5)--(0.5,1); \draw[very thick] (0.479,0.5)--(0.5,0.5); \fill[gray] (1.5,2) circle [radius=1.75pt]; \fill[gray] (1,2.5) circle [radius=1.75pt]; \fill[gray] (1.5,2) circle [radius=1.75pt];\fill[gray] (2,2) circle [radius=1.75pt]; \fill[gray] (1,1.5) circle [radius=1.75pt];\fill[gray] (1,2) circle [radius=1.75pt]; \fill[gray] (1.5,1.5) circle [radius=1.75pt];\fill[gray] (1,1) circle [radius=1.75pt]; \fill[gray] (2.5,3) circle [radius=1.75pt];\fill[gray] (2,3) circle [radius=1.75pt]; \fill[gray] (1.5,3) circle [radius=1.75pt];\fill[gray] (1,3) circle [radius=1.75pt]; \fill[gray] (2.5,2.5) circle [radius=1.75pt];\fill[gray] (2,2.5) circle [radius=1.75pt]; \fill[gray] (1.5,2.5) circle [radius=1.75pt];\fill[gray] (3,3) circle [radius=1.75pt]; \tikzset{cross/.style={cross out, draw=red, fill=none, minimum size=3, inner sep=0pt, outer sep=0pt}, cross/.default={2pt}} \node[draw,cross out,red,inner sep = 1pt,minimum size = 4.5pt] at (2,2){}; \draw[-latex,semithick] (1.5,2.0)--(1.5,1.5); \draw[-latex,semithick] (1.0,2.0)--(1.0,1.5); \draw[-latex,semithick] (2.5,2.5)--(2,2.5); \draw[-latex,semithick] (2.5,3.0)--(2,3.0); \footnotesize \draw[very thin] (2,0.55) -- (2,0.45)node[anchor=north ] {$k$}; \draw[very thin] (2.5,0.55) -- (2.5,0.45) \draw[very thin] (3,0.45) -- (3,0.55); \draw[very thin] (1,0.55) -- (1,0.45); \draw[very thin] (1.5,0.55) -- (1.5,0.45) \footnotesize \draw[very thin] (.55,2) -- (.45,2)node[anchor=east ] {$k$}; \draw[very thin] (.55,2.5) -- (.45,2.5) \draw[very thin] (.55,3) -- (0.45,3); \draw[very thin] (.55,1) -- (0.45,1); \draw[very thin] (.55,1.5) -- (.45,1.5) \node [anchor=east] at (2.75,1.25) {$\altmathcal{C}_k$}; \end{tikzpicture}\quad} \caption{Points in the persistence diagram of a zigzag module $\mathds{V}$ with a \textit{forward flow} at index $k$ move according to the arrows when the extroversion functor $\altmathcal{L}_k$ (left) or introversion functor $\altmathcal{C}_k$ (right) are applied. The point $(k,k)$ corresponding to the simple summand $\mathbb{I}([k,k])$ is annihilated by both of these functors.} \label{inex_diagram} \end{figure} \subsection{Removing Simple Summands}\label{sec:sanitation} In the next section we define the reflection distance. In order to avoid having the distance between a simple summand and the zero zigzag module be nonzero, we introduce a map sending a zigzag module $\mathbb{V}$ to a zigzag module $\mathbb{V}'$ obtained from $\mathbb{V}$ by removing simple summands. We define the map $\altmathcal{S}:n\textup{-Mod}\ton\textup{-Mod}$ as follows: given a zigzag module $\mathbb{V}$ with interval decomposition $\mathbb{V} = \bigoplus_{[b,d]\in \mathrm{Dgm}(\mathbb{V})}\mathbb{I}([b,d])$, we define \[\altmathcal{S}(\mathbb{V}):= \bigoplus_{\substack{[b,d]\in \mathrm{Dgm}(\mathbb{V})\\b\neq d}}\mathbb{I}([b,d]).\] That is, $\altmathcal{S}(\mathbb{V})$ is obtained from $\mathbb{V}$ by removing simple summands. If $\mathbb{V}$ is a direct sum of simple modules, then we set $\altmathcal{S}(\mathbb{V}) := \scalebox{.7}[1]{$\mathbb{O}$}$. We also set $\altmathcal{S}(\scalebox{.7}[1]{$\mathbb{O}$}) := \scalebox{.7}[1]{$\mathbb{O}$}$. By definition, $\altmathcal{S}$ sends zigzag modules of type $\tau$ to zigzag modules of type $\tau$ so that $\altmathcal{S}$ restricts to a map $\altmathcal{S}:\taumod{\tau}\to\taumod{\tau}$ for each type $\tau$. \begin{remarks}\label{sanitize_alternate_def} \begin{enumerate}[(1)] \item If $\,\mathbb{V} = \mathbb{V}_1\oplus\mathbb{V}_2$ where $\mathbb{V}_1$ is a direct sum of simple modules and $\mathbb{V}_2$ has no simple summands, then $\altmathcal{S}(\mathbb{V}) = \mathbb{V}_2$. \item By definition, $\altmathcal{S}(\mathbb{V})\preceq \mathbb{V}$ and hence $\altmathcal{S}(\mathbb{V})\precsim\mathbb{V}$ for any zigzag module $\mathbb{V}$. \item $\altmathcal{S}(\altmathcal{S}(\mathbb{V})) = \altmathcal{S}(\mathbb{V})$ for all zigzag modules $\mathbb{V}$. \item $\mathrm{Dgm}(\altmathcal{S}(\mathbb{V})) = \{\!\!\{(b,d)\in \mathrm{Dgm}(\mathbb{V}) \ | \ b\neq d\}\!\!\}$. That is, the diagram of $\altmathcal{S}(\mathbb{V})$ is that of $\,\mathbb{V}$ minus the points on the diagonal. \end{enumerate} \end{remarks} We also have the following analogues of Corollaries \ref{refsummand} and \ref{refprecsim}: \begin{proposition}\label{sanitize_summand} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V}$, $\mathbb{W}\in\taumod{\tau}$. If $\,\mathbb{W}\preceq \mathbb{V}$ then $\altmathcal{S}(\mathbb{W})\preceq\altmathcal{S}(\mathbb{V})$. \begin{proof}If $\mathbb{W}\preceq\mathbb{V}$ then $\mathbb{V} \cong \mathbb{W}\oplus\mathbb{U}$ for some $\mathbb{U}\in\taumod{\tau}$. Write \[\mathbb{W}= \mathbb{W}_1\oplus \mathbb{W}_2 \quad \textup{ and } \quad \mathbb{U} = \mathbb{U}_1\oplus\mathbb{U}_2,\] where $\mathbb{W}_1,\mathbb{U}_1$ are direct sums of simple modules and $\mathbb{W}_2,\mathbb{U}_2$ have no simple summands. By Remark \ref{sanitize_alternate_def}, $\altmathcal{S}(\mathbb{W}) = \mathbb{W}_2$ and $\altmathcal{S}(\mathbb{U}) = \mathbb{U}_2$. Moreover $\mathbb{V} \cong (\mathbb{W}_1\oplus\mathbb{U}_1)\oplus(\mathbb{W}_2\oplus\mathbb{U}_2)$, with $\mathbb{W}_1\oplus\mathbb{U}_1$ being a direct sum of simple modules and $\mathbb{W}_2\oplus\mathbb{U}_2$ having no simple summands. Then again by Remark \ref{sanitize_alternate_def}, \[ \altmathcal{S}(\mathbb{V}) \cong \mathbb{W}_2\oplus\mathbb{U}_2 = \altmathcal{S}(\mathbb{W})\oplus \altmathcal{S}(\mathbb{U}),\] and hence $\altmathcal{S}(\mathbb{W})\preceq \altmathcal{S}(\mathbb{V})$. \end{proof} \end{proposition} \begin{proposition}\label{sanitize_precsim} Fix $k\in\{1,\hdots,n\}$, let $\tau\in \altmathcal{T}_n$, and let $\mathbb{V}$, $\mathbb{W}\inn\textup{-Mod}$. If $\,\mathbb{W}\precsim \mathbb{V}$ then $\altmathcal{S}(\mathbb{W})\precsim\altmathcal{S}(\mathbb{V})$. \begin{proof} If $\mathbb{W}\precsim \mathbb{V}$ then $\mathbb{W}\sim\mathbb{W}'\preceq \mathbb{V}$ for some $\mathbb{W}'$. Now note that that if $\mathbb{W}\sim \mathbb{W}'$ then $\altmathcal{S}(\mathbb{W})\sim \altmathcal{S}(\mathbb{W}')$. This is because if an arrow representing an isomorphism is reversed, then the source and target of this arrow cannot be an index at which a simple summand appears. From this fact and Corollary \ref{sanitize_summand}, we have \[\altmathcal{S}(\mathbb{W})\sim\altmathcal{S}(\mathbb{W}')\preceq \altmathcal{S}(\mathbb{V}),\] and hence $\altmathcal{S}(\mathbb{W})\precsim\altmathcal{S}(\mathbb{V})$. \end{proof} \end{proposition} \begin{definition} Define an equivalence relation $\sim_\altmathcal{S}$ on $n\textup{-Mod}$ by \[ \mathbb{V}\sim_\altmathcal{S}\mathbb{W} \quad \iff \quad \altmathcal{S}(\mathbb{V})\precsim\mathbb{W} \quad \textup{ and } \quad \altmathcal{S}(\mathbb{W})\precsim\mathbb{V}.\] \end{definition} \noindent Reflexivity of $\sim_{\altmathcal{S}}$ follows from the fact that $\altmathcal{S}(\mathbb{V})\precsim \mathbb{V}$ and symmetry is immediate from the definition. To see that $\sim_{\altmathcal{S}}$ is transitive, suppose that $\mathbb{V}\sim_\altmathcal{S}\mathbb{W}$ and $\mathbb{W} \sim_{\altmathcal{S}}\mathbb{U}$. Then $$\altmathcal{S}(\mathbb{V})\precsim \mathbb{W},\quad \altmathcal{S}(\mathbb{W})\precsim \mathbb{V},\quad \altmathcal{S}(\mathbb{W})\precsim \mathbb{U}, \quad \textup{ and }\quad \altmathcal{S}(\mathbb{U})\precsim \mathbb{W}.$$ Using Proposition \ref{sanitize_precsim} together with the fact that $\altmathcal{S}(\altmathcal{S}(\mathbb{V})) = \altmathcal{S}(\mathbb{V})$ for all zigzag modules $\mathbb{V}$, we have \[\altmathcal{S}(\mathbb{V}) = \altmathcal{S}(\altmathcal{S}(\mathbb{V}))\precsim \altmathcal{S}(\mathbb{W})\precsim\mathbb{U} \quad \textup{ and } \quad \altmathcal{S}(\mathbb{U}) = \altmathcal{S}(\altmathcal{S}(\mathbb{U}))\precsim \altmathcal{S}(\mathbb{W})\precsim\mathbb{V}. \] By transitivity of the preorder $\precsim$, we have $\altmathcal{S}(\mathbb{V})\precsim \mathbb{U}$ and $\altmathcal{S}(\mathbb{U})\precsim \mathbb{V}$ so that $\mathbb{V} \sim_{\altmathcal{S}}\mathbb{U}$. Thus $\sim_{\altmathcal{S}}$ is indeed an equivalence relation on $n\textup{-Mod}$. \section{The Reflection Distance}\label{sec:ref-distance} Let $\mathfrak{R} = (\altmathcal{R}_{k_\ell},\altmathcal{R}_{k_{\ell-1}},\cdots,\altmathcal{R}_{k_1})$ be a sequence of $\ell$ reflection functors with $1\leq k_i\leq n$ for all $i$. For $\mathbb{V}\inn\textup{-Mod}$, we denote by $\mathfrak{R}(\mathbb{V})$ the zigzag module \[(\altmathcal{S}\circ\altmathcal{R}_{k_\ell}\circ\altmathcal{S}\circ\altmathcal{R}_{k_{\ell-1}}\circ\cdots \circ\altmathcal{S}\circ\altmathcal{R}_{k_1}\circ\altmathcal{S})(\mathbb{V}),\] where $\altmathcal{S}$ is the map defined in Section \ref{sec:sanitation}. That is, $\mathfrak{R}(\mathbb{V})$ is obtained from $\mathbb{V}$ by first removing simple summands from $\mathbb{V}$, applying the reflection $\altmathcal{R}_{k_1}$, removing simple summands from the resulting zigzag module, and so on. We also set $\varepsilon(\mathbb{V}) :=\altmathcal{S}(\mathbb{V})$, where $\varepsilon$ denotes the empty sequence. For each $p\in[1,\infty)$ define the \textit{$p$-cost} of a nonempty sequence $\mathfrak{R}$ of length $\ell$ by \[C_p(\mathfrak{R}):= \ell^{1/p}\] and define $C_p(\varepsilon) = 0$. \begin{definition}\label{reflection_distance} For each $p\in[1,\infty)$, we define a function $d_{\altmathcal{R}}^p:n\textup{-Mod}\timesn\textup{-Mod}\to \mathbb{R}$ by setting \[ d_{\altmathcal{R}}^p(\mathbb{V},\mathbb{W}):= \min_{(\mathfrak{R},\mathfrak{R}')}\big\{\max\{C_p(\mathfrak{R}),C_p(\mathfrak{R}')\} \ | \ \mathfrak{R}(\mathbb{V})\precsim \mathbb{W} \text{ and } \mathfrak{R}'(\mathbb{W})\precsim \mathbb{V}\big\},\] where the minimum is taken over all pairs $(\mathfrak{R},\mathfrak{R}')$ of sequences of reflection functors. \end{definition} Given two sequences $\mathfrak{R} = (\altmathcal{R}_{k_\ell},\altmathcal{R}_{k_{\ell-1}},\dots,\altmathcal{R}_{k_1})$, $\mathfrak{R}' = (\altmathcal{R}_{m_j},\altmathcal{R}_{m_{j-1}},\dots,\altmathcal{R}_{m_1})$ of reflection functors, denote by $\mathfrak{R}'\circ \mathfrak{R}$ the concatenation $(\altmathcal{R}_{m_j},\dots,\altmathcal{R}_{m_1},\altmathcal{R}_{k_\ell},\dots,\altmathcal{R}_{k_1})$. \begin{fact}\label{costadd} For any $p\in [1,\infty)$ we have $C_p(\mathfrak{R}' \circ \mathfrak{R})\leq C_p(\mathfrak{R}) + C_p(\mathfrak{R}')$. \begin{proof} This follows from the inequality $(a + b)^{1/p}\leq a^{1/p} + b^{1/p}$ for all nonnegative integers $a,b$. \end{proof} \end{fact} We will say that a sequence $\mathfrak{R}$ of reflection functors \textit{annihilates} a summand $\mathbb{W}$ of $\mathbb{V}$ if $\mathfrak{R}(\mathbb{W}) = \scalebox{.7}[1]{$\mathbb{O}$}$. Note that if $\mathbb{V} = \mathbb{W}\oplus \mathbb{U}$ and $\mathfrak{R}$ annihilates $\mathbb{W}$ then $\mathfrak{R}(\mathbb{V}) \cong \mathfrak{R}(\mathbb{U})$. \begin{proposition} For any $\mathbb{V}\inn\textup{-Mod}$ there exists a sequence of reflection functors $\mathfrak{R}$ which annihilates $\mathbb{V}$, i.e., $\mathfrak{R}(\mathbb{V}) = \scalebox{.7}[1]{$\mathbb{O}$}$. \begin{proof} Let $\tau = \textup{type}(\mathbb{V})$. It suffices to show that for any interval module $\mathbb{I}_\tau([b,d])$, there is a composition of reflections $\mathfrak{R}$ such that $\mathfrak{R}(\mathbb{I}_\tau([b,d])) = \scalebox{.7}[1]{$\mathbb{O}$}$. For if this is the case then we can iteratively annihilate interval summands of $\mathbb{V}$ until we arrive at $\scalebox{.7}[1]{$\mathbb{O}$}$. We demonstrate that this can be done as follows: the interval $\tau$-module $\mathbb{I}_\tau([b,d])$ has either a sink, source, or flow at index $d$. In any case, Theorems \ref{intervalreflections} and \ref{flow_reflection} guarantee that we can choose a reflection functor $\altmathcal{R}_d$ at index $d$ such that $\altmathcal{R}_d(\mathbb{I}([b,d]))$ is an interval module supported over $[b,d-1]$. In this way, we obtain a sequence $\mathfrak{R}_1 = (\altmathcal{R}_{b},\altmathcal{R}_{b+1},\cdots,\altmathcal{R}_{d-1},\altmathcal{R}_{d})$ such that $\mathfrak{R}_1(\mathbb{I}_\tau([b,d]))= \scalebox{.7}[1]{$\mathbb{O}$}$. Applying this procedure again to an interval summand of $\mathfrak{R}^1(\mathbb{V})$ yields a sequence $\mathfrak{R}_2$ of reflection functors which annihilates this summand. Since $\mathbb{V}\in n\textup{-Mod}$ has only finitely many summands, and since applying a reflection functor to a zigzag module can only reduce the number of summands, we obtain a sequence $(\mathfrak{R}_1,\dots,\mathfrak{R}_m)$ of sequences of reflection functors, where $m \leq|\mathrm{Dgm}(\mathbb{V})|$, for which the concatenation $\mathfrak{R} = \mathfrak{R}_m\circ\cdots\circ\mathfrak{R}_1$ satisfies $\mathfrak{R}(\mathbb{V}) = \scalebox{.7}[1]{$\mathbb{O}$}$. \end{proof} \end{proposition} \begin{remark}\label{finiteness_remark} By Theorems \ref{intervalreflections} and \ref{flow_reflection}, it is clear that it takes exactly $d-b$ reflections to annihilate an interval module supported over $[b,d]$. It follows that $d_{\altmathcal{R}}^p(\mathbb{I}([b,d]),\scalebox{.7}[1]{$\mathbb{O}$}) = (d-b)^{1/p}$. Moreover, if $|\mathrm{Dgm}(\mathbb{V})| = K$ for $\mathbb{V}\in n\textup{-Mod}$ then $d_{\altmathcal{R}}^p(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$}) \leq Kn$. This follows since we can iteratively annihilate each of the $K$ interval summands of $\,\mathbb{V}$ with at most $n$ reflections. \end{remark} \begin{theorem}\label{thm:ref_is_metric} For each $n\in \mathbb{N}$ and for all $p\in[1,\infty)$, the function $d_{\altmathcal{R}}^p$ is a pseudometric on $n\textup{-Mod}$. Moreover, $d_{\altmathcal{R}}^p(\mathbb{V},\mathbb{W}) = 0$ if and only if $\,\mathbb{V}\sim_{\altmathcal{S}}\mathbb{W}$ so that $d_{\altmathcal{R}}^p$ induces a metric on $n\textup{-Mod}/{\sim_\altmathcal{S}}$. \begin{proof} It follows from Remark \ref{finiteness_remark} that $d_\altmathcal{R}^p(\mathbb{V},\mathbb{W})$ is finite for all $\mathbb{V},\mathbb{W}\in n\textup{-Mod}$. The facts that $d_{\altmathcal{R}}^p$ is a non-negative and symmetric function follow immediately. Since $\epsilon(\mathbb{V}) = \altmathcal{S}(\mathbb{V})\precsim\mathbb{V}$ we have $d_{\altmathcal{R}}^p(\mathbb{V},\mathbb{V}) = 0$ for all $\mathbb{V}\in n\textup{-Mod}$. Next, we verify the triangle inequality. Fix $p\in [1,\infty)$ and let $\mathbb{V}_1$, $\mathbb{V}_2$, $\mathbb{V}_3\inn\textup{-Mod}$. Then there exist sequences of reflections $\mathfrak{R}_1, \mathfrak{R}_2$ such that \[\mathfrak{R}_1(\mathbb{V}_1)\precsim\mathbb{V}_2,\quad \mathfrak{R}_2(\mathbb{V}_2)\precsim\mathbb{V}_1 \quad \textup{ with } \quad C_p(\mathfrak{R}_1)\leq d_{\altmathcal{R}}^p(\mathbb{V}_1,\mathbb{V}_2), \quad C_p(\mathfrak{R}_2)\leq d_{\altmathcal{R}}^p(\mathbb{V}_1,\mathbb{V}_2). \] Similarly, there exist sequences of reflections $\mathfrak{R}_1', \mathfrak{R}_2'$ such that \[\mathfrak{R}_1'(\mathbb{V}_2)\precsim\mathbb{V}_3, \quad \mathfrak{R}_2'(\mathbb{V}_3)\precsim\mathbb{V}_2\quad \textup{ with } \quad C_p(\mathfrak{R}_1')\leq d_{\altmathcal{R}}^p(\mathbb{V}_2,\mathbb{V}_3), \quad C_p(\mathfrak{R}_2')\leq d_{\altmathcal{R}}^p(\mathbb{V}_2,\mathbb{V}_3).\] Then by Corollary \ref{refprecsim} and Proposition \ref{sanitize_precsim}, $\mathfrak{R}_1'\circ\mathfrak{R}_1(\mathbb{V}_1)\precsim\mathfrak{R}_1'(\mathbb{V}_2)\precsim \mathbb{V}_3$ and $\mathfrak{R}_2\circ\mathfrak{R}_2'(\mathbb{V}_3)\precsim\mathfrak{R}_2(\mathbb{V}_2)\precsim\mathbb{V}_1$. By transitivity of the preorder $\precsim$, we have \begin{equation}\label{composite_ref}\mathfrak{R}_1'\circ\mathfrak{R}_1(\mathbb{V}_1)\precsim \mathbb{V}_3\quad \textup{ and }\quad \mathfrak{R}_2\circ\mathfrak{R}_2'(\mathbb{V}_3)\precsim\mathbb{V}_1.\end{equation} By Fact \ref{costadd}, $C_p(\mathfrak{R}_1'\circ\mathfrak{R}_1)\leq C_p(\mathfrak{R}_1')+C_p(\mathfrak{R}_1)$ and $C_p(\mathfrak{R}_2\circ\mathfrak{R}_2')\leq C_p(\mathfrak{R}_2) + C_p(\mathfrak{R}_2')$ so that \begin{equation} \label{composite_cost}C_p(\mathfrak{R}_1'\circ\mathfrak{R}_1)\leq d_{\altmathcal{R}}^p(\mathbb{V}_1,\mathbb{V}_2) + d_{\altmathcal{R}}^p(\mathbb{V}_2,\mathbb{V}_3)\quad \textup{ and }\quad C_p(\mathfrak{R}_2\circ\mathfrak{R}_2')\leq d_{\altmathcal{R}}^p(\mathbb{V}_1,\mathbb{V}_2) + d_{\altmathcal{R}}^p(\mathbb{V}_2,\mathbb{V}_3).\end{equation} Equations (\ref{composite_ref}) and (\ref{composite_cost}) together imply \[d_{\altmathcal{R}}^p(\mathbb{V}_1,\mathbb{V}_3)\leq d_{\altmathcal{R}}^p(\mathbb{V}_1,\mathbb{V}_2) + d_{\altmathcal{R}}^p(\mathbb{V}_2,\mathbb{V}_3).\] Now if $\mathbb{V}\sim_{\altmathcal{S}}\mathbb{W}$ then $\epsilon(\mathbb{V}) = \altmathcal{S}(\mathbb{V})\precsim\mathbb{W}$ and $\epsilon(\mathbb{W}) = \altmathcal{S}(\mathbb{W})\precsim \mathbb{V}$ so that $d_{\altmathcal{R}}^p(\mathbb{V},\mathbb{W}) = 0$. Conversely, if $\mathbb{V} \not\sim_\altmathcal{S} \mathbb{W}$ then either $\altmathcal{S}(\mathbb{V})\not\precsim\mathbb{W}$ or $\altmathcal{S}(\mathbb{W})\not\precsim\mathbb{V}$. In any case, we must apply some non-trivial reflection to either $\mathbb{V}$ or $\mathbb{W}$, incurring a nonzero cost so that $d_{\altmathcal{R}}^p(\mathbb{V},\mathbb{W}) > 0$. It follows that if $\mathbb{V} \sim_\altmathcal{S}\mathbb{V}'$ and $\mathbb{W}\sim_\altmathcal{S} \mathbb{W}'$ then $d_{\altmathcal{R}}^p(\mathbb{V},\mathbb{W}) = d_{\altmathcal{R}}^p(\mathbb{V}',\mathbb{W}')$ so that $d_{\altmathcal{R}}^p$ induces a well-defined metric on the equivalence classes of $n\textup{-Mod}$ under $\sim_\altmathcal{S}$. \end{proof} \end{theorem} \begin{corollary}\label{corollary:distance_is_zero} If $\,\mathbb{V}\in n\textup{-Mod}$ is a direct sum of simple modules then $d_{\altmathcal{R}}^p(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$}) = 0$. \begin{proof} We have $\altmathcal{S}(\mathbb{V}) = \scalebox{.7}[1]{$\mathbb{O}$} \precsim \scalebox{.7}[1]{$\mathbb{O}$}$ and $\altmathcal{S}(\scalebox{.7}[1]{$\mathbb{O}$}) = \scalebox{.7}[1]{$\mathbb{O}$}\precsim \mathbb{V}$ so that $\mathbb{V}\sim_{\altmathcal{S}}\scalebox{.7}[1]{$\mathbb{O}$}$ and hence $d_{\altmathcal{R}}^p(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$}) = 0$ by the second statement of Theorem \ref{thm:ref_is_metric}. \end{proof} \end{corollary} \section{The $\ell^p$-Bottleneck Distance and the Proof of the Main Theorem}\label{sec:d-bottleneck} In this section we define a family of metrics on the space of persistence diagrams which we call the $\ell^p$-bottleneck distances. We show that the map which takes a zigzag module to its persistence diagram is 1-Lipschitz with respect to the 1-reflection distance and the $\ell^1$-bottleneck distance. \subsection{Matchings} Let $S$ and $T$ be sets. A \textit{matching} between $S$ and $T$ is a relation $M\subseteq S\times T$ such that \begin{enumerate}[(1)] \item For any $s\in S$, there is at most one $t\in T$ such that $(s,t)\in M$, \item For any $t\in T$, there is a most one $s\in S$ such that $(s,t)\in M$. \end{enumerate} We denote a matching by $M:S\nrightarrow T$. Equivalently, a matching is a bijection $M:S'\to T'$ for some subsets $S'\subseteq S$ and $T' \subseteq T$. From this point of view, $S'$ is called the \textit{coimage} of $M$, denoted $S' = \textup{coim} (M)$, and $T'$ is called the \textit{image} of $M$, denoted $T' = \textup{im} (M)$. A matching $M$ is said to be \textit{finite} if $|\textup{coim} (M)| = |\textup{im} (M)|$ is finite. If $(s,t)\in M$ then $s$ and $t$ are said to be \textit{matched}. Points in $S\sqcup T$ which are not matched are said to be \textit{unmatched}. The following classical theorem from matching theory is crucial to our proof of stability: \begin{lemma}[Matching Lemma, \cite{ore1962theory}]\label{matching_lemma} Let $S$ and $T$ be sets and let $f:S\nrightarrow T$ and $g:T\nrightarrow S$ be matchings. Then there exists a matching $M:S\nrightarrow T$ such that \begin{enumerate}[label=(\arabic*)] \item $\textup{coim} (f)\subseteq\textup{coim} (M)$, \item $ \textup{coim} (g)\subseteq\textup{im} (M)$, \item if $M(s) = t$ then either $f(s) = t$ or $g(t) = s$. \end{enumerate} \end{lemma} \noindent The Matching Lemma holds for arbitrary matchings but, for our purposes, we will only use the Matching Lemma in the case of finite matchings. A proof of the Matching Lemma \ref{matching_lemma} for the case of finite matchings is provided in Appendix \ref{app:matchings}. \subsection{The $\ell^p$-Bottleneck Distance} Let $S$ and $T$ be multisets of points from $\mathbb{R}^2$ and let $M:S\nrightarrow T$ be a matching. For each $p\in [1,\infty]$ define \begin{equation}\label{matching_cost} c_p(M) := \max\left\{\max_{(s,t)\in M}\|s-t\|_{p},\max_{r\in S\sqcup T \textup{ unmatched}}\frac{|r_y-r_x|}{2^{1-1/p}}\right\},\end{equation} where $r_x$ and $r_y$ denote the $x$ and $y$ coordinates, respectively, of the point $r\in \mathbb{R}^2$, and where $\|\cdot\|_p:\mathbb{R}^2\to \mathbb{R}$ denotes the usual $\ell^p$-norm on $\mathbb{R}^2$. Here we use the convention $1/\infty = 0$. We then define the \textit{$\ell^p$-bottleneck distance} between $S$ and $T$ by \[d_b^p(S,T) : = \inf_{M:S\nrightarrow T}c_p(M),\] where the infimum is taken over all matchings between $S$ and $T$. \begin{remark}\label{remark:db1_bottleneck_equiv} The bottleneck distance (as defined in \cite{bauer2014induced}) is just a special case of the $\ell^p$-bottleneck distance when $p = \infty$. It is simply the limit of $d_b^p$ as $p\to \infty$. Since $\|x\|_\infty\leq\|x\|_1\leq 2\|x\|_\infty$ for all $x\in \mathbb{R}^2$, we see that $d_b^\infty(S,T)\leq d_b^1(S,T)\leq 2d_b^\infty(S,T)$ for all multisets $S$ and $T$. \end{remark} Given a multisubset $S$ of points in $\mathbb{R}^2$, and given $p\in [1,\infty]$ and $\eta>0$, we define \[S^\eta_p := \left\{\!\!\!\left\{ s\in S \ \Big| \ \frac{|s_y-s_x|}{2^{1-1/p}} >\eta \right\}\!\!\!\right\}.\] \begin{lemma}\label{matching_lemma_two} Fix $p\in[1,\infty]$, let $S$ and $T$ be multisubsets of points in $\mathbb{R}^2$, and let $M:S\nrightarrow T$ be a matching such that \begin{enumerate}[(1)] \item $S^\eta_p \subseteq \textup{coim} (M)$, \item $T^\eta_p \subseteq \textup{im} (M)$, \item if $M(s) = t$ then $\|s-t\|_p\leq \eta$. \end{enumerate} Then $c_p(M)\leq \eta$. \begin{proof} Condition (3) guarantees that $\max_{(s,t)\in M}\|s-t\|_p\leq \eta$. If $r\in S\sqcup T$ is unmatched then $r\not\in S^\eta_p\sqcup T^\eta_p$ so that $\frac{|r_y-r_x|}{2^{1-1/p}}\leq \eta$. That $c_p(M)\leq \eta$ now follows from equation (\ref{matching_cost}). \end{proof} \end{lemma} \subsection{Stability of Persistence Diagrams with Respect to $d_\altmathcal{R}^1$} In this section, we show that the $\ell^1$-bottleneck distance between the persistence diagrams of two given zigzag modules is bounded above by the 1-reflection distance between the zigzag modules themselves. \begin{lemma}\label{cost_to_kill} Let $\tau\in\altmathcal{T}_n$ and $\mathbb{V}\in\taumod{\tau}$. If $\,\mathfrak{R}(\mathbb{V}) = \scalebox{.7}[1]{$\mathbb{O}$}$ then \[C_1(\mathfrak{R})\geq \displaystyle\max_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}(d-b).\] Moreover, if $\,\mathfrak{R}(\mathbb{I}([b,d]))=\mathbb{I}([b',d'])$ then \[C_1(\mathfrak{R})\geq |d'-d| + |b'-b|.\] \end{lemma} \begin{proof} By Theorem \ref{bigsumreflection} we have \[\scalebox{.7}[1]{$\mathbb{O}$} = \mathfrak{R}(\mathbb{V}) = \mathfrak{R}\left(\bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\mathbb{I}_{\tau}([b,d])\right) \cong \bigoplus_{(b,d)\in\mathrm{Dgm}(\mathbb{V})}\mathfrak{R}(\mathbb{I}_{\tau}([b,d]))\] so that $\mathfrak{R}(\mathbb{I}_{\tau}([b,d])) = \scalebox{.7}[1]{$\mathbb{O}$}$ for all $(b,d)\in\mathrm{Dgm}(\mathbb{V})$. Hence $C_1(\mathfrak{R})\geq d-b$ for all $(b,d)\in \mathrm{Dgm}(\mathbb{V})$, proving the first claim. To prove the second claim, suppose that $\mathfrak{R}$ is a sequence of $m$ reflections so that $C_1(\mathfrak{R}) = m$. By Theorems \ref{intervalreflections} and \ref{flow_reflection}, any reflection moves points in corresponding diagram a distance of at most $1$ in the $\ell^1$ norm. Thus if $\mathfrak{R}(\mathbb{I}([b,d]))=\mathbb{I}([b',d'])$ then $(b',d')$ lives in the $m$-ball of $(b,d)$ with respect to the $\ell^1$ norm. Hence \[ \|(b,d)-(b',d')\|_1 = |d'-d| + |b'-b| \leq m = C_1(\mathfrak{R}).\qedhere\] \end{proof} We can now prove our main theorem. \thmmain* \begin{proof} We show that for any pair of sequences of reflections $(\mathfrak{R}^1,\mathfrak{R}^2)$ with $\mathfrak{R}^1(\mathbb{V})\precsim \mathbb{W}$ and $\mathfrak{R}^2(\mathbb{W})\precsim \mathbb{V}$, there is a matching $M:\mathrm{Dgm}(\mathbb{W})\nrightarrow \mathrm{Dgm}(\mathbb{V})$ satisfying $c_1(M) \leq \max\{C_1(\mathfrak{R}^1),C_1(\mathfrak{R}^2)\}$. Let $\eta = \max\{C_1(\mathfrak{R}^1),C_1(\mathfrak{R}^2)\}$. Consider the multisubsets \[\altmathcal{V}_1^\eta = \{\!\!\{(b,d)\in\mathrm{Dgm}(\mathbb{V}) \ | \ |d-b|>\eta \}\!\!\},\] \[\altmathcal{W}_1^\eta = \{\!\!\{(b',d')\in\mathrm{Dgm}(\mathbb{W}) \ | \ |d'-b'| > \eta\}\!\!\}\] and \[\altmathcal{I} = \{\!\!\{(b,d)\in \mathrm{Dgm}(\mathbb{V}) \ | \ \mathfrak{R}^1(\mathbb{I}([b,d])) \neq \scalebox{.7}[1]{$\mathbb{O}$} \}\!\!\},\] \[ \altmathcal{J} = \{\!\!\{(b',d')\in \mathrm{Dgm}(\mathbb{W}) \ | \ \mathfrak{R}^2(\mathbb{I}([b',d'])) \neq \scalebox{.7}[1]{$\mathbb{O}$} \}\!\!\}.\] Note that by Lemma \ref{cost_to_kill}, $\altmathcal{V}_1^\eta\subseteq \altmathcal{I}$ and $\altmathcal{W}_1^\eta \subseteq \altmathcal{J}$. Let $\alpha_1:\altmathcal{I}\to \mathrm{Dgm}(\mathfrak{R}^1(\mathbb{V}))$ and $\alpha_2:\altmathcal{J} \to \mathrm{Dgm}(\mathfrak{R}^2(\mathbb{W}))$ be the injections given by $\alpha_i(b,d) = (b',d')$ if and only if $\mathfrak{R}^i(\mathbb{I}[b,d]) = \mathbb{I}[b',d']$ for each $i\in\{1,2\}$. By Theorem \ref{perssum} and Proposition \ref{simpers}, we have $\mathrm{Dgm}(\mathfrak{R}^1(\mathbb{V}))\subseteq \mathrm{Dgm}(\mathbb{W})$ and $\mathrm{Dgm}(\mathfrak{R}^2(\mathbb{W}))\subseteq\mathrm{Dgm}(\mathbb{V})$. Let $j_1$ and $j_2$ denote the respective inclusion maps. Then $M_1 := j_1\circ\alpha_1 :\altmathcal{I}\to \mathrm{Dgm}(\mathbb{W})$ and $M_2 := j_2\circ\alpha_2 : \altmathcal{J}\to \mathrm{Dgm}(\mathbb{V})$ are injective and hence can be viewed as matchings $M_1:\mathrm{Dgm}(\mathbb{V})\nrightarrow\mathrm{Dgm}(\mathbb{W})$ and $M_2:\mathrm{Dgm}(\mathbb{W})\nrightarrow\mathrm{Dgm}(\mathbb{V})$. Note that $\altmathcal{V}_1^\eta\subseteq \textup{coim} (M_1)$ and $\altmathcal{W}_1^\eta\subseteq \textup{coim} (M_2)$. Let $M:\mathrm{Dgm}(\mathbb{V})\nrightarrow\mathrm{Dgm}(\mathbb{W})$ be the matching constructed from $M_1$ and $M_2$ provided by the Matching Lemma \ref{matching_lemma}. This matching has the following properties: \begin{enumerate}[(1)] \item $\textup{coim} (M_1)\subseteq \textup{coim} (M)$, \item $\textup{coim} (M_2)\subseteq \textup{im} (M)$, \item if $M(b,d) = (b',d')$ then either $M_1(b,d) = (b',d')$ or $M_2(b,d) = (b',d')$. \end{enumerate} In particular, we have $\altmathcal{V}_1^\eta\subseteq \textup{coim} (M)$ and $\altmathcal{W}_1^\eta\subseteq \textup{im} (M)$. Also, if $M(b,d) = (b',d')$ then either $\mathfrak{R}^1(\mathbb{I}([b,d])) = \mathbb{I}([b',d'])$ or $\mathfrak{R}^2(\mathbb{I}([b,d])) = \mathbb{I}([b',d'])$. By Lemma \ref{cost_to_kill}, $\|(b,d)-(b',d')\|_1\leq \max\{C_1(\mathfrak{R}^1),C_2(\mathfrak{R}^2)\} = \eta$. Hence by Lemma \ref{matching_lemma_two}, $c_1(M)\leq \eta$, completing the proof. \end{proof} Since $d_b^\infty\leq d_b^1$, we have the following \begin{corollary}\label{bottleneckbound} For any $\mathbb{V},\mathbb{W}\inn\textup{-Mod}$ we have $d_{b}^\infty(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})) \leq d_{\altmathcal{R}}^1(\mathbb{V},\mathbb{W}).$ \end{corollary} \subsection{Comparison with $d_b^1$ and the Bottleneck Distance}\label{sec:comparison_with_bottleneck} The next proposition shows that restricting the reflection distance to $\taumod{\tau}$ for a fixed type $\tau$, we obtain bi-Lipschitz equivalence of $d_{\altmathcal{R}}^1$ and $d_b^1$. First, we need the following lemmas: \begin{lemma}\label{lem:bound_to_zero} Fix $n\in \mathbb{N}$. For $\mathbb{V}\inn\textup{-Mod}$ we have $d_\altmathcal{R}^1(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$})\leq n\cdot {n\choose 2} = n^2(n+1)/2$. \begin{proof} There are precisely ${n\choose k}$ intervals in $\{1,\dots, n\}$ corresponding to non-simple summands. Note that any sequence of reflections which annihilates an interval summand of $\mathbb{V}$ annihilates all copies of this summand. Since any interval summand of a zigzag module can be annihilated with no more than $n$ reflections, we have $d_\altmathcal{R}^1(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$})\leq n\cdot {n\choose 2}$. \end{proof} \end{lemma} \begin{lemma}\label{lem:d_b=0_implies_d_R=0} Fix $n\in \mathbb{N}$ and a type $\tau\in\altmathcal{T}_n$. For $\mathbb{V},\mathbb{W}\in\taumod{\tau}$, if $d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})) = 0$ then $d_\altmathcal{R}(\mathbb{V},\mathbb{W}) = 0$. \begin{proof} It is not hard to see that if $d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})) = 0$ then \[ \{\!\!\{(b,d)\in \mathrm{Dgm}(\mathbb{V}) \ | \ b\neq d\}\!\!\} = \{\!\!\{(b,d)\in \mathrm{Dgm}(\mathbb{W}) \ | \ b\neq d\}\!\!\},\] i.e. $\mathrm{Dgm}(\mathbb{V})$ and $\mathrm{Dgm}(\mathbb{W})$ have the same multisubsets of off-diagonal points. It follows that $\altmathcal{S}(\mathbb{V}) = \altmathcal{S}(\mathbb{W})$ so that $\epsilon(\mathbb{V}) = \altmathcal{S}(\mathbb{V}) = \altmathcal{S}(\mathbb{W}) \precsim \mathbb{W}$ and $\epsilon(\mathbb{W}) = \altmathcal{S}(\mathbb{W}) = \altmathcal{S}(\mathbb{V}) \precsim \mathbb{V}$. Hence $d_\altmathcal{R}^1(\mathbb{V},\mathbb{W}) = 0$. \end{proof} \end{lemma} \begin{proposition}[Bi-Lipschitz Equivalence of $d_{\altmathcal{R}}^1$ and $d_b^1$ for fixed $n$] \label{prop:db1_bilip}Fix $n\in \mathbb{N}$ and a type $\tau\in\altmathcal{T}_n$. For all zigzag modules $\mathbb{V},\mathbb{W}\in\taumod{\tau}$ we have \[ d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W}))\leq d_{\altmathcal{R}}^1(\mathbb{V},\mathbb{W})\leq n^2(n+1)\cdot d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})).\] \begin{proof} The first of these inequalities is the claim of Theorem \ref{thm:main}. Now let $\eta = d_{\altmathcal{R}}^1(\mathbb{V},\mathbb{W})$. If $\eta = 0$ then the second inequality holds trivially, so suppose that $\eta>0$. By (the contrapositive of) Lemma \ref{lem:d_b=0_implies_d_R=0}, we have $d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W}))\geq 1$. By Lemma \ref{lem:bound_to_zero} and the triangle inequality, \[ d_\altmathcal{R}^1(\mathbb{V},\mathbb{W}) \leq d_\altmathcal{R}^1(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$}) + d_\altmathcal{R}^1(\mathbb{W},\scalebox{.7}[1]{$\mathbb{O}$}) \leq n^2(n+1).\] Thus $d_\altmathcal{R}^1(\mathbb{V},\mathbb{W})\leq n^2(n+1)\leq n^2(n+1)\cdot d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W}))$. \end{proof} \end{proposition} \begin{corollary}\label{cor:bottleneck_bilip} Fix $n\in \mathbb{N}$ and a type $\tau\in \altmathcal{T}_n$. For any $\mathbb{V},\mathbb{W}\in \taumod{\tau}$ we have \[d_{b}^\infty(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})) \leq d_{\altmathcal{R}}^1(\mathbb{V},\mathbb{W})\leq 2n^2(n+1)\cdot d_b^\infty(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})).\] \begin{proof} This follows immediately from Proposition \ref{prop:db1_bilip} and the equivalence $d_b^\infty\leq d_b^1\leq 2d_b^1$ (see Remark \ref{remark:db1_bottleneck_equiv}). \end{proof} \end{corollary} \begin{remark} Bi-Lipschitz equivalence does not hold in general on $n\textup{-Mod}$. To see this, consider any two zigzag modules $\mathbb{V},\mathbb{W}\in n\textup{-Mod}$ for which $\mathrm{Dgm}(\mathbb{V}) = \mathrm{Dgm}(\mathbb{W})$ but $\mathbb{V}\not\sim_\altmathcal{S}\mathbb{W}$. Then $d_b^1(\mathrm{Dgm}(\mathbb{V}),\mathrm{Dgm}(\mathbb{W})) = 0$ but $d_\altmathcal{R}^1(\mathbb{V},\mathbb{W})>0$ by the second statement of Theorem \ref{thm:ref_is_metric}. For example, let $n = 3$, let $\tau = (\rightarrow,\rightarrow)$, and let $\tau' = (\rightarrow,\leftarrow)$. Define \[\mathbb{V} = \mathbb{I}_\tau([1,2]) \oplus \mathbb{I}_{\tau}([2,3]) \quad \textup{ and } \quad \mathbb{W} = \mathbb{I}_{\tau'}([1,2]) \oplus \mathbb{I}_{\tau'}([2,3]).\] By definition $\mathrm{Dgm}(\mathbb{V}) = \mathrm{Dgm}(\mathbb{W})$. On the other hand, we have $\altmathcal{S}(\mathbb{V}) = \mathbb{V}$ and $\altmathcal{S}(\mathbb{W}) = \mathbb{W}$ so that $\mathbb{V} \sim_{\altmathcal{S}} \mathbb{W}$ if and only if $\,\mathbb{V}\precsim \mathbb{W}$ and $\mathbb{W} \precsim \mathbb{V}$ which is true if and only if $\,\mathbb{V} \sim \mathbb{W}$. But $\mathbb{V}\not\sim \mathbb{W}$, as none of the structure maps of $\mathbb{V}$ nor $\mathbb{W}$ are isomorphisms. Thus $\mathbb{V}\not\sim_{\altmathcal{S}}\mathbb{W}$. \end{remark} \subsection{Comparison with Botnan and Lesnick's Distance} Botnan and Lesnick introduced a distance between zigzag modules in \cite{algstabzz}. Their distance is defined by considering extensions of zigzag modules to so-called \textit{block decomposable}, two-dimensional persistence modules. Their distance, denoted $d_I$, is defined to be the two-dimensional interleaving distance between these extensions. In this section we consider examples which show that in general there is not a bi-Lipschitz equivalence between $d_I$ and $d_\altmathcal{R}^1$. We refer the reader to \cite[Definition 4.4]{algstabzz} for details on $d_I$. \begin{example} Let $n = 3$ and let $\tau = (\leftarrow,\rightarrow)\in\altmathcal{T}_n$. Consider the interval modules $\mathbb{V}:= \mathbb{I}_\tau([1,3])$. We have $d_\altmathcal{R}^1(\mathbb{V},\scalebox{.7}[1]{$\mathbb{O}$}) = 2$ by Remark \ref{finiteness_remark}. On the other hand, we can view $\mathbb{V}$ as the zigzag module \[\widetilde{\mathbb{V}} = \begin{tikzcd}[column sep=2em] \cdots &\arrow[l,swap,"0"]0 \arrow[r,"0"]& \mathbb{F}& \mathbb{F} \arrow[r,"\textup{id}"]\arrow[l,swap,"\textup{id}"]& \mathbb{F} & 0 \arrow[l,swap,"0"]\arrow[r,"0"]& \cdots, \end{tikzcd}\] a zigzag module indexed by the integers obtained from $\mathbb{V}$ by extending with $0$. The block extension functor of Botnan and Lesnick sends $\widetilde{\mathbb{V}}$ to an unbounded block in $\mathbb{R}^2$ and hence $d_I(\widetilde{\mathbb{V}},\scalebox{.7}[1]{$\mathbb{O}$}) = \infty$ (see Figure 4 in \cite{algstabzz}). \end{example} The next example gives a sequence $(\mathbb{V}_n)_{n\in\mathbb{N}}$ of zigzag modules whose reflection distance to the zero module increases linearly but whose $d_I$ distance to the zero module is constant: \begin{example} For each $n\in \mathbb{N}$ let $\mathbb{V}_n$ be the zigzag module \[\mathbb{V}_n = \begin{tikzcd}[column sep=2em] \mathbb{F} &\arrow[l,swap,"\textup{id}"]\mathbb{F} \arrow[r,"0"]& 0& 0 \arrow[r,"0"]\arrow[l,swap,"0"]& \cdots\arrow[r,"0"]& \mathbb{F} &\arrow[l,swap,"\textup{id}"]\mathbb{F} \arrow[r,"0"]& 0& 0 \arrow[l,swap,"0"] \end{tikzcd}\] of length $4n$ ($\,\mathbb{V}_n$ is the ``concatenation" of $n$ copies of $\,\mathbb{F} \stackrel{\textup{id}}{\longleftarrow} \mathbb{F} \stackrel{0}{\longrightarrow} 0 \stackrel{0}{\longleftarrow}0$). It is not hard to see that $d_\altmathcal{R}^1(\mathbb{V}_n,\scalebox{.7}[1]{$\mathbb{O}$}) = n$. On the other hand, extending each $\mathbb{V}_n$ by 0 to obtain zigzag modules $\widetilde{V}_n$ indexed over the integers, we have $d_I(\widetilde{V}_n,\scalebox{.7}[1]{$\mathbb{O}$}) = 1/2$ for all $n$. To see this, note that the extension of $\widetilde{V}_n$ to a two-dimensional persistence module is a direct sum of block modules $I^{[1,2)_{\textup{BL}}},I^{[3,4)_{\textup{BL}}},\dots, I^{[2n-1,2n)_\textup{BL}}$ (see Section 4.1 of \cite{algstabzz} for notation). These block modules are infinite horizontal strips of height $1$ with vertical distance $1$ between adjacent blocks (see again Figure 4 from \cite{algstabzz}; $I^{[k,k+1)_\textup{BL}}$ is a vertical shift by $k$ of the block $I^{[1,2)_\textup{BL}}$). The zero morphism then serves as a $1/2$-interleaving between this direct sum of block modules and the zero module. Moreover, after a moment of reflection one sees that $1/2$ is the smallest possible interleaving constant. \end{example} \section{Discussion}\label{sec:discussion} Our definition of the reflection distance made use of the notion of reflection functors introduced by Gelfand et al \cite{bgp} which are transformations on zigzag modules that affect only a portion a given module. In our constructions, the effect of these transformations was restricted to a subdiagram of length at most $3$ (cf. Section \ref{sec:reflections}, equation (\ref{subdiagram})). This design choice could also potentially be altered, and its exploration may lead to other interesting distances. In particular, it is conceivable that similar ideas can be exported to the setting of persistence modules over graphs other than those inducing zigzag modules. Indeed, since Gelfand et al consider reflection functors on arbitrary graphs, the possibility of extending our reflection distance to such general setting appears interesting. One initial question arising from this is whether by applying a suitable sequence of reflections to a given persistence module defined on a graph, one can transform it into a summand of another such persistence module. Whereas we were able to show that this is always possible for zigzag modules, it is not clear that this can be done for persistence modules defined on arbitrary graphs. Our construction of a distance between zigzag modules took a route different from the one followed by Botnan and Lesnick in \cite{algstabzz}. We showed that these two distances are in general not bi-Lipschitz equivalent. It seems of interest to identify large families of zigzag persistence modules for which a bi-Lipschitz equivalence might be possible. Another topic where research would be welcome is the elucidation of the computational complexity associated to estimating distances between zigzag modules of a given length. In this direction, recent results by Botnan, Bjerkevik, and Kerber \cite{kerber} about the computational complexity of the two-dimensional interleaving distance might be relevant. \subsection*{Acknowledgements} We thank Peter Bubenik, Mike Catanzaro, Woojin Kim, and Jos\'e Perea for their useful feedback. We also thank Lauren Wickman for her careful reading of the late drafts. A.E. was supported by the NSF RTG \# 1547357. F.M. was supported by NSF RI \# 1422400 and NSF AF \#1526513. \bibliographystyle{alpha}

\section{Introduction} The theory of relativistic wave equations (RWE) is one of the oldest topics in theoretical physics. As usual, the theory of RWE closely relates with higher spin formalisms. However, at present there is no a fully adequate formalism for description of higher--spin fields (all widely accepted higher--spin formalisms such as Rarita--Schwinger approach \cite{RS41}, Bargmann--Wigner \cite{BW48} and Gel'fand--Yaglom \cite{GY48} multispinor theories, and also Joos--Weinberg $2(2j+1)$--component formalism \cite{Joo62,Wein} have many intrinsic contradictions and difficulties). On the other hand, it is known that the study of RWE leads naturally to the fields which depend on both Minkowski space coordinates and some continuous variables corresponding to spin degrees of freedom \cite{GT47,BW48,Yuk50,Shi51}. Wave functions of this type can be treated as the fields on homogeneous spaces of the Poincar\'{e} group. These fields was first studied by Finkelstein \cite{Fin55}, he gave a classification and explicit constructions of homogeneous spaces of the Poincar\'{e} group. The general form of these fields closely relates with the structure of the Lorentz and Poincar\'{e} group representations \cite{GMS,Nai58,BBTD88,GS01} and admits the following factorization $f(x,{\bf z})=\phi^n({\bf z})\psi_n(x)$, where $x\in T_4$ and $\phi^n({\bf z})$ form a basis in the representation space of the Lorentz group, $T_4$ is the group of four-dimensional translations. In addition to a general theory of relativistic wave equations, a search of solutions for RWE has a great importance (see, for example, \cite{BG89}). In the present work we study solutions of RWE in terms of the fields defined on the two-dimensional complex sphere. This sphere is a homogeneous space of the Lorentz group (in this case the field functions $f(x,{\bf z})$ reduce to $\phi^n({\bf z})$). In the previous work \cite{Var03}, solutions of RWE were obtained in the form of expansions in hyperspherical functions for the representations of the type $(l,0)\oplus(0,\dot{l})$. Solutions (wavefunctions on the group manifold ${\cal M}_{10}$) of the simplest RWE (Dirac $(1/2,0)\oplus(0,1/2)$ and Maxwell $(1,0)\oplus(0,1)$ fields) have been given in \cite{Var03c,Var03d}. In this work we study more general representations $(l_1,\dot{l}_2)\oplus(l_2,\dot{l}_1)$ (tensor representations of the proper orthocronous Lorentz group $\mathfrak{G}_+$) which correspond to arbitrary spin chains. In turn, arbitrary spin chains contain interlocking representations of the group $\mathfrak{G}_+$ as a particular case. It is known that the higher spin fields and correspondingly composite elementary particles can be formulated in terms of interlocking representations of $\mathfrak{G}_+$ (Bhabha-Gel'fand-Yaglom chains). The stable composite particle corresponds to an indecomposable equation defined within interlocking representation, and vice versa, unstable particles are described by decomposable equations. The present paper is organized as follows. In the section 2 we consider some basic facts concerning the group $SL(2,\hbox{\bb C})$. The main object of this section is a local isomorphism $SL(2,\hbox{\bb C})\simeq SU(2)\otimes SU(2)$. We introduce here hyperspherical functions for the tensor representations and further we study recurrence relations between these functions. It is shown that matrix elements of the representations $(l_1,\dot{l}_2)\oplus(l_2,\dot{l}_1)$ are expressed via the hyperspherical functions. In the section 3 we give a brief introduction to the fields on the Poincar\'{e} group ${\cal P}$. We consider fields on the group manifold ${\cal M}_{10}$ and on the eight-dimensional homogeneous space ${\cal M}_8=\R^{1,3}\times\dS^2$ of ${\cal P}$, where $\R^{1,3}$ is the Minkowski spacetime and $\dS^2$ is the two-dimensional complex sphere. In parallel, we consider basic facts concerning harmonic analysis on the homogeneous spaces ${\cal M}_{10}$, ${\cal M}_8$, and also on the group $SL(2,\hbox{\bb C})$ and on the sphere $\dS^2$. The following logical step consists in definition of the Lagrangian formalism and field equations on the homogeneous spaces of ${\cal P}$. The field equations for arbitrary spin are derived in the section 4 by the standard variation procedure from a selected Lagrangian. The explicit construction of RWE for the arbitrary spin chains of $\mathfrak{G}_+$ is given in the section 5. In the section 6 we separate variables in a relativistically invariant system via the recurrence relations between hyperspherical functions. First of all, we set up a boundary value problem for RWE, which can be treated as a Dirichlet problem for a complex ball. It is shown that solutions of this problem are expressed via Fourier type series on the two-dimensional complex sphere. \section{The group $SL(2,\hbox{\bb C})$}\label{SL2C} As is known, the group $SL(2,\hbox{\bb C})$ is an universal covering of the proper orthochronous Lorentz group $\mathfrak{G}_+$. The group $SL(2,\hbox{\bb C})$ of all complex matrices \[\ar \mathfrak{g}= \begin{pmatrix} \alpha & \beta\\ \gamma & \delta \end{pmatrix} \] of 2-nd order with the determinant $\alpha\delta-\gamma\beta=1$, is a {\it complexification} of the group $SU(2)$. The group $SU(2)$ is one of the real forms of $SL(2,\hbox{\bb C})$. The transition from $SU(2)$ to $SL(2,\hbox{\bb C})$ is realized via the complexification of three real parameters $\varphi,\,\theta,\,\psi$ (Euler angles). Let $\theta^c=\theta-i\tau$, $\varphi^c=\varphi-i\epsilon$, $\psi^c=\psi-i\varepsilon$ be complex Euler angles,\index{Euler angles!complex} where \begin{equation}\label{CEA} {\renewcommand{\arraystretch}{1.05} \begin{array}{ccccc} 0 &\leq&\mbox{\rm Re}\,\theta^c=\theta& \leq& \pi,\\ 0 &\leq&\mbox{\rm Re}\,\varphi^c=\varphi& <&2\pi,\\ -2\pi&\leq&\mbox{\rm Re}\,\psi^c=\psi&<&2\pi, \end{array}\quad\quad \begin{array}{ccccc} -\infty &<&\mbox{\rm Im}\,\theta^c=\tau&<&+\infty,\\ -\infty&<&\mbox{\rm Im}\,\varphi^c=\epsilon&<&+\infty,\\ -\infty&<&\mbox{\rm Im}\,\psi^c=\varepsilon&<&+\infty. \end{array}} \end{equation} The group $SL(2,\hbox{\bb C})$ has six one-parameter subgroups \begin{gather} a_1(t)=\begin{pmatrix} \cos\frac{t}{2} & i\sin\frac{t}{2}\\ i\sin\frac{t}{2} & \cos\frac{t}{2} \end{pmatrix},\quad a_2(t)=\begin{pmatrix} \cos\frac{t}{2} & -\sin\frac{t}{2}\\ \sin\frac{t}{2} & \cos\frac{t}{2} \end{pmatrix},\quad a_3(t)=\begin{pmatrix} e^{\frac{it}{2}} & 0\\ 0 & e^{-\frac{it}{2}} \end{pmatrix},\nonumber\\ b_1(t)=\begin{pmatrix} \cosh\frac{t}{2} & \sinh\frac{t}{2}\\ \sinh\frac{t}{2} & \cosh\frac{t}{2} \end{pmatrix},\quad b_2(t)=\begin{pmatrix} \cosh\frac{t}{2} & i\sinh\frac{t}{2}\\ -i\sinh\frac{t}{2} & \cosh\frac{t}{2} \end{pmatrix},\quad b_3(t)=\begin{pmatrix} e^{\frac{t}{2}} & 0\\ 0 & e^{-\frac{t}{2}} \end{pmatrix}.\nonumber \end{gather} The tangent matrices $A_i$ and $B_i$ of these subgroups are defined as follows \begin{eqnarray} A_1&=&\left.\frac{d a_1(t)}{dt}\right|_{t=0}=\frac{i}{2}\begin{pmatrix} 0 & 1\\ 1 & 0 \end{pmatrix},\nonumber\\ A_2&=&\left.\frac{d a_2(t)}{dt}\right|_{t=0}=\frac{1}{2}\begin{pmatrix} 0 & -1\\ 1 & 0 \end{pmatrix},\nonumber\\ A_3&=&\left.\frac{d a_3(t)}{dt}\right|_{t=0}=\frac{i}{2}\begin{pmatrix} 1 & 0\\ 0 & -1 \end{pmatrix},\nonumber\\ B_1&=&\left.\frac{d b_1(t)}{dt}\right|_{t=0}=\frac{1}{2}\begin{pmatrix} 0 & 1\\ 1 & 0 \end{pmatrix},\nonumber\\ B_2&=&\left.\frac{d b_2(t)}{dt}\right|_{t=0}=\frac{i}{2}\begin{pmatrix} 0 & 1\\ -1 & 0 \end{pmatrix},\nonumber\\ B_3&=&\left.\frac{d b_3(t)}{dt}\right|_{t=0}=\frac{1}{2}\begin{pmatrix} 1 & 0\\ 0 &-1 \end{pmatrix}.\nonumber \end{eqnarray} The elements ${\sf A}_i$ and ${\sf B}_i$ form a basis of Lie algebra $\mathfrak{sl}(2,\hbox{\bb C})$ and satisfy the relations \begin{equation}\label{Com1} \left.\begin{array}{lll} \left[{\sf A}_1,{\sf A}_2\right]={\sf A}_3, & \left[{\sf A}_2,{\sf A}_3\right]={\sf A}_1, & \left[{\sf A}_3,{\sf A}_1\right]={\sf A}_2,\\[0.1cm] \left[{\sf B}_1,{\sf B}_2\right]=-{\sf A}_3, & \left[{\sf B}_2,{\sf B}_3\right]=-{\sf A}_1, & \left[{\sf B}_3,{\sf B}_1\right]=-{\sf A}_2,\\[0.1cm] \left[{\sf A}_1,{\sf B}_1\right]=0, & \left[{\sf A}_2,{\sf B}_2\right]=0, & \left[{\sf A}_3,{\sf B}_3\right]=0,\\[0.1cm] \left[{\sf A}_1,{\sf B}_2\right]={\sf B}_3, & \left[{\sf A}_1,{\sf B}_3\right]=-{\sf B}_2, & \\[0.1cm] \left[{\sf A}_2,{\sf B}_3\right]={\sf B}_1, & \left[{\sf A}_2,{\sf B}_1\right]=-{\sf B}_3, & \\[0.1cm] \left[{\sf A}_3,{\sf B}_1\right]={\sf B}_2, & \left[{\sf A}_3,{\sf B}_2\right]=-{\sf B}_1. & \end{array}\right\} \end{equation} Let us consider the operators \begin{gather} {\sf X}_l=\frac{1}{2}i({\sf A}_l+i{\sf B}_l),\quad{\sf Y}_l=\frac{1}{2}i({\sf A}_l-i{\sf B}_l), \label{SL25}\\ (l=1,2,3).\nonumber \end{gather} Using the relations (\ref{Com1}), we find that \begin{equation}\label{Com2} \left[{\sf X}_k,{\sf X}_l\right]=i\varepsilon_{klm}{\sf X}_m,\quad \left[{\sf Y}_l,{\sf Y}_m\right]=i\varepsilon_{lmn}{\sf Y}_n,\quad \left[{\sf X}_l,{\sf Y}_m\right]=0. \end{equation} Further, introducing generators of the form \begin{equation}\label{SL26} \left.\begin{array}{cc} {\sf X}_+={\sf X}_1+i{\sf X}_2, & {\sf X}_-={\sf X}_1-i{\sf X}_2,\\[0.1cm] {\sf Y}_+={\sf Y}_1+i{\sf Y}_2, & {\sf Y}_-={\sf Y}_1-i{\sf Y}_2, \end{array}\right\} \end{equation} we see that in virtue of commutativity of the relations (\ref{Com2}) a space of an irreducible finite--dimensional representation of the group $SL(2,\hbox{\bb C})$ can be spanned on the totality of $(2l+1)(2\dot{l}+1)$ basis vectors $\mid l,m;\dot{l},\dot{m}\rangle$, where $l,m,\dot{l},\dot{m}$ are integer or half--integer numbers, $-l\leq m\leq l$, $-\dot{l}\leq \dot{m}\leq \dot{l}$. Therefore, \begin{eqnarray} &&{\sf X}_-\mid l,m;\dot{l},\dot{m}\rangle= \sqrt{(l+m)(l-m+1)}\mid l,m-1,\dot{l},\dot{m}\rangle \;\;(m>-l),\nonumber\\ &&{\sf X}_+\mid l,m;\dot{l},\dot{m}\rangle= \sqrt{(l-m)(l+m+1)}\mid l,m+1;\dot{l},\dot{m}\rangle \;\;(m<l),\nonumber\\ &&{\sf X}_3\mid l,m;\dot{l},\dot{m}\rangle= m\mid l,m;\dot{l},\dot{m}\rangle,\nonumber\\ &&{\sf Y}_-\mid l,m;\dot{l},\dot{m}\rangle= \sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)}\mid l,m;\dot{l},\dot{m}-1 \rangle\;\;(\dot{m}>-\dot{l}),\nonumber\\ &&{\sf Y}_+\mid l,m;\dot{l},\dot{m}\rangle= \sqrt{(\dot{l}-\dot{m})(\dot{l}+\dot{m}+1)}\mid l,m;\dot{l},\dot{m}+1 \rangle\;\;(\dot{m}<\dot{l}),\nonumber\\ &&{\sf Y}_3\mid l,m;\dot{l},\dot{m}\rangle= \dot{m}\mid l,m;\dot{l},\dot{m}\rangle.\label{Waerden} \end{eqnarray} From the relations (\ref{Com2}) it follows that each of the sets of infinitesimal operators ${\sf X}$ and ${\sf Y}$ generates the group $SU(2)$ and these two groups commute with each other. Thus, from the relations (\ref{Com2}) and (\ref{Waerden}) it follows that the group $SL(2,\hbox{\bb C})$, in essence, is equivalent locally to the group $SU(2)\otimes SU(2)$. It should be noted here that the representation basis, defined by the formulae (\ref{SL25})--(\ref{Waerden}), has an evident physical meaning. For example, in the case of $(1,0)\oplus(0,1)$--representation space there is an analogy with the photon spin states. Namely, the operators ${\sf X}$ and ${\sf Y}$ correspond to the right and left polarization states of the photon. For that reason we will call the canonical basis consisting of the vectors $\mid lm;\dot{l}\dot{m}\rangle$ as {\it a helicity basis}. In the work \cite{Var03} we studied RWE for the fields $(l,0)\oplus(0,\dot{l})$ which described within the representation $\boldsymbol{\tau}_{l,0}\oplus\boldsymbol{\tau}_{0,\dot{l}}$ of the group $\mathfrak{G}_+$. In this work we study more general representations of the type $\boldsymbol{\tau}_{l,\dot{l}}\oplus\boldsymbol{\tau}_{\dot{l},l}$ (tensor representations), where $\boldsymbol{\tau}_{l,\dot{l}}=\boldsymbol{\tau}_{l,0}\otimes \boldsymbol{\tau}_{0,\dot{l}}$ and $\boldsymbol{\tau}_{\dot{l},l}=\boldsymbol{\tau}_{\dot{l},0}\otimes \boldsymbol{\tau}_{0,l}$. It is obvious that these representations include $\boldsymbol{\tau}_{l,0}\oplus\boldsymbol{\tau}_{0,\dot{l}}$ as a particular case. In general, a tensor structure of the infinitesimal operators has the form \begin{eqnarray} {\sf A}^{l\dot{l}}_i&=&{\sf A}^l_i\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes{\sf A}^{\dot{l}}_i,\nonumber\\ {\sf B}^{l\dot{l}}_i&=&{\sf B}^l_i\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes{\sf B}^{\dot{l}}_i,\nonumber\\ \widetilde{{\sf A}}^{l\dot{l}}_i&=&{\sf A}^{\dot{l}}_i\otimes\boldsymbol{1}_{2l+1}- \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf A}^l_i,\nonumber\\ \widetilde{{\sf B}}^{l\dot{l}}_i&=&{\sf B}^{\dot{l}}_i\otimes\boldsymbol{1}_{2l+1}- \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf B}^l_i,\label{TenStr} \end{eqnarray} where ${\sf A}^l_i$, ${\sf B}^l_i$ and ${\sf A}^{\dot{l}}_i$, ${\sf B}^{\dot{l}}_i$ are infinitesimal operators of the representations $\boldsymbol{\tau}_{l,0}$ and $\boldsymbol{\tau}_{0,\dot{l}}$, respectively ($i=1,2,3$). Or, more explicitly, \begin{multline} {\sf A}^{l\dot{l}}_1\mid l,m;\dot{l},\dot{m}\rangle= -\frac{i}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle -\frac{i}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle+\\ +\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle +\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle, \label{Op1} \end{multline} \begin{multline} {\sf A}^{l\dot{l}}_2\mid l,m;\dot{l},\dot{m}\rangle= \frac{1}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle -\frac{1}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle-\\ -\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle +\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle, \label{Op2} \end{multline} \begin{equation}\label{Op3} {\sf A}^{l\dot{l}}_3\mid l,m;\dot{l},\dot{m}\rangle= -im\mid l,m;\dot{l},\dot{m}\rangle +i\dot{m}\mid l,m;\dot{l},\dot{m}\rangle, \end{equation} \begin{multline} {\sf B}^{l\dot{l}}_1\mid l,m;\dot{l},\dot{m}\rangle= -\frac{1}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle -\frac{1}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle-\\ -\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle -\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle, \label{Op4} \end{multline} \begin{multline} {\sf B}^{l\dot{l}}_2\mid l,m;\dot{l},\dot{m}\rangle= -\frac{i}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle +\frac{i}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle-\\ -\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle +\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle, \label{Op5} \end{multline} \begin{equation}\label{Op6} {\sf B}^{l\dot{l}}_3\mid l,m;\dot{l},\dot{m}\rangle= -m\mid l,m;\dot{l},\dot{m}\rangle -\dot{m}\mid l,m;\dot{l},\dot{m}\rangle, \end{equation} \begin{multline} \widetilde{{\sf A}}^{l\dot{l}}_1\mid l,m;\dot{l},\dot{m}\rangle= -\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle -\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle+\\ +\frac{i}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle +\frac{i}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle, \label{Op7} \end{multline} \begin{multline} \widetilde{{\sf A}}^{l\dot{l}}_2\mid l,m;\dot{l},\dot{m}\rangle= \frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle -\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle+\\ +\frac{1}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle +\frac{1}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle, \label{Op8} \end{multline} \begin{equation}\label{Op9} \widetilde{{\sf A}}^{l\dot{l}}_3\mid l,m;\dot{l},\dot{m}\rangle= -i\dot{m}\mid l,m;\dot{l},\dot{m}\rangle +im\mid l,m;\dot{l},\dot{m}\rangle, \end{equation} \begin{multline} \widetilde{{\sf B}}^{l\dot{l}}_1\mid l,m;\dot{l},\dot{m}\rangle= \frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle +\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle+\\ +\frac{1}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle +\frac{1}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle, \label{Op10} \end{multline} \begin{multline} \widetilde{{\sf B}}^{l\dot{l}}_2\mid l,m;\dot{l},\dot{m}\rangle= \frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\mid l,m;\dot{l},\dot{m}-1\rangle -\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\mid l,m;\dot{l},\dot{m}+1\rangle+\\ +\frac{i}{2}\boldsymbol{\alpha}^l_m\mid l,m-1;\dot{l},\dot{m}\rangle -\frac{i}{2}\boldsymbol{\alpha}^l_{m+1}\mid l,m+1;\dot{l},\dot{m}\rangle, \label{Op11} \end{multline} \begin{equation}\label{Op12} \widetilde{{\sf B}}^{l\dot{l}}_3\mid l,m;\dot{l},\dot{m}\rangle= \dot{m}\mid l,m;\dot{l},\dot{m}\rangle +m\mid l,m;\dot{l},\dot{m}\rangle. \end{equation} Further, let us define generators ${\sf X}^{l\dot{l}}_-$, ${\sf X}^{l\dot{l}}_+$, ${\sf Y}^{l\dot{l}}_-$, ${\sf Y}^{l\dot{l}}_+$ and $\widetilde{{\sf X}}^{l\dot{l}}_-$, $\widetilde{{\sf X}}^{l\dot{l}}_+$, $\widetilde{{\sf Y}}^{l\dot{l}}_-$, $\widetilde{{\sf Y}}^{l\dot{l}}_+$ following the rule (\ref{SL26}). It is easy to verify that these generators satisfy the relations of the type (\ref{Waerden}). On the group $SL(2,\hbox{\bb C})$ there exist the following Laplace-Beltrami operators: \begin{eqnarray} {\sf X}^2&=&{\sf X}^2_1+{\sf X}^2_2+{\sf X}^2_3=\frac{1}{4}({\sf A}^2-{\sf B}^2+2i{\sf A}{\sf B}),\nonumber\\ {\sf Y}^2&=&{\sf Y}^2_1+{\sf Y}^2_2+{\sf Y}^2_3= \frac{1}{4}(\widetilde{{\sf A}}^2-\widetilde{{\sf B}}^2- 2i\widetilde{{\sf A}}\widetilde{{\sf B}}).\label{KO} \end{eqnarray} At this point, we see that operators (\ref{KO}) contain the well known Casimir operators ${\sf A}^2-{\sf B}^2$, ${\sf A}{\sf B}$ of the Lorentz group. Using expressions (\ref{CEA}), we obtain an Euler parametrization of the Laplace-Beltrami operators, \begin{eqnarray} {\sf X}^2&=&\frac{\partial^2}{\partial\theta^c{}^2}+ \cot\theta^c\frac{\partial}{\partial\theta^c}+\frac{1}{\sin^2\theta^c}\left[ \frac{\partial^2}{\partial\varphi^c{}^2}- 2\cos\theta^c\frac{\partial}{\partial\varphi^c} \frac{\partial}{\partial\psi^c}+ \frac{\partial^2}{\partial\psi^c{}^2}\right],\nonumber\\ {\sf Y}^2&=&\frac{\partial^2}{\partial\dot{\theta}^c{}^2}+ \cot\dot{\theta}^c\frac{\partial}{\partial\dot{\theta}^c}+ \frac{1}{\sin^2\dot{\theta}^c}\left[ \frac{\partial^2}{\partial\dot{\varphi}^c{}^2}- 2\cos\dot{\theta}^c\frac{\partial}{\partial\dot{\varphi}^c} \frac{\partial}{\partial\dot{\psi}^c}+ \frac{\partial^2}{\partial\dot{\psi}^c{}^2}\right].\label{KO2} \end{eqnarray} Matrix elements $t^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})= \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi^c,\theta^c,\psi^c)$ of irreducible representations of $SL(2,\hbox{\bb C})$ are eigenfunctions of the operators (\ref{KO2}), \begin{eqnarray} \left[{\sf X}^2+l(l+1)\right]\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi^c,\theta^c,\psi^c)&=&0,\nonumber\\ \left[{\sf Y}^2+\dot{l}(\dot{l}+1)\right]\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}} (\dot{\varphi}^c,\dot{\theta}^c,\dot{\psi}^c)&=&0,\label{EQ} \end{eqnarray} where \begin{equation}\label{HF3'} \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})= e^{-i(m\varphi^c+n\psi^c)}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}} (\cos\theta^c,\cos\dot{\theta}^c)e^{i(\dot{m}\dot{\varphi}^c+ \dot{n}\dot{\psi}^c)}. \end{equation} Substituting the functions (\ref{HF3'}) into (\ref{EQ}) and taking into account the operators (\ref{KO2}) and substitutions $z=\cos\theta^c$, $\overset{\ast}{z}=\cos\dot{\theta}^c$, we come to the following differential equations (a complex analog of the Legendre equations): \begin{eqnarray} \left[(1-z^2)\frac{d^2}{dz^2}-2z\frac{d}{dz}- \frac{m^2+n^2-2mnz}{1-z^2}+l(l+1)\right] \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(z,\overset{\ast}{z})&=&0,\label{Leg1}\\ \left[(1-\overset{\ast}{z}{}^2)\frac{d^2}{d\overset{\ast}{z}{}^2}- 2\overset{\ast}{z}\frac{d}{d\overset{\ast}{z}}- \frac{\dot{m}^2+\dot{n}^2-2\dot{m}\dot{n}\overset{\ast}{z}} {1-\overset{\ast}{z}{}^2}+\dot{l}(\dot{l}+1)\right] \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(z,\overset{\ast}{z})&=&0.\label{Leg2} \end{eqnarray} The latter equations have three singular points $-1$, $+1$, $\infty$. Solutions of (\ref{Leg1})-(\ref{Leg2}) have the form \begin{equation} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\cos\theta^c,\cos\dot{\theta}^c)= Z^l_{mn}(\cos\theta^c)Z^{\dot{l}}_{\dot{m}\dot{n}}(\cos\dot{\theta}^c), \label{HS} \end{equation} where \begin{multline} Z^l_{mn}(\cos\theta^c)= \sum^l_{k=-l}i^{m-k} \sqrt{\Gamma(l-m+1)\Gamma(l+m+1)\Gamma(l-k+1)\Gamma(l+k+1)}\times\\ \cos^{2l}\frac{\theta}{2}\tan^{m-k}\frac{\theta}{2}\times\\[0.2cm] \sum^{\min(l-m,l+k)}_{j=\max(0,k-m)} \frac{i^{2j}\tan^{2j}\dfrac{\theta}{2}} {\Gamma(j+1)\Gamma(l-m-j+1)\Gamma(l+k-j+1)\Gamma(m-k+j+1)}\times\\[0.2cm] \sqrt{\Gamma(l-n+1)\Gamma(l+n+1)\Gamma(l-k+1)\Gamma(l+k+1)} \cosh^{2l}\frac{\tau}{2}\tanh^{n-k}\frac{\tau}{2}\times\\[0.2cm] \sum^{\min(l-n,l+k)}_{s=\max(0,k-n)} \frac{\tanh^{2s}\dfrac{\tau}{2}} {\Gamma(s+1)\Gamma(l-n-s+1)\Gamma(l+k-s+1)\Gamma(n-k+s+1)}.\label{HS0} \end{multline} and $Z^{\dot{l}}_{\dot{m}\dot{n}}$ is a complex conjugate function with respect to $Z^l_{mn}$. We will call the functions $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}$ in (\ref{HS}) as {\it tensor hyperspherical functions}. Using the formula (\ref{HS}), we find explicit expressions for the matrices $T_{l\dot{l}}(\theta,\tau)$ of the finite--dimensional representations $\boldsymbol{\tau}_{\frac{1}{2}\frac{1}{2}}$ and $\boldsymbol{\tau}_{1\frac{1}{2}}$: \begin{gather} T_{\frac{1}{2}\frac{1}{2}}(\theta,\tau)=\ar\begin{pmatrix} Z^{\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2}} & Z^{\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2}} \\ Z^{\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2}} & Z^{\frac{1}{2}}_{\frac{1}{2}\frac{1}{2}} \end{pmatrix}\otimes \begin{pmatrix} \overset{\ast}{Z}^{\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2}} & \overset{\ast}{Z}^{\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2}} \\ \overset{\ast}{Z}^{\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2}} & \overset{\ast}{Z}^{\frac{1}{2}}_{\frac{1}{2}\frac{1}{2}} \end{pmatrix}=\nonumber\\ \ar\begin{pmatrix} \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2};-\frac{1}{2}-\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2};\frac{1}{2}-\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2};-\frac{1}{2}-\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2};\frac{1}{2}-\frac{1}{2}}\\ \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2};-\frac{1}{2}\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2};\frac{1}{2}\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2};-\frac{1}{2}\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2};\frac{1}{2}\frac{1}{2}}\\ \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2};-\frac{1}{2}-\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2};\frac{1}{2}-\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}\frac{1}{2};-\frac{1}{2}-\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}\frac{1}{2};\frac{1}{2}-\frac{1}{2}}\\ \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2};-\frac{1}{2}\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2};\frac{1}{2}\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}\frac{1}{2};-\frac{1}{2}\frac{1}{2}}& \mathfrak{Z}^{\frac{1}{2}\frac{1}{2}}_{\frac{1}{2}\frac{1}{2};\frac{1}{2}\frac{1}{2}} \end{pmatrix}, \label{Tenrep1} \end{gather} \[ T_{1\frac{1}{2}}(\theta,\tau)= \ar\begin{pmatrix} Z^1_{-1-1} & Z^1_{-10} & Z^1_{-11}\\ Z^1_{0-1} & Z^1_{00} & Z^1_{01}\\ Z^1_{1-1} & Z^1_{10} & Z^1_{11} \end{pmatrix}\otimes \begin{pmatrix} \overset{\ast}{Z}^{\frac{1}{2}}_{-\frac{1}{2}-\frac{1}{2}} & \overset{\ast}{Z}^{\frac{1}{2}}_{\frac{1}{2}-\frac{1}{2}} \\ \overset{\ast}{Z}^{\frac{1}{2}}_{-\frac{1}{2}\frac{1}{2}} & \overset{\ast}{Z}^{\frac{1}{2}}_{\frac{1}{2}\frac{1}{2}} \end{pmatrix}= \] \begin{equation}\label{Tenrep2} \ar\begin{pmatrix} \mathfrak{Z}^{1\frac{1}{2}}_{-1-1;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-1-1;\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-10;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-10;\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-11;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-11;\frac{1}{2}-\frac{1}{2}} \\ \mathfrak{Z}^{1\frac{1}{2}}_{-1-1;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-1-1;\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-10;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-10;\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-11;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{-11;\frac{1}{2}\frac{1}{2}} \\ \mathfrak{Z}^{1\frac{1}{2}}_{0-1;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{0-1;\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{00;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{00;\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{01;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{01;\frac{1}{2}-\frac{1}{2}} \\ \mathfrak{Z}^{1\frac{1}{2}}_{0-1;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{0-1;\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{00;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{00;\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{01;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{01;\frac{1}{2}\frac{1}{2}} \\ \mathfrak{Z}^{1\frac{1}{2}}_{1-1;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{1-1;\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{10;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{10;\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{11;-\frac{1}{2}-\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{11;\frac{1}{2}-\frac{1}{2}} \\ \mathfrak{Z}^{1\frac{1}{2}}_{1-1;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{1-1;\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{10;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{10;\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{11;-\frac{1}{2}\frac{1}{2}} & \mathfrak{Z}^{1\frac{1}{2}}_{11;\frac{1}{2}\frac{1}{2}} \end{pmatrix}. \end{equation} \subsection{Recurrence relations between the functions $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\cos\theta^c,\cos\dot{\theta}^c)$} Between the hyperspherical functions $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}$ there exists a wide variety of recurrence relations. The part of them relates the hyperspherical functions of one and the same order (with identical $l$ and $\dot{l}$), other part relates the functions of different orders. In virtue of the representation (\ref{Waerden}), the recurrence formulae for the hyperspherical functions of one and the same order follow from the equalities \begin{eqnarray} {\sf X}^{l\dot{l}}_-\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}=\boldsymbol{\alpha}_n \mathfrak{M}^{l\dot{l}}_{m,n-1;\dot{m}\dot{n}},&\quad& {\sf X}^{l\dot{l}}_+\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}=\boldsymbol{\alpha}_{n+1} \mathfrak{M}^{l\dot{l}}_{m,n+1;\dot{m}\dot{n}},\label{XG}\\ {\sf Y}^{l\dot{l}}_-\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}=\boldsymbol{\alpha}_{\dot{n}} \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}-1},&\quad& {\sf Y}^{l\dot{l}}_+\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= \boldsymbol{\alpha}_{\dot{n}+1} \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}+1}.\label{YG} \end{eqnarray} As it is shown in \cite{Var03}, the generators (\ref{SL26}) can be expressed via the complex Euler angles as \begin{eqnarray} {\sf X}_+&=&\frac{e^{-i\psi^c}}{2}\left[i\frac{\partial}{\partial\theta}- \frac{1}{\sin\theta^c}\frac{\partial}{\partial\varphi} +\cot\theta^c\frac{\partial}{\partial\psi}-\frac{\partial}{\partial\tau}- \frac{i}{\sin\theta^c}\frac{\partial}{\partial\epsilon}+ i\cot\theta^c\frac{\partial}{\partial\varepsilon}\right],\label{SL27}\\ {\sf X}_-&=&\frac{e^{i\psi^c}}{2}\left[i\frac{\partial}{\partial\theta}+ \frac{1}{\sin\theta^c}\frac{\partial}{\partial\varphi} -\cot\theta^c\frac{\partial}{\partial\psi}-\frac{\partial}{\partial\tau}+ \frac{i}{\sin\theta^c}\frac{\partial}{\partial\epsilon}- i\cot\theta^c\frac{\partial}{\partial\varepsilon}\right],\label{SL28}\\ {\sf Y}_+&=&\frac{e^{-i\dot{\psi}^c}}{2}\left[i\frac{\partial}{\partial\theta}- \frac{1}{\sin\dot{\theta}^c}\frac{\partial}{\partial\varphi}+ \cot\dot{\theta}^c\frac{\partial}{\partial\psi}+\frac{\partial}{\partial\tau}+ \frac{i}{\sin\dot{\theta}^c}\frac{\partial}{\partial\epsilon}- i\cot\dot{\theta}^c\frac{\partial}{\partial\varepsilon}\right],\label{SL29}\\ {\sf Y}_-&=&\frac{e^{i\dot{\psi}^c}}{2}\left[i\frac{\partial}{\partial\theta}+ \frac{1}{\sin\dot{\theta}^c}\frac{\partial}{\partial\varphi} -\cot\dot{\theta}^c\frac{\partial}{\partial\psi}+\frac{\partial}{\partial\tau}- \frac{i}{\sin\dot{\theta}^c}\frac{\partial}{\partial\epsilon}+ i\cot\dot{\theta}^c\frac{\partial}{\partial\varepsilon}\right].\label{SL30} \end{eqnarray} Substituting the function $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= e^{-m(\epsilon+i\varphi)-n(\varepsilon+i\psi)} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau) e^{-\dot{m}(\epsilon-i\varpi)-\dot{n}(\varepsilon-i\psi)}$ into the relations (\ref{XG}) and taking into account the operators (\ref{SL27}) and (\ref{SL28}), we find that \begin{eqnarray} i\frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\tau}- \frac{2i(m-n\cos\theta^c)}{\sin\theta^c}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_n\mathfrak{Z}^{l\dot{l}}_{m,n-1;\dot{m}\dot{n}},\nonumber\\ i\frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(m-n\cos\theta^c)}{\sin\theta^c}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{n+1}\mathfrak{Z}^{l\dot{l}}_{m,n+1;\dot{m}\dot{n}}.\nonumber \end{eqnarray} Since the functions $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}$ are symmetric, that is, $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= \mathfrak{Z}^{l\dot{l}}_{nm;\dot{n}\dot{m}}$, then substituting $\mathfrak{Z}^{l\dot{l}}_{nm;\dot{n}\dot{m}}$ instead of $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}$ and replacing $m$ by $n$, and $n$ by $m$, we obtain \begin{eqnarray} i\frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\tau}- \frac{2i(n-m\cos\theta^c)}{\sin\theta^c}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_m\mathfrak{Z}^{l\dot{l}}_{m-1,n;\dot{m}\dot{n}},\label{RZ1}\\ i\frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(n-m\cos\theta^c)}{\sin\theta^c}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{m+1}\mathfrak{Z}^{l\dot{l}}_{m+1,n;\dot{m}\dot{n}}.\label{RZ2} \end{eqnarray} Analogously, from the relations (\ref{YG}) and generators (\ref{SL29})-- (\ref{SL30}), we have \begin{eqnarray} i\frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\theta}+ \frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(\dot{n}-\dot{m}\cos\dot{\theta}^c)}{\sin\dot{\theta}^c} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{\dot{m}}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}-1,\dot{n}}, \label{RZ3}\\ i\frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\theta}+ \frac{\partial\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}}{\partial\tau}- \frac{2i(\dot{n}-\dot{m}\cos\dot{\theta}^c)}{\sin\dot{\theta}^c} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{\dot{m}+1}\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}+1,\dot{n}}. \label{RZ4} \end{eqnarray} Further, for the conjugate representations we have \begin{eqnarray} \widetilde{{\sf X}}^{l\dot{l}}_-\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}} =\boldsymbol{\alpha}_n\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{m,n-1;\dot{m}\dot{n}},&\quad& \widetilde{{\sf X}}^{l\dot{l}}_+\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}} =\boldsymbol{\alpha}_{n+1}\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{m,n+1;\dot{m}\dot{n}}, \label{CXG}\\ \widetilde{{\sf Y}}^{l\dot{l}}_-\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}} =\boldsymbol{\alpha}_{\dot{n}}\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{m,n-1;\dot{m}\dot{n}}, &\quad& \widetilde{{\sf Y}}^{l\dot{l}}_+\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}} =\boldsymbol{\alpha}_{\dot{n}+1}\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{m,n+1;\dot{m}\dot{n}}, \label{CYG} \end{eqnarray} where $\widetilde{{\sf X}}^{l\dot{l}}_-\sim{\sf Y}^{l\dot{l}}_-$, $\widetilde{{\sf X}}^{l\dot{l}}_+\sim{\sf Y}^{l\dot{l}}_+$, $\widetilde{{\sf Y}}^{l\dot{l}}_-\sim{\sf X}^{l\dot{l}}_-$, $\widetilde{{\sf Y}}^{l\dot{l}}_+\sim{\sf X}^{l\dot{l}}_+$. Substituting now the function $\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= e^{-m(\epsilon-i\varphi)-n(\varepsilon-i\psi)} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau) e^{-\dot{m}(\epsilon+i\varphi)-\dot{n}(\varepsilon+i\psi)}$ into the relations (\ref{CXG})--(\ref{CYG}), we obtain \begin{eqnarray} i\frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\tau}+ \frac{2i(n-m\cos\dot{\theta}^c)}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{m} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{m-1,n;\dot{m}\dot{n}}, \label{RZ5}\\ i\frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\tau}- \frac{2i(n-m\cos\dot{\theta}^c)}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{m+1} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{m+1,n;\dot{m}\dot{n}}. \label{RZ6} \end{eqnarray} \begin{eqnarray} i\frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\tau}- \frac{2i(\dot{n}-\dot{m}\cos\theta^c)}{\sin\theta^c} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{\dot{m}} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}-1,\dot{n}}, \label{RZ7}\\ i\frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}} {\partial\tau}+ \frac{2i(\dot{n}-\dot{m}\cos\theta^c)}{\sin\theta^c} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}= 2\boldsymbol{\alpha}_{\dot{m}+1} \overset{\ast}{\mathfrak{Z}}{}^{l\dot{l}}_{mn;\dot{m}+1,\dot{n}}. \label{RZ8} \end{eqnarray} \section{Fields on the Poincar\'{e} group} Fields on the Poincar\'{e} group present itself a natural generalization of the concept of wave function. These fields (generalized wave functions) were introduced independently by several authors \cite{GT47,BW48,Yuk50,Shi51} mainly in connection with constructing relativistic wave equations. The following logical step was done by Finkelstein \cite{Fin55}, he suggested to consider the wave function depending both the coordinates on the Minkowski spacetime and some continuous variables corresponding to spin degrees of freedom (internal space). In essence, this generalization consists in replacing the Minkowski space by a larger space on which the Poincar\'{e} group acts. If this action is to be transitive, one is lead to consider the homogeneous spaces of the Poincar\'{e} group. All the homogeneous spaces of this type were listed by Finkelstein \cite{Fin55} and by Bacry and Kihlberg \cite{BK69} and the fields on these spaces were considered in the works \cite{Lur64,BN67,NB67,Kih68,Kih70,Tol96,GS01}. A homogeneous space ${\cal M}$ of a group $G$ has the following properties:\\ a) It is a topological space on which the group $G$ acts continuously, that is, let $y$ be a point in ${\cal M}$, then $gy$ is defined and is again a point in ${\cal M}$ ($g\in G$).\\ b) This action is transitive, that is, for any two points $y_1$ and $y_2$ in ${\cal M}$ it is always possible to find a group element $g\in G$ such that $y_2=gy_1$.\\ There is a one-to-one correspondence between the homogeneous spaces of $G$ and the coset spaces of $G$. Let $H_0$ be a maximal subgroup of $G$ which leaves the point $y_0$ invariant, $gy_0=y_0$, $g\in H_0$, then $H_0$ is called the stabilizer of $y_0$. Representing now any group element of $G$ in the form $g=g_cg_0$, where $g_0\in H_0$ and $g_c\in G/H_0$, we see that, by virtie of the transitivity property, any point $y\in{\cal M}$ can be given by $y=g_cg_0y_0=g_cy$. Hence it follows that the elements $g_c$ of the coset space give a parametrization of ${\cal M}$. The mapping ${\cal M}\leftrightarrow G/H_0$ is continuous since the group multiplication is continuous and the action on ${\cal M}$ is continuous by definition. The stabilizers $H$ and $H_0$ of two different points $y$ and $y_0$ are conjugate, since from $H_0g_0=g_0$, $y_0=g^{-1}y$, it follows that $gH_0g^{-1}y=y$, that is, $H=gH_0g^{-1}$. Coming back to the Poincar\'{e} group ${\cal P}$, we see that the enumeration of the different homogeneous spaces ${\cal M}$ of ${\cal P}$ amounts to an enumeration of the subgroups of ${\cal P}$ up to a conjugation. Following to Finkelstein, we require that ${\cal M}$ always contains the Minkowski space $\R^{1,3}$ which means that four parameters of ${\cal M}$ can be denoted by $x\,(x^\mu)$. This means that the stabilizer $H$ of a given point in ${\cal M}$ can never contain an element of the translation subgroup of ${\cal P}$. Thus, the stabilizer must be a subgroup of the proper Lorentz group $\mathfrak{G}_+$. In such a way, studying different subgroups of $\mathfrak{G}_+$, we obtain a full list of homogeneous spaces ${\cal M}={\cal P}/H$ of the Poincar\'{e} group. In the present paper we restrict ourselves by a consideration of the following two homogeneous spaces: \begin{eqnarray} {\cal M}_{10}&=&\R^{1,3}\times\mathfrak{L}_6,\quad H=0;\nonumber\\ {\cal M}_8&=&\R^{1,3}\times\dS^2,\quad H=\Omega^c_\psi;\nonumber \end{eqnarray} Hence it follows that a group manifold of the Poincar\'{e} group, ${\cal M}_{10}=\R^{1,3}\times\mathfrak{L}_6$, is a maximal homogeneous space of ${\cal P}$, $\mathfrak{L}_6$ is a group manifold of the Lorentz group. The fields on the manifold ${\cal M}_{10}$ were considered by Lur\c{c}at \cite{Lur64}. These fields depend on all the ten parameters of ${\cal P}$: \[ \boldsymbol{\psi}(x,\mathfrak{g})=\psi(x)\psi(\mathfrak{g})=\psi(x_0,x_1,x_2,x_3) \psi(\mathfrak{g}_1,\mathfrak{g}_2,\mathfrak{g}_3,\mathfrak{g}_4,\mathfrak{g}_5,\mathfrak{g}_6), \] where an explicit form of $\psi(x)$ is given by the exponentials, and the functions $\psi(\mathfrak{g})$ are expressed via the generalized hyperspherical functions $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})$ in the case of finite dimensional representations. The following eight-dimensional homogeneous space ${\cal M}_8=\R^{1,3}\times\dS^2$ is a direct product of the Minkowski space $\R^{1,3}$ and the complex two-sphere $\dS^2$. In this case the stabilizer $H$ consists of the subgroup $\Omega^c_\psi$ of the diagonal matrices $\begin{pmatrix} e^{\frac{i\psi^c}{2}} & 0\\ 0 & e^{-\frac{i\psi^c}{2}} \end{pmatrix}$. Bacry and Kihlberg \cite{BK69} claimed that the space ${\cal M}_8$ is the most suitable for a description of both half-integer and integer spins. The fields, defined in ${\cal M}_8$, depend on the eight parameters of ${\cal P}$: \begin{equation}\label{WF} \boldsymbol{\psi}(x,\varphi^c,\theta^c)=\psi(x)\psi(\varphi^c,\theta^c)= \psi(x_0,x_1,x_2,x_3)\psi(\varphi,\epsilon,\theta,\tau), \end{equation} where the functions $\psi(\varphi^c,\theta^c)$ are expressed via the associated hyperspherical functions defined on the surface of the complex two-sphere $\dS^2$. \subsection{Harmonic analysis on $SU(2)\otimes SU(2)\odot T_4$} In this subsection we will consider Fourier series on the Poincar\'{e} group ${\cal P}$. First of all, the group ${\cal P}$ has the same number of connected components as with the Lorentz group. Later on we will consider only the component ${\cal P}^\uparrow_+$ corresponding the connected component $L^\uparrow_+$ (so called special Lorentz group, see \cite{RF}). As is known, an universal covering $\overline{{\cal P}^\uparrow_+}$ of the group ${\cal P}^\uparrow_+$ is defined by a semidirect product $\overline{{\cal P}^\uparrow_+}=SL(2,\hbox{\bb C})\odot T_4\simeq\spin_+(1,3)\odot T_4$, where $T_4$ is a subgroup of four-dimensional translations. Since the Poincar\'{e} group is a 10-parameter group, then an invariant measure on this group has a form \[ d^{10}\boldsymbol{\alpha}=d^6\mathfrak{g} d^4x, \] where $d^6\mathfrak{g}$ is the Haar measure on the Lorentz group. Or, taking into account (\ref{HA1}), we obtain \begin{equation}\label{HMP} d\boldsymbol{\alpha}=\sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon dx_1 dx_2 dx_3 dx_4, \end{equation} where $x_i\in T_4$. Thus, an invariant integration on the group $SL(2,\hbox{\bb C})\odot T_4$ is defined by the formula \[ \int\limits_{SL(2,\hbox{\bb C})\odot T_4}f(\boldsymbol{\alpha})d^{10}\boldsymbol{\alpha}= \int\limits_{SL(2,\hbox{\bb C})}\int\limits_{T_4}f(x,\mathfrak{g})d^4xd^6\mathfrak{g}, \] where $f(\boldsymbol{\alpha})$ is a finite continuous function on $SL(2,\hbox{\bb C})\odot T_4$. In the case of finite-dimensional representations we come again to a local isomorphism $SL(2,\hbox{\bb C})\odot T_4\simeq SU(2)\otimes SU(2)\odot T_4$. In this case basis representation functions of the Poincar\'{e} group are defined by symmetric polynomials of the form \begin{gather} p(x,z,\bar{z})=\sum_{\substack{(\alpha_1,\ldots,\alpha_k)\\ (\dot{\alpha}_1,\ldots,\dot{\alpha}_r)}}\frac{1}{k!\,r!} a^{\alpha_1\cdots\alpha_k\dot{\alpha}_1\cdots\dot{\alpha}_r}(x) z_{\alpha_1}\cdots z_{\alpha_k}\bar{z}_{\dot{\alpha}_1}\cdots \bar{z}_{\dot{\alpha}_r},\label{SF2}\\ (\alpha_i,\dot{\alpha}_i=0,1)\nonumber \end{gather} where the coefficients $a^{\alpha_1\cdots\alpha_k\dot{\alpha}_1\cdots\dot{\alpha}_r}$ depend on the variables $x^\alpha$ ($\alpha=0,1,2,3$) (the parameters of $T_4$). The functions (\ref{SF2}) should be considered as {\it the functions on the Poincar\'{e} group}.\index{function!on the Poincar\'{e} group} Some applications of these functions contained in \cite{GS01,Vas96}. The group $T_4$ is an Abelian group formed by a direct product of the four one-dimensional translation groups, $T_1$, where $T_1$ is isomorphic to an additive group of real numbers $\R$ (usual Fourier analysis is formulated in terms of the group $\R$). Hence it follows that all irreducible representations of $T_4$ are one-dimensional and expressed via the exponentials. Thus, the basis functions (matrix elements) of the finite-dimensional representations of ${\cal P}$ have the form \begin{equation}\label{BFP} t^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\boldsymbol{\alpha})= e^{-ipx}\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g}), \end{equation} where $x=(x_1,x_2,x_3,x_4)$, and $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})$ is the generalized hyperspherical function (\ref{HF3'}). Let us consider now the configuration space ${\cal M}_8=\R^{1,3}\times \dS^2$. In this case the Fourier series on ${\cal M}_8$ can be defined as follows \begin{equation}\label{FM8} f(\boldsymbol{\alpha})=\sum^{+\infty}_{p=-\infty}e^{ipx}\sum^\infty_{l=0} \sum^l_{m=-l}\sum^{\dot{l}}_{\dot{m}=-\dot{l}} \alpha^{l\dot{l}}_{m\dot{m}}\mathfrak{M}^{m\dot{m}}_{l\dot{l}} (\varphi,\epsilon,\theta,\tau,0,0), \end{equation} where \[ \alpha^{l\dot{l}}_{m\dot{m}}=\frac{(-1)^m(2l+1)(2\dot{l}+1)}{32\pi^4} \int\limits_{\dS^2}\int\limits_{T_4}f(\boldsymbol{\alpha})e^{-ipx} \mathfrak{M}^{m\dot{m}}_{l\dot{l}}(\varphi,\epsilon,\theta,\tau,0,0)d^4xd^4\mathfrak{g}, \] and $d^4\mathfrak{g}=\sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon$ is the Haar measure on $\dS^2$, $f(\boldsymbol{\alpha})$ is the square integrable function on ${\cal M}_8$, such that \[ \int\limits_{\dS^2}\int\limits_{T_4}|f(\boldsymbol{\alpha})|^2d^4xd^4\mathfrak{g}<+\infty. \] \subsection{Harmonic analysis on the group $SL(2,\hbox{\bb C})$} First of all, on the group $SL(2,\hbox{\bb C})$ there exists an invariant measure $d\mathfrak{g}$, that is, such a measure that for any finite continuous function $f(\mathfrak{g})$ on $SL(2,\hbox{\bb C})$ the following equality \[ \int f(\mathfrak{g})d\mathfrak{g}=\int f(\mathfrak{g}_0\mathfrak{g})d\mathfrak{g}=\int f(\mathfrak{g}\fg_0)d\mathfrak{g}=\int f(\mathfrak{g}^{-1})d\mathfrak{g} \] holds. Now we express the Haar measure (left or right) in terms of the parameters (\ref{CEA}), \begin{equation}\label{HA1} d\mathfrak{g}=\sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon. \end{equation} Thus, an invariant integration on the group $SL(2,\hbox{\bb C})$ is defined by the formula \[ \int\limits_{SL(2,\hbox{\bb C})}f(g)d\mathfrak{g}=\frac{1}{32\pi^4} \int\limits^{+\infty}_{-\infty} \int\limits^{+\infty}_{-\infty} \int\limits^{+\infty}_{-\infty} \int\limits^{2\pi}_{-2\pi} \int\limits^{2\pi}_0 \int\limits^\pi_0 f(\theta,\varphi,\psi,\tau,\epsilon,\varepsilon) \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon. \] When we consider finite-dimensional (spinor) representations of $SL(2,\hbox{\bb C})$, we come naturally to a local isomorphism $SU(2)\otimes SU(2)\simeq SL(2,\hbox{\bb C})$ considered by many authors \cite{HS70,Ryd85}. Since a dimension of the spinor representation $\boldsymbol{\tau}_{l\dot{l}}$ of $SU(2)\otimes SU(2)$ is equal to $(2l+1)(2\dot{l}+1)$, then the functions $\sqrt{(2l+1)(2\dot{l}+1)}t^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})$ form a full orthogonal normalized system on this group with respect to the invariant measure $d\mathfrak{g}$. At this point, the indices $l$ and $\dot{l}$ run all possible integer or half-integer non-negative values, and the indices $m$, $n$ and $\dot{m}$, $\dot{n}$ run the values $-l,-l+1,\ldots,l-1,l$ and $-\dot{l},-\dot{l}+1,\ldots,\dot{l}-1,\dot{l}$. In virtue of (\ref{HF3'}) the matrix elements $t^{l\dot{l}}_{mn;\dot{m}\dot{n}}$ are expressed via the generalized hyperspherical function $t^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})= \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon, \theta,\tau,\psi,\varepsilon)$. Therefore, \begin{equation}\label{HA2} \int\limits_{SU(2)\otimes SU(2)}\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g}) \overline{\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})}d\mathfrak{g}= \frac{32\pi^4}{(2l+1)(2\dot{l}+1)} \delta(\mathfrak{g}^\prime-\mathfrak{g}), \end{equation} where $\delta(\mathfrak{g}^\prime-\mathfrak{g})$ is a $\delta$-function on the group $SU(2)\otimes SU(2)$. An explicit form of $\delta$-function is \begin{multline} \delta(\mathfrak{g}^\prime-\mathfrak{g})=\delta(\varphi^\prime-\varphi) \delta(\epsilon^\prime-\epsilon)\delta(\cos\theta^\prime\cosh\tau^\prime- \cos\theta\cosh\tau)\times\\ \times\delta(\sin\theta^\prime\sinh\tau^\prime-\sin\theta\sinh\tau) \delta(\psi^\prime-\psi)\delta(\varepsilon^\prime-\varepsilon).\nonumber \end{multline} Substituting into (\ref{HA2}) the expression \[ \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\mathfrak{g})=e^{-m(\epsilon+i\varphi)- n(\varepsilon+i\psi)} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau)e^{-\dot{m}(\epsilon-i\varphi)- \dot{n}(\varepsilon-i\psi)} \] and taking into account (\ref{HA1}), we obtain \begin{multline} \int\limits_{SU(2)\otimes SU(2)} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau) \overline{\mathfrak{Z}^{s\dot{s}}_{pq;\dot{p}\dot{q}}(\theta,\tau)} e^{-(m+p)\epsilon}e^{-i(m-p)\varphi}e^{-(\dot{m}+\dot{p})\epsilon} e^{i(\dot{m}-\dot{p})\varphi} \times e^{-(n+q)\varepsilon}\times\\ \times e^{-i(n-q)\psi}e^{-(\dot{n}+\dot{q})\varepsilon} e^{i(\dot{n}-\dot{q})\psi} \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon=\frac{32\pi^4\delta_{ls}\delta_{\dot{l}\dot{s}} \delta_{mp}\delta_{nq}\delta_{\dot{m}\dot{p}}\delta_{\dot{n}\dot{q}} \delta(\mathfrak{g}^\prime-\mathfrak{g})}{(2l+1)(2\dot{l}+1)}.\nonumber \end{multline} Thus, any square integrable function $f(\varphi^c,\theta^c,\psi^c)$ on the group $SU(2)\otimes SU(2)$, such that \[ \int\limits_{SU(2)\otimes SU(2)} |f(\varphi^c,\theta^c,\psi^c)|^2 \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon <+\infty, \] is expanded into a convergent (on an average) Fourier series on $SU(2)\otimes SU(2)$, \begin{multline} f(\varphi^c,\theta^c,\psi^c)=\sum^\infty_{l,\dot{l}=0} \sum^l_{m=-l}\sum^l_{n=-l}\sum^{\dot{l}}_{\dot{m}=-\dot{l}} \sum^{\dot{l}}_{\dot{n}=-\dot{l}} \alpha^{l\dot{l}}_{mn;\dot{m}\dot{n}}\times\\ \times e^{-m(\epsilon+i\varphi)-n(\varepsilon+i\psi)} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\cos\theta^c,\cos\dot{\theta}^c) e^{-\dot{m}(\epsilon-i\varphi)-\dot{n}(\varepsilon-i\psi)}, \label{CFS} \end{multline} where \begin{multline} \alpha^{l\dot{l}}_{mn;\dot{m}\dot{n}}= \frac{(-1)^{m-n}(2l+1)(2\dot{l}+1)}{32\pi^4}\times\\ \int\limits_{SU(2)\otimes SU(2)} f(\varphi^c,\theta^c,\psi^c)e^{i(m\varphi^c+n\psi^c)} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\cos\theta^c,\cos\dot{\theta}^c) e^{-i(\dot{m}\dot{\phi}^c+\dot{n}\dot{\psi}^c)} \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon. \nonumber \end{multline} The Parseval equality for the case of $SU(2)\otimes SU(2)$ is defined as follows \begin{multline} \sum^\infty_{l,\dot{l}=0}\sum^l_{m=-l}\sum^l_{n=-l} \sum^{\dot{l}}_{\dot{m}=-\dot{l}}\sum^{\dot{l}}_{\dot{n}=-\dot{l}} |\alpha^{l\dot{l}}_{mn;\dot{m}\dot{n}}|^2=\\ =\frac{(2l+1)(2\dot{l}+1)}{32\pi^4} \int\limits_{SU(2)\otimes SU(2)} |f(\varphi^c,\theta^c,\psi^c)|^2 \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\psi d\tau d\epsilon d\varepsilon. \nonumber \end{multline} About convergence of Fourier series of the type (\ref{CFS}) see \cite{BD70}. In like manner we can define Fourier series on the two-dimensional complex sphere via the associated hyperspherical functions. An expansion of the functions on the surface of the two-dimensional sphere has an important meaning for the subsequent physical applications. So, let $f(\varphi^c,\theta^c)$ be a function on the complex two-sphere $\dS^2$, such that \[ \int\limits_{\dS^2} |f(\varphi^c,\theta^c)|^2 \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon <+\infty, \] then $f(\varphi^c,\theta^c)$ is expanded into a convergent Fourier series on $\dS^2$, \[ f(\varphi^c,\theta^c)=\sum^\infty_{l,\dot{l}=0}\sum^l_{m=-l} \sum^{\dot{l}}_{\dot{m}=-\dot{l}} \alpha^{l\dot{l}}_{m\dot{m}}e^{-m(\epsilon+i\varphi)} \mathfrak{Z}^{m\dot{m}}_{l\dot{l}}(\cos\theta^c,\cos\dot{\theta}^c) e^{-\dot{m}(\epsilon-i\varphi)}, \] where \[ \alpha^{l\dot{l}}_{m\dot{m}}=\frac{(-1)^m(2l+1)(2\dot{l}+1)}{32\pi^4} \int\limits_{\dS^2} f(\varphi^c,\theta^c)e^{im\varphi^c} \mathfrak{Z}^{m\dot{m}}_{l\dot{l}}(\cos\theta^c,\cos\dot{\theta}^c) e^{-i\dot{m}\dot{\varphi}^c} \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon, \] and $\mathfrak{Z}^{m\dot{m}}_{l\dot{l}}(\cos\theta^c,\cos\dot{\theta}^c)$ is an associated hyperspherical function, $d\mathfrak{g}=\sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon$ is the Haar measure on the sphere $\dS^2$. Correspondingly, the Parseval equality on $\dS^2$ has the form \[ \sum^\infty_{l,\dot{l}=0}\sum^l_{m=-l}\sum^{\dot{l}}_{\dot{m}=-\dot{l}} |\alpha^{l\dot{l}}_{m\dot{m}}|^2= \frac{(2l+1)(2\dot{l}+1)}{32\pi^4} \int\limits_{\dS^2} |f(\varphi^c,\theta^c)|^2 \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon. \] \section{Lagrangian formalism and field equations on the\protect\newline Poincar\'{e} group} We will start with a more general homogeneous space of the group ${\cal P}$, ${\cal M}_{10}=\R^{1,3}\times\mathfrak{L}_6$ (group manifold of the Poincar\'{e} group). Let $\mathcal{L}(\boldsymbol{\alpha})$ be a Lagrangian on the group manifold ${\cal M}_{10}$ (in other words, $\mathcal{L}(\boldsymbol{\alpha})$ is a 10-dimensional point function), where $\boldsymbol{\alpha}$ is the parameter set of this group. Then an integral for $\mathcal{L}(\boldsymbol{\alpha})$ on some 10-dimensional volume $\Omega$ of the group manifold we will call {\it an action on the Poincar\'{e} group}: \[ A=\int\limits_\Omega d\boldsymbol{\alpha}\mathcal{L}(\boldsymbol{\alpha}), \] where $d\boldsymbol{\alpha}$ is a Haar measure on the group ${\cal P}$ (see (\ref{HMP})). Let $\boldsymbol{\psi}(\boldsymbol{\alpha})$ be a function on the group manifold ${\cal M}_{10}$ (now it is sufficient to assume that $\boldsymbol{\psi}(\boldsymbol{\alpha})$ is a square integrable function on the Poincar\'{e} group) and let \begin{equation}\label{ELE} \frac{\partial\mathcal{L}}{\partial\boldsymbol{\psi}}-\frac{\partial}{\partial\boldsymbol{\alpha}} \frac{\partial\mathcal{L}}{\partial\frac{\partial\boldsymbol{\psi}}{\partial\boldsymbol{\alpha}}}=0 \end{equation} be Euler-Lagrange equations on ${\cal M}_{10}$ (more precisely speaking, the equations (\ref{ELE}) act on the tangent bundle $T{\cal M}_{10}=\underset{\boldsymbol{\alpha}\in{\cal M}_{10}}{\cup}T_{\boldsymbol{\alpha}}{\cal M}_{10}$ of the manifold ${\cal M}_{10}$, see \cite{Arn}). Let us introduce a Lagrangian $\mathcal{L}(\boldsymbol{\alpha})$ depending on the field function $\boldsymbol{\psi}(\boldsymbol{\alpha})$ as follows \[ \mathcal{L}(\boldsymbol{\alpha})=-\frac{1}{2}\left(\boldsymbol{\psi}^\ast(\boldsymbol{\alpha})B_\mu \frac{\partial\boldsymbol{\psi}(\boldsymbol{\alpha})}{\partial\boldsymbol{\alpha}_\mu}- \frac{\partial\boldsymbol{\psi}^\ast(\boldsymbol{\alpha})} {\partial\boldsymbol{\alpha}_\mu}B_\mu\boldsymbol{\psi}(\boldsymbol{\alpha})\right) -\kappa\boldsymbol{\psi}^\ast(\boldsymbol{\alpha})B_{11}\boldsymbol{\psi}(\boldsymbol{\alpha}), \] where $B_\nu$ ($\nu=1,2,\ldots,10$) are square matrices. The number of rows and columns in these matrices is equal to the number of components of $\boldsymbol{\psi}(\boldsymbol{\alpha})$, $\kappa$ is a non-null real constant. Further, if $B_{11}$ is non-singular, then we can introduce the matrices \[ \Gamma_\mu=B^{-1}_{11}B_\mu,\quad \mu=1,2,\ldots,10, \] and represent the Lagrangian $\mathcal{L}(\boldsymbol{\alpha})$ in the form \begin{equation}\label{Lagrange} \mathcal{L}(\boldsymbol{\alpha})=-\frac{1}{2}\left(\overline{\boldsymbol{\psi}}(\boldsymbol{\alpha})\Gamma_\mu \frac{\partial\boldsymbol{\psi}(\boldsymbol{\alpha})}{\partial\boldsymbol{\alpha}_\mu}- \frac{\overline{\boldsymbol{\psi}}(\boldsymbol{\alpha})}{\partial\boldsymbol{\alpha}_\mu}\Gamma_\mu \boldsymbol{\psi}(\boldsymbol{\alpha})\right)- \kappa\overline{\boldsymbol{\psi}}(\boldsymbol{\alpha})\boldsymbol{\psi}(\boldsymbol{\alpha}), \end{equation} where \[ \overline{\boldsymbol{\psi}}(\boldsymbol{\alpha})=\boldsymbol{\psi}^\ast(\boldsymbol{\alpha})B_{11}. \] Varying independently $\psi(x)$ and $\overline{\psi}(x)$, we obtain from (\ref{Lagrange}) in accordance with (\ref{ELE}) the following equations: \begin{equation}\label{FET} \begin{array}{ccc} \Gamma_i\dfrac{\partial\psi(x)}{\partial x_i}+\kappa\psi(x)&=&0,\\ \Gamma^T_i\dfrac{\partial\overline{\psi}(x)}{\partial x_i}- \kappa\overline{\psi}(x)&=&0. \end{array}\quad(i=1,\ldots,4) \end{equation} Analogously, varying independently $\psi(\mathfrak{g})$ and $\overline{\psi}(\mathfrak{g})$ one gets \begin{equation}\label{FEL} \begin{array}{ccc} \Gamma_k\dfrac{\partial\psi(\mathfrak{g})}{\partial\mathfrak{g}_k}+\kappa\psi(\mathfrak{g})&=&0,\\ \Gamma^T_k\dfrac{\overline{\psi}(\mathfrak{g})}{\partial\mathfrak{g}_k}- \kappa\overline{\psi}(\mathfrak{g})&=&0, \end{array}\quad(k=1,\ldots,6) \end{equation} where \[ \psi(\mathfrak{g})=\begin{pmatrix} \psi(\mathfrak{g})\\ \dot{\psi}(\mathfrak{g}) \end{pmatrix},\quad \Gamma_k=\begin{pmatrix} 0 & \overset{\ast}{\Lambda}{}^{l\dot{l}}_k\\ \Lambda^{l\dot{l}}_k & 0 \end{pmatrix}. \] The doubling of representations, described by a bispinor $\psi(\mathfrak{g})=(\psi(\mathfrak{g}),\dot{\psi}(\mathfrak{g}))^T$, is the well known feature of the Lorentz group representations \cite{GMS,Nai58}. Since an universal covering $SL(2,\hbox{\bb C})$ of the proper orthochronous Lorentz group is a complexification of the group $SU(2)$ (see the section \ref{SL2C}), then it is more convenient to express six parameters $\mathfrak{g}_k$ of the Lorentz group via three parameters $a_1$, $a_2$, $a_3$ of the group $SU(2)$. It is obvious that $\mathfrak{g}_1=a_1$, $\mathfrak{g}_2=a_2$, $\mathfrak{g}_3=a_3$, $\mathfrak{g}_4=ia_1$, $\mathfrak{g}_5=ia_2$, $\mathfrak{g}_6=ia_3$. Then the first equation from (\ref{FEL}) can be written as \begin{eqnarray} \sum^3_{j=1}\Lambda^{l\dot{l}}_j\frac{\partial\psi}{\partial a_j}- i\sum^3_{j=1}\Lambda^{l\dot{l}}_j\frac{\partial\psi}{\partial a^\ast_j}+ \dot{k}^c\dot{\psi}&=&0,\nonumber\\ \sum^3_{j=1}\overset{\ast}{\Lambda}{}^{l\dot{l}}_j\frac{\partial\dot{\psi}} {\partial\widetilde{a}_j}+i\sum^3_{j=1} \overset{\ast}{\Lambda}{}^{l\dot{l}}_j\frac{\partial\dot{\psi}} {\partial\widetilde{a^\ast}_j}+k^c\psi&=&0.\label{CT} \end{eqnarray} where $a^\ast_1=-i\mathfrak{g}_4$, $a^\ast_2=-i\mathfrak{g}_5$, $a^\ast_3=-i\mathfrak{g}_6$, and $\widetilde{a}_j$, $\widetilde{a}^\ast_j$ are the parameters corresponding the dual basis. In essence, the equations (\ref{CT}) are defined in a three-dimensional complex space $\hbox{\bb C}^3$. In turn, the space $\hbox{\bb C}^3$ is isometric to a 6-dimensional bivector space $\R^6$ (a parameter space of the Lorentz group \cite{Kag26,Pet69}). The bivector space $\R^6$ is a tangent space of the group manifold $\mathfrak{L}_6$ of the Lorentz group, that is, the manifold $\mathfrak{L}_6$ in the each its point is equivalent locally to the space $\R^6$. Thus, for all $\mathfrak{g}\in\mathfrak{L}_6$ we have $T_{\mathfrak{g}}\mathfrak{L}_6\simeq\R^6$. There exists a close relationship between the metrics of the Minkowski spacetime $\R^{1,3}$ and the metrics of $\R^6$ defined by the formulae (see \cite{Pet69}) \begin{equation}\label{Metric} g_{ab}\longrightarrow g_{\alpha\beta\gamma\delta}\equiv g_{\alpha\gamma}g_{\beta\delta}-g_{\alpha\delta}g_{\beta\gamma}, \end{equation} where $g_{\alpha\beta}$ is a metric tensor of the spacetime $\R^{1,3}$, and collective indices are skewsymmetric pairs $\alpha\beta\rightarrow a$, $\gamma\delta\rightarrow b$. In more detail, if \[\ar g_{\alpha\beta}=\begin{pmatrix} -1 & 0 & 0 & 0\\ 0 & -1& 0 & 0\\ 0 & 0 & -1& 0\\ 0 & 0 & 0 & 1 \end{pmatrix}, \] then in virtue of (\ref{Metric}) for the metric tensor of $\R^6$ we obtain \begin{equation}\label{MetB} g_{ab}=\ar\begin{pmatrix} -1& 0 & 0 & 0 & 0 & 0\\ 0 & -1& 0 & 0 & 0 & 0\\ 0 & 0 & -1& 0 & 0 & 0\\ 0 & 0 & 0 & 1 & 0 & 0\\ 0 & 0 & 0 & 0 & 1 & 0\\ 0 & 0 & 0 & 0 & 0 & 1 \end{pmatrix}, \end{equation} where the order of collective indices in $\R^6$ is $23\rightarrow 0$, $10\rightarrow 1$, $20\rightarrow 2$, $30\rightarrow 3$, $31\rightarrow 4$, $12\rightarrow 5$. Let us write an invariance condition for the system (\ref{CT}). As it is shown in \cite{Kag26}, the Lorentz transformations can be represented by linear transformations of the space $\R^6$. Let $\mathfrak{g}: a^\prime=\mathfrak{g}^{-1} a$ be a transformation of the bivector space $\R^6$, that is, $a^\prime=\sum^6_{b=1}g_{ba} a_b$, where $a=(a_1,a_2,a_3,a^\ast_1,a^\ast_2,a^\ast_3)$ and $g_{ba}$ is the metric tensor (\ref{MetB}). We can write the tensor (\ref{MetB}) in the form $g_{ab}=\ar\begin{pmatrix} g^-_{ik} & \\ & g^+_{ik} \end{pmatrix}$, then $a^\prime=\sum^3_{k=1}g^-_{ki}a_k$, $a^\ast{}^\prime=\sum^3_{k=1}g^+_{ki}a^\ast_k$. Replacing $\psi$ via $T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime$, and differentiation on $a_k$ ($a^\ast_k$) by differentiation on $a^\prime_k$ ($a^\ast_k{}^\prime$) via the formulae \[ \frac{\partial}{\partial a_k}=\sum g^-_{ik} \frac{\partial}{\partial a^\prime_i},\quad \frac{\partial}{\partial a^\ast_k}=\sum g^+_{ik} \frac{\partial}{\partial a^\ast_i{}^\prime}, \] we obtain \begin{multline} \sum^3_{i=1}\left[ g^-_{i1}\Lambda^{l\dot{l}}_1\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime)} {\partial a^\prime_i}+ g^-_{i2}\Lambda^{l\dot{l}}_2\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime)} {\partial a^\prime_i}+ g^-_{i3}\Lambda^{l\dot{l}}_3\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime)} {\partial a^\prime_i}- \right.\\ \shoveright{\left. -ig^-_{i1}\Lambda^{l\dot{l}}_1\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime)} {\partial\overset{\ast}{a}_i{}^\prime}- ig^-_{i2}\Lambda^{l\dot{l}}_2\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime)} {\partial\overset{\ast}{a}_i{}^\prime}- ig^-_{i3}\Lambda^{l\dot{l}}_3\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime)} {\partial\overset{\ast}{a}_i{}^\prime}\right]+ \dot{\kappa}^cT^{-1}_{l\dot{l}}(\mathfrak{g})\psi^\prime=0,}\\ \shoveleft{ \sum^3_{i=1}\left[ g^+_{i1}\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \frac{\partial(\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \dot{\psi}^\prime)}{\partial\widetilde{a}^\prime_i}+ g^+_{i2}\overset{\ast}{\Lambda}{}^{l\dot{l}}_2 \frac{\partial(\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \dot{\psi}^\prime)}{\partial\widetilde{a}^\prime_i}+ g^+_{i3}\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \frac{\partial(\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \dot{\psi}^\prime)}{\partial\widetilde{a}^\prime_i}+\right.}\nonumber\\ \left. +ig^+_{i1}\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \frac{\partial(\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \dot{\psi}^\prime)}{\partial\widetilde{a^\ast}_i{}^\prime}+ ig^+_{i2}\overset{\ast}{\Lambda}{}^{l\dot{l}}_2 \frac{\partial(\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \dot{\psi}^\prime)}{\partial\widetilde{a^\ast}_i{}^\prime}+ ig^+_{i3}\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \frac{\partial(\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \dot{\psi}^\prime)}{\partial\widetilde{a^\ast}_i{}^\prime}\right]+ \kappa^c\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g})\psi^\prime=0.\nonumber \end{multline} Or, \begin{multline} \sum_i\left[ g^-_{i1}\Lambda^{l\dot{l}}_1T^{-1}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\psi^\prime}{\partial a^\prime_i}+ g^-_{i2}\Lambda^{l\dot{l}}_2T^{-1}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\psi^\prime}{\partial a^\prime_i}+ g^-_{i3}\Lambda^{l\dot{l}}_3T^{-1}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\psi^\prime}{\partial a^\prime_i}- \right.\\ \shoveright{\left. -ig^-_{i1}\Lambda^{l\dot{l}}_1T^{-1}_{l\dot{l}}(\mathfrak{g})\frac{\partial\psi^\prime} {\partial a^\ast_i{}^\prime}- ig^-_{i2}\Lambda^{l\dot{l}}_2T^{-1}_{l\dot{l}}(\mathfrak{g})\frac{\partial\psi^\prime} {\partial a^\ast_i{}^\prime}- ig^-_{i3}\Lambda^{l\dot{l}}_3T^{-1}_{l\dot{l}}(\mathfrak{g})\frac{\partial\psi^\prime} {\partial a^\ast_i{}^\prime}\right]+ \dot{\kappa}^cT^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\psi}^\prime=0,}\\ \shoveleft{ \sum_i\left[ g^+_{i1}\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \frac{\partial\dot{\psi}^\prime}{\partial\widetilde{a}^\prime_i}+ g^+_{i2}\overset{\ast}{\Lambda}{}^{l\dot{l}}_2 \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \frac{\partial\dot{\psi}^\prime}{\partial\widetilde{a}^\prime_i}+ g^+_{i3}\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \frac{\partial\dot{\psi}^\prime}{\partial\widetilde{a}^\prime_i}+\right.}\\ \left. +ig^+_{i1}\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \frac{\partial\dot{\psi}^\prime}{\partial\widetilde{a^\ast}_i{}^\prime}+ ig^+_{i2}\overset{\ast}{\Lambda}{}^{l\dot{l}}_2 \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \frac{\partial\dot{\psi}^\prime}{\partial\widetilde{a^\ast}_i{}^\prime}+ ig^+_{i3}\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g}) \frac{\partial\dot{\psi}^\prime}{\partial\widetilde{a^\ast}_i{}^\prime}\right]+ \kappa^c\overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g})\psi^\prime=0.\nonumber \end{multline} For coincidence of the latter system with (\ref{CT}) we must multiply this system by $T_{l\dot{l}}(\mathfrak{g})$ ($\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})$) from the left, \begin{eqnarray} \sum_i\sum_k g^-_{ik}T_{l\dot{l}}(\mathfrak{g})\Lambda^{l\dot{l}}_kT^{-1}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\psi^\prime}{\partial a^\prime_i} -i\sum_i\sum_k g^-_{ik}T_{l\dot{l}}(\mathfrak{g})\Lambda^{l\dot{l}}_kT^{-1}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\psi^\prime}{\partial a^\ast_i{}^\prime}+ \dot{\kappa}^c\dot{\psi}^\prime&=&0,\nonumber\\ \sum_i\sum_k g^+_{ik}\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \overset{\ast}{\Lambda}{}^{l\dot{l}}_k \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g})\frac{\partial\dot{\psi}^\prime} {\partial\widetilde{a}^\prime_i}+ i\sum_i\sum_k g^+_{ik}\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \overset{\ast}{\Lambda}{}^{l\dot{l}}_k \overset{\ast}{T}_{l\dot{l}}\!\!{}^{-1}(\mathfrak{g})\frac{\partial\dot{\psi}^\prime} {\partial\widetilde{a^\ast}_i{}^\prime}+ \kappa^c\psi^\prime&=&0.\nonumber \end{eqnarray} The requirement of invariance means that for any transformation $\mathfrak{g}$ between the matrices $\Lambda^{l\dot{l}}_k$ ($\overset{\ast}{\Lambda}{}^{l\dot{l}}_k$) the following relations hold: \begin{eqnarray} \sum_k g^-_{ik}T_{l\dot{l}}(\mathfrak{g})\Lambda^{l\dot{l}}_k T^{-1}_{l\dot{l}}(\mathfrak{g})&=&\Lambda^{l\dot{l}}_i,\nonumber\\ \sum_k g^+_{ik}\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \overset{\ast}{\Lambda}{}^{l\dot{l}}_k \overset{\ast}{T}_{l\dot{l}}\!\!\!{}^{-1}(\mathfrak{g}) &=&\overset{\ast}{\Lambda}{}^{l\dot{l}}_i,\label{IC} \end{eqnarray} where $\overset{\ast}{\Lambda}{}^{l\dot{l}}_i$ are the matrices of the equations in the dual representation space, $\kappa^c$ is a complex number, $\partial/\partial\widetilde{a}_i$ mean covariant derivatives in the dual space. \section{The structure of the matrices $\Lambda^{l\dot{l}}_i$} First of all, let us find commutation relations between the matrices $\Lambda^{l\dot{l}}_i$, $\overset{\ast}{\Lambda}{}^{l\dot{l}}_i$ and infinitesimal operators (\ref{Op1})--(\ref{Op12}) defined in the helicity basis. Let us represent transformations $T_{l\dot{l}}(\mathfrak{g})$ ($\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})$) in the infinitesimal form, \begin{eqnarray} {\sf I}+{\sf A}^{l\dot{l}}_i\xi+\ldots, &&\quad{\sf I}+{\sf B}^{l\dot{l}}_i\xi+\ldots, \nonumber\\ {\sf I}+\widetilde{{\sf A}}^{l\dot{l}}_i\xi+\ldots, &&\quad {\sf I}+ \widetilde{{\sf B}}^{l\dot{l}}_i\xi+\ldots. \nonumber \end{eqnarray} It is easy to see that the bivector space $\R^6$ contains two three-dimensional subspaces $\R^3_-$ and $\R^3_+$ with the metric tensors $g^-_{ik}$ and $g^+_{ik}$, respectively. Let us consider a rotation $\mathfrak{g}=e+a_1\xi+\ldots$ in the subspace $\R^3_-$. The matrix of this rotation can be represented in the form \[ \begin{pmatrix} 1 & 0 & 0\\ 0 & 1 &-\xi\\ 0 &\xi& 1 \end{pmatrix}. \] Substituting these transformations into invariance conditions (\ref{IC}), we obtain with an accuracy of the terms of second order the following three equalities: \begin{eqnarray} ({\sf I}+{\sf A}^{l\dot{l}}_1\xi)\Lambda^{l\dot{l}}_1({\sf I}-{\sf A}^{l\dot{l}}_1\xi)&=& \Lambda^{l\dot{l}}_1,\nonumber\\ ({\sf I}+{\sf A}^{l\dot{l}}_1\xi)(\Lambda^{l\dot{l}}_2-\xi\Lambda^{l\dot{l}}_3) ({\sf I}-{\sf A}^{l\dot{l}}_1\xi)&=&\Lambda^{l\dot{l}}_2,\nonumber\\ ({\sf I}+{\sf A}^{l\dot{l}}_1\xi)(\xi\Lambda^{l\dot{l}}_2+\Lambda^{l\dot{l}}_3) ({\sf I}-{\sf A}^{l\dot{l}}_1\xi)&=&\Lambda^{l\dot{l}}_3.\nonumber \end{eqnarray} Hence it follows that \begin{eqnarray} {\sf A}^{l\dot{l}}_1\Lambda^{l\dot{l}}_1-\Lambda^{l\dot{l}}_1{\sf A}^{l\dot{l}}_1&=&0, \nonumber\\ {\sf A}^{l\dot{l}}_1\Lambda^{l\dot{l}}_2-\Lambda^{l\dot{l}}_2{\sf A}^{l\dot{l}}_1- \Lambda^{l\dot{l}}_3&=&0,\nonumber\\ {\sf A}^{l\dot{l}}_1\Lambda^{l\dot{l}}_3-\Lambda^{l\dot{l}}_3{\sf A}^{l\dot{l}}_1+ \Lambda^{l\dot{l}}_2&=&0,\nonumber \end{eqnarray} or \begin{eqnarray} \left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_1\right]&=&0,\nonumber\\ \left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_2\right]&=&\Lambda^{l\dot{l}}_3,\nonumber\\ \left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right]&=&-\Lambda^{l\dot{l}}_2.\nonumber \end{eqnarray} Analogously, for a rotation $\mathfrak{g}=e+a_2\xi+\ldots$ with the matrix \[ \begin{pmatrix} 1 & 0 & \xi\\ 0 & 1 & 0\\ -\xi & 0 & 1 \end{pmatrix} \] we have \begin{eqnarray} (1+{\sf A}^{l\dot{l}}_2\xi)(\Lambda^{l\dot{l}}_1+\xi\Lambda^{l\dot{l}}_3) (1-{\sf A}^{l\dot{l}}_2\xi)&=&\Lambda^{l\dot{l}}_1,\nonumber\\ (1+{\sf A}^{l\dot{l}}_2\xi)\Lambda^{l\dot{l}}_2(1-{\sf A}^{l\dot{l}}_2\xi)&=& \Lambda^{l\dot{l}}_2,\nonumber\\ (1+{\sf A}^{l\dot{l}}_2\xi)(\xi\Lambda^{l\dot{l}}_1+\Lambda^{l\dot{l}}_3) (1-{\sf A}^{l\dot{l}}_2\xi)&=&\Lambda^{l\dot{l}}_3.\nonumber \end{eqnarray} From the latter relations we see that \begin{eqnarray} \left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_1\right]&=&-\Lambda^{l\dot{l}}_3,\nonumber\\ \left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_2\right]&=&0,\nonumber\\ \left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right]&=&\Lambda^{l\dot{l}}_1.\nonumber \end{eqnarray} Further, taking into account all possible transformations (rotations) in the subspaces $\R^3_-$ and $\R^3_+$, we obtain the following commutation relations: \begin{equation}\label{AL} \begin{array}{rcl} \left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_1\right] &=& 0,\\ \left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_1\right] &=&-\Lambda^{l\dot{l}}_3,\\ \left[{\sf A}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_1\right] &=& \Lambda^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_2\right] &=& \Lambda^{l\dot{l}}_3,\\ \left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_2\right] &=& 0,\\ \left[{\sf A}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_2\right] &=&-\Lambda^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right] &=&-\Lambda^{l\dot{l}}_2,\\ \left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right] &=& \Lambda^{l\dot{l}}_1,\\ \left[{\sf A}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_3\right] &=& 0. \end{array} \end{equation} \begin{equation}\label{BL} \begin{array}{rcl} \left[{\sf B}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_1\right] &=& 0,\\ \left[{\sf B}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_1\right] &=&i\Lambda^{l\dot{l}}_3,\\ \left[{\sf B}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_1\right] &=& -i\Lambda^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf B}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_2\right] &=& -i\Lambda^{l\dot{l}}_3,\\ \left[{\sf B}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_2\right] &=& 0,\\ \left[{\sf B}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_2\right] &=&i\Lambda^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf B}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right] &=&i\Lambda^{l\dot{l}}_2,\\ \left[{\sf B}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right] &=& -i\Lambda^{l\dot{l}}_1,\\ \left[{\sf B}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_3\right] &=& 0. \end{array} \end{equation} \begin{equation}\label{DAL'} {\renewcommand{\arraystretch}{1.2} \begin{array}{rcl} \left[\widetilde{{\sf A}}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_1\right] &=& 0,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_1\right] &=& -\Lambda^{l\dot{l}}_3,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_1\right] &=&\Lambda^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf A}}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_2\right] &=&\Lambda^{l\dot{l}}_3,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_2\right] &=& 0,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_2\right] &=& -\Lambda^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf A}}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right] &=& -\Lambda^{l\dot{l}}_2,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right] &=&\Lambda^{l\dot{l}}_1,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_3\right] &=& 0. \end{array}} \end{equation} \begin{equation}\label{DBL'} {\renewcommand{\arraystretch}{1.2} \begin{array}{rcl} \left[\widetilde{{\sf B}}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_1\right] &=& 0,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_1\right] &=&-i\Lambda^{l\dot{l}}_3,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_1\right] &=&i\Lambda^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf B}}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_2\right] &=&i\Lambda^{l\dot{l}}_3,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_2\right] &=& 0,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_2\right] &=&-i\Lambda^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf B}}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right] &=&-i\Lambda^{l\dot{l}}_2,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right] &=&i\Lambda^{l\dot{l}}_1,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_3,\Lambda^{l\dot{l}}_3\right] &=& 0. \end{array}} \end{equation} \begin{equation}\label{DAL} {\renewcommand{\arraystretch}{1.6} \begin{array}{rcl} \left[\widetilde{{\sf A}}^{l\dot{l}}_1,\overset{\ast}{\Lambda}\!{}^{l\dot{l}}_1\right] &=& 0,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_2, \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=& -\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_3, \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=&\overset{\ast}{\Lambda}{}^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf A}}^{l\dot{l}}_1, \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=&\overset{\ast}{\Lambda}^{l\dot{l}}_3,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& 0,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_3, \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& -\overset{\ast}{\Lambda}{}^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf A}}^{l\dot{l}}_1, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& -\overset{\ast}{\Lambda}{}^{l\dot{l}}_2,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_2, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=&\overset{\ast}{\Lambda}{}^{l\dot{l}}_1,\\ \left[\widetilde{{\sf A}}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& 0. \end{array}} \end{equation} \begin{equation}\label{DBL} {\renewcommand{\arraystretch}{1.2} \begin{array}{rcl} \left[\widetilde{{\sf B}}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=& 0,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_2, \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=&-i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_3, \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=&i\overset{\ast}{\Lambda}{}^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf B}}^{l\dot{l}}_1, \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=&i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& 0,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_3, \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=&-i\overset{\ast}{\Lambda}{}^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[\widetilde{{\sf B}}^{l\dot{l}}_1, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=&-i\overset{\ast}{\Lambda}{}^{l\dot{l}}_2,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_2, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=&i\overset{\ast}{\Lambda}{}^{l\dot{l}}_1,\\ \left[\widetilde{{\sf B}}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& 0. \end{array}} \end{equation} \begin{equation}\label{AL'} \begin{array}{rcl} \left[{\sf A}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=& 0,\\ \left[{\sf A}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=&-\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[{\sf A}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=& \overset{\ast}{\Lambda}{}^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf A}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& \overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[{\sf A}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& 0,\\ \left[{\sf A}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=&-\overset{\ast}{\Lambda}{}^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf A}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=&-\overset{\ast}{\Lambda}{}^{l\dot{l}}_2,\\ \left[{\sf A}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& \overset{\ast}{\Lambda}{}^{l\dot{l}}_1,\\ \left[{\sf A}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& 0. \end{array} \end{equation} \begin{equation}\label{BL'} \begin{array}{rcl} \left[{\sf B}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=& 0,\\ \left[{\sf B}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=&i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[{\sf B}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\right] &=& -i\overset{\ast}{\Lambda}{}^{l\dot{l}}_2, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf B}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& -i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[{\sf B}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=& 0,\\ \left[{\sf B}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_2\right] &=&i\overset{\ast}{\Lambda}{}^{l\dot{l}}_1, \end{array}\;\;\; \begin{array}{rcl} \left[{\sf B}^{l\dot{l}}_1,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=&i\overset{\ast}{\Lambda}{}^{l\dot{l}}_2,\\ \left[{\sf B}^{l\dot{l}}_2,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& -i\overset{\ast}{\Lambda}{}^{l\dot{l}}_1,\\ \left[{\sf B}^{l\dot{l}}_3,\overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right] &=& 0. \end{array} \end{equation} \begin{sloppypar} From the latter relations and definition (\ref{SL25}) it immediately follows that commutation relations between $\Lambda^{l\dot{l}}_3$, $\overset{\ast}{\Lambda}{}^{l\dot{l}}_3$ and generators ${\sf Y}^{l\dot{l}}_\pm$, ${\sf Y}^{l\dot{l}}_3$, ${\sf X}^{l\dot{l}}_\pm$, ${\sf X}^{l\dot{l}}_3$ are of the form \end{sloppypar} \begin{equation}\label{LX} {\renewcommand{\arraystretch}{1.5} \left.\begin{array}{l} \left[\ld\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right],{\sf X}^{l\dot{l}}_+\right] = 2\Lambda^{l\dot{l}}_3,\\ \left[\ld\Lambda^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_-\right],{\sf Y}^{l\dot{l}}_+\right]= 2\Lambda^{l\dot{l}}_3,\\ \left[\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_3\right] =0,\\ \left[\Lambda^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_3\right] =0, \end{array}\right. } {\renewcommand{\arraystretch}{1.5} \left.\begin{array}{l} \left[\ld\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_-\right], {\sf Y}^{l\dot{l}}_+\right] =2\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[\ld\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right], {\sf X}^{l\dot{l}}_+\right] =2\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\\ \left[\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_3\right] =0,\\ \left[\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_3\right] =0. \end{array}\right. } \end{equation} Using the relations (\ref{LX}) we will find an explicit form of the matrices $\Lambda^{l\dot{l}}_3$ and $\overset{\ast}{\Lambda}{}^{l\dot{l}}_3$, and further we will find $\Lambda^{l\dot{l}}_1$, $\Lambda^{l\dot{l}}_2$ and $\overset{\ast}{\Lambda}{}^{l\dot{l}}_1$, $\overset{\ast}{\Lambda}{}^{l\dot{l}}_2$. The wave function\index{function!wave} $\boldsymbol{\psi}$ is transformed within some representation $T_{l\dot{l}}(\mathfrak{g})$ of the group $\mathfrak{G}_+$. We assume that $T_{l\dot{l}}(\mathfrak{g})$ is decomposed into irreducible representations. The components of the function $\boldsymbol{\psi}$ we will numerate by the indices $l$, $\dot{l}$ and $m$, $\dot{m}$, where $l$ ($\dot{l}$) is a weight of irreducible representation, $m$ ($\dot{m}$) is a number of the components in the representation of the weight $l$ ($\dot{l}$). In the case when a representation with one and the same weight $l$ ($\dot{l}$) at the decomposition of $\boldsymbol{\psi}$ occurs more than one time, then with the aim to distinguish these representations we will add the index $k$ ($\dot{k}$), which indicates a number of the representations of the weight $l$ ($\dot{l}$). Denoting $\zeta_{lm;\dot{l}\dot{m}}=\mid lm;\dot{l}\dot{m}\rangle$ and coming to the helicity basis,\index{basis!helicity} we obtain a following decomposition for the wave function: \[ \boldsymbol{\psi}(a_1,a_2,a_3,a^\ast_1,a^\ast_2,a^\ast_3)= \sum_{l,m,k,\dot{l},\dot{m},\dot{k}} \psi^{k\dot{k}}_{lm;\dot{l}\dot{m}}(a_1,a_2,a_3,a^\ast_1,a^\ast_2,a^\ast_3) \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}, \] where $a_1,a_2,a_3,a^\ast_1,a^\ast_2,a^\ast_3$ are the coordinates of the complex space $\hbox{\bb C}^3\simeq\R^6$ (parameters of $SL(2,\hbox{\bb C})$)\footnote{Recall that the wave function $\boldsymbol{\psi}(a_j,a^\ast_j)$ is defined on the group manifold $\mathfrak{L}_6$, that is, $\boldsymbol{\psi}$ is a function on the Lorentz group.}. Analogously, for the dual representation we have \[ \dot{\boldsymbol{\psi}}(\widetilde{a}_1,\widetilde{a}_2,\widetilde{a}_3, \widetilde{a^\ast}_1,\widetilde{a^\ast}_2,\widetilde{a^\ast}_3)= \sum_{l,m,k,\dot{l},\dot{m},\dot{k}} \dot{\psi}^{k\dot{k}}_{lm;\dot{l}\dot{m}} (\widetilde{a}_1,\widetilde{a}_2,\widetilde{a}_3,\widetilde{a^\ast}_1, \widetilde{a^\ast}_2,\widetilde{a^\ast}_3) \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}. \] The transformation $\Lambda^{l\dot{l}}_3$ in the helicity basis has the form \[ \Lambda^{l\dot{l}}_3\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= \sum_{l^\prime,m^\prime,k^\prime,\dot{l}^\prime,\dot{m}^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime, \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}. \] Using the commutators (\ref{LX}), we will find the numbers $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime, \dot{l},\dot{m}^\prime\dot{m}}$. First of all, recalling that \begin{eqnarray} {\sf X}^{l\dot{l}}_-\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&\boldsymbol{\alpha}^l_m \zeta^{k\dot{k}}_{l,m-1;\dot{l}\dot{m}},\nonumber\\ {\sf X}^{l\dot{l}}_+\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&\boldsymbol{\alpha}^{l}_{m+1} \zeta^{k\dot{k}}_{l,m+1;\dot{l}\dot{m}},\nonumber\\ {\sf X}^{l\dot{l}}_3\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& m\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}},\nonumber \end{eqnarray} where $(\boldsymbol{\alpha}^l_m)^2=(l+m)(l-m+1)$, we obtain \begin{eqnarray} \Lambda^{l\dot{l}}_3{\sf X}^{l\dot{l}}_3 \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&m\Lambda^{l\dot{l}}_3 \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= m\sum_{l^\prime,m^\prime,k^\prime,\dot{l}^\prime,\dot{m}^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime, \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}, \nonumber\\ {\sf X}^{l\dot{l}}_3\Lambda^{l\dot{l}}_3 \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&{\sf X}^{l\dot{l}}_3 \sum_{l^\prime,m^\prime,k^\prime,\dot{l}^\prime,\dot{m}^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime, \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}= m^\prime\sum_{l^\prime,m^\prime,k^\prime,\dot{l}^\prime,\dot{m}^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime, \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}. \nonumber \end{eqnarray} From the second equation of (\ref{LX}) we have $(\Lambda^{l\dot{l}}_3{\sf X}^{l\dot{l}}_3- {\sf X}^{l\dot{l}}_3\Lambda^{l\dot{l}}_3)\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=0$. Therefore, \[ \sum_{l^\prime,m^\prime,k^\prime,\dot{l}^\prime,\dot{m}^\prime,\dot{k}^\prime} (m-m^\prime) c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime, \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime \dot{m}^\prime}=0. \] Hence it immediately follows that $m^\prime=m$. By this reason we can denote the coefficients $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l, m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}$ via $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l, m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}$. Analogously, from the relations \begin{eqnarray} {\sf Y}^{l\dot{l}}_-\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}} &=&\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\zeta^{k\dot{k}}_{lm;\dot{l},\dot{m}-1}, \nonumber\\ {\sf Y}^{l\dot{l}}_+\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}} &=&\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\zeta^{k\dot{k}}_{lm;\dot{l},\dot{m}+1}, \nonumber\\ {\sf Y}^{l\dot{l}}_3\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \dot{m}\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}\nonumber \end{eqnarray} and equation $(\Lambda^{l\dot{l}}_3{\sf Y}^{l\dot{l}}_3-{\sf Y}^{l\dot{l}}_3 \Lambda^{l\dot{l}}_3)\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=0$ we see that the coefficients $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}$ can be replaced by $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}$. Let us use now the first equation of the system (\ref{LX}): \begin{eqnarray} \Lambda^{l\dot{l}}_3{\sf X}^{l\dot{l}}_- \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&\Lambda^{l\dot{l}}_3\boldsymbol{\alpha}^l_m \zeta^{k\dot{k}}_{l,m-1;\dot{l}\dot{m}}= \boldsymbol{\alpha}^l_m\sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime\dot{l} \dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m-1;\dot{l}^\prime\dot{m}}, \nonumber\\ {\sf X}^{l\dot{l}}_-\Lambda^{l\dot{l}}_3 \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&{\sf X}^{l\dot{l}}_- \sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l},\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m;\dot{l}^\prime\dot{m}}= \sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} \boldsymbol{\alpha}^{l^\prime}_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l},\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m-1;\dot{l}^\prime\dot{m}}, \nonumber\\ \left[\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right] \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} \left[\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime\dot{l},\dot{m}}- \boldsymbol{\alpha}^{l^\prime}_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}\right] \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m-1;\dot{l}^\prime\dot{m}}.\nonumber \end{eqnarray} Further, \begin{eqnarray} \left[\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right]{\sf X}^{l\dot{l}}_+ \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \left[\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right] \boldsymbol{\alpha}^l_{m+1}\zeta^{k\dot{k}}_{l,m+1;\dot{l}\dot{m}}= \nonumber\\ &=&\boldsymbol{\alpha}^l_{m+1} \sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime}\left[ \boldsymbol{\alpha}^l_{m+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l},\dot{m}}- \boldsymbol{\alpha}^{l^\prime}_{m+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m+1;\dot{l}^\prime\dot{l}, \dot{m}}\right] \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m;\dot{l}^\prime\dot{m}},\nonumber\\ {\sf X}^{l\dot{l}}_+\left[\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right] \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& {\sf X}^{l\dot{l}}_+ \sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} \left[\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime\dot{l}, \dot{m}}-\boldsymbol{\alpha}^{l^\prime}_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}\right] \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m-1;\dot{l}^\prime\dot{m}}=\nonumber\\ &=&\sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} \boldsymbol{\alpha}^{l^\prime}_m \left[\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime\dot{l}, \dot{m}}- \boldsymbol{\alpha}^{l^\prime}_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}\right] \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m;\dot{l}^\prime\dot{m}},\nonumber \end{eqnarray} \begin{multline} \left[\ld\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right],{\sf X}^{l\dot{l}}_+\right] \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=\\ =\sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} \left\{ \left[(\boldsymbol{\alpha}^l_{m+1})^2+ (\boldsymbol{\alpha}^{l^\prime}_m)^2\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}- \boldsymbol{\alpha}^{l^\prime}_m\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime \dot{l},\dot{m}}-\right.\\ -\left.\boldsymbol{\alpha}^{l^\prime}_{m+1}\boldsymbol{\alpha}^l_{m+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m+1;\dot{l}^\prime\dot{l}, \dot{m}} \right\} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m;\dot{l}^\prime\dot{m}}. \nonumber \end{multline} Thus, the first commutator from (\ref{LX}) gives a system of equations with respect to the coefficients $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}$: \[ 2c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}= \left[(\boldsymbol{\alpha}^l_{m+1})^2+ (\boldsymbol{\alpha}^{l^\prime}_m)^2\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}- \boldsymbol{\alpha}^{l^\prime}_m\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime \dot{l},\dot{m}} -\boldsymbol{\alpha}^{l^\prime}_{m+1}\boldsymbol{\alpha}^l_{m+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m+1;\dot{l}^\prime\dot{l}, \dot{m}}. \] Or, substituting instead $\boldsymbol{\alpha}^l_m$ their values, we obtain \begin{multline} 2c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l, m;\dot{l}^\prime\dot{l}, \dot{m}} =\left[(l+m+1)(l-m)+(l^\prime+m)(l^\prime-m+1)\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}-\\ -\sqrt{(l^\prime+m)(l^\prime-m+1)(l+m)(l-m+1)} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m-1;\dot{l}^\prime\dot{l}, \dot{m}}-\\ -\sqrt{(l^\prime+m+1)(l^\prime-m)(l+m+1)(l-m)} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m+1;\dot{l}^\prime\dot{l}, \dot{m}}.\label{Sys} \end{multline} This system can be solved at the fixed indices $l^\prime$, $l$, $\dot{l}^\prime$, $\dot{l}$, $k^\prime$, $k$, $\dot{k}^\prime$, $\dot{k}$. Let us fix some indices $l^\prime$, $l$, $\dot{l}^\prime$, $\dot{l}$, $k^\prime$, $k$, $\dot{k}^\prime$, $\dot{k}$ and denote $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}$ via $c_m$. Then we obtain a system of homogeneous equations for $c_m$, where $-\min(l^\prime,l)\leq m \leq \min(l^\prime,l)$. We solve these equations using the Gauss method. When $m$ has the value $m_0=\min(l^\prime,l)$, we obtain an equation containing two unknown variables $c_{m_0}$ and $c_{m_0-1}$, from which $c_{m_0-1}$ is defined via $c_{m_0}$. Further, when $m$ has a value $m_0-1$, we obtain an equation with $c_{m_0-2}$, $c_{m_0-1}$, $c_{m_0}$, from which we can define $c_{m_0-2}$ via $c_{m_0}$ again. In doing so, we see that the coefficients $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}$ different from zero when $|l^\prime-l|\leq 1$, that is, at $l^\prime=l$, $l^\prime=l-1$ and $l^\prime=l+1$. For other values of $l^\prime$ the coefficients $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}$ are equal to zero. First, we take $l^\prime=l$ ($\dot{l}^\prime$, $\dot{l}$, $k^\prime$, $k$, $\dot{k}^\prime$, $\dot{k}$ are arbitrary), then the equations (\ref{Sys}) are rewritten as follows \begin{multline} \left[ 2-(l+m+1)(l-m)-(l-m)(l-m+1)\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m;\dot{l}^\prime\dot{l},\dot{m}}+\\ +(l+m)(l-m+1) c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m-1;\dot{l}^\prime\dot{l},\dot{m}}+ (l+m+1)(l-m) c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m+1;\dot{l}^\prime\dot{l},\dot{m}}=0. \nonumber \end{multline} Supposing $m=l$, we find that $(1-l)c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,l;\dot{l}^\prime\dot{l},\dot{m}}+ lc^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,l-1;\dot{l}^\prime\dot{l},\dot{m}}=0$. Whence $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,l;\dot{l}^\prime\dot{l},\dot{m}}= c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}^\prime\dot{l},\dot{m}}\cdot l$; $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,l-1;\dot{l}^\prime\dot{l},\dot{m}} =c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}^\prime\dot{l},\dot{m}}(l-1)$, where the constant $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}^\prime\dot{l},\dot{m}}$ does not depend on $m$. Supposing $m=l-1$, we find analogously that $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,l-2;\dot{l}^\prime\dot{l},\dot{m}}= c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}^\prime\dot{l},\dot{m}}(l-2)$. It is easy to verify that for any $m$ there is an equality \[ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m;\dot{l}^\prime\dot{l},\dot{m}}= c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}^\prime\dot{l},\dot{m}}\cdot m. \] Let us suppose now $l^\prime=l-1$, then the equations (\ref{Sys}) take the form \begin{multline} \left[ 2-(l+m+1)(l-m)-(l+m-1)(l-m)\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime\dot{l},\dot{m}}+\\ +\sqrt{(l+m-1)(l-m)(l+m)(l-m+1)} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1;\dot{l}^\prime\dot{l},\dot{m}}+\\ +\sqrt{(l+m)(l-m-1)(l+m+1)(l-m)} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1;\dot{l}^\prime\dot{l},\dot{m}} =0.\nonumber \end{multline} Making in these equations the substitutions \begin{eqnarray} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime\dot{l},\dot{m}}&=& \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime \dot{l},\dot{m}}\sqrt{(l+m)(l-m)},\nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1;\dot{l}^\prime\dot{l},\dot{m}} &=&\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1;\dot{l}^\prime \dot{l},\dot{m}}\sqrt{(l+m-1)(l-m+1)},\nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1;\dot{l}^\prime\dot{l},\dot{m}} &=&\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1;\dot{l}^\prime \dot{l},\dot{m}}\sqrt{(l+m+1)(l-m-1)},\nonumber \end{eqnarray} we obtain \begin{multline} \left[ 2-(l+m+1)(l-m)-(l+m-1)(l-m)\right] \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime \dot{l},\dot{m}}\sqrt{(l+m)(l-m)}+\\ +(l+m-1)(l-m+1) \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1;\dot{l}^\prime \dot{l},\dot{m}}\sqrt{(l-m)(l+m)}+\\ +(l-m-1)(l+m+1) \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1;\dot{l}^\prime \dot{l},\dot{m}}\sqrt{(l+m)(l-m)}=0. \nonumber \end{multline} Whence \begin{multline} 2\left[ 1-l^2+m^2\right]\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m; \dot{l}^\prime\dot{l},\dot{m}}+ \left[ l^2-(m-1)^2\right]\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1; \dot{l}^\prime\dot{l},\dot{m}}+\\ +\left[ l^2-(m+1)^2\right]\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1; \dot{l}^\prime\dot{l},\dot{m}}=0.\nonumber \end{multline} It is easy to verify that this system can be solved at $\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime \dot{l},\dot{m}}$ (this coefficient does not depend on $m$). For that reason we can suppose $\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime \dot{l},\dot{m}}=c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}^\prime \dot{l},\dot{m}}$. Coming back to the old variables, we find that \[ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime\dot{l},\dot{m}}= c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}^\prime\dot{l},\dot{m}} \sqrt{l^2-m^2}. \] Finally, let us suppose $l^\prime=l+1$. In this case the system (\ref{Sys}) takes the form \begin{multline} \left[ 2-(l+m+1)(l-m)-(l+m+1)(l-m+2)\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m;\dot{l}^\prime\dot{l},\dot{m}}+\\ +\sqrt{(l+m+1)(l-m+2)(l+m)(l-m+1)} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1;\dot{l}^\prime\dot{l},\dot{m}}+\\ +\sqrt{(l+m+2)(l-m+1)(l+m+1)(l-m)} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1;\dot{l}^\prime\dot{l},\dot{m}} =0.\nonumber \end{multline} Making the substitutions \begin{eqnarray} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m;\dot{l}^\prime\dot{l},\dot{m}}&=& \sqrt{(l+m)(l-m+1)} \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m;\dot{l}^\prime\dot{l}, \dot{m}},\nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1;\dot{l}^\prime\dot{l},\dot{m}} &=& \sqrt{(l+m)(l-m+2)}\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1; \dot{l}^\prime\dot{l},\dot{m}},\nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1;\dot{l}^\prime\dot{l},\dot{m}} &=& \sqrt{(l+m+2)(l-m)} \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1;\dot{l}^\prime \dot{l},\dot{m}},\nonumber \end{eqnarray} we obtain \begin{multline} \left[ 2-(l+m+1)(l-m)-(l+m+1)(l-m+2)\right]\sqrt{(l+m+1)(l-m+1)} \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m;\dot{l}^\prime \dot{l},\dot{m}}+\\ +(l+m)(l-m+2)\sqrt{(l+m+1)(l-m+1)} \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1;\dot{l}^\prime \dot{l},\dot{m}}+\\ +(l+m+2)(l-m)\sqrt{(l+m+1)(l-m+1)} \widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1;\dot{l}^\prime \dot{l},\dot{m}}=0. \nonumber \end{multline} Whence \begin{multline} 2\left[ m^2-l^2-2l\right]\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m; \dot{l}^\prime\dot{l},\dot{m}}+ \left[ l^2-m^2+2l+2m\right]\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l, m-1;\dot{l}^\prime\dot{l},\dot{m}}+\\ +\left[ l^2-m^2+2l-2m\right]\widetilde{c}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l, m+1;\dot{l}^\prime\dot{l},\dot{m}}=0.\nonumber \end{multline} A solution $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}^\prime\dot{l}, \dot{m}}$ of the latter equation does not depend on $m$ also. Thus, the action of the commutator $\left[\ld\Lambda^{l\dot{l}}_3, {\sf X}^{l\dot{l}}_-\right],{\sf X}^{l\dot{l}}_+\right]$ on the basis vectors $\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}$ gives us the following solutions: \begin{eqnarray} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}^\prime\dot{l},\dot{m}} &=&c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}^\prime\dot{l},\dot{m}} \sqrt{l^2-m^2},\nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m;\dot{l}^\prime\dot{l},\dot{m}}&=& c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}^\prime\dot{l},\dot{m}}m, \nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m;\dot{l}^\prime\dot{l},\dot{m}} &=&c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}^\prime\dot{l},\dot{m}} \sqrt{(l+1)^2-m^2}.\nonumber \end{eqnarray} With the aim to find the final form for non-zero elements of $\Lambda^{l\dot{l}}_3$ we must apply the commutator $\left[\ld\Lambda^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_-\right], {\sf Y}^{l\dot{l}}_+\right]$. In the result we have \begin{multline} \left[\ld\Lambda^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_-\right],{\sf Y}^{l\dot{l}}_+\right] \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=\\ =\sum_{l^\prime,k^\prime,\dot{l}^\prime,\dot{k}^\prime} \left\{ \left[(\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1})^2+ (\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}})^2\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}- \boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}} \boldsymbol{\alpha}^{\dot{l}}_{\dot{m}} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime \dot{l},\dot{m}-1}-\right.\\ -\left.\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}+1} \boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}+1} \right\} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m;\dot{l}^\prime\dot{m}}. \nonumber \end{multline} Hence it follows \[ 2c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}= \left[(\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1})^2+ (\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}})^2\right] c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}}- \boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}} \boldsymbol{\alpha}^{\dot{l}}_{\dot{m}} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime \dot{l},\dot{m}-1} -\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}+1} \boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}^\prime\dot{l}, \dot{m}+1}. \] Solutions of the latter system are derived by means of the analogous calculations presented in the previous case of the commutator $\left[\ld\Lambda^{l\dot{l}}_3,{\sf X}^{l\dot{l}}_-\right],{\sf X}^{l\dot{l}}_+\right]$. They have the form \begin{eqnarray} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime,l,m;\dot{l}-1, \dot{l},\dot{m}} &=&c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}-1,\dot{l}} \sqrt{\dot{l}^2-\dot{m}^2},\nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l},\dot{l},\dot{m}}&=& c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}\dot{l}}\dot{m}, \nonumber\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}+1,\dot{l},\dot{m}} &=&c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m;\dot{l}+1,\dot{l}} \sqrt{(\dot{l}+1)^2-\dot{m}^2}.\nonumber \end{eqnarray} Thus, matrix elements of $\Lambda^{l\dot{l}}_3$ are \begin{equation}\label{L3} {\renewcommand{\arraystretch}{1.5} \Lambda^{l\dot{l}}_3:\quad\left\{\begin{array}{ccc} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m;\dot{l}-1,\dot{l},\dot{m}} &=&c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)},\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m;\dot{l}\dot{l},\dot{m}}&=& c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}}m\dot{m},\\ c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m;\dot{l}+1,\dot{l},\dot{m}} &=&c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}. \end{array}\right.} \end{equation} All other elements of the matrix $\Lambda^{l\dot{l}}_3$ are equal to zero. Let us define now elements of the matrices $\Lambda^{l\dot{l}}_1$ and $\Lambda^{l\dot{l}}_2$. For the transformation $\Lambda^{l\dot{l}}_1$ in the helicity basis we have \[ \Lambda^{l\dot{l}}_1\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}. \] We will find the numbers $a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}$ using the relations $\Lambda^{l\dot{l}}_1=\left[{\sf A}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right]$ (or $\Lambda^{l\dot{l}}_1=i\left[{\sf B}^{l\dot{l}}_2,\Lambda^{l\dot{l}}_3\right]$) and (\ref{Op2}) (or (\ref{Op5})). Indeed, \begin{multline} \Lambda^{l\dot{l}}_1\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= {\sf A}^{l\dot{l}}_2\Lambda^{l\dot{l}}_3\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}- \Lambda^{l\dot{l}}_3{\sf A}^{l\dot{l}}_2\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=\\ ={\sf A}^{l\dot{l}}_2 \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime \dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime \dot{m}^\prime}- \frac{1}{2}\Lambda^{l\dot{l}}_3\left(\boldsymbol{\alpha}^l_m \zeta^{k\dot{k}}_{l,m-1;\dot{l},\dot{m}}- \boldsymbol{\alpha}^l_{m+1}\zeta^{k\dot{k}}_{l,m+1;\dot{l}\dot{m}}\right.-\\ \left.-\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}\zeta^{k\dot{k}}_{l,m;\dot{l},\dot{m}-1}+ +\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1}\zeta^{k\dot{k}}_{l,m;\dot{l},\dot{m}+1}\right)= \frac{1}{2} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime, k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} \left(\boldsymbol{\alpha}^{l^\prime}_{m^\prime} \zeta^{k^\prime}_{l^\prime,m^\prime-1;\dot{l}^\prime\dot{m}^\prime}-\right.\\ -\left.\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} \zeta^{k^\prime}_{l^\prime,m^\prime+1;\dot{l}^\prime\dot{m}^\prime} -\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime;\dot{l}^\prime,\dot{m}^\prime -1}+\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime;\dot{l}^\prime,\dot{m}^\prime +1}\right)-\\ -\frac{1}{2}\boldsymbol{\alpha}^l_m \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime,m-1;\dot{l}^\prime dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}+ \frac{1}{2}\boldsymbol{\alpha}^l_{m+1} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m+1;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime} +\\ +\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}-1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}- \frac{1}{2}\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}+1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}. \nonumber \end{multline} Dividing the first sum on the four and changing the summation index in the each sums obtained, we come to the following expression: \begin{multline} \Lambda^{l\dot{l}}_1\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=\frac{1}{2} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \left(\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime+1,m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}- \boldsymbol{\alpha}^{l^\prime}_{m^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime-1,m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}-\right.\\ -\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime+1,\dot{m}} +\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime-1,\dot{m}} -\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m-1;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}+ \boldsymbol{\alpha}^l_{m+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m+1;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}+\\ +\left.\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime,\dot{m}-1}- \boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime,\dot{m}+1}\right) \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}. \nonumber \end{multline} Therefore, elements of the matrix $\Lambda^{l\dot{l}}_1$ have the form \begin{multline} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}=\frac{1}{2} \left(\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime+1,m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}- \boldsymbol{\alpha}^{l^\prime}_{m^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime-1,m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} -\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime+1,\dot{m}}+\right.\\ +\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime-1,\dot{m}} -\boldsymbol{\alpha}^l_m c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m-1;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}+ \boldsymbol{\alpha}^l_{m+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime, m+1;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}+\\ +\left.\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime,\dot{m}-1}- \boldsymbol{\alpha}^{\dot{l}}_{\dot{m}+1} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime,\dot{m}+1}\right) \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}. \label{AM} \end{multline}\begin{sloppypar}\noindent Since $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}\neq 0$ only at $m^\prime=m$, $\dot{m}^\prime=\dot{m}$ and $l^\prime=l-1,\,l,\,l+1$, $\dot{l}^\prime=\dot{l}-1$, $\dot{l}$, $\dot{l}+1$, then at the fixed indices $m$, $\dot{m}$, $l$, $\dot{l}$ $k^\prime$, $k$, $\dot{k}^\prime$, $\dot{k}$ we have twelve numbers $a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}}$ which are different from zero. Substituting $\boldsymbol{\alpha}^l_m=\sqrt{(l+m)(l-m+1)}$, $\boldsymbol{\alpha}^{\dot{l}}_{\dot{m}}=\sqrt{(\dot{l}+\dot{m})(\dot{l}- \dot{m}+1)}$ into (\ref{AM}) and using $c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime lm;\dot{l}^\prime \dot{l},\dot{m}}$ from (\ref{L3}), we find that \end{sloppypar} \begin{eqnarray} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1,m;\dot{l}-1,\dot{l},\dot{m} \dot{m}}&=& -\frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m-1,m;\dot{l}\dot{l},\dot{m}\dot{m}} &=& \frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} \dot{m}\sqrt{(l+m)(l-m+1)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1,m;\dot{l}+1,\dot{l},\dot{m} \dot{m}}&=& \frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1,m;\dot{l}-1,\dot{l},\dot{m} \dot{m}}&=& \frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m+1,m;\dot{l}\dot{l},\dot{m} \dot{m}}&=& \frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} \dot{m}\sqrt{(l+m+1)(l-m)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1,m;\dot{l}+1,\dot{l},\dot{m} \dot{m}}&=& -\frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m;\dot{l}-1,\dot{l},\dot{m}-1, \dot{m}}&=& \frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m;\dot{l}\dot{l},\dot{m}-1,\dot{m}} &=& -\frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m;\dot{l}+1,\dot{l},\dot{m}-1, \dot{m}}&=& -\frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)},\nonumber \end{eqnarray} \begin{eqnarray} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m;\dot{l}-1,\dot{l},\dot{m}+1, \dot{m}}&=& -\frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m;\dot{l}\dot{l},\dot{m}+1, \dot{m}}&=& -\frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})},\nonumber\\ a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m;\dot{l}+1,\dot{l},\dot{m}+1, \dot{m}}&=& \frac{1}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)}.\label{L1} \end{eqnarray} Let us define elements of the matrix $\Lambda^{l\dot{l}}_2$. From $\Lambda^{l\dot{l}}_2=-\left[{\sf A}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right]$ (or $\Lambda^{l\dot{l}}_2=-i\left[{\sf B}^{l\dot{l}}_1,\Lambda^{l\dot{l}}_3\right]$) and \[ \Lambda^{l\dot{l}}_2\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}\dot{m}} \] and also the relations (\ref{Op1}) (or (\ref{Op4})), (\ref{L3}) it follows that \begin{eqnarray} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1,m;\dot{l}-1,\dot{l},\dot{m} \dot{m}}&=& -\frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m-1,m;\dot{l}\dot{l},\dot{m}\dot{m}} &=& \frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} \dot{m}\sqrt{(l+m)(l-m+1)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1,m;\dot{l}+1,\dot{l},\dot{m} \dot{m}}&=& \frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1,m;\dot{l}-1,\dot{l},\dot{m} \dot{m}}&=& -\frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m+1,m;\dot{l}\dot{l},\dot{m} \dot{m}}&=& -\frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} \dot{m}\sqrt{(l+m+1)(l-m)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1,m;\dot{l}+1,\dot{l},\dot{m} \dot{m}}&=& \frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m;\dot{l}-1,\dot{l},\dot{m}-1, \dot{m}}&=& \frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m;\dot{l}\dot{l},\dot{m}-1,\dot{m}} &=& -\frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m;\dot{l}+1,\dot{l},\dot{m}-1, \dot{m}}&=& -\frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)},\nonumber \end{eqnarray} \begin{eqnarray} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m;\dot{l}-1,\dot{l},\dot{m}+1, \dot{m}}&=& \frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m;\dot{l}\dot{l},\dot{m}+1, \dot{m}}&=& \frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}} m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})},\nonumber\\ b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m;\dot{l}+1,\dot{l},\dot{m}+1, \dot{m}}&=& -\frac{i}{2}c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)}.\label{L2} \end{eqnarray} Coming to the dual representations,\index{representation!dual} we find elements of the matrices $\overset{\ast}{\Lambda}{}^{l\dot{l}}_1$, $\overset{\ast}{\Lambda}{}^{l\dot{l}}_2$ and $\overset{\ast}{\Lambda}{}^{l\dot{l}}_3$. The dual transformations $\overset{\ast}{\Lambda}{}^{l\dot{l}}_i$ in the helicity basis are \begin{eqnarray} \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{lm;\dot{l}^\prime\dot{m}^\prime},\nonumber\\ \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{lm;\dot{l}^\prime\dot{m}^\prime},\nonumber\\ \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{lm;\dot{l}^\prime\dot{m}^\prime}.\nonumber \end{eqnarray} \begin{sloppypar}\noindent Calculating the commutators $\left[\overset{\ast}{\Lambda}{}^{l\dot{l}}_3, {\sf Y}^{l\dot{l}}_3\right]$, $\left[\ld\overset{\ast}{\Lambda}{}^{l\dot{l}}_3,{\sf Y}^{l\dot{l}}_-\right], {\sf Y}^{l\dot{l}}_+\right]$ with respect to the vectors $\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}$, we find elements of the matrix $\overset{\ast}{\Lambda}{}^{l\dot{l}}_3$: \begin{equation}\label{L3'} {\renewcommand{\arraystretch}{1.5} \overset{\ast}{\Lambda}{}^{l\dot{l}}_3:\quad\left\{\begin{array}{ccc} \overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m; \dot{l}-1,\dot{l},\dot{m}} &=&\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)},\\ \overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m; \dot{l}\dot{l},\dot{m}}&=& \overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}m\dot{m},\\ \overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m; \dot{l}+1,\dot{l},\dot{m}} &=&\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}. \end{array}\right.} \end{equation} Using the relations $\overset{\ast}{\Lambda}{}^{l\dot{l}}_1=\left[\widetilde{{\sf A}}^{l\dot{l}}_2, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right]$ (or $\overset{\ast}{\Lambda}{}^{l\dot{l}}_1=-i\left[\widetilde{{\sf B}}^{l\dot{l}}_2, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right]$) and (\ref{Op8}) (or (\ref{Op11})), we find elements $\overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime \dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}$ of the matrix $\overset{\ast}{\Lambda}{}^{l\dot{l}}_1$: \begin{eqnarray} \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1,m; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& \frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}\sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m-1,m; \dot{l}\dot{l},\dot{m}\dot{m}}&=& -\frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}\dot{m}\sqrt{(l+m)(l-m+1)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1,m; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& -\frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}\sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1,m; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& -\frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}\sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m+1,m; \dot{l}\dot{l},\dot{m}\dot{m}}&=& -\frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}\dot{m}\sqrt{(l+m+1)(l-m)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1,m; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& \frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}\sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m; \dot{l}-1,\dot{l},\dot{m}-1,\dot{m}}&=& -\frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}\sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m; \dot{l}\dot{l},\dot{m}-1,\dot{m}}&=& \frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m;\dot{l}+1, \dot{l},\dot{m}-1,\dot{m}}&=& \frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)},\nonumber \end{eqnarray} \begin{eqnarray} \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m;\dot{l}-1, \dot{l},\dot{m}+1,\dot{m}}&=& \frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m; \dot{l}\dot{l},\dot{m}+1,\dot{m}}&=& \frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})},\nonumber\\ \overset{\ast}{a}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m; \dot{l}+1,\dot{l},\dot{m}+1,\dot{m}}&=& -\frac{1}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)}.\label{L1'} \end{eqnarray} Further, from the relations $\overset{\ast}{\Lambda}{}^{l\dot{l}}_2=-\left[\widetilde{{\sf A}}^{l\dot{l}}_1, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right]$ (or $\overset{\ast}{\Lambda}^{l\dot{l}}_2=i\left[\widetilde{{\sf B}}^{l\dot{l}}_1, \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\right]$) we obtain elements $\overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime \dot{k}}_{l^prime l,m^\prime m;\dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}$ of $\overset{\ast}{\Lambda}{}^{l\dot{l}}_2$. All calculations are analogous to the calculations presented for the case of $\Lambda^{l\dot{l}}_i$. In the result we have \end{sloppypar} \begin{eqnarray} \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m-1,m; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& \frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}\sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m-1,m; \dot{l}\dot{l},\dot{m}\dot{m}}&=& -\frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}\dot{m}\sqrt{(l+m)(l-m+1)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m-1,m; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& -\frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}\sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m+1,m; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& \frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}\sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m+1,m; \dot{l}\dot{l},\dot{m}\dot{m}}&=& \frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}\dot{m}\sqrt{(l+m+1)(l-m)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m+1,m; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& -\frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}\sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)},\nonumber \end{eqnarray} \begin{eqnarray} \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m; \dot{l}-1,\dot{l},\dot{m}-1,\dot{m}}&=& -\frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m; \dot{l}\dot{l},\dot{m}-1,\dot{m}}&=& \frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m; \dot{l}+1,\dot{l},\dot{m}-1,\dot{m}}&=& \frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)},\nonumber \end{eqnarray} \begin{eqnarray} \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,m,m; \dot{l}-1,\dot{l},\dot{m}+1,\dot{m}}&=& -\frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}} \sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,m,m;\dot{l} \dot{l},\dot{m}+1,\dot{m}}&=& -\frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})},\nonumber\\ \overset{\ast}{b}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,m,m; \dot{l}+1,\dot{l},\dot{m}+1,\dot{m}}&=& \frac{i}{2}\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}} \sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)}.\label{L2'} \end{eqnarray} In general, the matrix $\Lambda^{l\dot{l}}_3$ must be a reducible representation of the proper Lorentz group $\mathfrak{G}_+$, and can always be written in the form \begin{equation}\label{Decomp} \Lambda^{l\dot{l}}_3=\begin{bmatrix} \Lambda^{l_1\dot{l}_1}_3 & & & &\\ & \Lambda^{l_2\dot{l}_2}_3 & &\text{\huge 0} &\\ & & \Lambda^{l_3\dot{l}_3}_3 & &\\ &\text{\huge 0} & & \ddots &\\ & & & & \Lambda^{l_n\dot{l}_n}_3 \end{bmatrix}, \end{equation} where $\Lambda^{l_i\dot{l}_i}_3$ is a spin block (the matrix $\overset{\ast}{\Lambda}{}^{l\dot{l}}_3$ has the same decompositions). It is obvious that the matrices $\Lambda^{l\dot{l}}_1$, $\Lambda^{l\dot{l}}_2$ and $\overset{\ast}{\Lambda}{}^{l\dot{l}}_1$, $\overset{\ast}{\Lambda}{}^{l\dot{l}}_2$ admit also the decompositions of the type (\ref{Decomp}) by definition. If the spin block $\Lambda^{l_i\dot{l}_i}_3$ has non--null roots, then the particle possesses the spin $s_i=|l_i-\dot{l}_i|$. The spin block $\Lambda^s_3$ in (\ref{Decomp}) consists of the elements $c^s_{\boldsymbol{\tau}\boldsymbol{\tau}^\prime}$, where $\boldsymbol{\tau}_{l_1,\dot{l}_2}$ and $\boldsymbol{\tau}_{l^\prime_1,\dot{l}^\prime_2}$ are interlocking irreducible representations\index{representation!interlocking} of the Lorentz group, that is, such representations, for which $l^\prime_1=l_1\pm\frac{1}{2}$, $\dot{l}^\prime_2=\dot{l}_2\pm\frac{1}{2}$. At this point, the block $\Lambda^s_3$ contains only the elements $c^s_{\boldsymbol{\tau}\boldsymbol{\tau}^\prime}$ corresponding to such interlocking representations $\boldsymbol{\tau}_{l_1,\dot{l}_2}$, $\boldsymbol{\tau}_{l^\prime_1,\dot{l}^\prime_2}$ which satisfy the conditions \[ |l_1-\dot{l}_2|\leq s\leq l_1+\dot{l}_2,\quad |l^\prime_1-\dot{l}^\prime_2|\leq s \leq l^\prime_1+\dot{l}^\prime_2. \] The interlocking irreducible representations of the Lorentz group also called as Bhabha--Gel'fand--Yaglom chains \cite{Bha45,GY48}. Corresponding to the decomposition (\ref{Decomp}), the wave function also decomposes into a direct sum of component wave functions which we write \[ \boldsymbol{\psi}=\psi_{l_1m_1;\dot{l}_1\dot{m}_1}+ \psi_{l_2m_2;\dot{l}_2\dot{m}_2}+\psi_{l_3m_3;\dot{l}_3\dot{m}_3}+\ldots. \] According to a de Broglie theory of fusion \cite{Bro43}, interlocking representations give rise to indecomposable RWE.\index{equation!wave!indecomposable} Otherwise, we have decomposable equations. \index{equation!wave!decomposable} As is known, the indecomposable RWE correspond to composite particles. A relation between indecomposable RWE and composite particles will be studied in a separate work. \section{Separation of variables in RWE} \subsection{Boundary value problem} Following to the classical methods of mathematical physics \cite{CH31}, it is quite natural to set up {\it a boundary value problem for the relativistic wave equations (relativistically invariant system)}. It is well known that all the physically meaningful requirements, which follow from the experience, are contained in the boundary value problem. Let us construct in $\hbox{\bb C}^3$ the two--dimensional complex sphere $\dS^2$ from the quantities $z_k=x_k+iy_k$, $\overset{\ast}{z}_k=x_k-iy_k$ as follows \begin{equation}\label{CS} {\bf z}^2=z^2_1+z^2_2+z^2_3={\bf x}^2-{\bf y}^2+2i{\bf x}{\bf y}=r^2 \end{equation} and its complex conjugate (dual) sphere $\dot{\dS}^2$, \begin{equation}\label{DS} \overset{\ast}{{\bf z}}{}^2=\overset{\ast}{z}_1{}^2+\overset{\ast}{z}_2{}^2+ \overset{\ast}{z}_3{}^2={\bf x}^2-{\bf y}^2-2i{\bf x}{\bf y}=\overset{\ast}{r}{}^2. \end{equation} For more details about the two-dimensional complex sphere see \cite{Hus70,HS70,SH70}. It is well-known that both quantities ${\bf x}^2-{\bf y}^2$, ${\bf x}{\bf y}$ are invariant with respect to the Lorentz transformations, since a surface of the complex sphere is invariant (Casimir operators\index{operator!Casimir} of the Lorentz group are constructed from such quantities, see also (\ref{KO})). Moreover, since the real and imaginary parts of the complex two-sphere transform like the electric and magnetic fields, respectively, the invariance of ${\bf z}^2\sim({\bf E}+i{\bf B})^2$ under proper Lorentz transformations is evident. At this point, the quantities ${\bf x}^2-{\bf y}^2$, ${\bf x}{\bf y}$ are similar to the well known electromagnetic invariants $E^2-B^2$, ${\bf E}{\bf B}$. This intriguing relationship between the Laplace-Beltrami operators (\ref{KO}), Casimir operators of the Lorentz group and electromagnetic invariants $E^2-B^2\sim{\bf x}^2-{\bf y}^2$, ${\bf E}{\bf B}\sim{\bf x}{\bf y}$ leads naturally to a Riemann-Silberstein representation of the electromagnetic field (see, for example, \cite{Web01,Sil07,Bir96}). In other words, the two-dimensional sphere, considered as a homogeneous space of the Poincar\'{e} group, is the most suitable arena for the subsequent investigations in quantum electrodynamics. We will set up a boundary value problem for the two-dimensional complex sphere $\dS^2$ (this problem can be considered as a relativistic generalization of the classical Dirichlet problem for the sphere $S^2$).\\ {\it Let $T$ be an unbounded region in $\hbox{\bb C}^3\simeq\R^6$ and let $\Sigma$ be a surface of the complex two-sphere (correspondingly, $\dot{\Sigma}$, for the dual two-sphere), then it needs to find a function $\boldsymbol{\psi}(\mathfrak{g})=(\psi_{m\dot{m}}(\mathfrak{g}), \dot{\psi}_{m\dot{m}}(\mathfrak{g}))^T$ satisfying the following conditions:\\ 1) $\boldsymbol{\psi}(\mathfrak{g})$ is a solution of the system \begin{eqnarray} \sum^3_{j=1}\Lambda^{l\dot{l}}_j\frac{\partial\psi}{\partial a_j}- i\sum^3_{j=1}\Lambda^{l\dot{l}}_j\frac{\partial\psi}{\partial a^\ast_j}+ \dot{\kappa}^c\dot{\psi}&=&0,\nonumber\\ \sum^3_{j=1}\overset{\ast}{\Lambda}{}^{l\dot{l}}_j\frac{\partial\dot{\psi}} {\partial\widetilde{a}_j}+i\sum^3_{j=1} \overset{\ast}{\Lambda}{}^{l\dot{l}}_j \frac{\partial\dot{\psi}}{\partial\widetilde{a}^\ast_j}+ \kappa^c\psi&=&0,\nonumber \end{eqnarray} in the all region $T$;\\ 2) $\boldsymbol{\psi}(\mathfrak{g})$ is a continuous function (everywhere in $T$), including the surfaces $\Sigma$ and $\dot{\Sigma}$;\\ 3) $\left.\phantom{\frac{x}{x}}\psi_{m\dot{m}}(\mathfrak{g})\right|_\Sigma= F_{m\dot{m}}(\mathfrak{g})$, $\left.\phantom{\frac{x}{x}}\dot{\psi}_{m\dot{m}}(\mathfrak{g})\right|_{\dot{\Sigma}}= \dot{F}_{m\dot{m}}(\mathfrak{g})$, where $F_{m\dot{m}}(\mathfrak{g})$ and $\dot{F}_{m\dot{m}}(\mathfrak{g})$ are square integrable functions defined on the surfaces $\Sigma$ and $\dot{\Sigma}$, respectively.} In particular, boundary conditions can be represented by constants, \[ \left.\phantom{\frac{x}{x}}\psi(\mathfrak{g})\right|_\Sigma=\text{const}=F_0,\quad \left.\phantom{\frac{x}{x}}\dot{\psi}(\mathfrak{g})\right|_{\dot{\Sigma}}=\text{const}= \dot{F}_0. \] It is obvious that an explicit form of the boundary conditions follows from the experience. For example, they can describe a distribution of energy in the experiment. With the aim to solve the boundary value problem we come to the complex Euler angles (\ref{CEA}) and represent the function $\boldsymbol{\psi}(r,\theta^c,\varphi^c)=(\psi_{m\dot{m}}(r,\theta^c,\varphi^c), \dot{\psi}_{m\dot{m}}(r^\ast,\dot{\theta}^c,\dot{\varphi}^c))^T$ in the form of following series: \begin{eqnarray} \psi_{m\dot{m}}(r,\theta^c,\varphi^c)&=&\sum^\infty_{l,\dot{l}=0} \sum_{k,\dot{k}} \boldsymbol{f}_{lmk;\dot{l}\dot{m}\dot{k}}(r)\sum^l_{n=-l} \sum^{\dot{l}}_{\dot{n}=-\dot{l}}\alpha^{m\dot{m}}_{ln;\dot{l}\dot{n}} \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0), \label{Fourier1}\\ \dot{\psi}_{m\dot{m}}(r^\ast,\dot{\theta}^c,\dot{\varphi}^c)&=& \sum^\infty_{l,\dot{l}=0}\sum_{k,\dot{k}} \overset{\ast}{\boldsymbol{f}}_{lmk;\dot{l}\dot{m}\dot{k}}(r^\ast) \sum^l_{n=-l} \sum^{\dot{l}}_{\dot{n}=-\dot{l}} \alpha^{m\dot{m}}_{ln;\dot{l}\dot{n}} \overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi, \epsilon,\theta,\tau,0,0), \label{Fourier2} \end{eqnarray} where \begin{eqnarray} \alpha^{m\dot{m}}_{ln;\dot{l}\dot{n}}&=& \frac{(-1)^n(2l+1)(2\dot{l}+1)}{32\pi^4} \int\limits_{\dS^2} F_{m\dot{m}}(\theta^c,\varphi^c) \mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0) \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon,\nonumber\\ \alpha^{m\dot{m}}_{ln;\dot{l}\dot{n}}&= &\frac{(-1)^{\dot{n}}(2l+1)(2\dot{l}+1)}{32\pi^4} \int\limits_{\dS^2} \dot{F}_{m\dot{m}}(\dot{\theta}^c,\dot{\varphi}^c) \overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0) \sin\theta^c\sin\dot{\theta}^cd\theta d\varphi d\tau d\epsilon,\nonumber \end{eqnarray} \begin{sloppypar}\noindent The indices $k$ and $\dot{k}$ numerate equivalent representations. $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)$ ($\overset{\ast}{\mathfrak{M}}{}^{l\dot{l}}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)$) are hyperspherical functions defined on the surface $\Sigma$ ($\dot{\Sigma}$) of the two-dimensional complex sphere of the radius $r$ ($r^\ast$), $\boldsymbol{f}_{lmk;\dot{l}\dot{m}\dot{k}}(r)$ and $\overset{\ast}{\boldsymbol{f}}_{lmk;\dot{l}\dot{m}\dot{k}}(r^\ast)$ are radial functions. It is easy to see that we come here to the harmonic analysis on the complex two-sphere, since the series (\ref{Fourier1}) and (\ref{Fourier2}) have the structure of the Fourier series on $\dS^2$. \end{sloppypar} Let us introduce now hyperspherical coordinates on the surfaces of the complex and dual spheres, \begin{equation}\label{HC} \begin{array}{ccl} z_1&=&r\sin\theta^c\cos\varphi^c,\\ z_2&=&r\sin\theta^c\sin\varphi^c,\\ z_3&=&r\cos\theta^c, \end{array}\quad \begin{array}{ccl} z^\ast_1&=&r^\ast\sin\dot{\theta}^c\cos\dot{\varphi}^c,\\ z^\ast_2&=&r^\ast\sin\dot{\theta}^c\sin\dot{\varphi}^c,\\ z^\ast_3&=&r^\ast\cos\dot{\theta}^c, \end{array} \end{equation} where $\theta^c$, $\varphi^c$ are the complex Euler angles. Let us show that solutions of the equations (\ref{CT}) can be found in therms of expansions in generalized hyperspherical functions considered in the previous section. With this end in view let us transform the system (\ref{CT}) as follows. First of all, let us define the derivatives $\dfrac{\partial}{\partial a_i}$, $\dfrac{\partial}{\partial a^\ast_i}$ on the surface of the two-dimensional complex sphere (\ref{CS}) and write them in the hyperspherical coordinates (\ref{HC}) as \begin{eqnarray} \frac{\partial}{\partial a_1}&=&-\frac{\sin\varphi^c}{r\sin\theta^c} \frac{\partial}{\partial\varphi}+\frac{\cos\varphi^c\cos\theta^c}{r} \frac{\partial}{\partial\theta}+\cos\varphi^c\sin\theta^c\frac{\partial} {\partial r},\label{CD1}\\ \frac{\partial}{\partial a_2}&=&\frac{\cos\varphi^c}{r\sin\theta^c} \frac{\partial}{\partial\varphi}+\frac{\sin\varphi^c\cos\theta^c}{r} \frac{\partial}{\partial\theta}+\sin\varphi^c\sin\theta^c\frac{\partial} {\partial r},\label{CD2}\\ \frac{\partial}{\partial a_3}&=&-\frac{\sin\theta^c}{r}\frac{\partial} {\partial\theta}+\cos\theta^c\frac{\partial}{\partial r}.\label{CD3} \end{eqnarray} \begin{eqnarray} \frac{\partial}{\partial a^\ast_1}&=&i\frac{\partial}{\partial a_1}= -\frac{\sin\dot{\varphi}^c}{r\sin\dot{\theta}^c}\frac{\partial} {\partial\epsilon}+ \frac{\cos\dot{\varphi}^c\sin\dot{\theta}^c}{r}\frac{\partial}{\partial\tau}+ i\cos\dot{\varphi}^c\sin\dot{\theta}^c\frac{\partial}{\partial r},\label{CD4}\\ \frac{\partial}{\partial a^\ast_2}&=&i\frac{\partial}{\partial a_2}= \frac{\cos\dot{\varphi}^c}{r\sin\dot{\theta}^c} \frac{\partial}{\partial\epsilon}+ \frac{\sin\dot{\varphi}^c\cos\dot{\theta}^c}{r}\frac{\partial}{\partial\tau}+ i\sin\dot{\varphi}^c\sin\dot{\theta}^c\frac{\partial}{\partial r},\label{CD5}\\ \frac{\partial}{\partial a^\ast_3}&=&i\frac{\partial}{\partial a_3}= -\frac{\sin\dot{\theta}^c}{r}\frac{\partial}{\partial\tau}+ i\cos\dot{\theta}^c\frac{\partial}{\partial r}.\label{CD6} \end{eqnarray} Analogously, on the surface of the dual sphere (\ref{DS}) we have \begin{eqnarray} \frac{\partial}{\partial\widetilde{a}_1}&=& -\frac{\sin\varphi^c}{r^\ast\sin\theta^c}\frac{\partial} {\partial\varphi}+\frac{\cos\varphi^c\cos\theta^c}{r^\ast} \frac{\partial}{\partial\theta}+\cos\varphi^c\sin\theta^c \frac{\partial}{\partial r^\ast},\label{CDD1}\\ \frac{\partial}{\partial\widetilde{a}_2}&=& \frac{\cos\varphi^c}{r^\ast\sin\theta^c}\frac{\partial} {\partial\varphi}+\frac{\sin\varphi^c\cos\theta^c}{r^\ast} \frac{\partial}{\partial\theta}+\sin\varphi^c\sin\theta^c \frac{\partial}{\partial r^\ast},\label{CDD2}\\ \frac{\partial}{\partial\widetilde{a}_3}&=& -\frac{\sin\theta^c}{r^\ast}\frac{\partial}{\partial\theta}+ \cos\theta^c\frac{\partial}{\partial r^\ast}.\label{CDD3} \end{eqnarray} \begin{eqnarray} \frac{\partial}{\partial\widetilde{a}^\ast_1}&=& -i\frac{\partial}{\partial\widetilde{a}_1}= \frac{\sin\dot{\varphi}^c}{r^\ast\sin\dot{\theta}^c}\frac{\partial} {\partial\epsilon}-\frac{\cos\dot{\varphi}^c\cos\dot{\theta}^c}{r^\ast} \frac{\partial}{\partial\tau}-i\cos\dot{\varphi}^c\sin\dot{\theta}^c \frac{\partial}{\partial r^\ast},\label{CDD4}\\ \frac{\partial}{\partial\widetilde{a}^\ast_2}&=& -i\frac{\partial}{\partial\widetilde{a}_2}= -\frac{\cos\dot{\varphi}^c}{r^\ast\sin\dot{\theta}^c}\frac{\partial} {\partial\epsilon}-\frac{\sin\dot{\varphi}^c\cos\dot{\theta}^c}{r^\ast} \frac{\partial}{\partial\tau}-i\sin\dot{\varphi}^c\sin\dot{\theta}^c \frac{\partial}{\partial r^\ast},\label{CDD5}\\ \frac{\partial}{\partial\widetilde{a}^\ast_3}&=& -i\frac{\partial}{\partial\widetilde{a}_3}= \frac{\sin\dot{\theta}^c}{r^\ast}\frac{\partial}{\partial\tau} -i\cos\dot{\theta}^c\frac{\partial}{\partial r^\ast}.\label{CDD6} \end{eqnarray} Coming back to the equations (\ref{CT}), we see that the matrices $\Lambda^{l\dot{l}}_j$ and $\overset{\ast}{\Lambda}{}^{l\dot{l}}_j$ inherit their tensor structures from the infinitesimal operators (\ref{TenStr}), \begin{eqnarray} \Lambda^{l\dot{l}}_j&=&\Lambda^l_j\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_j,\nonumber\\ \overset{\ast}{\Lambda}{}^{l\dot{l}}_j&=& \overset{\ast}{\Lambda}{}^{\dot{l}}_j\otimes\boldsymbol{1}_{2l+1}- \boldsymbol{1}_{2\dot{l}+1}\otimes\overset{\ast}{\Lambda}{}^l_j.\nonumber \end{eqnarray} Taking into account the latter expressions, we rewrite the system (\ref{CT}) as follows \begin{eqnarray} \sum^3_{j=1}\left(\Lambda^l_j\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_j\right) \frac{\partial\boldsymbol{\psi}}{\partial a_j}+ i\sum^3_{j=1}\left(\Lambda^l_j\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_j\right) \frac{\partial\boldsymbol{\psi}}{\partial a^\ast_j}+ \kappa^c\boldsymbol{\psi}&=&0,\nonumber\\ \sum^3_{j=1}\left(\overset{\ast}{\Lambda}{}^{\dot{l}}_j\otimes \boldsymbol{1}_{2l+1}-\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_j\right)\frac{\dot{\boldsymbol{\psi}}} {\partial\widetilde{a}_j}- i\sum^3_{j=1}\left(\overset{\ast}{\Lambda}{}^{\dot{l}}_j\otimes \boldsymbol{1}_{2l+1}-\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_j\right)\frac{\dot{\boldsymbol{\psi}}} {\partial\widetilde{a}^\ast_j}+\dot{\kappa}^c \dot{\boldsymbol{\psi}}&=&0.\nonumber \end{eqnarray} Substituting the functions $\boldsymbol{\psi}=T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime$ ($\dot{\boldsymbol{\psi}}=\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime$) and the derivatives (\ref{CD1})--(\ref{CD6}), (\ref{CDD1})--(\ref{CDD6}) into this system, and multiply by $T_{l\dot{l}}(\mathfrak{g})=T_{l\dot{l}}(\varphi^c,\theta^c,0)$ ($\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{T}_{l\dot{l}}(\dot{\varphi}^c,\dot{\theta}^c,0)$) from the left, we obtain \begin{multline} T_{l\dot{l}}(\mathfrak{g})\left(\Lambda^l_1\otimes\boldsymbol{1}_{2\dot{1}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_1\right) \left[-\frac{\sin\varphi^c}{r\sin\theta^c} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial\varphi}+\right.\\ \left.+\frac{\cos\varphi^c\cos\theta^c}{r} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial\theta}+ \cos\varphi^c\sin\theta^c \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial r}\right]+\\ +T_{l\dot{l}}(\mathfrak{g})\left(\Lambda^l_2\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_2\right) \left[\frac{\cos\varphi^c}{r\sin\theta^c}\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\varphi}+\right.\\ \left.+\frac{\sin\varphi^c\cos\theta^c}{r}\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\theta}+\sin\varphi^c\sin\theta^c \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial r}\right]+ \nonumber \end{multline} \begin{multline} T_{l\dot{l}}(\mathfrak{g})\Lambda^{l\dot{l}}_3 \left[-\frac{\sin\theta^c}{r} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial\theta}+ \cos\theta^c \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial r}\right]+\\ +iT_{l\dot{l}}(\mathfrak{g})\left(\Lambda^l_1\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_1\right) \left[-\frac{\sin\dot{\varphi}^c}{r\sin\dot{\theta}^c} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\epsilon}+\right.\\ \left.+\frac{\cos\dot{\varphi}^c\sin\dot{\theta}^c}{r} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\tau}+i\cos\dot{\varphi}^c\sin\dot{\theta}^c \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial r}\right]+ \nonumber \end{multline} \begin{multline} +iT_{l\dot{l}}(\mathfrak{g})\left(\Lambda^l_2\otimes\boldsymbol{1}_{2\dot{1}+1}- \boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_2\right) \left[-\frac{\cos\dot{\varphi}^c}{r\sin\dot{\theta}^c} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial\epsilon}+\right.\\ \left.+\frac{\sin\dot{\varphi}^c\cos\dot{\theta}^c}{r} \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial\theta}+ i\sin\dot{\varphi}^c\sin\dot{\theta}^c \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)} {\partial r}\right]+\\ +iT_{l\dot{l}}(\mathfrak{g})\Lambda^{l\dot{l}}_3 \left[-\frac{\sin\dot{\varphi}^c}{r}\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\tau}+ i\cos\dot{\theta}^c\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial r}\right] +\kappa^c\boldsymbol{\psi}=0, \nonumber \end{multline} \begin{multline} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left(\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1}-\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_1\right) \left[-\frac{\sin\varphi^c}{r^\ast\sin\theta^c} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\varphi}+\right.\\ \left.+\frac{\cos\varphi^c\cos\theta^c}{r^\ast} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\theta}+\cos\varphi^c\sin\theta^c \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast}\right]+\\ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left(\overset{\ast}{\Lambda}{}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1}-\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_2\right) \left[\frac{\cos\varphi^c}{r^\ast\sin\theta^c} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\varphi}+\right.\\ \left.+\frac{\sin\varphi^c\cos\theta^c}{r^\ast} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\theta}+\sin\varphi^c\sin\theta^c \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast}\right]+ \nonumber \end{multline} \begin{multline} +\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \left[-\frac{\sin\theta^c}{r^\ast} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\theta}+ \cos\theta^c \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast}\right]-\\ -i\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left(\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1}-\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_1\right) \left[\frac{\sin\dot{\varphi}^c}{r^\ast\sin\dot{\theta}^c} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\epsilon}-\right.\\ \left.-\frac{\cos\dot{\varphi}^c\cos\dot{\theta}^c}{r^\ast} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\tau}-i\cos\dot{\varphi}^c\sin\dot{\theta}^c \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast}\right]- \nonumber \end{multline} \begin{multline} -i\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left(\overset{\ast}{\Lambda}{}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1}-\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_2\right) \left[-\frac{\cos\dot{\varphi}^c}{r^\ast\sin\dot{\theta}^c} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\epsilon}-\right.\\ \left.-\frac{\sin\dot{\varphi}^c\cos\dot{\theta}^c}{r^\ast} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\tau}-i\sin\dot{\varphi}^c\sin\dot{\theta}^c \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast}\right]-\\ -i\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \left[\frac{\sin\dot{\theta}^c}{r^\ast} \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial\tau}- i\cos\dot{\theta}^c \frac{(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast} \right] +\dot{\kappa}^c\dot{\boldsymbol{\psi}}=0. \nonumber \end{multline} In virtue of the invariance conditions (\ref{IC}) we have \begin{gather} T_{l\dot{l}}(\mathfrak{g})\left[-\Lambda^{l\dot{l}}_1\sin\varphi^c+ \Lambda^{l\dot{l}}_2\cos\varphi^c\right] T^{-1}_{l\dot{l}}(\mathfrak{g})=\Lambda^{l\dot{l}}_1,\nonumber\\ T_{l\dot{l}}(\mathfrak{g})\left[\Lambda^{l\dot{l}}_1\cos\varphi^c\cos\theta^c+ \Lambda^{l\dot{l}}_2\sin\varphi^c \cos\theta^c-\Lambda^{l\dot{l}}_3\sin\theta^c\right] T^{-1}_{l\dot{l}}(\mathfrak{g})=\Lambda_2,\nonumber\\ T_{l\dot{l}}(\mathfrak{g})\left[2\Lambda^{l\dot{l}}_1\cos\varphi^c\sin\theta^c+ \Lambda^{l\dot{l}}_2\sin\varphi^c\sin\theta^c+ \Lambda^{l\dot{l}}_3\cos\theta^c\right]T^{-1}_{l\dot{l}}(\mathfrak{g})= \Lambda^{l\dot{l}}_3,\nonumber\\ T_{l\dot{l}}(\mathfrak{g})\left[-\Lambda^{l\dot{l}}_1\sin\dot{\varphi}^c+ \Lambda^{l\dot{l}}_2\cos\dot{\varphi}^c\right]T^{-1}_{l\dot{l}}(\mathfrak{g})= \Lambda^{l\dot{l}}_1,\nonumber\\ T_{l\dot{l}}(\mathfrak{g})\left[\Lambda^{l\dot{l}}_1\cos\dot{\varphi}^c\cos\dot{\theta}^c+ \Lambda^{l\dot{l}}_2\sin\dot{\varphi}^c \cos\dot{\theta}^c-\Lambda^{l\dot{l}}_3\sin\dot{\theta}^c\right] T^{-1}_{l\dot{l}}(\mathfrak{g})=\Lambda^{l\dot{l}}_2,\nonumber\\ T_{l\dot{l}}(\mathfrak{g})\left[\Lambda^{l\dot{l}}_1\cos\dot{\varphi}^c\sin\dot{\theta}^c+ \Lambda^{l\dot{l}}_2\sin\dot{\varphi}^c\sin\dot{\theta}^c+ \Lambda^{l\dot{l}}_3\cos\dot{\theta}^c\right] T^{-1}_{l\dot{l}}(\mathfrak{g})=\Lambda^{l\dot{l}}_3,\nonumber \end{gather} \begin{gather} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left[ -\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\sin\varphi^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\cos\varphi^c\right] \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{\Lambda}{}^{l\dot{l}}_1,\nonumber\\ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left[ \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\cos\varphi^c \cos\theta^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\sin\varphi^c\cos\theta^c- \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\sin\theta^c\right] \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{\Lambda}{}^{l\dot{l}}_2,\nonumber\\ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left[ \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\cos\varphi^c \sin\theta^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\sin\varphi^c\sin\theta^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\cos\theta^c\right] \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{\Lambda}{}^{l\dot{l}}_3,\nonumber\\ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left[ \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\sin\dot{\varphi}^c- \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\cos\dot{\varphi}^c\right] \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{\Lambda}{}^{l\dot{l}}_1,\nonumber\\ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left[ -\overset{\ast}{\Lambda}{}^{l\dot{l}}_1\cos\dot{\varphi}^c \cos\dot{\theta}^c- \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\sin\dot{\varphi}^c\cos\dot{\theta}^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\sin\dot{\theta}^c\right] \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{\Lambda}{}^{l\dot{l}}_2,\nonumber\\ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g})\left[ \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\cos\dot{\varphi}^c \sin\dot{\theta}^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\sin\dot{\varphi}^c\sin\dot{\theta}^c+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_3\cos\dot{\theta}^c\right] \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})= \overset{\ast}{\Lambda}{}^{l\dot{l}}_3.\nonumber \end{gather} Taking into account the latter relations we can write the system (\ref{CT}) as follows \begin{multline} \frac{1}{r\sin\theta^c}\Lambda^{l}_1\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g}) \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\varphi}- \frac{1}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_1 T_{l\dot{l}}(\mathfrak{g})\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\varphi}+\\ +\frac{i}{r\sin\theta^c}\Lambda^{l}_1\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g}) \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\epsilon}+ \frac{i}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_1 T_{l\dot{l}}(\mathfrak{g})\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\epsilon}-\\ -\frac{1}{r}\Lambda^{l}_2\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g})\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\theta}+ \frac{1}{r}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_2T_{l\dot{l}}(\mathfrak{g}) \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)}{\partial\theta}- \nonumber \end{multline} \begin{multline} -\frac{i}{r}\Lambda^{l}_2\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g})\frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g}) \boldsymbol{\psi}^\prime)}{\partial\tau}- \frac{i}{r}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_2T_{l\dot{l}}(\mathfrak{g}) \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)}{\partial\tau}- \\ -\Lambda^{l\dot{l}}_3T_{l\dot{l}}(\mathfrak{g}) \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)}{\partial r}+ i\Lambda^{l\dot{l}}_3T_{l\dot{l}}(\mathfrak{g}) \frac{\partial(T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime)}{\partial r}+ \kappa^c\boldsymbol{\psi}^\prime=0, \nonumber \end{multline} \begin{multline} \frac{1}{r^\ast\sin\theta^c}\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\varphi}- \frac{1}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_1 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\varphi}+ \\ +\frac{i}{r^\ast\sin\theta^c}\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\epsilon}+ \frac{i}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_1 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\epsilon}- \\ -\frac{1}{r^\ast}\overset{\ast}{\Lambda}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\theta}+\frac{1}{r^\ast} \boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_2\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\theta}- \nonumber \end{multline} \begin{multline} -\frac{i}{r^\ast}\overset{\ast}{\Lambda}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\tau}-\frac{i}{r^\ast} \boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_2\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial\tau}+\\ +\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast} -i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial(\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g}) \dot{\boldsymbol{\psi}}^\prime)} {\partial r^\ast} +\dot{\kappa}^c\dot{\boldsymbol{\psi}}^\prime=0. \label{Complex3} \end{multline} The matrices $T^{-1}_{l\dot{l}}(\mathfrak{g})$, $\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})$ depend on $\varphi$, $\epsilon$, $\theta$, $\tau$. Therefore, we must differentiate in $T^{-1}_{l\dot{l}}(\mathfrak{g})\boldsymbol{\psi}^\prime$ ($\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})\dot{\boldsymbol{\psi}}^\prime$) the both factors. After differentiation we come to the following system: \begin{multline} \frac{1}{r\sin\theta^c}\Lambda^{l}_1\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime} {\partial\varphi}- \frac{1}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_1 \frac{\partial\boldsymbol{\psi}^\prime}{\partial\varphi}+ \frac{i}{r\sin\theta^c}\Lambda^{l}_1\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime} {\partial\epsilon}-\\ -\frac{i}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_1 \frac{\partial\boldsymbol{\psi}^\prime}{\partial\epsilon}- \frac{1}{r}\Lambda^{l}_2\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime} {\partial\theta} +\frac{1}{r}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_2\frac{\partial\boldsymbol{\psi}^\prime} {\partial\theta}-\\ -\frac{i}{r}\Lambda^{l}_2\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime} {\partial\tau} -\frac{i}{r}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_2\frac{\partial\boldsymbol{\psi}^\prime} {\partial\tau} +\Lambda^{l\dot{l}}_3 \frac{\partial\boldsymbol{\psi}^\prime}{\partial r} +i\Lambda^{l\dot{l}}_3 \frac{\partial\boldsymbol{\psi}^\prime}{\partial r}+ \nonumber \end{multline} \begin{multline} +\left[ \frac{1}{r\sin\theta^c}\Lambda^{l}_1\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi}- \frac{1}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_1T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi}+\right.\\ +\frac{i}{r\sin\theta^c}\Lambda^{l}_1\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\epsilon}- \frac{i}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_1T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\epsilon}-\\ -\frac{1}{r}\Lambda^{l}_2\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\theta} +\frac{1}{r}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_2T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\theta}-\\ \left.-\frac{i}{r}\Lambda^{l}_2\otimes\boldsymbol{1}_{2\dot{l}+1} T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\tau} -\frac{i}{r}\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_2T_{l\dot{l}}(\mathfrak{g}) \frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\tau} +\kappa^cI\right] \boldsymbol{\psi}^\prime=0, \nonumber \end{multline} \begin{multline} \frac{1}{r^\ast\sin\theta^c}\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\varphi}- \frac{1}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_1 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\varphi}+ \frac{i}{r^\ast\sin\theta^c}\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\epsilon}+\\ +\frac{i}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_1 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\epsilon}- \frac{1}{r^\ast}\overset{\ast}{\Lambda}{}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\theta}+ \frac{1}{r^\ast}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_2 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\theta}-\\ -\frac{i}{r^\ast}\overset{\ast}{\Lambda}{}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\tau} -\frac{i}{r^\ast}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_2 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\tau}+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial r^\ast}- i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial r^\ast}+ \nonumber \end{multline} \begin{multline} +\left[ \frac{1}{r^\ast\sin\theta^c} \overset{\ast}{\Lambda}{}^{\dot{l}}_1\otimes\boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi} - \frac{1}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_1 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial \overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi}+\right.\\ +\frac{i}{r^\ast\sin\theta^c} \overset{\ast}{\Lambda}{}^{\dot{l}}_1\otimes\boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\epsilon} + \frac{i}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_1 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\epsilon}-\\ -\frac{1}{r^\ast}\overset{\ast}{\Lambda}{}^{\dot{l}}_2\otimes \boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\theta}+ \frac{1}{r^\ast}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_2 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\theta}\\ \left.-\frac{i}{r^\ast}\overset{\ast}{\Lambda}{}^{\dot{l}}_2\otimes \boldsymbol{1}_{2l+1} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\tau}- \frac{i}{r^\ast}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^{l}_2 \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\tau}+ \dot{\kappa}^cI\right]\dot{\boldsymbol{\psi}}^\prime=0. \label{Complex4} \end{multline} Let us show that the products $T_{l\dot{l}}(\mathfrak{g})\frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi}$, $\ldots$, $\overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})} {\partial\tau}$ are expressed via linear combinations of the infinitesimal operators. For example, let us consider the simplest tensor representation $\boldsymbol{\tau}_{\frac{1}{2},\frac{1}{2}}= \boldsymbol{\tau}_{\frac{1}{2},0}\otimes\boldsymbol{\tau}_{0,\frac{1}{2}}$. The representation $\boldsymbol{\tau}_{\frac{1}{2},\frac{1}{2}}$ is realized in the four-dimensional symmetric space $\Sym(1,1)$. The matrix of $\boldsymbol{\tau}_{\frac{1}{2},\frac{1}{2}}$ in the space $\Sym(1,1)$ has the following form: \[ T_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})=\ar\begin{pmatrix} e^\epsilon\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & -ie^{i\varphi}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & ie^{-i\varphi}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & e^{-\epsilon}\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ -ie^\epsilon\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & e^{i\varphi}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & e^{-i\varphi}\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & ie^{-\epsilon}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \\ ie^\epsilon\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & e^{i\varphi}\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & e^{-i\varphi}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & -ie^{-\epsilon}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ e^\epsilon\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & ie^{i\varphi}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & -ie^{-i\varphi}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & e^{-\epsilon}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \end{pmatrix}. \] \begin{sloppypar}\noindent This matrix is obtained from (\ref{Tenrep1}) via replacing all the functions $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau)$ by $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)= e^{-m(\epsilon+i\varphi)-\dot{m}(\epsilon-i\varphi)} \mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau)$. An inverse matrix for $T_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})$ is \end{sloppypar} \[ T^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})=\ar\begin{pmatrix} e^{-\epsilon}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & ie^{-\epsilon}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & -ie^{-\epsilon}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & e^{-\epsilon}\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ ie^{-i\varphi}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & e^{-i\varphi}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & e^{-i\varphi}\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & -ie^{-i\varphi}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \\ -ie^{i\varphi}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & e^{i\varphi}\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & e^{i\varphi}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & ie^{i\varphi}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ e^\epsilon\sin\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & -ie^{\epsilon}\sin\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & ie^{\epsilon}\cos\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & e^{\epsilon}\cos\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \end{pmatrix}. \] Infinitesimal operators of the representations $\boldsymbol{\tau}_{\frac{1}{2},0}$ and $\boldsymbol{\tau}_{0,\frac{1}{2}}$ are \[ {\sf A}^{\frac{1}{2}}_1=-\frac{i}{2}\begin{bmatrix} 0 & 1\\ 1 & 0 \end{bmatrix},\quad {\sf A}^{\frac{1}{2}}_2=\frac{1}{2}\begin{bmatrix} 0 & 1\\ -1 & 0 \end{bmatrix},\quad {\sf A}^{\frac{1}{2}}_3=\frac{1}{2}\begin{bmatrix} i & 0\\ 0 &-i \end{bmatrix}, \] \[ {\sf B}^{\frac{1}{2}}_1=-\frac{1}{2}\begin{bmatrix} 0 & 1\\ 1 & 0 \end{bmatrix},\quad {\sf B}^{\frac{1}{2}}_2=\frac{1}{2}\begin{bmatrix} 0 & -i\\ i & 0 \end{bmatrix},\quad {\sf B}^{\frac{1}{2}}_3=\frac{1}{2}\begin{bmatrix} 1 & 0\\ 0 &-1 \end{bmatrix}, \] \[ \widetilde{{\sf A}}^{\frac{1}{2}}_1=-\frac{i}{2}\begin{bmatrix} 0 & 1\\ 1 & 0 \end{bmatrix},\quad \widetilde{{\sf A}}^{\frac{1}{2}}_2=\frac{1}{2}\begin{bmatrix} 0 & 1\\ -1 & 0 \end{bmatrix},\quad \widetilde{{\sf A}}^{\frac{1}{2}}_3=\frac{1}{2}\begin{bmatrix} i & 0\\ 0 &-i \end{bmatrix}, \] \begin{equation}\label{Infhalf} \widetilde{{\sf B}}^{\frac{1}{2}}_1=\frac{1}{2}\begin{bmatrix} 0 & 1\\ 1 & 0 \end{bmatrix},\quad \widetilde{{\sf B}}^{\frac{1}{2}}_2=\frac{1}{2}\begin{bmatrix} 0 & i\\ -i & 0 \end{bmatrix},\quad \widetilde{{\sf B}}^{\frac{1}{2}}_3=\frac{1}{2}\begin{bmatrix} -1 & 0\\ 0 & 1 \end{bmatrix}. \end{equation} Taking into account the latter relations, we find \begin{multline} T_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})}{\partial\varphi}=\\ \frac{1}{2}\ar\begin{pmatrix} -i\cos\theta^c+i\cos\dot{\theta}^c & -\sin\dot{\theta}^c & -\sin\theta^c & 0\\ \sin\dot{\theta}^c & -i\cos\theta^c-i\cos\dot{\theta}^c & 0 & -\sin\theta^c\\ \sin\theta^c & 0 & i\cos\theta^c+i\cos\dot{\theta}^c & -\sin\dot{\theta}^c \\ 0 & \sin\theta^c & \sin\dot{\theta}^c & i\cos\theta^c-i\cos\dot{\theta}^c \end{pmatrix}=\\ =-({\sf A}_3\otimes\boldsymbol{1}_2\cos\theta^c- \boldsymbol{1}_2\otimes\widetilde{{\sf A}}_3\cos\dot{\theta}^c)- ({\sf A}_2\otimes\boldsymbol{1}_2\sin\theta^c+ \boldsymbol{1}_2\otimes\widetilde{{\sf A}}_2\sin\dot{\theta}^c),\label{Irel1} \end{multline} \begin{multline} T_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})}{\partial\epsilon}=\\ \frac{1}{2}\ar\begin{pmatrix} -\cos\theta^c-\cos\dot{\theta}^c & -i\sin\dot{\theta}^c & i\sin\theta^c & 0\\ i\sin\dot{\theta}^c & -\cos\theta^c+\cos\dot{\theta}^c & 0 & i\sin\theta^c\\ -i\sin\theta^c & 0 & \cos\theta^c-\cos\dot{\theta}^c & -i\sin\dot{\theta}^c \\ 0 & -i\sin\theta^c & i\sin\dot{\theta}^c & \cos\theta^c+\cos\dot{\theta}^c \end{pmatrix}=\\ =-({\sf B}_3\otimes\boldsymbol{1}_2\cos\theta^c- \boldsymbol{1}_2\otimes\widetilde{{\sf B}}_3\cos\dot{\theta}^c)- ({\sf B}_2\otimes\boldsymbol{1}_2\sin\theta^c+ \boldsymbol{1}_2\otimes\widetilde{{\sf B}}_2\sin\dot{\theta}^c),\label{Irel2} \end{multline} \begin{equation}\label{Irel3} T_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})}{\partial\theta}= \frac{1}{2}\ar\begin{pmatrix} 0 & i & -i & 0\\ i & 0 & 0 & -i\\ -i & 0 & 0 & i\\ 0 & -i & i & 0 \end{pmatrix}= {\sf A}_1\otimes\boldsymbol{1}_2-\boldsymbol{1}_2\otimes\widetilde{{\sf A}}_1, \end{equation} \begin{equation}\label{Irel4} T_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})}{\partial\tau}= -\frac{1}{2}\ar\begin{pmatrix} 0 & 1 & 1 & 0\\ 1 & 0 & 0 & 1\\ 1 & 0 & 0 & 1\\ 0 & 1 & 1 & 0 \end{pmatrix}= {\sf B}_1\otimes\boldsymbol{1}_2-\boldsymbol{1}_2\otimes\widetilde{{\sf B}}_1. \end{equation} Further, a matrix of the conjugate representation $\overset{\ast}{\boldsymbol{\tau}}_{\frac{1}{2},\frac{1}{2}}$ has the form \[ \overset{\ast}{T}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})=\ar\begin{pmatrix} e^\epsilon\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & ie^{-i\varphi}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & -ie^{i\varphi}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & e^{-\epsilon}\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} \\ ie^\epsilon\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & e^{-i\varphi}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & e^{i\varphi}\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & -ie^{-\epsilon}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} \\ -ie^\epsilon\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & e^{-i\varphi}\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & e^{i\varphi}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & ie^{-\epsilon}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} \\ e^\epsilon\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & -ie^{-i\varphi}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & ie^{i\varphi}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & e^{-\epsilon}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} \end{pmatrix}, \] and its inverse matrix is \[ \overset{\ast}{T}{}^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})=\ar\begin{pmatrix} e^{-\epsilon}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & -ie^{-\epsilon}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & ie^{-\epsilon}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & e^{-\epsilon}\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} \\ -ie^{i\varphi}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & e^{i\varphi}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & e^{i\varphi}\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & ie^{i\varphi}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} \\ ie^{-i\varphi}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & e^{-i\varphi}\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & e^{-i\varphi}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & -ie^{-i\varphi}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} \\ e^\epsilon\sin\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & ie^{\epsilon}\sin\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} & -ie^{\epsilon}\cos\frac{\dot{\theta}^c}{2}\sin\frac{\theta^c}{2} & e^{\epsilon}\cos\frac{\dot{\theta}^c}{2}\cos\frac{\theta^c}{2} \end{pmatrix}. \] In this case we have \begin{multline} \overset{\ast}{T}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})} {\partial\varphi}=\\ \frac{1}{2}\ar\begin{pmatrix} -i\cos\theta^c+i\cos\dot{\theta}^c & -\sin\theta^c & -\sin\dot{\theta}^c & 0\\ \sin\theta^c & i\cos\theta^c+i\cos\dot{\theta}^c & 0 & -\sin\dot{\theta}^c\\ \sin\dot{\theta}^c & 0 & -i\cos\theta^c-i\cos\dot{\theta}^c & -\sin\theta^c \\ 0 & \sin\dot{\theta}^c & \sin\theta^c & i\cos\theta^c-i\cos\dot{\theta}^c \end{pmatrix}=\\ =(\widetilde{{\sf A}}_3\otimes\boldsymbol{1}_2\cos\dot{\theta}^c- \boldsymbol{1}_2\otimes{\sf A}_3\cos\theta^c)- (\widetilde{{\sf A}}_2\otimes\boldsymbol{1}_2\sin\dot{\theta}^c+ \boldsymbol{1}_2\otimes{\sf A}_2\sin\theta^c),\label{Irel5} \end{multline} \begin{multline} \overset{\ast}{T}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})} {\partial\epsilon}=\\ \frac{1}{2}\ar\begin{pmatrix} -\cos\theta^c-\cos\dot{\theta}^c & i\sin\theta^c & -i\sin\dot{\theta}^c & 0\\ -i\sin\theta^c & \cos\theta^c-\cos\dot{\theta}^c & 0 & -i\sin\dot{\theta}^c\\ i\sin\dot{\theta}^c & 0 & -\cos\theta^c+\cos\dot{\theta}^c & i\sin\theta^c \\ 0 & i\sin\dot{\theta}^c & -i\sin\theta^c & \cos\theta^c+\cos\dot{\theta}^c \end{pmatrix}=\\ =(\widetilde{{\sf B}}_3\otimes\boldsymbol{1}_2\cos\dot{\theta}^c- \boldsymbol{1}_2\otimes{\sf B}_3\cos\theta^c)- (\widetilde{{\sf B}}_2\otimes\boldsymbol{1}_2\sin\dot{\theta}^c+ \boldsymbol{1}_2\otimes{\sf B}_2\sin\theta^c),\label{Irel6} \end{multline} \begin{equation}\label{Irel7} \overset{\ast}{T}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})} {\partial\theta}= \frac{1}{2}\ar\begin{pmatrix} 0 & -i & i & 0\\ -i & 0 & 0 & i\\ i & 0 & 0 & -i\\ 0 & i & -i & 0 \end{pmatrix}= -(\widetilde{{\sf A}}_1\otimes\boldsymbol{1}_2-\boldsymbol{1}_2\otimes{\sf A}_1), \end{equation} \begin{equation}\label{Irel8} \overset{\ast}{T}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{\frac{1}{2}\frac{1}{2}}(\mathfrak{g})} {\partial\tau}= -\frac{1}{2}\ar\begin{pmatrix} 0 & 1 & 1 & 0\\ 1 & 0 & 0 & 1\\ 1 & 0 & 0 & 1\\ 0 & 1 & 1 & 0 \end{pmatrix}= -(\widetilde{{\sf B}}_1\otimes\boldsymbol{1}_2-\boldsymbol{1}_2\otimes{\sf B}_1). \end{equation} In the following example we consider the first nontrivial tensor representation $\boldsymbol{\tau}_{1,\frac{1}{2}}= \boldsymbol{\tau}_{1,0}\otimes\boldsymbol{\tau}_{0,\frac{1}{2}}$. The representation $\boldsymbol{\tau}_{1,\frac{1}{2}}$ is realized in the six-dimensional symmetric space $\Sym(2,1)$. The matrix of $\boldsymbol{\tau}_{1,\frac{1}{2}}$ in the space $\Sym(2,1)$ has the following form: \begin{multline} T_{1\frac{1}{2}}(\mathfrak{g})=\left(\ar\begin{array}{ccc} \scriptstyle e^{\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{ie^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle -ie^{\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{ie^{\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{\frac{\epsilon-i\varphi}{2}} \cos\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{e^{\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{ie^{\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{\frac{\epsilon-i\varphi}{2}} \cos\theta^c\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle -e^{\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{ie^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle ie^{\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle -e^{\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} \end{array}\right.\\ \left.\ar\begin{array}{ccc} \scriptstyle\frac{e^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle -e^{-\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{-\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{ie^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{-\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle -e^{-\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle -ie^{\frac{-\epsilon+i\varphi}{2}} \cos\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{ie^{-\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{-\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle e^{\frac{-\epsilon+i\varphi}{2}} \cos\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{-\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{ie^{-\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{e^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{-\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{-\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{ie^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{-\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{-\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \end{array}\right).\nonumber \end{multline} In turn, this matrix is obtained from (\ref{Tenrep2}) via replacing all the functions $\mathfrak{Z}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\theta,\tau)$ by $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)$. An inverse matrix for $T_{1\frac{1}{2}}(\mathfrak{g})$ is \begin{multline} T^{-1}_{1\frac{1}{2}}(\mathfrak{g})=\left(\ar\begin{array}{ccc} \scriptstyle e^{-\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{-\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle-\frac{ie^{-\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle ie^{-\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{-\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{-\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle-\frac{ie^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{\frac{-\epsilon+i\varphi}{2}} \cos\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{e^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle-\frac{ie^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{\frac{\epsilon-i\varphi}{2}} \cos\theta^c\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle -e^{\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle-\frac{ie^{\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle -ie^{\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle -e^{\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} \end{array}\right.\\ \left.\ar\begin{array}{ccc} \scriptstyle\frac{e^{-\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle -e^{-\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{-\frac{3\epsilon+i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle-\frac{ie^{-\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle -ie^{-\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle -e^{-\frac{\epsilon+3i\varphi}{2}} \sin^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle ie^{\frac{-\epsilon+i\varphi}{2}} \cos\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle-\frac{ie^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{-\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle e^{\frac{\epsilon-i\varphi}{2}} \cos\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle\frac{e^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle-\frac{ie^{\frac{\epsilon-i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} \\ \scriptstyle\frac{e^{\frac{\epsilon+3i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{\frac{\epsilon+3i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} \\ \scriptstyle-\frac{ie^{\frac{3\epsilon+i\varphi}{2}}}{\sqrt{2}} \sin\theta^c\cos\frac{\dot{\theta}^c}{2} & \scriptstyle ie^{\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\sin\frac{\dot{\theta}^c}{2} & \scriptstyle e^{\frac{3\epsilon+i\varphi}{2}} \cos^2\frac{\theta^c}{2}\cos\frac{\dot{\theta}^c}{2} \end{array}\right).\nonumber \end{multline} In turn, infinitesimal operators of the representation $\boldsymbol{\tau}_{1,0}$ are \[ {\sf A}^1_1=-\frac{i}{\sqrt{2}}\begin{bmatrix} 0 & 1 & 0\\ 1 & 0 & 1\\ 0 & 1 & 0 \end{bmatrix},\quad {\sf A}^1_2=\frac{1}{\sqrt{2}}\begin{bmatrix} 0 & 1 & 0\\ -1 & 0 & 1\\ 0 &-1 & 0 \end{bmatrix},\quad {\sf A}^1_3=\begin{bmatrix} i & 0 & 0\\ 0 & 0 & 0\\ 0 & 0 &-i \end{bmatrix}, \] \[ {\sf B}^1_1=-\frac{1}{\sqrt{2}}\begin{bmatrix} 0 & 1 & 0\\ 1 & 0 & 1\\ 0 & 1 & 0 \end{bmatrix},\quad {\sf B}^1_2=\frac{1}{\sqrt{2}}\begin{bmatrix} 0 &-i & 0\\ i & 0 &-i\\ 0 & i & 0 \end{bmatrix},\quad {\sf B}^1_3=\begin{bmatrix} 1 & 0 & 0\\ 0 & 0 & 0\\ 0 & 0 &-1 \end{bmatrix}, \] \[ \widetilde{{\sf A}}^1_1=-\frac{i}{\sqrt{2}}\begin{bmatrix} 0 & 1 & 0\\ 1 & 0 & 1\\ 0 & 1 & 0 \end{bmatrix},\quad \widetilde{{\sf A}}^1_2=\frac{1}{\sqrt{2}}\begin{bmatrix} 0 & 1 & 0\\ -1 & 0 & 1\\ 0 & -1 & 0 \end{bmatrix},\quad \widetilde{{\sf A}}^1_3=\begin{bmatrix} i & 0 & 0\\ 0 & 0 & 0\\ 0 & 0 & -i \end{bmatrix}, \] \[ \widetilde{{\sf B}}^1_1=\frac{1}{\sqrt{2}}\begin{bmatrix} 0 & 1 & 0\\ 1 & 0 & 1\\ 0 & 1 & 0 \end{bmatrix},\quad \widetilde{{\sf B}}^1_2=\frac{1}{\sqrt{2}}\begin{bmatrix} 0 & i & 0\\ -i & 0 & i\\ 0 & -i & 0 \end{bmatrix},\quad \widetilde{{\sf B}}^1_3=\begin{bmatrix} -1 & 0 & 0\\ 0 & 0 & 0\\ 0 & 0 & 1 \end{bmatrix}. \] Taking into account the latter expressions and the operators (\ref{Infhalf}), we obtain \begin{gather} T_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\varphi}=\nonumber\\ \ar\begin{pmatrix} \scriptstyle -i\cos\theta^c+\frac{i}{2}\cos\dot{\theta}^c & \scriptstyle -\frac{1}{2}\sin\dot{\theta}^c & \scriptstyle -\frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle\frac{1}{2}\sin\dot{\theta}^c & \scriptstyle -i\cos\theta^c-\frac{i}{2}\cos\dot{\theta}^c & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle \frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle \frac{i}{2}\cos\dot{\theta}^c & \scriptstyle -\frac{1}{2}\sin\dot{\theta}^c & \scriptstyle -\frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 \\ \scriptstyle 0 & \scriptstyle \frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle \frac{1}{2}\sin\dot{\theta}^c & \scriptstyle -\frac{i}{2}\cos\dot{\theta}^c & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}}\sin\theta^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle i\cos\theta^c+\frac{i}{2}\cos\dot{\theta}^c & \scriptstyle -\frac{1}{2}\sin\dot{\theta}^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{1}{\sqrt{2}}\sin\theta^c & \scriptstyle \frac{1}{2}\sin\dot{\theta}^c & \scriptstyle i\cos\theta^c-\frac{i}{2}\cos\dot{\theta}^c \end{pmatrix}=\nonumber\\ =-({\sf A}^1_3\otimes\boldsymbol{1}_2\cos\theta^c- \boldsymbol{1}_3\otimes\widetilde{{\sf A}}^{\frac{1}{2}}_3\cos\dot{\theta}^c)- ({\sf A}^1_2\otimes\boldsymbol{1}_2\sin\theta^c+ \boldsymbol{1}_3\otimes\widetilde{{\sf A}}^{\frac{1}{2}}_2\sin\dot{\theta}^c), \label{Irel9} \end{gather} \begin{gather} T_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\epsilon}=\nonumber\\ \ar\begin{pmatrix} \scriptstyle -\cos\theta^c-\frac{1}{2}\cos\dot{\theta}^c & \scriptstyle -\frac{i}{2}\sin\dot{\theta}^c & \scriptstyle -\frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle\frac{i}{2}\sin\dot{\theta}^c & \scriptstyle -\cos\theta^c+\frac{1}{2}\cos\dot{\theta}^c & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle -\frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle -\frac{1}{2}\cos\dot{\theta}^c & \scriptstyle -\frac{i}{2}\sin\dot{\theta}^c & \scriptstyle \frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 \\ \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle \frac{i}{2}\sin\dot{\theta}^c & \scriptstyle \frac{1}{2}\cos\dot{\theta}^c & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}}\sin\theta^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle 0 & \scriptstyle \cos\theta^c-\frac{1}{2}\cos\dot{\theta}^c & \scriptstyle -\frac{i}{2}\sin\dot{\theta}^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}}\sin\theta^c & \scriptstyle \frac{i}{2}\sin\dot{\theta}^c & \scriptstyle \cos\theta^c+\frac{1}{2}\cos\dot{\theta}^c \end{pmatrix}=\nonumber\\ =-({\sf B}^1_3\otimes\boldsymbol{1}_2\cos\theta^c- \boldsymbol{1}_3\otimes\widetilde{{\sf B}}^{\frac{1}{2}}_3\cos\dot{\theta}^c)- ({\sf B}^1_2\otimes\boldsymbol{1}_2\sin\theta^c+ \boldsymbol{1}_3\otimes\widetilde{{\sf B}}^{\frac{1}{2}}_2\sin\dot{\theta}^c), \label{Irel10} \end{gather} \begin{equation}\label{Irel11} T_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\theta}= \ar\begin{pmatrix} \scriptstyle 0 & \scriptstyle \frac{i}{2} & \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle \frac{i}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{2} & \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle 0\\ \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle \frac{i}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{2}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}} & \scriptstyle \frac{i}{2} & \scriptstyle 0 \end{pmatrix}= {\sf A}^1_1\otimes\boldsymbol{1}_2- \boldsymbol{1}_3\otimes\widetilde{{\sf A}}^{\frac{1}{2}}_1, \end{equation} \begin{equation}\label{Irel12} T_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial T^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\tau}= \ar\begin{pmatrix} \scriptstyle 0 & \scriptstyle -\frac{1}{2} & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle -\frac{1}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{2} & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0\\ \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle -\frac{1}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{2}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle -\frac{1}{2} & \scriptstyle 0 \end{pmatrix} ={\sf B}^1_1\otimes\boldsymbol{1}_2- \boldsymbol{1}_3\otimes\widetilde{{\sf B}}^{\frac{1}{2}}_1. \end{equation} Further, the conjugate representation $\overset{\ast}{\boldsymbol{\tau}}_{1,\frac{1}{2}}$ acts in the space $\Sym(1,2)$ and we have the following relations: \begin{gather} \overset{\ast}{T}_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial \overset{\ast}{T}{}^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\varphi}=\nonumber\\ \ar\begin{pmatrix} \scriptstyle i\cos\dot{\theta}^c-\frac{i}{2}\cos\theta^c & \scriptstyle -\frac{1}{2}\sin\theta^c & \scriptstyle -\frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle\frac{1}{2}\sin\theta^c & \scriptstyle i\cos\dot{\theta}^c+\frac{i}{2}\cos\theta^c & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle \frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle -\frac{i}{2}\cos\theta^c & \scriptstyle -\frac{1}{2}\sin\theta^c & \scriptstyle -\frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 \\ \scriptstyle 0 & \scriptstyle \frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle \frac{1}{2}\sin\theta^c & \scriptstyle \frac{i}{2}\cos\theta^c & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}}\sin\dot{\theta}^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle -i\cos\dot{\theta}^c-\frac{i}{2}\cos\theta^c & \scriptstyle -\frac{1}{2}\sin\theta^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{1}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle \frac{1}{2}\sin\theta^c & \scriptstyle -i\cos\dot{\theta}^c+\frac{i}{2}\cos\theta^c \end{pmatrix}=\nonumber\\ =(\widetilde{{\sf A}}^1_3\otimes\boldsymbol{1}_2\cos\dot{\theta}^c- \boldsymbol{1}_3\otimes{\sf A}^{\frac{1}{2}}_3\cos\theta^c)- (\widetilde{{\sf A}}^1_2\otimes\boldsymbol{1}_2\sin\dot{\theta}^c+ \boldsymbol{1}_3\otimes{\sf A}^{\frac{1}{2}}_2\sin\theta^c), \label{Irel13} \end{gather} \begin{gather} \overset{\ast}{T}_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial \overset{\ast}{T}{}^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\epsilon}=\nonumber\\ \ar\begin{pmatrix} \scriptstyle -\cos\dot{\theta}^c-\frac{1}{2}\cos\theta^c & \scriptstyle \frac{i}{2}\sin\theta^c & \scriptstyle -\frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle-\frac{i}{2}\sin\theta^c & \scriptstyle -\cos\dot{\theta}^c+\frac{1}{2}\cos\theta^c & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle 0 \\ \scriptstyle \frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle -\frac{1}{2}\cos\theta^c & \scriptstyle \frac{i}{2}\sin\theta^c & \scriptstyle -\frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 \\ \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle -\frac{i}{2}\sin\theta^c & \scriptstyle \frac{1}{2}\cos\theta^c & \scriptstyle 0 & \scriptstyle -\frac{i}{\sqrt{2}}\sin\dot{\theta}^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle 0 & \scriptstyle \cos\dot{\theta}^c-\frac{1}{2}\cos\theta^c & \scriptstyle \frac{i}{2}\sin\theta^c \\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}}\sin\dot{\theta}^c & \scriptstyle -\frac{i}{2}\sin\theta^c & \scriptstyle \cos\dot{\theta}^c+\frac{1}{2}\cos\theta^c \end{pmatrix}=\nonumber\\ =(\widetilde{{\sf B}}^1_3\otimes\boldsymbol{1}_2\cos\dot{\theta}^c- \boldsymbol{1}_3\otimes{\sf B}^{\frac{1}{2}}_3\cos\theta^c)- (\widetilde{{\sf B}}^1_2\otimes\boldsymbol{1}_2\sin\dot{\theta}^c+ \boldsymbol{1}_3\otimes{\sf B}^{\frac{1}{2}}_2\sin\theta^c), \label{Irel14} \end{gather} \begin{equation}\label{Irel15} \overset{\ast}{T}_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial \overset{\ast}{T}{}^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\theta}= \ar\begin{pmatrix} \scriptstyle 0 & \scriptstyle -\frac{i}{2} & \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle -\frac{i}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{2} & \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle 0\\ \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle -\frac{i}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{i}{2}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle \frac{i}{\sqrt{2}} & \scriptstyle -\frac{i}{2} & \scriptstyle 0 \end{pmatrix} =-(\widetilde{{\sf A}}^1_1\otimes\boldsymbol{1}_2- \boldsymbol{1}_3\otimes{\sf A}^{\frac{1}{2}}_1), \end{equation} \begin{equation}\label{Irel16} \overset{\ast}{T}_{1\frac{1}{2}}(\mathfrak{g}) \frac{\partial \overset{\ast}{T}{}^{-1}_{1\frac{1}{2}}(\mathfrak{g})}{\partial\tau}= \ar\begin{pmatrix} \scriptstyle 0 & \scriptstyle -\frac{1}{2} & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle -\frac{1}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0\\ \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{2} & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0\\ \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle -\frac{1}{2} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{2}\\ \scriptstyle 0 & \scriptstyle 0 & \scriptstyle 0 & \scriptstyle -\frac{1}{\sqrt{2}} & \scriptstyle -\frac{1}{2} & \scriptstyle 0 \end{pmatrix} =-(\widetilde{{\sf B}}^1_1\otimes\boldsymbol{1}_2- \boldsymbol{1}_3\otimes{\sf B}^{\frac{1}{2}}_1. \end{equation} It is easy to verify that relations of the type (\ref{Irel1})--(\ref{Irel16}) take place for any representation $T_{l\dot{l}}(\mathfrak{g})$ of the group $\mathfrak{G}_+$. Therefore, \begin{multline} T_{l\dot{l}}(\mathfrak{g})\frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi}= -({\sf A}^l_3\otimes\boldsymbol{1}_{2\dot{l}+1}\cos\theta^c- \boldsymbol{1}_{2l+1}\otimes{\sf A}^{\dot{l}}_3\cos\dot{\theta}^c)-\\ -({\sf A}^l_2\otimes\boldsymbol{1}_{2\dot{l}+1}\sin\theta^c+ \boldsymbol{1}_{2l+1}\otimes{\sf A}^{\dot{l}}_2\sin\dot{\theta}^c), \nonumber \end{multline} \begin{multline} T_{l\dot{l}}(\mathfrak{g})\frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\epsilon}= -({\sf B}^l_3\otimes\boldsymbol{1}_{2\dot{l}+1}\cos\theta^c- \boldsymbol{1}_{2l+1}\otimes{\sf B}^{\dot{l}}_3\cos\dot{\theta}^c)-\\ -({\sf B}^l_2\otimes\boldsymbol{1}_{2\dot{l}+1}\sin\theta^c+ \boldsymbol{1}_{2l+1}\otimes{\sf B}^{\dot{l}}_2\sin\dot{\theta}^c), \nonumber \end{multline} \[ T_{l\dot{l}}(\mathfrak{g})\frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\theta}= {\sf A}^l_1\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes{\sf A}^{\dot{l}}_1, \] \[ T_{l\dot{l}}(\mathfrak{g})\frac{\partial T^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\tau}= {\sf B}^l_1\otimes\boldsymbol{1}_{2\dot{l}+1}- \boldsymbol{1}_{2l+1}\otimes{\sf B}^{\dot{l}}_1, \] \begin{multline} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\varphi}= ({\sf A}^{\dot{l}}_3\otimes\boldsymbol{1}_{2l+1}\cos\dot{\theta}^c- \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf A}^l_3\cos\theta^c)-\\ -({\sf A}^{\dot{l}}_2\otimes\boldsymbol{1}_{2l+1}\sin\dot{\theta}^c+ \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf A}^l_2\sin\theta^c), \nonumber \end{multline} \begin{multline} \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\epsilon}= ({\sf B}^{\dot{l}}_3\otimes\boldsymbol{1}_{2l+1}\cos\dot{\theta}^c- \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf B}^l_3\cos\theta^c)-\\ -({\sf B}^{\dot{l}}_2\otimes\boldsymbol{1}_{2l+1}\sin\dot{\theta}^c+ \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf B}^l_2\sin\theta^c), \nonumber \end{multline} \[ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\theta}= -({\sf A}^{\dot{l}}_1\otimes\boldsymbol{1}_{2l+1}- \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf A}^l_1), \] \[ \overset{\ast}{T}_{l\dot{l}}(\mathfrak{g}) \frac{\partial\overset{\ast}{T}{}^{-1}_{l\dot{l}}(\mathfrak{g})}{\partial\tau}= -({\sf B}^{\dot{l}}_1\otimes\boldsymbol{1}_{2l+1}- \boldsymbol{1}_{2\dot{l}+1}\otimes{\sf B}^l_1). \] Substituting these relations into the system (\ref{Complex4}), we obtain \begin{multline} \frac{1}{r\sin\theta^c}\Lambda^l_1\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime} {\partial\varphi}- \frac{1}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_1 \frac{\partial\boldsymbol{\psi}^\prime}{\partial\varphi}+ \frac{i}{r\sin\theta^c}\Lambda^l_1\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime} {\partial\epsilon}+\\ +\frac{i}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_1 \frac{\partial\boldsymbol{\psi}^\prime} {\partial\epsilon}- \frac{1}{r}\Lambda^l_2\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime}{\partial\theta}+ \frac{1}{r}\boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_2 \frac{\partial\boldsymbol{\psi}^\prime}{\partial\theta}-\\ -\frac{i}{r}\Lambda^l_2\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\boldsymbol{\psi}^\prime}{\partial\tau} -\frac{i}{r}\boldsymbol{1}_{2l+1}\otimes\Lambda^{\dot{l}}_2 \frac{\partial\boldsymbol{\psi}^\prime}{\partial\tau} +(1+i)\Lambda^{l\dot{l}}_3\frac{\partial\boldsymbol{\psi}^\prime}{\partial r}+\\ +\frac{1}{r}\left[\Lambda^{l\dot{l}}_1{\sf A}^{l\dot{l}}_2- \Lambda^{l\dot{l}}_2{\sf A}^{l\dot{l}}_1+\Lambda^{l\dot{l}}_1{\sf B}^{l\dot{l}}_2- \Lambda^{l\dot{l}}_2{\sf B}^{l\dot{l}}_1- 2\cot\theta^c\Lambda^l_1\otimes\boldsymbol{1}_{2\dot{l}+1}{\sf A}^l_3\otimes \boldsymbol{1}_{2\dot{l}+1}-\right. \\ \left.-2\cot\dot{\theta}^c\boldsymbol{1}_{2l+1}\otimes \Lambda^{\dot{l}}_1\boldsymbol{1}_{2l+1}\otimes{\sf A}^{\dot{l}}_3 \right] \boldsymbol{\psi}^\prime+ \kappa^c\boldsymbol{\psi}=0, \nonumber \end{multline} \begin{multline} \frac{1}{r^\ast\sin\theta^c}\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime} {\partial\varphi}- \frac{1}{r^\ast\sin\dot{\theta}^c} \boldsymbol{1}_{2\dot{l}+1}\otimes\overset{\ast}{\Lambda}{}^l_1 \frac{\partial\dot{\boldsymbol{\psi}}^\prime} {\partial\varphi}+ \frac{i}{r^\ast\sin\theta^c}\overset{\ast}{\Lambda}{}^{\dot{l}}_1 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\epsilon}+\\ +\frac{i}{r^\ast\sin\dot{\theta}^c} \boldsymbol{1}_{2\dot{l}+1}\otimes\overset{\ast}{\Lambda}{}^l_1 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\epsilon}- \frac{1}{r^\ast}\overset{\ast}{\Lambda}{}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\theta}+ \frac{1}{r^\ast} \boldsymbol{1}_{2\dot{l}+1}\otimes\overset{\ast}{\Lambda}{}^l_2 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\theta}-\\ -\frac{i}{r^\ast}\overset{\ast}{\Lambda}{}^{\dot{l}}_2 \otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\tau}- \frac{i}{r^\ast} \boldsymbol{1}_{2\dot{l}+1}\otimes\overset{\ast}{\Lambda}{}^l_2 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial\tau}+ (1-i)\overset{\ast}{\Lambda}{}^{l\dot{l}}_3 \frac{\partial\dot{\boldsymbol{\psi}}^\prime}{\partial r^\ast}+\\ +\frac{1}{r^\ast}\left[\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \widetilde{{\sf A}}^{l\dot{l}}_2- \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\widetilde{{\sf A}}^{l\dot{l}}_1+ \overset{\ast}{\Lambda}{}^{l\dot{l}}_1\widetilde{{\sf B}}^{l\dot{l}}_2- \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\widetilde{{\sf B}}^{l\dot{l}}_1- 2\cot\theta^c\overset{\ast}{\Lambda}{}^{\dot{l}}_1\otimes \boldsymbol{1}_{2l+1}{\sf A}^{\dot{l}}_3\otimes \boldsymbol{1}_{2l+1}- \right. \\ \left.- 2\cot\dot{\theta}^c\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{\Lambda}{}^l_1\boldsymbol{1}_{2\dot{l}+1}\otimes{\sf A}^l_3 \right] \dot{\boldsymbol{\psi}}^\prime+ \dot{\kappa}^c\dot{\boldsymbol{\psi}}=0. \label{Complex5} \end{multline} Now we can separate the variables in the relativistically invariant system. Namely, we represent the each component $\psi^{k\dot{k}}_{lm;\dot{l}\dot{m}}$ of the wave function $\boldsymbol{\psi}$ in the form of an expansion in the generalized hyperspherical functions $\mathfrak{M}^{l\dot{l}}_{mn;\dot{m}\dot{n}}$.\index{function!hypersperical!generalized} This procedure gives rise to separation of variables, that is, it reduces the relativistically invariant system\index{relativistically invariant system} to the system of ordinary differential equations. Preliminarily, we will calculate elements of the matrices $D=\Lambda^{l\dot{l}}_1{\sf A}^{l\dot{l}}_2-\Lambda^{l\dot{l}}_2{\sf A}^{l\dot{l}}_1$, $E=\Lambda^{l\dot{l}}_1{\sf B}^{l\dot{l}}_2-\Lambda^{l\dot{l}}_2{\sf B}^{l\dot{l}}_1$, $\overset{\ast}{D}=\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \widetilde{{\sf A}}^{l\dot{l}}_2- \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\widetilde{{\sf A}}^{l\dot{l}}_1$, $\overset{\ast}{E}=\overset{\ast}{\Lambda}{}^{l\dot{l}}_1 \widetilde{{\sf B}}^{l\dot{l}}_2- \overset{\ast}{\Lambda}{}^{l\dot{l}}_2\widetilde{{\sf B}}^{l\dot{l}}_1$. First of all, let us find elements of the matrix $D=\Lambda^{l\dot{l}}_1{\sf A}^{l\dot{l}}_2- \Lambda^{l\dot{l}}_2{\sf A}^{l\dot{l}}_1$. Using the relations (\ref{AL}), we can write $D=2\Lambda^{l\dot{l}}_3+{\sf A}^{l\dot{l}}_2\Lambda^{l\dot{l}}_1-{\sf A}^{l\dot{l}}_1 \Lambda^{l\dot{l}}_2$. As usual, the action of the transformation $D$ in the helicity basis has the following form: \[ D\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} d^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m;\dot{l}^\prime \dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime;\dot{l}^\prime\dot{m}^\prime}. \] Taking into account (\ref{L3}), (\ref{L1}), (\ref{L2}) and (\ref{Op1})--(\ref{Op2}), we obtain \begin{multline} D\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}=( 2\Lambda^{l\dot{l}}_3+{\sf A}^{l\dot{l}}_2\Lambda^{l\dot{l}}_1-{\sf A}^{l\dot{l}}_1 \Lambda^{l\dot{l}}_2) \zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}} =2\sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime \dot{m}^\prime}+\\ +{\sf A}^{l\dot{l}}_2 \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}- {\sf A}^{l\dot{l}}_1 \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}=\\ =2\sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}+ \frac{1}{2} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \left(\boldsymbol{\alpha}^{l^\prime}_{m^\prime} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime-1; \dot{l}^\prime\dot{m}^\prime}-\right.\\ -\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime+1; \dot{l}^\prime\dot{m}^\prime} \left.-\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime, \dot{m}^\prime-1}+ \boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime, \dot{m}^\prime+1}\right)-\\ -\frac{i}{2} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}} \left(-\boldsymbol{\alpha}^{l^\prime}_{m^\prime} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime-1; \dot{l}^\prime\dot{m}^\prime}- \boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime,m^\prime+1; \dot{l}^\prime\dot{m}^\prime}\right.+\\ \left.+\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime, \dot{m}^\prime-1}+ \boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime, \dot{m}^\prime+1}\right). \nonumber \end{multline} Dividing the each of the two latter sums on the four and changing the summation index in the each eight sums obtained, we come to the following expression: \begin{multline} D\zeta^{k\dot{k}}_{lm;\dot{l}\dot{m}}= \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \left( 2c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}+ \frac{1}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime+1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}-\right.\\ \left.-\frac{1}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime-1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}- \frac{1}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{m}^\prime+1,\dot{m}}+\right.\\ +\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime-1,\dot{m}}+ \frac{i}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime+1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}+\\ +\frac{i}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime-1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}- \frac{i}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime+1,\dot{m}}-\\ \left.-\frac{i}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime-1,\dot{m}}\right) \zeta^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}.\nonumber \end{multline} Therefore, a general element of the matrix $D$ has the form \begin{multline} d^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}= 2c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}+ \frac{1}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime+1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}-\\ -\frac{1}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime-1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}- \frac{1}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{m}^\prime+1,\dot{m}}+\\ +\frac{1}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} a^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime-1,\dot{m}}+ \frac{i}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime+1} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime+1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}+\\ +\frac{i}{2}\boldsymbol{\alpha}^{l^\prime}_{m^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime-1,m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}- \frac{i}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime+1} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime+1,\dot{m}}- \frac{i}{2}\boldsymbol{\alpha}^{\dot{l}^\prime}_{\dot{m}^\prime} b^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime-1,\dot{m}} \nonumber \end{multline} Using the formulae (\ref{L3}), (\ref{L1}) and (\ref{L2}), we find that \begin{equation}\label{VM} D:\quad {\renewcommand{\arraystretch}{1.3} \left\{\begin{array}{ccc} d^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,mm;\dot{l}-1,\dot{l}, \dot{m}\dot{m}}&=& c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l;\dot{l}-1,\dot{l}} (l+\dot{l})\sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)},\\ d^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,mm;\dot{l}+1,\dot{l}, \dot{m}\dot{m}}&=& -c^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l;\dot{l}+1,\dot{l}} (l+\dot{l}+2)\sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}. \end{array}\right.} \end{equation} All other elements $d^{k^\prime k;\dot{k}^\prime\dot{k}}_{l^\prime l,m^\prime m; \dot{l}^\prime\dot{l},\dot{m}^\prime\dot{m}}$ are equal to zero. Analogously, using the relations (\ref{BL}), (\ref{DAL})--(\ref{DBL}) and the operators (\ref{Op1})--(\ref{Op12}), we find that elements of the matrices $E=-2i\Lambda^{l\dot{l}}_3+{\sf B}^{l\dot{l}}_2\Lambda^{l\dot{l}}_1- {\sf B}^{l\dot{l}}_1\Lambda^{l\dot{l}}_2$, $\overset{\ast}{D}=2\overset{\ast}{\Lambda}{}^{l\dot{l}}_3+ \widetilde{{\sf A}}^{l\dot{l}}_2\overset{\ast}{\Lambda}{}^{l\dot{l}}_1- \widetilde{{\sf A}}^{l\dot{l}}_1\overset{\ast}{\Lambda}{}^{l\dot{l}}_2$, $\overset{\ast}{E}=2i\overset{\ast}{\Lambda}{}^{l\dot{l}}_3+ \widetilde{{\sf B}}^{l\dot{l}}_2\overset{\ast}{\Lambda}{}^{l\dot{l}}_1- \widetilde{{\sf B}}^{l\dot{l}}_1\overset{\ast}{\Lambda}{}^{l\dot{l}}_2$ are \begin{equation}\label{UM} E:\quad {\renewcommand{\arraystretch}{1.3} \left\{\begin{array}{ccc} e^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,mm; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& ic^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}(\dot{l}-l-2) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)},\\ e^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,mm; \dot{l}\dot{l},\dot{m}\dot{m}}&=& -2ic^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll;\dot{l}\dot{l}}m\dot{m},\\ e^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,mm; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& ic^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}(l-\dot{l}-2) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}. \end{array}\right.} \end{equation} \begin{equation}\label{DM} \overset{\ast}{D}:\quad {\renewcommand{\arraystretch}{1.3} \left\{\begin{array}{ccc} \overset{\ast}{d}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,mm; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& \overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}(l+\dot{l}) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)},\\ \overset{\ast}{d}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,mm; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& -\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}(l+\dot{l}+2) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}. \end{array}\right.} \end{equation} \begin{equation}\label{EM} \overset{\ast}{E}:\quad {\renewcommand{\arraystretch}{1.3} \left\{\begin{array}{ccc} \overset{\ast}{e}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l,mm; \dot{l}-1,\dot{l},\dot{m}\dot{m}}&=& i\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l-1,l; \dot{l}-1,\dot{l}}(\dot{l}-l+2) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)},\\ \overset{\ast}{e}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll,mm; \dot{l}\dot{l},\dot{m}\dot{m}}&=& 2i\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{ll; \dot{l}\dot{l}}m\dot{m},\\ \overset{\ast}{e}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l,mm; \dot{l}+1,\dot{l},\dot{m}\dot{m}}&=& i\overset{\ast}{c}{}^{k^\prime k;\dot{k}^\prime\dot{k}}_{l+1,l; \dot{l}+1,\dot{l}}(l-\dot{l}+2) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}. \end{array}\right.} \end{equation} The system (\ref{Complex5}) in the components $\psi^{k\dot{k}}_{lm;\dot{l}\dot{m}}$ can be written as \begin{multline} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \left[ \frac{1}{r\sin\theta^c} a^{kk^\prime}_{ll^\prime,mm^\prime}\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\varphi}- \frac{1}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes a^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\varphi}+\right.\\ +\frac{i}{r\sin\theta^c} a^{kk^\prime}_{ll^\prime,mm^\prime}\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\epsilon}+ \frac{i}{r\sin\dot{\theta}^c}\boldsymbol{1}_{2l+1}\otimes a^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\epsilon}-\\ -\frac{1}{r} b^{kk^\prime}_{ll^\prime,mm^\prime}\otimes\boldsymbol{1}_{2\dot{l}+1} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\theta}+ \frac{1}{r}\boldsymbol{1}_{2l+1}\otimes b^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\theta}- \nonumber \end{multline} \begin{multline} -\frac{i}{r}b^{kk^\prime}_{ll^\prime,mm^\prime}\otimes \boldsymbol{1}_{2\dot{l}+1} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\tau}- \frac{i}{r}\boldsymbol{1}_{2l+1}\otimes b^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial\tau}+\\ \left.+(1+i) c^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll^\prime,mm^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \frac{\partial\psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}}{\partial r}+ \frac{1}{r} d^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll^\prime,mm^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}+ \frac{1}{r} e^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll^\prime,mm^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}+\right.\\ \left.+\frac{2i}{r}\cot\theta^c a^{kk^\prime}_{ll^\prime,mm^\prime}\otimes\boldsymbol{1}_{2\dot{l}+1} m^\prime \psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}+ \frac{2i}{r}\cot\dot{\theta}^c\boldsymbol{1}_{2l+1}\otimes a^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \dot{m}^\prime \psi^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}\right]+\\ +\kappa^c\psi^{k\dot{k}}_{lm;\dot{l}\dot{m}}=0, \nonumber \end{multline} \begin{multline} \sum_{l^\prime,m^\prime,\dot{l}^\prime,\dot{m}^\prime,k^\prime,\dot{k}^\prime} \left[ \frac{1}{r^\ast\sin\theta^c} \overset{\ast}{a}{}^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime, \dot{m}\dot{m}^\prime}\otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\varphi}- \frac{1}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{a}{}^{kk^\prime}_{ll^\prime, mm^\prime} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\varphi}+\right.\\ +\frac{i}{r^\ast\sin\theta^c} \overset{\ast}{a}{}^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime, \dot{m}\dot{m}^\prime}\otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\epsilon}+ \frac{i}{r^\ast\sin\dot{\theta}^c}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{a}{}^{kk^\prime}_{ll^\prime, mm^\prime} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\epsilon}- \\ -\frac{1}{r^\ast} \overset{\ast}{b}{}^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime, \dot{m}\dot{m}^\prime}\otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\theta}+ \frac{1}{r^\ast}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{b}{}^{kk^\prime}_{ll^\prime, mm^\prime} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\theta}- \nonumber \end{multline} \begin{multline} -\frac{i}{r^\ast} \overset{\ast}{b}{}^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime, \dot{m}\dot{m}^\prime}\otimes\boldsymbol{1}_{2l+1} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\tau}- \frac{i}{r^\ast}\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{b}{}^{kk^\prime}_{ll^\prime, mm^\prime} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial\tau}+\\ +(1-i) \overset{\ast}{c}{}^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll^\prime,mm^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \frac{\partial\dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}} {\partial r^\ast} +\frac{1}{r^\ast} \overset{\ast}{d}{}^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll^\prime,mm^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}+ \frac{1}{r^\ast} \overset{\ast}{e}{}^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll^\prime,mm^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime} \dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}+\\ \left.+\frac{2i}{r^\ast}\cot\theta^c \overset{\ast}{a}{}^{\dot{k}\dot{k}^\prime}_{\dot{l}\dot{l}^\prime, \dot{m}\dot{m}^\prime}\otimes\boldsymbol{1}_{2l+1}\dot{m}^\prime \dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime;\dot{l}^\prime\dot{m}^\prime}+ \frac{2i}{r^\ast}\cot\dot{\theta}^c\boldsymbol{1}_{2\dot{l}+1}\otimes \overset{\ast}{a}{}^{kk^\prime}_{ll^\prime, mm^\prime}m^\prime \dot{\psi}^{k^\prime\dot{k}^\prime}_{l^\prime m^\prime; \dot{l}^\prime\dot{m}^\prime}\right]+\\ +\dot{\kappa}^c\dot{\psi}^{k\dot{k}}_{lm;\dot{l}\dot{m}}=0, \label{Complex6} \end{multline}\begin{sloppypar}\noindent where the coefficients $a^{k k^\prime;\dot{k}\dot{k}^\prime}_{l l^\prime,m m^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime}$, $\ldots$, $\overset{\ast}{e}{}^{k k^\prime;\dot{k}\dot{k}^\prime}_{l l^\prime, m m^\prime;\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime}$ are defined by the formulae (\ref{L1}), (\ref{L2}), (\ref{L3}), (\ref{VM}), (\ref{L1'}), (\ref{L2'}), (\ref{L3'}), (\ref{UM}), respectively. \end{sloppypar} With the view to separate the variables in (\ref{Complex6}) let us assume that \begin{eqnarray} \psi^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=&\boldsymbol{f}^{l_0\dot{l}_0}_{lmk; \dot{l}\dot{m}\dot{k}}(r) \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0),\nonumber\\ \dot{\psi}^{k\dot{k}}_{lm;\dot{l}\dot{m}}&=& \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk;\dot{l}\dot{m}\dot{k}} (r^\ast) \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0),\label{F} \end{eqnarray} where $l_0\ge l$, $-l_0\le m,n\le l_0$ and $\dot{l}_0\ge\dot{l}$, $-\dot{l}_0\le\dot{m},\dot{n}\le\dot{l}_0$. Substituting the functions (\ref{F}) into the system (\ref{Complex6}) and taking into account values of the coefficients $a^{k k^\prime;\dot{k}\dot{k}^\prime}_{l l^\prime,m m^\prime; \dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime}$, $\ldots$, $\overset{\ast}{e}{}^{k k^\prime;\dot{k}\dot{k}^\prime}_{l l^\prime, m m^\prime;\dot{l}\dot{l}^\prime,\dot{m}\dot{m}^\prime}$, we collect together the terms with identical radial functions.\index{function!radial} In the result we obtain \begin{multline} \sum_{k^\prime,\dot{k}^\prime} c^{kk^\prime;\dot{k}\dot{k}^\prime}_{l-1,l;\dot{l}-1,\dot{l}} \left\{ \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)}\left[ (1+i)\frac{\partial\boldsymbol{f}^{l_0\dot{l}_0}_{l-1,m,k^\prime; \dot{l}-1,\dot{m}\dot{k}^\prime}} {\partial r}+\right.\right.\\ +\left.\left.\frac{1}{r}(l-il+\dot{l}+i\dot{l}-2) \boldsymbol{f}^{l_0}_{l-1,m,k^\prime}\right] \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0) +\right.\\ +\frac{1}{2r}\sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,m-1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} \left[-\frac{1}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\varphi}- \right.\\ -\frac{i}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\epsilon} \left.+i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\theta}- \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\tau}- \frac{2i(m-1)\cos\theta^c}{\sin\theta^c} \mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]+ \nonumber \end{multline} \begin{multline} +\frac{1}{2r}\sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,m+1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} \left[\frac{1}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\varphi}+ \right.\\ +\frac{i}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\epsilon}+ \left. \left.+i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(m+1)\cos\theta^c}{\sin\theta^c} \mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]\right.+ \\ +\frac{1}{2r}\sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,mk^\prime;\dot{l}-1,\dot{m}-1,\dot{k}^\prime} \left[-\frac{1}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\varphi}+ \right. \nonumber \end{multline} \begin{multline} \left.+\frac{i}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\epsilon} +i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\theta}+ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\tau}+ \frac{2i(\dot{m}-1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}\right]+ \\ +\frac{1}{2r}\sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,mk^\prime;\dot{l}-1,\dot{m}+1,\dot{k}^\prime} \left[\frac{1}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\varphi}- \right.\\ -\frac{i}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\epsilon}+ \left. \left.+i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\theta}+ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\tau}- \frac{2i(\dot{m}+1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}\right]\right\}+ \nonumber \end{multline} \[ +\sum_{k^\prime,\dot{k}^\prime} c^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll;\dot{l}\dot{l}}\left\{\left[ (1+i)m\dot{m} \frac{\partial\boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l}\dot{m}\dot{k}^\prime}}{\partial r} -\frac{2i}{r}m\dot{m} \boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l}\dot{m}\dot{k}^\prime}\right] \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)+\right. \] \begin{multline} +\frac{1}{2r}\dot{m}\sqrt{(l+m)(l-m+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l,m-1,k^\prime;\dot{l}\dot{m}\dot{k}^\prime} \left[\frac{1}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\varphi}+ \frac{i}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\epsilon} \right.-\\ \left.-i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\theta}+ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(m-1)\cos\theta^c}{\sin\theta^c} \mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]+ \\ +\frac{1}{2r}\dot{m}\sqrt{(l+m+1)(l-m)} \boldsymbol{f}^{l_0\dot{l}_0}_{l,m+1,k^\prime;\dot{l}\dot{m}\dot{k}^\prime} \left[\frac{1}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\varphi}+ \frac{i}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\epsilon}+ \right. \nonumber \end{multline} \begin{multline} \left. \left.+i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(m+1)\cos\theta^c}{\sin\theta^c} \mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]\right\}+ \\ +\frac{1}{2r}m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l},\dot{m}-1,\dot{k}^\prime} \left[\frac{1}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\varphi}- \frac{i}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\epsilon} \right.-\\ \left.-i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\theta}- \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\tau}- \frac{2i(\dot{m}-1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}\right]+ \nonumber \end{multline} \begin{multline} +\frac{1}{2r}m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})} \boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l},\dot{m}+1,\dot{k}^\prime} \left[\frac{1}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\varphi}- \frac{i}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\epsilon}+ \right.\\ \left. \left.+i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\theta}+ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\tau}- \frac{2i(\dot{m}+1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}\right]\right\}+ \nonumber \end{multline} \begin{multline} +\sum_{k^\prime,\dot{k}^\prime} c^{kk^\prime;\dot{k}\dot{k}^\prime}_{l+1,l;\dot{l}+1,\dot{l}}\left\{ \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}\left[ (1+i)\frac{\partial \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,m,k^\prime;\dot{l}+1,\dot{m}\dot{k}^\prime}} {\partial r}+\right.\right.\\ +\left.\left.\frac{1}{r}(il-l-\dot{l}-i\dot{l}-2) \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,m,k^\prime;\dot{l}+1,\dot{m}\dot{k}^\prime} \right] \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)+\right.\\ +\frac{1}{2r}\sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,m-1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} \left[\frac{1}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\varphi}+ \right.\nonumber \end{multline} \begin{multline} \left. +\frac{i}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\epsilon} -i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\theta}+ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\tau}+ \frac{2i(m-1)\cos\theta^c}{\sin\theta^c} \mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]+ \\ +\frac{1}{2r}\sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,m+1,k^\prime;\dot{l}+1,\dot{m}\dot{k}^\prime} \left[-\frac{1}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\varphi}- \right.\\ -\frac{i}{\sin\theta^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\epsilon}- \left. \left.-i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\theta}+ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\tau}- \frac{2i(m+1)\cos\theta^c}{\sin\theta^c} \mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]\right.+ \nonumber \end{multline} \begin{multline} +\frac{1}{2r}\sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,mk^\prime;\dot{l}-1,\dot{m}-1,\dot{k}^\prime} \left[\frac{1}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\varphi}- \right.\\ -\frac{i}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\epsilon} \left.-i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\theta}- \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}}{\partial\tau}+ \frac{2i(\dot{m}-1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}\right]+ \\ +\frac{1}{2r}\sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,mk^\prime;\dot{l}+1,\dot{m}+1,\dot{k}^\prime} \left[-\frac{1}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\varphi}+ \right. \nonumber \end{multline} \begin{multline} +\frac{i}{\sin\dot{\theta}^c} \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\epsilon}- \left. \left.-i\frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\theta}- \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}}{\partial\tau}+ \frac{2i(\dot{m}+1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}\right]\right\}+ \\ +\kappa^c\boldsymbol{f}^{l_0\dot{l}_0}_{lm;\dot{l}\dot{m}} \mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)=0, \nonumber \end{multline} \begin{multline} \sum_{\dot{k}^\prime}\overset{\ast}{c}{}^{kk^\prime;\dot{k} \dot{k}^\prime}_{l-1,l;\dot{l}-1,\dot{l}} \left\{ \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)}\left[(1-i) \frac{\partial\overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{ l-1,mk^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime}} {\partial r^\ast}-\right.\right.\\ -\left.\left.\frac{1}{r^\ast}(l-il+\dot{l}+i\dot{l}+2) \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{ l-1,mk^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime}\right] \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)+\right.\\ +\frac{1}{2r^\ast}\sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{ l-1,mk^\prime;\dot{l}-1,\dot{m}-1,\dot{k}^\prime}\left[ -\frac{1}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\varphi}-\right.\\ -\frac{i}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\epsilon} \left.+i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1, \dot{n}}}{\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\tau}- \frac{2i(\dot{m}-1)\cos\theta^c}{\sin\theta^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}\right]+ \nonumber \end{multline} \begin{multline} +\frac{1}{2r^\ast}\sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,mk^\prime; \dot{l}-1,\dot{m}+1,\dot{k}^\prime}\left[ \frac{1}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\varphi}+\right.\\ +\frac{i}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\epsilon} \left. \left.+i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1, \dot{n}}}{\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\tau}+ \frac{2i(\dot{m}+1)\cos\theta^c}{\sin\theta^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}\right]\right.+ \\ +\frac{1}{2r^\ast}\sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{ l-1,m-1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime}\left[ -\frac{1}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\varphi}+\right. \nonumber \end{multline} \begin{multline} +\frac{i}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\epsilon} \left.+i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m} \dot{n}}}{\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\tau}+ \frac{2i(m-1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]+ \\ +\frac{1}{2r^\ast}\sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,m+1,k^\prime; \dot{l}-1,\dot{m}\dot{k}^\prime}\left[ \frac{1}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\varphi}-\right.\\ -\frac{i}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\epsilon} \left. \left.+i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m} \dot{n}}}{\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\tau}- \frac{2i(m+1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]\right\}+ \\ +\sum_{k^\prime,\dot{k}^\prime} \overset{\ast}{c}^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll; \dot{l}\dot{l}}\left\{\left[ (1-i)m\dot{m} \frac{\partial\overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l},\dot{m},\dot{k}^\prime}} {\partial r^\ast}+\frac{2i}{r^\ast}m\dot{m} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l},\dot{m},\dot{k}^\prime}\right] \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m},\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)+\right. \nonumber \end{multline} \begin{multline} +\frac{1}{2r^\ast}m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^prime; \dot{l},\dot{m}-1,\dot{k}^\prime}\left[ \frac{1}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\varphi}+ \frac{i}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\epsilon}-\right.\\ \left.-i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1, \dot{n}}}{\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\tau}+ \frac{2i(\dot{m}-1)\cos\theta^c}{\sin\theta^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}\right]+ \\ +\frac{1}{2r^\ast}m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l},\dot{m}+1, \dot{k}^\prime}\left[ \frac{1}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\varphi}+ \frac{i}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\epsilon}+\right. \nonumber \end{multline} \begin{multline} \left. \left.+i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1, \dot{n}}}{\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\tau}+ \frac{2i(\dot{m}+1)\cos\theta^c}{\sin\theta^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}\right]\right.+ \\ +\frac{1}{2r^\ast}\dot{m}\sqrt{(l+m)(l-m+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l,m-1,k^prime; \dot{l}\dot{m}\dot{k}^\prime}\left[ \frac{1}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\varphi}- \frac{i}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\epsilon}-\right.\\ \left.-i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m} \dot{n}}}{\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\tau}- \frac{2i(m-1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]+ \nonumber \end{multline} \begin{multline} +\frac{1}{2r^\ast}\dot{m}\sqrt{(l+m+1)(l-m)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l,m+1,k^\prime;\dot{l}\dot{m} \dot{k}^\prime}\left[ \frac{1}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\varphi}- \frac{i}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\epsilon}+\right. \\ \left. \left.+i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m} \dot{n}}}{\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\tau}- \frac{2i(m+1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]\right\}+ \\ +\sum_{k^\prime,\dot{k}^\prime} \overset{\ast}{c}^{kk^\prime;\dot{k}\dot{k}^\prime}_{l+1,l; \dot{l}+1,\dot{l}}\left\{\sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)}\left[ (1-i) \frac{\partial\overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m},\dot{k}^\prime}} {\partial r^\ast}+\right.\right. \\ +\left.\left.\frac{1}{r^\ast}(il-l-\dot{l}-i\dot{l}-2+i2) \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m},\dot{k}^\prime}\right] \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)+\right.\\ +\frac{1}{2r^\ast}\sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m}-1,\dot{k}^\prime}\left[ \frac{1}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\varphi}+\right. \nonumber \end{multline} \begin{multline} +\frac{i}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\epsilon} \left.-i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn; \dot{m}-1,\dot{n}}}{\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}} {\partial\tau}+ \frac{2i(\dot{m}-1)\cos\theta^c}{\sin\theta^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}\right]+ \\ +\frac{1}{2r^\ast}\sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m}+1,\dot{k}^\prime}\left[ -\frac{1}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\varphi}-\right.\\ -\frac{i}{\sin\theta^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\epsilon} \left. \left.-i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn; \dot{m}+1,\dot{n}}}{\partial\theta}+ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}} {\partial\tau}- \frac{2i(\dot{m}+1)\cos\theta^c}{\sin\theta^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}\right]\right.+ \nonumber \end{multline} \begin{multline} +\frac{1}{2r^\ast}\sqrt{((l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,m-1,k^\prime; \dot{l}+1,\dot{m}\dot{k}^\prime}\left[ \frac{1}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\varphi}-\right.\\ -\frac{i}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\epsilon} \left.-i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n; \dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\tau}- \frac{2i(m-1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]+ \\ +\frac{1}{2r^\ast}\sqrt{((l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,m+1,k^\prime; \dot{l}+1,\dot{m}\dot{k}^\prime}\left[ -\frac{1}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\varphi}+\right.\\ +\frac{i}{\sin\dot{\theta}^c} \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\epsilon} \left. \left.-i\frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n; \dot{m}\dot{n}}}{\partial\theta}- \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\tau}+ \frac{2i(m+1)\cos\dot{\theta}^c}{\sin\dot{\theta}^c} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]\right\}+ \\ +\dot{\kappa}^c \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lm;\dot{l}\dot{m}} \overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)=0. \label{Complex7} \end{multline} \begin{sloppypar} \noindent The each equation of the system obtained contains five generalized hyperspherical functions\index{function!hyperspherical!generalized} $\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}$, $\mathfrak{M}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}$, $\mathfrak{M}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}$, $\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}-1,\dot{n}}$, $\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}+1,\dot{n}}$ and their conjugate. Let us recall that $\mathfrak{M}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)= e^{-n(\epsilon+i\varphi)-\dot{n}(\epsilon-i\varphi)} \mathfrak{Z}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}(\theta,\tau)$, $\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}(\varphi,\epsilon,\theta,\tau,0,0)= e^{-n(\epsilon+i\varphi)-\dot{n}(\epsilon-i\varphi)} \mathfrak{Z}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}(\theta,\tau)$ and $\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)= e^{-n(\epsilon-i\varphi)-\dot{n}(\epsilon-i\varphi)} \overset{\ast}{\mathfrak{Z}}{}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}(\theta,\tau)$, $\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}} (\varphi,\epsilon,\theta,\tau,0,0)= e^{-n(\epsilon-i\varphi)-\dot{n}(\epsilon-i\varphi)} \overset{\ast}{\mathfrak{Z}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}(\theta,\tau)$. Therefore,\end{sloppypar} \[ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}}{\partial\varphi}= (-in+i\dot{n})\mathfrak{M}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}},\quad \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}}{\partial\epsilon}= -(n+\dot{n})\mathfrak{M}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}, \] \[ \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}}{\partial\varphi}= (-in+i\dot{n})\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}},\quad \frac{\partial\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}}{\partial\epsilon}= -(n+\dot{n})\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}, \] \[ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}} {\partial\varphi}= (in-i\dot{n})\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}, \quad \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}} {\partial\epsilon}= -(n+\dot{n})\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{m\pm 1,n;\dot{m}\dot{n}}, \] \[ \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}} {\partial\varphi}= (in-i\dot{n})\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}, \quad \frac{\partial\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}} {\partial\epsilon}= -(n+\dot{n})\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\pm 1,\dot{n}}. \] We apply now the recurrence relations (\ref{RZ1})--(\ref{RZ4}), (\ref{RZ5})--(\ref{RZ8}) to square brackets containing the hyperspherical functions. For example, in virtue of (\ref{RZ2}) the second bracket in (\ref{Complex7}) can be written as \begin{multline} e^{-n(\epsilon+i\varphi)-\dot{n}(\epsilon-i\varphi)}\left[ i\frac{\partial\mathfrak{Z}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}} {\partial\theta}- \frac{\partial\mathfrak{Z}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}}{\partial\tau}+\right.\\ \left. +\frac{2i(n-(m-1)\cos\theta^c)}{\sin\theta^c} \mathfrak{Z}^{l_0\dot{l}_0}_{m-1,n;\dot{m}\dot{n}}\right]= 2\sqrt{(l_0+m)(l_0-m+1)}\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}.\label{RR1} \end{multline} Further, in virtue of (\ref{RZ1}) for the third bracket we have \begin{multline} e^{-n(\epsilon+i\varphi)-\dot{n}(\epsilon-i\varphi)}\left[ i\frac{\partial\mathfrak{Z}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}} {\partial\theta}- \frac{\partial\mathfrak{Z}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}}{\partial\tau}-\right.\\ \left. -\frac{2i(n-(m+1)\cos\theta^c)}{\sin\theta^c} \mathfrak{Z}^{l_0\dot{l}_0}_{m+1,n;\dot{m}\dot{n}}\right]= 2\sqrt{(l_0+m+1)(l_0-m)}\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}.\label{RR2} \end{multline} \begin{sloppypar}\noindent and so on. In doing so, we replace all the square brackets in the system (\ref{Complex7}) via the relations of the type (\ref{RR1})--(\ref{RR2}) and cancel all the equations by $\mathfrak{M}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}$ ($\overset{\ast}{\mathfrak{M}}{}^{l_0\dot{l}_0}_{mn;\dot{m}\dot{n}}$). In the result we see that the relativistically invariant system is reduced to a system of ordinary differential equations,\end{sloppypar} \begin{multline} \sum_{k^\prime,\dot{k}^\prime}c^{kk^\prime;\dot{k}\dot{k}^\prime}_{l-1,l; \dot{l}-1,\dot{l}}\left[(1+i) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)} \frac{d\boldsymbol{f}^{l_0\dot{l}_0}_{l-1,mk^\prime;\dot{l}-1, \dot{m}\dot{k}^\prime}}{d r}+\right.\\ +\frac{1}{r}(l-il+\dot{l}+i\dot{l}-2i) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,mk^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} +\\ +\frac{1}{r}\sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)} \sqrt{(l_0+m)(l_0-m+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,m-1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} +\\ \left.+\frac{1}{r}\sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)} \sqrt{(l_0+m+1)(l_0-m)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,m+1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} \right.+\\ +\frac{1}{r}\sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)} \sqrt{(\dot{l}_0+\dot{m})(\dot{l}_0-\dot{m}+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,mk^\prime;\dot{l}-1,\dot{m}-1,\dot{k}^\prime} +\\ \left.+\frac{1}{r}\sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)} \sqrt{(\dot{l}_0+\dot{m}+1)(\dot{l}_0-\dot{m})} \boldsymbol{f}^{l_0\dot{l}_0}_{l-1,mk^\prime;\dot{l}-1,\dot{m}+1,\dot{k}^\prime} \right]+ \nonumber \end{multline} \begin{multline} +\sum_{k^\prime,\dot{k}^\prime}c^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll; \dot{l}\dot{l}}\left[(1+i) m\dot{m}\frac{d\boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l}\dot{m}\dot{k}}} {d r}- \frac{2i}{r}m\dot{m}\boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l}\dot{m} \dot{k}^\prime}-\right. \\ -\frac{1}{r}\dot{m}\sqrt{(l+m)(l-m+1)} \sqrt{(l_0+m)(l_0-m+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l,m-1,k^\prime;\dot{l}\dot{m}\dot{k}^\prime}+\\ \left.+\frac{1}{r}\dot{m}\sqrt{(l+m+1)(l-m)} \sqrt{(l_0+m+1)(l_0-m)} \boldsymbol{f}^{l_0\dot{l}_0}_{l,m+1,k^\prime;\dot{l}\dot{m}\dot{k}^\prime} \right.+\\ -\frac{1}{r}m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)} \sqrt{(\dot{l}_0+\dot{m})(\dot{l}_0-\dot{m}+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l},\dot{m}-1,\dot{k}^\prime}+\\ \left.+\frac{1}{r}m\sqrt{(\dot{l}+\dot{m}+1)}{(\dot{l}-\dot{m})} \sqrt{(\dot{l}_0+\dot{m}+1)(\dot{l}_0-\dot{m})} \boldsymbol{f}^{l_0\dot{l}_0}_{lmk^\prime;\dot{l},\dot{m}+1,\dot{k}^\prime} \right]+ \nonumber \end{multline} \begin{multline} +\sum_{k^\prime,\dot{k}^\prime}c^{kk^\prime;\dot{k}\dot{k}^\prime}_{l+1,l; \dot{l}+1,\dot{l}}\left[(1+i) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)} \frac{d\boldsymbol{f}^{l_0\dot{l}_0}_{l+1,mk^\prime;\dot{l}+1,\dot{m} \dot{k}^\prime}}{d r}+\right.\\ +\frac{1}{r}(il-l-\dot{l}-i\dot{l}-2) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,mk^\prime;\dot{l}+1,\dot{m}\dot{k}^\prime} -\\ -\frac{1}{r}\sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)} \sqrt{(l_0+m)(l_0-m+1)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,m-1,k^\prime;\dot{l}-1,\dot{m}\dot{k}^\prime} -\\ \left.-\frac{1}{r}\sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)} \sqrt{(l_0+m+1)(l_0-m)} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,m+1,k^\prime;\dot{l}+1,\dot{m}\dot{k}^\prime} \right.-\\ -\frac{1}{r}\sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)} \sqrt{(\dot{l}_0+\dot{m})(\dot{l}_0-\dot{m})} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,mk^\prime;\dot{l}-1,\dot{m}-1,\dot{k}^\prime} -\\ \left.-\frac{1}{r}\sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)} \sqrt{(\dot{l}_0+\dot{m}+1)(\dot{l}_0-\dot{m})} \boldsymbol{f}^{l_0\dot{l}_0}_{l+1,mk^\prime;\dot{l}+1,\dot{m}+1,\dot{k}^\prime} \right]+ \\ +\kappa^c\boldsymbol{f}^{l_0}_{lmk}(r)=0,\nonumber \end{multline} \begin{multline} \sum_{k^\prime,\dot{k}^\prime} \overset{\ast}{c}{}^{kk^\prime;\dot{k}\dot{k}^\prime}_{l-1,l; \dot{l}-1,\dot{l}}\left[(1-i) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)} \frac{d\overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,mk^\prime \dot{l}-1,\dot{m}\dot{k}^\prime}} {d r^\ast}+\right.\\ +\frac{1}{r^\ast}(l-il+\dot{l}+i\dot{l}+2i) \sqrt{(l^2-m^2)(\dot{l}^2-\dot{m}^2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,mk^\prime; \dot{l}-1,\dot{m}\dot{k}^\prime}+\\ +\frac{1}{r^\ast}\sqrt{(l^2-m^2)(\dot{l}+\dot{m})(\dot{l}+\dot{m}-1)} \sqrt{(\dot{l}_0+\dot{m})(\dot{l}_0-\dot{m}+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,mk^\prime; \dot{l}-1,\dot{m}-1,\dot{k}^\prime}+\\ \left. +\frac{1}{r^\ast}\sqrt{(l^2-m^2)(\dot{l}-\dot{m})(\dot{l}-\dot{m}-1)} \sqrt{(\dot{l}_0+\dot{m}+1)(\dot{l}_0-\dot{m})} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,mk^\prime; \dot{l}-1,\dot{m}+1,\dot{k}^\prime}\right.+\\ +\frac{1}{r^\ast}\sqrt{(l+m)(l+m-1)(\dot{l}^2-\dot{m}^2)} \sqrt{(l_0+m)(l_0-m+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,m-1,k^\prime; \dot{l}-1,\dot{m}\dot{k}^\prime}+\\ \left.+\frac{1}{r^\ast}\sqrt{(l-m)(l-m-1)(\dot{l}^2-\dot{m}^2)} \sqrt{(l_0+m+1)(l_0-m)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l-1,m+1,k^\prime; \dot{l}-1,\dot{m}\dot{k}^\prime}\right]+\\ \nonumber \end{multline} \begin{multline} +\sum_{k^\prime,\dot{k}^\prime} \overset{\ast}{c}{}^{kk^\prime;\dot{k}\dot{k}^\prime}_{ll;\dot{l}\dot{l}}\left[ (1-i)m\dot{m} \frac{d\overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l}\dot{m}\dot{k}^\prime}} {d r^\ast} +\frac{2i}{r^\ast}m\dot{m} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l}\dot{m}\dot{k}^\prime}-\right. \\ -\frac{1}{r^\ast}m\sqrt{(\dot{l}+\dot{m})(\dot{l}-\dot{m}+1)} \sqrt{(\dot{l}_0+\dot{m})(\dot{l}_0-\dot{m})} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l},\dot{m}-1,\dot{k}^\prime}+\\ \left. +\frac{1}{r^\ast}m\sqrt{(\dot{l}+\dot{m}+1)(\dot{l}-\dot{m})} \sqrt{(\dot{l}_0+\dot{m}+1)(\dot{l}_0-\dot{m})} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lmk^\prime; \dot{l},\dot{m}+1,\dot{k}^\prime}\right.- \\ -\frac{1}{r^\ast}\dot{m}\sqrt{(l+m)(l-m+1)} \sqrt{(l_0+m)(l_0-m+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l,m-1,k^\prime; \dot{l}\dot{m}\dot{k}^\prime}+\\ \left. +\frac{1}{r^\ast}\dot{m}\sqrt{(l+m+1)(l-m)} \sqrt{(l_0+m+1)(l_0-m)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l,m+1,k^\prime; \dot{l}\dot{m}\dot{k}^\prime}\right]+ \nonumber \end{multline} \begin{multline} +\sum_{k^\prime,\dot{k}^\prime} \overset{\ast}{c}{}^{kk^\prime;\dot{k}\dot{k}^\prime}_{l+1,l; \dot{l}+1,\dot{l}}\left[(1-i) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)} \frac{d\overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m}\dot{k}^\prime}} {d r^\ast}+\right.\\ +\frac{1}{r^\ast}(il-l-\dot{l}-i\dot{l}-2+2i) \sqrt{((l+1)^2-m^2)((\dot{l}+1)^2-\dot{m}^2)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m}\dot{k}^\prime}+\\ +\frac{1}{r^\ast}\sqrt{((l+1)^2-m^2)(\dot{l}-\dot{m}+1)(\dot{l}-\dot{m}+2)} \sqrt{(\dot{l}_0+\dot{m})(\dot{l}_0-\dot{m}+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m}-1,\dot{k}^\prime}-\\ \left. -\frac{1}{r^\ast}\sqrt{((l+1)^2-m^2)(\dot{l}+\dot{m}+1)(\dot{l}+\dot{m}+2)} \sqrt{(\dot{l}_0+\dot{m}+1)(\dot{l}_0-\dot{m})} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,mk^\prime; \dot{l}+1,\dot{m}+1,\dot{k}^\prime}\right.- \\ -\frac{1}{r^\ast}\sqrt{(l-m+1)(l-m+2)((\dot{l}+1)^2-\dot{m}^2)} \sqrt{(l_0+m)(l_0-m+1)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,m-1,k^\prime; \dot{l}+1,\dot{m}\dot{k}^\prime}-\\ \left. -\frac{1}{r^\ast}\sqrt{(l+m+1)(l+m+2)((\dot{l}+1)^2-\dot{m}^2)} \sqrt{(l_0+m+1)(l_0-m)} \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{l+1,m+1,k^\prime; \dot{l}+1,\dot{m}\dot{k}^\prime}\right]+ \\ +\dot{\kappa}^c \overset{\ast}{\boldsymbol{f}}{}^{l_0\dot{l}_0}_{lm; \dot{l}\dot{m}}=0. \label{RFS} \end{multline} Substituting solutions of this system into the series (\ref{Fourier1}) and (\ref{Fourier2}), we obtain a solution of the boundary value problem.

One day while working in the intensive care unit, our infection control nurse asked me about a patient who developed pneumonia while he was intubated through orotracheal route. Patient was a case of hypertensive intracerebral bleed and he was intubated for airway protection in view of low Glasgow coma scale. He developed pneumonia while he was intubated, but was not put on mechanical ventilator. Later, he was placed on mechanical ventilation in view of increasing oxygen requirement due to pneumonia. The infection control nurse asked me whether the patient developed pneumonia before or after initiation of mechanical ventilation because she wanted to know whether it was ventilator-associated pneumonia (VAP). Her question was interesting since VAP is a confusing terminology for most of the healthcare workers and actually it has nothing to do with ventilator. It is important to understand that terminology becomes important since strategies to prevent VAP are likely to be successful if they are based upon a sound understanding of pathogenesis.^[@B1]^ In terms of pathophysiology, VAP is a misnomer because its occurrence is not related to ventilator *per se* but to the presence of endotracheal tube (ETT). The ETT allows direct access to the lower respiratory tract, impairing the cough reflex and mucociliary clearance, and provides incomplete sealing to secretions above the cuff. Microaspiration of colonized secretions from oropharyngeal cavities, sinuses, and gastrointestinal tract and biofilm formation are the two most important mechanisms in the development of VAP.^[@B2]^ The oropharynx becomes rapidly colonized by aerobic gram-negative bacteria after illness, antibiotic administration, and hospital admission. These contaminated secretions pool above the ETT cuff and slowly gain access to the lower airway through a fold in the wall of the cuff. A bacterial biofilm, which is impervious to antibiotics, gradually forms on the inner surface of the tube and serves as a nidus for infection.^[@B3]^ So it is important to understand that ETT is major contributor to VAP pathogenesis, not ventilator per se, since many of our patients are on ETT or tracheal tube without any ventilator support. **Source of support:** Nil **Conflict of interest:** None

INTRODUCTION {#s1} ============ Desmoid tumours comprise 0.03% of all neoplasms, less than 3% of all soft tissue tumours ([@R1]), with an incidence of 2-4 cases / 1'000'000 per year ([@R2]). They are usually detected relatively early on in life (10-40 years of age), with a slight female predominance. While normally occurring sporadically, they are more common in patients with familial adenomatous polyposis (FAP), where the incidence is 850 times higher than in the general population ([@R3]). Most desmoid tumours typically tend to grow slowly and do not metastasise, although their growth pattern remains quite aggressive, with organ infiltration, local tissue destruction and organ compression/displacement. The tumours often grow without causing any clinical symptoms and are therefore usually only detected relatively late or during a check-up for another ailment, when they may have already reached substantial sizes of 15cm in diameter or more. Problems associated with intestinal, vascular or neural compression are quite common. CASE REPORT {#s2} =========== During a routine check-up for hypertension, the 43-year old female patient's general practitioner noted a painless mass in the abdomen. The patient reported increased fatigue and a reduced overall performance for the previous two years. Further questioning revealed occasional dyspepsia and an increased urge to urinate, although micturition was otherwise without any associated discomfort. She was otherwise in good general health and without previous abdominal surgery. Routine haematology and biochemistry were normal. An abdominal computed tomography revealed a very large, well-defined intra-abdominal mass, resulting in the displacement of most intra-abdominal organs to the right of the patient's abdominal cavity ([Fig. 1](#fig1){ref-type="fig"}). {#fig1} Radiological evaluation did not make it possible to tell from which organ the tumour originated. Intra-operative findings revealed a tumour spanning from the symphysis to the xyphoid ([Fig. 2](#fig2){ref-type="fig"}) and showed multiple adherences between the omentum and the caudal pole of the spleen. In order to achieve an R0-resection, a partial splenectomy and partial removal of the omentum was necessary. The tumour weighed 5.6kg and measured 28cm. {#fig2} DISCUSSION {#s3} ========== Diagnostic imaging includes abdominal sonography, computed tomography or magnetic resonance imaging (MRI). MRI usually is the examination of choice as it provides the surgeon with a better evaluation of exact tumour size, location and infiltration in to the surrounding tissues with tumour recurrences being more easily detected. Upon diagnosis of a desmoid tumour all patients should have a colonoscopy and a gastroscopy to exclude FAP. Histologically, desmoid tumours are benign and characterised by a monoclonal fibroblastic proliferation typically arising from muscular or aponeurotic structures. Although upon initial inspection an intact capsule seems to surround the periphery, the tumour characteristically extends beyond this pseudo-capsule. Tumours are classified as intra- or extra-abdominal, with extra-abdominal tumours most often located around the neck and shoulders, the pelvis, or within the thoracic wall. Sporadic desmoid tumours are mainly located extra-abdominal and are relatively easy to resect. High recurrence rates and difficult locations make treatment of the intra-abdominal sub-group particularly difficult. Treatment of desmoid tumours is challenging and the optimal therapeutic strategy remains a matter of debate. While spontaneous tumour regression has been observed, usually in menopausal women or post- oophorectomy, suggesting a certain estrogen-dependency ([@R4],[@R5]), rapid growing tumours require an aggressive therapy if cure is to be sought. Their high recurrence rate and their ability and tendency to infiltrate vital organs and structures can result in difficult and highly complex operations. Treatment for small, non-infiltrating tumours can be delayed in completely asymptomatic patients and / or in patients unfit for surgery, as the tumours can remain stable over years without any significant morbidity. However, these patients require routine radiological imaging to exclude rapid growth or new organ infiltration. Surgery remains the main therapeutic option, with complete tumour resection providing the only potentially curative treatment. Although, as with classical cancer surgery, the aim is to achieve an R0 resection margin, surgery should nevertheless be kept to the required minimum, avoiding any mutilating intervention. If resected early on, the removal of abdominal-wall desmoid tumours is associated with a low morbidity and mortality. Overall recurrence rates lie by 20% to 75% ([@R6],[@R7]) and are more frequent for the rarer extra-abdominal tumours. Surgical interventions for intra-abdominal tumours are often associated with a high mortality and morbidity. Many are found at the root of the small bowel mesenterium and up to 65% of the tumours are unresectable upon intraoperative inspection ([@R8]). There is little evidence for the benefit of pharmacological therapies, which include the use of non-steroidal anti-inflammatory drugs, such as sulindac or anti-estrogens such as tamoxifen ([@R9]). The aim is tumour regression / growth stop and it can be used as a first-line treatment for multiple, intra-abdominal tumours. The high recurrence rate can be reduced by adjuvant and / or neoadjuvant radiotherapy ([@R10]) both for abdominal wall and extra-abdominal tumours. The usefulness of alternative therapies, such as radio frequency ablation is currently under scrutiny. Other options include placement of a percutaneous catheter for use in chemical ablation using acids. The possibility of gene transfer is also subject to intense research. Desmoid tumours are rare and are either found sporadically or more commonly associated with FAP. A classification into abdominal wall tumours as well as extra- and intra-abdominal tumours is important with regard to further treatment and outcome. Clear guidelines outlining possible therapy options -- including helpful algorithms -- are missing. However, symptomatic tumours, especially if located intra-abdominal, should be treated aggressively. Surgery still remains the treatment of choice, although recurrence rates are high, even after R0 resection.

\section{Introduction} Second-order optimality conditions of Karush-Kuhn-Tuc\-ker type and Fritz John one play essential role in development of the contemporary vector optimization. We mention several important works which concern problems with inequality constraints: $[^1]$, $[^2]$, $[^3]$, $[^4]$, $[^5]$. In the present paper, we deal with optimality criteria of Karush-Kuhn-Tucker type and Fritz John one for the nonlinear programming problem with inequality constraints \medskip \noindent Minimize$\quad f(x)\quad$subject to $g(x)\leqq0$,\hfill ${\rm (P)}$ \medskip \noindent where $f:X\to\R^n$ and $g:X\to\R^m$ are given vector real-valued functions, defined on some open set $X$ in the finite-dimensional space $\R^s$. Such results were obtained for scalar problems by Ginchev and Ivanov $[^6]$ We generalize them to multi-objective programming problems. All results given here are obtained for nonsmooth problems in terms of the standard second-order directional derivative. We derive second-order necessary conditions for a (weak) local efficiency for problems with C$^1$ (i.e. continuously differentiable) data. We introduce a new second-order constraint qualification (for short CQ). Its form is similar to Mangasarian-Fromovitz CQ $[^7]$ and we call it second-order Mangasarian-Fromovitz CQ, because it includes second-order derivatives. If the constraints satisfy the Mangasarian-Fromovitz CQ, then they satisfy the second-order one. We show that the necessary conditions without any constraint qualification are sufficient for global efficiency in some problems of generalized convexity type, which are introduced in the paper. We obtain also that the same conditions are necessary and sufficient for weak global efficiency under different generalized convexity assumptions. In the sufficient conditions, we suppose that all components of the objective function are second-order pseudoconvex, a notion recently introduced by Ginchev and Ivanov $[^8]$ \bigskip {\bf 2. Necessary and sufficient conditions for an efficient solution.} In this section, we derive necessary and sufficient optimality conditions for efficiency for the problem {\rm{(P)}}. We begin with some preliminary definitions. \smallskip Denote by $\R$ the set of reals and let $\RR=\R\cup\{-\infty\}\cup\{+\infty\}$. Let the function $\f$ with an open domain $X\subset\R^n$ be differentiable at the point $x\in X$. Then the second-order directional derivative $f^{\pr\pr}(x,u)$ of $f$ at the point $x\in X$ in direction $u\in\E$ is defined as element of $\RR$ by \centerline{$f^{\pr\pr}(x,u)=\lim_{t\to +0}\;2 t^{-2}[f(x+tu)-f(x)-t\nabla f(x)u].$} The function $f$ is called second-order directionally differentiable on $X$ if the derivative $f^{\pr\pr}(x,u)$ exists for each point $x\in X$ and any direction $u\in\R^n$. \smallskip Recall that a scalar function $f:X\to\R$ is said to be quasiconvex at the point $x\in X$ (with respect to $X\subset\R^s$) $[^9]$ if the conditions $y\in X$, $f(y)\leqq f(x)$, $t\in[0,1]$, $(1-t)x+ty\in X$ imply $ f((1-t)x+ty)\leqq f(x)$. If the set $X$ is convex, then the function $f$ is called quasiconvex on $X$ when for all $x,\;y\in X$ and $t\in [0,\,1]$ it holds $ f((1-t)x+ty)\leqq\max\left(f(x),\,f(y)\right). $ Throughout this paper we use the following notations comparing the vectors $x$ and $y$ with components $x_i$ and $y_i$ in finite-dimensional spaces: \centerline{$x<y\quad\textrm{if}\quad x_i<y_i\quad\textrm{for all indices }i;$} \centerline{$x\leqq y\quad\textrm{if}\quad x_i\leqq y_i\quad\textrm{for all indices }i;$} \centerline{$x\le y\quad\textrm{if}\quad x_i\leqq y_i\quad\textrm{for all indices }\quad i\quad\textrm{with at least one being strict.}$} A feasible point $\bar x\in S$ is called a (weak) local Pareto minimizer, or (weakly) efficient if there exists a neigbourhood $U\ni\bar x$ such that there is no $x\in U\cap S$ with $f(x)\le f(\bar x)$ ($f(x)<f(\bar x)$). The point $\bar x\in S$ is called a (weak) global Pareto minimizer if there does not exist $x\in S$ with $f(x)\le f(\bar x)$ ($f(x)<f(\bar x)$). The following lemma is due to Kanniappan $[^{10}]$. \begin{lemma}\label{kan} A point $\bar x\in X$ is Pareto optimal for (P) if and only if $\bar x$ minimizes each component $f_i$ of the vector function $f$ over the constraint set \begin{equation}\label{1} C_i=\{x\in X\mid g(x)\leqq 0,\; f_j(x)\le f_j(\bar x), j\ne i\}. \end{equation} \end{lemma} Consider the following scalar problem: \medskip \noindent Minimize$\quad h_0(x)\quad$subject to$\quad h_i(x)\leqq0$, $i=1,2,...m$,\hfill ${\rm (SP)}$ \medskip \noindent where $h_i:X\to\R$, $i=0,1,...,m$ are given real-valued functions, defined on some open set $X\subset\R^s$. For fixed vectors $\bar x\in\E$ and $d\in\E$, let \centerline{$I_0(\bar x,d):=\{i\in \{0\}\cup I(\bar x)\mid\nabla h_i(\bar x)d=0\}.$} The following theorem is due to Ginchev and Ivanov $[^6]$: \begin{lemma}[Second-order necessary conditions for the scalar problem (SP)]\label{NKT} Let $X$ be an open set in the space $\R^n$, the functions $h_i$ $(i=0,1,...,m)$ be defined on $X$. Suppose that the feasible point $\bar x$ is a local minimizer of the problem {\rm{(SP)}}, the functions $h_i$ belong to the class C$^1$, $i\in\{0\}\cup I(\bar x)\}$, they are second-order directionally differentiable at $\bar x$ in every critical direction $d\in\R^n$, and the functions $h_i$ $(i\notin I(\bar x))$ are continuous at $\bar x$. Then corresponding to any critical direction $d$ there exist nonnegative multipliers $\lambda_0,\lambda_1,...,\lambda_m$, not all zero, such that \[ \lambda_i h_i(\bar x)=0,\;i=1,2,...,m,\quad\nabla L(\bar x)=0, \] \[ \lambda_i\nabla h_i(\bar x)d=0,\quad i\in\{0\}\cup I(\bar x), \] \begin{equation}\label{4} L^{\pr\pr}(\bar x,d)=\sum_{i\in\{0\}\cup I(\bar x)}\lambda_i h_i^{\pr\pr}(\bar x,d)\geqq 0. \end{equation} \end{lemma} Consider the problem {\rm{(P)}}. Denote \centerline{$S:=\{x\in X\mid g_j(x)\le 0,\; j=1,2,\ldots,m\}.$} \noindent For every feasible point $x\in S$ let $A(x)$ be the set of active constraints \centerline{$A(x):=\{j\in\{1,2,\ldots,m\}\mid g_j(x)=0\}.$} A direction $d$ is called critical at the point $x\in S$ if \centerline{$\nabla f_i(x)d\leqq0\quad\textrm{for all}\quad i\in\{1,2,\ldots,n\}\quad\textrm{and}\quad \nabla g_j(x)\leqq 0\quad\textrm{for all}\quad j\in A(x).$} For every critical direction $d$ at the feasible point $x$ denote by $I(x,d)$ and $J(x,d)$ the following index sets: \centerline{$I(x,d):=\{i\in\{1,2,\ldots,n\}\mid\nabla f_i(x)d=0\},\; J(x,d):=\{j\in A(x)\mid\nabla g_j(x)d=0\}.$} It is said that Mangasarian-Fromovitz constraint qualification $[^7]$ holds at the point $\bar x$ if there exists a vector $u$ such that \centerline{$\nabla g_j(\bar x)u>0,\quad j\in A(\bar x).$} \noindent It is usually supposed the following equivalent form: $\nabla g_j(\bar x)u<0$, $j\in A(\bar x)$. We introduce the following more general constraint qualification: \begin{definition} We say that the active constraints satisfy the second-order Man\-ga\-sa\-rian-Fromovitz CQ if for every direction $d$ with \begin{equation}\label{CQ} \nabla g_j(\bar x)d\le 0,\; j\in A(\bar x),\; d\ne 0 \end{equation} there exist a vector $u$ and a number $\omega\ge 0$ which depend on $d$ such that \begin{equation}\label{SOCQ} \nabla g_j(\bar x)u>\omega g_j^{\pr\pr}(\bar x,d),\quad j\in A(\bar x). \end{equation} \end{definition} \begin{proposition} If the active constraints satisfy Mangasarian-Fromovitz CQ and they are second-order directionally differentiable at $\bar x$ in every direction $d$ which satisfies $(\ref{CQ})$, then the second-order Mangasarian-Fromovitz CQ holds. \end{proposition} {\bf Proof.} To show that the second-order CQ holds for every direction such that (\ref{CQ}) holds, we should keep the same vector $u$ and choose $\omega=0$. \qed The following simple example shows that the second-order Mangasarian-Fro\-mo\-vitz CQ is really more general for differentiable and second-order directionally differentiable problems than Mangasarian-Fro\-mo\-vitz CQ. \begin{example} Consider the problem (P) where $X\equiv\R^2$, $f$ is an arbitrary vector or scalar function of two variables whose minimum is attained at $\bar x=(0,0)$ and $g$ is the scalar function \centerline{$g(x_1,x_2)=-x_1^2-x_2^2.$} \noindent The inequality constraint is active at $\bar x$. We have $\nabla g(\bar x)=(0,0)$ and $g^{\prime\prime}(\bar x,d)=-2\norm{d}^2<0$. Therefore, the Mangasarian-Fromovitz CQ is not satisfied, but the second-order CQ holds. \end{example} \begin{remark} We do not suppose that the inequality in the second-order CQ is satisfied for $d=0$. Really, if we assume that the second-order CQ is fulfilled for $d=0$, then both the first-order and second-order constraint qualifications become equivalent, because the second-order CQ when $d=0$ implies that the Mangasarian-Fromovitz CQ holds. \end{remark} \begin{theorem}[Second order necessary conditions for the vector problem (P)]\label{th1} Let $X$ be an open set in the space $\R^s$, the functions $f_i$ $(i=1,2,\ldots,n)$, $g_j$ $(j=1,2,\ldots,m)$ be defined on $X$. Suppose that $\bar x$ is a local Pareto minimizer of the problem {\rm{(P)}}, the functions $g_j$ $(j\notin A(\bar x))$ are continuous at $\bar x$, the vector function $f$ and the functions $g_j$ $(j\in A(\bar x))$ are continuously differentiable, and the functions $f_i$ $(i\in I(\bar x,d))$, $g_j$ $(j\in J(\bar x,d))$ are second-order directionally differentiable at $\bar x$ in any critical direction $d\in\R^n$. Then for every critical direction $d\in\E$, there exist nonnegative Lagrange multipliers $\lambda=(\lambda_1,\lambda_2,\ldots ,\lambda_n)$, $\mu=(\mu_1,\mu_2,\ldots,\mu_m)$ with $(\lambda,\mu)\ne 0$ such that \begin{equation}\label{9} \mu_j g_j(\bar x)=0,\;j=1,2,...,m,\quad\nabla L(\bar x)=0, \end{equation} \[ \lambda_i\nabla f_i(\bar x)d=0,\quad\mu_j\nabla g_j(\bar x)d=0,\; j\in A(\bar x), \] \begin{equation}\label{10} L^{\pr\pr}(\bar x,d)=\sum_{i=1}^n\lambda_i f_i^{\pr\pr}(\bar x,d)+ \sum_{j\in I(\bar x)}\mu_j g_j^{\pr\pr}(\bar x,d) \geqq 0. \end{equation} Assume further that the second-order Mangasarian-Fromovitz constraint qualification holds. Then we could suppose that $\lambda\ne 0$. \end{theorem} {\bf Proof.} The theorem follows directly from Lemmas \ref{kan} and \ref{NKT}, taking into account that the constraints $f_j(x)-f_j(\bar x)\leqq 0$ are active at $\bar x$. We prove that $\lambda\ne 0$ under the assumption that the second-order Man\-ga\-sa\-rian-Fromovitz CQ holds. Suppose the contrary that $\lambda=0$. It follows from (\ref{9}), (\ref{10}), second-order Mangasarian-Fromovitz CQ and $\mu\ne 0$ that \[ 0=[\sum_{j\in A(\bar x)}\mu_j\nabla g_j(\bar x)]u>\omega\sum_{j\in A(\bar x)}\mu_j g_j^{\prime\prime}(\bar x,d)\ge 0 \] \noindent which is impossible. Therefore $\lambda\ne 0$. \qed We introduce the following definition: \begin{definition} We call the problem {\rm (P)} second-order KT pseudoconvex at the point $x\in S$ if the following implication holds: \[ \left. \begin{array}{l} y\in S, \\ f(y)\le f(x) \end{array}\right] \;\Rightarrow\;\left[ \begin{array}{l} \nabla f_i(x)(y-x)\leqq 0,\; i=1,2,\dots,n \\ \nabla f_i(x)(y-x)= 0\quad\textrm{implies}\quad f^{\pr\pr}_i(x;y-x)<0,\\ \nabla g_j(x)(y-x)\leqq 0\;\textrm{ for all }\; j\in A(x), \\ \nabla g_j(x)(y-x)=0,\; j\in A(x)\quad\textrm{imply}\quad g_j^{\pr\pr}(x;y-x)\le 0, \\ \end{array} \right. \] provided that all necessary derivatives exist. We call {\rm (P)} second-order KT pseudoconvex if it is second-order KT pseudoconvex at each $x\in S$. \end{definition} In the case when $n=1$ this notion reduces to the notion of KT pseudoconvex scalar problem (see Ivanov $[^{11}]$). \begin{theorem}\label{th2} Let the functions $f$ and $g$ be differentiable and second-order differentiable in every critical direction. Suppose that {\rm (P)} is second-order KT pseudoconvex. Then every second-order Kuhn-Tucker stationary point is a global Pareto minimizer. \end{theorem} {\bf Proof.} Suppose that $x\in S$ is a second-order Kuhn-Tucker stationary point, but it is not a global Pareto minimizer. Therefore, there exists $y\in S$ such that $f(y)\le f(x)$. It follows from second-order pseudoconvexity of (P) that \centerline{$\nabla f_i(x)(y-x)\leqq 0$} \noindent for all $i=1,2,\dots, n$ and $\nabla g_j(x)(y-x)\leqq 0$ for all active indices $j$, that is $j\in A(x)$. We obtain from here that the direction $y-x$ is critical at the point $x$. According to second-order Kuhn-Tucker conditions we have $\nabla L(x)=0$, which implies that $\lambda_i\nabla f_i(x)(y-x)=0$ and $\mu_j\nabla g_j(x)(y-x)=0$. We conclude from here that $\nabla f_i(x)(y-x)=0$ and $\nabla g_j(x)(y-x)=0$ for all $i$ and $j$ such that $\lambda_i>0$ and $\mu_j>0$. Then it follows from second-order pseudoconvexity of (P) that $f^{\pr\pr}_i(x;y-x)<0$ and $g_j^{\pr\pr}(x;y-x)\le 0$ for all $i$ and $j\in A(x)$ such that $\lambda_i>0$ and $\mu_j>0$. By $\lambda\ne 0$ we get the inequality \[ L^{\pr\pr}(\bar x,y-x)=\sum_{i=1}^n\lambda_i f_i^{\pr\pr}(\bar x,y-x)+ \sum_{j\in I(\bar x)}\mu_j g_j^{\pr\pr}(\bar x,y-x)< 0, \] \noindent which contradicts the second-order Kuhn-Tucker condition (\ref{10}). \qed \bigskip {\bf 3. Necessary and sufficient conditions for a weak global minimum.} The following result is known and it could be found, for instance, in the book $[^9]$ (see Theorem 9.1.4]). \begin{lemma}\label{lema3} Let $X$ be an open set in $\R^s$, and let $f$ be a real scalar function defined on $X$ which is both differentiable and quasiconvex at the point $x\in X$. Then the following implication holds: \centerline{$y\in X,\,f(y)\leqq f(x)\quad\Longrightarrow\quad\nabla f(x)(y-x)\leqq0.$} \end{lemma} \smallskip Let the scalar function $f: X\to\R$ with an open domain $X\subset\R^s$ be differentiable at the point $x\in X$. Then $f$ is said to be pseudoconvex at $x\in X$ if $y\in X$ and $f(y)<f(x)$ imply $\nabla f(x)(y-x)<0$. If $f$ is differentiable on $X$, then it is called pseudoconvex on $X$ when $f$ is pseudoconvex at each $x\in X$. \smallskip The following definition is due to Ginchev and Ivanov $[^8]$. Consider a function $\f$ with an open domain $X$, which is differentiable at $x\in X$ and second-order directionally differentiable at $x\in X$ in every direction $y-x$ such that $y\in X$, $f(y)<f(x)$, $\nabla f(x)(y-x)=0$. We call $f$ second-order pseudoconvex (for short, 2-pseudoconvex) at $x\in X$ if for all $y\in X$ the following implications hold: \centerline{$f(y)<f(x)\quad\mbox{implies}\quad\nabla f(x)(y-x)\leqq0;$} \centerline{$f(y)<f(x),\;\nabla f(x)(y-x)=0\quad\mbox{imply}\quad f^{\pr\pr}(x,y-x)<0.$} Suppose that $f$ is differentiable on $X$ and second-order directionally differentiable at every $x\in X$ in each direction $y-x$ such that $y\in X$, $f(y)<f(x)$, $\nabla f(x)(y-x)=0$. We call $f$ 2-pseudoconvex on $X$ if it is 2-pseudoconvex at every $x\in X$. It follows from this definition that every differentiable pseudoconvex function is 2-pseudoconvex. The converse does not hold. The following result is a particular case of Theorem 4 in $[^8]$: \begin{lemma}\label{lema2} Every radially lower semicontinuous second-order pseudoconvex function, which is defined on some convex set $X\subset\R^n$, is semistrictly quasiconvex on $X$. \end{lemma} The following result is due to Luc and Schaible $[^{12}]$: \begin{lemma}\label{lema1} Let the functions $f_i$ $(i=1,2,\ldots,n)$ be quasiconvex and semistrictly quasiconvex. Then the point $\bar x$ is a weak Pareto minimizer if and only if there exists an index $k\in\{1,2,\ldots,n\}$ such that $\bar x$ minimizes some component $f_k$ of the vector function $f$ over the constraint set $C_k$, defined by Equation {\rm (\ref{1})}. \end{lemma} \begin{theorem} Let the point $\bar x$ be feasible for the problem {\rm (P)}. Suppose that the vector function $f$ and the functions $g_j$, $j\in I(\bar x)$ are continuously differentiable, the functions $g_j$, $j\notin I(\bar x)$ are continuous at $\bar x$, $f$ and $g$ are second-order directionally differentiable at $\bar x$ in every critical direction $d$, all components of $f$ are radially lower semicontinuous and second-order pseudoconvex, all components of $g$ are quasiconvex. Assume that $X$ is convex and the second-order Mangasarian-Fromovitz CQ holds. Then $\bar x$ is a weak global Pareto minimizer if and only if for every critical direction $d$ there exist nonnegative Lagrange multipliers $\lambda=(\lambda_1,\lambda_2,\ldots ,\lambda_n)$, $\mu=(\mu_1,\mu_2,\ldots,\mu_m)$ with $\lambda\ne 0$ such that Conditions (\ref{9}) and (\ref{10}) are satisfied. \end{theorem} {\bf Proof.} Suppose that there exist $\lambda$ and $\mu$ which satisfy Conditions (\ref{9}) and (\ref{10}). We prove that $\bar x$ is a weak Pareto global minimum. Assume the contrary that is there exists $x\in S$ with $f(x)<f(\bar x)$. We prove that $x-\bar x$ is a critical direction. By $2$-pseudoconvexity $\nabla f_i(\bar x)(x-\bar x)\leqq0$ for all $i=1,2,\dots,n$. Due to quasiconvexity, and $g_j(x)\leqq g_j(\bar x)$, $j\in I(\bar x)$, by Lemma \ref{lema3}, we have \centerline{$\nabla g_j(\bar x)(x-\bar x)\leqq0$} \noindent for all $j\in I(\bar x)$, which implies that $x-\bar x$ is critical. Using the assumptions of the theorem we obtain that there exist nonnegative multipliers $\lambda$ and $\mu$ with $\mu_j g_j(\bar x)=0$, $j=1,...,m$ and $\nabla L(\bar x)(x-\bar x)=0$ such that $L^{\pr\pr}(\bar x,x-\bar x)\geqq 0$. Therefore $\mu_j=0$ when $j\notin I(\bar x)$. Using that $x-\bar x$ is critical we obtain \[ \nabla L(\bar x)(x-\bar x)=\sum_{i=1}^n\lambda_i\nabla f_i(\bar x)(x-\bar x)+ \sum_{i\in I(\bar x)}\mu_i\,\nabla g_i(\bar x)(x-\bar x)\leqq 0. \] \noindent Hence $\lambda_i\nabla f_i(\bar x)(x-\bar x)=0$ for all indexes $i=1,2,\ldots,n$ and $\mu_j\nabla g_j(\bar x)(x-\bar x)=0$ for all $j\in I(\bar x)$. Then $\nabla g_j(\bar x)(x-\bar x)=0$ when $\mu_j>0$. It follows from $2$-pseudoconvexity that $f_i^{\pr\pr}(\bar x,x-\bar x)<0$ for all $i=1,2,\ldots,n$ such that $\lambda_i\ne 0$. It follows from quasiconvexity of $g$ that \centerline{$g_j^{\pr\pr}(\bar x,x-\bar x) = \lim_{t\to+0}\frac{g_j(\bar x+t(x-\bar x))-g_j(\bar x)}{t^2/2}\leqq 0\;\textrm{ for all }\; j\in I(\bar x)\;\textrm{ with }\;\mu_j>0.$} \noindent We conclude from $\lambda\ne 0$ that \[ L^{\pr\pr}(\bar x,x-\bar x)=\sum_{i=1}^n\lambda_i f_i^{\pr\pr}(\bar x,x-\bar x) +\sum_{j\in I(\bar x),\;\mu_j>0}\mu_j\,g_j^{\pr\pr}(\bar x,x-\bar x)<0, \] \noindent which is a contradiction. Conversely, suppose that $\bar x$ is a weak global minimum. We prove that Conditions (\ref{9}) and (\ref{10}) hold. It follows from Lemma \ref{lema2} that all components of $f$ are semistrictly quasiconvex and quasiconvex. Then, it follows from Lemma \ref{lema1} that there exists $k\in\{1,2,\ldots,m\}$ such that $\bar x$ minimizes the function $f_k$ over the set $C_k$. According to Lemma \ref{NKT} conditions (\ref{9}) and (\ref{10}) are satisfied. \qed \bigskip {\bf Acknowledgements.} This research is partially supported by the TU Varna Grant No 18/2012. \ \bigskip\bigskip \newpage \centerline{REFERENCES} \bigskip \noindent $[^1]$ A{\scriptsize GHEZZAF} B., M. H{\scriptsize ACHIMI}. J. Optim. Theory Appl. {\bf 102}, 1999, 37--50. \noindent $[^2]$ G{\scriptsize EORGIEV} P. G., N. P. Z{\scriptsize LATEVA}. Set-Valued Anal. {\bf 4}, 1996, 101--117. \noindent $[^3]$ G{\scriptsize UERRAGGIO} A., D. T. L{\scriptsize UC}. J. Optim. Theory Appl. {\bf 116}, 2003, 117--129. \noindent $[^4]$ G{\scriptsize UTIERREZ} C., B. J{\scriptsize IMENEZ}, V. N{\scriptsize OVO}. Math. Program. Ser. B. {\bf 123}, 2010, 199-223. \noindent $[^5]$ M{\scriptsize AEDA} T. J. Optim. Theory Appl. {\bf 122}, 2004, 521--538. \noindent $[^6]$ G{\scriptsize INCHEV} I., V. I. I{\scriptsize VANOV}. J. Math. Anal. Appl. {\bf 340}, 2008, 646--657. \noindent $[^7]$ M{\scriptsize ANGASARIAN} O.L., S. F{\scriptsize ROMOVITZ.} J. Math. Anal. Appl. {\bf 17}, 1967, 33--47. \noindent $[^8]$ G{\scriptsize INCHEV} I., V. I. I{\scriptsize VANOV}, Lecture Notes in Econom. and Math. Systems, Vol. 583, Springer, Berlin - Heidelberg - New York, 2007, pp. 247--264. \noindent $[^9]$ M{\scriptsize ANGASARIAN} O.L. Nonlinear Programming, New York, McGraw-Hill, 1969. \noindent $[^{10}]$ K{\scriptsize ANNIAPPAN} P. J. Optim. Theory Appl. {\bf 40}, 1983, 167-174. \noindent $[^{11}]$ I{\scriptsize VANOV}. J. Global Optim. {\bf 46}, 2010, 31--47. \noindent $[^{12}]$ L{\scriptsize UC} D.T., S. S{\scriptsize CHAIBLE.} J. Optim. Theory Appl. {\bf 94}, 1997, 147--153. \end{document} 

1. Introduction {#sec1-cells-08-00593} =============== An essential function of the mammalian small intestine is nutrient absorption \[[@B1-cells-08-00593]\]. While the intestinal epithelium is composed of multiple specialized cell types including goblet cells, enteroendocrine cells, Paneth cells, and enterocytes \[[@B2-cells-08-00593]\], only enterocytes are responsible for nutrient absorption from the intestinal lumen. The enterocytes comprise of undifferentiated crypt cells, which proliferate and differentiate to mature villus cells \[[@B3-cells-08-00593]\]. Nutrient, electrolyte and fluid absorption primarily occur through the villus cells while the crypt cells are thought to be primarily secretory. During the differentiation process, enterocytes acquire more transport properties, and are physiologically able to absorb more nutrients compared to undifferentiated crypt cells. Among the acquired transporters, Na-K-ATPase, a basolateral membrane (BLM) transporter, plays a vital role in regulating ionic homeostasis, cell volume and maintaining membrane potential \[[@B4-cells-08-00593]\]. Na-K-ATPase transports three Na^+^ out of the cell in exchange for two K^+^ into the cell, thus maintaining a high level of intracellular K^+^ and low intracellular Na^+^ concentration \[[@B5-cells-08-00593]\]. This establishes a Na^+^ gradient that is responsible for driving other secondary transport processes across the brush border membrane (BBM). Transport of different ions (H^+^, Ca^2+^, Cl^--^, PO~4~^2−^, SO~4~^2−^), nutrients like glucose, amino acids and vitamins, certain nucleic acids, bile acids, and neurotransmitters across the plasma membrane are all dependent on Na-K-ATPase activity \[[@B6-cells-08-00593],[@B7-cells-08-00593]\]. Previous studies have shown that Na-nutrient co-transport processes in the BBM, namely Na-glucose (SGLT1), Na-alanine (ATB0), Na-glutamine (B0AT1), Na-taurocholate (ASBT) and Na-adenosine (DMT1) are dependent on the BLM Na-K-ATPase for their optimal activity. Furthermore, their inhibition in villus cells in a rabbit model of inflammatory bowel disease (IBD) \[[@B8-cells-08-00593],[@B9-cells-08-00593],[@B10-cells-08-00593]\] is at least in part secondary to altered Na extruding capacity of the cell due to diminished Na-K-ATPase activity. Therefore, Na-K-ATPase plays a crucial role in nutrient absorption and maintenance of intestinal health. The Na-K-ATPase enzyme consists of various subunits: alpha (α), beta (β) and gamma (γ) subunits \[[@B5-cells-08-00593]\]. The α and β subunits are ubiquitously present and are important for the activity of Na-K-ATPase \[[@B11-cells-08-00593],[@B12-cells-08-00593]\], while the γ-subunit is an optional subunit and its expression is restricted to certain tissues \[[@B13-cells-08-00593]\]. Of these subunits, α subunit does the catalytic function of the transporter \[[@B12-cells-08-00593]\], whereas other subunits (β and γ) are responsible for the regulatory function of the α subunit \[[@B14-cells-08-00593],[@B15-cells-08-00593]\]. The α subunit is a transmembrane protein having intracellular Na^+^, ATP and phosphate binding sites, and extracellular K^+^ binding sites \[[@B5-cells-08-00593]\]. The β subunit is important for the proper folding of α subunit and its translocation to the plasma membrane \[[@B16-cells-08-00593],[@B17-cells-08-00593]\]. At least four isoforms of α subunit (α1, α2, α3, and α4) and three isoforms of β (β1, β2, and β3) subunit are known. These isoforms are expressed in a tissue-specific manner \[[@B5-cells-08-00593],[@B18-cells-08-00593],[@B19-cells-08-00593]\]. The combination of different isoforms of α and β subunits make up a series of Na-K-ATPase isoenzymes \[[@B20-cells-08-00593]\]. Each isoenzyme has a different functional property, and they are expressed differentially in tissue- and cell specific-manner \[[@B21-cells-08-00593]\]. Of these isoforms, α1 and β1 are ubiquitously present in epithelial cells and are also present in the mucosa of the intestine \[[@B22-cells-08-00593]\]. Na-K-ATPase activity may change according to the physiological requirements of the cell. Several mechanisms may regulate Na-K-ATPase activity. For example, availability of the substrate (Na^+^, K^+^ and ATP), the amount of the enzyme at the plasma membrane, which can be modified by changes in the rate of synthesis or degradation of the individual Na pump polypeptide, and also by the movement of the pump from cytoplasmic vesicles to the plasma membrane by exo/endocytotic vesicular transport \[[@B23-cells-08-00593]\]. Besides these mechanisms, Na-K-ATPase activity at the cell surface is directly regulated by direct phosphorylation and dephosphorylation by protein kinases and protein phosphatases \[[@B24-cells-08-00593],[@B25-cells-08-00593],[@B26-cells-08-00593]\]. Alterations in the phosphorylation levels of α and regulatory γ subunits are known to change the affinity and, therefore, the activity of the transporter. Studies using in vivo models have shown that the Na-K-ATPase functional activity varies depending on the cell type and function: rabbit villus cells that have the primary function of sodium-dependent nutrient absorption have twice the amount of Na-K-ATPase function compared to the crypt cells that have minimal absorptive capacity \[[@B27-cells-08-00593]\]. However, the mechanisms underlying the alteration of Na-K-ATPase function during maturation of crypts to villus cells is not known. Therefore, this study was aimed to determine the mechanism(s) underlying the alteration of Na-K-ATPase activity during growth and maturation of intestinal epithelial cells. 2. Results {#sec2-cells-08-00593} ========== 2.1. Alkaline Phosphatase (ALP) Levels during Cell Maturation {#sec2dot1-cells-08-00593} ------------------------------------------------------------- To determine cell growth and maturation of intestinal epithelial cells (IEC-18), alkaline phosphatase activity was measured from cells collected from 0-day to 4-day post confluence ([Figure 1](#cells-08-00593-f001){ref-type="fig"}). The highest level of alkaline phosphatase activity was observed at 4-days post-confluence (1536 ± 44.16 picomole/min·mL), whereas the lowest level of activity (1092 ± 5.432 picomole/min·mL) was observed at 0-day post-confluence. Thus, increasing alkaline phosphatase activity demonstrates that cells mature as they grow from 0-day post confluence (crypt-like) to 4-day post confluence (villus-like). 2.2. Na-Dependent Glucose Uptake during Cell Maturation {#sec2dot2-cells-08-00593} ------------------------------------------------------- As enterocytes mature from crypt to villus, physiological alterations are accompanied by the appearance of different transporters in the BBM. Specifically, the Na-glucose co-transporter SGLT1 appears as enterocytes mature from crypt to villus. Similar to in vivo observations \[[@B28-cells-08-00593]\], we also found that Na-dependent glucose uptake increased almost three-fold as cells matured from crypt-like to villus-like. [Figure 2](#cells-08-00593-f002){ref-type="fig"} shows that minimal SGLT1 activity (92.53 ± 13.35 picomole/mg protein·min) was seen at 0-day post-confluence. There was a steady and robust increase in SGLT1 activity from 0- to 4-day post-confluence in IEC-18 cells (376 ± 57.71). This phenomenon of increasing SGLT1 activity may be due to increasing cellular maturation, cellular polarity and/or an increase in the number of transporters itself, and is comparable to what is seen in vivo during crypt to villus maturation \[[@B29-cells-08-00593]\]. 2.3. Na-K-ATPase Activity Levels during Cell Maturation {#sec2dot3-cells-08-00593} ------------------------------------------------------- As enterocytes mature, along with the appearance of BBM transporters, BLM Na-K-ATPase activity increases to provide the favorable Na gradient necessary to absorb nutrients. Similar to the in vivo observation in rabbit intestine \[[@B27-cells-08-00593]\], Na-K-ATPase activity, as determined by inorganic phosphate (*P~i~*) release, increased gradually in IEC-18 cellular homogenates from 0--4 days post-confluence ([Figure 3](#cells-08-00593-f003){ref-type="fig"}A: 5.01 ± 0.12 nanomole/mg protein·min in 0-day, 5.74 ± 0.47 in 1-day, 14.48 ± 1.07 in 2-day, 14.58 ± 0.90 in 3-day, 13.72 ± 0.23 in 4-day). Since Na-K-ATPase is functional only as a membrane protein, plasma membrane was prepared from IEC-18 cells and Na-K-ATPase activity was measured. In plasma membrane preparations also, Na-K-ATPase activity increased gradually from 0-day to 4-day ([Figure 3](#cells-08-00593-f003){ref-type="fig"}B: 49.81 ± 02.74 nanomole/mg protein·min in 0-day, 81.36 ± 1.775 in 1-day, 86.53 ± 1.617 in 2-day, 102.8 ± 4.01 in 3-day, 115.6 ± 3.58 in 4-day). Finally, since radiolabeled rubidium (^86^Rb^+^) uptake is more a functional measure of Na-K-ATPase in live cells, it was also carried out. ^86^Rb^+^ uptake also increased gradually from 0-day to 4-day ([Figure 4](#cells-08-00593-f004){ref-type="fig"}: 649.2 ± 6.95 nanomole/mg protein·min in 0-day, 958.6 ± 32.52 in 1-day, 1285 ± 22.28 in 2-day, 1335 ± 39.27 in 3-day, 1565 ± 8,17 in 4-day). Thus, these data indicate that as IEC-18 cells mature from crypt-like at 0-day post-confluence to villus-like at 4-day post-confluence, Na-K-ATPase activity increased comparable to in vivo crypt and villus cells. 2.4. Kinetic Studies of Na-K-ATPase Activity during Cell Maturation {#sec2dot4-cells-08-00593} ------------------------------------------------------------------- To determine the mechanism of the increase in Na-K-ATPase activity from 0--4 days post-confluence, we performed ^86^Rb^+^ kinetics in IEC-18 cells. As the concentration of extracellular Rb^+^ was increased, ^86^Rb^+^ uptake was stimulated and subsequently became saturated in all conditions. Kinetic parameters showed that there was no significant change in *V~max~* among the groups. However, there was a significant difference in *K~m~* from 0--4 days post-confluence. Affinity (1/*K~m~*) of Na-K-ATPase activity increased gradually as the cells matured from 0-4 days post-confluence as shown in [Table 1](#cells-08-00593-t001){ref-type="table"}. 2.5. Na-K-ATPase α1 and Na-K-ATPase β1 Subunit mRNA Abundance during Cell Maturation {#sec2dot5-cells-08-00593} ------------------------------------------------------------------------------------ The α subunit primarily provides Na-K-ATPase functional activity whereas the β subunit does not have pumping activity, but contributes for proper transportation of α subunit to the plasma membrane to make the entire protein fully functional. Therefore, to determine whether the change in Na-K-ATPase activity may be transcriptionally regulated, we performed quantitative real-time polymerase chain reaction (qRT-PCR) analysis. There was no significant difference in the relative expression of Na-K-ATPase α1 mRNA ([Figure 5](#cells-08-00593-f005){ref-type="fig"}A) between different groups (0--4 days). Similarly, the Na-K-ATPase β1 subunit mRNA abundance ([Figure 5](#cells-08-00593-f005){ref-type="fig"}B) was also not statistically different between the groups (0--4 days). 2.6. Na-K-ATPase α1 and Na-K-ATPase β1 Subunit Protein Expression during Cell Maturation {#sec2dot6-cells-08-00593} ---------------------------------------------------------------------------------------- Since mRNA levels of Na-K-ATPase α1 and Na-K-ATPase β1 subunits may not definitely correlate with protein expression, Western blot analysis of protein expression of Na-K-ATPase α1 and Na-K-ATPase β1 subunits were performed. Immunoreactive protein levels of Na-K-ATPase α1 and Na-K-ATPase β1 subunits were determined in cellular homogenates and plasma membrane fractions. Densitometric analysis of relative protein expression revealed that the level of Na-K-ATPase α1 protein expression was not statistically different in IEC-18 cells from 0--4 days post-confluence in cellular homogenates ([Figure 6](#cells-08-00593-f006){ref-type="fig"}A,B) or plasma fractions ([Figure 7](#cells-08-00593-f007){ref-type="fig"}A,B). Similarly, there was also no significant difference in Na-K-ATPase β1 subunit protein expression among IEC-18 cells from 0--4 days post-confluence in cellular homogenates ([Figure 6](#cells-08-00593-f006){ref-type="fig"}A,C) and plasma membrane fractions ([Figure 7](#cells-08-00593-f007){ref-type="fig"}A,C). Protein expression of Na-K-ATPase α1 and Na-K-ATPase β1 subunits correlates with mRNA expression of their respective subunits. Furthermore, these studies are consistent with kinetic parameters determined above. Therefore, the expression of Na-K-ATPase α1 and Na-K-ATPase β1 subunits are unaltered in cellular homogenates and plasma membrane fractions in IEC-18 cells during maturation. Thus, the increase in Na-K-ATPase activity as IEC-18 cells mature from crypt-like to villus-like cells is secondary to increased affinity of the protein, which may be due to altered phosphorylation of the α1 Na-K-ATPase subunit. 2.7. Protein Phosphorylation Levels of Na-K-ATPase α1 Subunit during Cell Maturation {#sec2dot7-cells-08-00593} ------------------------------------------------------------------------------------ To examine whether changes in Na-K-ATPase activity during different days of post confluence may be due to alteration in phosphorylation of Na-K-ATPase α1 subunit, thus affecting its affinity, immunoprecipitation studies were performed. Na-K-ATPase α1 subunit was immunoprecipitated from the plasma membrane on different days of post-confluence and probed with antibodies specific for phosphorylated amino acids (serine, threonine, and tyrosine). As shown in the representative blot and densitometric analysis in [Figure 8](#cells-08-00593-f008){ref-type="fig"}, immunoprecipitated Na-K-ATPase α1 subunit from different days of post confluence had a gradual increase in the levels of phosphorylated tyrosine ([Figure 8](#cells-08-00593-f008){ref-type="fig"}A,B; 100% in 0-day and 170 ± 21.41 in 4-day post-confluence). Similarly, phosphorylated serine levels also increased ([Figure 8](#cells-08-00593-f008){ref-type="fig"}A,C; 100% in 0-day and 179 ± 5.08 in 4-day post confluence). However, phosphorylated threonine levels remained unchanged ([Figure 8](#cells-08-00593-f008){ref-type="fig"}A,D). To determine which specific residues might be the target of phosphorylation during maturation, protein extracted from different days were probed with anti-p-Na-K-ATPase α1-Ser^23^ antibody, the target of which has been previously implicated in alterations in Na-K-ATPase activity. As shown in the representative blot and densitometric analysis in [Figure 9](#cells-08-00593-f009){ref-type="fig"}, p-Na-K-ATPase α1-Ser^23^ had a gradual increase (100% in 0-day and 241.3 ± 5.5 in 4-day post-confluence) in its phosphorylation levels during IEC-18 cell maturation. These data confirmed that change in Na-K-ATPase activity during cell maturation is likely due to changes in phosphorylation of Na-K-ATPase α1 subunit, specifically at serine and tyrosine amino acid residues and possibly involving p-Na-K-ATPase α1-Ser^23^ residue. 3. Discussion {#sec3-cells-08-00593} ============= Villus and crypt cells have been extensively investigated concerning absorption mechanisms of Na-dependent nutrient co-transporters at the level of the co-transporters in the BBM. At the cellular level, all of these Na-dependent co-transport processes are regulated secondary to alterations in BLM Na-K-ATPase since it provides the necessary Na gradient for their secondary active transport processes. During the maturation of enterocytes from crypts to villus, Na-K-ATPase activity doubles to accommodate for the increased BBM Na-dependent nutrient absorption by the villus cells. However, the Na-K-ATPase located in the BLM that drives these nutrient transporters have not been previously studied for their capacity to adapt themselves to support the increasing nutritional absorptive capacity that the intestinal epithelial cells acquire during their growth and maturation. Therefore, the present study was undertaken to understand the mechanism of Na-K-ATPase regulation during the maturation of intestinal epithelial cells from the crypt to the villus. During the maturation of intestinal cells from the crypt to the villus, significant morphological and physiological changes occur. Morphological changes include changes in composition and increases in brush border membrane surface area. Along with these changes, physiologically mature cells acquire more transporters and, therefore, their capacity to absorb nutrients increases. A study of the crypt-villus axis of piglets has shown that there is upregulation of proteins associated with glycolysis/gluconeogenesis, fatty acid metabolism, amino acid metabolism, and citrate cycle as cells mature from crypt to villus \[[@B30-cells-08-00593]\]. Additionally, it has been reported that alkaline phosphatase (ALP), aminopeptidase, sucrose, lactase, and Na-K-ATPase activities increased along the crypt-villus axis in jejunum and ileum of neonatal pigs \[[@B31-cells-08-00593]\]. Similar to these results, our study has shown that there is an increase in ALP and SGLT1 activities as IEC-18 cells mature from crypt-like to villus-like cells in vitro \[[@B30-cells-08-00593],[@B31-cells-08-00593]\]. There are conflicting reports in the literature addressing the levels of Na-K-ATPase activity during the maturation of cells \[[@B32-cells-08-00593],[@B33-cells-08-00593],[@B34-cells-08-00593]\]. For example, in one study a reduction in activity was found to be due to the shedding of Na-K-ATPase into exosomes or due to degradation \[[@B35-cells-08-00593]\]. The present study, however, has shown that there is an increase in Na-K-ATPase activity as IEC-18 cells mature from crypt-like to villus-like cells. The increase or decrease in activity may be the result of various mechanisms: firstly, due to an increase or decrease in the protein levels of the subunits α and β \[[@B36-cells-08-00593],[@B37-cells-08-00593],[@B38-cells-08-00593],[@B39-cells-08-00593]\]. The increased Na-K-ATPase activity could, therefore, be due to increase in the protein levels of these subunits by increased mRNA and protein synthesis. Secondly, increased translocation of alpha and beta subunits from the cytoplasm to the plasma membrane can increase Na-K-ATPase activity \[[@B24-cells-08-00593],[@B25-cells-08-00593],[@B40-cells-08-00593]\]. Among different isoforms of α and β subunits, α1 and β1 isoforms are expressed ubiquitously and are known to perform a housekeeping role in most epithelial cell types including intestinal epithelial cells \[[@B22-cells-08-00593],[@B41-cells-08-00593]\]. During the development of neonates to newborn, Na-K-ATPase enzyme activity in erythrocytes was found to decrease with maturation, correlating with a decrease in α1 subunit \[[@B42-cells-08-00593]\]. Similarly, in conjunction with Na-K-ATPase activity, the α1 isoform also increased during the maturation of cochlea \[[@B43-cells-08-00593]\]. However, in the present study, we did not find any significant changes in mRNA expression of the α1 isoform or its protein levels although there was an increase in Na-K-ATPase activity. For Na-K-ATPase to be fully functional as a transporting unit in the plasma membrane, assembly of β subunit with α subunit is essential. The association of β subunit with α subunit is required for maturation, membrane insertion and transport activity of the alpha subunit \[[@B44-cells-08-00593]\]. Na-K-ATPase α and β heterodimer are selective in their subunit isoform association, where α1 mostly assembles with β1 isoform \[[@B44-cells-08-00593]\]. Although there was elevated Na-K-ATPase activity, the expression levels of α1 and β1 mRNA and protein in cellular homogenate remain unchanged during maturation. Therefore, the level of Na-K-ATPase α1 and β1 mRNA and protein in cellular homogenate confirmed that Na-K-ATPase is regulated post-translationally. Correspondingly, there was also no alteration in Na-K-ATPase α1 and β1 expression in the plasma membrane. Therefore, the unaltered level of Na-K-ATPase α1 and β1 protein expression in both cellular homogenate and plasma membrane also indicate that increased Na-K-ATPase activity during maturation is not likely due to altered membrane trafficking. The increase in Na-K-ATPase activity could be due to change in the affinity of Na-K-ATPase for Na^+^, K^+^ and ATP \[[@B45-cells-08-00593],[@B46-cells-08-00593]\]. In the present study, there was an increase in the affinity for K^+^ of Na-K-ATPase during cell maturation. This change in affinity may be due to alteration in the level of the phosphorylation of the protein. Considerable evidence indicates that the Na-K-ATPase activity is regulated by phosphorylation of Na-K-ATPase α1 subunit on serine, threonine \[[@B47-cells-08-00593],[@B48-cells-08-00593],[@B49-cells-08-00593],[@B50-cells-08-00593]\] and tyrosine residues \[[@B51-cells-08-00593],[@B52-cells-08-00593]\]. It has been reported that there are at least 21 phosphorylation sites in rat Na-K-ATPase α1 \[[@B53-cells-08-00593]\], including 6 tyrosine, 1 threonine, and 14 serine phosphorylation sites. Protein kinases A and C (PKA, PKC) have been reported to phosphorylate α1 subunit specifically on serine and threonine residues \[[@B47-cells-08-00593],[@B48-cells-08-00593],[@B49-cells-08-00593],[@B50-cells-08-00593]\], whereas various tyrosine kinases phosphorylate at tyrosine residues thus regulating Na-K-ATPase activity \[[@B51-cells-08-00593],[@B52-cells-08-00593]\]. Multiple studies have reported that the phosphorylation of the Na-K-ATPase α1 subunit on tyrosine residues causes different responses on Na-K-ATPase activity \[[@B51-cells-08-00593],[@B52-cells-08-00593],[@B54-cells-08-00593]\]. While some reports have demonstrated that increased tyrosine phosphorylation of Na-K-ATPase α1 in rat kidney tubules \[[@B51-cells-08-00593]\] and human skeletal muscle cells \[[@B54-cells-08-00593]\] increase the activity of the Na-K-ATPase, others have reported that increased tyrosine phosphorylation of Na-K-ATPase α1 in human and pig renal cells reduces the activity of Na-K-ATPase \[[@B52-cells-08-00593]\]. This differential regulation of Na-K-ATPase pump activity by tyrosine phosphorylation might be due to phosphorylation at different amino acid residues or due to the differences in both tissue and species. Our study demonstrates that there is an increase in Na-K-ATPase α1 tyrosine phosphorylation in conjunction with its increase in activity as cells mature. The different results suggest that, depending on the tissue studied, and the experimental conditions, phosphorylation of serine and tyrosine leads to either activation or inhibition of Na-K-ATPase \[[@B40-cells-08-00593]\]. However, in the present study, we observed that there was an increase in serine and tyrosine phosphorylation whereas there was no change in threonine phosphorylation of Na-K-ATPase α1, as cells matured. 4. Materials and Methods {#sec4-cells-08-00593} ======================== 4.1. Cell Culture {#sec4dot1-cells-08-00593} ----------------- Rat small intestinal epithelial cells (ATCC CRL-1589, American Type Culture Collection, Manassas, VA, USA), between passages 5 and 20, were used for all of the experiments. Cells were maintained in Dulbecco's modified Eagle's medium (DMEM), supplemented with 10% (*v*/*v*) fetal bovine serum, 100 U/L human insulin, 0.25 mM β-hydroxybutyric acid and 100 U/mL of penicillin and streptomycin. These cells were cultured in a humidified atmosphere of 10% CO~2~ at 37 °C. The cells were fed with fresh medium every other day. When the cells reached 100% confluence, it was considered as 0 day, and cells were grown until 4 days post-confluence. Experiments were conducted on 0-day, 1-day, 2-day, 3-day, and 4-day post-confluent cells. 4.2. Alkaline Phosphatase Activity (ALP) Assay {#sec4dot2-cells-08-00593} ---------------------------------------------- The activities of ALP were determined using an enzyme assay kit, according to the manufacturer's protocol (Alkaline phosphatase assay kit, colorimetric (Abcam\#83369) Abcam PLC, Cambridge, MA, USA). ALP activity was calculated as picomole/min/mL from the cells that were harvested between 0 and 4-day post-confluence. 4.3. Na-Dependent Glucose Uptake Studies {#sec4dot3-cells-08-00593} ---------------------------------------- To determine Na-dependent glucose transport in IEC-18 cells, glucose uptake studies were performed on cells grown on transwell inserts (0.4 μm; \#140620, Thermo Fisher Scientific Inc., Waltham, MA, USA) in 24 well plates. Cells were rinsed once with warm wash buffer (130 mM NaCl, 5 mM KCl, 1mM MgSO~4~, 2 mM CaCl~2~, 20 mM HEPES \[4-(2-Hydroxyethyl)-1-poperazineethanesulfonic acid\] and pH adjusted to 7.4) and incubated with the same buffer for 10 min. After the incubation, wash buffer was removed from the wells and replaced with the reaction mixture containing a trace amount of \[^3^H\] OMG in Na-HEPES buffer with and without phlorizin (SGLT1 inhibitor). The reaction was stopped after two minutes with cold wash buffer (130 mM NaCl, 5mM KCl, 1mM MgSO~4~, 2 mM CaCl~2~, 20 mM HEPES, 10 mM D-Glucose and pH adjusted to 7.4), and the cells were then washed twice with the same buffer. Cells were lysed by incubating them with 1M NaOH for 20 min at 70 °C. The lysed cells from each well was then mixed with 5 mL of Ecoscint A (National Diagnostics, Atlanta, GA, USA). The vials were kept in darkness overnight and radioactivity retained by the cells was determined in a Beckman Coulter 6500 scintillation counter (Beckman Coulter Inc., Brea, CA, USA). 4.4. Crude Plasma Membrane Isolation {#sec4dot4-cells-08-00593} ------------------------------------ Plasma membrane (crude) was prepared from cells according to the method of Havrankova et al. \[[@B55-cells-08-00593]\]. Briefly, 100 mg cells were homogenized in 5 mL of 1 mM NaHCO~3~ and centrifuged at 600× *g* for 30 min. The supernatant was collected and centrifuged for 30 min at 20,000× *g*. After washing membrane pellets with 1mM NaHCO~3~ for a couple of times~,~ final pellet was resuspended in 0.04 M Tri-HCl buffer (pH 7.4) containing 0.1% bovine serum albumin (BSA). All the procedures mentioned above were carried out at 4 °C. 4.5. Na-K-ATPase Activity Assay {#sec4dot5-cells-08-00593} ------------------------------- Na-K-ATPase activity was measured as *P~i~* liberated in cellular homogenates and crude plasma membrane fractions from the cells that were harvested 0-4 days post-confluence according to the protocol of Forbush et al. \[[@B56-cells-08-00593]\]. Enzyme-specific activity was expressed as nanomoles of *P~i~* released per milligram protein per minute. 4.6. Determination of Na-K-ATPase Activity and Kinetics by ^86^Rb^+^ Uptake {#sec4dot6-cells-08-00593} --------------------------------------------------------------------------- Uptake studies were performed in cells grown on transwell inserts in 24 well plates. IEC-18 cells were plated with approximately 1 × 10^5^ cells. Uptake studies for Na-K-ATPase were done using radioactive Rubidium (^86^Rb^+^; PerkinElmer, Inc., Waltham, Massachusetts, U.S.). Cells were incubated for 1 hr in serum-free DMEM media at conditions as mentioned above. The cells were subsequently washed with serum-free media and incubated for 10 min at 37 °C with 20 μM monensin (in serum-free DMEM) added to both the apical and basolateral sides. The cells were then washed with serum-free DMEM. Uptake of ^86^Rb^+^ was started by incubating the cells for 15 min in a reaction mixture (serum-free DMEM) containing \~1 μCi/ of ^86^Rb^+^ added to the basolateral side of the cells in the presence and absence of ouabain (1mM). For Na-K-ATPase kinetics, after incubation with serum-free DMEM for 1hr, the cells were subsequently washed with incubation buffer (130 mM NaCl, 5 mM KCl, 1 mM MgSO~4~, 2 mM CaCl~2~ and 20 mM HEPES) and incubated for 10 min at 37 °C in the incubation buffer with 20 μM monensin on both sides of the membrane. Then cells were washed with wash buffer (130 mM NaCl, 1 mM MgSO~4~, 2 mM CaCl~2~, 20 mM HEPES and 5 mM mannitol) on both sides. Na-K-ATPase kinetic studies were then performed by incubating the cells for 30 s with reaction mixtures (130 mM tetramethylammonium chloride, 1 mM MgSO~4~, 2 mM CaCl~2~, 20 mM HEPES, 5 mM mannitol and ^86^Rb^+^ (\~1 μCi/ well) containing different concentrations (0.05--2 mM) of RbCl added to the basolateral side of the membrane, in the presence and absence of ouabain (1 mM). For both uptake and kinetics experiment, the reaction was then stopped by the addition of ice-cold MgCl~2~. The cells were then washed three times with MgCl~2~ and lysed with 800 μL of 1N NaOH by incubating them for 30 min at 70 °C. The lysed cells from each well were then mixed with 5 mL of Ecoscint A (National Diagnostics, Atlanta, GA, USA). The vials were kept in darkness overnight and radioactivity retained by the cells was determined in a Beckman Coulter 6500 scintillation counter. The net Na-K-ATPase activity was calculated by subtracting ^86^Rb^+^ uptake in the presence of ouabain from that in the absence of ouabain. 4.7. RNA Isolation and Quantitative Real-Time Polymerase Chain Reaction (qRT-PCR) {#sec4dot7-cells-08-00593} --------------------------------------------------------------------------------- RNA was isolated from 0--4 days post-confluent cells by using the RNeasy mini kit obtained from Qiagen Inc. (Germantown, MD, USA). qRT-PCR was performed using isolated total RNA by a two-step method. First, total RNA was used to synthesize cDNA using SuperScript III (Invitrogen Life Technologies, Carlsbad, CA, USA). Then, newly synthesized cDNA was used as a template to perform real-time PCR using TaqMan universal PCR master mix from Applied Biosystems (Foster City, CA, USA) according to the manufacturer's protocol. Rat-specific Na-K-ATPase α1 and Na-K-ATPase β1 primers and probes were used for the qRT-PCR studies. Also, the rat-specific β-actin primer was used to normalize the expression levels of mRNA in the samples. 4.8. Immunoprecipitation {#sec4dot8-cells-08-00593} ------------------------ Immunoprecipitation (IP) assays were conducted to investigate the phosphorylation level of Na-K-ATPase α1 present in the plasma membrane. Cells were first rinsed with PBS, scraped and collected in PBS, and pelleted by brief low-speed centrifugation (800 rpm). Then collected pellet was further used for plasma membrane isolation as mentioned above. The prepared plasma membrane fraction was sonicated and dissolved in 200 uL of lysis buffer (T-per Buffer; Thermo Fisher Scientific Inc., Waltham, MA, USA) with protease inhibitors (Thermo Fisher Scientific Inc., Waltham, MA, USA). Plasma membrane protein was further diluted to 1 μg/μL in a microcentrifuge tube with T-per buffer and placed on a rocking shaker at 4 °C with 2 μg of Anti-Na-K-ATPase α1 antibody overnight. Antibody-protein complexes formed were captured by adding 50 μL of washed protein G agarose bead slurry (Catalog \# 16-266; Thermo Fisher Scientific Inc., Waltham, MA, USA) with rotation at 4 °C for 2 h. The agarose beads with the protein--antibody complex were collected by centrifugation (14,000× *g*) and washed three times with ice-cold lysis buffer. The agarose beads were resuspended in 2× sample loading buffer and incubated at 37 °C for 15 min. The beads in sample loading buffer were vortexed, and the supernatant was used for Western blot analysis. 4.9. Western Blot Analysis {#sec4dot9-cells-08-00593} -------------------------- Western blot analysis was performed in whole cell lysate and plasma membrane preparations. Equal amounts of protein (20 μg) were denatured in sample buffer (Laemmli sample buffer, BIO-RAD \#3100010639; Bio-Rad Laboratories, Hercules, CA, USA) and separated by electrophoresis on an 8%--12% gradient gel. Proteins on the gel were transferred to a polyvinylidene fluoride membrane which was blocked with 5% dry milk or BSA in TBS (20 mM Tris pH 7.5, 150 mM NaCl) with 0.1% Tween-20 and then incubated with one of the following primary antibodies overnight at 4 °C: Na-K-ATPase α1 (Millipore, \# 05-369); Na-K-ATPase β1 (ab2873; Abcam PLC, Cambridge, MA, USA); p-Serine (ab9332; Abcam PLC, Cambridge, MA, USA); p-Threonine (ab179530; Abcam PLC, Cambridge, MA, US); p-Tyrosine (ab9337; Abcam PLC, Cambridge, MA, US); p-Na-K-ATPase α1-Ser^23^ (Cell Signaling Technology, \#4006) and p-Na-K-ATPase α1-Ser^16^ (Cell Signaling Technology, \#4020). The membrane was washed three times with Tris Buffered Saline with Tween 20 (TBST) followed by TBS. The membrane was then incubated with secondary antibody for 1hr. ECL western blotting detection reagent (GE Healthcare Bio-Sciences, Piscataway, NJ, USA) was used to detect the immobilized protein. The chemiluminescence was detected using a FluorChem M instrument (Alpha Innotech, San Leandro, CA, USA) and its software analyzed the intensity of the bands. β-actin was used to normalize the expression levels in cellular homogenate. 4.10. Protein Determination {#sec4dot10-cells-08-00593} --------------------------- Total protein was measured by the Lowry method using the Bio-Rad protein assay kit (Hercules, CA, USA). BSA was used as a standard. 4.11. Statistical Analysis {#sec4dot11-cells-08-00593} -------------------------- All groups presented have at least *n* = 3 per group repeated with a different passage. The values are presented as mean ± standard error of the mean (SEM), and significant values of *p* \< 0.05 or less were taken to indicate statistical significance. All of the data were analyzed using one-way or two-way analysis of variance (ANOVA) using GraphPad Prism 7 (GraphPad Software Inc., San Diego, CA, USA). 5. Conclusions {#sec5-cells-08-00593} ============== In summary, this study demonstrated that as epithelial cells mature from secretory crypt-like to absorptive villus-like cells, the Na-K-ATPase activity increases to support the increasing absorptive capacity of the enterocytes. Na-K-ATPase activity increased secondary to increases in affinity rather than an increase in transporter numbers as cells matured from crypt to villus. Specifically, altered affinity was found to be likely due to increased phosphorylation of α1 subunit of Na-K-ATPase. The authors thank Usha Murughiyan for editorial assistance in the preparation of this manuscript. Conceptualization, U.S.; Funding acquisition, U.S.; Investigation, N.N.; Methodology, N.N., S.A., Writing, review and editing, N.N., S.A., and U.S. This work was supported by National Institutes of Health grants DK-67420, DK-108054, P20GM121299-01A1 and Veteran's Administration Merit Review grant BX003443-01 to U. Sundaram. The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results. {#cells-08-00593-f001} ![Increase in Na-dependent glucose uptake as IEC-18 cells matured. Uptake was performed in the presence and absence of phlorizin (1 mM) in reaction medium containing \[^3^H\]-OMG tracer. Values are represented as means ± SEM, *n* = 6 independent experiments. Values not sharing common superscripted letters are significantly different at *p* \< 0.001.](cells-08-00593-g002){#cells-08-00593-f002} {#cells-08-00593-f003} {#cells-08-00593-f004} {#cells-08-00593-f005} {#cells-08-00593-f006} {#cells-08-00593-f007} {#cells-08-00593-f008} {#cells-08-00593-f009} cells-08-00593-t001_Table 1 ###### Kinetic parameters of Na-K-ATPase transporters in IEC-18 cells. ------------------------------------------------------------------------------------------------------- 0 Day 1 Day 2 Day 3 Day 4 Day ---------------------------- -------------- -------------- -------------- -------------- -------------- *V~ma~*~x~\ 742.5 ± 62 714 ± 83.5 718 ± 58.4 713 ± 27.9 733 ± 44.8 (picomole/mg protein·30 S) *K~m~*\ 0.68 ± 0.11a 0.61 ± 0.15a 0.39 ± 0.08b 0.37 ± 0.04b 0.33 ± 0.06b (mM) ------------------------------------------------------------------------------------------------------- Uptake of ^86^Rb^+^ as a function of varying concentration of extracellular Rb^+^ (RbCl). As extracellular Rb^+^ concentration was increased, uptake of ^86^Rb^+^ was stimulated and subsequently became saturated in all groups. Kinetic parameters \[maximal uptake (*V~max~*) and affinity (1/*K~m~*)\] were obtained after the analysis of data with GraphPad Prism 7 software (GraphPad Software Inc., San Diego, CA, USA). Values are represented as mean ± SEM, *n* = 6. Values not sharing common superscripted letters are significantly different at *p* \< 0.01.

1. Introduction =============== With the improvement in living standards, health check-ups are receiving increasing attention. Chest computed tomography (CT) is superior to conventional chest radiography for detecting lesions and determining their position and qualitative aspects, making it the prime choice for lesion detection. However, the radiation associated with the process is becoming a problem. The ICRP advocates the principle of "as low as reasonably achievable (ALARA)," which requires CT examination be on the premise of image quality to meet the demand of diagnosis as far as possible to reduce the radiation dose.^\[[@R1]\]^ Previously, the chest low-dose scanning techniques mainly used a fixed pipe and voltage, low current, and real-time dynamic exposure dose to regulate the tube current, but methods to further reduce the radiation dose are being studied. The CARE kV technique combined with iterative reconstruction is a new technique with lower radiation dose that the conventional techniques, which has attracted the attention of medical radiation workers. In this paper, we discuss the application of the CARE kV technique combined with iterative reconstruction to chest CT scanning for health examination patients. 2. Materials and methods ======================== 2.1. General information ------------------------ A total of 130 patients suspected with chest disease undergoing a chest CT examination from Dec, 2012 to April, 2013 in our hospital were randomized into two groups: the experimental group and control group. The former consisted of 60 patients; 31 were males and 29 were females. The average age was 49.36 + 18.41 years and the average body mass index (BMI; BMI= weight/height^2^ (kg/m^2^)) was 23.1 ± 3.43. The latter consisted of 70 patients; 34 were males and 36 were females. The average age was 50.03 ± 17.09 years and the BMI was 22.56 ± 3.38. There were no significant differences in gender and age in the two groups (*P* \> 0.05) (Table [1](#T1){ref-type="table"}).The clinical trial was approved by the Institute\'s Medical Ethics Committee of Kunming Medical University, and consent forms were signed by all the examination patients. ###### Patients characteristics.  2.2. CT protocols ----------------- Siemens Somatom Definition Flash CT system (SOMATONM Definition Flash, Siemens Healthcare, Forchheim, Germany) was used to perform a scan from the thoracic inlet to the end of the lung. Scanning parameters for the experimental group: The reference kV and mAs were 100 kV and 80 mAs, respectively. Real-time dynamic exposure dose adjustment tube current was used (CARE Dose 4D) pitching 3.0. The tube rotation time was 0.28 s. The collimation was 64 × 0.6 mm. The layer thickness and spacing were 8 mm. CARE kV scanning technology was first used followed by SAFIRE iterative reconstruction of the experimental raw data (value = 3). Convolution kernel: Window of lung fields I70f very sharp ASA, mediastinal window I31f medium smooth. Scanning parameters for the control group: The reference kV and mAs were 120 kV and 80 mAs, respectively. The CARE kV scanning technology and SAFIRE iterative reconstruction were not used. The rest of the scanning parameters are the same as for the experimental group. 2.3. Imaging analysis --------------------- "Viewing" software on the Siemens CT postprocessing workstation, the 8 mm thick layer of cross-sectional images of each patient were loaded into the software. The region of interest (ROI) in the mediastinum window was defined in three levels: the apex level on the right side of the pectoralis major, tracheal bifurcation level of the descending aorta, and left atrium level. Artifacts, calcification, and other regions of interference were excluded. The measurement area was ∼60 mm^2^; hence, each ROI area was set at the maximum and location was equal. Different levels for the determination of the average CT value and background noise (SD) were measured. The SNR and CNR were calculated as follows: SNR = mean CT value/SD and CNR = (descending aorta average CT value -- muscle average CT value)/SD.^\[[@R2]\]^ 2.4. Image quality analysis --------------------------- Two attending physicians analyzed the obtained images independently, following the double-blind principle, and experienced doctors were called to reach an agreement in case of a disagreement. The main observable elements were the mediastinum, great vessels, pleura, chest wall tissue, and lung segments. The definition of the bronchial segment revealed anatomical details, particle size, and artifacts. The subjective image-quality scores were as follows:^\[[@R3]\]^ 5 points: Anatomical details and lesions are simple and clearly defined; 4 points: Anatomical structure and details and lesions are less clear and can be defined, but not very well; 3 points: Most of the anatomic structures and lesions can satisfy diagnosis requirements, but a few images cannot be defined; 2 points: Anatomic structures and lesions are unclear and details cannot be obtained; and 1 point: Anatomical structures and lesions are too fuzzy to diagnose. A score of 3 points or higher meets the diagnostic requirements. The anatomical details used in the mediastinal and pulmonary window analysis included the lesion form display, size, and edge. 2.5. Radiation dose ------------------- The effective current, effective voltage, CT dose index of the volume (CTDIvol), and dose length product (DLP) for each patient were automatically generated during scanning and recorded. The CTDIvol reflected the average dose of the whole scanning volume. The DLP was used to evaluate the total radiation dose for a patient undergoing complete CT scanning. The effective dose (ED) was calculated using ED = DLP × *K*, where *K* is the conversion factor, using the European standard guidelines of CT quality. The chest average value was 0.014.^\[[@R4]\]^ 2.6. Displayed lesion results ----------------------------- \[Normal, nodule, effusion and (or) consolidation, fiber cords, calcification, cavity, pleural effusion and lymph node\] of the two groups were used for statistical analysis. 2.7. Statistical analysis ------------------------- An SPSS 17.0 statistical package was used. The image quality, CTDIvol, DLP, mean CT value of the image, SD, and SNR of the two groups were compared with two independent sample tests. For *P* \< 0.05, there were statistically significant differences. The *Kappa* test was used to measure the consistency of the grading results of different doctors; ≥0.75 showed a good consistency; 0.4--0.76 showed an average consistency, and \<0.4 showed a poor consistency. 3. Results ========== 3.1. Comparison of mean CT value -------------------------------- The mean CT values at the apex level of the right pectoralis major, tracheal bifurcation level of the descending aorta, and left central atrial of the experimental group were compared with those of the control group (Table [2](#T2){ref-type="table"}). No statistical difference was found (*P *\> 0.05). ###### Comparison of mean CT^∗^ values of experimental group and control group.  3.2. Comparison of SD --------------------- The SDs of the apex level of the right pectoralis major, tracheal bifurcation level of the descending aorta, and left central atrial of the control group are higher than those of the experimental group (Table [3](#T3){ref-type="table"}). The difference in the SD was found to have statistical significance (*P *\> 0.05). ###### Comparison of SD^∗^ of experimental group and control group.  3.3. Comparison of SNR and CNR ------------------------------ The SNRs of the apex level of the right pectoralis major, tracheal bifurcation level of the descending aorta, and left central atrial of the experimental group are higher than those in the control group (Table [4](#T4){ref-type="table"}). The difference in the SNR is of statistical significance (*P*\<0.05). The difference in the CNR of the tracheal bifurcation level of the descending aorta of the two groups has no statistical significance (*P *\> 0.05). ###### Comparison of SNR^∗^ and CNR^∗^ of experimental group and control group.  3.4. Comparison of subjective scoring of image quality ------------------------------------------------------ The image-quality scores of the mediastinal window in the experimental group and control group were 4.75 ± 0.07 and 4.62 ± 0.09 (*t* = 1.12, *P* = 0.248), respectively. The scores for the pulmonary window were 4.87 ± 0.08 and 4.75 ± 0.11 *(t* = 2.02, *P* = 0.056), respectively. There was no statistically significant difference in the two groups and all the image-quality scores were greater than 4.5 points, which meets the requirement of diagnosis according to the image-quality score given by two attending physician (Kappa = 0.786) (Fig. [1](#F1){ref-type="fig"}A and B). {#F1} 3.5. Lesions in the display picture ----------------------------------- The difference in the constituent ratio of lesions in the experimental group and control group has no statistical significance (Table [5](#T5){ref-type="table"}). ###### Comparison of lesion of experimental group and control group.  3.6. Comparison of radiation dose --------------------------------- In the experimental group, 48 patients were subject to 80-kV tube voltage scanning and 12 were subject to 100-kV tube voltage scanning (Table [6](#T6){ref-type="table"}). In the control group, a tube voltage of 120 kV was used for all patients. Compared with that of the control group, the CTDIvol of the experimental group was 2.01 mGy (37.92%) less; the DLP was ∼52.70 mGy cm (31.44%) less; and the ED was ∼0.74 mSv (31.49%) less. All the differences are statistically significant. ###### Comparison of radiation dose.  4. Discussion ============= Although the radiation dose received by a patient in one CT scan is negligible, a cancer risk may exist for frequent scanning or large dosage.^\[[@R5]\]^ In 1990, Naidich et al^\[[@R6]\]^ first proposed the concept of low-dose CT, that is, reduce the radiation dose and radiation hazard, but still obtain the same image quality. Currently, dose-reduction includes the use of real-time dynamic exposure dose adjustment tube current, fixed low tube voltage, fix low tube current, and big pitch heart switch control technology, which lower the radiation dose in different levels.^\[[@R7]--[@R9]\]^ CARE kV technology and SAFIRE iterative reconstruction---new methods to reduce radiation dose---are breaking new ground for radiation dose control. CARE kV technology can automatically lower the radiation dose by setting conditions corresponding to the areas and purpose of different settings, which in turn can be used to set the standard quality reference voltage according to the image quality. CARE kV technology can automatically calculate the reference values and variation curve for different tube voltages according to the automatically defined body type, calculate the CTDIvol, compare, vary the tube voltage from low to high, and chose the lowest voltage scanning if CT ball pipe system hardware is allowed.^\[[@R10]\]^ To guarantee and improve the image quality, a constant or increased CNR as the standard in CARE kV technology must be used to lower the tube voltage and increase the tube current simultaneously. Reduction of the tube voltage will result in an increased photoelectric effect of x-rays on patients, which will increase the absorption by organs with different densities, especially tissues with a high atomic number will increase the contrast of blood vessels and the CT values of image, significantly reduce the radiation dose, and improve the image quality.^\[[@R11]\]^ Winklehne et al^\[[@R10]\]^ scanned the head and chest of which body mass index (BMI) in the same level using CARE kV scanning technology. Their results showed that the contrast of the image increased by 21%, the CNR increased by 17.3% without the increase of image noise of background. Meanwhile, the radiation dose reduced by 58% and 25.3%. Although lowering the tube voltage can significantly reduce the radiation dose, a low tube voltage reduces x-ray penetration; hence, fewer photons reach the detector, which results in increased image noise. In order to reduce the image artifacts caused by noise, which can affect the image density resolution, this study combined CARE kV scanning technology with the SAFIRE iterative reconstruction technique. SAFIRE iterative reconstruction technology uses a support noise model of the original data, which can be applied into the image noise and inhibit and eliminate it in each iteration. The image is a combination of iterative reconstructions; hence, the technique can improve the image quality and reduce the image noise with the potential of reducing the radiation dose.^\[[@R12],[@R13]\]^ This study used CARE kV scanning technology for the chest scanning, but SAFIRE iterative reconstruction techniques were applied to the experimental group. The SD of the experimental groups is lower than that of the control group and the SNR is higher than that of the control group, indicating that the SAFIRE iterative reconstruction algorithm can reduce the noise level and improve the SNR simultaneously. Furthermore, the CT value and image quality score of the lesions of the experimental group and control group showed no statistical differences. All the image scores were greater than 4.5 points, which meets the diagnosis requirements. The reason for this is that the lung is a gas-filled organ, which results in a tissue natural contrast, and structural differences and overlapping within intrathoracic tissues lead to density distribution, making it a good anatomical basis for CARE kV chest scanning technology. The differences in the CTDIvol, DLP, and ED for the experimental group and control group all have statistical significance. The radiation dose with the CTDIvol of the experimental group was ∼2.01 mGy (38%) less than that of the control group. In the experimental group, all 60 patients were subjected to 100-kV scanning, while the control group did not use CARE kV technology and all 70 patients were subjected to 120-kV scanning, further proving that CARE kV scanning technology can significantly reduce the radiation dose. In conclusion, in chest CT scanning of health examination patients, CARE kV scanning technology alone will cause increased noise and reduced SNR, and scanning dose reduction without improving the algorithm may lead to diagnostic information loss and enhance increase the risk of incorrect diagnosis. However, SAFIRE iterative reconstruction alone can result in higher quality images with the potential of reducing the radiation dose, even though the reduction it limited. Hence, for chest CT scanning, CARE kV technology combined with SAFIRE iterative reconstruction technology can significantly reduce the radiation dose and improve the image quality, which has potential clinical application. However, this study has some limitations. For example, no measurement and analysis of the detailed morphology of lesions and CT value were performed. In addition, the results were not classified according to the features of the lesions. Finally, group studies on the relationship between the BMI and voltage were not included, which needs to be further investigated. Abbreviations: ALARA = as low as reasonably achievable, BMI = body mass index, CNR = contrast-to-noise ratio, CT = computed tomography, CTDIvol = CT dose index of the volume, DLP = dose length product, ED = effective dose, ICRP = International Commission on Radiological Protection, SAFIRE = sinogram-affirmed iterative reconstruction, SNR = signal-to-noise ratio. BY and Z-LL contributed equally to this work. Sources of funding: Medical Imaging Department, the First Affiliated Hospital, Kunming Medical University, key subjects for medical imaging The authors have no conflicts of interest to disclose.

Introduction {#s1} ============ Myotonia is a phenomenon of skeletal muscle hyperexcitability that impairs muscle relaxation following contraction or percussion ([@B1]--[@B4]). Myotonia can be clinically evident with visually appreciable delays in muscle relaxation, but can also be subclinical with myotonic discharges noted on electromyography (EMG) in the absence of overt clinical signs. Etiologies of myotonia include monogenic inherited causes as well as acquired causes such as hypothyroidism, denervation, inflammatory myopathies, or toxic myopathies ([@B5]). Inherited myotonic disorders are generally divided into two major categories, dystrophic and non-dystrophic, with the major differentiating factor being the presence of progressive muscle degeneration in patients with dystrophic forms of myotonia ([@B1], [@B3], [@B6], [@B7]). Dystrophic myotonic disorders include myotonic dystrophy type I (DM1), characterized by prominent distal limb weakness and multisystem disease, and myotonic dystrophy type II (DM2, also known as proximal myopathic myotonia or PROMM), characterized by proximal limb weakness with less prevalent multisystem disease ([@B3]). DM1 is caused by a pathogenic CTG expansion in *DMPK* (\>50 repeats) which results in altered splicing of *CLCN1*, as well as other genes, leading to reduced chloride conductance and myotonia ([@B8]--[@B10]). Intergenerational expansion of repeat size may be observed, with an increased repeat size correlating with earlier age of onset and increased severity of symptoms, a phenomenon known as anticipation ([@B9]--[@B11]). The most severe form of DM1 is seen in individuals with congenital symptoms, typically caused by a repeat size of \>1000 ([@B10], [@B11]). DM2 is caused by a pathogenic CCTG expansion in *CNBP* and also affects splicing of *CLCN1*. Genetic anticipation and congenital onset of disease are not observed in DM2 ([@B9]--[@B13]). Typically, DM2 is milder and has a later age of onset than DM1. Additionally, repeat length is not correlated with severity of symptoms or age of onset ([@B9]--[@B12]). Pain is a more prominent feature in many individuals with DM2 and may be the presenting symptom, often leading to misdiagnoses ([@B13]). Patients with DM1 and DM2 also experience systemic complications including progressive cardiac conduction defects, respiratory insufficiency, premature cataracts, daytime hypersomnolence, gastrointestinal and endocrine dysfunction, cognitive and behavioral deficits, and increased malignancy rates ([@B9]--[@B11]). The non-dystrophic myotonias are caused by mutations of specific skeletal muscle ion channels and are usually categorized on the basis of the ion channel affected, inheritance pattern, and clinical features. The two skeletal muscle ion channels that are associated with non-dystrophic disorders include chloride (*CLCN1)* and sodium (*SCN4A)* channels. The non-dystrophic myotonias are known to be highly variable in expression, leading to missed or delayed diagnosis in many cases ([@B1], [@B14]). Dominant and recessive pathogenic variants in *CLCN1* cause myotonia congenita (MC), the most common inherited muscle channelopathy ([@B6], [@B7]). Variants lead to loss-of-function and dominant negative effects in the CLC-1 channel causing reduced chloride conductance and membrane hyperexcitability ([@B2]--[@B4]). Both forms are typically characterized by childhood onset muscle stiffness ([@B1], [@B3]). Stiffness can be triggered by emotional surprise and aggravated by cold temperatures and pregnancy. Repetitive motion or "warm up" can alleviate these symptoms ([@B1], [@B2], [@B6], [@B7], [@B14]). Recessively inherited MC often causes more severe symptoms and may cause a slowly progressive muscle weakness that may be identified on clinical examination, as well as transient weakness ([@B1], [@B3], [@B14]). Paramyotonia congenita (PMC), sodium channel myotonia (SCM), and hyperkalemic periodic paralysis (HyperPP) are all caused by dominant gain-of-function variants in *SCN4A* ([@B15]). These variants lead to an excessive inward sodium ion current which causes muscle hyperexcitability, leading to myotonia, or transient loss of excitability, causing periodic paralysis ([@B16]). Symptoms of PMC typically begin in the first decade of life and include myotonia that is induced by cold or repeated muscle activity (paradoxical myotonia or paramyotonia), as well as episodic weakness triggered by exercise, cold, potassium ingestion, or fasting ([@B6], [@B7], [@B12], [@B15], [@B16]). On clinical examination patients with variants in *SCN4A* often display eyelid myotonia that worsens with repetitive eyelid closure ([@B1]). SCM is characterized by pure, often painful, myotonia without episodic weakness and typically not triggered by cold temperatures ([@B1], [@B6], [@B7], [@B15]). HyperPP is characterized by recurrent episodes of weakness triggered by exercise, potassium ingestion, or emotional stress that can last hours to days. Over time, a slowly progressive permanent weakness may occur in HyperPP ([@B16]). Hypokalemic periodic paralysis (HypoPP) is also caused by mutations in *SCN4A* and is usually not associated with myotonia, although rarely is has been described in patients with homozygous loss-of-function variants ([@B17]). Treatment of myotonic disorders is generally influenced by symptom severity and ability to control symptoms through avoidance of triggers ([@B6]). Pharmacological treatment may not be necessary in all patients, but patients who accept medical treatment may experience significant improvement in myotonia symptoms, including pain ([@B5]). There are currently no FDA approved medications for the treatment of myotonia; however, mexiletine has been approved in the EU as an antimyotonia agent. A multinational study has found that only 40% of patients received treatment for this symptom ([@B14]). Off-label antimyotonia treatments include anti-arrhythmic, anti-epileptic, and anti-depressant medications, which have shown clinical benefit, usually in small case series or single case reports. Mexiletine has demonstrated efficacy in multiple controlled studies in non-dystrophic myotonia and in DM1 ([@B18], [@B19]). In addition, a recent n-of-1 aggregate study also showed efficacy in a study cohort of 27 patients with non-dystrophic myotonia ([@B20]). Similarly, lamotrigine has been demonstrated to be effective in non-dystrophic myotonia ([@B21]). There is a lack of published data detailing usage and efficacy of antimyotonia agents in clinical practice. In this study, we retrospectively reviewed patient-reported symptoms and clinical data of patients with genetically defined myotonic disorders seen at a large tertiary center. We aimed to characterize the phenotypic profiles of each disorder by comparing symptom profiles between the disorders as well as the usage of commonly prescribed antimyotonia agents. Materials and Methods {#s2} ===================== A retrospective chart review was performed of patients with inherited myotonic disorders seen at The Ohio State University Wexner Medical Center from March 2009 to December 2018. This study and waiver of consent was approved by the Ohio State University institutional review board. Subjects -------- Initial patient search was conducted to identify all patients with myotonia-related diagnostic codes. A complete list of codes utilized are found in [Table 1](#T1){ref-type="table"}. Patients were included in the review if they or a family member had a documented variant in *CLCN1* or *SCN4A* or a pathogenic expansion in *DMPK* or *CNBP*. Patients were excluded if they had absence of myotonia both clinically and electrically or if they had a pathogenic or likely pathogenic variant identified in a second gene related to neuromuscular disease. ###### List of diagnostic codes utilized in the search for candidates meeting inclusion criteria. **ICD-9/ICD-10** **Diagnosis** ---------------------------------------- --------------------------------------------------------------------------------- 359.39/G71.19 Myotonia fluctuans 359.21/G71.11 Myotonia atrophica 359.23/G71.13 Myotonia chondrodystrophica 359.22/G71.12 Myotonia congenita 728.85, 319, 756.50/M62.89, F79, Q78.9 Myotonia with intellectual disability and skeletal anomaly 359.24, E980.5/G71.14 Myotonia, drug-induced 359.3/G72.3 Periodic myotonia 794.17/R94.131 Myotonic changes present on EMG V83.89/Z14.8 Carrier of myotonic dystrophy 271/E74.02 Pompe disease 796.4/R89.0 Low acid maltase in muscle determined by biopsy 792.9/R89.0 Low acid maltase levels in fibroblasts 359.0, V84.89/G71.2, Z15.89 Autosomal dominant centronuclear myopathy associated with mutation in DMN2 gene 359.89, 359.0/G72.89, G71.2 Myofibrillar myopathy Chart Review ------------ Data were collected on patient demographics, patient-reported symptoms (including stiffness, weakness, pain, cramping, and exacerbating factors), family history, and medication history. Physical examination data were documented, including presence of clinical myotonia and weakness, creatine kinase levels, and EMG data. Presence of periodic paralysis and paradoxical myotonia were not ascertained. Genetic testing results, including genetic tests completed and complete variant data, including pathogenic classification (pathogenic, likely pathogenic, or uncertain), were ascertained via review of laboratory reports. For patient-reported symptoms and clinical examination data, each symptom was recorded as present only if the chart note specifically stated that the patient had that symptom. To avoid ascertainment bias, if a symptom was not recorded as present or absent in the chart note, it was listed as "unknown" and data on that symptom in that patient was not utilized in the statistical analysis. Exacerbating factors were recorded as present if documented in the patient\'s chart, and recorded as absent if it was not documented or if specifically noted as not being present. Muscle weakness was considered present on clinical examination if a patient scored a "4" or less in any muscle group at any clinical visit during manual muscle testing by a neuromuscular specialist. If multiple creatine kinase levels were available, the highest value was recorded. Reference ranges used were 30-220 U/L for men and 30-184 U/L for women. Values recorded from electromyography (EMG) studies included number of muscles tested and number of muscles with abnormal spontaneous activity, including myotonia, positive sharp waves, and fibrillation potentials. If multiple EMG studies were performed, the number of muscles tested and the number of muscles with abnormal spontaneous activity were summed across all studies for each patient in each of these categories. Usage of the following medications was recorded: Acetazolamide, Clomipramine, Diazapam, Dichlorphenamide, Dispyramide, Imipramine, Lamotrigine, Mexiletine, Nifedipine, Phenytoin, Procainamide, Quinine, Ranolazine, Taurine, Thiazides, and Tocainide. Statistical Analysis -------------------- Descriptive statistics were utilized to create the phenotypic profile for each genotype. Mean values were utilized for participant age and age of symptom onset. For the remaining chart review data, percentages were utilized to depict the incidence of a given phenotypic characteristic by genotype. Any characteristic that was not explicitly recorded as present or absent was not utilized in the statistical analysis for that characteristic. Comparison of symptoms across genotype groups and dystrophic versus non-dystrophic groups were made utilizing a Fisher\'s exact test. Descriptive statistics were utilized to determine the proportion of each genotype group who had trialed and were currently taking an antimyotonia agent. Patients currently taking a medication at the time of review were divided by genotype and type of medication utilized. Results {#s3} ======= Demographics ------------ A total of 142 patients were included in this study: 27 had one or more variants in *CLCN1*, 15 had a variant in *SCN4A*, 89 had an expansion in *DMPK*, and 11 had an expansion in *CNBP*. The average age of symptom onset for individuals with *CLCN1, SCN4A, DMPK*, and *CNBP* variants was 16.5 years \[standard deviation (*SD*) 12.1\], 23.6 years (*SD* 21.0), 28.0 years (*SD* 15.9), and 40.7 (*SD* 12.0), respectively. [Figure 1](#F1){ref-type="fig"} depicts symptom onset by age category (neonatal being defined as \<1 year of age, childhood as 1--18 years old, and adulthood being greater than 18 years old) and genotype. Most participants were Caucasian (83.8%) and female (81.5% *CLCN1*, 73.3% *SCN4A*, 60.7% *DMPK*, 45.5% *CNBP*). A total of 13 individuals included in the study were deceased at the time of chart review. Eleven of these individuals had *DMPK* expansions, with the average age of death among these individuals being 54.3 years (*SD* 10.4). The other two deaths both occurred at the age of 68, one in an individual with an expansion in *CNBP* and the other in an individual with a variant in *CLCN1*. {#F1} Genetic Variants ---------------- Of the 27 individuals with at least one variant in *CLCN1*, 23 individuals had one variant identified and 4 had two variants with apparent autosomal recessive inheritance of disease. A total of 13 different variants were identified in the *CLCN1* cohort with eight being classified as pathogenic, two as likely pathogenic, and three as variants of uncertain significance (VUS). No individuals with an *SCN4A* variant were identified to have more than one variant in this gene on their genetic lab report. In the *SCN4A* cohort, 9 different variants were identified with eight being classified as pathogenic and one as a VUS. A full list of genotypes identified and corresponding myotonia phenotypes are summarized in [Tables 2](#T2){ref-type="table"}--[4](#T4){ref-type="table"}. Additional details on the genetic testing performed for patients with unclear results (including those with single variants previously associated with recessive disease, those with variants of uncertain significance and those possessing variants with conflicting interpretations) are summarized in [Supplementary Table 1](#SM1){ref-type="supplementary-material"}. Expansion length in *DMPK* ranged from 74 repeats to 2,450 repeats. Among individuals in this cohort, 5.6% (*n* = 5) were identified as having \<100 repeats, 67.4% (*n* = 60) were identified as having between 100 and 1000 repeats, and 27% (*n* = 24) were identified as having greater than 1,000 repeats in at least part of the sample (in the case of mosaicism). Due to the retrospective nature of this study, we were unable to ascertain the subtype of DM1 for each patient; however, only two patients had onset of symptoms in infancy, while the remaining had childhood or adult onset symptoms. ###### List of patients with one variant identified in *CLCN1*. **c**. **p**. **Lab reported classification** **ClinVar classification** **Apparent inheritance pattern** **Clinical myotonia** **Spontaneous activity on EMG** ---------------- ------------------- --------------------------------- ---------------------------- ---------------------------------- ----------------------- --------------------------------- c.469delC p.Leu157Phefs\*13 Pathogenic Pathogenic Negative family Hx No Yes (8/11) c.501C\>G p.Phe167Leu VUS Conflicting Negative family Hx Yes Yes (4/10) c.592C\>G p.Leu198Val VUS Conflicting AD Yes Yes (3/5) c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD Yes Yes (3/3) c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD Yes Unknown c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD Unknown Yes (3/9) c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD Yes Unknown c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD No Unknown c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD Yes Yes (3/3) c.689G\>A p.Gly230Glu Pathogenic Pathogenic AD Yes Unknown c.689G\>A p.Gly230Glu Pathogenic Pathogenic Unknown Unknown Unknown c.689G\>A p.Gly230Glu Pathogenic Pathogenic Unclear Yes Unknown c.929C\>T p.Thr310Met Pathogenic Pathogenic AD No Unknown c.929C\>T p.Thr310Met Pathogenic Pathogenic AD Unknown Unknown c.937G\>A p.Ala313Thr Pathogenic Pathogenic AD Unknown Yes (8/17) c.937G\>A p.Ala313Thr Pathogenic Pathogenic AD Yes Yes c.1167-10 T\>C Intronic Likely pathogenic Likely pathogenic Unclear Yes Yes (8/8) c.1444G\>C p.Gly482Arg Likely pathogenic Likely pathogenic AD No Yes (10/10) c.1655A\>G p.Gln552Arg Pathogenic Conflicting (LP/P) AD Yes Yes c.2680C\>T p.Arg894Ter Pathogenic Conflicting (LP/P) Unknown No Unknown c.2680C\>T p.Arg894Ter Pathogenic Conflicting (LP/P) Unknown Yes Unknown c.2848G\>A p.Glu950Lys VUS N/a AD Yes Yes (3/5) *For each patient, the classification of the variant as listed on the original laboratory report as well as the classification in the ClinVar database are included. Apparent inheritance is based on family history information identified in the chart. Patients with family members with a diagnosis of myotonia congenita (with or without genetic confirmation), EMG positive for myotonia or reported symptoms of myotonia or stiffness were categorized as autosomal dominant inheritance ("AD"). Patients with family history information available which was negative for any of these symptoms were categorized as "negative family hx." Patients with no family history information listed in the chart were categorized as "unknown" and patients with family history of symptoms of unclear diagnostic significance (pain, cramping or weakness) were categorized as "unclear." Presence of clinical myotonia is listed as positive, negative or unknown for each patient. Spontaneous activity on EMG is listed as positive, negative or unknown. In parentheses, the number of muscles affected with spontaneous activity divided by the total muscles tested on EMG for that individual is recorded (if known)*. ###### List of patients with two variants identified in *CLCN1*. **c**. **p**. **Lab reported classification** **ClinVar classification** **Apparent inheritance pattern** **Clinical myotonia** **Spontaneous activity on EMG** ------------ ------------- --------------------------------- ---------------------------- ---------------------------------- ----------------------- --------------------------------- c.689G\>A p.Gly230Glu Pathogenic; Pathogenic; AD Yes Unknown c.1444G\>C p.Gly482Arg Likely pathogenic Likely pathogenic c.568G\>A p.Gly190Arg VUS; Conflicting; Negative family Hx Unknown Yes c.1238T\>G p.Phe413Cys Pathogenic Pathogenic c.979G\>A p.Val327Ile Pathogenic; Pathogenic; Negative family Hx Yes Unknown c.1262G\>T p.Arg421Leu VUS Likely pathogenic c.1238T\>G p.Phe413Cys Pathogenic; Pathogenic; Negative family Hx Yes Yes (3/5) c.2680C\>T p.Arg894Ter Pathogenic Conflicting (LP/P) c.409T\>G p.Tyr137Asp Likely pathogenic; Likely pathogenic; Negative family Hx Yes Yes (18/18) c.1238T\>G p.Phe413Cys Pathogenic Pathogenic *For each patient, the classification of the variant as listed on the original laboratory report as well as the classification in the ClinVar database are included. Apparent inheritance is based on family history information identified in the chart. Patients with family members with a diagnosis of myotonia congenita (with or without genetic confirmation), EMG positive for myotonia or reported symptoms of myotonia or stiffness were categorized as autosomal dominant inheritance ("AD"). Patients with family history information available which was negative for any of these symptoms were categorized as "negative family hx." Patients with no family history information listed in the chart were categorized as "unknown" and patients with family history of symptoms of unclear diagnostic significance (pain, cramping or weakness) were categorized as "unclear." Presence of clinical myotonia is listed as positive, negative or unknown for each patient. Spontaneous activity on EMG is listed as positive, negative or unknown. In parentheses, the number of muscles affected with spontaneous activity divided by the total muscles tested on EMG for that individual is recorded (if known)*. ###### List of patients with variants in *SCN4A*. **c**. **p**. **Lab reported classification** **ClinVar classification** **Apparent inheritance pattern** **Clinical myotonia** **Spontaneous activity on EMG** ------------ -------------- --------------------------------- ---------------------------- ---------------------------------- ----------------------- --------------------------------- c.1333G\>A p.Val445Met Pathogenic Pathogenic AD Yes Yes (6/6) c.1333G\>A p.Val445Met Pathogenic Pathogenic AD Yes Yes (2/2) c.1333G\>A p.Val445Met Pathogenic Pathogenic AD No Yes (3/3) c.2078T\>C p.Ile693Thr Pathogenic Pathogenic AD Yes Unknown c.3917G\>C p.Gly1306Ala Pathogenic Pathogenic Unclear No Unknown c.3917G\>C p.Gly1306Ala Pathogenic Pathogenic Unknown Yes Unknown c.3917G\>C p.Gly1306Ala Pathogenic Pathogenic AD Yes Unknown c.3938C\>T p.Thr1313Met Pathogenic Pathogenic Negative Family Hx Yes Yes (7/7) c.4343G\>A p.Arg1448His Pathogenic Pathogenic AD Yes Yes (5/5) c.4343G\>A p.Arg1448His Pathogenic Pathogenic Unknown No Unknown c.4343G\>A p.Arg1448His Pathogenic Pathogenic AD Yes Yes (2/2) c.4372G\>T p.Val1458Phe Likely Pathogenic VUS Negative Family Hx Unknown Yes (16/22) c.4386C\>G p.Ile1462Met Pathogenic VUS Unknown Unknown Yes c.4765G\>A p.Val1589Met Pathogenic Pathogenic AD Yes Yes (2/2) c.5126A\>G p.Asn1709Ser VUS VUS Unclear Unknown Yes (2/4) *For each patient, the classification of the variant as listed on the original laboratory report as well as the classification in the ClinVar database are included. Apparent inheritance is based on family history information identified in the chart. Patients with family members with a diagnosis of paramyotonia congenita or sodium channel myotonia (with or without genetic confirmation), EMG positive for myotonia or reported symptoms of myotonia or stiffness were categorized as autosomal dominant inheritance ("AD"). Patients with family history information available which was negative for any of these symptoms were categorized as "negative family hx." Patients with no family history information listed in the chart were categorized as "unknown" and patients with family history of symptoms of unclear diagnostic significance (pain, cramping or weakness) were categorized as "unclear." Presence of clinical myotonia is listed as positive, negative or unknown for each patient. Spontaneous activity on EMG is listed as positive, negative or unknown. In parentheses, the number of muscles affected with spontaneous activity divided by the total muscles tested on EMG for that individual is recorded (if known)*. Patient Reported Symptoms ------------------------- Stiffness was reported by all individuals in our *CLCN1* cohort and was reported in \~80% of the remaining three groups. Weakness was reported in a similar proportion of individuals with non-dystrophic myotonias, 65.2% (*n* = 15) of individuals with *CLCN1* variants, and 69.2% (*n* = 9) *SCN4A* variants. Weakness was more often reported in individuals with dystrophic myotonia than non-dystrophic myotonia, with 90.8% (*n* = 79) of individuals with *DMPK* expansions and 90% (*n* = 9) with *CNBP* expansions reporting this as a symptom. History of pain was reported by all individuals with *CNBP* expansions (*n* = 11), while it was found in 50--70% of the remaining three groups. Presence of muscle cramping was similar across all groups with 50% (*n* = 6) of individuals with *SCN4A* variants and 30--40% of the remaining three groups reporting this. Cold as an exacerbating factor for symptoms was most commonly reported in individuals with non-dystrophic myotonias, with 40.7% (*n* = 11) of individuals with *CLCN1* variants and 46.7% (*n* = 7) of individuals with *SCN4A* variants reporting this compared to 36.4% (*n* = 4) of individuals with *CNBP* expansions and in 7.9% (*n* = 7) of individuals with *DMPK* expansions. A summary of patient reported symptom data is in [Table 5](#T5){ref-type="table"}. ###### Summary of patient reported symptoms by genotype. **Symptom** **Non-dystrophic** **Dystrophic** **Total** ***p*-value** ------------------- --------- -------------------- ---------------- ------------ --------------- ------------- ---------- Stiffness Yes 24 (100.0%) 11 (78.6%) 62 (78.5%) 8 (80.0%) 105 (82.7%) 0.0371 Unknown 3 1 10 1 15 Weakness Yes 15 (65.2%) 9 (69.2%) 79 (90.8%) 9 (90.0%) 112 (84.2%) 0.0072 Unknown 4 2 2 1 9 Pain Yes 18 (69.2%) 8 (53.3%) 50 (59.5%) 10 (100.0%) 86 (63.7%) 0.0417 Unknown 1 0 5 1 7 Cramping Yes 8 (34.8%) 6 (50.0%) 30 (37.5%) 3 (30.0%) 47 (37.6%) 0.8021 Unknown 4 3 9 1 17 Cold exacerbation Yes 11 (40.7%) 7 (46.7%) 7 (7.9%) 4 (36.4%) 29 (20.4%) \<0.0001 *P-values were obtained using Fisher\'s exact test*. Clinical Examination -------------------- Most individuals in the overall cohort had at least one form of clinical myotonia with 69--94% of each group having this present on physical examination. Hand grip myotonia was least common in individuals with expansions in *CNBP* at 16.7% (*n* = 1) while both hand grip and percussion myotonia were most commonly identified in individuals with *DMPK* expansions at 85.1% (*n* = 63) and 94.6% (*n* = 70), respectively. Muscle weakness was most common in the individuals with *DMPK* expansions at 91.8% (*n* = 78) followed by individuals with *CNBP* expansions at 54.5% (*n* = 6). However, 16% (*n* = 4) of individuals with *CLCN1* variants and 26.7% (*n* = 4) of individuals with *SCN4A* variants were also found to have weakness in at least one muscle group. Distribution of weakness on MMT was most commonly identified as proximal only in *SCN4A* patients, was evenly split between proximal only and both proximal and distal in *CLCN1* and *CNBP* patients and was most commonly identified as both proximal and distal in patients with expansions in *DMPK*. A summary of clinical examination data is in [Table 6](#T6){ref-type="table"}. ###### Summary of clinical examination findings by genotype. **Symptom** **Non-dystrophic** **Dystrophic** **Total** ***p*-value** -------------------------- --------- -------------------- ---------------- ------------ --------------- ------------- ---------- Clinical myotonia Yes 17 (73.9%) 9 (69.2%) 75 (93.8%) 7 (77.8%) 108 (86.4%) 0.0067 Unknown 4 2 9 2 17 Hand grip myotonia Yes 14 (70.0%) 9 (75.0%) 63 (85.1%) 1 (16.7%) 87 (77.7%) 0.0022 Unknown 7 3 15 5 30 Percussion myotonia Yes 12 (70.6%) 4 (57.1%) 70 (94.6%) 5 (62.5%) 91 (85.8%) 0.0006 Unknown 10 8 15 3 36 Muscle weakness (on MMT) Yes 4 (16.0%) 4 (26.7%) 78 (91.8%) 6 (54.5%) 92 (67.6%) \<0.0001 Unknown 2 0 4 0 6 *P-values were obtained using Fisher\'s exact test*. EMG Results ----------- In this cohort, individuals with *DMPK* expansions were least likely to have had an EMG performed at 43.8% of the group compared to 70--100% of the remaining three groups. The proportion of muscles with spontaneous activity was similar among the four groups ranging from 71 to 88%. Proportion of muscles with spontaneous activity by genotype is depicted in [Figure 2](#F2){ref-type="fig"}. {#F2} Creatine Kinase Levels ---------------------- CK levels were available for 81 participants. Values were similar for the non-dystrophic groups, with the average being 157.8 U/L (*SD* 153.6) for individuals with *CLCN1* variants and 168.6 U/L (*SD* 137.8) for individuals with *SCN4A* variants. Values were higher for the individuals with *DMPK* and *CNBP* expansions with the averages being 243.4 and 345.6 U/L, respectively. The proportion of individuals with an abnormal CK level was 23.1% (*n* = 3) of individuals with *CLCN1* variants, 37.5% (*n* = 3) of individuals with *SCN4A* variants, 54% (*n* = 27) of individuals with *DMPK* expansions, and 40% (*n* = 4) of individuals with *CNBP* expansions. Medication Usage ---------------- Patients with *DMPK* expansions were the least likely to have trialed an antimyotonia medication at 27% (*n* = 24). Medications had been trialed in 85.2% of individuals with *CLCN1* variants (*n* = 23), 93.3% of individuals with *SCN4A* variants (*n* = 14), and 63.6% of individuals with *CNBP* expansions (*n* = 7). Genotype specific medication trialing data is summarized in [Figure 3](#F3){ref-type="fig"}. Of those who trialed medications, 56.5% (*n* = 13) of *CLCN1* patients, 66.7% (*n* = 10) of *SCN4A* patients, and 8.3% (*n* = 2) of *DMPK* patients had tried more than one medication for myotonia. No patients with *CNBP* expansion had trialed multiple medications. The most commonly trialed medication across all four groups was mexiletine at 40.4% of individuals who trialed any medication (*n* = 42). Currently utilized medications by genotype are depicted by [Figure 4](#F4){ref-type="fig"}. More individuals with non-dystrophic myotonias were currently taking at least one medication for myotonia with 51.9% (*n* =14) of individuals with *CLCN1* variants and 80.0% (*n* = 12) of individuals with *SCN4A* variants compared to 13.5% (*n* = 12) of individuals with *DMPK* expansions, and 18.2% (*n* = 2) of individuals with *CNBP* expansions. Out of the patients trialing at least one medication, none with *SCN4A* variants had discontinued all medication use for myotonia. Patients with dystrophic myotonias more commonly had discontinued all antimyotonia agents with 50.0% (*n* = 12) patients with *DMPK* expansions and 71.4% (*n* = 5) patients with *CNBP* expansions having done so compared to 17.4% (*n* = 4) of individuals with *CLCN1* variants. Reason for discontinuation was given for 30 (60%) of the medications stopped. The frequency that each medication was stopped due to cost, drug interactions, efficacy, or side effects is presented recorded in [Table 7](#T7){ref-type="table"}. The most common reasons for discontinuation overall were lack of efficacy (32%, *n* = 16) and side effects (20%, *n* = 10). For mexiletine, reason for discontinuation was known in nine of cases with 21.2% (*n* = 4) being due to lack of efficacy and 26.3% (*n* = 5) due to side effects. {#F3} {#F4} ###### Summary of reasons patients discontinued commonly utilized antimyotonia agents. **Reason** **Procainamide** **Phenytoin** **Quinine** **Mexiletine** **Acetazolamide** **Ranolazine** **Lamotrigine** **Total** ------------------- ------------------ --------------- ------------- ---------------- ------------------- ---------------- ----------------- ------------ Cost -- -- 1 (33.3%) -- -- 2 (16.7%) -- 3 (6.0%) Drug interactions -- -- -- -- -- 1 (8.3%) -- 1 (2.0%) Efficacy 1 (100.0%) 3 (42.9%) -- 4 (21.1%) 4 (80.0%) 4 (33.4%) -- 16 (32.0%) Side effects -- 1 (14.3%) -- 5 (26.3%) 1 (20.0%) 2 (16.7%) 1 (50.0%) 10 (20.0%) \*No patients discontinued medications due to allergic reaction. *\*\*Not all fields will add to 100% as not all individuals discontinuing a medication had a reason in their chart*. Dystrophic (D) vs. Non-dystrophic (ND) -------------------------------------- Many clinical features differentiated the *D* and *ND* cohorts. The average age of onset of symptoms was significantly younger in the *ND* cohort compared to the *D* cohort at 18.6 years (*SD* 15.1) and 29.9 years (*SD* 16.0), respectively (*p* = 0.0037). Cold exacerbation of symptoms was less commonly reported in the *D* cohort (11 vs. 42.9%, *p* \< 0.0001) and muscle weakness was more commonly reported (90.7 vs. 66.7%, *p* = 0.0022). On examination, clinical myotonia (92.1 vs. 72.2%, *p* = 0.0073), percussion myotonia (91.5 vs. 66.7%, *p* = 0.0051), and muscle weakness (87.5 vs. 20%, *p* \< 0.0001) were more common in the *D* cohort. Average CK levels among the two cohorts (161.9 U/L in *ND* vs. 261.1 U/L in *D, p* = 0.0708) and proportion of individuals with an abnormal CK level (28.6% *ND* vs. 51.7% *D, p* = 0.0797) appeared to differ but were not statistically significant. Patient reported symptoms and clinical examination findings between these groups are summarized in [Figure 5](#F5){ref-type="fig"}. From a treatment standpoint, ND patients were significantly more likely to have trialed (88.1% *ND* vs. 31% *D, p* \< 0.001) and to be currently taking an antimyotonia agent than were the individuals with dystrophic myotonia (61.9% *ND* vs. 14% *D, p* \< 0.0001). Furthermore, *ND* patients were more likely to remain on a medication after trialing than *D* patients (70.3% *ND* vs. 45.2% *D, p* \< 0.0001). {#F5} CLCN1 vs. SCN4A --------------- The only significant difference identified between these two cohorts was a greater proportion of individuals reporting stiffness in the *CLCN1* cohort. In our cohort, 100% (*n* = 24) of individuals with *CLCN1* variants reported stiffness compared to 78.6% (*n* = 11) of individuals with *SCN4A* variants (*p* = 0.0431). Surprisingly, weakness (65.2% *CLCN1* vs. 69.2% *SCN4A, p* = 1.0000) and pain (69.2% *CLCN1* vs. 53.3% *SCN4A, p* = 0.3357) were not significantly different between these cohorts. Patient reported symptoms and clinical examination findings between these groups are summarized in [Figure 6](#F6){ref-type="fig"}. {#F6} Discussion {#s4} ========== In this retrospective study, we performed a comprehensive review of medical record data from a large group of patients with genetically confirmed dystrophic and non-dystrophic myotonic disorders. Our goals included an improved understanding of the phenotypic presentation of hereditary myotonic disorders, as well as characterization of medication use in affected persons. Utilizing this data, we are able to summarize symptom profiles and compare phenotypic features between genotypes. Additionally, we reviewed antimyotonia treatment usage, which has been understudied for this group of disorders. Muscle stiffness is the primary clinical symptom attributed to myotonia ([@B3]). Interestingly, the prevalence of stiffness was the only feature that differed significantly among patients with non-dystrophic myotonias, with more individuals with *CLCN1* variants reporting this symptom than individuals with *SCN4A* variants. In contrast, Trivedi et al., found that 100% of both *CLCN1* and *SCN4A* patients reported stiffness ([@B23]). Although eyelid myotonia has been previously identified as a hallmark of *SCN4A*-related myotonia, we were unable to characterize its prevalence in this cohort, as it was not commonly commented on in the charts reviewed ([@B1], [@B6], [@B7]). We found that the majority of patients with non-dystrophic myotonic disorders reported weakness and that there was no significant difference in the occurrence of these symptoms with respect to genotype. The high prevalence of reported weakness is particularly notable given that the majority (85%) of our *CLCN1* cohort had a single variant identified, and dominant CLCN1-related myotonia has not been classically associated with weakness. Patient-reported weakness in non-dystrophic myotonia has also been identified in two prospective studies; however, the proportion of individuals manifesting weakness differed by genotype. Trip et al. found that patient-reported weakness was almost twice as common in individuals with *CLCN1* variants (75%) compared to *SCN4A* variants (36.7%) while Trivedi et al. found episodic weakness to be approximately twice as common in individuals with *SCN4A* variants (76.5%) compared to individuals with *CLCN1* variants (37.5%) ([@B22], [@B23]). These disparate findings may reflect the different proportion of individuals with dominant versus recessive inheritance in the *CLCN1* cohorts. The majority of the Trip et al. cohort had recessive *CLCN1* variants (previously reported to be associated with a higher incidence of muscle weakness), while the cohort in the Trivedi et al. study had approximately equivalent proportions of dominant and recessive variants ([@B22], [@B23]). Alternatively, these variable findings could be due to differences in symptom ascertainment. Trip et al. documented the presence of "muscle weakness," while Trivedi et al. documented "episodic weakness." In future studies, it may be helpful to ascertain patient-reported weakness in several different ways to better characterize the spectrum of weakness and its functional impact for affected persons. MMT data revealed that 20% of our non-dystrophic cohort had weakness that was identified via clinical examination. In a *SCN4A* cohort reported by Matthews et al., four out of seventeen (23.5%) had weakness, none with strength of \<4/5 ([@B24]). Similarly, we identified clinical weakness in 26.7% of our *SCN4A* cohort. Data correlating weakness identified on MMT with *CLCN1* variants has not been published, but within our cohort was present in 16% (*n* = 4). The presence of muscle weakness on MMT is considered uncommon in individuals with non-dystrophic myotonia, despite the fact that a large proportion of these patients report weakness as a symptom ([@B5], [@B6], [@B22], [@B23]). Additionally, weakness that is episodic or associated with a specific trigger may be missed during standard strength testing, which could account for some of the discrepancy between patient report and clinical examination. Pain was reported in a high proportion of all cohorts, affecting 63.7% overall. Furthermore, we found that similar proportions of patients with *CLCN1* variants and *SCN4A* variants experience pain, at 69.2 and 53.3%, respectively. Although painful myotonia is an accepted clinical feature of *SCN4A*-related myotonia, *CLCN1-*related myotonia was originally described as being painless ([@B1]). Prior studies have reported pain in 57--82% of individuals with *SCN4A* variants and in 28--53% of individuals with *CLCN1* variants ([@B22], [@B23]). We found that pain was reported in all DM2 patients and 59.5% of patients with DM1. Presence of pain is often considered more common in individuals with DM2, which our data supports ([@B25], [@B26]). Identification of pain and weakness in a high proportion of patients, particularly in those with disease not historically associated with these symptoms, is important in understanding phenotype, disease burden, and potential treatment opportunities. Another aim of our study was to understand the treatment approaches in patients with inherited myotonic disorders. Currently, there are no medications that are approved for the treatment of myotonia, but a number of medications are used off label. In our study cohort, 47.9% of patients trialed at least one medication for myotonia and 28.2% were taking an antimyotonia medication at the time of chart review. Trivedi et al. similarly found that 60.6% of their cohort (*CLCN1, SCN4A, CNBP)* was currently taking an antimyotonia agent, but did not specify medication usage by genotype; and DM1 patients were not studied ([@B19]). In our cohort, 83% of all patients reported stiffness and 64% reported pain. This suggests that some symptomatic individuals are not being treated, and lack of medication utilization cannot be explained by lack of symptoms. Individuals with non-dystrophic myotonia were significantly more likely to have trialed and remained on an antimyotonia agent. Of those who tried an antimyotonia medication, non-dystrophic patients were also more likely to have trialed two or more medications than the dystrophic patients. Given that clinical myotonia was more common in the dystrophic cohort, and that patient-reported myotonia was similar in the dystrophic and non-dystrophic cohorts, the different rates of medication usage are evidently not due to differences in symptom prevalence. Although it is possible that the non-dystrophic cohort experienced less symptoms due higher medication use, charts were reviewed at multiple time points, including the initial patient visit, to reduce the effect of medication on these data points. Other possible explanations for this discrepancy include: non-dystrophic patients may have better therapeutic response to currently available medications, non-dystrophic patients may be more compliant in taking medications or due to the lack of systemic symptoms in non-dystrophic patients, the focus for therapy may be on their myotonia symptoms rather than symptoms in other body systems. None the less, this data suggests that there could be a gap between patients who may benefit from use of antimyotonia agents and those who are actually treated. Ideally, treatable symptoms should be ascertained and medication offered for patients who may benefit. Further study of medical treatment of myotonia, including genotype-specific treatment, dose, duration, side effects, reasons for discontinuation, and non-compliance is necessary to optimize symptom management for patients affected with these disorders. This study has limitations associated with retrospective chart review and single center bias. These include limitations in type of data that is charted and available to review, lack of consistency in the type and depth of information recorded in the medical record, variability in number of appointments and time between appointments for patients included and the possibility that patients may be not reporting all symptoms to a physician or falsely reporting presence or absence of symptoms and/or treatment response. Consistency of type of data recorded, and format in which it was recorded, was controlled by utilization of a standardized database which aided in consistency during the data entry process. Inherited myotonic disorders present diagnostic and treatment challenges in the clinical setting. Different types of myotonic disorders may be difficult to distinguish clinically, emphasizing the importance of comprehensive genetic testing. In patients suspected of having a non-dystrophic myotonia, the most expeditious, and cost-effective approach is panel testing including *CLCN1* and *SCN4A* sequencing. We found that the majority of patients with myotonic disorders had symptoms of pain, weakness, and stiffness. This includes individuals with non-dystrophic myotonia, a group where pain and weakness have not always been considered common symptoms. Despite this, only about half of the patients in this study were treated with antimyotonia agents. Advances in pharmacologic treatments for myotonia are needed, as there are no medications specifically approved for the indication of myotonia. Future studies should be designed to investigate relationships between specific genetic variants, clinical phenotypes, and symptom profiles, as well as response to different potential antimyotonia treatments. Data Availability Statement {#s5} =========================== All datasets generated for this study are included in the article/[Supplementary Material](#SM1){ref-type="supplementary-material"}. Ethics Statement {#s6} ================ This project was approved by the Ohio State Biomedical Sciences IRB. Author Contributions {#s7} ==================== AM performed chart reviews and data entry. RS and AM created the database. RS and DK performed statistical analysis. AM, JR, SL, and WA wrote the manuscript. All authors were involved in the design of this study. Conflict of Interest {#s8} ==================== WA received funding from Gilead Sciences, the Neuroscience Research Institute at The Ohio State University, and has served as a paid consultant to Genentech and La Hoffmann Roche. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. **Funding.** This study was funded in part by The Ohio State University Center for Clinical and Translational Sciences (CCTS) and The Ohio State Neuroscience Research Institute. Supplementary Material {#s9} ====================== The Supplementary Material for this article can be found online at: <https://www.frontiersin.org/articles/10.3389/fneur.2020.00593/full#supplementary-material> ###### Click here for additional data file. [^1]: Edited by: Emma Matthews, University College London, United Kingdom [^2]: Reviewed by: Karen Joan Suetterlin, University College London, United Kingdom; Lorenzo Maggi, Fondazione IRCCS Istituto Neurologico Carlo Besta, Milan, Italy; Masanori P. Takahashi, Osaka University, Japan [^3]: This article was submitted to Neuromuscular Diseases, a section of the journal Frontiers in Neurology

\section{Introduction} Topological data analysis (TDA) \citep{eh,carlsson} is, as the name suggests, the application of topology to data analysis. Persistent homology \citep{elz,zc} is one of the most important tools for TDA. In persistent homology, by encoding information on length scales in filtrations, we can capture characteristic geometric features with multiple length scales. By using filtrations, persistent homology is also robust to noise. Homology itself is translation and rotation invariant, and so persistent homology is similarly invariant. These properties are suitable for the analysis of shapes of data, and persistent homology is applied in various practical data analysis contexts in domains such as biology~\citep{virus}, image processing~\citep{Hu2019}, and materials science~\citep{Hiraoka28062016,granular,PhysRevE.95.012504,Kimura2018}. To describe our problem, we first define persistent homology. Persistent homology is defined on a filtration, an increasing sequence, of topological spaces. We consider the following filtration: \begin{align*} \mathbb{X} & = (X_t)_{t \in T}, \\ X_t & \subset X_s \mbox{ if $t \leq s$}, \\ X_0 &= \emptyset, \end{align*} where $T$ is $\{0, 1, \cdots, N\}$ or $\mathbb{R}_+$. The $q$th \emph{persistent homology} $H_q(\mathbb{X}; \Bbbk)$ with a coefficient ring $\Bbbk$ is defined as follows: \begin{align*} H_q(\mathbb{X}; \Bbbk) &\mbox{ is a pair of } \\ & \{H_q(X_t; \Bbbk)\}_{t \in T}, \\ & \{\phi_s^t:H_q(X_s;\Bbbk) \to H_q(X_t; \Bbbk) \}_{s \leq t}, \end{align*} where $\phi_s^t:H_q(X_s;\Bbbk) \to H_q(X_t; \Bbbk)$ is the homology map induced by the inclusion map $X_s \hookrightarrow X_t$. In standard homology theory, we use $\mathbb{Z}$ as a coefficient ring since the universal coefficient theorem ensures that $\mathbb{Z}$-homology provides the most information about homology. However, in the theory of persistent homology a field is used instead of $\mathbb{Z}$ since the interval decomposition described below is crucial for analysis of persistent homology and the decomposition is guaranteed \emph{only when $\Bbbk$ is a field}. Indeed, the structural theory of persistent homology ensures the existence and uniqueness of the following decomposition of $H_q(\mathbb{X}; \Bbbk)$ called the \emph{interval decomposition} if $\Bbbk$ is a field: \begin{align*} H_q(\mathbb{X}; \Bbbk) &= \bigoplus_{i=1}^L I(b_m, d_m), \\ I(b, d) &= \{\varphi_s^t: U_s \to U_t\}_{s \leq t}, \\ U_s &= \left\{ \begin{array}{ll} \Bbbk & \mbox{ if $b \leq s < d$}, \\ 0 & \mbox{ otherwise}, \end{array}\right. \\ \varphi_s^t &= \left\{ \begin{array}{ll} \mathrm{id} & \mbox{ if $b \leq s \leq t < d$}, \\ 0 & \mbox{ otherwise}, \end{array}\right. \end{align*} where $0 < b_m < d_m \leq \infty$. This $I(b, d)$ is called an \emph{interval indecomposable}. This theorem depends on the fact that $\Bbbk[z]$, the polynomial ring with a field coefficient, is PID~\citep{zc}, and so this theorem does not hold for $\Bbbk = \mathbb{Z}$. When this interval decomposition is given, we define the $q$th \emph{persistence diagram} (PD) $D_q(\mathbb{X}; \Bbbk)$ as a multiset of pairs of endpoints of the intervals. That is, $D_q(\mathbb{X}; \Bbbk) = \{(b_m, d_m)\}_{i=1}^L$. Each pair is called a \emph{birth-death pair}. Each $b_m and d_m$ are \emph{birth time} and \emph{death time}, respectively, and $d_m - b_m$ is called a \emph{lifetime}. Since a birth-death pair with a long lifetime corresponds to a ``stable'' homological structure in the filtration, we can use lifetimes to compare the significance of birth-death pairs. Normally, we choose $\Bbbk$ as one of $\mathbb{R}, \mathbb{Q}$, and $\mathbb{Z}_p$ for a prime $p$. $\mathbb{Z}_2$ is most often used since it is amenable to a fast algorithm and an intuitive interpretation. Here we face the problem of the choice of $\Bbbk$. If any $\Bbbk$ gives the same PD, there is no problem. However, this is not practical, because the dimensions of homology vector spaces for the same topological space are different when the $\mathbb{Z}$-homology group of the space has non-zero torsion. If a Klein bottle appears in a filtration, the PDs for $\Bbbk = \mathbb{Z}_2$ and $\Bbbk = \mathbb{R}$ are clearly different. For analysis of persistent homology with torsions, \cite{BJDMC} proposed an efficient algorithm to compute PD for multiple coefficient fields by utilizing the Chinese remainder theorem. Then, the following questions naturally arise. \begin{itemize} \item What condition ensures the independence of the choice of the field $\Bbbk$? \item Is there an efficient algorithm to check the above condition? \item How often does $D_q(\mathbb{X}; \Bbbk)$ change as the field changes $\Bbbk$? \item When $D_q(\mathbb{X}; \Bbbk)$ changes depending of the choice of $\Bbbk$, how does $D_q(\mathbb{X}; \Bbbk)$ change? \end{itemize} In this paper, we offer complete answers for the first and second questions, and partial answers for the third and fourth questions. \subsection{Results} To describe the results of the paper, we give some assumptions. We always assume the finiteness of the filtration. A filtration is \emph{finite} if $X = \cup_t X_t$ is a finite simplicial/cell/cubical complex. This condition ensures the existence and uniqueness of the interval decomposition~\citep{zc}. This assumption is reasonable since an infinite filtration cannot be represented on a computer and we cannot use such a filtration in practical applications. To consider field choice problems, we always restrict the candidates of a field to $\mathbb{C}, \mathbb{R}, \mathbb{Q},$ and $\mathbb{Z}_p$ for a prime $p$. \begin{question}\label{q:cond} When is $D_q(\mathbb{X}; \Bbbk)$ independent of the choice of the field $\Bbbk$? \end{question} To consider Question~\ref{q:cond}, it is desirable for the following proposition to hold since $H_q(X_t; \mathbb{Z})$ is often free for every $t$ in practical cases. \begin{proposition}[Incorrect!] If $H_q(X_t; \mathbb{Z})$ is free for every $t \in T$, the persistent homology $H_q(\mathbb{X}; \Bbbk)$ has the same decomposition for any field $\Bbbk$. \end{proposition} However, we have a counterexample of this proposition (Fig.~\ref{fig:mobius}). Let $M$ be a M\"obius strip and $\partial M$ be its boundary. Both $H_1(\partial M; \mathbb{Z})$ and $H_1(M; \mathbb{Z})$ are isomorphic to $\mathbb{Z}$, and the homomorphism \begin{align*} H_1(\partial M; \mathbb{Z}) \to H_1(M; \mathbb{Z}) \end{align*} is isomorphic to \begin{align*} n \in \mathbb{Z} \mapsto 2n \in \mathbb{Z} \end{align*} and the interval decomposition on $\mathbb{R}$ and $\mathbb{Z}_2$ gives the different decomposition as follows: \begin{align*} &\mathbb{X} : X_0 = \emptyset \subset X_1 = \partial M \subset X_2 = M, \\ &H_1(\mathbb{X}; \mathbb{Z}_2) = I(1, 2) \oplus I(2, \infty), \\ &H_1(\mathbb{X}; \mathbb{R}) = I(1, \infty). \end{align*} In this example, both $H_1(\partial M; \mathbb{Z})$ and $H_1(M; \mathbb{Z})$ are free, but $H_1(M, \partial M; \mathbb{Z}) \simeq \mathbb{Z}_2$ and this is not free. This fact is key to the different diagrams. Section~\ref{sec:example} shows some other examples. \begin{figure}[hbtp] \centering \includegraphics[width=0.7\hsize]{mobius.pdf} \caption{M\"obius strip and its boundary} \label{fig:mobius} \end{figure} We present the following theorem. \begin{theorem}\label{thm:allfree} $D_q(\mathbb{X}; \Bbbk)$ is independent of the choice of $\Bbbk$ if $H_q(X_n, X_m; \mathbb{Z})$ is free for any $0\leq m < n \in T$ and $H_{q-1}(X_n; \mathbb{Z})$ is free for any $n \in T$. \end{theorem} This theorem yields the following corollaries. \begin{corollary}\label{cor:M_0} $D_0(\mathbb{X}; \Bbbk)$ is always independent of the choice of $\Bbbk$. \end{corollary} \begin{corollary}\label{cor:M_munus_1} When $\mathbb{X}$ is a filtration of finite cell/simplicial/cubical complexes embedded in $\mathbb{R}^M$, the $(M-1)$th persistent homology gives the same PD among any fields $\Bbbk$. \end{corollary} Corollary~\ref{cor:M_0} derives from the fact that $H_{-1}(\cdot) = 0$ and $H_0(X_n, X_m; \mathbb{Z})$ is free for any $n > m$. Corollary~\ref{cor:M_munus_1} is proved in Section~\ref{subsec:M_minus_1}. The above two corollaries ensure that if a filtration is embedded in $\mathbb{R}^2$, all non-trivial persistence diagrams $D_0$ and $D_1$ do not depend on the choice of the coefficient field. We also have the following theorem which provides the sufficient condition for the freeness of $H_q(X_n, X_m; \mathbb{Z})$. \begin{theorem}\label{thm:allfree2} For a given $q$, $H_q(X_n, X_m; \mathbb{Z})$ are free for any $0\leq m < n \in T$ if $D_q(\mathbb{X}; \Bbbk)$ is independent of the choice of $\Bbbk$ and $H_{q-1}(X_n; \mathbb{Z})$ is free for any $n \in T$. \end{theorem} From the above two theorems, we have the following corollary. \begin{corollary}\label{cor:allfree3} Let $M$ be a non-negative integer. $D_q(\mathbb{X}; \Bbbk)$ is independent of the choice of $\Bbbk$ for all $q = 0, \ldots, M$ if and only if $H_q(X_n, X_m; \mathbb{Z})$ are free for any $0\leq m < n \in T$, and $q = 0, \ldots, M$. \end{corollary} From the above discussion another question arises. \begin{question} Is there an efficient algorithm for checking the condition of Corollary~\ref{cor:allfree3}? \end{question} Such an algorithm would be useful to provide information as to whether we should be concerned about field choice. Of course, we can compute relative homology groups for all $m < n$ on a computer, but that would be cumbersome and inefficient because the number of possible pairs $(m, n)$ is $(N+1)N/2$. The computation cost (time complexity) is $O(N^2 G)$, where $G$ is the average cost of computing $H_q(X_n, X_m; \mathbb{Z})$. It is known that the time complexity of computing a PD is $O(G)$ \footnote{To compute the torsion subgroup of a homology group, we need to compute the Smith normal form of the boundary matrix, and the computational cost of the Smith normal form is $O(n^{\theta})$ in the worst case where $n$ is the number of simplices and $\theta \approx 2.376$ is a constant. The constant $\theta$ comes from the time complexity of the multiplication of two $n \times n$ matrices. The time complexity of persistent homology is also $O(n^{\theta})$ in the worst case. See \citep{Milosavljevic:2011:ZPH:1998196.1998229,Storjohann} for further details.}. To describe the algorithm, we assume the following condition. \begin{cond}\label{cond:finite} $X = \{\sigma_1, \sigma_2, \ldots, \sigma_N\}$ is a finite simplicial, cubical, or cell complex and the subset $X_k = \{\sigma_1, \ldots, \sigma_k\}$ is a sub-complex of $X$. \end{cond} With this setting, we consider the filtration of complexes $\mathbb{X}: \emptyset = X_0 \subset X_1 \subset \cdots \subset X_N$. Since the filtration is finite, we can transform the persistence decomposition problem into the problem under Condition~\ref{cond:finite}. The following theorem is proved in Section~\ref{sec:algorithm}. \begin{theorem}\label{thm:allfreealg} There is an algorithm for judging the condition in Corollary~\ref{cor:allfree3} whose time complexity is the same as the algorithm for computing a PD. \end{theorem} The algorithm is shown in Algorithm~\ref{alg:allfree} in Section~\ref{sec:algorithm}. In Section~\ref{sec:implementation}, we apply the algorithm to some examples shown in Section~\ref{sec:example} and demonstrate that it performs well. A performance benchmark is also covered in that section. We now pose the following additional question. \begin{question} How often do we face filtrations with non-trivial torsion subgroups? \end{question} We can construct such an example by a M\"obius strip as shown above, but would we often face such a filtration? To demonstrate the probability of torsions we conduct a numerical experiment for random data in $\mathbb{R}^3$. From this experiment, we show that filtrations with non-trivial torsion subgroups are very rare. This suggests that, in practical terms, if the data is in $\mathbb{R}^3$, we do not need to be particularly concerned about the torsion problem. We also conduct another numerical experiment for random filtrations in high dimensional simplex. The second experiment shows that the filtrations with non-trivial torsion subgroups are usual when the space is high dimensional. The following question is also important. \begin{question} When $D_q(\mathbb{X}; \Bbbk)$ changes depending of the choice of $\Bbbk$, how does $D_q(\mathbb{X}; \Bbbk)$ change? \end{question} In the above example about a M\"obius strip, a long interval $I(1, \infty)$ is split into two shorter intervals, $ I(1, 2)$ and $I(2, \infty)$, when $\Bbbk$ changes from $\mathbb{R}$ to $\mathbb{Z}_2$. From the example, we expect that a long interval indecomposable tends to be split into shorter intervals when $\Bbbk$ changes from $\mathbb{R}$ to $\mathbb{Z}_p$. The following theorem proved in Section~\ref{sec:convex} partially answers the question. \begin{theorem}\label{thm:r-wasserstein} Assume that $H_q(X_t; \mathbb{Z})$ and $H_{q-1}(X_t; \mathbb{Z})$ are free for all $t$ and $H_q(\cup_t X_t) = 0$. Let $f$ be a $C^2$ convex function on $[0, \infty)$ with $f(0) = 0$. Then the following inequality holds: \begin{align*} \sum_{(b, d) \in D_q(\mathbb{X}; \mathbb{R})} f(d - b) \geq \sum_{(b, d) \in D_q(\mathbb{X}; \mathbb{Z}_p)} f(d - b). \end{align*} When $f$ is strictly convex, the equality holds if and only if $D_q(\mathbb{X}; \mathbb{R}) = D_q(\mathbb{X}; \mathbb{Z}_p)$. \end{theorem} For $f(x) = x^r$ with $r > 1$, the inequality means \begin{align*} W_r(D_q(\mathbb{X}; \mathbb{R}), \emptyset) \geq W_r(D_q(\mathbb{X}; \mathbb{Z}_p), \emptyset), \end{align*} where $W_r$ is the $r$-Wasserstein distance. In some sense, the $r$-Wasserstein distance from the empty diagram indicates the information richness of the diagram. Therefore, $D_q(\mathbb{X}; \mathbb{R})$ contains richer information than $D_q(\mathbb{X}; \mathbb{Z}_p)$ under the condition of the theorem. The remainder of the paper is organized as follows. Section~\ref{sec:ph} reviews the basic concepts of persistent homology. Section~\ref{sec:example} shows some examples which exhibit the dependency of PDs to their coefficient fields. Section~\ref{sec:proof-allfree} and Section~\ref{sec:proof-allfree2} prove Theorem~\ref{thm:allfree} and Theorem~\ref{thm:allfree2}. Section~\ref{sec:algorithm} presents an algorithm which permits efficient judgement and the proof which testifies to the correctness of the algorithm. Section~\ref{sec:implementation} introduces an implementation of the algorithm in HomCloud. This section also shows the performance benchmark. Section~\ref{sec:probability} presents numerical experiments to measure the probability of the appearance of non-trivial torsions in random filtrations. Section~\ref{sec:convex} contains the proof of Theorem~\ref{thm:r-wasserstein} and, finally, conclusions are offered in Section~\ref{sec:conclusion}. \section{Persistent homology}\label{sec:ph} In this section, we prepare some fundamental concepts for persistent homology. \subsection{Filtrations} A filtration is an increasing sequence of topological spaces. One typical filtration is the union of $r$-balls constructed from a pointcloud in $\mathbb{R}^{M}$. For a pointcloud, a set of finite points $\{x_i\}$, $X_r$ is defined as \begin{equation} X_r = \cup_i B_{x_i}(r), \end{equation} where $B_{x}(r)$ is the closed ball whose center is $x$ and radius is $r$. The sequence of $X_r$ parameterized by $r$, $\{X_r\}_{r \geq 0}$, is obviously a filtration. This filtration is used to investigate the shape formed by the pointcloud. For a practical application of persistent homology, we usually use finite simplicial or cubical filtrations since such filtrations are practical to consider on a computer. One well-known filtration is a \v{C}ech filtration. The \emph{\v{C}ech complex} $\mathrm{\v{C}ech}(P, r)$ of a pointcloud $P = \{x_i\}$ with radius parameter $r \geq 0$ is defined as follows: \begin{equation} \mathrm{\v{C}ech} (P, r) = \{ \{x_{i_1}, \ldots, x_{i_k}\} \subset P \mid \bigcap_{n=1}^k B_{x_{i_n}}(r) \not = \emptyset \}. \end{equation} The filtration $\{\mathrm{\v{C}ech}(P, r)\}_{r \geq 0}$ is called a \emph{\v{C}ech filtration}. From the nerve theorem, $\mathrm{\v{C}ech}(P, r)$ is homotopic to $\cup_i B_{x_i}(r)$ and we can use the \v{C}ech filtration to investigate the union of $r$-balls. There are many simplices in a \v{C}ech complex for a large pointcloud and we usually use an \emph{alpha complex} \citep{alphashape1,alphashape2} instead since the alpha complex is homotopic to the \v{C}ech complex and the number of simplices of the alpha complex is much smaller than the \v{C}ech complex. The alpha complex has another advantage in that it can be embedded in $\mathbb{R}^M$ but such embedding is impossible for the \v{C}ech complex. When a filtration is finite, it is essentially time-discrete even if $T = \mathbb{R}_+$. Therefore we assume $T = \{0, \ldots, N\}$ for the proofs of this paper except Theorem~\ref{thm:r-wasserstein}. In addition, under this assumption, it is straightforward to configure a filtration satisfying Condition~\ref{cond:finite} by ordering simplices appropriately; hence, we can assume the condition without loss of generality. Since Condition~\ref{cond:finite} is useful to describe algorithms, we sometimes assume this and consider the filtration $X_0 \subset X_1 \subset \cdots \subset X_N$ where $X_k = \{\sigma_1, \ldots, \sigma_k\}$. \subsection{Computation of a persistence diagram} Under Condition~\ref{cond:finite}, Algorithm~\ref{alg:pd} computes the PD of the filtration~\citep{elz,zc,Otter2017}. To simplify the algorithm, all simplices of all dimensions are mixed and in the output all birth-death pairs of all degrees are also mixed. In this algorithm, $L_B(j)$ means \begin{equation} L_B(j) = \left\{ \begin{array}{ll} \max \{i \mid B_{ij} \not = 0\} & \text{if column $j$ of $B$ is nonzero},\\ -\infty & \text{if column $j$ of $B$ is zero}, \end{array} \right. \end{equation} where $B$ is a matrix and $j$ is an integer. Furthermore, in this algorithm, matrix $B$ is reduced from left column to right column. After terminating the algorithm, the PD is computed as follows: \begin{equation}\label{eq:pd} \begin{aligned} D(\mathbb{X}) =& \{ (L_{\hat{B}}(j), j) \mid L_{\hat{B}}(j) \not = -\infty \} \\ &\cup \{(j, \infty) \mid L_{\hat{B}}(j) = -\infty \text{ and } \forall i, L_{\hat{B}}(i) \not = j\}, \end{aligned} \end{equation} where $\hat{B}$ is the matrix returned by the algorithm. The $q$th PD is given from $D(\mathbb{X})$ as follows: \begin{equation} D_q(\mathbb{X}) = \{(i, j) \in D(\mathbb{X}) \mid \dim \sigma_i = q\}. \end{equation} \begin{algorithm}[ht] \caption{Algorithm to compute persistence diagrams}\label{alg:pd} \begin{algorithmic} \State $B \leftarrow $ the boundary matrix with respect to the basis $\{\sigma_1, \ldots, \sigma_N\}$ \For{$j=1, \ldots, N$} \While{there exists $i < j$ with $L_B(i) = L_B(j) \not = -\infty$ } \State let $s = B_{L_B(j), j} / B_{L_B(i), i}$ \State add $(-s) \times (\textrm{column}\ i)$ to column $j$ of $B$ \Comment left-to-right reduction \EndWhile \EndFor \State \Return B \end{algorithmic} \end{algorithm} Justification for the algorithm is provided in Appendix \ref{sec:proof-alg1}. Indeed, Theorem~\ref{thm:allfreealg} shows that the algorithm for judging the condition of Corollary~\ref{cor:allfree3} is given by restricting Algorithm~\ref{alg:pd} to integer coefficients. Therefore, the time complexity of the Theorem~\ref{thm:allfreealg} algorithm is as per Algorithm~\ref{alg:pd}. \subsection{Persistent Betti number} From the definition of a PD, we have the following relationship between the map $H_q(X_m; \Bbbk) \to H_q(X_n; \Bbbk)$ and a PD: \begin{equation}\label{eq:persistence-betti-number} \begin{aligned} \beta_m^n(\Bbbk) :=& \rank\left(H_q(X_m; \Bbbk) \to H_q(X_n; \Bbbk)\right) \\ =& \#\{ (b, d) \in D_q(\mathbb{X}; \Bbbk) \mid b \leq m \leq n < d \}. \end{aligned} \end{equation} This $\beta_m^n(\Bbbk)$ is called a \emph{persistent Betti number} or a \emph{rank invariant}. Hence, the following identity holds: \begin{equation}\label{eq:multiplicity-1} (\text{multiplicity of $D_q(\mathbb{X}; \Bbbk)$ at $(b, d)$}) = \beta^{d-1}_b(\Bbbk) - \beta^{d-1}_{b-1}(\Bbbk) - \beta^{d}_{b}(\Bbbk) + \beta^{d}_{b-1}(\Bbbk). \end{equation} When $d = \infty$, the following equation holds instead: \begin{equation}\label{eq:multiplicity-2} (\text{multiplicity of $D_q(\mathbb{X}; \Bbbk)$ at $(b, \infty)$}) = \beta^{N}_b(\Bbbk) - \beta^{N}_{b-1}(\Bbbk). \end{equation} The next lemma follows directly from the foregoing. \begin{lemma}\label{lem:pbn} $D_q(\mathbb{X}, \Bbbk) = D_q(\mathbb{X}, \Bbbk')$ if and only if $\beta_m^n(\Bbbk) = \beta_m^n(\Bbbk')$ for all $0 \leq m \leq n \leq N$. \end{lemma} \subsection{Universal Coefficient Theorem} The universal coefficient theorem is fundamental for homology theory and plays an important role in this paper. We review the theorem here to foreground what follows. The \emph{universal coefficient theorem for homology} is as follows~\citep{AT}. \begin{theoremext} Let $X$ be a topological space, $\Bbbk$ a field, and $q \geq 0$. The following sequence is a natural short exact sequence: \begin{equation} \begin{aligned} 0 \to H_q(X; \mathbb{Z}) \otimes \Bbbk \to H_q(X; \Bbbk) \to \mathrm{Tor}(H_{q-1}(X; \mathbb{Z}), \Bbbk) \to 0. \end{aligned} \end{equation} Furthermore, this sequence splits, though not naturally. \end{theoremext} We use the above theorem in the following form. \begin{theoremext}\label{thm:ucm} Let $X$ and $Y$ be topological spaces, $f:X \to Y$ a continuous map, $\Bbbk$ a field, and $q \geq 0$. If $H_{q-1}(X; \mathbb{Z})$ and $H_{q-1}(Y; \mathbb{Z})$ are free, the following commutative diagram holds: \begin{equation} \vcenter{ \xymatrix{ H_q(X; \mathbb{Z})\otimes \Bbbk \ar[r]^{\simeq} \ar[d]_{f_*\otimes \mathrm{id}_{\Bbbk}} & H_q(X; \Bbbk) \ar[d]_{f_*} \\ H_q(Y; \mathbb{Z})\otimes \Bbbk \ar[r]^{\simeq} & H_q(Y; \Bbbk). \ar@{}[lu]|{\circlearrowright} } } \end{equation} \end{theoremext} This theorem states that the induced map $f_* : H_q(X; \Bbbk) \to H_q(Y; \Bbbk)$ is completely described by $f_*\otimes \mathrm{id}_\Bbbk : H_q(X; \mathbb{Z}) \otimes \Bbbk \to H_q(Y; \mathbb{Z})\otimes \Bbbk$ if $H_{q-1}(X; \mathbb{Z})$ and $H_{q-1}(Y; \mathbb{Z})$ are free. We use the theorem for an inclusion map between simplicial/cell/cubical complexes. \section{Examples of diagrammatic changes induced by coefficient field changes}\label{sec:example} In this section, we will give some examples of persistent homology, whose interval decomposition depends on the choice of coefficient field. \begin{example} \label{exm:bouquet} Let $S^1$ be a circle. We consider a filtration $\mathbb{X} : S^1 \xrightarrow{f} S^1 \vee S^1 \xrightarrow{g} S^1$, where $S^1 \vee S^1$ is a bouquet of $2$-circles, $f=\begin{pmatrix}1 \\ 1 \end{pmatrix}$ and $g=\begin{pmatrix}1&1 \end{pmatrix}$. By taking the 1st homology of this filtration, we obtain the 1st persistent homology $$ H_1 (\mathbb{X} ; \mathbb{Z})= \xymatrix{ \mathbb{Z} \ar[r]^{\begin{pmatrix}1 \\ 1 \end{pmatrix}} & \mathbb{Z}^2 \ar[r]^{\begin{pmatrix}1&1 \end{pmatrix}} & \mathbb{Z}, } $$ with a coefficient ring $\mathbb{Z}$. Then $H_1 (\mathbb{X} ; \mathbb{Z}_2) = H_1(\mathbb{X} ; \mathbb{Z}) \otimes_\mathbb{Z} \mathbb{Z}_2$ has the interval decomposition $I(1,2) \oplus I(2,3)$. On the other hand, $H_1 (\mathbb{X} ; \mathbb{R}) = H_1 (\mathbb{X} ; \mathbb{Z}) \otimes_\mathbb{Z} \mathbb{R}$ has the interval decomposition $I(1,3) \oplus I(2,2)$. Thus, the interval decomposition of the 1st persistent homology of $\mathbb{X}$ depends on the choice of coefficient field. \begin{figure}[h] \centering \includegraphics[width=1.0\hsize]{g3404.png} \\ \caption{Visualization of $\mathbb{X} : S^1 \xrightarrow{f} S^1 \vee S^1 \xrightarrow{g} S^1$} \end{figure} \end{example} Note that if we consider a bouquet of $p$-circles for a prime $p$, then we obtain the 1st persistent homology, which has different decompositions over $\mathbb{Z}_p$ and $\mathbb{R}$. By using Example \ref{exm:bouquet}, we can consider the 1st persistent homology, whose interval decomposition depends on the choice of characteristic $p>0$. \begin{example} \label{exm:charp} Let $$ M= \xymatrix{ \mathbb{Z} \ar[r]^{\begin{pmatrix}1 \\ 1 \end{pmatrix}} & \mathbb{Z}^2 \ar[r]^{\begin{pmatrix}1&1 \end{pmatrix}} & \mathbb{Z} \ar[r]^{\begin{pmatrix}1 \\ 1 \\1 \end{pmatrix}} & \mathbb{Z}^3 \ar[r]^{\begin{pmatrix}1 & 1 & 1 \end{pmatrix}} & \mathbb{Z} } $$ be the 1st persistent homology. Then $M$ has the following interval decomposition: $$ M \cong \left\{ \begin{array}{cc} I(1,2) \oplus I(2,5) \oplus I(4,4)^2 & \text{if } p=2 \\ I(1,4) \oplus I(2,2) \oplus I(4,4) \oplus I(4,5) & \text{if } p=3 \\ \end{array}\right. . $$ \begin{figure}[h] \centering \includegraphics[width=1.0\hsize]{g2655_2.png} \\ \caption{Visualization of $\mathbb{X}$ with $M=H_1 (\mathbb{X})$} \end{figure} \end{example} Other examples are double and triple loop pointclouds. Figure~\ref{fig:doubletripleloop} (a) shows the double loop pointcloud. The pointcloud is located on the boundary of a M\"obius strip. We compute the 1st PDs of the double loop pointcloud with fields $\mathbb{Z}_2, \mathbb{Z}_3,$ and $\mathbb{Z}_5$. The alpha filtration of the pointcloud is used for the computation. The diagrams are shown in Figure~\ref{fig:doubletripleloop} (b), (c), and (d). Note that (c) and (d) are the same diagram. The difference between (b) and (c) is only two birth-death pairs. Figure~\ref{fig:doubletripleloop} (e) shows the triple loop pointcloud and (f), (g), and (h) show the 1st PDs of the triple loop pointcloud with fields $\mathbb{Z}_2, \mathbb{Z}_3,$ and $\mathbb{Z}_5$. To be expected, (f) and (h) are the same diagram and (g) is different from (f) and (h). \begin{figure}[h] \centering \includegraphics[width=1.0\hsize]{doubleloop.pdf} \\ \includegraphics[width=1.0\hsize]{tripleloop.pdf} \caption{1st PDs with various fields for double and triple loop pointclouds. (a) A double loop point cloud. (b) The PD of the double loop with $\mathbb{Z}_2$. (c) The PD of the double loop with $\mathbb{Z}_3$. (d) The PD of the double loop with $\mathbb{Z}_5$. (e) A triple loop point cloud. (f) The PD of the triple loop with $\mathbb{Z}_2$. (g) The PD of the triple loop with $\mathbb{Z}_3$. (h) The PD of the triple loop with $\mathbb{Z}_5$ } \label{fig:doubletripleloop} \end{figure} \section{Proof of Theorem \ref{thm:allfree}}\label{sec:proof-allfree} The following proposition is required to prove the theorem. \begin{proposition}\label{prop:freefree} If $H_q(X_n, X_m; \mathbb{Z})$ is free, $\coker(\phi_m^n: H_q(X_m; \mathbb{Z}) \to H_q(X_n; \mathbb{Z}))$ is also free. \end{proposition} \begin{proof} We have the following long exact sequence for the pair $(X_n, X_m)$: \begin{align} \cdots \to H_q(X_m; \mathbb{Z}) \xrightarrow{\phi_m^n} H_q(X_n; \mathbb{Z}) \xrightarrow{\psi_m^n} H_q(X_n, X_m; \mathbb{Z}) \to \cdots, \end{align} where $\psi_m^n$ is induced by canonical projection. Therefore, we have the following relationship between $\coker(\phi_m^n)$ and $H_q(X_n, X_m; \mathbb{Z})$. \begin{align} \coker(\phi_m^n) = H_q(X_n; \mathbb{Z}) / \im \phi_m^n = H_q(X_n; \mathbb{Z}) / \ker \psi_m^n \simeq \im \psi_m^n \subset H_q(X_n, X_m; \mathbb{Z}) \end{align} To complete the proof, we show that $\im \psi_m^n$ is free, and this derives from the following well-known theorem. \begin{theoremext}\label{thmext:submodule} Any sub-module of a free $\mathbb{Z}$-module is also free. \end{theoremext} \end{proof} From the assumption of the theorem, $H_q(X_m; \mathbb{Z}) = H_q(X_m, X_0; \mathbb{Z})$ is free for all $m$. Hence, $\phi_m^n: H_q(X_m; \mathbb{Z}) \to H_q(X_n; \mathbb{Z})$ is a homomorphism between two finitely generated free $\mathbb{Z}$-modules and the map has a Smith normal form (SNF). That is, by taking an appropriate basis, $\phi_m^n$ can be represented by the following $\mathbb{Z}$ matrix: \begin{align}\label{eq:snf} \begin{pmatrix} \alpha_1 & 0 & 0 & \cdots & 0 & \\ 0 & \alpha_2 & 0 & \cdots & 0 & \\ 0 & 0 & \ddots & & &\ \ O \ \ \\ \vdots & \vdots & & \ddots & & \\ 0 & 0 & & & \alpha_K & \\ \\ & & O & & & \ \ O \ \ \\ \\ \end{pmatrix}, \end{align} where $0 < \alpha_k \in \mathbb{Z}$ and $\alpha_k \mid \alpha_{k+1}$ for any $k$. Then from Theorem~\ref{thm:ucm}, the following relationship holds: \begin{equation} \begin{aligned} \beta_m^n(\Bbbk) & = \rank (\phi_m^n: H_q(X_m; \Bbbk) \to H_q(X_n; \Bbbk)) \\ & = \rank (\phi_m^n \otimes \mathrm{id}_\Bbbk: H_q(X_m; \mathbb{Z}) \otimes \Bbbk \to H_q(X_n; \mathbb{Z}) \otimes \Bbbk). \end{aligned}\label{eq:beta-mn} \end{equation} From \eqref{eq:snf} and \eqref{eq:beta-mn}, we know that $\beta_m^n(\Bbbk)$ is independent of the choice of $\Bbbk$ if and only if $\alpha_1 = \cdots = \alpha_K = 1$. From SNF, we also have the following: \begin{align*} \coker \phi_m^n \simeq &\bigoplus_{k=k_0}^K(\mathbb{Z} / \alpha_k\mathbb{Z}) \oplus \mathbb{Z}^{L - K}, \\ \text{where } & k_0 = 1 + \max\{i \mid \alpha_m = 1\}, \\ & L = \rank H_q(X_n; \mathbb{Z}). \end{align*} This means that $\alpha_1 = \cdots = \alpha_K = 1$ if and only if $\coker(\phi_m^n)$ is free and the condition is shown from Prop~\ref{prop:freefree} and the assumption of the theorem. \subsection{Proof of Corollary~\ref{cor:M_munus_1}}\label{subsec:M_minus_1} Standard homology theory~\citep[Corollary 3.46, pp. 256]{AT} shows that $H_{M-1}(X_n; \mathbb{Z})$ and $H_{M-2}(X_n; \mathbb{Z})$ are free under the condition of this corollary. The above corollary is shown by using the Alexander duality. $H_{M-1}(X_n, X_m; \mathbb{Z})$ is also free since this relative homology group is isomorphic to a 0th relative cohomology group because of the Alexander duality. Therefore, from Theorem~\ref{thm:allfree}, this can be applicable to the filtration. \section{Proof of Theorem~\ref{thm:allfree2}}\label{sec:proof-allfree2} The proof is similar to that of Theorem~\ref{thm:allfree}, but slightly more complex. We prepare the following proposition. \begin{proposition}\label{prop:quotfree} $H_q(X_n, X_m; \mathbb{Z})$ is free if $\coker(\phi_m^n: H_q(X_m; \mathbb{Z}) \to H_q(X_n; \mathbb{Z}))$ and $H_{q-1}(X_m; \mathbb{Z})$ are free. \end{proposition} \begin{proof} From the long exact sequence for the pair $(X_n, X_m)$, \begin{equation} \begin{aligned} \cdots \to H_q(X_m; \mathbb{Z}) \xrightarrow{\phi_m^n} H_q(X_n; \mathbb{Z}) \xrightarrow{\psi_m^n} H_q(X_n, X_m; \mathbb{Z}) \xrightarrow{\partial} H_{q-1}(X_m; \mathbb{Z}) \to \cdots, \end{aligned} \end{equation} we have the following facts: \begin{equation} \begin{aligned} \coker(\phi_m^n) & \simeq \im \psi_m^n \subset H_q(X_n, X_m; \mathbb{Z}), \\ \im \partial & \simeq H_q(X_n, X_m; \mathbb{Z}) / \im \psi_m^n . \end{aligned} \end{equation} $\im \partial$ is free since $H_{q-1}(X_m)$ is free. We complete the proof by the following theorem from standard algebra. \begin{theoremext} Let $M$ be a module over $\mathbb{Z}$ and $N$ be a sub-module of $M$. $M$ is finitely generated and free if $N$ and $M/N$ are finitely generated and free. \end{theoremext} \end{proof} \subsection{Proof of Theorem~\ref{thm:allfree2}} We assume that $D_q(\mathbb{X}; \Bbbk)$ is independent of the choice of $\Bbbk$. Then from Lemma~\ref{lem:pbn}, $\beta_m^n(\Bbbk)$ is independent of $\Bbbk$ for any $m$ and $n$. Especially, for any $n$ and $q$, $\beta_n^n(\Bbbk) = \dim H_q(X_n; \Bbbk)$ is independent of $\Bbbk$ and therefore $H_q(X_n; \mathbb{Z})$ is free due to the universal coefficient theorem since $H_{q-1}(X_n; \mathbb{Z})$ is free. Then $\phi_m^n$ has SNF and $\coker \phi_m^n$ is free because of the discussion of the proof of Theorem~\ref{thm:allfree}. From the above fact and Proposition~\ref{prop:quotfree}, we conclude that $H_q(X_n, X_m; \mathbb{Z})$ is free. \subsection{Proof of Corollary~\ref{cor:allfree3}} From Theorem~\ref{thm:allfree}, it is straightforward to show that $D_q(\mathbb{X}; \Bbbk)$ is independent of the choice of $\Bbbk$ for all $q = 0, \ldots, M$ if $H_q(X_n, X_m; \mathbb{Z})$ are free for any $0\leq m < n \in T$, and $q = 0, \ldots, M$. We can show the converse by induction on $q$. For $q = 0$, it is trivial that $H_0(X_n, X_m; \mathbb{Z})$ is free and the induction process proceeds by using Theorem~\ref{thm:allfree2}. \section{Algorithm to determine the dependency of $D_q(\mathbb{X}; \Bbbk)$ on $\Bbbk$}\label{sec:algorithm} In this section, we explore an algorithm to judge the existence of non-zero torsion and prove Theorem~\ref{thm:allfreealg}. See Algorithm~\ref{alg:allfree}. Now we prove the following facts. \begin{itemize} \item If the algorithm returns ``independent'', the given filtration satisfies the condition of Corollary~\ref{cor:allfree3}. Therefore, $D_q(\mathbb{X}; \Bbbk)$ is independent of the choice of $\Bbbk$. \item If the algorithm returns ``dependent'', the given filtration does not satisfy the condition of Corollary~\ref{cor:allfree3}. Therefore, $D_q(\mathbb{X}; \Bbbk)$ depends on the choice of $\Bbbk$. \end{itemize} \begin{algorithm}[ht] \caption{Algorithm to determine the dependency of $D_q(\mathbb{X}; \Bbbk)$ on $\Bbbk$}\label{alg:allfree} \begin{algorithmic} \State let $B$ be the matrix representation of the boundary operator \For{$j=1, \ldots, N$} \Comment (OUTERLOOP) \While{there exists $i < j$ with $L_B(i) = L_B(j) \not = -\infty$} \Comment (INNERLOOP) \State let $s = - B_{L_B(j), j} / B_{L_B(i), i}$ \Comment (A) \State add $s \times (\textrm{column}\ i)$ to column $j$ of $B$ \EndWhile \If{$L(j) \not = -\infty$ and $B_{L_{B}(j), j} \not \in \{\pm 1\}$} \Comment (B) \State print $|B_{L_B(j), j}|$ \State \Return ``dependent'' \EndIf \EndFor \State \Return ``independent'' \end{algorithmic} \end{algorithm} We remark that $B_{L_{B}(i), i}$ in this algorithm is always $\pm 1$ at (A), so the division at (A) always applies. This is because the condition is checked at (B). For the proof, we use ideas presented in the Appendix~\ref{sec:proof-alg1}. We also use Notation~\ref{notation:appendix} in Appendix~\ref{sec:proof-alg1} and check whether $H(X_n, X_m; \mathbb{Z}) = \bigoplus_{q=0}^{M} H_q(X_n, X_m; \mathbb{Z})$ has a non-zero torsion subgroup for every $m < n$. We assume that $\dim(\mathbb{X}) \leq M + 1$ by removing all simplices whose dimensions exceed $M + 1$. First we prove $T(H_q(X_n, X_m; \mathbb{Z})) = 0$ for every $m < n$ when the algorithm returns ``independent''. Let $\hat{B}$ be the matrix $B$ just before returning ``independent''. Since division is not used in the proof of Fact~\ref{fact:algpd-basis} in Section~\ref{subsec:proof-algpd-basis}, the discussion can also apply to $\mathbb{Z}$-homology and we can find a basis of $C(X_N; \mathbb{Z})$, $\{\tilde{\sigma}_1, \ldots, \tilde{\sigma}_N\}$, satisfying Condition~\ref{cond:phbasis} (i), (ii), (iii), and (iv). We can explicitly write the bases of $C(X_m; \mathbb{Z})$ and $C(X_n, X_m; \mathbb{Z}) $ as follows: \begin{equation} \begin{aligned} &\{\tilde{\sigma}_{1}, \ldots, \tilde{\sigma}_{m} \} \text{ is a basis of } C(X_m; \mathbb{Z}), \\ &\{\tilde{\sigma}_{m+1} + C(X_m;\mathbb{Z}), \ldots, \tilde{\sigma}_{n} + C(X_m;\mathbb{Z}) \} \text{ is a basis of } C(X_n, X_m; \mathbb{Z}). \\ \end{aligned} \end{equation} Let $\partial_{n,m}: C(X_n, X_m; \mathbb{Z}) \to C(X_n, X_m; \mathbb{Z}) $ be the boundary operator on relative chain complexes. From Condition~\ref{cond:phbasis} (ii), (iii), and (iv), $\ker \partial_{n,m}$ and $\im \partial_{n,m}$ are both $\mathbb{Z}$ free modules and the bases are \begin{equation} \begin{aligned} &\{\tilde{\sigma}_k + C(X_m; \mathbb{Z}) \mid m < k \leq n, k \in D' \sqcup E\} \\ \sqcup &\{\tilde{\sigma}_k + C(X_m; \mathbb{Z}) \mid m < k \leq n, k \in D, L_{\hat{B}}(k) < m\}, \end{aligned} \end{equation} and \begin{equation} \begin{aligned} &\{\tilde{\sigma}_{L(k)} + C(X_m; \mathbb{Z}) \mid m < k \leq n, k \in D, m < L_{\hat{B}}(k)\} \\ =&\{\tilde{\sigma}_{k} + C(X_m; \mathbb{Z}) \mid m < k \leq n, k \in D', L_{\hat{B}}^{-1}(k) \leq n \}. \end{aligned} \end{equation} Therefore, the basis of $H(X_n, X_m; \mathbb{Z}) = \ker \partial_{n,m} / \im \partial_{n,m}$ can be written as follows: \begin{equation} \begin{aligned} & \{[\tilde{\sigma}_k + C(X_m; \mathbb{Z})] \mid m < k \leq n, k \in E \} \\ \sqcup & \{[\tilde{\sigma}_k + C(X_m; \mathbb{Z})] \mid m < k \leq n, k \in D', n < L_{\hat{B}}^{-1}(k) \}, \end{aligned} \end{equation} where $[z + C(X_m; \mathbb{Z})]$ is a homology class of $z + C(X_m; \mathbb{Z})$ in $H(X_n, X_m; \mathbb{Z})$. Therefore, $H(X_n, X_m; \mathbb{Z})$ is a free $\mathbb{Z}$-module and we complete the proof for the ``independent'' case. Next we show that there is a pair $(m, n)$ such that $T(H(X_n, X_m; \mathbb{Z})) \not = 0$ if the algorithm returns ``dependent''. In that case, condition (B) in the algorithm is true for one $j$, so let $n$ be that $j$ and $\hat{B}$ be the matrix $B$ at that time. We consider the subfiltration $\mathbb{X}_n = \emptyset = X_0 \subset \cdots \subset X_{n-1}$. Algorithm~\ref{alg:allfree} applies on the subfiltration, so we can find a basis of $C(X_{n-1})$ satisfying Condition~\ref{cond:phbasis}(i)-(iv). That is, there is a basis of $C(X_{n-1})$, $\{\tilde{\sigma}_1, \ldots, \tilde{\sigma}_{n-1}\}$, and a decomposition of $\{1, \ldots, n-1\}$, $D_n \sqcup D_n' \sqcup E_n$, such that the following conditions hold. \begin{enumerate}[(a)] \item\label{enum:basis} $\{\tilde{\sigma}_1,\ldots, \tilde{\sigma}_k\}$ is a basis of $C(X_k; \mathbb{Z})$ for any $1 \leq k < n$. \item $\partial \tilde{\sigma}_j \not = 0$ for $j \in D_n$. \item $L_{\hat{B}}$ is a bijection from $D_n$ to $D'_n$ and $\partial \tilde{\sigma}_j = \tilde{\sigma}_{L_{\hat{B}}(j)}$ for any $j \in D_n$. \item $\partial \tilde{\sigma}_i = 0$ for $i \in D'_n \sqcup E_n$. \end{enumerate} Furthermore, since the loop (INNERLOOP) also terminates at $j = n$, $L_{\hat{B}}(n)$ is not $-\infty$ and there exists $\hat{\sigma}_n \in C(X_n; \mathbb{Z})$ such that the following conditions hold: \begin{equation} \begin{aligned} C(X_n) & = C(X_{n-1}) \oplus \left< \hat{\sigma}_n \right>, \\ \partial \hat{\sigma}_n &= \hat{B}_{L_{\hat{B}}(n), n} \sigma_{L(n)} + \sum_{1 \leq i < L_{\hat{B}}(n)} \hat{B}_{ij} \sigma_i. \end{aligned} \end{equation} By the same discussion in Section~\ref{subsec:proof-algpd-basis}, there exist integers $\{\beta_i\}_{i=1}^{L_{\hat{B}}(n)-1}$ such that \begin{equation}\label{eq:p-mult} \partial \hat{\sigma}_n = \hat{B}_{L_{\hat{B}}(n), n} \tilde{\sigma}_{L_{\hat{B}}(n)} + \sum_{1 \leq i < L_{\hat{B}}(n)} \beta_i \tilde{\sigma}_i, \end{equation} holds. We can also show $\partial \sigma_{L_{\hat{B}}(n)} = 0 $ in the same manner as in the proof of the claim in Appendix~\ref{subsec:proof-algpd-basis}. Therefore, $L_{\hat{B}}(n) \in D'_n \sqcup E_n$, but since $L_{\hat{B}}(n) \not = L_{\hat{B}}(i)$ for any $1 \leq i < n$, $L_{\hat{B}}(n) \not \in D'_n$ and we have $L_{\hat{B}}(n) \in E_n$. Now let $m := L_{\hat{B}}(n) - 1$. From (a)-(d) and \eqref{eq:p-mult}, the bases of $\ker \partial_{n,m}$ and $\im \partial_{n,m}$ can be explicitly written as follows: \begin{equation} \begin{aligned} &\{\tilde{\sigma}_k + C(X_m; \mathbb{Z}) \mid m < k < n, k \in D'_n \sqcup E_n\} \\ \sqcup &\{\tilde{\sigma}_k + C(X_m; \mathbb{Z}) \mid m < k < n, k \in D_n, L_{\hat{B}}(k) < m\}, \end{aligned} \end{equation} and \begin{equation} \begin{aligned} &\{\tilde{\sigma}_{L_{\hat{B}}(k)} + C(X_m; \mathbb{Z}) \mid m < k < n, k \in D_n, m < L_{\hat{B}}(k)\} \\ \sqcup & \{ p \tilde{\sigma}_{m+1} + C(X_m; \mathbb{Z}) \}, \end{aligned} \end{equation} where $p = |\hat{B}_{m+1, n}|$. Using $m + 1 \in E_n$, finally we have \begin{equation} \begin{aligned} H(X_n, X_m; \mathbb{Z}) =& \left< [\tilde{\sigma}_{k} + C(X_m; \mathbb{Z})] \mid m < k < n, k \in D'_n, L_{\hat{B}}^{-1}(k) \leq n \right> \\ \oplus & \left< [\tilde{\sigma}_{k} + C(X_m; \mathbb{Z})] \mid m < k < n, k \in E_n\backslash\{ m + 1\} \right> \\ \oplus& \left< [\tilde{\sigma}_{m+1} + C(X_m; \mathbb{Z})] \right> \end{aligned} \end{equation} and \begin{equation} \left< [\tilde{\sigma}_{m+1} + C(X_m; \mathbb{Z})] \right> \simeq \mathbb{Z}_p. \end{equation} The proof for the ``dependent'' case is completed. Here, when the algorithm returns ``dependent'', the number $p$ is displayed. This facilitates understanding the dependency of $D(\mathbb{X}; \mathbb{Z}_{p'})$ to a prime $p'$ which is a divisor of $p$. \section{Algorithm implementation}\label{sec:implementation} The judgement algorithm is implemented in HomCloud\footnote{\url{https://www.wpi-aimr.tohoku.ac.jp/hiraoka_labo/homcloud/index.en.html}}. The twist algorithm introduced by \cite{Chen11persistenthomology} is used for faster computations. The program correctly judges the existence of the torsion for pointclouds shown in Fig.~\ref{fig:doubletripleloop} (a) and (e). \subsection{Performance benchmark} In this section, we explore the performance of Algorithm~\ref{alg:allfree}. We compare the program implemented in HomCloud and Phat \citep{phat}\footnote{\url{https://bitbucket.org/phat-code/phat/}}. The Phat code is straightforward and efficient. The input filtration for the performance comparison is an alpha filtration constructed from random 5000 and 50000 points in $\mathbb{R}^3$. Five trials were undertaken and the average computation time is shown. In Phat, we use twist-algorithm with \verb|bit_tree_pivot_column|, as recommended by \cite{phat}. The benchmark is executed on a PC with a 1.5 GHz Intel(R) Core(TM) i7-8500Y CPU, 16 GB of memory, and the Debian 10.0 operating system. Both programs run on a single core. Results are shown in Table~\ref{tab:performance}. \begin{table}[htbp] \centering \begin{tabular}{c|c|c} \hline & 5000 points & 50000 points \\ \hline Phat & 0.0282 sec & 0.446 sec \\ \hline HomCloud & 0.0323 sec & 0.550 sec \\ \hline \end{tabular} \caption{Performance benchmark results} \label{tab:performance} \end{table} According to the benchmark, our new program is c.~$\times$1.20 slower than Phat. Phat uses $\mathbb{Z}_2$ as a coefficient field and implements fast arithmetic operations by using bit-wise operations. The technique likely renders Phat faster and we conclude that the performance of our program is roughly as efficient as Phat. \section{Probability of torsion appearance}\label{sec:probability} Here we measured the probability of the appearance of torsions of random filtrations by a numerical experiment. Corollary~\ref{cor:M_0} and Corollary~\ref{cor:M_munus_1} already ensures the independence of persistence diagrams from $\Bbbk$ for a filtration embedded in $\mathbb{R}^2$. Therefore we started from filtrations in $\mathbb{R}^3$. We generated a random filtration from a pointcloud sampled from a Poisson point process in $[0, 1]^3$. The average number of points is 1000. Thus, a random number $k$ is sampled from the Poisson distribution whose parameter is 1000 and $k$ points are uniformly randomly sampled in $[0, 1]^3$. An alpha filtration was computed from the generated pointcloud and the condition was judged by HomCloud. Here, 10000 trials were carried out. Only one filtration had non-trivial torsion; thus, 9999 filtrations had trivial torsion\footnote{ Run-time errors occurred two times in these 10000 trials. When an error occurred, we disposed of the input data and retried random sampling. The cause of the errors is probably the violation of the general position condition of the randomly generated pointcloud. }. In sum, it can be stated that a filtration with non-trivial torsion is possible, but very rare. This numerical experiment suggests that there is some mathematical mechanism explaining why a random filtration with non-trivial torsion is quite rare. Exploring this further here is beyond the scope of the current paper. In contrast \cite{doi:10.1080/10586458.2018.1473821} experimentally showed that torsion subgroups often appeared in random $d$-complex $Y \sim Y_d(n, p)$, introduced by \cite{Linial*2006}. We apply our algorithm to random filtrations used in the paper. Let $\bar{Y}(n)$ be a simplex on $n$ vertices and $Y_0$ the $(d-1)$-skeleton of $\bar{Y}(n)$. $Y_k$ for $k=1, \ldots, m$ is randomly generated by adding a $d$-simplex to $Y_{k-1}$. The $d$-simplex is uniformly randomly sampled from all $d$-simplices in $\bar{Y}(n) \backslash Y_{k-1}$. We apply the algorithm to the filtration $Y_0 \subset Y_1 \subset \cdots \subset Y_m$. We used $d=2, n=75, m=5000$. The number of random flirtations was 10000. In the experiment we found that all 10000 random filtrations have non-trivial torsion. The above two experiments are contrasting. We expect that the difference comes from the dimension of the space. In the first experiment a filtration is embedded in $\mathbb{R}^3$ and in the second experiment $\bar{Y}(75)$ can be embedded in $\mathbb{R}^{74}$. The above experiments suggest that a random filtration embedded in a higher dimensional space has more non-trivial torsion subgroups in the relative homology groups than a filtration embedded in a lower dimensional space. Exploring the problem further here is also beyond the scope of the current paper. Anyway, the experiment also suggests that we do not need to be concerned about the coefficient field in most cases if the space is $\mathbb{R}^3$. Thus, if there are concerns about the field choice problem in future research contexts, our proposed algorithm would be helpful. \section{Proof of Theorem~\ref{thm:r-wasserstein}}\label{sec:convex} In this section, we assume the following conditions. \begin{cond} There exists $0 = r_0 < r_1 < \ldots < r_N < r_{N+1} = \infty$ such that $r_k \leq r < r_{k+1}$ implies $X_r = X_{r_k}$. \end{cond} This means that the filtration is assumed to be right-continuous. This condition is not essential and we can prove the theorem if the filtration is left-continuous. We assume right-continuous for expository purposes. From this condition and $H_q( \cup_t X_t) = 0$, all birth-death pairs can be written as $(r_k, r_\ell)$ for $0 \leq k < \ell \leq N$. Hence, using persistent Betti numbers, we have the following equation: \begin{equation} \begin{aligned} & \sum_{(b, d) \in D_q(\mathbb{X}; \mathbb{R})} f(d - b) \\ =& \sum_{0 \leq k < \ell \leq N}f(r_\ell - r_k)\left( \beta_{r_k}^{r_{\ell-1}}(\Bbbk) - \beta_{r_{k-1}}^{r_{\ell-1}}(\Bbbk) - \beta_{r_k}^{r_{\ell}}(\Bbbk) + \beta_{r_{k-1}}^{r_{\ell}}(\Bbbk) \right) \\ =& \sum_{0 \leq k < \ell \leq N} \beta_{r_{k}}^{r_{\ell}}(\Bbbk)( f(r_{\ell+1} - r_{k}) - f(r_{\ell+1} - r_{k+1}) - f(r_{\ell} - r_{k}) + f(r_{\ell} - r_{k+1}) ). \end{aligned}\label{eq:f_sum} \end{equation} Now we prove the following inequality: \begin{equation} f(r_{\ell+1} - r_{k}) - f(r_{\ell+1} - r_{k+1}) - f(r_{\ell} - r_{k}) + f(r_{\ell} - r_{k+1}) ) \geq 0. \label{eq:f_convex} \end{equation} In addition, if $f$ is strictly convex, the left-hand side is strictly positive. First we prove \eqref{eq:f_convex} under the condition of $r_{\ell+1} - r_{k+1} \geq r_{\ell} - r_{k}$. In this case, \begin{equation} \begin{aligned} & f(r_{\ell+1} - r_{k}) - f(r_{\ell+1} - r_{k+1}) - f(r_{\ell} - r_{k}) + f(r_{\ell} - r_{k+1}) \\ = & \left( f(r_{\ell+1} - r_{k}) - f(r_{\ell+1} - r_{k+1}) \right) - \left( f(r_{\ell} - r_{k}) - f(r_{\ell} - r_{k+1}) \right) \\ =& f'(\zeta_1) (r_{k+1} - r_{k}) - f'(\zeta_2)(r_{k+1} - r_{k}) \\ =& (r_{k+1} - r_{k})(f'(\zeta_1) - f'(\zeta_2)), \end{aligned} \end{equation} where $r_{\ell+1} - r_{k+1} \leq \zeta_1 \leq r_{\ell+1} - r_{k}$ and $r_{\ell} - r_{k+1} \leq \zeta_2 \leq r_{\ell} - r_{k}$. Here, from the assumption of $r_{\ell+1} - r_{k+1} \geq r_{\ell} - r_{k}$ and the convexity of $f$, we have $f'(\zeta_1) - f'(\zeta_2) \geq 0$ and the inequality \eqref{eq:f_convex}. The strict positivity from strict convexity is trivial. When $r_{\ell+1} - r_{k+1} \leq r_{\ell} - r_{k}$, we can prove the inequality in a similar way by exchanging the role of $f(r_{\ell+1} - r_{k+1})$ and $f(r_{\ell} - r_{k})$ in the foregoing. From the discussion in Section~\ref{sec:proof-allfree}, we have \begin{equation} \beta_{r_{k}}^{r_{\ell}}(\mathbb{R}) \geq \beta_{r_{k}}^{r_{\ell}}(\mathbb{Z}_p), \label{eq:beta_ineq_1} \end{equation} for any $p$, $k$, and $\ell$. Furthermore, if $D_q(\mathbb{X}; \mathbb{R}) \not = D_q(\mathbb{X}; \mathbb{Z}_p)$, there exists $k < \ell$ such that \begin{equation} \beta_{r_{k}}^{r_{\ell}}(\mathbb{R}) > \beta_{r_{k}}^{r_{\ell}}(\mathbb{Z}_p) \label{eq:beta_ineq_2} \end{equation} holds. From \eqref{eq:f_sum}, \eqref{eq:f_convex}, \eqref{eq:beta_ineq_1}, and \eqref{eq:beta_ineq_2}, we complete the proof of the theorem. \section{Conclusions}\label{sec:conclusion} In this paper, we focus on mathematical phenomena concerning the change of the coefficient field in persistent homology. We show that the torsion of relative homology groups $H_q(X_n, X_m; \mathbb{Z})$ plays an essential role for the phenomena. We also propose an algorithm to judge the independence of the field change. The algorithm is implemented in software, HomCloud. Using the algorithm, we show that the probability of persistence diagrams changing as a result of field changes is not zero, but very low for random pointclouds in $\mathbb{R}^3$. This suggests that we do not need to be particularly concerned about the choice of the field in most practical persistent homology contexts if we approach persistence diagrams in statistical terms. To assuage researchers' future concerns about this issue, the torsion condition can be checked by the algorithm. Of course, where torsion structures are important, such as Klein bottles or M\"obius strip, the choice of the coefficient field is important. We also suggest that the choice of the coefficient field is important for high dimensional data by the numerical experiment on $\bar{Y}(75)$. In such contexts, further study is required into the torsion on the filtrations. Further, the results herein suggest that the ``difficulty'' of computation of $D_q(\mathbb{X}; \Bbbk)$ depends on the torsion. If all torsions are zero, $D_q(\mathbb{X}; \Bbbk)$ for any $\Bbbk$ is computable by computing $D_q(\mathbb{X}; \Bbbk)$ for only one $\Bbbk$, for example, $\mathbb{Z}_2$. If not, to compute $D_q(\mathbb{X}; \Bbbk)$ for many $\Bbbk$ is more onerous. In that case, we can apply the algorithm in \cite{BJDMC} for faster computation; however, that algorithm simultaneously computes $D_q(\mathbb{X}; \Bbbk)$ for multiple, but not for all, $\Bbbk$. This phenomenon is not dissimilar to a theorem in \cite{optimal-Day}. Those authors proved that the difficulty of computing a kind of optimization problem on homology algebra depends on the existence of the non-zero torsion subgroup of the relative homology group. Integer programming on homology algebra can be solved by linear programming if the torsion-free condition holds. Integer programming requires much more time than linear programming in the sense of computational complexity theory. Of course, our paper and \cite{optimal-Day} concern different problems, but the results are similar because of the shared focus on the torsions of relative homology. These results suggest that the existence of non-trivial torsion subgroups in relative homology renders the problems of computational homology more difficult.

\section{Introduction}\label{S_intro} Let $G=(V,E)$ be a graph on $V=[n]$. A {\it tournament} on $G$ is an orientation of $E$. Intuitively, for each edge $(i,j)\in E$, teams $i$ and $j$ play a match, and then this edge is directed towards the winner. The {\it score sequence} $s=(s_1,\ldots,s_n)$ lists the number of wins by each team. Motivated by observations by Allee \cite{All38} on pecking orders in animal populations, Rapoport \cite{Rap1,Rap2,Rap3} and Landau \cite{Lan1,Lan2,Lan3} pioneered the mathematical study of dominance relations. Applications include paired comparisons, elections and sporting events, see e.g.\ \cite{Rys64,HM66,Moon68}. Landau is well-known for his characterization \cite{Lan3} of score sequences in the case that $G=K_n$ is the complete graph. Note that, in such a tournament, a single match is held between each pair of teams. \begin{thm}[Landau's Theorem]\label{T_landau} Any $s\in{\mathbb Z}^n$ is the score sequence for a tournament on $K_n$ if and only if $\sum_i s_i={n\choose2}$ and $\sum_{i\in A} s_i\ge{k\choose2}$ for all $A\subset[n]$ of size $k$. \end{thm} The necessity of these conditions is clear, since the total wins by any $k$ teams is at least the number of matches held between them. Sufficiency is less obvious; however, many proofs have since appeared, see e.g.\ \cite{BS77,T81,BK09,C14,BF15}. Ten years later, Moon \cite{M63} discovered that the same conditions are necessary and sufficient for random tournaments on $K_n$. \begin{defn} A {\it random tournament} on $G$ is a collection of real numbers $p_{ij}\in[0,1]$, $i<j$, one for each edge $(i,j)\in E$. The {\it mean score sequence} $x=(x_1,\ldots,x_n)$ of a random tournament is given by \[ x_i = \sum_{\substack{j>i \\ (i,j) \in E}} p_{ij} + \sum_{\substack{j<i \\ (i,j) \in E}} (1 - p_{ji}). \] \end{defn} Intuitively, each edge in $G$ corresponds to a match between two teams $i< j$ and the value $p_{ij}$ is the probability that team $i$ wins and $1 - p_{ij}$ is the probability that team $j$ wins. The entry $x_i$ of the mean score sequence is the expected number of wins for team $i$. \begin{thm}[Moon's Theorem]\label{T_moon} Any $x\in{\mathbb R}^n$ is the mean score sequence for a random tournament on $K_n$ if and only if $\sum_i x_i={n\choose2}$ and $\sum_{i\in A} x_i\ge{k\choose2}$ for all $A\subset[n]$ of size $k$. \end{thm} Generalizations of these results have been studied in e.g.\ \cite{H65,C14,BF15}. Our purpose is to provide short and natural proofs of Landau's and Moon's Theorems for {\it any} graph $G$ using zonotopes from convex geometry. We were led in this direction by realizing that the hyperplane description of the permutahedron, the graphic zonotope $Z_G$ when $G=K_n$ (see below for definitions), coincides with the conditions in \cref{T_landau,T_moon}. It appears that this connection has not been fully capitalized on in the literature. For instance, seen in this light, \cref{T_moon} is immediate by earlier work of Rado \cite{R52}. We think the zonotopal perspective will be useful for studying further properties of tournaments. Finally, we mention here that our arguments extend immediately to any multigraph $M$, however, as this becomes notationally cumbersome, we leave this to the interested reader. Throughout, we fix a graph $G=(V,E)$ with $V=[n]$ and $|E|=m$. \section{A zonotopal proof of Moon's Theorem} A {\it zonotope} \cite{Z95} is an affine image of a cube. In particular, the {\it graphic zonotope} $Z_G$ of the graph $G=(V,E)$ is the polytope given by the Minkowski sum \begin{equation}\label{E_ZG} Z_G = \sum_{\substack{(i,j) \in E \\ i<j}} [e_i, e_j]. \end{equation} In this case, $Z_G$ is the image of the cube $[0,1]^{m}$, where recall we let $|E|=m$. To understand the projection map, let $e_{ij}$, for $i<j$ and $(i,j) \in E$, denote a basis vector of ${\mathbb R}^{m}$ and let $e_i$ denote a basis vector of ${\mathbb R}^n$. Then the projection map $\pi: {\mathbb R}^{m} \to {\mathbb R}^n$ satisfies \begin{equation}\label{E_pi} \pi(a_{ij} e_{ij}) = a_{ij}e_i + (1- a_{ij}) e_j. \end{equation} Hence, for any $\{a_{ij}:i<j\}\in[0,1]^{m}$, the image under $\pi$ of any $\sum_{i<j}a_{ij} e_{ij}$ is the vector with $i$th coordinate $\sum_{j>i}a_{ij}+\sum_{j<i}(1-a_{ji})$. Hence Moon's Theorem essentially follows by identifying $\{a_{ij}:i<j\}$ with a random tournament. Then the projection map $\pi$ is simply the expectation operator. \begin{thm}[Generalized Moon's Theorem]\label{thm: moon's theorem} For any $ A \subset [n]$ let $\phi(A)$ be the number of edges in the subgraph $G|_A$ of $G$ induced by $A$. Then any $x = (x_1,\ldots,x_n) \in {\mathbb R}^n$ is the mean score sequence of a random tournament on $G$ if and only if $\sum_{i} x_i =m$ and $\sum_{i\in A}x_i\ge\phi(A)$ for all $A\subset [n]$. \end{thm} \begin{proof} Identify the cube $[0,1]^{m}$ with the set of random tournaments on $G$ by mapping $\{a_{ij}: i < j\}$ to the random tournament $X$ on $G$ where $p_{ij} = a_{ij}$. As discussed, it follows by \eqref{E_pi} that the image of $X$ under $\pi$ is its mean score sequence $x$. Therefore, $x \in {\mathbb R}^n$ is a mean score sequence if and only if $x \in \operatorname{image}(\pi) = Z_G$. To conclude, we use the following hyperplane description of the graphic zonotope $Z_G$, which follows from \cite{AA17}: \begin{equation}\label{E_HP} Z_G = \{ x \in {\mathbb R}^n : \sum_{i}x_i = m, \; \sum_{i \in A} x_i \geq \phi(A), \; \forall A \subset [n] \}. \end{equation} This description is obtained from the one in \cite{AA17} by realizing that $\phi(A) = \mu([n]) - \mu(A)$, where $\mu$ is the submodular function that defines $Z_G$. \end{proof} \begin{remark} In the classical case, when $G=K_n$, the conditions above coincide with $x$ being majorized, as in \cite{MOA11}, by $(0,1,\ldots,n-1)$, and \eqref{E_HP} is a result of Rado \cite{R52}. \end{remark} \section{A refinement} Next, by combining zonotopal tilings with the theory of mixed subdivisions, we obtain a refinement of \cref{thm: moon's theorem} that implies a generalization of Landau's theorem. Informally, we find that any mean score sequence $x$ can be realized by a tournament with at most a ``forest's worth of randomness.'' \begin{thm}\label{T_tree} For any $x \in Z_G$ there exists a forest $F \subset G$ and a random tournament $X$ on $G$ with mean score sequence $x$ such that for every edge $(i,j) \not \in E(F)$, we have $p_{ij} = 0$ or $1$ in the tournament $X$. Furthermore, if $x\in{\mathbb Z}^n$, then the same is true for $(i,j) \in E(F)$. Hence, in this case, $x$ is the score sequence of a (deterministic) tournament on $G$. \end{thm} We obtain the following immediately. \begin{corollary}[Generalized Landau's Theorem] Any $s = (s_1,\ldots,s_n) \in {\mathbb Z}^n$ is the score sequence of a tournament on $G$ if and only if $\sum_{i} s_i = m$ and $\sum_{i\in A}s_i\ge\phi(A)$ for all $A\subset [n]$. \end{corollary} To prove these results, we will need to recall two different types of subdivisions of polytopes. First, a {\it zonotopal subdivision} of a zonotope $P$ is a collection of zonotopes $\{P_i\}$ such $\bigcup P_i = P$ and any two zonotopes $P_i$ and $P_j$ intersect properly; i.e., $P_i$ and $P_j$ intersect at a face of both or not all, and their intersection is also in the collection $\{P_i\}$. We call the zonotopes $P_i$ the {\it tiles} of the subdivision. The following lemma is implicit in the proof of Theorem 2.2 in Stanley \cite{S90}. \begin{lemma}\label{lem: subdivision} There are vectors $v_F \in {\mathbb R}^n$ for each forest $F \subset G$ such that $\{v_F + Z_F\}$ is a zonotopal subdivision of $Z_G$. The full-dimensional tiles are those corresponding to spanning forests. Furthermore, every lattice point in $Z_G$ appears as a vertex of a zonotope $v_F + Z_F$ for some spanning forest $F$. \end{lemma} The last statement of this lemma is true since the number of lattice points contained in the half-open parallelopiped generated by a linearly independent set of vectors is given by the determinant of the matrix whose columns are vectors in the set. In the case of graphic zonotopes, each half-open paralleopiped is generated by vectors corresponding to an edge in a forest and the corresponding matrix has determinant $1$. See \cite{S90} for more details. The second type of subdivison comes from the theory of {\it mixed subdivisions}. Let $P = P_1 + \cdots + P_k$ be the Minkowski sum of polytopes. A {\it mixed cell} $\sum_{i} B_i$ is a Minkowski sum of polytopes, where the vertices of $B_i$ are contained in the vertices of $P_i$. A {\it mixed subdivision} of $P$ is a collection of mixed cells which cover $P$ and intersect properly; i.e., for any two mixed cells $\sum B_i$ and $\sum B_i'$ the polytopes $B_i$ and $B_j'$ intersect at a face of both, or not at all. In the case of graphic zonotopes, the subdivision given by \cref{lem: subdivision} is a mixed subdivision, see Lemma 9.2.10 in De Loera et al.\ \cite{DRS10}. This means that every tile $v_F + Z_F$ can be written as $\sum_{(i,j) \in E} B_{ij}$, where $B_{ij}$ is a face of the line segment $[e_i, e_j]$. The only faces of these segments are the vectors $e_i$ and $e_j$ and the entire segment $[e_i,e_j]$. Since $Z_F = \sum_{(i,j) \in E(F)} [e_i, e_j]$ this means that $E$ can be partitioned $A\cup B\cup E(F)=E$ in such a way that \begin{equation} \label{eq: mixed subdivision description} v_F + Z_F = \sum_{\substack{(i,j) \in A \\ i < j}} e_i + \sum_{\substack{(i,j) \in B \\ i < j}} e_j + \sum_{\substack{(i,j) \in E(F)\\ i<j}} [e_i, e_j]. \end{equation} With this at hand, we prove our main result. \begin{proof}[Proof of \cref{T_tree}] If $x \in Z_G$, then it is contained in one of the full-dimensional tiles of the subdivision of $Z_G$ given by \cref{lem: subdivision}. Let $F$ be a forest corresponding to one of the tiles $v_F + Z_F$ that contains $x$. Then \eqref{eq: mixed subdivision description} tells us that $x$ is of the form \[x =\sum_{\substack{(i,j) \in A \\ i < j}} e_i + \sum_{\substack{(i,j) \in B \\ i < j}} e_j + \sum_{\substack{(i,j) \in E(F)\\ i<j}} [a_{ij}e_i + (1- a_{ij})e_j]. \] for some $0 \leq a_{ij} \leq 1$. Let $X$ be the random tournament where, for $i<j$, \[ p_{ij} = \left \{ \begin{array}{ll} 0 & \text{if $(i,j) \in A$} \\ 1 & \text{if $(i,j) \in B$} \\ a_{ij} & \text{otherwise.} \end{array} \right.\] Then the mean score sequence of $X$ is $x$. If $x\in{\mathbb Z}^n$ is integer-valued, then in fact (also by \cref{lem: subdivision}) all $a_{ij}\in\{0,1\}$, in which case $X$ is a deterministic tournament. \end{proof} \begin{remark} Geometrically, this theorem states that every lattice point of $Z_G$ is the image of some vertex of the cube $[0,1]^{m}$ under the map $\pi$. \end{remark} \subsection*{Acknowledgements} We thank David Aldous, Federico Ardila, Persi Diaconis, Jim Pitman and Richard Stanley for helpful conversations. \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

INTRODUCTION {#s1} ============ Breast-conserving treatment (BCT) is the standard treatment for patients with early breast cancer, and radiotherapy is the important component of BCT. For radiotherapy of the breast, whole breast irradiation (WBI) is the basic mode \[[@RRS143C1]--[@RRS143C3]\]. Whole breast intensity-modulated radiotherapy (IMRT) increased the dose homogeneity to the treated breast, meanwhile improving normal tissue-sparing \[[@RRS143C4]\]. In addition, the long-term follow-up results showed that WBI delivered by IMRT not only decreased acute radiation dermatitis, but also decreased the late breast fibrosis, chronic pulmonary and cardiac toxicities \[[@RRS143C5]\]. As the IMRT dose distributions are highly conformal to the target, the target motions during setup and respiration may lead to a change in the dose distributions. With the development of the image-guided online and offline setup verification and correction techniques, interfraction setup error has been reduced significantly during delivery of irradiation for breast cancer patients \[[@RRS143C6], [@RRS143C7]\]. However, respiration-induced motion during radiation delivery becomes one of the main geometrical uncertainties, which may affect the treatment accuracy. So the dosimetric variances of the target resulting from the intrafraction motion become relatively more significant. For the measurement of respiration-induced motion in the breast target and the chest wall, passive markers placed on the skin (skin marker), and implanted clips placed in the lumpectomy cavity could be considered as the surrogates for the breast \[[@RRS143C8]--[@RRS143C11]\]. Several image-guided therapy techniques have been applied to breast radiotherapy, such as electronic portal image devices (EPID), cone beam computed tomography (CBCT), four-dimensional computed tomography (4DCT) combined with real-time position management system (RPM), and so on \[[@RRS143C9], [@RRS143C12], [@RRS143C13]\]. The respiration-induced motion amplitude of the breast during free breathing was generally \< 4 mm in all directions \[[@RRS143C8], [@RRS143C9]\]. Chopra *et al*. \[[@RRS143C8]\] found that during normal breathing, breast movement judged by five surface markers was 1.07, 1.94 and 1.86 mm in the mediolateral, superoinferior and anteroposterior dimensions, respectively. Qi *et al*. \[[@RRS143C9]\] examined the respiratory motion for the target between the two extreme phases, and the maximum centroid movement ranged from 1.1--3.9 mm for the treated breast. The respiration-induced motion was small, but the effect of the dose distribution resulted from respiration-induced motion within the breast tissue and the organ at risk (OAR) were inconsistently reported \[[@RRS143C9], [@RRS143C14]--[@RRS143C17]\]. On the other hand, IMRT based on three-dimensional computed tomography (3DCT) simulation is the current main mode in most countries of the world. However, the variability of the specified dosimetric parameters between 4DCT and 3DCT IMRT planning for WBI is unknown. The purpose of this study was to evaluate the respiratory motion-induced change in the specified dosimetric parameters of the irradiated breast and OARs during free breathing, in the tangential field technique with static multileaf collimator segments (SMLC) IMRT planning, and to compare the specified dosimetric changes between 4DCT and 3DCT IMRT plannings. MATERIALS AND METHODS {#s2} ===================== Patient selection and instruction {#s2a} --------------------------------- A total of 17 breast cancer patients who were prescribed to receive adjuvant whole breast radiotherapy (WBRT) after breast-conserving surgery were enrolled in this study between June 2009 and May 2011. Patients with restricted arm movement after surgery and poor pulmonary function were excluded. Of these 17 patients, 10 had right-sided breast cancer, and the remaining 7 had left-sided breast cancer. Written informed consents were obtained from all the patients, and the study was approved by the institutional research ethics board of the hospital. 3DCT and 4DCT data acquisition {#s2b} ------------------------------ The patients were immobilized in supine position on a breast board using an arm support (with both arms above the head to expose the breast adequately) and a knee support. The 3DCT and 4DCT data sets were acquired from the 17 patients on a 16-slice CT scanner (Philips Brilliance Bores CT, Netherlands) during free breathing. Three laser alignment lines were marked on the patient before CT acquisition. The 3DCT scan, in which 12 contiguous slices with a thickness of 2 mm were produced per gantry rotation (1 s) and interval (1.8 s) between rotations was acquired in sequential mode, and the 4DCT scan was acquired in helical mode with the scanning pitch between 0.09 and 0.15. The respiratory signal was recorded with the Varian real-time position management (RPM) System (Varian Medical Systems, Palo Alto, CA), by measuring the displacement of the infrared markers placed on the epigastric region of the patient\'s abdomen. GE Advantage 4D software (GE Healthcare, Waukesha, WI) sorted the reconstructed 4DCT images into 10 respiratory phases labeled as 0--90% on the basis of triggered signal. Phase 0% denoted the maximum end inspiration (EI) and phase 50% denoted the maximum end expiration (EE). The 4DCT images were reconstructed using a thickness of 2 mm and then transferred to the Eclipse treatment planning system (TPS) (Eclipse 8.6, Varian Medical Systems, Palo Alto, CA) for structure delineation and treatment planning generation. Treatment planning and dosimetric evaluation {#s2c} -------------------------------------------- The whole breast (CTV), ipsilateral lung (IPSL) and heart were delineated on the image data sets separately from 3DCT images, EE and EI phase of 4DCT images. The delineation of the CTV and OARs were done by one clinician using the same window and level setting. The CTV was delineated based on the visible glandular breast tissue seen on the CT images, taking into consideration the anatomic references, which were defined as medially at the sternal-rib junction, inferiorly at the inframammary fold, superiorly at the inferior edge of the medial head of the clavicle, and laterally at mid-axillary line typically. The anterior margin of the CTV was shrunk by 5 mm below the skin surface, and the posterior margin was the junction of the breast tissue and the pectoralis muscles. The planning target volume (PTV) was generated using a 5-mm margin around the CTV and shrunk by 5 mm below skin surface. Treatment planning was established at the 3DCT images, using the tangential field technique with static multileaf collimator segments (SMLC) IMRT, with two parallel opposed tangential fields. To reduce the lung volume within the treatment field and the volumes of hot spots in the treatment field, 2--5 segmented fields were set up in each direction. The prescription dose was 50 Gy in 25 fractions (2 Gy per fraction) to the PTV using 6 MV photon beams, which was defined as the 90% isodose line. The criteria of the SMLC-IMRT planning was to ensure at least 95% of the PTV volume received the prescription dose. The segmented MLCs were manipulated to shield the areas of PTV receiving a dose \>103% of the isodose line, and to keep the dose delivered to OARs such as the IPSL and heart within normally accepted tolerances. The treatment planning designing based on the 3DCT images was copied and applied to the EI and EE phase images with the same gantry angles, collimator angles, primary field size, monitor units delivered per beam, and so on. The dose distribution was calculated separately in all the three treatment plannings, and dose--volume histogram (DVH) parameters for the CTV, PTV, IPSL and heart were calculated for each treatment planning in all patients. The CTVs were evaluated on the basis of the volumes receiving 50 Gy and 55 Gy. The parameters such as mean dose (D~mean~), homogeneity index (HI), and conformal index (CI) were evaluated in PTV. HI was defined as in which *D*~*2*~ and *D*~*98*~ represent the dose covered 2% and 98% of the target volume, and *D*~*T*~ is the prescription dose \[[@RRS143C18]\]. CI is defined as where *PTV*~*ref*~ represents the volume of PTV that is covered by prescription dose, *V*~*PTV*~ is defined as the planning target volume and *V*~*ref*~ represents the volume enclosed by the prescribed isodose \[[@RRS143C19], [@RRS143C20]\]. OARs were evaluated using D~mean~ and the volumes receiving ≥ 5, 10, 20, 30, 40, or 50 Gy (V~5~, V~10~, V~20~, V~30~, V~40~, V~50~). Statistical analysis {#s2d} -------------------- Statistical analysis was performed with the SPSS statistical analysis software package. The Wilcoxon Signed Ranks Test was used for each dosimetric parameter. Data were regarded as statistically significant at *P* \< 0.05. RESULTS {#s3} ======= The mean CTV volume based on EI, EE and 3DCT images were 680.05 ± 251.55 cm^3^, 678.88 ± 256.28 cm^3^ and 668.12 ± 241.93 cm^3^, respectively. The mean volume of the maximum volumetric difference among the three CTVs of these patients was only 11.93 ± 28.64 cm^3^ (1.42 ± 3.79 %). The mean percentage volume of CTV receiving 50 Gy was 98 ± 1.15%, 98.49 ± 1.07% and 98.76 ± 0.66% for EI, EE and 3DCT treatment planning, respectively. The percentage volume of CTV (mean ± SD) receiving 55 Gy in the EI, EE and 3DCT treatment planning was 7.90 ± 4.19%, 7.83 ± 4.18%, and 7.78% ± 4.02%, respectively. For the SMLC-IMRT technique, only a minor difference was observed: the percentage differences in the volume of CTV receiving 50 Gy and 55 Gy among different scans were all \< 0.8%. The mean PTV doses were 5303.95 cGy (SD: 32.36) and 5306.69 cGy (SD: 34.89) for the EI phase planning and EE phase planning, respectively (*Z* =--0.26, *P* = 0.795). However, the D~mean~ at 3DCT planning for PTV was 5314.12 cGy (SD: 27.97), and compared with the EI and EE phases, significant differences were found (*Z* =--2.58,--2.34; *P* = 0.010, 0.019). For the PTV, the variations in CI and HI for the 17 patients were shown in Fig. [1](#RRS143F1){ref-type="fig"}. The HI was 0.139 ± 0.02 for the EI phase planning and 0.141 ± 0.02 for the EE phase planning (*Z* =--0.02, *P* = 0.981). The CI was 0.716 ± 0.07 for the EI phase planning, and 0.712 ± 0.06 for the EE phase planning (*Z* =--0.17, *P* = 0.287). There was no significant difference between the EE and EI phase plannings for either CI or HI. Significant differences did exist for the HI between the 3DCT planning (0.127 ± 0.01) and the EI and EE phase plannings (*Z* =--3.34,--2.77; *P* = 0.001, 0.006). The CI increased significantly more in 3DCT scan planning (0.74 ± 0.07) compared with the EI and EE phase plannings (*Z* =--2.817,--2.817; *P* = 0.005, 0.005). Fig. 1.The homogeneity index (HI) and conformal index (CI) for the 17 patients in the three different treatment plannings. Table [1](#RRS143TB1){ref-type="table"} showed the variation of the specified dose and volume parameters for the ipsilateral lung and heart for the three treatment plannings. A comparison of these parameters for the treatment planning based on the three different CT images is detailed in Table [2](#RRS143TB2){ref-type="table"}. In our observation, during free respiration cycle, changes in the lung volume followed respiratory patterns and the smaller lung volume was most likely to be associated with a lower dose to the IPSL. For the percent volume receiving a high dose (V~30~, V~40~ and V~50~), significant differences were observed between EE phase and EI phase plannings. Average lung volume varied by up to 8.72%, which suggested that the IPSL volume change during respiration might needed to be considered. For the 7 left-sided breast cancer patients, no differences in the percentage volume receiving the specified dose and D~mean~ were observed between the EI, EE and 3DCT plannings, and the heart volume varied by only ≤ 2.68% during the whole respiration cycle. Table 1.Dosimetric and volume parameters of the ipsilateral lung and heart for each CT scan planningV~5~ (%)V~10~ (%)V~20~ (%)V~30~ (%)V~40~ (%)V~50~ (%)D~mean~ cGyV cm^3^IPSL EI31.45 ± 4.3822.60 ± 4.5517.97 ± 4.6215.47 ± 4.5512.67 ± 4.463.69 ± 3.021016.39 ± 211.351204.09 ± 275.45 EE31.39 ± 4.6322.40 ± 4.6517.56 ± 4.7314.98 ± 4.6912.13 ± 4.563.15 ± 3.08997.60 ± 217.721099.14 ± 280.12 3D31.39 ± 4.5522.50 ± 4.6918.04 ± 4.4715.30 ± 4.6912.45 ± 4.663.54 ± 2.981010.96 ± 219.671178.79 ± 282.13 HT EI24.08 ± 7.1317.74 ± 7.2514.59 ± 6.2212.42 ± 5.3210.44 ± 5.824.92 ± 3.04858.57 ± 285.89518.37 ± 99.49 EE25.02 ± 7.1218.44 ± 7.1915.16 ± 6.2412.91 ± 5.3910.86 ± 4.945.11 ± 3.22887.74 ± 287.67505.97 ± 99.23 3D24.29 ± 6.5717.82 ± 6.8014.69 ± 6.0012.53 ± 5.2710.53 ± 4.875.03 ± 3.31864.21 ± 276.97504.49 ± 90.33[^1] Table 2.Dose and volume evaluation of the ipsilateral lung and heart for three different planningsParametersEI-EEEI-3DEE-3Dmean*ZP* valuemean*ZP* valuemean*ZP* valueIPSL V~5~ (%)--0.04 ± 0.90−0.810.421−0.07 ± 1.21--1.420.8870.11 ± 1.24--0.260.795 V~10~ (%)0.20 ± 0.750.6390.5230.11 ± 1.14--1.870.061--0.09 ± 1.26--0.240.813 V~20~ (%)0.41 ± 0.69--1.870.061--0.07 ± 0.89--0.240.813--0.48 ± 0.95--1.870.061 V~30~ (%)0.49 ± 0.66--2.530.0110.17 ± 1.030.540.586--0.32 ± 1.11--1.540.124 V~40~ (%)0.54 ± 0.67--2.720.0060.22 ± 0.94--1.320.187--0.32 ± 1.05--1.350.177 V~50~ (%)0.54 ± 0.48--3.360.0010.15 ± 0.49--0.190.052--0.39 ± 0.55--2.490.013D~mean~ (cGy)--18.78 ± 32.50--1.870.0425.43 ± 49.90--0.1420.887--13.35 ± 54.271.300.193 V (cm^3^)104.95 ± 95.644.530.00025.29 ± 107.350.970.346--79.65 ± 50.24--6.540.000HT V~5~ (%)--1.04 ± 1.38--1.520.128--0.21 ± 0.9610.6120.82 ± 1.091.520.128 V~10~ (%)--0.70 ± 1.18--1.180.237--0.07 ± 0.88--0.170.8660.63 ± 1.11--1.520.128 V~20~ (%)--0.57 ± 1.03--1.180.237--0.10 ± 0.73--0.340.7350.46 ± 0.99--1.520.128 V~30~ (%)--0.50 ± 0.94--1.180.237--0.11 ± 0.67--0.340.7350.39 ± 0.92--1.350.176 V~40~ (%)--0.42 ± 0.83--1.180.237--0.09 ± 0.66--0.420.6720.33 ± 0.88--1.350.176 V~50~ (%)--0.19 ± 0.60--1.010.672--0.11 ± 0.58--0.420.672--0.08 ± 0.62--0.340.735D~mean~ (cGy)--29.17 ± 50.00--1.180.237--5.64 ± 37.47--0.340.73523.53 ± 48.26--1.350.176 V (cm^3^)12.40 ± 29.071.230.30213.89 ± 26.281.400.2121.49 ± 13.100.300.774 DISCUSSION {#s4} ========== For whole breast IMRT, the dose distribution was highly conformal to the target in the treatment planning, therefore, part of the target could move out of the treatment field due to patient breathing. Volume variance of the breast treated is one of the geometrical uncertainties affecting treatment. In addition, a small breathing motion may lead to a large change in the dose distribution. There was a small difference between the EI, EE and 3DCT scans in terms of the breast volume, and the mean volume variation of the maximum volumetric difference was generally within 1.42 ± 3.79%. Previous study has suggested that the intraobserver variability in breast target volume delineation decreases according to a standard contouring protocol \[[@RRS143C21]\]. In our study, all the delineations were performed by one radiation oncologist referring to the same detailed delineation criteria. Breast volume did not show a significant change throughout free respiration, which indicated that the breast deformation during respiration could be ignored for WBI. Qi *et al*. \[[@RRS143C9]\] reported that, for a series of 18 patients, during normal breathing, the dosimetric impact of respiratory motion was clinically insignificant, with the exception of internal mammary nodes. In our investigation, although there was no significant difference between EE and EI plannings for the D~mean~ of PTV, significant differences were found between 3DCT planning and EI, and between 3DCT and EE phase plannings. These differences may derive from the variations in the treatment margin, the technique adopted in our study and the non-linear trajectories due to hysteresis during CT datasets acquisition. For the PTV, a 5-mm margin in all directions was added to the CTV and the whole breast treated with SMLC-IMRT. Whereas Qi *et al*. used conventional 3D planning with tangential beams for the whole breast and boost to the lumpectomy bed PTV. We also demonstrated statistically significant differences between 3DCT planning and EI/EE plannings in CI and HI for PTV. During treatment, the decrease of the target dose homogeneity significantly increased acute skin toxicity \[[@RRS143C22]\]. In addition, compared with conventional radiotherapy, hypofractionated whole breast irradiation (HF-WBI) which increasing the prescription dose per fraction delivered has been investigated recently \[[@RRS143C23]--[@RRS143C25]\]. The dose inhomogeneities and inconformities in irradiated volumes strongly influence radiation-induced toxicities, hence the target position during treatment with HF-WBI has to be determined with high precision. Some studies have found that segmented and wedged IMRT planning were not sensitive to breathing motion, which suggested that 4D treatment planning was not required for daily clinical practice in postoperative segmented or wedged radiotherapy after breast-conserving surgery \[[@RRS143C15]--[@RRS143C17]\]. We explored the different dosimetric impact of treatment planning based on conventional 3DCT scans and the two extreme phases of 4DCT scans. In our study, we have verified that respiratory motion-induced CTV volume variance is minimal; in addition, for the mean PTV dose, a significant difference between the 3DCT scans and both the two extreme scans (EE and EI) was found. This might be induced by the fact that the conventional scans were composed of all phases of the respiratory cycle collected in the fixed interval. Another possible explanation was, in spite of the good compliance and active participation of all patients, the different scanning methods and the target envelope variance during the respiratory cycle might be considered to be the significant relevant factors. For the HI and CI, the differences between the 3DCT planning and EI/EE phase plannings also supported our conclusions that during free breathing dose distributions and target coverage during the intrafraction treatment might be affected by the scanning mode of the simulation CT and the choice of the treatment planning CT. There were significant differences in statistics for some of the PTV parameters between 3DCT and EI, and between 3DCT and EE, but the small absolute differences indicated that the conventional 3D treatment planning using the tangential field technique with SMLC was sufficient for daily radiotherapy, but could decrease the precision of IMRT. The use of IMRT for whole breast treatment can minimize the volume of lung and heart being irradiated and decrease the acute and late radiation reaction. The volume, V~30~, V~40~, V~50~ and D~mean~ for IPSL in EI phase were all higher than in EE phase, which showed that although the target movement was small during free breathing, the thorax expansion was significant between the extreme respiratory phases, and thus induced the dose variation for the ipsilateral lung. The respiratory motion-induced dose variations for heart were not statistically significant, and the absolute difference was rather small. This is, in part, due to the intrinsic cardiac contractions. Qi *et al*. \[[@RRS143C9]\] showed the heart normal tissue complication probability (NTCP) was \< 10% for the patients calculated, with approximately 1% of NTCP variations respiratory motion-induced. During free breathing, the dosimetric impact of respiratory motion is clinically insignificant for heart. It should be noted that these findings may not necessarily be extrapolated to all patients due to the minimal number of patients enrolled in this study. CONCLUSION {#s5} ========== Conventional 3D-treatment planning using the tangential field technique with SMLC is sufficient for daily radiotherapy, because 3DCT images are taken disregarding the respiratory cycle during free breathing. The variance of the ipsilateral lung volume receiving high dose irradiation and mean lung dose (MLD) between the EE and EI phase plannings showed that 4D-treatment planning with consideration of breathing motion, breath-hold or respiratory gate would achieve a high-precision treatment delivery, especially for the patients with an irregular breathing pattern. FUNDING {#s6} ======= This work was supported by the National Natural Science Foundation of China (Grant No. 30870742). [^1]: IPSL = ipsilateral lung, HT = heart, EI = end inspiration, EE = end expiration.

Background {#Sec1} ========== In a network aimed at the regionalization of perinatal care, high-risk pregnancies should be transferred "in utero" in order to minimize risks to both the mother and the neonate \[[@CR1], [@CR2]\]. However, it is not always feasible to predict and prevent all the conditions possibly requiring neonatal care that cannot be provided in the referral center and there will always be a number of neonates who undergo acute inter-facility transport \[[@CR3]\]. Neonatal transport represents an additional risk factor for a critically ill patient, especially for very preterm infants \[[@CR1], [@CR4]\]. Therefore, the presence of a well-organized Neonatal Emergency Transport Service (NETS) is mandatory for a perinatal regional network, as it represents the link between birth centers and neonatal intensive care units (NICUs) and can reduce risks of transportation, especially for very preterm infants \[[@CR5], [@CR6]\]. In the past years, a program of regionalization of perinatal care was implemented and NETS have been progressively activated in Italy \[[@CR7], [@CR8]\]. Regional governments in Italy have the autonomy to legislate issues regarding healthcare, thus resulting in regional variations in resources and models of organizations with differences in health care quality and outcomes, as measured by several indicators and reported by the Organisation for Economic Co-operation and Development (OECD) and the European Observatory on Health Systems and Policies \[[@CR9], [@CR10]\]. The aim of this study was to describe neonatal transfer practices and to assess the current organization of NETS in Italy. For this purpose, in 2015, a national survey was carried out by the neonatal transport study group of the Italian Society of Neonatology (SIN) under the auspices of the Italian Ministry of Health. Methods {#Sec2} ======= Survey {#Sec3} ------ A survey regarding NETS activity was conducted in 2015, at all 44 existing Italian NETS. An exploratory, descriptive design, including a survey questionnaire, was adopted to maximize sample size and facilitate data collection. A multiple-choice questionnaire consisting of 20 questions regarding NETS organization and activity data for the previous year (2014) was designed by the SIN neonatal transport study group. The aim of the questionnaire was mainly to identify and describe NETS status, organization, coverage and activity; therefore, questions regarding outcomes of the transferred newborns were not included in this survey. The questionnaire included questions about: a) annual volume of NETS activity, including the number of primary and back transports, number of transported newborns ≤28 week gestational age (GA), number of transports of infants over 28 days of life or 44 weeks of corrected GA for preterm infants, and average time of each transport; b) type of organization (i.e., number of unit-based teams, dedicated teams, and free-standing independent transport services); c) policies for quality evaluation, training and education d) vehicles used, including the availability of a helicopter for air transports. The survey received approval from the SIN institutional review committee. The web-based survey tool Survey Monkey (http://[www.surveymonkey.com/](http://www.surveymonkey.com/)) was used and emails were sent to the person in charge of the NETS at each institution asking them to participate in the survey and to fill in the online questionnaire. Filling in the questionnaire implied consent to participate. Filled in questionnaires were checked for invalid responses or missing data. Requests for missing information were made to the NETS directors by the secretary of the SIN Neonatal Transport study group by telephone. When all the questionnaires were ready the data were transferred to an electronic database and evaluated by the SIN neonatal transport study group. Demographic data were obtained from governmental sources (ISTAT, Istituto Nazionale di Statistica - National Statistical Institute, available at <http://www.istat.it/>). Data about the number of birth centers and the number of birth centers with less than 500 births/year operating in 2014 were obtained, for each region, by the document of the Italian Ministry of Health: "Attuazione delle azioni previste dall'accordo del 16 dicembre 2010. Linee di indirizzo per la promozione e miglioramento della qualità, della sicurezza e dell'appropriatezza degli interventi assistenziali nel percorso nascita e per la riduzione del taglio cesareo. Monitoraggio al 31 dicembre 2014", with the exception of Lazio and Campania, where data obtained from the available regional reports have been used \[[@CR11]--[@CR13]\]. Italian administrative organization {#Sec4} ----------------------------------- Italy is subdivided into 20 regions, five of which have a special autonomous status. The Italian State has run a universal public national healthcare system since 1978, which is managed by the Ministry of Health and administered on a devolved regional basis. Each regional government is responsible for organizing its own healthcare system. The results were grouped by region to allow a comparison among regions and not among individual NETS. Statistical analysis {#Sec5} -------------------- Statistical analysis consisted of descriptive statistics and the evaluation of categorical variables; continuous variables are presented as mean values ± standard deviation (SD), whereas they are presented as absolute and relative frequencies where needed. Data are provided both as absolute numbers and relative frequencies. The χ^2^ test or Fisher's exact test were used to compare differences in categorical variables between groups. A *p*-value less than 0.05 was considered statistically significant, and all *p*-values were based on two-tailed tests. Statistical analysis was performed using SPSS for Windows (SPSS, Inc., Chicago, IL). Results {#Sec6} ======= The response rate for the questionnaire was 100% (44/44). After a single request for missing data the 44 questionnaires were fully filled in, thus all of them were included in the analysis. The results, including demographic data for each region, are reported in Table [1](#Tab1){ref-type="table"}.Table 1Summary of the nationwide survey results on Italian NETS. All data refers to the year 2014RegionOverall BirthsInhabitants at 31st December 2014Surface (km^2^)30-day mortality rate ‰NICUsNICUs with \> 90 VLBW (\<  1500 g) admissions per yearNETSBirths per NICUBirths per NETSNETS per km^2^Maternity wardsMaternity wards with ≤ 500 births per yearRatio Maternity wards per NETSTotal Neonatal transportsPrimary Neonatal transportsNeonatal Back-transportsNeonatal transports of infants with GA ≤ 28 weeksNTI (%)NTI for Primary transports (%)Average time of transports (minutes)Type of Ambulance (Dedicated-not Dedicated)NETS Helicopter availabilityPiemonte34,6374,406,86025,3872.57174948494836262974.143633144981.040.901255--2NoValle d'Aosta1119127,39032602.60--0------10------------------Lombardia86,23910,004,79423,8632.71721150727839216970106.3693285181441.080.981054--7YesTrentino-Alto Adige10,3791,058,50513,6063.52025189518968031376.501731314211.661.261670--2YesVeneto40,6294,916,81518,4072.5722580420,314920339919.50424283141561.040.691480--2YesFriuli-Venezia Giulia91771,060,00082401.62024588458841201125.501571213621.711.31902--0YesLiguria10,7491,573,83754162.611110,74910,7495416110112431647992.261.521501--0YesEmilia-Romagna36,6684,446,58422,4523.5802458318,33411,22629814.5036261020.090.07800--2NoToscana29,1183,745,98322,9873.26134853970676622558.3336335761701.241.221021--2YesUmbria7015891,84884643.32003507----117------------0----Marche12,3631,544,23794012.410112,36312,3639401141142127613651.710.611900--1YesLazio50,3605,886,97717,2073.651210,07225,180860344122213311316151342.642.611202--0YesAbruzzo10,5341,327,17110,8314.410010,534----123------------------Molise2213312,48444604.01012213221344603232623301.171.03600--1NoCampania51,2435,852,72913,6704.41003512417,0814556672022.331597158314583.113.081183--0NoPuglia33,1914,079,27819,5404.9702474116,595977035617.50157157040.470.47652--0YesBasilicata4123574,25110,0734.51014123412310,0736261614200.380.011201--0YesCalabria16,4901,972,14915,2215.32008245----151------------------Sicilia44,8765,077,48725,8324.5604747911,219645856171435733621290.790.74952--2YesSardegna11,4731,658,64924,1003.43003824----179------------------ITALY502,59660,518,028302,4263.489844564711,6886903508128\ (25.2%)11.5563875752635522 (8.17%)1.281.14123 ± 8023--2133--11 Regional coverage {#Sec7} ----------------- Results showed that in 2014, among the 20 Italian regions only 12 were fully covered by NETS (Piemonte, Lombardia, Trentino-Alto Adige, Veneto, Friuli-Venezia Giulia, Liguria, Toscana, Marche, Lazio, Molise, Campania, and Basilicata), 3 regions (Emilia-Romagna, Puglia e Sicilia) were partially covered, while neonatal transport was not available in 5 regions (Valle d'Aosta, Umbria, Abruzzo, Calabria e Sardegna) (Fig. [1](#Fig1){ref-type="fig"}, panel a).Fig. 1Geographical distribution and amount of activity of the 44 Italian NETS. Panel **a.** Geographical distribution of NETS in Italy, in 2014. Regional borders are shown. White regions were fully covered by NETS (Piemonte, Lombardia, Trentino-Alto Adige, Veneto, Friuli-Venezia Giulia, Liguria, Toscana, Marche, Lazio, Molise, Campania, and Basilicata); grey regions were partially covered (Emilia-Romagna, Puglia e Sicilia); NETS was not available in black regions (Valle d'Aosta, Umbria, Abruzzo, Calabria e Sardegna). Panel **b.** Number of transports per year of the 44 NETS, in 2014. All the three NETS organized on a dedicated model carried out more than 200 transports per year NETS organization {#Sec8} ----------------- The results of the survey showed how all 44 NETS guaranteed 24/7 service coverage. 41/44 NETS were organized as an on-call service, while three were fully dedicated services. Each of the forty-one on-call NETS was linked to a NICU which provided a senior neonatologist and a neonatal nurse for each transport. NETS activity {#Sec9} ------------- Overall, in 2014, the 44 NETS in Italy transported a total of 6387 infants, 522 (8.17%) of whom were of GA \< 28 weeks, and 635 (9.94%) who were back-transported infants. Median regional transport time was 123 min (minimum-maximum range 60--190). Among the 41 on-call NETS, 3 carried out fewer than 20 transports per year, 7 carried out between 21 and 40, 10 between 41 and 80, 6 between 81 and 100, a further 10 between 101 and 200 and only 5 on-call NETS carried out more than 200 transports per year. All the 3 dedicated services carried out more than 200 transports per year (Fig. [1](#Fig1){ref-type="fig"}, panel b). The number of total transports and back-transports varied widely within the Italian territory (Fig. [2](#Fig2){ref-type="fig"}).Fig. 2Distribution of transports in Italy. Total number of transports within various Italian NETS in 2014; note that the four Turin NETS were grouped. Panel **a**: total number of transports. Panel **b**: total number back-transports 21/ 44 NETS also provided transportation for infants and children older than 28 days of life or 44 weeks of corrected GA for preterm. These patients were transferred both from the same hospitals where the services were based (9% from the NICUs, 23% from other units) and from other hospitals (68% of cases). T Most of them weighed less than 6 kg (67%) and in 48% of cases were infants below 3 months of age. However, 14% of these transports involved children weighing more than 10 kg and who were over 1 year of age. The neonatal transport index (NTI) i.e., the number of neonates transferred per 100 live births (back-transports included or excluded), is reported in Table [1](#Tab1){ref-type="table"}. We performed a broad statistical analysis and found that the comparison between the maternity wards per NETS ratio vs NTI for primary transports and the comparison between the rate of hospitals with ≤500 births per year vs 30-day mortality was significant (*p*-value \< 0.00001 and 0.00002, respectively) while the comparison between the number of neonatal transports vs 30-day mortality was slightly significant (p-value 0.007). No other statistically significant results were observed. Vehicles and air transport {#Sec10} -------------------------- Dedicated ambulances for neonatal transport were owned by 23/44 NETS (52.2%). Fifteen of the 21 NETS based at a level III NICU, where a dedicated ambulance for neonatal transport was not available, reported that they were unable to purchase one owing to financial constraints. Air transport by helicopter was not available for all NETS, as it was only carried out in 10 of the 15 regions covered by NETS. Fixed wing air transport was not available on a routine basis all over Italy. However, the Italian Air Force allowed NETS to use their aircraft for urgent cases, mainly for transfers of neonates with congenital heart defects or surgical emergencies.. This occurred especially for transports from the islands of Sicilia and Sardegna towards cities with hospitals providing pediatric/neonatal surgery, cardiothoracic surgery or other special care. Quality evaluation, training and education {#Sec11} ------------------------------------------ A dedicated NETS database was available in 42/44 services; specific guidelines were edited by 43/44 NETS and in one case by the "112 Emergency Service" (118 at the time of the Survey), (i.e., 911 in the USA and 999 in the UK). Regular auditing was performed by 36/44 NETS; in two services through an agreement with the "112 Emergency Services". No auditing was performed on a regular basis by the remaining 8 NETS. Training and education activities were provided at 40/44 NETS, at 2 NICUs linked to NETS, and at one 112 Emergency Service. Discussion {#Sec12} ========== Organized NETS activity in Italy became available during the eighties \[[@CR14]\], but it was during the nineties that NETS coverage mainly spread throughout Italy, with 10 of the 20 Italian regions reaching regional coverage \> 50% in 1999 \[[@CR9]\]. The agreement of the State-Regions Conference in December 2010 (also known as "birth path" or "percorso nascita") provided guidance to standardize pregnancy and childbirth practices at regional level, including the adoption of perinatal networks based on the "hub and spoke" organizational model that guarantee the presence of Maternal and Neonatal Emergency Transport Services \[[@CR15]\]. However, our study shows how in 2014 neonatal transport was still not available in an organized form in five Italian regions, though a NETS is not strictly required in Valle d'Aosta since there is only one perinatal center in that region. The NTI is considered an indicator of the quality of regionalization of perinatal care \[[@CR16]\]. In countries where perinatal care is highly regionalized, like in the UK, the NTI can be as low as 1% \[[@CR17], [@CR18]\]. Whereas in the past, in areas with heterogeneous distribution of obstetrics units, the NTI reached values of about 10%, e.g., in the Loire-Atlantique region in France \[[@CR18]\] or in Portugal \[[@CR19]\]. Our survey showed that regions which were fully covered by NETS, in 2014, had a mean NTI for primary transports (back-transports excluded) of 1.27 ± 0.84 and that the three other regions with partial NETS coverage had a mean NTI for primary transports of 0.43 ± 0.34. The two regions with the highest NTI for primary transports were Campania (NTI 3.08) and Lazio (NTI 2.61), both regions with full NETS coverage. This result can be explained both by the high number of birth centers that are currently active in these regions (67 in Campania and 44 in Lazio) and by the number of maternity wards carrying out ≤500 births per year (20 in Campania and 12 in Lazio, i.e., 29.9 and 27.3%, of birth centers, respectively). In 2014, in Italy, 25.2% of birth centers had an activity of less than 500 births per year, this data increases to 35.7% if we consider only the five regions without NETS coverage (Table [1](#Tab1){ref-type="table"}). The results of this survey also showed that most NETS (59.1%) carried out fewer than 100 transports per year (Fig. [1](#Fig1){ref-type="fig"}, panel b), which is a relatively low degree of activity when trying to provide its personnel with an adequate level of skills and experience and good cost performance \[[@CR20]\]. Moreover, half of the level III perinatal centers did not have a specially equipped ambulance for neonatal transport (23/21, dedicated/non-dedicated), thus leading to difficulties in providing adequate transport of these vulnerable patients. Non-dedicated ambulances are usually not equipped with dedicated infant incubators, therefore, neonatal transport is usually carried out by simply placing an infant incubator on a stretcher designed for transporting adults \[[@CR21]--[@CR23]\]. It is well known that a lack of adequate equipment and low skill levels of the personnel involved in the transport increases the risk of serious adverse events in the transport of severely ill newborns \[[@CR6]\]. Dedicated mechanical neonatal respirators, resuscitation equipments, standard vital-sign monitors, as well asdedicated twin-newborn \[[@CR24], [@CR25]\] and iNO \[[@CR26]\] devices and are nowadays available and can improve safety and quality of care during transport. The use of protocols, quality evaluation procedures and training and education programmes became widely adopted among NETS, with the exception of regular audits, which were not performed in 20% of NETS. This survey demonstrated that there is no agreement regarding the limits to be applied to age and weight for neonatal transport, i.e., if the transport must be limited to the first 30 days of life, or if a broader approach is possible. We report that as much as 10 kg of weight or 1 year of age was the upper limits of transported patients, thus changing neonatal transport into pediatric transport. This study has however some limitations that have to be pointed out. Collection and analysis of survey data have been laborious and required a considerable amount of time. NETS status is very recently changed in some Italian Regions. Abruzzo, Umbria and Calabria have activated their on-call NETS. Meanwhile for Sardegna NETS activations is foreseen soon, having already obtained deliberations from their regional governments \[[@CR8]\]. The reported missing data together with the recent changes of NETS presence in some regions could represent several of the possible avenues for future researches. Conclusion {#Sec13} ========== The inter-facility transport of severely ill and premature newborns is commonly performed in Italy by organized, generally well-equipped and well-trained NETS. The implementation of NETS, together with programs of regionalization of perinatal care have likely played a major role in reducing the neonatal mortality rate that has been observed in Italy in the last two decades (from 5.2‰ in 1998 to 3.4‰ in 2014). However, NETS coverage is still lacking in many densely populated areas and, where it is present, most of the NETS seem to have a volume of activity that is not sufficient to provide the personnel with an adequate level of skills and an appropriate cost-efficiency ratio. Moreover, many NETS report that, mainly due to financial constraints, there is still a lack of vehicles which are fully dedicated to neonatal transport, that would be of help to improve safety and quality of services. Differences found in neonatal mortality rates, NTI, NETS availability, and NETS costs among regions seems to reflect the presence of different regional perinatal systems of organizations related to the autonomy of regional governments to legislate in matter of healthcare. It is envisaged that the full application of the State-Regions Conference agreement, along with the implementation of minimal standards of care and education programs driven by the national scientific societies, would lead toward the harmonization of perinatal care in Italy. These measures, however, constitute only a part of the continuing efforts required to improve outcomes and reduce the regional differences in the availability and quality of perinatal services. Periodical national and local audits are needed to evaluate and drive these quality improvement processes. GA : Gestational Age ISTAT : Italian National Statistical Institute NETS : Neonatal Emergency Transport Services NICU : Neonatal Intensive Care Unit NTI : Neonatal Transport Index OECD : Organisation for Economic Co-operation and Development SD : Standard deviation SIN : Italian Society of Neonatology VLBW : Very low birth weight Thanks are due to all the NETS Directors who filled in the survey questionnaire; the following list refers to the name of the director and hospital. S. Schettini (H. San Carlo, Potenza), F. Raimondi (H. Federico II, Naples), M. Carpentieri (H. Santobono, Naples), M. Panico (H. San Sebastiano, Caserta), S. Demarini (H. Burlo Garofalo, Trieste), L. Cattarossi (H. Santa Maria della Misericordia, Udine), F. Ferrari (H. Modena, Modena), G. Gargano (H. Santa Maria Nuova, Reggio Emilia), M. Gente (Policlinico Umberto I, Rome), A. Dotta (H. Bambino Gesù, Rome), C. Bellini (Gaslini Institute, Genoa), F. Mosca (H. Cà Granda Policlinico, Milan), S. Martinelli (H. Niguarda, Milan), E.P. Villani (H. Carlo Poma, Mantua), L. Bernardo (H. Macedonio Melloni, Milan), P. Tagliabue (H. San Gerarado, Monza), M. Stronati (Policlinico San Matteo, Pavia), M. Barbarini (H. Sant'Anna, Como), R. Bellù (H. Lecco, Lecco), M. Agosti (H. Varese, Varese), G. Chirico (Spedali Civili Brescia, Brescia), G. Mangili (H. Papa Giovanni XXIII, Bergamo), C. Flumini (H. Riuniti Ancona, Ancona), V. Santillo (H. Cardarelli, Campobasso), F. Ferrero (H. Maggiore della Carità, Novara), D. Farina, M. Vivalda, and E. Bertino (H. Sant'Anna, Turin), G. Guala (H. Maria Vittoria, Turin), A. Marra (H. Moncalieri, Moncalieri, Turin), P. Gancia (H. Croce e Carle, Cuneo), D. Gazzolo (H. Arrigo, Alessandria), R. Magaldi (H. Riuniti Foggia, Foggia), A. Gatta and M. Bisceglia (H. Casa Sollievo e Sofferenza, San Giovanni Rotondo, Foggia), F. Tumminelli (ASL4, Enna), I. Barbieri and A. Arco (H. Martino, Messina), G. Sulliotti (H. Cervello, Palermo), C. Cacace (H. Barone Romeo, Patti, Messina), M. Moroni (H. Meyer, Florence), B. Tomassini (H. Santa Maria alle Scotte, Siena), A. Boldrini and P. Biver (H. Pisa, Pisa), H. Messner (H. Bolzano, Bolzano), E. Baraldi and D. Trevisanuto (H. Padova, Padua), and E.M. Padovani (Policlinico Rossi, Verona). Funding {#FPar1} ======= None. Availability of data and materials {#FPar2} ================================== The datasets used and/or analyzed during the current study data are available from the Neonatal Transport Study Group of the Italian Society of Neonatology (SIN) on reasonable request. Some preliminary limited data of this survey have been already partially published in: Gente M, Fedeli T. STEN. I risultati della Survey nazionale SIN. SIN INFORMA - Magazine della Società Italiana di Neonatologia. 2016; 32:1--2. (access to this newsletter is restricted to SIN members only); Bellini C, Gente M. STEN in Italia, situazione Attuale. In: "Organizzazione del sistema di trasporto di Emergenza neonatale (STEN): raccomandazioni del gruppo di studio di trasporto neonatale, Società Italiana di neonatologia." Società Italiana di neonatologia 2018 (access to this recommendations is restricted to SIN members only); Daussac A, Leslie A, Roth-Kleiner M, Zwissig M, Aufieri R, Gente M, Jourdain G. Le transfert néonatal par un SMUR pédiatrique, de la salle de naissance à la réanimation, en France ... et ailleurs. Neonatal Transfer by a Retrieval Team, from the Delivery Room to the NICU, in France ... and Abroad. Rev. Méd. Périnat. 2018;10,50--8 (in this paper have been reported only limited national activity data needed for comparison of Neonatal Transport System activity with other European nations). All authors read and approved the final manuscript; all authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. MG (Gente): organized and performed the survey; is responsible for conceiving the study and writing the article; is a former secretary of the Neonatal Transport Study Group, SIN. RA (Aufieri): organized and performed the survey; is responsible for conceiving the study and reviewing the article. RA (Agostino): organized and performed the survey; is responsible for conceiving the study and reviewing the article; is a former secretary of the Neonatal Transport Study Group, SIN. TF (Fedeli): organized and performed the survey; is responsible for conceiving the study. MGC (Calevo): wrote the statistical analysis plan, cleaned and analyzed the data. PM (Massirio): reviewed the article. CB (Bellini): organized and performed the survey; is responsible for conceiving the study and writing the article; is presently the secretary of the Neonatal Transport Study Group, SIN. Ethics approval and consent to participate {#FPar3} ========================================== Not applicable. Consent for publication {#FPar4} ======================= Not applicable. Competing interests {#FPar5} =================== The authors declare that they have no competing interests. Publisher's Note {#FPar6} ================ Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

1. Introduction {#sec1} =============== Human Immunodeficiency Virus (HIV) and tuberculosis (TB) disease represent health issues of major public importance worldwide. Approximately one in three individuals is infected with*Mycobacterium tuberculosis globally*, and HIV is the main risk factor of active tuberculosis disease, increasing the risk of latent TB reactivation 20-fold \[[@B1]\]. Tuberculosis kills more than 1 million women per year, and it is estimated that 646 million women and girls are already infected with tuberculosis \[[@B2]\]. Among women aged 15--44 years in developing countries, tuberculosis is the third most common cause of morbidity and mortality combined and kills more women than any other infectious disease, including malaria and AIDS \[[@B3]\]. HIV-TB coinfected individuals are at greater risk for relapse following treatment as well as increased chances of drug resistance to anti-TB drugs \[[@B4]--[@B8]\]. Even in industrialized settings with good healthcare access and antiretroviral therapy (ART) availability, HIV-TB coinfection is associated with high mortality \[[@B9]\]. A recent study suggests that the rate of TB among pregnant women in the United States (US) is climbing, although the increase appears to be as a result of an upsurge in extrapulmonary disease as well as inclusion of pregnant women with reported history of tuberculosis \[[@B10]\]. This poses a concern regarding the effects of TB disease in HIV-positive mothers. Although HIV and TB among pregnant women have been studied separately, the impact of HIV-TB coinfection among affected mothers remains poorly defined or understood. The aim of this study is to determine the association between TB, HIV, and HIV-TB coinfection with pregnancy complications, alcohol use, drug abuse, and depression using a highly reliable and validated population-based dataset. Drawing from more than 57 million pregnancy-related and delivery admissions, this study is the largest in the world on HIV-TB coinfection given the sample size of the source population. 2. Materials and Methods {#sec2} ======================== We used data from the Nationwide Inpatient Sample (NIS) covering the period from January 1, 2002, through December 31, 2014, for this analysis. The NIS dataset is made available by the Healthcare Cost and Utilization Project (HCUP) and currently constitutes the largest all-payer, publicly available inpatient database in the US \[[@B11]\]. To create the sample annually, HCUP employs a two-stage cluster sampling design that first stratifies all nonfederal community hospitals from participating states by five major hospital characteristics: rural/urban location, number of beds, geographic region, teaching status, and ownership. Then, 20% of hospitals from each unique stratum were selected using a systematic random sampling technique \[[@B12]\]. In the second stage, all inpatient discharges from hospitals selected during stage 1 were selected for inclusion in the NIS. Beginning in 2012, the NIS sampling strategy was changed from keeping all discharges from a sample of hospitals to drawing a sample of discharges from all hospitals. The NIS approximates a 20% systematic sample that is representative of the population of all discharges on critical hospital and patient characteristics \[[@B12]\]. To assess the study\'s primary exposures, we scanned ICD-9-CM codes (the principal diagnosis and up to 24 secondary diagnoses) in each woman\'s discharge record for an indication of HIV and TB status. Individual-level sociodemographic and behavioral characteristics were also extracted from the NIS databases. Maternal age in years was classified into three categories: 13--24, 25--34, and 35--49. Self-reported maternal race/ethnicity was first based on ethnicity (Hispanic or non-Hispanic \[NH\]), and the NH group was further subdivided by race (white, black, or other). Median household income, which served as a proxy for socioeconomic status, was estimated using the patient\'s zip code and subsequently grouped into quartiles. We classified the primary payer for hospital admission into three categories: government (Medicare/Medicaid), private (commercial carriers, private health maintenance organization \[HMOs\], and preferred provider organization \[PPOs\]), and other sources (e.g., self-pay and charity). Due to their strong associations with HIV infection, we also used ICD-9-CM codes to ascertain information on alcohol and drug use during pregnancy. Our definition of alcohol abuse included indications of alcohol-induced mental disorders (291.0--9), alcohol dependence syndrome (303.00--93), nondependent alcohol abuse such as binge drinking (305.00--03), alcoholic cardiomyopathy (425.5), and alcohol that has affected the fetus or newborn via the placenta or breast milk (760.71). Drug abuse included drug-induced mental disorders (291.0--9), drug dependence (304.00--93), nondependent drug abuse (305.20--93), drug dependence complicating pregnancy or childbirth (648.30--34), suspected damage to the fetus from drugs (655.50--53), drugs affecting the fetus or newborn via the placenta or breast milk (760.72-73, 760.75), and poisoning by opiates and related narcotics (965.00--09). We also considered several hospital characteristics including teaching status (teaching versus nonteaching), location (urban versus rural), and US region (Northeast, Midwest, South, or West). 2.1. Statistical Analysis {#sec2.1} ------------------------- Descriptive statistics including frequencies and percentages were used to describe the distribution of index delivery hospitalizations across patient- and hospital-level characteristics and stratified by exposure groups (HIV and TB negative, HIV monoinfection, TB monoinfection, and HIV-TB coinfection). Since national estimates were desired, all statistical analyses were weighted using an HCUP-provided discharge-level weight that accounts for the NIS\'s sampling design. Multivariable survey logistic regression was then used to generate adjusted odds ratios (OR) that quantified the magnitude of the association between infection status (i.e., HIV monoinfection, TB monoinfection, and HIV-TB coinfection) and the main outcomes of the study (pregnancy complications, alcohol use, and drug abuse). Statistical analyses were performed with SAS, version 9.4 (SAS Institute, Inc., Cary, NC); we assumed a 5% type I error rate for all hypothesis tests (two-sided). Due to the deidentified, publicly available nature of NIS data, the analyses performed for this study were considered exempt by the Baylor College of Medicine Institutional Review Board. It is pertinent to mention that some of the results of our analyses contained small numbers which must be suppressed in accordance with guidelines set forth by the Healthcare Cost and Utilization Project. The reason is to prevent possible identification of these individuals, and, in such cases, we described the findings in text without displaying the actual values. 3. Results {#sec3} ========== We analyzed a total of 57,393,459 hospitalizations related to pregnancy and delivery over the study period. The diagnosis of active TB monoinfection was documented in 4,053 mothers yielding a rate of 7.06 per 100,000. In the entire population of mothers, the prevalence of HIV monoinfection was 12.76 per 10,000 (*N* = 73,223 HIV-infected mothers). Although the prevalence of HIV-TB coinfection was 1.9 per million (*N* = 110) in the entire population of pregnant mothers, the rate of TB among HIV-positive mothers was 150.23/100,000. This was 21 times as high when compared to the rate of TB in the entire population of pregnant women. [Table 1](#tab1){ref-type="table"} and Figures [1(a)](#fig1){ref-type="fig"}, [1(b)](#fig1){ref-type="fig"}, and [1(c)](#fig1){ref-type="fig"} represent a summary of sociodemographic characteristics by infection status. HIV-TB monoinfection and HIV-TB coinfection correlated positively with ascending age, with notable racial/ethnic differences. The rates of HIV and HIV-TB coinfection were highest among black mothers, while TB monoinfection was most prevalent in Hispanics. Whereas the proportion of HIV monoinfection, TB, and HIV-TB coinfection was 13-, 5-, and 20-fold as high among black mothers when compared to their white counterparts, respectively, the corresponding proportion among Hispanics was 2, 7, and 4 times as high as that of their white counterparts, respectively. Low socioeconomic status and having public or no health insurance were the greatest risks for HIV-TB monoinfection and HIV-TB coinfection. Our data also portrayed geographical differences in the rates of infection among mothers. The Northeast had the highest rates for HIV-TB monoinfection and HIV-TB coinfection, while the Western part of the US had relatively low prevalence of HIV and HIV-TB coinfection. Across health facilities, the rates for HIV-TB monoinfection and HIV-TB coinfection were highest among urban teaching hospitals and lowest for hospital facilities located in rural areas. [Table 2](#tab2){ref-type="table"} summarizes the rates of pregnancy complications, alcohol use, drug abuse, and depression among pregnancy-related hospital admissions for the entire period of the study. Note that some of the values in the table are suppressed due to small numbers in accordance with guidelines set forth by the Healthcare Cost and Utilization Project. The reason is to prevent possible identification of these individuals, and, in such cases, we described the findings in text without displaying the actual values. Pregnancy complications as a composite outcome were highest among HIV-TB coinfected mothers followed by those with TB monoinfection. Since we only observed 110 cases of HIV-TB coinfection out of the more than 57 million hospital admissions, it was to be expected that not all pregnancy complications would occur in the 110 HIV-TB coinfected mothers. Four (eclampsia, placenta accreta, postpartum hemorrhage, and anemia) of the 10 pregnancy complications were detected among women with HIV-TB coinfections. When the rates of these complications were compared across the categories of infection status, three of the four complications (eclampsia, placenta accreta, and anemia) occurred with greatest frequency among HIV-TB coinfected mothers. The only exception (postpartum hemorrhage) was highest among women with TB monoinfection. We then compared the other remaining six complications that had zero occurrence among HIV-TB coinfected mothers in disease-free (no TB and no HIV), HIV-monoinfected, and TB-monoinfected mothers. We found similar rates of preeclampsia across the three groups. Diabetes mellitus, placental abruption, and other sources of antepartum hemorrhages were highest among HIV-monoinfected mothers while placenta previa and sepsis were greatest among mothers with TB monoinfection. Alcohol abuse and drug abuse rates were consistently greatest among mothers with HIV-TB coinfection followed by those with HIV monoinfection while the prevalence of depression was most pronounced among HIV-monoinfected mothers. [Table 3](#tab3){ref-type="table"} describes the association between infection status and obstetrics complications, alcohol abuse, drug abuse, and mental health. The risk for pregnancy complications, even after adjustment for alcohol use, drug abuse, and depression, was greatest among women with HIV-TB coinfection \[OR = 2.00 (0.83, 4.79)\] although this 2-fold elevated risk level did not reach statistical significance. Mothers with HIV monoinfection exhibited the lowest likelihood for pregnancy complications \[OR = 1.40 (1.32, 1.47)\]. We did observe moderation of the effect of HIV/TB status on pregnancy complications by substance abuse behaviors and depression. Whereas women without any indication of alcohol use, drug abuse, or depression had associations between HIV/TB status and pregnancy complications that were similar to the entire study population, measures of association were greatly attenuated among women with substance abuse or depression (data not shown). The likelihood for alcohol abuse was more than twice as high among HIV-positive patients \[OR = 5.12 (4.42, 5.93)\] as compared to TB-monoinfected patients \[OR = 2.10 (0.79, 5.55)\]. That risk gap more than doubled with HIV-TB coinfection \[OR = 11.39 (1.38, 94.06)\]. Mothers with HIV or HIV-TB coinfection experienced elevated risks for drug abuse and depression almost twice as much as those with TB monoinfection only. 4. Discussion {#sec4} ============= The prevalence of 1.9 HIV-TB coinfection cases per one million pregnant women in our study is lower than that for resource-limited nations, although our denominator (more than 57 million admissions) is the largest reported. It is noteworthy that the rate of TB among HIV-positive mothers was 150.23/100,000 in the population we analyzed, which was 21 times as high when compared to the rate of TB in the entire population of pregnant women analyzed. This demonstrates the extremely high vulnerability of HIV-infected persons to TB disease. People infected with HIV are at high risk for developing active TB due to either the reactivation of a latent infection or the progression of newly acquired tubercle bacilli infections. Among HIV-positive individuals, the immune responses to tubercle bacilli may not be effective due to underlying HIV-induced immunological aberrations which weaken effector immune responses \[[@B13], [@B14]\]. Results from immune phenotyping studies indicate that peripheral CD4+T cells from a subset of HIV-positive patients on ART regimens still continue to manifest characteristic features of pronounced cellular exhaustion \[[@B15]\], turnover \[[@B16]\], and senescence \[[@B17]\]. These aberrant immunological patterns will explain the raised likelihood of tuberculosis disease and reactivation of latent tuberculosis in HIV-positive pregnant mothers as reported in this study. We also found HIV-TB coinfection to be associated with the greatest risk for pregnancy complications as compared to HIV and TB monoinfection status. However, the almost 2-fold elevated risk among HIV-TB coinfected mothers had a wide confidence interval that encompassed the null. This might be due to the small number of cases. Nonetheless, our results are in congruence with previous studies in low-resource high-HIV prevalence settings which demonstrated significant maternal morbidity and mortality in pregnant women with HIV-TB coinfection \[[@B18]--[@B20]\]. In a recent analysis by Canadian authors utilizing the same dataset covering the period 2003 to 2011, the investigators reported elevated risks of chorioamnionitis, preterm labor, postpartum anemia, blood transfusion, and pneumonia as well as maternal mortality among pregnant women with tuberculosis disease \[[@B10]\]. Our study showed that HIV-coinfected women were particularly at high risk for eclampsia, placenta accreta, and anemia. The heightened risks of maternal complications among mothers with HIV-TB coinfection represent a pathway for the increased risk of maternal mortality consistently reported among HIV-positive mothers with TB coinfection. We also observed elevated risk for alcohol use which was more than twice as high among HIV-positive patients \[OR = 5.12 (4.42, 5.93)\] as compared to TB-monoinfected patients \[OR = 2.10 (0.79, 5.55)\]. That risk gap more than doubled with HIV-TB coinfection \[OR = 11.39 (1.38, 94.06)\]. Mothers with HIV or HIV-TB coinfection experienced elevated risks for drug abuse and depression, almost twice as much as those with TB monoinfection only. A potential mechanism that explains the strong positive association between HIV positivity and excessive alcohol use, drug abuse, and depression is perceived stress \[[@B21]\]. HIV-positive persons who experience the greatest stress in their daily lives are those who disengage behaviorally/emotionally in coping with their illness and those who approach their interpersonal relationships in a less secure or more anxious style \[[@B22]\]. By contrast, studies indicate that, for people with HIV or AIDS, those individuals who are more satisfied with their relationships, securely engaged with others, and more directly engaged with their illness are more likely to experience positive adjustment \[[@B23]--[@B25]\]. It is noteworthy to mention that reverse causality does exist in the relationships between HIV on one hand and alcohol use/drug abuse on the other, since drug abusers and alcohol users tend to engage in behaviors that put them at elevated risk for HIV infection \[[@B26], [@B27]\]. The nature of social support could also influence the likelihood of risk behaviors among HIV-positive individuals. Previous investigators have noted the importance of the lack of positive social support or the existence of "negative" support as an important factor that encourages sexual and drug-related HIV risk behavior such that illegal drug use by members of a social network often provides peer support for continued risk behavior \[[@B28]\]. Our results tend to suggest that the concomitant existence of TB disease among HIV-positive individuals increased the risk for alcohol use and drug abuse although the magnitude of the excess risk was more pronounced for alcohol use. The results of our analysis underscore the importance of augmenting and enhancing social and structural support systems for HIV-TB coinfected pregnant women because of the increased risk for pregnancy complications in addition to the associated raised likelihood of alcohol use and drug abuse. Given the significant findings in this study, tailored interventions are needed to maximize the effectiveness of management modalities of HIV-TB coinfected pregnant mothers in order to prevent or mitigate the health consequences of concomitant alcohol use, drug abuse, and depression. It is recommended that HIV-TB coinfected mothers be supported with sufficient provider-patient communication and contact time before discharge in order to reinforce understanding of their condition and link them with community assets to cater for their future needs and care related to sustained pregnancy complications. Even though the dataset used for this study is nationally representative and rich with an array of clinical and epidemiologic measures, certain limitations merit mentioning. It is logical to assume that the quality of care received for the management of HIV and TB among the mothers in this study would vary considerably across the diverse population settings in the United States. We lack a marker/measure of quality of care or adherence with HIV care protocols in the dataset to assess variation in quality of management of HIV and TB conditions which could have influenced the occurrence of the maternal pregnancy complications examined in this study. It is notable that 60% of the pregnancy complications were not detected among women with HIV-TB coinfection due to the rarity of HIV-TB coinfection (1.9 cases per million). Another shortcoming in this study is the lack of biological markers (e.g., CD4 cell count and HIV load) to predict mothers that were likely to develop pregnancy complications, or to determine the association of these biologic markers with alcohol use, drug abuse, or depression. Despite these limitations, our analysis is based on the most powerful data so far in terms of sample size (57 million) for the source population of HIV-TB coinfected pregnant women and the range of outcomes examined in our study. Research funding support was provided by the US Department of Health and Human Services, Health Resources and Services Administration for the Maternal and Child Health Pipeline Training Program: TSU-BCM Maternal and Child Health Student Training for Academic Readiness and Success (MCH STARS) Undergraduate Fellowship Program, Grant no. T16MC29831. Conflicts of Interest ===================== The authors declare that they have no conflicts of interest. {#fig1} ###### Distribution of sociodemographic and health facility characteristics among hospital admissions of pregnant women, by HIV/TB status, United States, 2002--2014.   HIV/TB status --------------------- --------------- ------ ------ ------ *Age*          13--24 33.9 26.4 31.3 c  25--34 51.4 56.5 50.8 70.3  35--49 14.7 17.2 17.8 21.4 *Race/ethnicity*          NH-White 42.1 14.5 11.3 c  NH-Black 11.8 54.8 16.0 49.9  Hispanic 18.9 12.3 34.8 16.7  Other 8.6 5.4 23.0 12.7  Not reported 18.7 13.0 14.9 12.4 *Income*          Lowest 27.1 47.3 36.3 44.7  2nd 24.8 21.0 24.3 12.1  3rd 24.1 14.6 20.9 25.7  Highest 22.2 8.1 16.1 8.7 *Payer*          Government 43.0 72.7 62.6 66.4  Private 50.5 18.8 23.6 21.3  Other 6.5 8.5 13.8 12.3 *Hospital region*          Northeast 16.5 25.1 24.0 30.3  Midwest 21.2 13.1 18.3 c  South 37.9 56.8 34.0 53.2  West 24.4 5.1 23.7 c *Hospital type*          Rural 11.2 3.3 4.8 c  Urban, nonteaching 40.4 16.7 20.0 16.4  Urban, teaching 48.0 79.5 74.6 79.3 ^a^The estimated number of records with HIV--, TB-- is 57,316,293. ^b^Displayed percentages are column percentages. Percentages may not add to 100% due to missing data. ^c^Data are suppressed due to small numbers in accordance with guidelines set forth by the Healthcare Cost and Utilization Project. ###### Rates (per 1,000 hospitalizations) of alcohol use, drug abuse, depression, and complications of pregnancy among pregnancy-related admissions, by HIV/TB status, United States, 2002--2014.   HIV/TB status ------------------------------- --------------- ------- ------- -------- Alcohol use 1.7 16.1 4.7 b Drug abuse 16.0 114.6 25.3 119.08 Depression 20.1 53.2 25.8 b Pregnancy complications (all) 166.4 272.2 292.1 331.28  Anemia 102.7 214.1 216.6 206.33  Diabetes 11.3 24.6 16.7 b  Preeclampsia 23.4 23.8 24.0 b  Eclampsia 1.0 1.4 2.8 b  Placental abruption 10.6 11.6 8.9 b  Placenta previa 5.2 5.5 9.9 b  Placenta accreta 3.3 2.6 1.3 b  Others^c^ 3.6 4.9 3.6 b  Postpartum hemorrhage 25.8 21.6 45.5 b  Sepsis 1.3 5.4 8.3 b ^a^The estimated number of records with HIV−, TB− is 57,316,293. ^b^Data are suppressed in accordance with guidelines set forth by the Healthcare Cost and Utilization Project. ^c^Other antepartum hemorrhages. ###### Adjusted odds ratios for the association between HIV infection, TB disease, and HIV-TB coinfection status and obstetrics complications, alcohol abuse, drug abuse, and mental health among pregnancy-related admissions in the Unites States. Status Outcome ----------- ----------------------- ------------------------- ------------------------ ----------------------- HIV−, TB− 1.00 (reference) 1.00 (reference) 1.00 (reference) 1.00 (reference) HIV+, TB− **1.40 (1.32, 1.47)** **5.12 (4.42, 5.93)** **4.63 (4.19, 5.12)** **2.83 (2.60, 2.39)** TB+, HIV− **1.91 (1.64, 2.23)** 2.10 (0.79, 5.55) 1.39 (0.92, 2.12) 1.54 (1.00, 2.39) HIV+, TB+ 2.00 (0.83, 4.79) **11.39 (1.38, 94.06)** **5.08 (1.43, 18.04)** 2.55 (0.34, 19.09) *Note*. Statistically significant adjusted odds ratios (95% confidence interval does not include 1) appear in bold text. Models for all outcomes are adjusted for age, race, income, payer, year of hospitalization, and hospital region; the model with pregnancy complications as the outcome is also adjusted for alcohol use, drug abuse, and depression. [^1]: Academic Editor: Fabio Facchinetti

\section{Introduction} Power transmission over long distances with low losses is one of the main challenges in today's power transmission systems. As the share of renewables rises, so does the need to balance generation and consumption mismatches, often over large geographical areas. Due to the high resistive losses in AC cables, high-voltage direct current (HVDC) power transmission is a commonly used technology for long-distance power transmission. The higher investment cost of an HVDC transmission system compared to an AC transmission system is compensated by the lower resistive losses for sufficiently long distances \citep{melhem2013electricity}. The break-even point, i.e., when the total construction and operation costs of overhead HVDC and AC lines equal, is typically 500--800 km \citep{padiyar1990hvdc}. However, for cables, the break-even point is typically less than 50 km \citep{Hertem2010technical}. Increased use of HVDC technologies for electrical power transmission suggests that future HVDC transmission systems are likely to consist of multiple terminals connected by several HVDC transmission lines \citep{Haileselassie2013Power}. Such systems are referred to as multi-terminal HVDC (MTDC) systems in the literature. Many existing AC grids are connected through legacy HVDC links, which are typically used for bulk power transfer between AC areas, rather than balancing generation mismatches. The fast operation of the DC converters however also enables frequency regulation of one of the connected AC grids through the HVDC link. One example is the frequency regulation of the island of Gotland in Sweden, which is connected to the Nordic grid through an HVDC cable \citep{axelsson2001gotland}. However, since the Nordic grid has orders of magnitudes higher inertia than the AC grid of Gotland, the influence of the frequency regulation on the Nordic grid is negligible. By connecting multiple AC grids by an MTDC system, the frequency regulation reserves in each AC grid can be shared with connected AC grids, which reduces the need for frequency regulation reserves in the individual AC systems \citep{li2008frequency}. In recent years, this idea has gained increasing interest in the literature. \cite{dai2011voltage} and \cite{silva2012provision} employ several controllers with decentralized structure to share frequency control reserves. In \citep{silva2012provision} no stability analysis of the closed-loop system is performed, whereas \citep{dai2011voltage} guarantees stability provided that the connected AC areas have identical parameters and the voltage dynamics of the HVDC system are neglected. In \citep{taylordecentralized}, an optimal decentralized controller for AC systems connected by HVDC systems is derived. In contrast to the aforementioned references, \cite{andreasson2014distributed} consider the dynamics of connected AC systems as well as the dynamics of the MTDC system. By connecting the AC areas with a communication network, the performance of the controller can be further improved. \cite{dai2010impact} consider a distributed controller relying on a communication network to share frequency control reserves of asynchronous AC transmission systems connected through an MTDC system. However, the controller requires a slack bus to control the DC voltage, and is thus only able to share the generation reserves of the non-slack AC areas. Another distributed controller is proposed by \cite{dai2013voltage}. Stability is guaranteed, and the need for a slack bus is eliminated. The voltage dynamics of the MTDC system are however neglected. Moreover the implementation of the controller requires every controller to access measurements of the DC voltages of all other MTDC terminals. \citep{andreasson2014distributedSecondary} propose a distributed secondary generation controller. In contrast to the aforementioned references, the MTDC dynamics are explicitly modeled, and the DC voltages are controlled in addition to the frequencies. Despite the coordination through a communication network, the distributed controller by \cite{andreasson2014distributedSecondary} fails in eliminating the static errors of the frequencies. In this paper we address this issue by introducing a distributed secondary controller also for the power fed into the MTDC grid from the AC grids. This allows us to eliminate any static errors in the frequency deviations. Furthermore, quadratic cost functions of the power generation and the voltage deviations are minimized. The remainder of this paper is organized as follows. In Section \ref{sec:model}, the system model and control objectives are defined. In Section \ref{sec:secondary_frequency_control}, a distributed secondary frequency controller for sharing frequency control and restoration reserves is presented, and is shown to satisfy the control objectives. In Section \ref{sec:simulations}, simulations of the controller on a six-terminal MTDC test system are provided, showing the effectiveness of the proposed controller. The paper ends with concluding remarks in Section~\ref{sec:discussion}. \section{Model and problem setup} \label{sec:model} \subsection{Notation} \label{subsec:prel} Let $\mathcal{G}$ be a static, undirected graph. Denote by $\mathcal{V}=\{ 1,\hdots, n \}$ the vertex set of $\mathcal{G}$, and by $\mathcal{E}=\{ 1,\hdots, m \}$ the edge set of $\mathcal{G}$. Let $\mathcal{N}_i$ be the set of neighboring vertices to $i \in \mathcal{V}$. Denote by $\mathcal{B}$ the vertex-edge incidence matrix of $\mathcal{G}$, and let $\mathcal{\mathcal{L}_W}=\mathcal{B}W\mathcal{B}^T$ be the weighted Laplacian matrix of $\mathcal{G}$, with edge-weights given by the elements of the diagonal matrix $W$. We denote the space of real-valued $n\times m$-valued matrices by $\mathbb{R}^{n\times m}$. Let $\mathbb{C}^-$ denote the open left half complex plane, and $\bar{\mathbb{C}}^-$ its closure. We denote by $c_{n\times m}$ a vector or matrix of dimension $n\times m$, whose elements are all equal to $c$. For simplifying notation, we write $c_n$ for $c_{n\times 1}$. For a symmetric matrix $A$, $A>0 \;(A\ge 0)$ is used to denote that $A$ is positive (semi) definite. $I_{n}$ denotes the identity matrix of dimension $n$. For simplicity, we will often drop the notion of time dependence of variables, i.e., $x(t)$ will be denoted $x$. Let $\norm{\cdot}_\infty$ denote the maximal absolute value of the elements of a vector. \subsection{Model} \label{subsec:model} We will give here a unified model for an MTDC system interconnected with several mutually asynchronous AC systems. We consider an MTDC transmission system consisting of $n$ converters, denoted $i=1, \dots, n$, each connected to an AC system, i.e., there are no pure DC nodes of the MTDC grid. The converters are assumed to be connected by an MTDC transmission grid. The dynamics of converter $i$ is assumed to be given by \begin{align} \begin{aligned} C_i \dot{V}_i &= -\sum_{j\in \mathcal{N}_i} \frac{1}{R_{ij}}(V_i -V_j) + I_i^{\text{inj}} , \end{aligned} \label{eq:hvdc_coordinated_voltage} \end{align} where $V_i$ is the voltage of converter $i$, $C_i>0$ is its capacitance, and $I_i^{\text{inj}}$ is the injected current from an AC grid connected to the DC converter. The constant $R_{ij}$ denotes the resistance of the HVDC transmission line connecting the converters $i$ and $j$. The graph corresponding to the HVDC line connections is assumed to be connected. The AC system is assumed to consist of a single generator which is connected to the corresponding DC converter, representing an aggregate model of the AC grid. The dynamics of the AC system are given by the swing equation \citep{machowski2008power}: \begin{align} m_i \dot{\omega}_i &= P^\text{gen}_i + P_i^{{m}} - P_i^{\text{inj}}, \label{eq:frequency} \end{align} where $m_i>0$ is its moment of inertia. The constant $P^\text{gen}_i$ is the generated power, $P^m_i$ is the power load and $P_i^{\text{inj}}$ is the power injected to the DC system through converter $i$, respectively. The powers are all assumed to be deviations from a nominal operation point. The control objective can now be stated as follows. \begin{objective} \label{obj:1_hvdc_coordinated} The frequency deviations are asymptotically equal to zero, i.e., \begin{align} \lim_{t\rightarrow \infty} \omega_i(t)-\omega^{\text{ref}} = 0 \quad i = 1, \dots, n, \label{eq:hvdc_coordinated_frequency_objective} \end{align} where $\omega^{\text{ref}}$ is the nominal frequency. The total quadratic cost of the power generation is minimized asymptotically, i.e., $\lim_{t\rightarrow \infty} P_i^\text{gen} = P_i^{\text{gen}*}, \forall i=1, \dots, n$, where \begin{align} [P_1^{\text{gen}*}, \dots, P_n^{\text{gen}*}] = \argmin_{P_1, \dots, P_n} \frac 12 \sum_{i=1}^n f^P_i P_i^2 \label{eq:hvdc_coordinated_generation_objective} \end{align} subject to $P^\text{gen}_i + P_i^{{m}} - P_i^{\text{inj}} = 0, \; \forall i=1, \dots, n$ and $\sum_{i=1}^n P_i^\text{inj} = 0$, i.e., power balance both in the AC grids and in the MTDC grid, in the absence of power losses. The positive constants $f^P_i$ represent the local cost of generating power. Finally, the voltages are such that the a quadratic cost function of the voltage deviations is minimized asymptotically, i.e., $\lim_{t\rightarrow \infty} V_i = V_i^*, \forall i=1, \dots, n$, where \begin{align} [V_1^*, \dots, V_n^*] = \argmin_{V_1, \dots, V_n} \frac 12 \sum_{i=1}^n f^V_i (V_i - V_i^\text{ref})^2 \label{eq:hvdc_coordinated_voltage_objective} \end{align} subject to \eqref{eq:hvdc_coordinated_frequency_objective}--\eqref{eq:hvdc_coordinated_generation_objective}, and where the $f^V_i$ is a positive constant reflecting the local cost of voltage deviations, and $V_i^\text{ref}$ is the nominal voltage of converter $i$. \end{objective} \begin{remark} The minimization of \eqref{eq:hvdc_coordinated_generation_objective} is equivalent to power sharing, where the generated power of AC area $i$ is asymptotically inverse proportional to the cost $f_i^P$. The cost $f_i^P$ can be chosen to reflect the available generation capacity of area $i$. \end{remark} \begin{remark} It is in general not possible that $\lim_{t\rightarrow \infty} V_i(t) = V_i^\text{ref} \; \forall i=1, \dots, n$, since this does not allow for the currents between the HVDC converters to change by \eqref{eq:hvdc_coordinated_voltage}. Note that the optimal solution to \eqref{eq:hvdc_coordinated_generation_objective} fixes the relative voltages, leaving only the ground voltage as a decision variable of \eqref{eq:hvdc_coordinated_voltage_objective}. Note also that the reference voltages $V_i^\text{ref}, \; i=1, \dots, n$, are generally not uniform, as is the reference frequency $\omega^\text{ref}$. \end{remark} \begin{remark} Note that Objective \ref{obj:1_hvdc_coordinated} does not include constraints of, e.g., generation and line capacities. This requires that the perturbations from the operating point are sufficiently small, to guarantee that these constraints are not violated. In case of a large disturbance, e.g., a fault, the constraints might be violated. \end{remark} \section{Coordinated secondary frequency control} \label{sec:secondary_frequency_control} \subsection{Controller structure} \label{subsec:controlelr_structure} In this section we propose a distributed secondary frequency controller. In addition to the generation controller proposed by \cite{andreasson2014distributedSecondary}, we also propose a secondary controller for the voltage injections into the HVDC grid. The distributed generation controller of the AC systems, which was first given by \cite{andreasson2014distributedSecondary}, is given by \begin{align} P^\text{gen}_i &=- K_i^{\text{droop}} (\omega_i-\omega^{\text{ref}}) - \frac{K^V_i}{K^\omega_i} K^\text{droop, I}_i \eta_i \nonumber \\ \dot{\eta}_i &= K_i^{\text{droop,I}}(\omega_i-\omega^{\text{ref}}) - \sum_{j\in \mathcal{N}_i} c^\eta_{ij} (\eta_i-\eta_j), \label{eq:hvdc_coordinated_droop_control_secondary_distributed} \end{align} where $K_i^\text{droop}$ and $K_i^\text{droop, I}$ are positive controller parameters. Moreover $c^\eta_{ij} = c^\eta_{ji}>0$, i.e., the communication graph is undirected. The above controller can be interpreted as a distributed PI-controller, with a distributed consensus filter acting on the integral states $\eta_i$. The proposed converter controllers governing the incremental power injections from the AC systems into the HVDC grid are given by \begin{align} \label{eq:voltage_control_secondary} P_i^{\text{inj}} &= K_i^{{\omega}} (\omega_i - \omega^{\text{ref}}) + K_i^{{V}}(V_i^{\text{ref}}-V_i) \nonumber \\ & \;\;\;\;+ \sum_{j\in \mathcal{N}_i} c^\phi_{ij} (\phi_i - \phi_j) \nonumber \\ \dot{\phi}_i &= \frac{K^\omega_i}{K^V_i} \omega_i - \gamma \phi_i, \end{align} where $K_i^V$ and $K_i^\omega$ are positive controller parameters, $\gamma>0$ and $c^\phi_{ij} = c^\phi_{ji}>0$. The converter controller \eqref{eq:voltage_control_secondary} can be interpreted as an emulation of an AC network between the isolated AC areas. The auxiliary controller variables $\phi_i$ can be thought of as the phase angles of AC area $i$, governing the power transfer between the areas, if these were connected by AC lines rather than an MTDC grid. In contrast to an AC system, however, the power is fed into the HVDC grid instead of being directly transferred to the connected AC areas. The HVDC converter is assumed to be perfect and instantaneous, i.e., injected power on the AC side is immediately and losslessly converted to DC power. Furthermore, the dynamics of the converter are ignored, implying that the converter is assumed to track the output of controller \eqref{eq:voltage_control_secondary} instantaneously. This assumption is reasonable due to the dynamics of the converter typically being orders of magnitudes faster than the primary frequency control dynamics of the AC system \citep{kundur1994power}. The relation between the injected HVDC current and the injected AC power is thus given by \begin{align} V_iI_i^{\text{inj}} = P_i^{\text{inj}}. \label{eq:power-current_nonlinear} \end{align} By assuming $V_i=V^{\text{nom}}\; \forall i=1, \dots, n$, where $V^{\text{nom}}$ is a global nominal voltage, we obtain \begin{align} V^{\text{nom}}I_i^{\text{inj}} = P_i^{\text{inj}}. \label{eq:power-current} \end{align} \subsection{Stability analysis} \label{sec:secondary_frequency_control_stability} We are now ready to analyze the stability of the closed-loop system. Define the stacked state vectors as $\hat{\omega}=\omega - \omega^\text{ref}1_n$ and $\hat{V}=V - V^\text{ref}$, where $\omega=[\omega_1, \dots, \omega _n]^T$, $V=[V_1, \dots, V_n]^T$, $V^\text{ref} = [V_1^\text{ref}, \dots, V_n^\text{ref}]^T$, $\eta = [\eta_1, \dots, \eta_n]^T$ and $\phi = [\phi_1, \dots, \phi_n]$. Combining the voltage dynamics \eqref{eq:hvdc_coordinated_voltage}, the frequency dynamics \eqref{eq:frequency} and the generation control \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed}, the converter controller \eqref{eq:voltage_control_secondary} and the power-current relationship \eqref{eq:power-current}, we obtain the closed-loop dynamics \begin{IEEEeqnarray}{lcl} \dot{\hat{\omega}} &=& M \Big(- (K^{\text{droop}} + K^\omega) \hat{\omega} + K^V \hat{V} \nonumber \\ && - {K^V}(K^\omega)^{-1} K^\text{droop, I} \eta - \mathcal{L}_\phi \phi + P^{{m}} \Big) \nonumber \\ \dot{\hat{V}} &=& \frac{1}{V^{\text{nom}}}E {K}^\omega \hat{\omega} -E\left(\mathcal{L}_R + \frac{K^V}{V^{\text{nom}}} \right) \hat{V} + \frac{1}{V^\text{nom}} E \mathcal{L}_\phi \phi \nonumber \\ \dot{\eta} &=& K^{\text{droop,I}}\hat{\omega} - \mathcal{L}_\eta \eta \nonumber \\ \dot{\phi} &=& (K^V)^{-1}{K}^\omega \hat{\omega} -\gamma \phi, \label{eq:cl_dynamics_vec_delta_int_coordinated} \end{IEEEeqnarray} where $M=\diag({m_1}^{-1}, \hdots , {m_n}^{-1})$ is a matrix of inverse generator inertia, $E=\diag(C_1^{-1}, \dots, C_n^{-1})$ is a matrix of electrical elastances, $\mathcal{L}_R$ is the weighted Laplacian matrix of the MTDC grid with edge-weights $1/R_{ij}$, $\mathcal{L}_\eta$ and $\mathcal{L}_\phi$ are the weighted Laplacian matrices of the communication graphs with edge-weights $c^\eta_{ij}$ and $c^\phi_{ij}$, respectively, and $P^m=[P^m_1, \dots, P^m_n]^T$. We define the diagonal matrices of the controller gains as $K^\omega = \diag(K^\omega_1, \dots, K^\omega_n)$, etc. Let $y=[\hat{\omega}^T, \hat{V}^T]^T$ define the output of \eqref{eq:cl_dynamics_vec_delta_int_coordinated}. Clearly the linear combination $1_n^T\phi$ is unobservable and marginally stable with respect to the dynamics \eqref{eq:cl_dynamics_vec_delta_int_coordinated}, as it lies in the nullspace of $\mathcal{L}_\phi$. In order to facilitate the stability analysis, we will perform a state-transformation to this unobservable mode. Consider the following state-transformation: \begin{align} \phi' = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n^T \\ S^T \end{bmatrix} \phi \qquad \phi = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n & S \end{bmatrix} \phi' \label{eq:transformation_phi} \end{align} where $S$ is an $n\times(n-1)$ matrix such that $\left[\frac{1}{\sqrt{n}}1_n \; S\right]$ is orthonormal. By applying the state-transformation \eqref{eq:transformation_phi} to \eqref{eq:cl_dynamics_vec_delta_int_coordinated}, we obtain dynamics where it can be shown that the state $\phi_1'$ is unobservable with respect to the defined output. Hence, omitting $\phi_1'$ does not affect the output dynamics. Thus, we define $\phi''=[\phi_2', \dots, \phi_n']$, and obtain the dynamics \begin{IEEEeqnarray}{lcl} \dot{\hat{\omega}} &=& M \Big(- (K^{\text{droop}} + K^\omega) \hat{\omega} + K^V \hat{V} \nonumber \\ && - {K^V}(K^\omega)^{-1} K^\text{droop, I} \eta - \mathcal{L}_\phi S \phi'' + P^{{m}} \Big) \nonumber \\ \dot{\hat{V}} &=& \frac{1}{V^{\text{nom}}}E {K}^\omega \hat{\omega} -E\left(\mathcal{L}_R + \frac{K^V}{V^{\text{nom}}} \right) \hat{V} + \frac{1}{V^\text{nom}} E \mathcal{L}_\phi S\phi'' \nonumber \\ \dot{\eta} &=& K^{\text{droop,I}}\hat{\omega} - \mathcal{L}_\eta \eta \nonumber \\ \dot{\phi}'' &=& S^T(K^V)^{-1}{K}^\omega \hat{\omega} -\gamma \phi''. \label{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} \end{IEEEeqnarray} Provided that the system matrix, denoted $A$, of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} is full-rank, then \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} has a unique equilibrium. Denote this equilibrium $x_0=[\omega_0^T, V_0^T, \eta_0^T, \phi_0''^T]^T$. Define $\bar{x}\triangleq [\bar{\omega}^T, \bar{V}^T, \bar{\eta}^T, \bar{\phi}^T]^T =[\hat{\omega}^T, \hat{V}^T, \eta^T, \phi''^T]^T - [\omega_0^T, V_0^T, \eta_0^T, \phi_0''^T]^T$. Hence, \begin{align} \dot{\bar{x}} = A \bar{x}, \label{eq:dynamics_A_coordinated_shifted} \end{align} with the origin as its unique equilibrium. We are now ready to show the main stability result of this section. We first make the following assumptions. \begin{assumption} \label{ass:L_phi_coordinated} The Laplacian matrix satisfies $\mathcal{L}_\phi = k_\phi \mathcal{L}_R$. \end{assumption} Assumption \ref{ass:L_phi_coordinated} can be interpreted as the emulated AC dynamics of \eqref{eq:voltage_control_secondary} having the same susceptance ratios as the conductance ratios of the HVDC lines. \begin{assumption} \label{ass:gamma_coordinated} The gain $\gamma$ satisfies $\gamma > \frac{k_\phi}{4V^\text{nom}}$. \end{assumption} Assumption \ref{ass:gamma_coordinated} lower bounds for the damping coefficient $\gamma$. \begin{theorem} \label{th:stability_A_coordinated} If Assumptions \ref{ass:L_phi_coordinated} and \ref{ass:gamma_coordinated} hold, the origin of \eqref{eq:dynamics_A_coordinated_shifted} is globally asymptotically stable. \end{theorem} \begin{pf} Consider the Lyapynov function candidate: \begin{align} W(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) &= \frac 12 \bar{\omega}^T K^\omega (K^V)^{-1} M^{-1}\bar{\omega} + \frac{V^\text{nom}}{2} \bar{V}^T C \bar{V} \nonumber \\ &\;\;\;\; + \frac 12 \bar{\eta}^T \bar{\eta} + \frac 12 \bar{\phi}^T S^T \mathcal{L}_\phi S \bar{\phi}, \label{eq:lyap_hvdc_coordinated} \end{align} where $C=\diag(C_1, \dots, C_n)$. Clearly $W(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi})$ is positive definite and radially unbounded. Differentiating \eqref{eq:lyap_hvdc_coordinated} along trajectories of \eqref{eq:dynamics_A_coordinated_shifted}, we obtain \begin{align*} \dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) &= \bar{\omega}^T K^\omega (K^V)^{-1} M^{-1}\dot{\bar{\omega}} + V^\text{nom} \bar{V}^T C \dot{\bar{V}} \nonumber \\ &\;\;\;\; + \bar{\eta}^T \dot{\bar{\eta}} + \bar{\phi}^T S^T \mathcal{L}_\phi S \dot{\bar{\phi}} \\ &= \bar{\omega}^T \big( -K^\omega (K^V)^{-1}(K^\omega + K^\text{droop})\bar{\omega} \\ &\;\;\;\; + K^\omega \bar{V} - K^\text{droop, I} \bar{\eta} - K^\omega (K^V)^{-1} \mathcal{L}_\phi S \bar{\phi} \big) \\ &\;\;\;\; + \bar{V}^T \Big( K^\omega \bar{\omega} - (V^\text{nom}\mathcal{L}_R {+} K^V)\bar{V} + \mathcal{L}_\phi S \bar{\phi} \Big) \\ &\;\;\;\; + \bar{\eta}^T \Big( K^\text{droop, I} \bar{\omega} - \mathcal{L}_\eta \eta \Big) \\ &\;\;\;\; + \bar{\phi}^T S^T \mathcal{L}_\phi S \Big( S^T (K^V)^{-1}K^\omega \bar{\omega} - \gamma \bar{\phi} \Big) \\ &= -\bar{\omega}^T \big( -K^\omega (K^V)^{-1}(K^\omega + K^\text{droop})\bar{\omega} \\ &\;\;\;\; + 2 \bar{\omega}^T K^\omega \bar{V} - \bar{V}^T (V^\text{nom}\mathcal{L}_R + K^V)\bar{V} \\ &\;\;\;\; + \bar{V}^T \mathcal{L}_\phi S \bar{\phi} - \bar{\eta}^T \mathcal{L}_\eta \bar{\eta} - \gamma \bar{\phi}^T S^T \mathcal{L}_\phi S \bar{\phi}, \end{align*} since $SS^T = I_n - \frac 1n 1_{n\times n}$, so $\mathcal{L}_\phi SS^T = \mathcal{L}_\phi$. By defining \begin{align*} \bar{V}' = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n^T \\ S^T \end{bmatrix} & \bar{V} \qquad \bar{V} = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n & S \end{bmatrix} \bar{V}', \end{align*} we obtain $\bar{V}^T\mathcal{L}_R \bar{V} = \bar{V}''^T S^T\mathcal{L}_R S \bar{V}'' $ and $\bar{V}^T\mathcal{L}_\phi S \bar{\phi} = \bar{V}''^T S^T\mathcal{L}_\phi S \bar{\phi}$, where $\bar{V}'' = [\bar{V}'_2, \dots, \bar{V}'_n]^T$. By invoking Assumption~\ref{ass:L_phi_coordinated}, we obtain \begin{align*} &\dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) \\ &= - \begin{bmatrix} \bar{\omega}^T & \bar{V}^T \end{bmatrix} \underbrace{\begin{bmatrix} K^\omega (K^V)^{-1}(K^\omega + K^\text{droop}) & -K^\omega \\ -K^\omega & K^V \end{bmatrix}}_{\triangleq Q_1} \begin{bmatrix} \bar{\omega} \\ \bar{V} \end{bmatrix} \\ &\;\;\;\; - \begin{bmatrix} \bar{V}''^T & \bar{\phi}^T \end{bmatrix} \underbrace{\begin{bmatrix} V^\text{nom} S^T\mathcal{L}_R S & -\frac{k_\phi}{2} S^T\mathcal{L}_R S \\ -\frac{k_\phi}{2} S^T\mathcal{L}_R S & \gamma k_\phi S^T\mathcal{L}_R S \end{bmatrix}}_{\triangleq Q_2} \begin{bmatrix} \bar{V}'' \\ \bar{\phi} \end{bmatrix} \\ &\;\;\;\; - \bar{\eta}^T \mathcal{L}_\eta \bar{\eta}. \end{align*} By applying the Schur complement condition for positive definiteness, we see that $Q_1$ is positive definite iff \begin{eqnarray*} K^\omega (K^V)^{-1}(K^\omega + K^\text{droop}) - K^\omega (K_V)^{-1} K^\omega \\ = K^\omega (K^V)^{-1} K^\text{droop} > 0. \end{eqnarray*} Hence $Q_1$ is positive definite. By applying the same argument to $Q_2$, $Q_2$ is positive definite iff \begin{align*} \Bigg( \gamma k_\phi- \frac{k_\phi^2}{4 V^\text{nom}} \Bigg) S^T \mathcal{L}_R S > 0. \end{align*} Clearly the above matrix inequality holds under Assumption~\ref{ass:gamma_coordinated}, since $S^T \mathcal{L}_R S \ge 0$, and $Sx\ne k 1_n$ for $k\ne 0$. Thus $\dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) \le 0$, and the set where $W(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi})$ is non-decreasing is given by \begin{align*} G &= \{ (\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) | \dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) =0 \} \\ &= \{ (\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) | \bar{\eta} = k 1_n \}, \end{align*} for any $k\in \mathbb{R}$. Clearly the largest invariant set in $G$ is the origin, and thus $k=0$. By LaSalle's theorem, the origin is globally asymptotically stable under the dynamics \eqref{eq:dynamics_A_coordinated_shifted}. \end{pf} We now turn our attention to the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime}. If the stability condition in Theorem~\ref{th:stability_A_coordinated} is met, then the equilibrium is stable, and $A$ is thus Hurwitz. Hence \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} must have a unique equilibrium. In the following corollary we show that the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} satisfies Objective~\ref{obj:1_hvdc_coordinated}. \begin{corollary} \label{cor:hvdc_coordinated_equilibrium} Let Assumption~\ref{ass:L_phi_coordinated} hold and let $\gamma$, $k_\phi$ be given such that Assumption~\ref{ass:gamma_coordinated} holds. Let $K^V, K^\omega$ and $K^\text{droop}$ be such that $(F^P)^{-1} = K^V(K^\omega)^{-1}K^\text{droop}$ and $F^V=K^V$, where $F^P = \diag(f^P_1, \dots, f^P_n)$ and $F^V = \diag(f^V_1, \dots, f^V_n)$. Then Objective~\ref{obj:1_hvdc_coordinated} is satisfied in the limit when $\norm{(K^\omega)^{-1} K^V}_\infty \rightarrow 0$. \end{corollary} \begin{pf} By Theorem~\ref{th:stability_A_coordinated}, \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} has a unique and stable equilibrium. The last $n-1$ rows of the equilibrium imply \begin{align*} S^T(K^V)^{-1}K^\omega \hat{\omega} - \gamma \phi'' = 0_{n-1}. \end{align*} Now $\norm{(K^\omega)^{-1} K^V}_\infty \rightarrow 0$ in the above equation implies that $S^T\hat{\omega} = 0 \Leftrightarrow \hat{\omega} = k_1 1_n$ for some $k_1\in \mathbb{R}$. Consider the $(2n+1)$th to $3n$th rows of the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime}: \begin{align*} K^\text{droop, I} \hat{\omega} - \mathcal{L}_\eta \eta = 0_n. \end{align*} By inserting $\hat{\omega} = k_1 1_n$ and premultiplying the above equation with $1_n^T$, we obtain that $k_1=0$, so $\hat{\omega}=0_n$ and Equation~\eqref{eq:hvdc_coordinated_frequency_objective} of Objective~\ref{obj:1_hvdc_coordinated} is satisfied. This implies that $\eta = k_2 1_n$ for some $k_2\in \mathbb{R}$. Finally we consider the $(n+1)$th to $2n$th rows of the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime}: \begin{align} \frac{1}{V^\text{nom}} K^\omega \hat{\omega} - \Big( \mathcal{L}_R + \frac{K^V}{V^\text{nom}} \Big) \hat{V} + \frac{1}{V^\text{nom}} \mathcal{L}_\phi S \phi'' = 0_n \label{eq:cl_dynamics_vec_delta_int_coordinated_double_prime_n+1:2n} \end{align} Inserting $\hat{\omega}=0_n$ and premultiplying \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime_n+1:2n} with $1_n^T$ yield \begin{align} 1_n^TK^V\hat{V}=0. \label{eq:hvdc_coordinated_voltage_equilibrium} \end{align} Inserting $\hat{\omega}=0_n$ and $\eta = k_2 1_n$ in \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} yields \begin{align} P^\text{gen} = - k_2 K^V(K^\omega)^{-1} K^\text{droop, I} 1_n, \label{eq:hvdc_coordinated_eta_equilibrium} \end{align} where $P^\text{gen} = [P^\text{gen}_1, \dots, P^\text{gen}_n]^T$. It now remains to show that the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} minimizes the cost functions \eqref{eq:hvdc_coordinated_generation_objective} and \eqref{eq:hvdc_coordinated_voltage_objective}. Consider first \eqref{eq:hvdc_coordinated_generation_objective}, with the constraints $P^\text{gen}_i + P_i^{{m}} - P_i^{\text{inj}} = 0 \; \forall i=1, \dots, n$ and $\sum_{i=1}^n P_i^\text{inj} = 0$. By summing the first constraints we obtain $\sum_{i=1}^n P^\text{gen}_i = - \sum_{i=1}^n P_i^{{m}} $. The KKT condition of \eqref{eq:hvdc_coordinated_generation_objective} is \begin{align} F^P P^\text{gen} = - k_3 1_n. \label{eq:hvdc_coordinated_KKT_generation} \end{align} Since $(F^P)^{-1} = K^V(K^\omega)^{-1}K^\text{droop}$, \eqref{eq:hvdc_coordinated_eta_equilibrium} and \eqref{eq:hvdc_coordinated_KKT_generation} are identical for $k_2=k_3$. We conclude that \eqref{eq:hvdc_coordinated_generation_objective} is minimized. Since $P^\text{gen} = - K^V(K^\omega)^{-1} K^\text{droop} \eta = - k_2 K^V(K^\omega)^{-1} K^\text{droop} 1_n$ and $\hat{\omega}=0_n$, premultiplying the first $n$ rows of the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} with $M^{-1}$, and adding to the $(n+1)$th to $2n$th rows premultiplied with $V^\text{nom}E^{-1}$ yields \begin{align*} - V^\text{nom} \mathcal{L}_R \hat{V} - k_2 K^V(K^\omega)^{-1} K^\text{droop} 1_n = P^m. \end{align*} Premultiplying the above equation with $1^T_n$ yields $k_2 = - \sum_{i=1}^n P^m_i \sum_{i=1}^n \frac{K^\omega_i}{K^V_i K^\text{droop}_i}$. Additionally, $\mathcal{L}_R \hat{V}$ is uniquely determined. Now consider \eqref{eq:hvdc_coordinated_voltage_objective}. Note that $P_i^\text{inj}$ and hence $I_i^\text{inj}$, are uniquely determined by \eqref{eq:hvdc_coordinated_generation_objective}. By the equilibrium of \eqref{eq:hvdc_coordinated_voltage}, $\mathcal{L}_R \hat{V} = I^\text{inj}$, where $I^\text{inj} = [I^\text{inj}_1, \dots, I^\text{inj}_n]^T$. Thus, the KKT condition of \eqref{eq:hvdc_coordinated_voltage_objective} is \begin{align} F^V \hat{V} = \mathcal{L}_R r, \label{eq:hvdc_coordinated_KKT_voltage_raw} \end{align} where $r\in \mathbb{R}^n$. Since $\mathcal{L}_R \hat{V}$ is uniquely determined, we premultiply \eqref{eq:hvdc_coordinated_KKT_voltage_raw} with $1_n^T$ and obtain the equivalent condition \begin{align} 1^T_n F^V \hat{V} = 0. \label{eq:hvdc_coordinated_KKT_voltage} \end{align} Since $F^V = K^V$, \eqref{eq:hvdc_coordinated_voltage_equilibrium} and \eqref{eq:hvdc_coordinated_KKT_voltage} are equivalent. Hence \eqref{eq:hvdc_coordinated_voltage_objective} is minimized, so Objective~\ref{obj:1_hvdc_coordinated} is satisfied. \end{pf} \begin{remark} Corollary \ref{cor:hvdc_coordinated_equilibrium} provides insight in choosing the controller gains of \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} and \eqref{eq:voltage_control_secondary}, to satisfy Objective~\ref{obj:1_hvdc_coordinated}. \end{remark} \section{Simulations} \label{sec:simulations} In this section, simulations are conducted on a test system to validate the performance of the proposed controllers. The simulation was performed in Matlab, using a dynamic phasor approach based on \citep{demiray2008}. The test system is illustrated in Figure~\ref{fig:testsystem}. \begin{figure}[ht] \def1.25\columnwidth{1.25\columnwidth} \begin{center} \input{ACandDCGrid2.pdf_tex} \captionsetup{width=\linewidth} \caption{MTDC test system, consisting of a 6-terminal MTDC grid. Each terminal is connected to an IEEE 14 bus AC grid, sketched as octagons.} \label{fig:testsystem} \end{center} \end{figure} The line parameters of the MTDC grid are given in Table~\ref{tab:HVDCgridParameter}. Note that we in the simulation also consider the inductances $L_{ij}$ and capacitances $C_{ij}$ of the HVDC lines. The capacitances of the terminals are assumed to be given by $C_i=0.375\times 10^{-3}$ p.u., and are chosen uniformly for all VSC stations. The AC grid parameters were obtained from \citep{Milano}. The generators are modeled as a 6{th} order machine model controlled by an automatic voltage controller and a governor \citep{kundur1994power}. Each load in the grid is assumed to be equipped with an ideal power controller. \begin{table}[htb] \captionsetup{width=\linewidth} \caption{HVDC grid line parameters} \label{tab:HVDCgridParameter} \begin{center} \begin{tabular}{ccccc} $i$&$j$&$R_{ij}$ [p.u.]&$L_{ij}$ [$10^{-3}$ p.u.]& $C_{ij}$ [p.u.]\\ \hline 1 & 2 & 0.0586 & 0.2560 &0.0085 \\ 1 & 3 & 0.0586 & 0.2560 &0.0085\\ 2 & 3 & 0.0878 & 0.3840 &0.0127\\ 2 & 4 & 0.0586 & 0.2560 &0.0085\\ 2 & 5 & 0.0732 & 0.3200 &0.0106\\ 2 & 6 & 0.1464 & 0.6400 &0.0212\\ 3 & 4 & 0.0586 & 0.2560 &0.0085\\ 3 & 5 & 0.1464 & 0.6400 &0.0212\\ 4 & 5 & 0.0732 & 0.3200 &0.0106\\ 5 & 6 & 0.1464 & 0.6400 &0.0212 \end{tabular} \end{center} \end{table} \begin{table}[htb] \begin{center} \captionsetup{width=\linewidth} \caption{Controller parameters} \label{tab:ControllerParameter} \begin{tabular}{ccccccc} $K^{\omega}_i$ &$K^{V}_i$&$K^\text{droop}_i$& $K^\text{droop, I}_i$ &$c^{\eta}_{ij}$ &$c^{\phi}_{ij}$ & $\gamma$ \\ \hline 1000 & 100 & 9 & 3.35 & ${5}/{R_{ij}}$ & ${15}/{R_{ij}}$ & 0 \end{tabular} \end{center} \end{table} The test grid was controlled with the controllers \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} and \eqref{eq:voltage_control_secondary}, with parameters given in Table~\ref{tab:ControllerParameter}. The communication network of \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} and \eqref{eq:voltage_control_secondary} is illustrated by the dashed lines in Figure~\ref{fig:testsystem}. Note that we have set $\gamma=0$, and that Theorem~\ref{th:stability_A_coordinated} thus does not guarantee stability of the equilibrium. The resulting matrix of the closed-loop system is however verified to be Hurwitz. At time $t=1$ the output of one generator in area 1 was reduced by $0.2$ p.u., simulating a fault. Figure~\ref{fig:ACFrequency} shows the average frequencies of the AC grids, while Figure~\ref{fig:DCVoltage} shows the DC voltages of the terminals. Figure~\ref{fig:Generators} shows the total increase of the generated power within each AC area. It can be noted that immediately after the fault, the average frequency of the AC area of the fault drops. The frequency drop is followed by a voltage drop in all terminals, and a subsequent frequency drop in the remaining AC areas. The frequencies converge to the nominal frequency, while the voltages converge to their new stationary values after approximately $30$ s. We note that the frequencies are restored to the nominal frequency, as predicted by Corollary~\ref{cor:hvdc_coordinated_equilibrium}. Furthermore the incremental generated power of the AC areas converge to the same value, as a consequence of Corollary~\ref{cor:hvdc_coordinated_equilibrium} and the controller parameters being equal. \begin{figure}[t!] \centering \tikzsetnextfilename{Frequencies} \begin{tikzpicture} \begin{axis} [cycle list name=Necsys, xlabel={$t$ [s]}, ylabel={$\omega(t)$ [p.u.]}, xmin=0, xmax=40, xtick={0,5,...,40}, ymin=0.99935, ymax=1.00005, yticklabel style={/pgf/number format/.cd, fixed, precision=4}, grid=major, height=6.5cm, width=0.96\columnwidth, legend cell align=left, legend pos= south east, ] \foreach \x in {1,2,3,4,5,6}{ \addplot table[x index=0,y index=\x,col sep=tab]{PlotData_coordinated/Necsys_Frequency.txt}; } \legend{Area 1, Area 2, Area 3, Area 4, Area 5, Area 6} \end{axis} \end{tikzpicture} \caption{Average frequencies in the AC areas.} \label{fig:ACFrequency} \end{figure} \begin{figure}[htb] \centering \tikzsetnextfilename{DCVoltages} \begin{tikzpicture} \begin{axis}[cycle list name=Necsys, xlabel={$t$ [s]}, ylabel={$V(t)$ [p.u.]}, xmin=0, xmax=40, xtick={0,5,...,40}, ymin=0.982, ymax=1.004, yticklabel style={/pgf/number format/.cd, fixed, precision=4}, grid=major, height=6.5cm, width=0.975\columnwidth, legend pos= south east, ] \foreach \x in {1,2,3,4,5,6}{ \addplot table[x index=0,y index=\x,col sep=tab] {PlotData_coordinated/Necsys2_Voltage.txt}; } \legend{Area 1, Area 2, Area 3, Area 4, Area 5, Area 6} \end{axis} \end{tikzpicture} \caption{DC terminal voltages.} \label{fig:DCVoltage} \end{figure} \begin{figure}[t!] \centering \tikzsetnextfilename{Generators} \begin{tikzpicture} \begin{axis}[cycle list name=Necsys, xlabel={$t$ [s]}, ylabel={$P^\text{gen}(t)$ [p.u.]}, xmin=0, xmax=40, xtick={0,5,...,40}, ymin=-0.005, ymax=0.076, scaled ticks=false, tick label style={/pgf/number format/fixed, /pgf/number format/precision=2}, ytick={0,0.02,...,0.101}, grid=major, height=6.5cm, width=0.975\columnwidth, legend pos= north east, ] \foreach \x in {1,2,3,4,5,6}{ \addplot table[x index=0,y index=\x,col sep=tab] {PlotData_coordinated/Necsys2_Generators.txt}; } \legend{Area 1, Area 2, Area 3, Area 4, Area 5, Area 6} \end{axis} \end{tikzpicture} \caption{Increase of generated power in the AC areas.} \label{fig:Generators} \end{figure} \section{Discussion and Conclusions} \label{sec:discussion} In this paper we have studied a distributed secondary controller for sharing frequency control reserves of asynchronous AC systems connected through an MTDC system. Under certain conditions, the proposed controller stabilizes the closed-loop system consisting of the interconnected AC systems and the MTDC grid. The frequencies in all AC grids are shown to converge to the nominal frequency. Furthermore, quadratic cost functions of the voltage deviations of the MTDC terminals and of the generated power, are minimized asymptotically. Finally, the results were validated on a six-terminal MTDC system with connected AC systems. \section{Introduction} Power transmission over long distances with low losses is one of the main challenges in today's power transmission systems. As the share of renewables rises, so does the need to balance generation and consumption mismatches, often over large geographical areas. Due to the high resistive losses in AC cables, high-voltage direct current (HVDC) power transmission is a commonly used technology for long-distance power transmission. The higher investment cost of an HVDC transmission system compared to an AC transmission system is compensated by the lower resistive losses for sufficiently long distances \citep{melhem2013electricity}. The break-even point, i.e., when the total construction and operation costs of overhead HVDC and AC lines equal, is typically 500--800 km \citep{padiyar1990hvdc}. However, for cables, the break-even point is typically less than 50 km \citep{Hertem2010technical}. Increased use of HVDC technologies for electrical power transmission suggests that future HVDC transmission systems are likely to consist of multiple terminals connected by several HVDC transmission lines \citep{Haileselassie2013Power}. Such systems are referred to as multi-terminal HVDC (MTDC) systems in the literature. Many existing AC grids are connected through legacy HVDC links, which are typically used for bulk power transfer between AC areas, rather than balancing generation mismatches. The fast operation of the DC converters however also enables frequency regulation of one of the connected AC grids through the HVDC link. One example is the frequency regulation of the island of Gotland in Sweden, which is connected to the Nordic grid through an HVDC cable \citep{axelsson2001gotland}. However, since the Nordic grid has orders of magnitudes higher inertia than the AC grid of Gotland, the influence of the frequency regulation on the Nordic grid is negligible. By connecting multiple AC grids by an MTDC system, the frequency regulation reserves in each AC grid can be shared with connected AC grids, which reduces the need for frequency regulation reserves in the individual AC systems \citep{li2008frequency}. In recent years, this idea has gained increasing interest in the literature. \cite{dai2011voltage} and \cite{silva2012provision} employ several controllers with decentralized structure to share frequency control reserves. In \citep{silva2012provision} no stability analysis of the closed-loop system is performed, whereas \citep{dai2011voltage} guarantees stability provided that the connected AC areas have identical parameters and the voltage dynamics of the HVDC system are neglected. In \citep{taylordecentralized}, an optimal decentralized controller for AC systems connected by HVDC systems is derived. In contrast to the aforementioned references, \cite{andreasson2014distributed} consider the dynamics of connected AC systems as well as the dynamics of the MTDC system. By connecting the AC areas with a communication network, the performance of the controller can be further improved. \cite{dai2010impact} consider a distributed controller relying on a communication network to share frequency control reserves of asynchronous AC transmission systems connected through an MTDC system. However, the controller requires a slack bus to control the DC voltage, and is thus only able to share the generation reserves of the non-slack AC areas. Another distributed controller is proposed by \cite{dai2013voltage}. Stability is guaranteed, and the need for a slack bus is eliminated. The voltage dynamics of the MTDC system are however neglected. Moreover the implementation of the controller requires every controller to access measurements of the DC voltages of all other MTDC terminals. \citep{andreasson2014distributedSecondary} propose a distributed secondary generation controller. In contrast to the aforementioned references, the MTDC dynamics are explicitly modeled, and the DC voltages are controlled in addition to the frequencies. Despite the coordination through a communication network, the distributed controller by \cite{andreasson2014distributedSecondary} fails in eliminating the static errors of the frequencies. In this paper we address this issue by introducing a distributed secondary controller also for the power fed into the MTDC grid from the AC grids. This allows us to eliminate any static errors in the frequency deviations. Furthermore, quadratic cost functions of the power generation and the voltage deviations are minimized. The remainder of this paper is organized as follows. In Section \ref{sec:model}, the system model and control objectives are defined. In Section \ref{sec:secondary_frequency_control}, a distributed secondary frequency controller for sharing frequency control and restoration reserves is presented, and is shown to satisfy the control objectives. In Section \ref{sec:simulations}, simulations of the controller on a six-terminal MTDC test system are provided, showing the effectiveness of the proposed controller. The paper ends with concluding remarks in Section~\ref{sec:discussion}. \section{Model and problem setup} \label{sec:model} \subsection{Notation} \label{subsec:prel} Let $\mathcal{G}$ be a static, undirected graph. Denote by $\mathcal{V}=\{ 1,\hdots, n \}$ the vertex set of $\mathcal{G}$, and by $\mathcal{E}=\{ 1,\hdots, m \}$ the edge set of $\mathcal{G}$. Let $\mathcal{N}_i$ be the set of neighboring vertices to $i \in \mathcal{V}$. Denote by $\mathcal{B}$ the vertex-edge incidence matrix of $\mathcal{G}$, and let $\mathcal{\mathcal{L}_W}=\mathcal{B}W\mathcal{B}^T$ be the weighted Laplacian matrix of $\mathcal{G}$, with edge-weights given by the elements of the diagonal matrix $W$. We denote the space of real-valued $n\times m$-valued matrices by $\mathbb{R}^{n\times m}$. Let $\mathbb{C}^-$ denote the open left half complex plane, and $\bar{\mathbb{C}}^-$ its closure. We denote by $c_{n\times m}$ a vector or matrix of dimension $n\times m$, whose elements are all equal to $c$. For simplifying notation, we write $c_n$ for $c_{n\times 1}$. For a symmetric matrix $A$, $A>0 \;(A\ge 0)$ is used to denote that $A$ is positive (semi) definite. $I_{n}$ denotes the identity matrix of dimension $n$. For simplicity, we will often drop the notion of time dependence of variables, i.e., $x(t)$ will be denoted $x$. Let $\norm{\cdot}_\infty$ denote the maximal absolute value of the elements of a vector. \subsection{Model} \label{subsec:model} We will give here a unified model for an MTDC system interconnected with several mutually asynchronous AC systems. We consider an MTDC transmission system consisting of $n$ converters, denoted $i=1, \dots, n$, each connected to an AC system, i.e., there are no pure DC nodes of the MTDC grid. The converters are assumed to be connected by an MTDC transmission grid. The dynamics of converter $i$ is assumed to be given by \begin{align} \begin{aligned} C_i \dot{V}_i &= -\sum_{j\in \mathcal{N}_i} \frac{1}{R_{ij}}(V_i -V_j) + I_i^{\text{inj}} , \end{aligned} \label{eq:hvdc_coordinated_voltage} \end{align} where $V_i$ is the voltage of converter $i$, $C_i>0$ is its capacitance, and $I_i^{\text{inj}}$ is the injected current from an AC grid connected to the DC converter. The constant $R_{ij}$ denotes the resistance of the HVDC transmission line connecting the converters $i$ and $j$. The graph corresponding to the HVDC line connections is assumed to be connected. The AC system is assumed to consist of a single generator which is connected to the corresponding DC converter, representing an aggregate model of the AC grid. The dynamics of the AC system are given by the swing equation \citep{machowski2008power}: \begin{align} m_i \dot{\omega}_i &= P^\text{gen}_i + P_i^{{m}} - P_i^{\text{inj}}, \label{eq:frequency} \end{align} where $m_i>0$ is its moment of inertia. The constant $P^\text{gen}_i$ is the generated power, $P^m_i$ is the power load and $P_i^{\text{inj}}$ is the power injected to the DC system through converter $i$, respectively. The powers are all assumed to be deviations from a nominal operation point. The control objective can now be stated as follows. \begin{objective} \label{obj:1_hvdc_coordinated} The frequency deviations are asymptotically equal to zero, i.e., \begin{align} \lim_{t\rightarrow \infty} \omega_i(t)-\omega^{\text{ref}} = 0 \quad i = 1, \dots, n, \label{eq:hvdc_coordinated_frequency_objective} \end{align} where $\omega^{\text{ref}}$ is the nominal frequency. The total quadratic cost of the power generation is minimized asymptotically, i.e., $\lim_{t\rightarrow \infty} P_i^\text{gen} = P_i^{\text{gen}*}, \forall i=1, \dots, n$, where \begin{align} [P_1^{\text{gen}*}, \dots, P_n^{\text{gen}*}] = \argmin_{P_1, \dots, P_n} \frac 12 \sum_{i=1}^n f^P_i P_i^2 \label{eq:hvdc_coordinated_generation_objective} \end{align} subject to $P^\text{gen}_i + P_i^{{m}} - P_i^{\text{inj}} = 0, \; \forall i=1, \dots, n$ and $\sum_{i=1}^n P_i^\text{inj} = 0$, i.e., power balance both in the AC grids and in the MTDC grid, in the absence of power losses. The positive constants $f^P_i$ represent the local cost of generating power. Finally, the voltages are such that the a quadratic cost function of the voltage deviations is minimized asymptotically, i.e., $\lim_{t\rightarrow \infty} V_i = V_i^*, \forall i=1, \dots, n$, where \begin{align} [V_1^*, \dots, V_n^*] = \argmin_{V_1, \dots, V_n} \frac 12 \sum_{i=1}^n f^V_i (V_i - V_i^\text{ref})^2 \label{eq:hvdc_coordinated_voltage_objective} \end{align} subject to \eqref{eq:hvdc_coordinated_frequency_objective}--\eqref{eq:hvdc_coordinated_generation_objective}, and where the $f^V_i$ is a positive constant reflecting the local cost of voltage deviations, and $V_i^\text{ref}$ is the nominal voltage of converter $i$. \end{objective} \begin{remark} The minimization of \eqref{eq:hvdc_coordinated_generation_objective} is equivalent to power sharing, where the generated power of AC area $i$ is asymptotically inverse proportional to the cost $f_i^P$. The cost $f_i^P$ can be chosen to reflect the available generation capacity of area $i$. \end{remark} \begin{remark} It is in general not possible that $\lim_{t\rightarrow \infty} V_i(t) = V_i^\text{ref} \; \forall i=1, \dots, n$, since this does not allow for the currents between the HVDC converters to change by \eqref{eq:hvdc_coordinated_voltage}. Note that the optimal solution to \eqref{eq:hvdc_coordinated_generation_objective} fixes the relative voltages, leaving only the ground voltage as a decision variable of \eqref{eq:hvdc_coordinated_voltage_objective}. Note also that the reference voltages $V_i^\text{ref}, \; i=1, \dots, n$, are generally not uniform, as is the reference frequency $\omega^\text{ref}$. \end{remark} \begin{remark} Note that Objective \ref{obj:1_hvdc_coordinated} does not include constraints of, e.g., generation and line capacities. This requires that the perturbations from the operating point are sufficiently small, to guarantee that these constraints are not violated. In case of a large disturbance, e.g., a fault, the constraints might be violated. \end{remark} \section{Coordinated secondary frequency control} \label{sec:secondary_frequency_control} \subsection{Controller structure} \label{subsec:controlelr_structure} In this section we propose a distributed secondary frequency controller. In addition to the generation controller proposed by \cite{andreasson2014distributedSecondary}, we also propose a secondary controller for the voltage injections into the HVDC grid. The distributed generation controller of the AC systems, which was first given by \cite{andreasson2014distributedSecondary}, is given by \begin{align} P^\text{gen}_i &=- K_i^{\text{droop}} (\omega_i-\omega^{\text{ref}}) - \frac{K^V_i}{K^\omega_i} K^\text{droop, I}_i \eta_i \nonumber \\ \dot{\eta}_i &= K_i^{\text{droop,I}}(\omega_i-\omega^{\text{ref}}) - \sum_{j\in \mathcal{N}_i} c^\eta_{ij} (\eta_i-\eta_j), \label{eq:hvdc_coordinated_droop_control_secondary_distributed} \end{align} where $K_i^\text{droop}$ and $K_i^\text{droop, I}$ are positive controller parameters. Moreover $c^\eta_{ij} = c^\eta_{ji}>0$, i.e., the communication graph is undirected. The above controller can be interpreted as a distributed PI-controller, with a distributed consensus filter acting on the integral states $\eta_i$. The proposed converter controllers governing the incremental power injections from the AC systems into the HVDC grid are given by \begin{align} \label{eq:voltage_control_secondary} P_i^{\text{inj}} &= K_i^{{\omega}} (\omega_i - \omega^{\text{ref}}) + K_i^{{V}}(V_i^{\text{ref}}-V_i) \nonumber \\ & \;\;\;\;+ \sum_{j\in \mathcal{N}_i} c^\phi_{ij} (\phi_i - \phi_j) \nonumber \\ \dot{\phi}_i &= \frac{K^\omega_i}{K^V_i} \omega_i - \gamma \phi_i, \end{align} where $K_i^V$ and $K_i^\omega$ are positive controller parameters, $\gamma>0$ and $c^\phi_{ij} = c^\phi_{ji}>0$. The converter controller \eqref{eq:voltage_control_secondary} can be interpreted as an emulation of an AC network between the isolated AC areas. The auxiliary controller variables $\phi_i$ can be thought of as the phase angles of AC area $i$, governing the power transfer between the areas, if these were connected by AC lines rather than an MTDC grid. In contrast to an AC system, however, the power is fed into the HVDC grid instead of being directly transferred to the connected AC areas. The HVDC converter is assumed to be perfect and instantaneous, i.e., injected power on the AC side is immediately and losslessly converted to DC power. Furthermore, the dynamics of the converter are ignored, implying that the converter is assumed to track the output of controller \eqref{eq:voltage_control_secondary} instantaneously. This assumption is reasonable due to the dynamics of the converter typically being orders of magnitudes faster than the primary frequency control dynamics of the AC system \citep{kundur1994power}. The relation between the injected HVDC current and the injected AC power is thus given by \begin{align} V_iI_i^{\text{inj}} = P_i^{\text{inj}}. \label{eq:power-current_nonlinear} \end{align} By assuming $V_i=V^{\text{nom}}\; \forall i=1, \dots, n$, where $V^{\text{nom}}$ is a global nominal voltage, we obtain \begin{align} V^{\text{nom}}I_i^{\text{inj}} = P_i^{\text{inj}}. \label{eq:power-current} \end{align} \subsection{Stability analysis} \label{sec:secondary_frequency_control_stability} We are now ready to analyze the stability of the closed-loop system. Define the stacked state vectors as $\hat{\omega}=\omega - \omega^\text{ref}1_n$ and $\hat{V}=V - V^\text{ref}$, where $\omega=[\omega_1, \dots, \omega _n]^T$, $V=[V_1, \dots, V_n]^T$, $V^\text{ref} = [V_1^\text{ref}, \dots, V_n^\text{ref}]^T$, $\eta = [\eta_1, \dots, \eta_n]^T$ and $\phi = [\phi_1, \dots, \phi_n]$. Combining the voltage dynamics \eqref{eq:hvdc_coordinated_voltage}, the frequency dynamics \eqref{eq:frequency} and the generation control \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed}, the converter controller \eqref{eq:voltage_control_secondary} and the power-current relationship \eqref{eq:power-current}, we obtain the closed-loop dynamics \begin{IEEEeqnarray}{lcl} \dot{\hat{\omega}} &=& M \Big(- (K^{\text{droop}} + K^\omega) \hat{\omega} + K^V \hat{V} \nonumber \\ && - {K^V}(K^\omega)^{-1} K^\text{droop, I} \eta - \mathcal{L}_\phi \phi + P^{{m}} \Big) \nonumber \\ \dot{\hat{V}} &=& \frac{1}{V^{\text{nom}}}E {K}^\omega \hat{\omega} -E\left(\mathcal{L}_R + \frac{K^V}{V^{\text{nom}}} \right) \hat{V} + \frac{1}{V^\text{nom}} E \mathcal{L}_\phi \phi \nonumber \\ \dot{\eta} &=& K^{\text{droop,I}}\hat{\omega} - \mathcal{L}_\eta \eta \nonumber \\ \dot{\phi} &=& (K^V)^{-1}{K}^\omega \hat{\omega} -\gamma \phi, \label{eq:cl_dynamics_vec_delta_int_coordinated} \end{IEEEeqnarray} where $M=\diag({m_1}^{-1}, \hdots , {m_n}^{-1})$ is a matrix of inverse generator inertia, $E=\diag(C_1^{-1}, \dots, C_n^{-1})$ is a matrix of electrical elastances, $\mathcal{L}_R$ is the weighted Laplacian matrix of the MTDC grid with edge-weights $1/R_{ij}$, $\mathcal{L}_\eta$ and $\mathcal{L}_\phi$ are the weighted Laplacian matrices of the communication graphs with edge-weights $c^\eta_{ij}$ and $c^\phi_{ij}$, respectively, and $P^m=[P^m_1, \dots, P^m_n]^T$. We define the diagonal matrices of the controller gains as $K^\omega = \diag(K^\omega_1, \dots, K^\omega_n)$, etc. Let $y=[\hat{\omega}^T, \hat{V}^T]^T$ define the output of \eqref{eq:cl_dynamics_vec_delta_int_coordinated}. Clearly the linear combination $1_n^T\phi$ is unobservable and marginally stable with respect to the dynamics \eqref{eq:cl_dynamics_vec_delta_int_coordinated}, as it lies in the nullspace of $\mathcal{L}_\phi$. In order to facilitate the stability analysis, we will perform a state-transformation to this unobservable mode. Consider the following state-transformation: \begin{align} \phi' = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n^T \\ S^T \end{bmatrix} \phi \qquad \phi = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n & S \end{bmatrix} \phi' \label{eq:transformation_phi} \end{align} where $S$ is an $n\times(n-1)$ matrix such that $\left[\frac{1}{\sqrt{n}}1_n \; S\right]$ is orthonormal. By applying the state-transformation \eqref{eq:transformation_phi} to \eqref{eq:cl_dynamics_vec_delta_int_coordinated}, we obtain dynamics where it can be shown that the state $\phi_1'$ is unobservable with respect to the defined output. Hence, omitting $\phi_1'$ does not affect the output dynamics. Thus, we define $\phi''=[\phi_2', \dots, \phi_n']$, and obtain the dynamics \begin{IEEEeqnarray}{lcl} \dot{\hat{\omega}} &=& M \Big(- (K^{\text{droop}} + K^\omega) \hat{\omega} + K^V \hat{V} \nonumber \\ && - {K^V}(K^\omega)^{-1} K^\text{droop, I} \eta - \mathcal{L}_\phi S \phi'' + P^{{m}} \Big) \nonumber \\ \dot{\hat{V}} &=& \frac{1}{V^{\text{nom}}}E {K}^\omega \hat{\omega} -E\left(\mathcal{L}_R + \frac{K^V}{V^{\text{nom}}} \right) \hat{V} + \frac{1}{V^\text{nom}} E \mathcal{L}_\phi S\phi'' \nonumber \\ \dot{\eta} &=& K^{\text{droop,I}}\hat{\omega} - \mathcal{L}_\eta \eta \nonumber \\ \dot{\phi}'' &=& S^T(K^V)^{-1}{K}^\omega \hat{\omega} -\gamma \phi''. \label{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} \end{IEEEeqnarray} Provided that the system matrix, denoted $A$, of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} is full-rank, then \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} has a unique equilibrium. Denote this equilibrium $x_0=[\omega_0^T, V_0^T, \eta_0^T, \phi_0''^T]^T$. Define $\bar{x}\triangleq [\bar{\omega}^T, \bar{V}^T, \bar{\eta}^T, \bar{\phi}^T]^T =[\hat{\omega}^T, \hat{V}^T, \eta^T, \phi''^T]^T - [\omega_0^T, V_0^T, \eta_0^T, \phi_0''^T]^T$. Hence, \begin{align} \dot{\bar{x}} = A \bar{x}, \label{eq:dynamics_A_coordinated_shifted} \end{align} with the origin as its unique equilibrium. We are now ready to show the main stability result of this section. We first make the following assumptions. \begin{assumption} \label{ass:L_phi_coordinated} The Laplacian matrix satisfies $\mathcal{L}_\phi = k_\phi \mathcal{L}_R$. \end{assumption} Assumption \ref{ass:L_phi_coordinated} can be interpreted as the emulated AC dynamics of \eqref{eq:voltage_control_secondary} having the same susceptance ratios as the conductance ratios of the HVDC lines. \begin{assumption} \label{ass:gamma_coordinated} The gain $\gamma$ satisfies $\gamma > \frac{k_\phi}{4V^\text{nom}}$. \end{assumption} Assumption \ref{ass:gamma_coordinated} lower bounds for the damping coefficient $\gamma$. \begin{theorem} \label{th:stability_A_coordinated} If Assumptions \ref{ass:L_phi_coordinated} and \ref{ass:gamma_coordinated} hold, the origin of \eqref{eq:dynamics_A_coordinated_shifted} is globally asymptotically stable. \end{theorem} \begin{pf} Consider the Lyapynov function candidate: \begin{align} W(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) &= \frac 12 \bar{\omega}^T K^\omega (K^V)^{-1} M^{-1}\bar{\omega} + \frac{V^\text{nom}}{2} \bar{V}^T C \bar{V} \nonumber \\ &\;\;\;\; + \frac 12 \bar{\eta}^T \bar{\eta} + \frac 12 \bar{\phi}^T S^T \mathcal{L}_\phi S \bar{\phi}, \label{eq:lyap_hvdc_coordinated} \end{align} where $C=\diag(C_1, \dots, C_n)$. Clearly $W(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi})$ is positive definite and radially unbounded. Differentiating \eqref{eq:lyap_hvdc_coordinated} along trajectories of \eqref{eq:dynamics_A_coordinated_shifted}, we obtain \begin{align*} \dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) &= \bar{\omega}^T K^\omega (K^V)^{-1} M^{-1}\dot{\bar{\omega}} + V^\text{nom} \bar{V}^T C \dot{\bar{V}} \nonumber \\ &\;\;\;\; + \bar{\eta}^T \dot{\bar{\eta}} + \bar{\phi}^T S^T \mathcal{L}_\phi S \dot{\bar{\phi}} \\ &= \bar{\omega}^T \big( -K^\omega (K^V)^{-1}(K^\omega + K^\text{droop})\bar{\omega} \\ &\;\;\;\; + K^\omega \bar{V} - K^\text{droop, I} \bar{\eta} - K^\omega (K^V)^{-1} \mathcal{L}_\phi S \bar{\phi} \big) \\ &\;\;\;\; + \bar{V}^T \Big( K^\omega \bar{\omega} - (V^\text{nom}\mathcal{L}_R {+} K^V)\bar{V} + \mathcal{L}_\phi S \bar{\phi} \Big) \\ &\;\;\;\; + \bar{\eta}^T \Big( K^\text{droop, I} \bar{\omega} - \mathcal{L}_\eta \eta \Big) \\ &\;\;\;\; + \bar{\phi}^T S^T \mathcal{L}_\phi S \Big( S^T (K^V)^{-1}K^\omega \bar{\omega} - \gamma \bar{\phi} \Big) \\ &= -\bar{\omega}^T \big( -K^\omega (K^V)^{-1}(K^\omega + K^\text{droop})\bar{\omega} \\ &\;\;\;\; + 2 \bar{\omega}^T K^\omega \bar{V} - \bar{V}^T (V^\text{nom}\mathcal{L}_R + K^V)\bar{V} \\ &\;\;\;\; + \bar{V}^T \mathcal{L}_\phi S \bar{\phi} - \bar{\eta}^T \mathcal{L}_\eta \bar{\eta} - \gamma \bar{\phi}^T S^T \mathcal{L}_\phi S \bar{\phi}, \end{align*} since $SS^T = I_n - \frac 1n 1_{n\times n}$, so $\mathcal{L}_\phi SS^T = \mathcal{L}_\phi$. By defining \begin{align*} \bar{V}' = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n^T \\ S^T \end{bmatrix} & \bar{V} \qquad \bar{V} = \begin{bmatrix} \frac{1}{\sqrt{n}} 1_n & S \end{bmatrix} \bar{V}', \end{align*} we obtain $\bar{V}^T\mathcal{L}_R \bar{V} = \bar{V}''^T S^T\mathcal{L}_R S \bar{V}'' $ and $\bar{V}^T\mathcal{L}_\phi S \bar{\phi} = \bar{V}''^T S^T\mathcal{L}_\phi S \bar{\phi}$, where $\bar{V}'' = [\bar{V}'_2, \dots, \bar{V}'_n]^T$. By invoking Assumption~\ref{ass:L_phi_coordinated}, we obtain \begin{align*} &\dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) \\ &= - \begin{bmatrix} \bar{\omega}^T & \bar{V}^T \end{bmatrix} \underbrace{\begin{bmatrix} K^\omega (K^V)^{-1}(K^\omega + K^\text{droop}) & -K^\omega \\ -K^\omega & K^V \end{bmatrix}}_{\triangleq Q_1} \begin{bmatrix} \bar{\omega} \\ \bar{V} \end{bmatrix} \\ &\;\;\;\; - \begin{bmatrix} \bar{V}''^T & \bar{\phi}^T \end{bmatrix} \underbrace{\begin{bmatrix} V^\text{nom} S^T\mathcal{L}_R S & -\frac{k_\phi}{2} S^T\mathcal{L}_R S \\ -\frac{k_\phi}{2} S^T\mathcal{L}_R S & \gamma k_\phi S^T\mathcal{L}_R S \end{bmatrix}}_{\triangleq Q_2} \begin{bmatrix} \bar{V}'' \\ \bar{\phi} \end{bmatrix} \\ &\;\;\;\; - \bar{\eta}^T \mathcal{L}_\eta \bar{\eta}. \end{align*} By applying the Schur complement condition for positive definiteness, we see that $Q_1$ is positive definite iff \begin{eqnarray*} K^\omega (K^V)^{-1}(K^\omega + K^\text{droop}) - K^\omega (K_V)^{-1} K^\omega \\ = K^\omega (K^V)^{-1} K^\text{droop} > 0. \end{eqnarray*} Hence $Q_1$ is positive definite. By applying the same argument to $Q_2$, $Q_2$ is positive definite iff \begin{align*} \Bigg( \gamma k_\phi- \frac{k_\phi^2}{4 V^\text{nom}} \Bigg) S^T \mathcal{L}_R S > 0. \end{align*} Clearly the above matrix inequality holds under Assumption~\ref{ass:gamma_coordinated}, since $S^T \mathcal{L}_R S \ge 0$, and $Sx\ne k 1_n$ for $k\ne 0$. Thus $\dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) \le 0$, and the set where $W(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi})$ is non-decreasing is given by \begin{align*} G &= \{ (\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) | \dot{W}(\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) =0 \} \\ &= \{ (\bar{\omega}, \bar{V}, \bar{\eta}, \bar{\phi}) | \bar{\eta} = k 1_n \}, \end{align*} for any $k\in \mathbb{R}$. Clearly the largest invariant set in $G$ is the origin, and thus $k=0$. By LaSalle's theorem, the origin is globally asymptotically stable under the dynamics \eqref{eq:dynamics_A_coordinated_shifted}. \end{pf} We now turn our attention to the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime}. If the stability condition in Theorem~\ref{th:stability_A_coordinated} is met, then the equilibrium is stable, and $A$ is thus Hurwitz. Hence \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} must have a unique equilibrium. In the following corollary we show that the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} satisfies Objective~\ref{obj:1_hvdc_coordinated}. \begin{corollary} \label{cor:hvdc_coordinated_equilibrium} Let Assumption~\ref{ass:L_phi_coordinated} hold and let $\gamma$, $k_\phi$ be given such that Assumption~\ref{ass:gamma_coordinated} holds. Let $K^V, K^\omega$ and $K^\text{droop}$ be such that $(F^P)^{-1} = K^V(K^\omega)^{-1}K^\text{droop}$ and $F^V=K^V$, where $F^P = \diag(f^P_1, \dots, f^P_n)$ and $F^V = \diag(f^V_1, \dots, f^V_n)$. Then Objective~\ref{obj:1_hvdc_coordinated} is satisfied in the limit when $\norm{(K^\omega)^{-1} K^V}_\infty \rightarrow 0$. \end{corollary} \begin{pf} By Theorem~\ref{th:stability_A_coordinated}, \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} has a unique and stable equilibrium. The last $n-1$ rows of the equilibrium imply \begin{align*} S^T(K^V)^{-1}K^\omega \hat{\omega} - \gamma \phi'' = 0_{n-1}. \end{align*} Now $\norm{(K^\omega)^{-1} K^V}_\infty \rightarrow 0$ in the above equation implies that $S^T\hat{\omega} = 0 \Leftrightarrow \hat{\omega} = k_1 1_n$ for some $k_1\in \mathbb{R}$. Consider the $(2n+1)$th to $3n$th rows of the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime}: \begin{align*} K^\text{droop, I} \hat{\omega} - \mathcal{L}_\eta \eta = 0_n. \end{align*} By inserting $\hat{\omega} = k_1 1_n$ and premultiplying the above equation with $1_n^T$, we obtain that $k_1=0$, so $\hat{\omega}=0_n$ and Equation~\eqref{eq:hvdc_coordinated_frequency_objective} of Objective~\ref{obj:1_hvdc_coordinated} is satisfied. This implies that $\eta = k_2 1_n$ for some $k_2\in \mathbb{R}$. Finally we consider the $(n+1)$th to $2n$th rows of the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime}: \begin{align} \frac{1}{V^\text{nom}} K^\omega \hat{\omega} - \Big( \mathcal{L}_R + \frac{K^V}{V^\text{nom}} \Big) \hat{V} + \frac{1}{V^\text{nom}} \mathcal{L}_\phi S \phi'' = 0_n \label{eq:cl_dynamics_vec_delta_int_coordinated_double_prime_n+1:2n} \end{align} Inserting $\hat{\omega}=0_n$ and premultiplying \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime_n+1:2n} with $1_n^T$ yield \begin{align} 1_n^TK^V\hat{V}=0. \label{eq:hvdc_coordinated_voltage_equilibrium} \end{align} Inserting $\hat{\omega}=0_n$ and $\eta = k_2 1_n$ in \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} yields \begin{align} P^\text{gen} = - k_2 K^V(K^\omega)^{-1} K^\text{droop, I} 1_n, \label{eq:hvdc_coordinated_eta_equilibrium} \end{align} where $P^\text{gen} = [P^\text{gen}_1, \dots, P^\text{gen}_n]^T$. It now remains to show that the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} minimizes the cost functions \eqref{eq:hvdc_coordinated_generation_objective} and \eqref{eq:hvdc_coordinated_voltage_objective}. Consider first \eqref{eq:hvdc_coordinated_generation_objective}, with the constraints $P^\text{gen}_i + P_i^{{m}} - P_i^{\text{inj}} = 0 \; \forall i=1, \dots, n$ and $\sum_{i=1}^n P_i^\text{inj} = 0$. By summing the first constraints we obtain $\sum_{i=1}^n P^\text{gen}_i = - \sum_{i=1}^n P_i^{{m}} $. The KKT condition of \eqref{eq:hvdc_coordinated_generation_objective} is \begin{align} F^P P^\text{gen} = - k_3 1_n. \label{eq:hvdc_coordinated_KKT_generation} \end{align} Since $(F^P)^{-1} = K^V(K^\omega)^{-1}K^\text{droop}$, \eqref{eq:hvdc_coordinated_eta_equilibrium} and \eqref{eq:hvdc_coordinated_KKT_generation} are identical for $k_2=k_3$. We conclude that \eqref{eq:hvdc_coordinated_generation_objective} is minimized. Since $P^\text{gen} = - K^V(K^\omega)^{-1} K^\text{droop} \eta = - k_2 K^V(K^\omega)^{-1} K^\text{droop} 1_n$ and $\hat{\omega}=0_n$, premultiplying the first $n$ rows of the equilibrium of \eqref{eq:cl_dynamics_vec_delta_int_coordinated_double_prime} with $M^{-1}$, and adding to the $(n+1)$th to $2n$th rows premultiplied with $V^\text{nom}E^{-1}$ yields \begin{align*} - V^\text{nom} \mathcal{L}_R \hat{V} - k_2 K^V(K^\omega)^{-1} K^\text{droop} 1_n = P^m. \end{align*} Premultiplying the above equation with $1^T_n$ yields $k_2 = - \sum_{i=1}^n P^m_i \sum_{i=1}^n \frac{K^\omega_i}{K^V_i K^\text{droop}_i}$. Additionally, $\mathcal{L}_R \hat{V}$ is uniquely determined. Now consider \eqref{eq:hvdc_coordinated_voltage_objective}. Note that $P_i^\text{inj}$ and hence $I_i^\text{inj}$, are uniquely determined by \eqref{eq:hvdc_coordinated_generation_objective}. By the equilibrium of \eqref{eq:hvdc_coordinated_voltage}, $\mathcal{L}_R \hat{V} = I^\text{inj}$, where $I^\text{inj} = [I^\text{inj}_1, \dots, I^\text{inj}_n]^T$. Thus, the KKT condition of \eqref{eq:hvdc_coordinated_voltage_objective} is \begin{align} F^V \hat{V} = \mathcal{L}_R r, \label{eq:hvdc_coordinated_KKT_voltage_raw} \end{align} where $r\in \mathbb{R}^n$. Since $\mathcal{L}_R \hat{V}$ is uniquely determined, we premultiply \eqref{eq:hvdc_coordinated_KKT_voltage_raw} with $1_n^T$ and obtain the equivalent condition \begin{align} 1^T_n F^V \hat{V} = 0. \label{eq:hvdc_coordinated_KKT_voltage} \end{align} Since $F^V = K^V$, \eqref{eq:hvdc_coordinated_voltage_equilibrium} and \eqref{eq:hvdc_coordinated_KKT_voltage} are equivalent. Hence \eqref{eq:hvdc_coordinated_voltage_objective} is minimized, so Objective~\ref{obj:1_hvdc_coordinated} is satisfied. \end{pf} \begin{remark} Corollary \ref{cor:hvdc_coordinated_equilibrium} provides insight in choosing the controller gains of \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} and \eqref{eq:voltage_control_secondary}, to satisfy Objective~\ref{obj:1_hvdc_coordinated}. \end{remark} \section{Simulations} \label{sec:simulations} In this section, simulations are conducted on a test system to validate the performance of the proposed controllers. The simulation was performed in Matlab, using a dynamic phasor approach based on \citep{demiray2008}. The test system is illustrated in Figure~\ref{fig:testsystem}. \begin{figure}[ht] \def1.25\columnwidth{1.25\columnwidth} \begin{center} \input{ACandDCGrid2.pdf_tex} \captionsetup{width=\linewidth} \caption{MTDC test system, consisting of a 6-terminal MTDC grid. Each terminal is connected to an IEEE 14 bus AC grid, sketched as octagons.} \label{fig:testsystem} \end{center} \end{figure} The line parameters of the MTDC grid are given in Table~\ref{tab:HVDCgridParameter}. Note that we in the simulation also consider the inductances $L_{ij}$ and capacitances $C_{ij}$ of the HVDC lines. The capacitances of the terminals are assumed to be given by $C_i=0.375\times 10^{-3}$ p.u., and are chosen uniformly for all VSC stations. The AC grid parameters were obtained from \citep{Milano}. The generators are modeled as a 6{th} order machine model controlled by an automatic voltage controller and a governor \citep{kundur1994power}. Each load in the grid is assumed to be equipped with an ideal power controller. \begin{table}[htb] \captionsetup{width=\linewidth} \caption{HVDC grid line parameters} \label{tab:HVDCgridParameter} \begin{center} \begin{tabular}{ccccc} $i$&$j$&$R_{ij}$ [p.u.]&$L_{ij}$ [$10^{-3}$ p.u.]& $C_{ij}$ [p.u.]\\ \hline 1 & 2 & 0.0586 & 0.2560 &0.0085 \\ 1 & 3 & 0.0586 & 0.2560 &0.0085\\ 2 & 3 & 0.0878 & 0.3840 &0.0127\\ 2 & 4 & 0.0586 & 0.2560 &0.0085\\ 2 & 5 & 0.0732 & 0.3200 &0.0106\\ 2 & 6 & 0.1464 & 0.6400 &0.0212\\ 3 & 4 & 0.0586 & 0.2560 &0.0085\\ 3 & 5 & 0.1464 & 0.6400 &0.0212\\ 4 & 5 & 0.0732 & 0.3200 &0.0106\\ 5 & 6 & 0.1464 & 0.6400 &0.0212 \end{tabular} \end{center} \end{table} \begin{table}[htb] \begin{center} \captionsetup{width=\linewidth} \caption{Controller parameters} \label{tab:ControllerParameter} \begin{tabular}{ccccccc} $K^{\omega}_i$ &$K^{V}_i$&$K^\text{droop}_i$& $K^\text{droop, I}_i$ &$c^{\eta}_{ij}$ &$c^{\phi}_{ij}$ & $\gamma$ \\ \hline 1000 & 100 & 9 & 3.35 & ${5}/{R_{ij}}$ & ${15}/{R_{ij}}$ & 0 \end{tabular} \end{center} \end{table} The test grid was controlled with the controllers \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} and \eqref{eq:voltage_control_secondary}, with parameters given in Table~\ref{tab:ControllerParameter}. The communication network of \eqref{eq:hvdc_coordinated_droop_control_secondary_distributed} and \eqref{eq:voltage_control_secondary} is illustrated by the dashed lines in Figure~\ref{fig:testsystem}. Note that we have set $\gamma=0$, and that Theorem~\ref{th:stability_A_coordinated} thus does not guarantee stability of the equilibrium. The resulting matrix of the closed-loop system is however verified to be Hurwitz. At time $t=1$ the output of one generator in area 1 was reduced by $0.2$ p.u., simulating a fault. Figure~\ref{fig:ACFrequency} shows the average frequencies of the AC grids, while Figure~\ref{fig:DCVoltage} shows the DC voltages of the terminals. Figure~\ref{fig:Generators} shows the total increase of the generated power within each AC area. It can be noted that immediately after the fault, the average frequency of the AC area of the fault drops. The frequency drop is followed by a voltage drop in all terminals, and a subsequent frequency drop in the remaining AC areas. The frequencies converge to the nominal frequency, while the voltages converge to their new stationary values after approximately $30$ s. We note that the frequencies are restored to the nominal frequency, as predicted by Corollary~\ref{cor:hvdc_coordinated_equilibrium}. Furthermore the incremental generated power of the AC areas converge to the same value, as a consequence of Corollary~\ref{cor:hvdc_coordinated_equilibrium} and the controller parameters being equal. \begin{figure}[t!] \centering \tikzsetnextfilename{Frequencies} \begin{tikzpicture} \begin{axis} [cycle list name=Necsys, xlabel={$t$ [s]}, ylabel={$\omega(t)$ [p.u.]}, xmin=0, xmax=40, xtick={0,5,...,40}, ymin=0.99935, ymax=1.00005, yticklabel style={/pgf/number format/.cd, fixed, precision=4}, grid=major, height=6.5cm, width=0.96\columnwidth, legend cell align=left, legend pos= south east, ] \foreach \x in {1,2,3,4,5,6}{ \addplot table[x index=0,y index=\x,col sep=tab]{PlotData_coordinated/Necsys_Frequency.txt}; } \legend{Area 1, Area 2, Area 3, Area 4, Area 5, Area 6} \end{axis} \end{tikzpicture} \caption{Average frequencies in the AC areas.} \label{fig:ACFrequency} \end{figure} \begin{figure}[htb] \centering \tikzsetnextfilename{DCVoltages} \begin{tikzpicture} \begin{axis}[cycle list name=Necsys, xlabel={$t$ [s]}, ylabel={$V(t)$ [p.u.]}, xmin=0, xmax=40, xtick={0,5,...,40}, ymin=0.982, ymax=1.004, yticklabel style={/pgf/number format/.cd, fixed, precision=4}, grid=major, height=6.5cm, width=0.975\columnwidth, legend pos= south east, ] \foreach \x in {1,2,3,4,5,6}{ \addplot table[x index=0,y index=\x,col sep=tab] {PlotData_coordinated/Necsys2_Voltage.txt}; } \legend{Area 1, Area 2, Area 3, Area 4, Area 5, Area 6} \end{axis} \end{tikzpicture} \caption{DC terminal voltages.} \label{fig:DCVoltage} \end{figure} \begin{figure}[t!] \centering \tikzsetnextfilename{Generators} \begin{tikzpicture} \begin{axis}[cycle list name=Necsys, xlabel={$t$ [s]}, ylabel={$P^\text{gen}(t)$ [p.u.]}, xmin=0, xmax=40, xtick={0,5,...,40}, ymin=-0.005, ymax=0.076, scaled ticks=false, tick label style={/pgf/number format/fixed, /pgf/number format/precision=2}, ytick={0,0.02,...,0.101}, grid=major, height=6.5cm, width=0.975\columnwidth, legend pos= north east, ] \foreach \x in {1,2,3,4,5,6}{ \addplot table[x index=0,y index=\x,col sep=tab] {PlotData_coordinated/Necsys2_Generators.txt}; } \legend{Area 1, Area 2, Area 3, Area 4, Area 5, Area 6} \end{axis} \end{tikzpicture} \caption{Increase of generated power in the AC areas.} \label{fig:Generators} \end{figure} \section{Discussion and Conclusions} \label{sec:discussion} In this paper we have studied a distributed secondary controller for sharing frequency control reserves of asynchronous AC systems connected through an MTDC system. Under certain conditions, the proposed controller stabilizes the closed-loop system consisting of the interconnected AC systems and the MTDC grid. The frequencies in all AC grids are shown to converge to the nominal frequency. Furthermore, quadratic cost functions of the voltage deviations of the MTDC terminals and of the generated power, are minimized asymptotically. Finally, the results were validated on a six-terminal MTDC system with connected AC systems.

\section{Introduction} Let $G=\left(V, A\right)$ be an undirected graph where~$V$ is the set of~$N< \infty$ nodes and~$A\in \left\{0,1\right\}^{N\times N}$ is the underlying adjacency matrix. Our goal in this document is to construct a preorder on the (finite) set of nodes~$V$ of~$G$ that conveys non-trivial structure: 1) it is preserved by the logistic dynamical system \begin{equation} \frac{d}{dt}y_{i}(t) = \left(\sum_{j\sim i}\gamma y_{j}(t)\right)\left(1-y_i(t)\right)-y_i(t)\label{eq:logistic} \end{equation} for $i=1,\ldots,N$, in a sense to be made precise; and 2) its underlying equivalence relation coincides with the coarsest equitable partition of a graph. The logistic dynamics consists of a standard set of coupled Ordinary Differential Equations (ODEs) that model the spread of epidemics in networks (e.g.,~\cite{porterdynamical,newman,augusto_qualitative,kurtzi,Melanie}). The parameter~$\gamma$ stands for the rate of infection of a virus in the network; the state variable~$y_i(t)$ models the likelihood of infection of an individual~$i$ at time~$t$, due to transmission from its peers. The state-variable~$y_i(t)$ can be also interpreted as the fraction of individuals infected in a community (or sub-graph click\footnote{In a click, all nodes are connected to all other nodes.})~$i$ in a network of interconnected communities (a.k.a., super-network). The sum in equation~\eqref{eq:logistic} runs over the neighbors of~$i$ and that is what accounts for the underlying network structure in the dynamics. A preorder on the set of nodes is a binary relation $\succeq$ on $V$ that is reflexive and transitive, i.e., $\bullet$ $i \succeq i$ (reflexive) $\bullet$ $i \succeq i' \succeq i'' \Rightarrow i\succeq i''$ (transitive). Any preorder on a set $V$ naturally induces a quotient on $V$ given by the following equivalence relation \begin{equation} i\cong j \overset{{\sf def.}}\Longleftrightarrow i\succeq j \mbox{ and } j\succeq i.\label{eq:equivale} \end{equation} Since `$\succeq$' is not a partial order, the right hand side of~\eqref{eq:equivale} does not imply equality~$i=j$ (up to isomorphism), but a weaker equivalence relation. In fact, we will show that our preorder construct induces (in light of equation~\eqref{eq:equivale}) the so-called coarsest equitable partition (CEP), which is an important coloring or symmetry on a graph and that we will introduce momentarily. It is clear that the logistic dynamical system preserves the automorphisms of a graph -- in that, if isomorphic nodes are initialized with the same initial condition, then their state evolve in synchrony -- but it is non-trivial to show that it preserves the CEP of the graph (this will follow as a corollary to the fact that it preserves the preorder~`$\succeq$' constructed in this paper). \textbf{Outline of the paper.} Section~\ref{sec:symmetry} introduces and contrast the definitions of isomorphism and the coarsest equitable partition on graphs; Section~\ref{sec:construct} constructs the preorder; Section~\ref{sec:preserva} shows the implications of the preorder on dynamical systems such as the logistic system, in particular, the logistic dynamical system preserves the preorder and hence admits a non-trivial invariant set; Section~\ref{sec:conclusion} concludes the paper. \vspace*{2ex} \textbf{Preliminary Notation.} \vspace*{2ex} \begin{itemize} \item $\mathbb{N}=\left\{1,2,3,\ldots\right\}$; \item $V(G)=$ set of nodes of graph~$G$ -- sometimes denoted simply by~$V$; \item $\mathcal{N}(i)=$ set of neighbors to the node $i$ (also named neighborhood to $i$) in the undirected graph $G=\left(V,E\right)$; \item $d(i)=$ degree of the node $i$ in the undirected graph $G=\left(V,E\right)$; \item $\mathbf{d}(\mathbf{p})=$ vector collecting the degrees of the nodes stacked in the vector $\mathbf{p}\in V^{\ell}$, assuming that $G=\left(V,E\right)$ is undirected; \item $\pi_{I}^{\ell}\,:\,V^{\ell+1}\rightarrow V^{\left|I\right|}:$ is the canonical projection on the coordinates indexed by the set~$I\subset \left\{1,\ldots,\ell\right\}$. \end{itemize} \section{Symmetries on Graphs}\label{sec:symmetry} In this Section, we define some notions of symmetry on graphs and present the exact notion that fits our purposes, namely, the coarsest equitable partition (CEP). We start by introducing the definition of isomorphism on graphs. \begin{definition}[Graph Isomorphism]\label{def:isomorphgraph} We say that two graphs~$\mathcal{G}_1=\left(V_1,E_1\right)$ and~$\mathcal{G}_2=\left(V_2,E_2\right)$ are isomorphic, and represent it by~$\mathcal{G}_1\cong \mathcal{G}_2$, whenever there exists a bijection~$f\,:\,V_1\rightarrow V_2$ with \begin{equation} i\sim j \Leftrightarrow f(i)\sim f(j)\nonumber \end{equation} where~`$i \sim j$' means that~$i$ and~$j$ are neighbors in the graph. \end{definition} \begin{definition}[Graph Automorphism]\label{def:automorphgraph} We call~$f\,:V\,\rightarrow V$ an automorphism on~$G=\left(V,E\right)$ if~$f$ is an isomorphism from~$G$ onto~$G$, i.e.,~$f$ is bijective and \begin{equation} i\sim j \Leftrightarrow f(i)\sim f(j).\nonumber \end{equation} \end{definition} We say that two nodes~$i$ and~$j$ in a graph are isomorphic, and denote it as~$i\approx j$, whenever there exists an automorphism~$f$ with~$f(i)=j$ or~$f(j)=i$. The set of automorphisms on a graph~$G$ endowed with the operation of composition conforms to a group that we refer to as~${\sf Aut}(G)$. \begin{definition}[Orbits]\label{def:orbit} Let~$i$ be a node in~$G$. We refer to \begin{equation} [i]\overset{\Delta}={\sf Aut}(G).i=\left\{f(i)\,:\,f\in {\sf Aut}(G)\right\}\nonumber \end{equation} as the orbit of the vertex~$i$. In words, two nodes are in the same orbit if and only if they are isomorphic. \end{definition} The orbits of a graph in definition~\ref{def:orbit} define an equivalence relation between nodes which allows to quotient the set of nodes into equivalence classes \begin{equation} \left[i\right]\overset{\Delta}=\left\{j\in V\,:\, j\approx i\right\}.\nonumber \end{equation} Fig.~\ref{fig:quotient} illustrates the chromatic partition of a graph according to this equivalence relation. Monochromatic graphs under this equivalence relation are referred to as vertex-transitive graphs. \begin{figure} [hbt] \begin{center} \includegraphics[scale= 0.5]{network_classes.pdf} \caption{Nodes with the same color are isomorphic nodes.}\label{fig:quotient} \end{center} \end{figure} It is natural to expect that the partition illustrated in Fig.~\ref{fig:quotient} is preserved by the logistic dynamical system in the sense that if all nodes have the same initial condition, i.e.,~$\mathbf{y}(0)=\mathbf{1}_N y_0$, then, there is no reason to expect that eventually, say one of the indigo nodes will increase its state (e.g., degree of infection in an epidemics) faster than the other indigo node. Indeed, and more formally, we can show that such dynamical system preserves the partition induced by the orbits of the automorphism group of a graph. In particular, the underlying dynamical system admits a lower dimensional version quotiented by the automorphism symmetries of the graph. Fig.~\ref{fig:frucht} illustrates a regular graph with degree~$3$, so-called Frucht graph, that entails some asymmetry. For instance, its group of automorphisms is the trivial one (for more details, refer to~\cite{implementing}), i.e., no two nodes are isomorphic in the Frucht graph. In other words, the Frucht graph is a regular graph that is not vertex-transitive. From such asymmetry, it is not clear whether nodes initialized at the same state will evolve evenly under an epidemics modeled by the logistic dynamics. \begin{figure} [hbt] \begin{center} \includegraphics[scale= 0.4]{Frucht.pdf} \caption{Frucht Graph.}\label{fig:frucht} \end{center} \end{figure} Despite the asymmetry, all nodes are equivalent to each other in the CEP (or also in our preorder construct) sense. Therefore, as corollary to the results proved in this paper, nodes departing from the same state, in any regular graph (not necessarily vertex-transitive), will evolve in synchrony -- in particular, a regular network admits a $1$-dimensional dynamics version. \textbf{Remark.} Note that the synchrony induced by the CEP on the logistic dynamics shall not be taken for granted as one can find counter-examples of dynamical systems whose qualitative properties depend on the nodes even for a regular graph (that is not vertex-transitive), refer, e.g., to~\cite{aldous1989}. The previous discussion illustrates that even though the orbits of an automorphism are preserved by the logistic and the like dynamical systems, this is not the coarsest coloring that is preserved. Next, we introduce the coarsest equitable partition (for more details refer to~\cite{fractional}). First, let~$\mathcal{P}=\left\{\mathcal{P}_1,\ldots,\mathcal{P}_k\right\}$ be a partition on~$V$. Define \begin{equation} s_{\ell}(i) = \left|\mathcal{N}(i)\cap \mathcal{P}_{\ell}\right|\nonumber \end{equation} as the number of neighbors of~$i$ in the partition~$\ell$, and define~ \begin{equation} \mathbf{s}(i)=\left(s_1(i),\ldots,s_{d(i)}(i)\right),\nonumber \end{equation} as the distribution of the~$d(i)$ neighbors of~$i$ across the~$\ell$ elements of the partition. \begin{definition}[Equitable partition] The partition~$\mathcal{P}$ is called equitable whenever \begin{equation} i\approx_{\mathcal{P}} j \Leftrightarrow \mathbf{s}(i)=\mathbf{s}(j).\nonumber \end{equation} \end{definition} In words, two nodes~$i$ and $j$ lie in the same class if and only if they have the same distribution of neighbors across classes. In particular, they have the same degree. The partition induced by the orbits of the group of automorphisms in a graph is an equitable partition, but not the coarsest one in the graph. As referred to before, no two nodes in the Frucht graph in Fig.~\ref{fig:frucht}, which is a regular graph, are equivalent w.r.t. the group of automorphism whereas all nodes are equivalent w.r.t. our preorder construct equivalence, which is equitable and corresponds to the coarsest equitable partition. In certain cases, e.g., when the underlying graph~$G$ is a tree, then the orbits of the automorphism group and the coarsest equitable partition coincide. The next lemma asserts that the \textbf{coarsest} equitable partition exists in any (finite) graph. \begin{lemma} Let~$\mathcal{P}$ and~$\mathcal{Q}$ be two equitable partitions on~$G$. Then, there exists the finest equitable partition that is coarser than~$\mathcal{P}$ and~$\mathcal{Q}$ and it is denoted as \begin{equation} \mathcal{P}\vee\mathcal{Q}\nonumber \end{equation} \end{lemma} We now present the characterization of the coarsest equitable partition on a graph~$G$, given by the multisets of the so-called iterated degree sequence of the nodes. First, let \begin{eqnarray} d^{(1)}(i) & \overset{\Delta}= & \left\{d(v)\,:\,v\in\mathcal{N}(i)\right\}\nonumber\\ d^{(2)}(i) & \overset{\Delta}= & \left\{d^{(1)}(v)\,:\,v\in\mathcal{N}(i)\right\}\nonumber\\ & \vdots & \nonumber\\ d^{(k+1)}(i) & \overset{\Delta}= & \left\{d^{(k)}(v)\,:\,v\in\mathcal{N}(i)\right\}\nonumber \end{eqnarray} For instance, from Fig.\ref{fig:quotient} we have \begin{eqnarray} d^{(1)}(2) & = & \left\{2,3\right\}\nonumber\\ d^{(2)}(2) & = & \left\{\left\{2,2\right\},\,\left\{1,2,2\right\}\right\}\nonumber\\ \end{eqnarray} One can extend it to the infinitary sequence, so-called ultimate degree sequence \begin{equation} D\left(i\right)\overset{\Delta}=\left(d(i),d^{(1)}(i),d^{(2)}(i),\ldots,d^{(k)}(i),\ldots\right)\nonumber \end{equation} Now,~$\mathcal{P}$ is the coarsest equitable partition in a graph~$G$ if and only if \begin{equation} i\approx_{\mathcal{P}} j \Leftrightarrow D(i)=D(j).\nonumber \end{equation} Observe in particular that all nodes in a regular graph fall in the same class with respect to the coarsest equitable partition. We rephrase a corollary to Theorem~$6.5.1$ in~\cite{fractional} as follows. \begin{corollary} Let~$\mathcal{P}$ be the partition induced on the set of nodes~$V$ by the iterated degree sequence on the graph~$G=\left(V,E\right)$. Let~$\mathcal{Q}$ be an equitable partition on~$V$. Then,~$\mathcal{P}$ is coarser than~$\mathcal{Q}$. \end{corollary} There are efficient algorithms to determine the coarsest equitable partition in~$O(N\log(N))$. Determining the finer partition induced by the orbits of the automorphism group is NP-hard. We reserve the symbol~$\approx$ as the equivalence relation associated to the coarsest equitable coloring. \section{Preorder-construct on simple undirected graphs}\label{sec:construct} In this Section, we construct a preorder on the set of nodes of an undirected graph~$G=\left(V,E\right)$. The idea of the preorder is to formalize the following domination concept: a node~$i$ is greater or dominates a node~$j$ whenever the degree of~$i$ is greater than the degree of~$j$; and the degrees of the neighbors of~$i$ are greater than the degrees of the neighbors of~$j$; and the degrees of the neighbors of the neighbors of~$i$ are greater than the degrees of the neighbors of the neighbors of~$j$ and so forth. This idea must be clearly formalized to make any sense. Define \begin{equation} \mathcal{P}_{\ell}(i)\overset{\Delta}= \left\{\mathbf{p}^{\ell}=\left(i,i_1,\ldots,i_{\ell}\right)\in V^{\ell+1}\,:\,i_{k}\sim i_{k+1}\,\,\forall{k\leq \ell-1}\right\}\nonumber, \end{equation} as the set of all paths, on the graph~$G$, departing from the node~$i$ and with length~$\ell$. Now, we introduce one of the main concepts for the preorder construct. \begin{definition}[$\ell$-adapted function] Let~$i,j\in V(G)$ and~$\ell \in \mathbb{N}$. We define the class of $\ell$-adapted functions on~$G$ at the pair~$\left(j,i\right)$ as the set of functions \begin{equation} \begin{array}{rcl} f_{ji}^{\ell}: \mathcal{P}_{\ell}(j) & \to & \mathcal{P}_{\ell}(i)\\ \left(j,j_1,\ldots,j_{\ell}\right) & \mapsto & \left(i,i_1,\ldots,i_{\ell}\right) \end{array}\nonumber \end{equation} fulfilling the next two properties \begin{eqnarray} \pi^{\ell}_{\leq m}\left(\mathbf{p}^{\ell}\right) & = & \pi^{\ell}_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber\\ & \Downarrow & \\ \pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right) & = & \pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right),\nonumber \end{eqnarray} and \begin{eqnarray} \pi^{\ell}_{\leq m}\left(\mathbf{p}^{\ell}\right) & \neq & \pi^{\ell}_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber\\ & \Downarrow & \label{eq:notequal}\\ \pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right) & \neq & \pi^{\ell}_{\leq m}\left(f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right),\nonumber \end{eqnarray} for all~$\mathbf{p}^{\ell},\mathbf{\widetilde{p}}^{\ell}\in \mathcal{P}_{\ell}(j)$, where we defined \begin{equation} \pi^{\ell}_{\leq m}\,:\,V^{\ell+1}\longrightarrow V^{m+1}\nonumber \end{equation} with~$\pi^{\ell}_{\leq m}\left(x_0,x_1,\ldots,x_{\ell}\right)= \left(x_0,x_1,\ldots,x_{m}\right)$, i.e.,~$\pi_{\leq m}$ returns the first~$m\leq \ell$ coordinates from the input vector. \end{definition} In words, if the first~$m$ coordinates of a path~$\mathbf{p}^{\ell}$ coincide with the first~$m$ coordinates of a path~$\mathbf{\widetilde{p}}^{\ell}$, then the same should hold for the images of~$\mathbf{p}^{\ell}$ and~$\mathbf{\widetilde{p}}^{\ell}$ by the map~$f^{\ell}_{ij}$. For simplicity, from now one, we will commit an abuse of notation and we will write~$\pi_{\leq m}$ instead of~$\pi_{\leq m}^{\ell}$. We assume throughout that~$f^{\ell}_{ii}$ is the identity map for any~$i\in V(G)$, which is clearly an adapted function. We now observe two important properties. \begin{lemma}[``Group" property] Let~$g^{\ell}_{jk}$ and~$h^{\ell}_{ki}$ be two $\ell$-adapted functions at~$\left(j,k\right)$ and~$\left(k,i\right)$, respectively. Then, \begin{equation} f^{\ell}_{ji}\overset{\Delta}=h^{\ell}_{ki}\circ g^{\ell}_{jk}\nonumber \end{equation} is $\ell$-adapted at~$\left(j,i\right)$. \end{lemma} \begin{proof} Indeed, let~$\mathbf{p}^{\ell}, \mathbf{\widetilde{p}}^{\ell}\in \mathcal{P}_{\ell}(j)$ with \begin{equation} \pi_{\leq m}\left(\mathbf{p}^{\ell}\right)= \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber \end{equation} for some~$m\leq \ell$, then \begin{equation} \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)= \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\nonumber \end{equation} as~$g^{\ell}_{jk}$ is $\ell$-adapted and thus, \begin{equation} \pi_{\leq m}\left(f^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)= \pi_{\leq m}\left(f^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\nonumber \end{equation} as~$h^{\ell}_{jk}$ is $\ell$-adapted. Similarly, we can establish the property in equation~\eqref{eq:notequal}. \end{proof} \begin{lemma}\label{lem:invertible} If~$f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is an $\ell$-adapted invertible function, then~$g^{\ell}_{ij}\overset{\Delta}=\left(f^{\ell}_{ji}\right)^{-1}$ is an $\ell$-adapted function. \end{lemma} \begin{proof} Assume that~$g^{\ell}_{ij}$ is not an adapted function. Then, either one of the cases below should hold true \begin{enumerate} \item $\pi_{\leq m}\left(\mathbf{p}^{\ell}\right)= \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)$, but $\pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)\neq \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)$ for some~$m$ and $\mathbf{p}^{\ell}, \mathbf{\widetilde{p}}^{\ell}\in\mathcal{P}_{\ell}(i)$. \item $\pi_{\leq m}\left(\mathbf{p}^{\ell}\right)\neq \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)$, but $\pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)= \pi_{\leq m}\left(g^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)$ for some~$m$ and $\mathbf{p}^{\ell}, \mathbf{\widetilde{p}}^{\ell}\in\mathcal{P}_{\ell}(i)$. \end{enumerate} Now, case 1) implies that \begin{eqnarray} \pi_{\leq m}\left(f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\right) & = & \pi_{\leq m}\left(\mathbf{p}^{\ell}\right)\nonumber\\ & = & \pi_{\leq m}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber\\ & = & \pi_{\leq m}\left(f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\right)\nonumber \end{eqnarray} which is a contradiction as \begin{equation} \pi_{\leq m}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\neq \pi_{\leq m}\left(g^{\ell}_{ij}\left(\mathbf{\widetilde{p}}^{\ell}\right)\right)\nonumber \end{equation} and~$f^{\ell}_{ji}$ is an adapted function. Case 2) similarly leads to a contradiction, therefore the inverse map~$g^{\ell}_{ij}$ is an $\ell$-adapted function. \end{proof} We now observe that if~$f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is an adapted function, then its restriction to the $(\ell-1)$-paths~$\mathcal{P}_{\ell-1}(j)$, denoted as~$f^{\ell-1}_{ji}$, is an $(\ell-1)$-adapted function. Moreover,~$f^{\ell}_{ji}$ extends~$f^{\ell-1}_{ji}$ in that if~$\mathbf{p}^{\ell-1}=\pi_{\leq \ell-1}\left(\mathbf{p}^{\ell}\right)$, then \begin{equation} \pi_{\leq \ell-1}\left(f^{\ell}_{ji}(\mathbf{p}^{\ell})\right)=f^{\ell-1}_{ji}\left(\mathbf{p}^{\ell-1}\right)\label{eq:extends} \end{equation} or else, if~$\mathbf{p}^{\ell-1} \neq \pi_{\leq \ell-1}\left(\mathbf{p}^{\ell}\right)$, then \begin{equation} \pi_{\leq \ell-1}\left(f^{\ell}_{ji}(\mathbf{p}^{\ell})\right)\neq f^{\ell-1}_{ji}\left(\mathbf{p}^{\ell-1}\right)\label{eq:extends2} \end{equation} \begin{definition}[Adapted sequence] We call \begin{equation} \left(f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)\right)_{\ell}\nonumber \end{equation} an adapted sequence (on~$\ell$) of adapted functions at the pair~$\left(j,i\right)$ if~$f^{\ell+1}_{ji}$ extends~$f^{\ell}_{ji}$ for any~$\ell \in \mathbb{N}$ in the sense of equations~\eqref{eq:extends}-\eqref{eq:extends2}. \end{definition} The following Lemma remarks that if the adapted sequence is comprised of invertible maps, then the inverse maps sequence is an adapted sequence as well. \begin{lemma}[Inverse sequence]\label{lem:inversesequence} Let~$\left(f^{\ell}_{ji}\right)_{\ell}$ be an adapted sequence with $f^k_{ji}$ being an invertible $k$-adapted function for all~$k$. Then, the sequence~$\left(g^{\ell}_{ij}\right)_{\ell}$, where~$g^{\ell}_{ij}\overset{\Delta}=\left(f^{\ell}_{ji}\right)^{-1}$, is an adapted sequence. \end{lemma} \begin{proof} Note that each~$g^{\ell}_{ij}$ is an adapted function from Lemma~\ref{lem:invertible}. We are just left to prove that~$g^{\ell+1}_{ij}$ extends~$g^{\ell}_{ij}$ for all~$\ell$, in the sense of equations~\eqref{eq:extends}-\eqref{eq:extends2}. Suppose that~$\left(g^{\ell}_{ij}\right)_{\ell}$ is not an adapted sequence. Then, there exists an~$\ell\in \mathbb{N}$ so that~$g^{\ell+1}_{ij}$ does not extend~$g^{\ell}_{ij}$, or in other words, one of the conditions below hold true \begin{enumerate} \item $\mathbf{p}^{\ell}=\pi_{\leq \ell}\left(\mathbf{p}^{\ell+1}\right)$, but $\pi_{\leq \ell}\left(g^{\ell+1}_{ji}(\mathbf{p}^{\ell+1})\right) \neq g^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)$; \item $\mathbf{p}^{\ell}\neq \pi_{\leq \ell}\left(\mathbf{p}^{\ell+1}\right)$, but $\pi_{\leq \ell}\left(g^{\ell+1}_{ji}(\mathbf{p}^{\ell+1})\right) = g^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)$. \end{enumerate} Define~$\mathbf{\widetilde{p}^{\ell}}\overset{\Delta}=g_{ij}^{\ell}\left(\mathbf{p}^{\ell}\right)$ and~$\mathbf{\widetilde{p}}^{\ell+1}\overset{\Delta}=g_{ij}^{\ell+1}\left(\mathbf{p}^{\ell+1}\right)$. Case 1) implies that \begin{equation} f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)\neq \pi_{\leq \ell}\left(f^{\ell+1}_{ji}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)\right),\nonumber \end{equation} but~$f^{\ell}_{ji}\left(\mathbf{\widetilde{p}}^{\ell}\right)=\mathbf{p}^{\ell}$ and~$f^{\ell+1}_{ji}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)=\mathbf{p}^{\ell+1}$ which contradicts the assumption. Case 2) leads similarly to a contradiction. Therefore, the inverse sequence~$\left(g^{\ell}_{ij}\right)_{\ell}$ is an adapted sequence. \end{proof} \begin{theorem}[``Group" property for adapted sequences]\label{th:groupadaptedseq} Let~$\left(h^{\ell}_{ki}\right)$ and~$\left(g^{\ell}_{jk}\right)$ be two adapted sequences. Then, the pointwise composition \begin{equation} \left(f^{\ell}_{ji}\overset{\Delta}= h^{\ell}_{ki} \circ g^{\ell}_{jk}\right)_{\ell}\nonumber \end{equation} is an adapted sequence. \end{theorem} \begin{proof} The sequence~$\left(f^{\ell}_{ji}\right)$ must fulfill the defining conditions in equations~\eqref{eq:extends}-\eqref{eq:extends2}. Let~$\mathbf{p}^{\ell}=\pi_{\leq \ell}\left(\mathbf{p}^{\ell+1}\right)$, with~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$ and~$\mathbf{p}^{\ell+1}\in\mathcal{P}_{\ell+1}(j)$. Then, \begin{equation} \pi_{\leq \ell}\left(g^{\ell+1}_{jk}\left(\mathbf{p}^{\ell+1}\right)\right)=g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\label{eq:g} \end{equation} as~$\left(g^{\ell}_{jk}\right)_{\ell}$ is an adapted sequence. Moreover, define~$\mathbf{\widetilde{p}}^{\ell+1}\overset{\Delta}= g^{\ell+1}_{jk}\left(\mathbf{p}^{\ell+1}\right)$ and~$\mathbf{\widetilde{p}}^{\ell}\overset{\Delta}= g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)$. Thus, equation~\eqref{eq:g} can be written as \begin{equation} \pi_{\leq \ell}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)=\mathbf{\widetilde{p}}^{\ell}\nonumber \end{equation} which implies that \begin{equation} \pi_{\leq \ell}\left(h^{\ell+1}_{jk}\left(\mathbf{\widetilde{p}}^{\ell+1}\right)\right)=h^{\ell}_{jk}\left(\mathbf{\widetilde{p}}^{\ell}\right)\nonumber \end{equation} and hence, \begin{equation} \pi_{\leq \ell}\left(f^{\ell+1}_{ji}\left(\mathbf{p}^{\ell+1}\right)\right)=f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right).\nonumber \end{equation} The second case in equation~\eqref{eq:extends2} holds similarly and~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence. \end{proof} Now, we introduce the main construct of this document. \begin{definition} [Preorder]\label{def:preorder} Let~$i,j\in V(G)$. We say that~$i\succeq j$, whenever there exists an adapted sequence \begin{equation} \left(f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)\right)_{\ell}\nonumber \end{equation} where both properties hold \begin{itemize} \item $f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is injective; \item $\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\geq \mathbf{d}\left(\mathbf{p}^{\ell}\right)$ \end{itemize} for all~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$ and all~$\ell \in\mathbb{N}$. \end{definition} We say that an adapted sequence~$\left(f^{\ell}_{ji}\right)$ \textbf{explains} the inequality~$i\succeq j$, whenever~$\left(f^{\ell}_{ji}\right)$ fulfills the two conditions in the definition~\ref{def:preorder}. We may also refer that $i\succeq j$ is \textbf{explained} by~$\left(f^{\ell}_{ji}\right)$ under these conditions. \begin{theorem} The binary relation~$\succeq$ in definition~\ref{def:preorder} is a preorder on the set of nodes~$V(G)$ of the graph~$G$, i.e., it is reflexive and transitive: \begin{itemize} \item \textbf{(Reflexivity)} $i\succeq i$; \item \textbf{(Transitivity)} $i \succeq k \succeq j \Rightarrow i \succeq j$. \end{itemize} \end{theorem} \begin{proof} \underline{\textbf{Reflexivity:}} For any node~$i\in V(G)$, consider the sequence of identity maps \begin{equation} f^{\ell}_{ii}\,:\,\mathcal{P}_{\ell}(i) \rightarrow \mathcal{P}_{\ell}(i)\nonumber \end{equation} and note that this is an adapted sequence, with injective maps that preserve the degree. \underline{\textbf{Transitivity:}} Let~$\left(h^{\ell}_{ki}\right)$ and~$\left(g^{\ell}_{jk}\right)$ explain~$i\succeq k$ and~$k\succeq j$, respectively, and define the sequence \begin{equation} \left(f^{\ell}_{ji}\overset{\Delta}= h^{\ell}_{ki} \circ g^{\ell}_{jk}\right)_{\ell}.\nonumber \end{equation} First, note that from Theorem~\ref{th:groupadaptedseq}, the sequence~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence. Since the composition of injective maps is injective, then each~$f^{\ell}_{ji}$ is injective. Now, \begin{eqnarray} \mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right) & = & \mathbf{d}\left(h^{\ell}_{ki}\left(g^{\ell}_{jk}\left(\mathbf{p}^{\ell}\right)\right)\right)\nonumber\\ & \geq & \mathbf{d}\left(g^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\nonumber\\ & \geq & \mathbf{d}\left(\mathbf{p}^{\ell}\right)\nonumber \end{eqnarray} and thus,~$i\succeq j$ (explained by~$\left(f^{\ell}_{ji}\right)$). \end{proof} Now, we remark that to each preorder, there is an associated equivalence relation defined as \begin{equation} i\cong j \Leftrightarrow \left(i\succeq j \wedge j\succeq i \right).\nonumber \end{equation} Checking that this is an equivalence relation is trivial. The next Theorem, provides a more explicit characterization for our preorder~$\succeq$. \begin{theorem}\label{th:characterization} $i \cong j$, if and only if there exists an adapted sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ such that \begin{itemize} \item $f^{\ell}_{ji}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ is bijective; \item $\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)=\mathbf{d}\left(\mathbf{p}^{\ell}\right)$; \end{itemize} for all~$\mathbf{p}^{\ell} \in \mathcal{P}_{\ell}(j)$ and for all~$\ell$. \end{theorem} \begin{proof} The \textbf{if} part is trivial: it follows as a corollary to Lemma~\ref{lem:inversesequence}. We are left to prove the \textbf{only if} part. Let~$i\cong j$. Then, by definition,~$i\succeq j$ and~$j\succeq i$. In other words, there exist two adapted sequences~$\left(f_{ji}^{\ell}\right)$ and~$\left(g_{ij}^{\ell}\right)$ that explain~$i\succeq j$ and~$j\succeq i$, respectively. First, note that since~$f_{ji}^{\ell}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ and~$g_{ij}^{\ell}\,:\,\mathcal{P}_{\ell}(i)\rightarrow \mathcal{P}_{\ell}(j)$ are injective maps, then~$\left|\mathcal{P}_{\ell}(i)\right|=\left|\mathcal{P}_{\ell}(j)\right|$ for all~$\ell$ (Schr\"{o}der-Bernstein Theorem) and thus, since the graph is finite, we have~$\left|\mathcal{P}_{\ell}(i)\right|=\left|\mathcal{P}_{\ell}(j)\right|< \infty$, and the injective maps~$f_{ji}^{\ell}\,:\,\mathcal{P}_{\ell}(j)\rightarrow \mathcal{P}_{\ell}(i)$ and~$g_{ij}^{\ell}\,:\,\mathcal{P}_{\ell}(i)\rightarrow \mathcal{P}_{\ell}(j)$ are in fact bijections. We are to show that~$f^{\ell}_{ji}$ (or~$g^{\ell}_{ij}$) preserves the degree sequence. Define the following interlaced dynamics on~$\mathcal{P}_{\ell}(i)\cup \mathcal{P}_{\ell}(j)$ \begin{itemize} \item If $\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$: $\mathbf{p}^{\ell}\rightarrow f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\rightarrow g^{\ell}_{ij}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\rightarrow f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)\right)\rightarrow \ldots$ \item If $\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(i)$: $\mathbf{p}^{\ell}\rightarrow g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\rightarrow f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\rightarrow g^{\ell}_{ij}\left(f^{\ell}_{ji}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)\right)\rightarrow \ldots$ \end{itemize} We first remark that the vector degree is monotonous over the orbits of the above interlaced dynamical system. Therefore, if~$\mathbf{p}^{\ell}$ is periodic, then \begin{itemize} \item $\mathbf{d}\left(\mathbf{p}^{\ell}\right)=\mathbf{d}\left(f^{\ell}_{ji}\left(\mathbf{p}^{\ell}\right)\right)$, if~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(j)$; \item $\mathbf{d}\left(\mathbf{p}^{\ell}\right)=\mathbf{d}\left(g^{\ell}_{ij}\left(\mathbf{p}^{\ell}\right)\right)$, if~$\mathbf{p}^{\ell}\in\mathcal{P}_{\ell}(i)$. \end{itemize} In other words, if~$\mathbf{p}^{\ell}\in \mathcal{P}_{\ell}(i)\cup \mathcal{P}_{\ell}(j)$ is a periodic point (or path) w.r.t. the interlaced dynamics, then the degree is preserved by the bijections~$f^{\ell}_{ji}$ and~$g^{\ell}_{ij}$. We claim that any point (or path)~$\mathbf{p}$ is periodic. Indeed, let~$\mathbf{p}\in \mathcal{P}_{\ell}(i)$ and assume that~$\mathbf{p}$ is not periodic. This in particular implies that all the points in the backwards iterates \begin{equation} \left(f^{\ell}_{ji}\right)^{-1}(\mathbf{p}),\left(g^{\ell}_{ji}\right)^{-1}\left(\left(f^{\ell}_{ji}\right)^{-1}(\mathbf{p})\right),\ldots\nonumber \end{equation} are not periodic as well. In other words, the set of all backwards iterates has infinite cardinality, which is a contradiction as~$\left|\mathcal{P}_{\ell}(i)\cup \mathcal{P}_{\ell}(j)\right|<\infty$. Therefore, we conclude that~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence of bijections preserving the vector degree. \end{proof} \begin{theorem} The equivalence relation~$\cong$ is equal to the coarsest equitable partition (CEP). \end{theorem} \begin{proof} We first show that~$\cong$ conforms to an equitable relation. Then, we show that any other equitable partition in the graph is finer then the coloring of~$\cong$. \textbf{Part I: $\cong$ is equitable.} Let~$i\cong j$. Take an arbitrary neighbour~$j_1\in\mathcal{N}(j)$ of~$j$ and let us show that there exists~$i_1\in \mathcal{N}(i)$ with~$j_1\cong i_1$. This establishes the first part. Indeed, let~$i_1\overset{\Delta}=\pi_{2}\left(f^{1}_{ji}\left(j,j_1\right)\right)$. Our claim is that~$j_1\cong i_1$. Define~$g^{\ell}_{j_1i_1}(\mathbf{p})\overset{\Delta}=\pi_{2\leq \cdot \leq \ell+1}\left(f^{\ell+1}_{ji}(j,\mathbf{p})\right)$ for all~$\mathbf{p}\in \mathcal{P}_{\ell}(j_1)$. We observe that~$\left(g^{\ell}_{j_1i_1}\right)$ fulfills the conditions of Theorem~\ref{th:characterization} and thus,~$j_1\cong i_1$. \textbf{Part II: $\cong$ is the coarsest equitable relation.} We will prove that given another equitable relation~$\approx$, then \begin{equation} i\approx j \Rightarrow i\cong j.\nonumber \end{equation} For each pair~$n,m$ of $\approx$-equivalent nodes~$n\approx m$, define \begin{equation} f_{nm}\,:\,\mathcal{N}(n)\rightarrow \mathcal{N}(m)\nonumber \end{equation} to be any bijection preserving the $\approx$-classes, i.e., $f_{nm}(k)\approx k$ for all~$k\in\mathcal{N}(n)$. Note that this is possible as~$\approx$ is equitable: $n,m$ have the same degree (hence we can choose a bijection), and have the same number of neighbors per class (hence we can choose a bijection that preserves classes). Now, assume~$i\approx j$ and define the sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ by induction as \begin{equation} \begin{array}{rcl} f_{ji}^{1}: \mathcal{P}_{1}(j) & \to & \mathcal{P}_{1}(i)\\ \left(j,j_1\right) & \mapsto & \left(i,i_1\right) \end{array}\nonumber \end{equation} where~$i_1=f_{ji}(j_1)$, with~$j_1\in \mathcal{N}(j)$ and \begin{equation} \begin{array}{rcl} f_{ji}^{n}: \mathcal{P}_{n}(j) & \to & \mathcal{P}_{n}(i)\\ \left(j,j_1,\ldots,j_{n}\right) & \mapsto & \left(i,i_1,\ldots,i_n\right) \end{array}\nonumber \end{equation} where, \begin{equation} i_1=f_{ji}(j_1),\,i_2=f_{j_{1}i_{1}}(j_2),\,\ldots,i_n=f_{j_{n-1}i_{n-1}}(j_n)\nonumber \end{equation} with~$j_1\in \mathcal{N}(j)$, and~$j_k\in\mathcal{N}(j_{k-1})$ for all~$k\leq n$, i.e., \begin{equation} \left(j,j_1,\ldots,j_n\right)\in\mathcal{P}_{n}(j).\nonumber \end{equation} Observe first that~$f_{ji}^{n}$ is well constructed as~$j_n\approx i_n$ for each $n$ above -- as one can show by induction -- and thus,~$f_{j_ni_n}$ makes sense. Moreover, the sequence~$\left(f^{\ell}_{ji}\right)$ is an adapted sequence of bijections that preserves the vector degree. By Theorem~\ref{th:characterization}, we conclude that~$i\cong j$. \end{proof} The next Theorem provides an alternative practical characterization for the preorder thus constructed: it is an inductive preorder. In fact, in order to prove that the logistic dynamical system preserves the preorder, we need the result in Theorem~\ref{th:induction}. \begin{theorem}[Induction property of $\succeq$]\label{th:induction} \begin{equation}\label{eq:inductive} i\succeq j \Leftrightarrow \exists{f\,:\,\mathcal{N}(j)\rightarrow \mathcal{N}(i) \mbox{ injective:}} f(k)\succeq k\,\, \forall{k\in\mathcal{N}(j)} \end{equation} \end{theorem} \begin{proof} \textbf{We start by proving the implication~`$\Rightarrow$'.} Define~$f(k)\overset{\Delta}=\pi_{2}\left(f^{1}_{ji}(j,k)\right)$. Note that~$f\,:\,\mathcal{N}(j)\rightarrow \mathcal{N}(i)$ is injective. Also, define the following sequence \begin{equation} g^{\ell}_{kf(k)}\left(k,k_1,k_2,\ldots,k_{\ell}\right)=\pi_{2\leq \cdot \leq \ell+2}\left(f^{\ell+1}_{ji}(j,k,k_1,\ldots,k_{\ell})\right)\nonumber \end{equation} and note that~$\left(g^{\ell}_{kf(k)}\,:\, \mathcal{P}_{\ell}(k)\rightarrow \mathcal{P}_{\ell}(f(k))\right)_{\ell}$ is an adapted sequence. Note also that~$g^{\ell}_{kf(k)}\,:\,\mathcal{P}_{\ell}(k)\rightarrow \mathcal{P}_{\ell}(f(k))$ is injective for all~$\ell$. The monotonicity on the vector degree is clearly inherited by~$f^{\ell}_{ji}$. Therefore,~$f(k)\succeq k$. \textbf{Now, we prove the implication~`$\Leftarrow$'.} By definition, we have~$f(k)\succeq k$ if and only if there exists an adapted sequence~$\left(g^{\ell}_{kf(k)}\right)$ that explains the inequality. Now, define \begin{equation} f^{\ell}_{ji}\left(j, j_1,j_2,\ldots,j_{\ell}\right)=\left(i, g^{\ell-1}_{j_1 f(j_1)}\left(j_1,j_2,\ldots,j_{\ell}\right)\right). \end{equation} $\left(f^{\ell}_{ji}\right)$ thus defined is an adapted sequence that explains~$i\succeq j$. \end{proof} In fact, our preorder is not only inductive in the sense of equation~\eqref{eq:inductive}, but it is the maximum preorder amongst the inductive preorders as explained next. Let~$\mathcal{O}$ be the set of inductive preorders on a graph~$G$ endowed with the following partial order~$\leq$ \begin{equation} \widetilde{\succeq}\,\, \leq \,\, \succeq \,\,\Longleftrightarrow \,\,\left(i \,\,\widetilde{\succeq}\,\, j \Rightarrow i \succeq j\right).\nonumber \end{equation} It is trivial to check that~$\leq$ is a partial order on~$\mathcal{O}$. \begin{theorem} $\succeq$ is the greatest element (a.k.a. maximum) on~$\mathcal{O}$ with respect to the partial order~$\leq$, \begin{equation} \widetilde{\succeq}\,\, \leq \,\,\succeq\,\,\forall{\widetilde{\succeq}\in \mathcal{O}}, \end{equation} i.e., any other preorder~$\widetilde{\succeq}\in \mathcal{O}$ is not only comparable to~$\succeq$, but it is upperbounded by it. \end{theorem} \begin{proof} Let~$\widetilde{\succeq}$ be another inductive preorder. For each pair of nodes~$\left(n,m\right)$ with~$n \,\,\widetilde{\succeq}\,\, m$, choose an injective function \begin{equation} f_{mn}\,:\,\mathcal{N}(m)\rightarrow \mathcal{N}(n)\nonumber \end{equation} that preserves~$\widetilde{\succeq}$, i.e.,~$f(k)\,\widetilde{\succeq} k$ for all~$k\in \mathcal{N}(m)$. Now, let~$i \widetilde{\succeq} j$ and define the sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ by induction: \begin{equation} \begin{array}{rcl} f_{ji}^{1}: \mathcal{P}_{1}(j) & \to & \mathcal{P}_{1}(i)\\ \left(j,j_1\right) & \mapsto & \left(i,i_1\right) \end{array}\nonumber \end{equation} where~$i_1=f_{ji}(j_1)$, with~$j_1\in \mathcal{N}(j)$ and \begin{equation} \begin{array}{rcl} f_{ji}^{n}: \mathcal{P}_{n}(j) & \to & \mathcal{P}_{n}(i)\\ \left(j,j_1,\ldots,j_{n}\right) & \mapsto & \left(i,i_1,\ldots,i_n\right) \end{array}\nonumber \end{equation} where, \begin{equation} i_1=f_{ji}(j_1),\,i_2=f_{j_{1}i_{1}}(j_2),\,\ldots,i_n=f_{j_{n-1}i_{n-1}}(j_n)\nonumber \end{equation} with~$j_1\in \mathcal{N}(j)$, and~$j_k\in\mathcal{N}(j_{k-1})$ for all~$k\leq n$, i.e., \begin{equation} \left(j,j_1,\ldots,j_n\right)\in\mathcal{P}_{n}(j).\nonumber \end{equation} Note that~$f^{\ell}_{ji}$ is well defined. Also, each~$f^{\ell}_{ji}$ is injective and preserves the degree sequence. Moreover, the sequence~$\left(f^{\ell}_{ji}\right)_{\ell}$ is adapted. \end{proof} \section{Logistic dynamics preserves the preorder}\label{sec:preserva} In this section, we show that the logistic dynamics~\eqref{eq:logistic} preserves the preorder~$\succeq$. As a result, this leads to the emergence of a non-trivial invariant set for this type of dynamical system over networks, whose characterization is tied to the underlying graph structure. \begin{theorem}\label{th:preservesineq} Let~$\mathbf{y}(0)\in\left[0,1\right]^{N}$ be so that \begin{equation} y_{i}(0)\geq y_{j}(0)\,\forall{i\succeq j}.\label{eq:assumption} \end{equation} Then, \begin{equation} y_{i}(t,\mathbf{y}(0))\geq y_{j}\left(t,\mathbf{y}(0)\right)\,\,\forall{t\geq 0}. \end{equation} \end{theorem} Equivalently, the set \begin{equation} S\overset{\Delta}=\left\{\mathbf{y}\in\left[0,1\right]^{N} \,:\, y_i\geq y_k,\,\forall{i\succeq k} \right\}\label{eq:invariant} \end{equation} is invariant to the logistic dynamical system. \begin{proof} To show that~$S$ (defined in equation~\eqref{eq:invariant}) is invariant under the logistic dynamics (either continuous or discrete-time), we just need to check the qualitative behavior of the trajectory~$\left(\mathbf{y}(t)\right)$ at the border of~$S$ to establish that it does not escape from the set. We focus attention on the continuous time case. Let~$i\succeq j$ with~$y_i(s)=y_j(s)$, for some time~$s$, i.e., the trajectory is at the border at time~$s$. Let the adaptive sequence~$\left(f_{ji}^{\ell}\right)$ explain the inequality~$i\succeq j$. For simplicity, denote~$f_{ji}:=f_{ji}^{1}$ with \begin{equation} f_{ji}\,:\,\mathcal{N}(j)\rightarrow \mathcal{N}(i).\nonumber \end{equation} Assume first that \begin{equation} y_{k}(s) > y_{f_{ji}(k)}(s) \end{equation} for some neighbor~$k\in\mathcal{N}(i)$. Given the induction property of the preorder in Theorem~\ref{th:induction} and the assumption in equation~\eqref{eq:assumption}, we conclude, by inspection on the ODE equation~\eqref{eq:logistic}, that \begin{equation} \overset{\cdot}y_i(s)>\overset{\cdot} y_j(s)\nonumber \end{equation} and thus, there exists~$\delta>0$ such that~$y_i(t)>y_j(t)$ for all $t\in\left.\left[s,s+\delta\right.\right)$. In other words, the trajectory does not scape the set. In the case where the two nodes have the same configuration in a first order neighborhood, i.e.,~$f_{ji}$ is bijective and \begin{equation} y_k(s)=y_{f_{ji}(k)}(s) \end{equation} for all~$k\in \mathcal{N}(i)$, then the previous argument does not apply since~$\overset{\cdot}y_i(s)=\overset{\cdot}y_j(s)$ and we need to consider higher order derivatives. The equations for the $(n+1)$th order derivatives of the logistic system are given by \begin{eqnarray} \overset{(n+1)}y_i(t) & = & \gamma \sum_{k\sim i} \left[\overset{(n)}y_k(t)\left(1-y_i(t)\right)- \sum_{m=0}^{n-1}\left(\begin{array}{c} n \\ m \end{array}\right)\overset{(m)}y_k(t) \overset{(n-m)}y_i(t)\right]-\overset{(n)}y_i(t)\label{eq:high}\\ \overset{(n+1)}y_j(t) & = & \gamma \sum_{\ell\sim j} \left[\overset{(n)}y_{\ell}(t)\left(1-y_j(t)\right)- \sum_{m=0}^{n-1}\left(\begin{array}{c} n \\ m \end{array}\right)\overset{(m)}y_{\ell}(t) \overset{(n-m)}y_j(t)\right]-\overset{(n)}y_j(t).\label{eq:high2} \end{eqnarray} Given a vector of indexes~$I=\left(I_1,I_2,\ldots,I_N\right)\in\mathbb{N}^N$, define~$y_I:=\left(y_{I_1},\ldots,y_{I_N}\right)$. Assume, for some~$N\geq 1$, that the adapted sequence~$\left(f_{ji}^{\ell}\right)$ is such that~$f_{ji}^{m}\,:\,\mathcal{P}_{m}(j)\rightarrow \mathcal{P}_{m}(i)$ is bijective for all~$m\leq N$ and \begin{equation} y_{p^{m}}=y_{f^m_{ji}\left(p^{m}\right)}, \end{equation} for all~$p^m\in \mathcal{P}^m(j)$, and~$m\leq N$, but at~$N+1$ we have (the strict dominance) \begin{equation} f_{ji}^{N+1}\mbox{ injective and not bijective, or } y_{p_{N+1}^{N+1}}<y_{f_{ji}\left(p^{N+1}_{N+1}\right)}.\nonumber \end{equation} Then, by inspection on the higher order derivative equations~\eqref{eq:high}-\eqref{eq:high2}, and applying induction on~$N$, we conclude that \begin{eqnarray} \overset{(n)}y_i(s) & = & \overset{(n)}y_j(s),\,\forall{n\leq N}\\ \overset{(N+1)}y_i(s) & > & \overset{(N+1)}y_j(s) \end{eqnarray} and, thus, there exists~$\delta>0$ such that~$y_i(t)>y_j(t)$ for all $t\in\left.\left[s,s+\delta\right.\right)$, that is, the trajectory does not escape the set~$S$. \end{proof} \begin{corollary}[LD preserves CEP]\label{co:preservesCEP} The homogeneous logistic dynamical system preserves the CEP of the graph, i.e., if~~$\mathbf{y}(0)\in\left[0,1\right]^{N}$ is so that \begin{equation} y_{i}(0)= y_{j}(0)\,\forall{i\cong j}. \end{equation} Then, \begin{equation} y_{i}(t,\mathbf{y}(0))= y_{j}\left(t,\mathbf{y}(0)\right)\,\,\forall{t\geq 0}\,\,\,\,\,\forall{i\cong j}. \end{equation} \end{corollary} Corollary~\ref{co:preservesCEP} follows from Theorems~\ref{th:characterization} and~\ref{th:preservesineq}. \textbf{Remark.} The proof of Theorem~\ref{th:preservesineq} naturally extends to establish the result for a broader class of dynamical systems. In particular, let~$\overline{F}\,:\,\mathbb{R}\times\mathbb{R}^{N-1}\rightarrow \mathbb{R}$ be an analytic real valued function invariant under permutations on the second vector coordinate, i.e., \begin{equation} \overline{F}\left(x_1;x_2,\ldots,x_N\right)= \overline{F}\left(x_1;p\left(x_2,\ldots,x_N\right)\right) \end{equation} for all permutations~$p\,:\,\mathbb{R}^{N-1}\rightarrow \mathbb{R}^{N-1}$ and for all~$\mathbf{x}=\left(x_1,\ldots,x_N\right)\in\mathbb{R}^{N}$. Assume also that~$\overline{F}$ is monotonous in the second vector coordinate, i.e., \begin{equation} \mathbf{x} \geq \mathbf{y}\in\mathbb{R}^{N-1}\Rightarrow \overline{F}\left(x_1;\mathbf{x}\right)\geq \overline{F}\left(x_1;\mathbf{y}\right). \end{equation} Given a graph~$G$, let~$F_G=\left(F_1,F_2,\ldots,F_N\right)$ be the vector field induced by~$\overline{F}$ as follows \begin{equation}\label{eq:construct2} F_i\left(x\right)= \overline{F}\left(x_i, x_{i_1},\ldots,x_{i_k},0,\ldots,0\right) \end{equation} where,~$\mathcal{N}(i)\overset{\Delta}=\left\{j\sim i\,:\, j\in V\right\}=\left\{i_1,\ldots,i_k\right\}$ is the set of neighbors to~$i$ in the graph~$G$ (i.e.,~$F$ conveys the graph structure of~$G$). Note that the ordering of~$x_{i_1},\ldots,x_{i_k}$ in equation~\eqref{eq:construct2} is not relevant as~$\overline{F}$ is invariant under permutation of the associated coordinates. By evoking Fa\`{a} di Bruno's formula (refer, e.g., to~\cite{faadi}) for higher order derivatives of~$F_G$ and resorting to a proof similar to the one in Theorem~\ref{th:preservesineq}, one can establish the result in Theorem~\ref{th:preservesineq} to the class of dynamical systems \begin{equation} \frac{d}{dt}\mathbf{y}(t)=F_G(\mathbf{y}(t)) \end{equation} with vector field~$F_G$. Note that~$\mathbf{F}$ is analytic and, thus, there exists unique solution~$\left(\mathbf{y}\left(t,\mathbf{y}(0)\right)\right)$ for each initial condition~$\mathbf{y}(0)\in\mathbb{R}^N$. \section{Concluding remarks}\label{sec:conclusion} In this paper, we constructed a preorder on the set of nodes of a graph. Such preorder entails the classical coarsest equitable partition of the graph. Moreover, it is preserved by certain monotonous dynamical systems over networks, which leads to the existence of an invariant set whose characterization relies on the graph structure of the dynamical system. Note that in light of Corollary~\ref{co:preservesCEP}, such dynamical systems admit a lower dimensional version with underlying dimension given by the number of colors associated with the CEP of the graph, i.e., by having the same initial conditions for nodes with same color, the state of such nodes will be synchronized for all time. Also, Theorem~\ref{th:preservesineq} tells us that one can lower-bound and/or upper-bound solutions to such dynamical systems by properly lower-bounding and/or upper-bounding the corresponding initial conditions. These observations provide with an application to approximate such dynamical systems over complex networks by lower-dimensional ones: i) find the CEP of the graph (which has a cost of about~$n \log n$, with~$n$ being the number of nodes); and ii) initialize evenly the state of nodes with same color so to lower-bound and/or upper-bound the arbitrary initial conditions of interest. In this case, the solutions (with arbitrary initial conditions) can be lower-bounded and/or upper-bounded by its lower dimensional versions for all time~$t$. \small \bibliographystyle{IEEEtran}

\section{Introduction} Colloidal suspensions of superparamagnetic iron-oxide nanoparticles, specifically those of magnetite ({\fo}) and maghemite ({\gfo}), have been extensively investigated for their potential applications such as cell separation, use as contrast agents for magnetic resonance imaging, targeted drug delivery and magnetic fluid hyperthermia (MFH) \cite{PankhurstJPhysD36, PankhurstJPhysD42}. Among these applications, MFH has piqued the interest of researchers from various disciplines including biophysics, biomedicine and oncology due to its potential applications in various technologies for the treatment of cancer without the side effects inherent to radiation and chemotherapy-based methods \cite{DutzNanotech25}. MFH involves the excitation of magnetic nanoparticles suspended in a fluid medium (a ferrofluid) using an oscillating magnetic field of the form $H(t)=H_{0}{\cos}(2{\pi}ft)$, where $H_{0}$ is the field amplitude (typically between 5 - 30 kA/m) and $f$ is the frequency (typically in the range from 150 kHz - 350 kHz). For a single-domain nanoparticle, hysteresis losses are absent and energy absorption from the field can occur via two excitation mechanisms: {\Neel} and Brownian relaxation. This absorbed magnetic energy is eventually transformed to thermal energy and the temperature of the ferrofluid rises. {\Neel} relaxation \cite{BrownPhysRev130} involves the alignment of the nanoparticle moment with the external field within a fixed, non-rotating nanoparticle. As the excitation occurs against the anisotropy energy barrier of the particle the relaxation time depends strongly on the nanoparticle's magnetic volume, $V_{m}$, magnetocrystalline anisotropy constant, $K$, temperature, $T$ and characteristic relaxation time, ${\tau}_{0}$. While there are a few models to describe this relaxation, the most commonly used has the form \cite{BrownPhysRev130} \begin{equation} \tau_{N}=\frac{\sqrt{{\pi}}}{2}{\tau}_{0}{\exp}\left(\frac{KV_{m}}{k_{B}T}\right)\sqrt{\frac{KV_{m}}{k_{b}T}}. \label{EQ:NeelRelax} \end{equation} Brownian relaxation \cite{DebyePolarMolecules} involves the alignment of the particle's moment with the external field via the physical rotation of a fixed-moment nanoparticle within the carrier fluid. This alignment is affected by the hydrodynamic properties of the carrier fluid and nanoparticles, it is described by \cite{FrenkelBook} \begin{equation} {\tau}_{B}=\frac{3{\eta}V_{H}}{k_{B}T}, \label{EQ:BrownRelax} \end{equation} where ${\eta}$ is the viscosity of the carrier fluid and $V_{H}$ is the hydrodynamic volume of the composite particle which includes any surfactant layer used to give colloidal stability to the ferrofluid. When both mechanisms are active these processes occur in parallel and the effective relaxation time which describes the energy transfer rate is given by \begin{equation} {\tau}=\frac{{\tau}_{N}{\tau}_{B}}{{\tau}_{N}+{\tau}_{B}}. \label{EQ:EffectiveRelax} \end{equation} From this relation, it is evident that the shorter relaxation time dominates the overall dissipative characteristics and thus determines the heating rate of the sample. The {\Neel} mechanism plays a dominant role in the relaxational characteristics of the nanoparticles particularly when they are embedded in a cancerous tissue undergoing MFH treatment, since in the tissue's environment, due to immobilization of the particles, the Brownian mechanism is highly damped. Several studies have shown that if the magnetic nanoparticles are internalized by the cancer cells, they are either locked in the cell plasma or adhere to the cell walls \cite{TellingACSnano9}. Consequently, the role of Brownian relaxation is insignificant for nanoparticles used in the hyperthermia treatment of cancer when the ferrofluid is applied directly to the tumor tissues via injection. In such situations, {\Neel} relaxation produces the local heating of the cancerous mass; however, one must be careful about comparing experiments done in the laboratory with therapies that will be performed under real-life conditions. The experimental studies of the heating rates of ferrofluids in the presence of alternating magnetic fields can be broadly placed into one of two categories: those focusing on varying only the intrinsic physical and magnetic parameters of the nanoparticles \cite{DewhirstCancerRes60,TackettJMMM320} and those that investigate the influence of varying only the parameters of the applied field (i.e. $H_{0}$ and $f$) \cite{WeitschiesJPhysCondMat18}. Among these two broad classifications, studies of how the intrinsic parameters of nanoparticles influence magnetic heating still remain an active area of research. However, as the effects of the intrinsic parameters are reflected through the relaxation processes ({\Neel} and Brownian) which act in parallel, it is difficult to understand the contribution of each process on the heating characteristics of the ferrofluid. Knowing the significance of the {\Neel} mechanism in hyperthermia, many studies have been performed to estimate this parameter using ac and dc magnetic susceptibility measurements \cite{FanninJPhysD27,KolitzJMMM149}; however, these methods often require large volumes of sample, accurate determinations of sample volume and mass and costly equipment. In this paper, we present a new method to determine this parameter using a simple induction heating system. This method requires only measurements of the sample heating rate under the influence of an oscillating magnetic field of constant amplitude at two different frequencies. Additional measurements of sample mass, volume and magnetic anisotropy are not needed. The goal of this investigation is to determine the temperature dependence of the {\Neel} relaxation time (${\tau}_{N}$) when the Brownian mechanism is completely quenched. This suppression of the Brownian relaxation is achieved by performing the experiments in the temperature range from 160 K to 210 K where the carrier fluid (DI water) is completely frozen, thus locking the nanoparticles in place. Experimentally, the heating rate in MFH is expressed in terms of the specific absorption rate ($SAR$) which is defined to be the power absorbed per unit mass of the nanoparticles in the ferrofluid. This can be calculated using the thermodynamic relation for rate of heat energy absorbed by the sample as \cite{BabincovaJMMM225} \begin{equation} SAR_{heating}=\frac{M_{sample}}{m_{Fe_{3}O_{4}}}C(T)\frac{{\Delta}T_{heating}}{{\Delta}t}, \label{EQ:SARheating} \end{equation} where $M_{sample}$ is the overall mass of the sample, $m_{Fe_{3}O_{4}}$ is the mass of the {\fo} nanoparticles, $C(T)$ is the temperature dependent specific heat of the carrier fluid and ${\Delta}T_{heating}/{\Delta}t$ is the time rate of change of the sample's temperature as it absorbs energy from the applied magnetic field. When the sample begins to warm under the influence of the applied magnetic field, it is important to note that heat exchange with the environment is also occurring due to the temperature differential between the sample and its surroundings. For accurate determination of $SAR$, it is important to estimate this heat exchange between the sample and the environment due to convective, conductive and radiative processes. While this process is quite hard to do from a calculational standpoint, we can experimentally estimate the collective effects of heat exchange via these processes by measuring the heating of the sample placed into the experimental apparatus while keeping the external field turned off. From this data, we determine the specific power gain (SPG) as heat is transferred from the environment to the sample as \cite{NemalaJAP116} \begin{equation} SPG_{warming}=\frac{M_{sample}}{m_{Fe_{3}O_{4}}}C(T)\frac{{\Delta}T_{warming}}{{\Delta}t}, \label{EQ:SPGwarming} \end{equation} where ${\Delta}T_{warming}/{\Delta}t$ is the warming rate of the sample due only to heat exchange between the sample and the environment. Once this SPG is determined, we can correct the measured $SAR$ for the sample as a function of temperature \cite{NemalaJAP116} \begin{equation} SAR=SAR_{heating}-SPG_{warming} \label{EQ:SARcorrected} \end{equation} Due to the suppression of the Brownian mechanism, the values of $SAR$ obtained over this temperature interval are dependent only upon {\Neel} relaxation. This relaxation time can be determined using a theoretical model based on a direct relationship between the power density dissipated by the nanoparticles and the out-of-phase, dissipative component of the ferrofluid's susceptibility \cite{RosensweigJMMM252} \begin{equation} P={\pi}{\mu}_{0}H_{0}^{2}{\chi}_{0}f\frac{2{\pi}f{\tau}_{N}}{1+(2{\pi}f{\tau}_{N})^{2}}, \label{EQ:PowerDensity} \end{equation} where ${\chi}_{0}$ is the equilibrium susceptibility and ${\mu}_{0}$ is the permeability constant. In this model, the energy dissipation by the nanoparticles, which is related to the out-of-phase component of the magnetic susceptibility, is expressed in terms of the equilibrium susceptibility and the {\Neel} relaxation time ${\tau}_{N}$. Measuring the $SAR$ at two different frequencies $f_{1}$ and $f_{2}$ while keeping the field amplitude ($H_{0}$) fixed, we find ${\tau}_{N}$ from Equation \ref{EQ:PowerDensity} as \begin{equation} {\tau}_{N}=\frac{1}{2{\pi}f_{1}f_{2}}\sqrt{\frac{f_{1}^{2}-{\alpha}f_{2}^{2}}{{\alpha}-1}}, \label{EQ:TauNeelAlpha} \end{equation} where \begin{equation} \alpha=\frac{SAR_{f_{1}}}{SAR_{f_{2}}}=\frac{\left(\frac{{\Delta}T_{heating}}{{\Delta}t}-\frac{{\Delta}T_{warming}}{{\Delta}t}\right)_{f_{1}}}{\left(\frac{{\Delta}T_{heating}}{{\Delta}t}-\frac{{\Delta}T_{warming}}{{\Delta}t}\right)_{f_{1}}}. \label{EQ:alpha} \end{equation} It is interesting to note that ${\tau}_{N}$ is determined only from the values of $f_{1}$, $f_{2}$ and the sample heating rates (corrected for heat exchange with the environment) and do not require measurement of the sample mass, nanoparticle mass, or specific heat values. \section{Experimental Details} Iron oxide nanoparticles were synthesized using a standard co-precipitation technique in which an aqueous solution of FeCl$_{3}\cdot$6H$_{2}$O and FeCl$_{2}\cdot$4H$_{2}$O were mixed in a 2:1 molar ratio and {\fo} nanoparticles were precipitated by the drop-wise addition of 1M NH$_{4}$OH. During precipitation, N$_{2}$ gas was bubbled through the solution to protect against oxidation of the Fe$^{2+}$ ions into Fe$^{3+}$ ions. The precipitate was separated from the solution by a strong magnet, washed with DI water and re-suspended in a metastable 0.5 M NaOH solution. In order to suspend the precipitated nanoparticles in a carrier solution (DI water) they were coated in dextran by the drop-by-drop addition of the metastable solution of {\fo} to a solution of 15-20 kDa dextran (MP Biomedicals) in 0.5M NaOH while simultaneously probe sonicating. The product was rinsed and resulted in a water-based suspension of {\fo} nanoparticles with a concentration of 20 mg of {\fo} per mL of solution. A portion of the sample was lypholized and characterized via x-ray diffraction (Rigaku MiniFlex 600) and transmission electron microscopy (JEOL HR TEM 2010 operating at 200 keV). For calorimetric measurements in the 160 K to 210 K range, the ferrofluid sample was first cooled to 77 K via the immersion of seald vials in liquid nitrogen. The sample was then allowed to warm under ambient conditions to 160 K at which time a ~20 kA/m (250 Oe) ac magnetic field was applied using an Ambrell EasyHeat 2.4 kW induction heating system with a water-cooled, 8-turn, 2-cm-diameter coil. Temperature versus time data were collected in the 160 K to 210 K region using an Optocon FOTEMP1-H fiber optic temperature monitoring system equipped with a TS5 optical temperature sensor with 0.1 K accuracy. The sample vial was thermally insulated using cotton padding and foam rubber to minimize heat exchange with the environment; however, as thermal interaction with the environment cannot be completely suppressed, experiments were performed to determine the extent of the heat exchange with the surrounding environment and this estimation was accounted for in all measurements as discussed above. \section{Results and Discussion} The left panel of Figure \ref{FIG:XRDmag} shows the powder x-ray diffraction (XRD) spectrum taken from the lypholized dextran coated {\fo} nanoparticles. The open symbols represent the observed counts recorded for different $d$-spacing values between 1.25 and 3.25 {\AA}. The solid line is a full-profile Le Bail fit to the data, the vertical bars indicated the $d$-spacing positions of the Bragg reflections and the lower trace is the difference curve between the observed and calculated counts. The fit confirms that the sample consists of a single nanocrystalline phase of {\fo} with cubic $Fm\bar{3}m$ symmetry and lattice constant $a$ = 8.36 {\AA}. Using the full-width at half maximum of the (311) reflection in Scherrer's equation, an average nanoparticle diameter of 14 nm was estimated for this sample. The right panel of Figure \ref{FIG:XRDmag} shows the magnetic field dependence of the ferrofluid's magnetization measured at 150 K which exhibits no hysteresis thus confirming the superparamagnetic nature of the nanoparticles within the experimental temperature range between 160 K and 210 K. Transmission electron microscopy (TEM) was performed to determine the mean diameter and size distribution of the nanoparticles in addition to confirming the crystallographic information determined via XRD. Figure \ref{FIG:TEM}(a) shows a histogram of particle sizes determined from the bright-field TEM micrograph shown in Figure \ref{FIG:TEM}(b) and similar images at the same magnification. The histogram shows a log-normal particle size distribution having mean diameter ${\langle}D{\rangle}$ = 13.4 nm and standard deviation from the mean of ${\sigma}_{D}$ = 4.7 nm. This number is in agreement with the 14 nm average diameter determined using Scherrer's equation in conjunction with the collected XRD spectrum. Figure \ref{FIG:TEM}(c) shows a high-resolution TEM micrograph of a portion of a single nanoparticle. The visible lattice planes in the image show a spacing of 2.9 {\AA} marking them as the (220) set of planes in {\fo}. Lastly, Figure \ref{FIG:TEM}(d) shows a selected area electron diffraction (SAED) pattern which was used to to confirm the cubic symmetry determined using XRD. The left panel of Figure \ref{FIG:HeatingCurves} shows the temperature versus time data collected while heating the ferrofluid using an alternating magnetic field of amplitude $H_{0}$ = 20 kA/m at 171 kHz (open circles), 232 kHz (open triangles) and 343 kHz (open squares). All three curves were fit to a polynomial (solid lines) and differentiated in order to determine the slope, ${\Delta}T/{\Delta}t$. To account for any heat exchange with the environment, a control experiment was performed in which the ferrofluid was placed inside the field coil and allowed to warm in the absence of the magnetic field. This data is plotted in the right panel of Figure \ref{FIG:HeatingCurves} and will be referred to as the ambient curve. From this data, we were able to calculate the specific absorption rate ($SAR_{heating}$) for each frequency and the specific power gain ($SPG_{warming}$) for the ambient experiment all as functions of temperature using the temperature-dependent specific heat of ice \cite{CRRELreport81} and Equations \ref{EQ:SARheating} and \ref{EQ:SPGwarming}, respectively. From these calculations, we were able to determine the specific absorption rate ($SAR$) due only to energy absorbed from the alternating magnetic field using Equation \ref{EQ:SARcorrected}. The results of these calculations are shown in Figure \ref{FIG:SARcurves} and yield average values of $SAR$ over the temperature interval of 57 W/g, 88 W/g and 130 W/g at 171 kHz, 232 kHz and 343 kHz, respectively. Additionally, using the temperature dependent slopes of the data presented in Figure \ref{FIG:HeatingCurves} found through differentiation of the polynomial fits, we are able to use Equations \ref{EQ:TauNeelAlpha} and \ref{EQ:alpha} to calculate the temperature dependence of the {\Neel} relaxation time over the temperature regime in question. The results of this calculation are shown in Figure \ref{FIG:TauCurves} and show a value near 5 $\times$ 10$^{-7}$ s for all possible permutations of $f_{1}$ and $f_{2}$. Using Equation \ref{EQ:NeelRelax} with this determined value of ${\tau}_{N}$, the experimentally determined particle diameter of 13.7 nm, and assuming a characteristic relaxation time of ${\tau}_{0}$ = 10$^{-9}$ s we find the room temperature magnetocrystalline anisotropy constant, $K$, for this sample to be 40 kJ/m$^{3}$, which agrees with the range of values previously reported in the literature \cite{GoyaJAP94, CaruntuJPhysD40}. It is important to emphasize that for quantitative determination of ${\tau}_{N}$ using Equation \ref{EQ:TauNeelAlpha}, one needs the time derivatives of the temperature at two different frequencies and for the ambient (no applied field) case. Using these values in Equation \ref{EQ:alpha}, we determine the temperature dependence of the {\Neel} relaxation time using Equation \ref{EQ:TauNeelAlpha}. In our case, by taking data at three different frequencies, we were able to provide three different estimations of the temperature dependence of ${\tau}_{N}$ and found, as expected, that the value was roughly the same for each pair of frequencies. We see in Figure \ref{FIG:TauCurves} a convergence of the values of ${\tau}_{N}$ as we approach 210 K. In addition, it is important to note that the magnitude of the {\Neel} relaxation time is on the order of 10$^{-7}$ s as expected for a system of non-interacting magnetic nanoparticles having characteristic time constant ${\tau}_{0}{\sim}10^{-9} - 10^{-13}$ s \cite{BrownPhysRev130,WernsdorferPRL78}. The non-interacting nature of this ensemble was confirmed through calculation of the relative variation of the blocking temperature per frequency decade (${\phi}={\Delta}T/T{\log}_{10}f$) in which we found a value of ${\phi}=0.18$, an order of magnitude larger than that found in interacting samples exhibiting spin-glass-like transitions \cite{KodamaPRL77,TholenceSSComm49,DuttaJPhysD43,PisaneJMMM384}. The details of this investigation will be published in a subsequent paper \cite{TackettTBA}. At the macroscopic level, ${\tau}_{N}$, as described by Equation \ref{EQ:NeelRelax}, references the relaxation of an individual nanoparticle of volume $V_{m}$ having magnetocrystalline anisotropy constant $K$ and characteristic relaxation time ${\tau}_{0}=\sqrt{{\pi}KV_{m}/4k_{B}T}$; however, in a polydisperse system these parameters may vary from particle to particle and consequently their {\Neel} relaxation times do vary. In $SAR$ measurements, one measures the temperature rise of the ferrofluid due to the average thermal energy of its constituents (i.e. nanoparticles, surfactant and carrier fluid) in contrast to the relaxation time described by Equation \ref{EQ:NeelRelax} which refers to a single particle. Thus, our measured ${\tau}_{N}$ represents an average {\Neel} relaxation time of the ensemble of particles--this is what is responsible for the measured values of $SAR$. However, for a monodisperse system, Equations \ref{EQ:NeelRelax} and \ref{EQ:TauNeelAlpha} may be applied as what is true for any one particle is true for any other in the system. The use of Equation \ref{EQ:NeelRelax}, however, requires an accurate determination of the particle volume, $V_{m}$, and the magnetocrystalline anisotropy constant, $K$. On the other hand, the method outlined in this paper requires only the value of the applied field frequencies and the derivatives of the temperature versus time curves which can be accurately determined using simple to understand methodology. \section{Conclusion} We studied the temperature dependence of the specific absorption rate ($SAR$) of a frozen ferrofluid at different frequencies at a constant field amplitude. From the determined $SAR$ values we were able to calculate the temperature dependence of the {\Neel} relaxation time using a novel and simple approach. This method is useful for cases in which there are uncertainties related to the magnetocrystalline anisotropy constant, $K$, and nanoparticle size. In our experimental temperature window between 160 K and 210 K, our measured $SAR$ decreases with temperature; however, this trend may not continue as the experimental window is widened to include temperatures above 210 K where other effects may begin to play a role in the relaxation characteristics of the nanoparticles. We believe that this investigation offers a new and simpler method of determining an average {\Neel} relaxation time in a colloidal suspension of magnetic nanoparticles. \section{Acknowledgements} This work was supported by the NSF though DMR-1337615 which provided funding for the Miniflex 600 powder x-ray diffractometer. RJT, REK, CR and PPV would like to thank Dr. James Zhang for his support through the Faculty Research Fellowship. In addition, PPV would like to thank Dr. Zhang for his support through the Rodes Professorship. Lastly, REK would like to thank the Society of Physics Students for the support. \newpage

\section{Introduction} \label{sec:INTRO} Mott insulators with spin-orbit (SO) coupling have recently been topics of great interest because of the plethora of novel phases and behaviors they are expected to exhibit \cite{Jackeli2009, Shitade2009, Pesin2010}. Iridium based transition metal oxides are ideal systems to investigate the novel behaviors predicted to arise due to the interplay of electron correlations and SO coupling \cite{Pesin2010,Wan2011}. In recent years honeycomb lattice iridates A$_{2}$IrO$ _{3} $ (A = Na,Li) have been subjects of intense scrutiny which was fuelled initially by the suggestion of exotic topological properties and Quantum Spin Hall effect ~\cite{Kitaev2006, Shitade2009} and by suggestions that these could be realizations of the Kitaev-Heisenberg model \cite{Jackeli2009, Chaloupka2010}. Na$_2$IrO$_3$ was found to undergo novel magnetic ordering at low temperatures suggesting that it wasn't situated in the strong Kitaev limit where a spin liquid was expected \cite{Singh2010, Singh2012}. Recently however, evidence for dominant bond-directional magnetic exchange and real space-magnetic moment locking has been found in Na$_2$IrO$_3$ \cite{Chun2015}. For ruthenates, the spin-orbit coupling is expected to be comparitively smaller. Nevertheless the compound $\alpha$--RuCl$_3$, which has a network of Ru$^{3+}$ $S = 1/2$ moments on a honeycomb lattice has recently been studied in the quest for a candidate Kitaev material \cite{Banerjee2016}. Observations of a quasi-continuum of excitations in Raman scattering for both Na$_2$IrO$_3$ and $\alpha$-RuCl$_3$ has been argued to be evidence for proximity to the quantum spin liquid state in the dominant Kitaev limit \cite{Gupta2014, Sandilands2015}. More recently, when Ir$^{4+}$ was partially replaced by Ru$^{4+}$ in A$_{2}$Ir$_{1-x}$Ru$_{x}$O$ _{3} $ (A = Na,Li), the materials were found to remain insulating and a spin-glass state is observed at low temperatures highlighting the presence of competing interactions and phases in the parent iridate compounds \cite{Mehlawat2015,Lei2014}. The ruthenate family A$_{2}$RuO$ _{3}$ (A = Na,Li) is also known to adopt a honeycomb lattice structure but with nominal $S = 1$ moments arising from the low-spin state of Ru$ ^{4+}$. Polycrystalline samples of Na$_2$RuO$_3$ were reported to crystallize in the $C2/c$ structure \cite{Mogare2004} similar to early reports on Na$_2$IrO$_3$ \cite{Singh2010}. More recently, single crystals of Na$_2$RuO$_3$ were synthesized and found to crystallize in the related but more symmetric $C2/m$ structure \cite{Wang2014}. Single crystal Na$_2$RuO$_3$ was found to be a local moment magnet which orders antiferromagnetically below $T_N = 30$~K \cite{Wang2014}. The structure and magnetic properties of Li$_2$RuO$_3$ seem to be very sensitive to synthesis conditions and quality of samples. Initial reports on polycrystalline samples suggested a room temperature $C2/c$ monoclinic structure and metallic paramagnetic behavior below $T = 300$~K \cite{Dulac1970, James1988, Felner2002}. Later a comprehensive study on polycrystalline samples of Li$_{2}$RuO$ _{3}$ revealed an unusual second order structural phase transition near $T \approx 540$~K from a nearly perfect honeycomb lattice $C2/m$ structure at high temperature to a low temperature structure with a distorted honeycomb lattice P2$ _{1}/m$ \cite{Miura2007}. This structural transition was acompanied by an increase in resistance and loss of magnetization. Nearly perfect hexagons of the high temperature $C2/m$ phase undergo strong distortion, leading to a low temperature structure with significant shortening of one of the three inequivalent Ru-Ru bonds on each honeycomb \cite{Miura2007}. Based on DFT calculations on the low and high temperature structures it was proposed that Li$_{2}$RuO$ _{3}$ undergoes a transition from a highly correlated metal to a molecular orbital insulator involving Ru-Ru dimerization and spin-singlet formation \cite{Miura2007, Miura2009}. An alternative mechanism of spin-singlet formation driven by magnetoelastic coupling has also been proposed \cite{Jackeli2008}. The evolution of the structural Ru-Ru dimers across the phase transition has been studied recently using pair distribution function (PDF) analysis of high energy powder X-ray data. The PDF analysis allows the tracking of short-ranged structural order. It was found that dynamically fluctuating dimers survive at temperatures well above the transition temperature $T \approx 540$~K \cite{Kimber2014}. This suggests a scenario where a valence bond crystal in the low temperature phase melts into a valence bond liquid at high temperatures. Such a scenario is supported by recent Ru site dilution experiments \cite{Park2016}. An electronic structure study has highlighted the importance of electronic correlations and proposed that a combination of local-moment behavior and molecular orbital formation could be the correct picture for this material \cite{Pchelkina2015} Recently a careful study of the effect of synthesis conditions on the structure and magnetic behavior of polycrysatalline samples has been carried out \cite{Segura2016}. It was found that all samples crystallized in the $P2_1/m$ structure at room temperature and showed the Ru-dimerization transition at high temperatures. However, the details of the structure and the magnetic properties strongly depends on the synthesis conditions. The best quality samples revealed that the magneto-structural transition is first-order in nature with a much higher onset temperature of $\approx 550$~K \cite{Segura2016}. Lastly, single crystals of Li$_2$RuO$_3$ have recently been synthesized. The crystals are found to crystallize at room temperature in either the $C2/m$ or the $P2_1/m$ structures depending on synthesis conditions. However, in complete contrast to all existing polycrystalline work \cite{Miura2007, Kimber2014, Park2016, Segura2016}, neither of these crystals show the magneto-structural transitions at high temperature. They instead show Curie-Weiss behavior below $300$~K and magnetic ordering at low temperatures into supposedly antiferromagnetic states \cite{Wang2014}. In this work we report the first crystal growth of Li$_2$RuO$_3$ and $5\%$ Na substituted Li$_2$RuO$_3$ crystallizing in the $P2_1/m$ structure at room temperature and showing the magneto-structural transition at high temperatures. We are therefore able to study for the first time the magnetic anisotropy across the high temperature magneto-structural transition. We observe that for Li$_2$RuO$_3$ the transition might occur in two steps with a first-order structural transition occurring first (onset $\approx 570$~K) which then drives the magnetic Ru-Ru dimerization transition ($\approx 540$~K). Replacing just $5\%$ Li by Na leads to a reversal of the magnetic anisotropy although the first-order magneto-structural transition is still seen at $\approx 540$~K\@. Room temperature structural studies show that the Ru-Ru structural dimerization arrangement is also changed in the Na substituted samples. While the Li$_2$RuO$_3$ shows 2 short and 4 long Ru-Ru bonds on each honeycomb in an armchair pattern as previously seen \cite{Miura2007,Jackeli2008}, the Na doped samples show 4 short and 2 long Ru-Ru bonds on each honeycomb in an arrangement which can be viewed as two inter-penetrating armchair patterns. \section{EXPERIMENTAL DETAILS} \label{sec:EXPT} The single crystalline samples of (Li$_{1-x}$Na$_{x}$)$ _{2} $RuO$_3$ ($x = 0, 0.05$) have been synthesized. The starting materials were Li$_2$CO$_3$ (99.995$\%$ Alfa Aesar, Na$_2$CO$_3$ (99.995$\%$ Alfa Aesar) and Ru metal powder (99.95$\%$ Alfa Aesar). Single crystals were grown using a self flux growth method. Off-stoichiometric amounts of starting materials were mixed and placed in an alumina crucible with a lid, heated to $750~^\circ$C for 24~h for calcination and then furnace cooled to room temperature. Crystal growth was done by keeping the calcined mixture for long periods ($70$--$80$~h) at temperatures between $1000~^\circ$C -- $1050~^\circ$C after which the furnace is turned off and allowed to cool to room temperature. Shiny plate like single crystals (size $\sim 0.5 \times 0.5 \times 0.03$) were found to grow on top of semi-melted polycrystalline powder. Growth of crystals with higher Na concentrations were tried but were not successful. The structure and composition of the resulting samples were checked by single-crystal and powder x-ray diffraction (PXRD), and chemical analysis using energy dispersive x-ray (EDX) analysis with a JEOL scanning electron microscope (SEM). The PXRD was obtained by a Rigaku diffractometer with Cu K$_{\alpha}$ radiation in 2$ \theta $ range from 10$^\circ$ to 90$^\circ$ with 0.02$^\circ$ step size. Anisotropic magnetic susceptibility measurements upto $T = 1000$~K were measured on a collection of co-aligned crystals with total mass $\approx 12$~mg using the VSM Oven option on a Quantum Design physical property measurement system. \section{RESULTS} \begin{figure}[t] \includegraphics[width= 3 in]{Fig1.eps} \caption{(Color online) Rietveld refinements of powder x-ray diffraction data for (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$. The solid circles represent the observed data, the solid lines through the data represent the fitted pattern, the vertical bars represent the peak positions, and the solid curve below the vertical bars is the difference between the observed and the fitted patterns. \label{Fig-1}} \end{figure} \subsection{Crystal Structure and Chemical Analysis} From room temperature single crystal and powder x-ray diffraction, we conclude that all samples adopt the $P2 _{1} $/m space group. A full single crystal refinement was not possible because the crystals have multiple twins rotated around the $c^*$ axis. However, it was possible to determine the space group and cell parameters using single crystal diffraction. Cell parameters were also obtained by performing Rietveld refinements of the PXRD patterns obtained on the crushed crystals. Fig.\ref{Fig-1} shows representative results of Rietveld refinement of the PXRD patterns for (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$. The fractional atomic positions obtained from the refinement are given in Table~\ref{Table-Structural-parameters}. The unit cell parameters and the relevant bond lengths extracted from Rietveld refinement of the powder diffraction data are listed in Table~\ref{Table-lattice-parameters}. The cell parameters change significantly (specially the $a$-axis) as Li is partially replaced by Na. The presence of Na in the doped crystals and its concentration relative to Ru was confirmed using energy dispersive X-ray spectroscopy on several spots on the same crystal and on several crystals and was found to be close to the nominal concentration targeted in the starting material. The room temperature crystal structure of Li$_2$RuO$_3$ and Li viewed perpendicular to the honeycomb planes is shown in Fig.~\ref{Fig-structure} to highlight the Ru-Ru dimerization pattern. For Li$_2$RuO$_3$ we find, consistent with previous work, that one ($d_2$) out of the three inequivalent Ru-Ru bonds is considerably shorter compared to the other two which are of similar lengths. Surprisingly, for just 5\% Na substitution for Li, the dimerization pattern changes and we now have two short ($d_2$ and $d_3$) and one long bonds. The Ru-Ru bond lengths are given in Table~\ref{Table-lattice-parameters} and the dimerization patterns shown in Figs.~\ref{Fig-structure}~(a) and ~(b), respectively. For Li$_2$RuO$_3$ as observed \cite{Miura2007} and explained \cite{Jackeli2008} previously the dimers on the $d_2$ bond form an armchair pattern. For the Na substituted sample, both $d_2$ and $d_3$ bonds dimerize and form inter-penetrating armchairs which run along the $a$-axis. \begin{table} \caption{Wyckoff position for (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ obtained from Rietveld refinements of polycrystal x-ray data at 300 K}. \begin{ruledtabular} \begin{tabular}{|ccccccc|} $Atom$ & $Wyckoff$ & $x$ & $y$ & $z $& $Occ$ & $B$ $({\AA})$ \\ \hline Ru& 4f & 0.2467(7) & 0.0776(8) & -0.0038(7) & 1 & 0.0265 \\ Li1 & 2e & 0.7857(5) & 0.25 & -0.0295(8) & 0.95 & 0.0800 \\ Na1 & 2e & 0.7857(5) & 0.25 & -0.0295(8) & 0.05 & 0.0900 \\ Li2 & 4f & 0.0661(3) & 0.25 & 0.6213(7) & 1 & 0.0034 \\ Li3 & 2e & 0.6887(3) & 0.0523(5) & 0.4685(6) & 1 & 0.0020 \\ O1 & 4f & 0.7812(6) & 0.0644(7) & 0.2831(8) & 1 & 0.0043 \\ O2 & 4f & 0.7502(5) & 0.0957(7) & 0.7931(2) & 1 & 0.0060 \\ O3 & 2e & 0.3124(7) & 0.25 & 0.2688(5) & 1 & 0.0088 \\ O4 & 2e & 0.2396(8) & 0.25 & 0.2373(4) & 1 & 0.0080 \\ \end{tabular} \end{ruledtabular} \label{Table-Structural-parameters} \end{table} \begin{table} \caption{Summary of Lattice Parameters and relevant bond lengths of (Li$_{1-x}$Na$_{x}$)$ _{2} $RuO$_3$ ($x \approx 0, 0.05 $)} \begin{ruledtabular} \begin{tabular}{|c|c|c|} Sample & Li$ _{2} $RuO$_3$ & (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ \\ \hline $Space $ $Group$ & P2$ _{1} $ /m & P2$ _{1} $ /m \\ \hline $a $ $({\AA}) $ & 4.920(4) & 4.934(5) \\ \hline $b $ $({\AA})$ & 8.781(7) & 8.774(4) \\ \hline $c $ $({\AA})$ & 5.893(3) & 5.895(6) \\ \hline $\beta $ $(deg)$ & 124.36(4) & 124.42(6) \\ \hline $V$ $({\AA})$ & 210.452( 5)& 210.452(5) \\ \hline $Ru-Ru$ $({\AA})$ & & \\ \hline d$_{1}$ & 3.024 & 3.025 \\ \hline d$_{2}$ & 2.632 & 2.812 \\ \hline d$_{3}$ & 2.937 & 2.823 \\ \end{tabular} \end{ruledtabular} \label{Table-lattice-parameters} \end{table} \begin{figure}[t] \includegraphics[width= 3 in]{Fig2.eps} \caption{(Color online) Room temperature structure of (a) Li$_2$RuO$_3$ and (b) (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ viewed perpendicular to the Ru honeycomb network in the $ab$-plane. There are three inequivalent Ru-Ru bonds in the honeycomb network labled as $d_1$ (blue), $d_2$ (red), and $d_3$ (black). For Li$_2$RuO$_3$, the Ru-Ru dimerization happens on the $d_2$ bonds (shown as the thicker red bonds in (a)) which are considerably shorter than $d_1$ and $d_3$ which are of similar length. The armchair pattern observed for Li$_2$RuO$_3$ is consistent with that observed earlier \cite{Miura2007}. For (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ the dimerization pattern changes and there are two short bonds $d_2$ (thick red) and $d_3$ (thick black) and one long bond $d_1$ (thin blue). The dimer arrangement can be viewed as two inter-penetrating armchair patterns. \label{Fig-structure}} \end{figure} \begin{figure}[t] \includegraphics[width= 3 in]{Fig3a.eps} \includegraphics[width= 3 in]{Fig3b.eps} \caption{(Color online) (a) Anisotropic Magnetic susceptibility $\chi_{||}$ and $\chi_{\perp}$ versus $T$ measured at in a magnetic field of $5$~T for Li$_2$RuO$_3$ between $T = 2$~K and 1000~K\@. (b) $\chi_{||}$ and $\chi_{\perp}$ versus $T$ in the temperature range $510$~K to $590$~K to highlight the behaviour near the transition. \label{Fig-2}} \end{figure} \begin{figure}[t] \includegraphics[width= 3 in]{Fig4.eps} \caption{(Color online) Anisotropic Magnetic susceptibility $\chi_{||}$ and $\chi_{\perp}$ versus $T$ measured at in a magnetic field of $5$~T for (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ between $T = 300$~K and 1000~K\@. The inset shows the $\chi_{||}$ and $\chi_{\perp}$ versus $T$ in the temperature range $510$~K to $570$~K to highlight the behaviour near the transition. \label{Fig-3}} \end{figure} \subsection{ Magnetic susceptibility} \subsubsection{Li$_2$RuO$_3$} The magnetic susceptibility $\chi$ versus T data for Li$ _{2} $RuO$_3$ measured in an applied magnetic field $H=5$~T applied parallel to the honeycomb plane ($\chi_{||}$) or perpendicular to the honeycomb plane ($\chi_{\perp}$) are shown in Fig.~\ref{Fig-2}. Figure~\ref{Fig-2}(a) shows the $\chi_{\perp}$ data from $300$~K to $1000$~K and the $\chi_{||}$ data from $2$~K to $1000$~K\@. Both sets of data were measured while cooling down from $1000$~K\@. The first thing to note is that $\chi_{||}>\chi_{\perp}$ for all temperatures. The $\chi(T)$ behavior at high temperatures is not Curie-Weiss like as expected for a paramagnet. Instead the $\chi(T)$ behavior is consistent with a quasi-two-dimensional magnetic system having stronger in-plane interactions. We also see evidence for a transition involving an abrupt drop in $\chi$ below about $550$~K\@. This is a signature of the magneto-structural transition observed previously for polycrystalline samples \cite{Miura2007}. The magneto-structural transition has been previously reported to involve a structural change below $540$~K from $C2/m$ to $P2_1/m$ symmetry and a simultaneous Ru-Ru dimerizations with spin-singlet formation \cite{Miura2007}. The abrupt drop in $\chi$ at the transition is consistent with Ru-Ru spin-singlet formation. The magnitude of the drop can be quantified by $\chi_{min}/\chi_{max}$ and is $\approx 0.45$ for both $\chi_{||}$ and $\chi_{\perp}$. Below $300$~K, the $\chi(T)$ is $T$ independent and small but finite. This $T$ independent finite value ($\chi_{||} \approx 2.75 \times 10^{-4}$~cm$^3$/mol) is most likely a Van Vleck paramagnetic contribution \cite{Khaliullin2013}. In Fig.~\ref{Fig-2}~(b) we show the $\chi_{\perp}$ and $\chi_{||}$ data on an expanded scale around the region of the transition. Data were recorded while warming from $300$~K to $1000$~K and then while cooling back again at a rate of $5$~K/min. We see that there is a thermal hysteresis between the warming and cooling data indicating the first order nature of the phase transition. The transition temperatures obtained by taking derivatives of the data (not shown) are listed in Table~\ref{Table-Transition Temperature}. For $\chi_{||}$ we get the transition temperatures $544$~K for warming and maybe a double transition at $540$~K and $546$~K for cooling measurements. The thermal hysteresis is about $5$~K\@. For $\chi_{\perp}$ the situation is more complex. The transition clearly happens in two steps as indicated by the vertical arrows close to the data in Fig.~\ref{Fig-2}(b) signalling the onset of the two transitions. A derivative of the $\chi_{\perp}$ data shows two peaks which are taken as the approximate transition temperatures and listed in Table~\ref{Table-Transition Temperature}. We note that the lower transition is sharp and is not accompanied by any significant thermal hysteresis whereas the higher temperature transition is broad and clearly hysteretic. The hysteresis in the higher temperature transition is about $6$~K as observed for the $\chi_{||}$ data. We will return to a discussion of these data in a later section. \begin{table} \caption{ Temperatures of the peaks in $d\chi/dT$ for $\chi_{||}$ and $\chi_{\perp}$ of single crystalline Li$_{2} $RuO$_3$ at $H = 5$~T} \begin{ruledtabular} \begin{tabular}{|ccc|} Magnetic susceptibilities & $T_{1}$ & $T_{2}$ \\ \hline Li$_2$RuO$_3$ \\ \hline $\chi_{||}$ (heating) & $540$~K & $546$~K \\ $\chi_{||}$ (cooling) & & 544$K$ \\ $\chi_{\perp}$ (heating) & $538.3$~K & $561$~K \\ $\chi_{\perp}$ (cooling) & $537.8$~K & $555$~K \\ \hline (Li$_{0.095}$Na$_{0.05}$)$_2$RuO$_3$ \\ \hline $\chi_{||}$ (heating) & $542$~K & \\ $\chi_{||}$ (cooling) & $535$~K & \\ $\chi_{\perp}$ (heating) & $538$~K & \\ $\chi_{\perp}$ (cooling) & $531$~K & \\ \end{tabular} \end{ruledtabular} \label{Table-Transition Temperature} \end{table} \subsubsection{(Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$} The magnetic susceptibility $\chi$ versus T data for (Li$_{0.95}$Na$_{0.05}$)$ _{2}$RuO$_3$ measured between $300$~K to $1000$~K in an applied magnetic field $H=5$~T applied parallel to the honeycomb plane ($\chi_{||}$) or perpendicular to the honeycomb plane ($\chi_{\perp}$) are shown in the main panel in Fig.~\ref{Fig-3}. Surprisingly, with only a $5\%$ Na substitution for Li, the anisotropy is reversed ($\chi_{\perp} > \chi_{||}$) compared to what was observed for Li$_2$RuO$_3$. The magneto-structural transition can be seen in both sets of data. The Fig.~\ref{Fig-3}~inset shows the $\chi_{||}$ and $\chi_{\perp}$ data in the temperature range $510$~K to $570$~K to highlight the transition. Data were recorded while warming from $300$~K to $1000$~K and then while cooling back again at a rate of $5$~K/min. We see that there is a thermal hysteresis between the warming and cooling data indicating that the first order nature of the phase transition persists in Na substituted samples. Peaks in the derivatives of the $\chi(T)$ data are taken as the approximate transition temperatures and are given in Table~\ref{Table-Transition Temperature}. \section{SUMMARY AND DISCUSSION} We have grown the first single crystals of (Li$_{1-x}$Na$_{x}$)$ _{2} $RuO$_3$ ($x = 0, 0.05$) crystallizing in the $P2_1/m$ structure at room temperature and showing a magneto-structural transition at high temperatures. Using magnetic susceptibility $\chi$ measurements for temperatures $T \leq 1000$~K we observe that for Li$_2$RuO$_3$, $\chi_{||} > \chi_{\perp}$. Additionally, we observe a first-order high temperature coupled magneto-structural transition which seems to occur in two steps. This is most evident in the $\chi_{\perp}$ data. The higher temperature transition has an onset as high as $T > 570$~K and a mid-point around $T \approx 561$~K as seen by the peak in $d\chi_{\perp}/dT$ measured while warming up to $1000$~K\@. This high temperature transition is hysteretic with a thermal hysteresis of $6$~K indicating its first-order nature. The lower temperature transition in $\chi_{\perp}$ occurs at $T \approx 538$~K, is very sharp, is accompanied by an abrupt fall in $\chi$, and with almost no thermal hysteresis. These observations suggest that the higher temperature, hysteretic transition is the structural dimerization transition while the lower temperature transition where we observe a sharp fall in $\chi$ is the magnetic transition involving Ru-Ru singlet formation. Thus for Li$_2$RuO$_3$ the two transitions most likely occur at slightly different temperatures with the structural dimerization transition occuring first and triggering the magnetic Ru-Ru singlet formation. The onset temperature of $570$~K is much higher than previously observed ($\approx 540$~K) and indicates the high quality of the samples. Just a $5\%$ substitution of Na for Li leads to interesting magnetic and structural changes. The high temperature $\chi(T)$ data show that the magnetic anisotropy is reversed compared to Li$_2$RuO$_3$ with $\chi_{\perp} > \chi_{||}$ for (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$. The arrangement of Ru-Ru dimers on the honeycomb lattice also changes. For Li$_2$RuO$_3$ Rietveld refinements of room temperature powder X-ray data reveal that one ($d_2$) out of the three inequivalent Ru-Ru bonds on the honeycomb lattice is shortened compared to the other two which are almost equal to each other as can be seen in Table~\ref{Table-lattice-parameters}. For (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ we find that two ($d_2$ and $d_3$) out of the three Ru-Ru bonds are smaller and almost equal while the third is much larger. The armchair arrangement of the dimers in Li$_2$RuO$_3$ is consistent with previous reports \cite{Miura2007, Jackli2008}. The dimer arrangement in (Li$_{0.95}$Na$_{0.05}$)$ _{2} $RuO$_3$ can be viewed as two inter-penetrating armchairs formed on the $d_2$ and $d_3$ bonds, respectively. This suggests a possible change in the orbital ordering pattern for the Na substituted sample. \paragraph{Acknowledgments.--} We thank the X-ray facility at IISER Mohali for powder XRD measurements. YS acknowledges DST, India for support through Ramanujan Grant \#SR/S2/RJN-76/2010 and through DST grant \#SB/S2/CMP-001/2013. KM acknowledges UGC India for a fellowship.

Seino Y., Terauchi Y., Osonoi T. et al. Safety and efficacy of semaglutide once weekly vs sitagliptin once daily as monotherapy in Japanese people with type 2 diabetes. Diabetes Obes Metab. 2018;20:378--388. <https://doi.org/10.1111/dom.13082> 28786547 **Funding information** This trial was supported by Novo Nordisk A/S; NCT02254291. 1. INTRODUCTION {#dom13082-sec-0005} =============== The prevalence of type 2 diabetes (T2D) and prediabetes increased significantly from the 1980s to the 2000s in the Japanese population, with increasing obesity and declining physical activity.[1](#dom13082-bib-0001){ref-type="ref"} Achieving glycated haemoglobin (HbA1c) targets is a challenge for many patients with T2D.[2](#dom13082-bib-0002){ref-type="ref"} Avoidance of weight gain and hypoglycaemia are important considerations in selecting appropriate treatment and individualizing treatment goals.[3](#dom13082-bib-0003){ref-type="ref"}, [4](#dom13082-bib-0004){ref-type="ref"} Unlike most other T2D therapies, glucagon‐like peptide‐1 (GLP‐1) receptor agonists (RAs) can achieve glycaemic control and reduce body weight, with a low risk of hypoglycaemia.[5](#dom13082-bib-0005){ref-type="ref"} Semaglutide is a GLP‐1 analogue in development for the treatment of T2D, with a 94% homology to native GLP‐1.[6](#dom13082-bib-0006){ref-type="ref"} A half‐life of \~1 week makes it appropriate for once‐weekly administration.[6](#dom13082-bib-0006){ref-type="ref"}, [7](#dom13082-bib-0007){ref-type="ref"} Sitagliptin, a once‐daily dipeptidyl peptidase‐4 (DPP‐4) inhibitor, is a widely used oral antidiabetic drug (OAD), and the most commonly used DPP‐4 inhibitor in Japan.[8](#dom13082-bib-0008){ref-type="ref"} DPP‐4 inhibitors achieve glycaemic control by inhibiting DPP‐4‐dependent inactivation of both GLP‐1 and gastric inhibitory polypeptide (GIP), thereby enhancing GLP‐1 and GIP receptor signalling, and thus are distinct from GLP‐1 RAs, which stably activate GLP‐1 receptor signalling.[9](#dom13082-bib-0009){ref-type="ref"} Semaglutide has been evaluated in the "Semaglutide Unabated Sustainability in Treatment of Type 2 Diabetes" (SUSTAIN) programme, consisting of six phase III global clinical trials that evaluated the efficacy and safety (including cardiovascular outcomes) of semaglutide vs a range of comparators, including sitagliptin.[10](#dom13082-bib-0010){ref-type="ref"}, [11](#dom13082-bib-0011){ref-type="ref"}, [12](#dom13082-bib-0012){ref-type="ref"}, [13](#dom13082-bib-0013){ref-type="ref"}, [14](#dom13082-bib-0014){ref-type="ref"}, [15](#dom13082-bib-0015){ref-type="ref"} Two additional phase III clinical trials in the SUSTAIN programme investigated the effect of semaglutide in Japanese populations.[16](#dom13082-bib-0016){ref-type="ref"}, [17](#dom13082-bib-0017){ref-type="ref"} In the present trial we evaluated the safety and efficacy of 30 weeks of treatment with once‐weekly semaglutide (0.5 and 1.0 mg) vs once‐daily sitagliptin (100 mg), both as monotherapy, in Japanese adults with T2D who were previously stable on diet/exercise or OAD monotherapy.[17](#dom13082-bib-0017){ref-type="ref"} 2. MATERIALS AND METHODS {#dom13082-sec-0006} ======================== 2.1. Trial design {#dom13082-sec-0007} ----------------- This was a phase III randomized (1:1:1), open‐label, parallel‐group, active‐controlled, single‐country (Japan), multicentre trial (NCT02254291; Figure [S1](#dom13082-supitem-0001){ref-type="supplementary-material"}). This trial was conducted to meet the "Guideline for Clinical Evaluation of Oral Hypoglycaemic Agents" issued by the Japanese Ministry of Health, Labour and Welfare,[18](#dom13082-bib-0018){ref-type="ref"} which advises that new drugs are assessed as monotherapy to investigate their isolated effects. A specific comparator is not stipulated in the guideline; however, sitagliptin was chosen as an active comparator because it is widely available in Japan,[8](#dom13082-bib-0008){ref-type="ref"} and has a well‐known efficacy and safety profile. The trial was conducted in compliance with the International Conference on Harmonisation Good Clinical Practice guidelines[19](#dom13082-bib-0019){ref-type="ref"} and the Declaration of Helsinki.[20](#dom13082-bib-0020){ref-type="ref"} Written informed consent was obtained from all participants. 2.2. Trial population {#dom13082-sec-0008} --------------------- Japanese men and women were eligible for inclusion if they were aged ≥20 years, diagnosed with T2D, and treated with either diet and exercise therapy in addition to OAD monotherapy if their HbA1c levels were 6.5% to 9.5% (48--80 mmol/mol; hereafter called OAD monotherapy), or treated with diet and exercise therapy only if their HbA1c levels were 7.0% to 10.5% (53--91 mmol/mol) for at least 30 days before screening. Key exclusion criteria were: treatment with glucose‐lowering agents (except for pre‐trial OAD in participants treated with OAD monotherapy) in the 60 days before screening (except for short‐term use of insulin in connection with intercurrent illness of ≤7 days); history of chronic or idiopathic acute pancreatitis; screening calcitonin value ≥50 ng/L; personal or family history of medullary thyroid carcinoma or multiple endocrine neoplasia syndrome type 2; impaired renal function (estimated glomerular filtration rate \< 60 mL/min/1.73 m^2^ per modification of diet in renal disease formula \[4‐variable version\]); acute coronary or cerebrovascular event within 90 days before randomization; or heart failure, New York Heart Association class IV. Full inclusion and exclusion criteria are listed in Table [S1](#dom13082-supitem-0001){ref-type="supplementary-material"}. 2.3. Randomization and masking {#dom13082-sec-0009} ------------------------------ Participants were randomly assigned using an interactive voice/web response system in a 1:1:1 ratio to receive once‐weekly subcutaneous (s.c.) semaglutide (0.5 or 1.0 mg) or once‐daily oral sitagliptin 100 mg. Randomization was stratified according to pre‐trial treatment at screening (diet and exercise therapy or OAD monotherapy). 2.4. Drug administration {#dom13082-sec-0010} ------------------------ After an 8‐week washout (OAD monotherapy group) or 2‐week screening period (diet and exercise only group), participants received s.c. semaglutide 0.5 or 1.0 mg once weekly or sitagliptin 100 mg once daily for 30 weeks, followed by a 5‐week follow‐up period (Figure [S1](#dom13082-supitem-0001){ref-type="supplementary-material"}). Participants in the semaglutide arms followed a fixed dose‐escalation regimen of semaglutide 0.5 mg (maintenance dose reached after 4 weeks of 0.25 mg semaglutide once weekly) or semaglutide 1.0 mg (maintenance dose reached after 4 weeks of 0.25 mg semaglutide, followed by 4 weeks of 0.5 mg semaglutide). In case of a safety concern or unacceptable intolerability, the trial product could be discontinued at the investigator's discretion and was not to be reinitiated. Participants discontinuing trial product prematurely were asked to continue with the scheduled site contact (if necessary, in order to retain the participant in the trial, site visits were replaced by phone contacts after discontinuation) and, as a minimum, were asked to attend the visits for end of treatment and follow‐up at the time of the scheduled completion of the trial. 2.5. Trial endpoints and assessments {#dom13082-sec-0011} ------------------------------------ The primary endpoint was the number of treatment‐emergent adverse events (TEAEs) during 30 weeks of treatment. Supportive secondary efficacy endpoints included change from baseline to week 30 in: HbA1c; fasting plasma glucose (FPG); self‐measured plasma glucose (SMPG \[measurements performed with capillary blood were automatically calibrated to plasma‐equivalent glucose values, shown on the display of the blood glucose (BG) meter and documented by the trial participant\]; mean 7‐point profile and mean postprandial increment, over all meals); body weight; body mass index (BMI) and waist circumference; fasting blood lipid levels (free fatty acids \[FFAs\], total cholesterol, HDL cholesterol, LDL cholesterol, VLDL cholesterol and triglycerides); and systolic and diastolic blood pressure (BP). Other efficacy endpoints included the proportion of participants reaching HbA1c targets of ≤6.5% (American Association of Clinical Endocrinologists \[AACE\]),[21](#dom13082-bib-0021){ref-type="ref"} \<7.0% (American Diabetes Association \[ADA\] and Japan Diabetes Society \[JDS\])[22](#dom13082-bib-0022){ref-type="ref"}, [23](#dom13082-bib-0023){ref-type="ref"} and \<7.0% without severe or BG‐confirmed symptomatic hypoglycaemia (defined as an episode that was severe according to ADA classification[22](#dom13082-bib-0022){ref-type="ref"} or BG‐confirmed by plasma glucose \<3.1 mmol/L, with symptoms consistent with hypoglycaemia) and no weight gain, and proportion of participants achieving weight loss of ≥5% or ≥10% at week 30. Supportive secondary safety endpoints included: the number of treatment‐emergent severe or BG‐confirmed symptomatic hypoglycaemic episodes during 30 weeks of treatment; change from baseline to week 30 in laboratory and physical examination variables; and occurrence of anti‐semaglutide antibodies. An external Event Adjudication Committee (EAC) validated selected adverse events (AEs), according to predefined diagnostic criteria, in an independent, blinded manner (Table [S2](#dom13082-supitem-0001){ref-type="supplementary-material"}). Events were adjudicated according to Food and Drug Administration guidance and requirements. 2.6. Statistical analysis {#dom13082-sec-0012} ------------------------- Trial sample size was determined based on the "Guideline for Clinical Evaluation of Oral Hypoglycaemic Agents."[18](#dom13082-bib-0018){ref-type="ref"} With the planned number of randomized participants of 306, a total of 81 participants randomized to each group were expected to complete treatment, assuming a treatment discontinuation rate of 20%. Together with the other trials in the SUSTAIN programme, ≥300 Japanese participants were expected to complete ≥6 months of treatment with semaglutide monotherapy. Randomized participants receiving at least 1 dose of trial product comprised the full analysis set and the safety analysis set. A TEAE was defined as an event with onset from first exposure to the follow‐up visit scheduled 5 weeks (+1 week visit window) after the last trial product dose. Main efficacy evaluations were based on assessments collected in the period where participants were treated with trial product, without rescue medication ("on‐treatment without rescue medication" period). Supportive analyses of efficacy and safety were based on assessments collected in the period where participants, after randomization, were considered trial participants and where data were to be collected systematically ("in‐trial" period). Continuous efficacy endpoints assessed over time were analysed using a mixed model for repeated measurements (MMRM) with treatment and pre‐trial treatment at screening as fixed factors and baseline value as a covariate, all nested within visits. With the exception of HbA1c, body weight, FPG, SMPG, BMI, waist circumference and BP, values were log‐transformed before analysis. Efficacy continuous endpoints assessed at baseline and week 30 were analysed using an analysis of covariance (ANCOVA) model. For the HbA1c target endpoints (\<7% and ≤6.5%) and weight loss response endpoints (≥5% and ≥10%), missing data at week 30 were imputed from the MMRM, used for the corresponding continuous endpoint, and subsequently classified. Treatment comparison was based on a logistic regression model including the same fixed factors and associated baseline value(s) as covariate, except for the weight‐loss response of ≥10%, which was compared using a Fisher's exact test, as no participants achieved this response in the sitagliptin group. The sensitivity of the main results of HbA1c and body weight was assessed using complete case analyses (MMRM‐based), ANCOVA using the last observation carried forward (LOCF) imputation, and comparator‐based multiple imputation. Data collected throughout the trial, regardless of whether participants discontinued treatment prematurely or initiated rescue medication, were also analysed ("in‐trial" analysis; MMRM based); thus, the amount of missing data was expected to be small. 3. RESULTS {#dom13082-sec-0013} ========== 3.1. Participant disposition and baseline characteristics {#dom13082-sec-0014} --------------------------------------------------------- In total, 308 adults with T2D were randomized to receive 1 of the 2 semaglutide maintenance doses or sitagliptin, all of whom were exposed to trial product (Figure [1](#dom13082-fig-0001){ref-type="fig"}). A higher proportion of participants discontinued treatment prematurely in the semaglutide 1.0 mg group (14.7%) compared with the semaglutide 0.5 mg and sitagliptin groups (each 2.9%). All semaglutide 0.5 mg‐treated participants completed the trial, whereas 3 participants in the semaglutide 1.0 mg group and 2 in the sitagliptin group withdrew prematurely. Rescue medication was provided to 1 participant in the semaglutide 0.5 mg group and 5 participants in the sitagliptin group. {#dom13082-fig-0001} Baseline characteristics were generally well balanced between the three groups (Table [1](#dom13082-tbl-0001){ref-type="table-wrap"}). The majority (76.3%) of participants were men, mean HbA1c was 8.1%, mean diabetes duration was 8.0 years, and the proportion of participants randomized while on pre‐trial OAD treatment was 29.9%. Mean body weight was 69.3 kg, although participants in the semaglutide 1.0 mg group were 3.0 kg heavier than in the semaglutide 0.5 mg group (70.8 vs 67.8 kg). ###### Baseline characteristics of trial populations Semaglutide 0.5 mg Semaglutide 1.0 mg Sitagliptin 100 mg Total ---------------------------------------------------------------- -------------------- -------------------- -------------------- ------- ----------- ------ ----------- ------ Age, years 58.8 10.4 58.1 11.6 57.9 10.1 58.3 10.7 Male/female, % 76.7/23.3 ‐ 73.5/26.5 ‐ 78.6/21.4 ‐ 76.3/23.7 ‐ HbA1c, % 8.2 1.0 8.0 0.9 8.2 0.9 8.1 0.9 FPG, mmol/L 9.2 2.1 9.2 1.8 9.5 2.0 9.3 2.0 Diabetes duration, years 8.0 5.2 7.8 6.9 8.1 6.7 8.0 6.3 Body weight, kg 67.8 11.7 70.8 16.4 69.4 12.9 69.3 13.8 BMI, kg/m^2^ 25.1 3.8 26.1 5.2 25.1 3.6 25.4 4.3 Previously treated with: Diet and exercise therapy, % 69.9 ‐ 70.6 ‐ 69.9 ‐ 70.1 ‐ OAD therapy, % 30.1 ‐ 29.4 ‐ 30.1 ‐ 29.9 ‐ Biguanides 11.7 ‐ 12.7 ‐ 11.7 ‐ 12.0 ‐ SU 4.9 ‐ 2.0 ‐ 4.9 ‐ 3.9 ‐ α‐GI 1.9 ‐ 2.9 ‐ 1.9 ‐ 2.3 ‐ TZDs 1.9 ‐ 2.0 ‐ 2.9 ‐ 2.3 ‐ DPP‐4 inhibitors 7.8 ‐ 6.9 ‐ 5.8 ‐ 6.8 ‐ Other BG‐lowering drugs[a](#dom13082-note-0003){ref-type="fn"} 1.9 ‐ 2.9 ‐ 2.9 ‐ 2.6 ‐ Abbreviations: α‐GI, α‐glucosidase inhibitor; BG, blood glucose; BMI, body mass index; DPP‐4, dipeptidyl peptidase‐4; eGFR, estimated glomerular filtration rate; FPG, fasting plasma glucose; OAD, oral antidiabetic drug; s.d., standard deviation; SU, sulphonylurea; TZD, thiazolidinedione. Excluding insulin. Summary is based on the full analysis set. 3.2. Primary endpoint {#dom13082-sec-0015} --------------------- Overall, the proportion of participants reporting TEAEs during the trial (primary endpoint) was higher with semaglutide (0.5 mg: 74.8%; 1.0 mg: 71.6%) than with sitagliptin (66.0%) (Table [2](#dom13082-tbl-0002){ref-type="table-wrap"}). AEs were mainly mild or moderate in severity (Table [2](#dom13082-tbl-0002){ref-type="table-wrap"}). The proportion of participants discontinuing treatment because of AEs was relatively low for semaglutide 0.5 mg and sitagliptin (2.9% and 1.9%, respectively), but higher for semaglutide 1.0 mg (10.8%) (Figure [S2](#dom13082-supitem-0001){ref-type="supplementary-material"}). Serious AEs (SAEs) were reported by 5.8%, 2.0% and 1.9% of participants treated with semaglutide 0.5 mg, 1.0 mg and sitagliptin, respectively. Events were distributed among multiple system organ classes, with no clustering (Table [S3](#dom13082-supitem-0001){ref-type="supplementary-material"}). No deaths were reported in the trial (Table [2](#dom13082-tbl-0002){ref-type="table-wrap"}). ###### Treatment‐emergent adverse events overview Semaglutide 0.5 mg Semaglutide 1.0 mg Sitagliptin 100 mg ------------------------------------------------ -------------------- -------------------- -------------------- ------- ----- -------- ----- ------- ----- -------- ----- ------- Number of participants 103 102 103 AEs: total 77 (74.8) 228 331.8 73 (71.6) 197 312.6 68 (66.0) 186 267.4 Fatal 0 ‐ ‐ ‐ 0 ‐ ‐ ‐ 0 ‐ ‐ ‐ Serious 6 (5.8) 7 10.2 2 (2.0) 2 3.2 2 (1.9) 3 4.3 Severity of AEs Severe 2 (1.9) 2 2.9 1 (1.0) 1 1.6 2 (1.9) 5 7.2 Moderate 13 (12.6) 23 33.5 9 (8.8) 12 19.0 10 (9.7) 10 14.4 Mild 73 (70.9) 203 295.4 68 (66.7) 184 292.0 67 (65.0) 171 245.8 Leading to premature treatment discontinuation 3 (2.9) 5 7.3 11 (10.8) 15 23.8 2 (1.9) 4 5.8 Abbreviations: E, number of events; N, number of participants experiencing at least one event; R, event rate per 100 years of exposure. TEAEs include events with onset from first exposure to the follow‐up visit scheduled 5 weeks (+1 week visit window) after the last trial product dose. The most frequently reported AEs were gastrointestinal (GI) events; namely constipation (14.6% and 11.8% for semaglutide 0.5 and 1.0 mg, respectively, vs 3.9% with sitagliptin), nausea (10.7% and 12.7%, vs none with sitagliptin) and diarrhoea (6.8% and 8.8%, vs 1.9% with sitagliptin; Table [S4](#dom13082-supitem-0001){ref-type="supplementary-material"}). GI events diminished over time (Figure [S3](#dom13082-supitem-0001){ref-type="supplementary-material"}). 3.3. Secondary efficacy endpoints {#dom13082-sec-0016} --------------------------------- ### 3.3.1. Glycaemic control {#dom13082-sec-0017} At week 30, mean HbA1c (baseline 8.1%) decreased by 1.9% and 2.2% with semaglutide 0.5 and 1.0 mg, respectively, vs 0.7% with sitagliptin (estimated treatment difference \[ETD\] vs sitagliptin −1.13%, 95% confidence interval \[CI\] −1.32; −0.94 and −1.44%, 95% CI −1.63; −1.24; both *P* \< .0001 \[Figure [2](#dom13082-fig-0002){ref-type="fig"}A,B and Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}\]). Statistical sensitivity analyses for change in HbA1c at week 30 supported the main analysis result (Figure [S4A](#dom13082-supitem-0001){ref-type="supplementary-material"}). The cumulative distribution of changes in HbA1c from baseline to week 30 indicated that greater proportions of participants treated with semaglutide 0.5 and 1.0 mg achieved improvements in glycaemic control compared with sitagliptin‐treated participants; nearly all participants treated with semaglutide attained improvements in glycaemic control (Figure [S5A](#dom13082-supitem-0001){ref-type="supplementary-material"}). {#dom13082-fig-0002} ###### Key outcomes by treatment group at week 30 Semaglutide 0.5 mg Semaglutide 1.0 mg Sitagliptin 100 mg -------------------------- --------------- -------------------- ---------------------- -------------------- ------------ ---------------------- --------- ------------ HbA1c (%) 8.1 (±0.9) −1.9 (0.1) −1.13 (−1.32; −0.94) \<.0001 −2.2 (0.1) −1.44 (−1.63; −1.24) \<.0001 −0.7 (0.1) Body weight (kg) 69.3 (±13.8) −2.2 (0.3) −2.22 (−3.02; −1.42) \<.0001 −3.9 (0.3) −3.88 (−4.70; −3.07) \<.0001 0.0 (0.3) BMI (kg/m^2^) 25.4 (±4.3) −0.8 (0.1) −0.84 (−1.13; −0.54) \<.0001 −1.4 (0.1) −1.44 (−1.74; −1.14) \<.0001 0.0 (0.1) Waist circumference (cm) 89.7 (±10.8) −2.4 (0.4) −2.38 (−3.40; −1.37) \<.0001 −3.8 (0.4) −3.78 (−4.82; −2.74) \<.0001 −0.1 (0.4) BP (mmHg) Systolic 129.1 (±14.8) −5.3 (1.1) −2.54 (−5.64; 0.55) .1067 −8.8 (1.2) −6.01 (−9.16; −2.85) .0002 −2.8 (1.1) Diastolic 77.6 (±10.3) −1.5 (0.7) 0.12 (−1.97; 2.21) .9072 −3.6 (0.8) −1.99 (−4.13; 0.16) .0690 −1.6 (0.8) Pulse rate, bpm 71.9 (±10.8) 4.5 (0.8) 3.41 (1.22; 5.59) .0024 6.1 (0.8) 4.94 (2.69; 7.19) \<.0001 1.1 (0.8) Semaglutide 0.5 mg Semaglutide 1.0 mg Sitagliptin 100 mg ---------------------------------------------------------------------------------------------------------- ---- -------------------------------------------- --------- ---- -------------------------------------------- --------- ---- -------------------------------------------- HbA1c targets \<7.0% 87 \(84\) \<.0001 97 \(95\) \<.0001 36 ([35](#dom13082-bib-0035){ref-type="ref"}) ≤6.5% 73 \(71\) \<.0001 89 \(87\) \<.0001 16 ([16](#dom13082-bib-0016){ref-type="ref"}) HbA1c target \<7% with no severe or BG‐confirmed symptomatic hypoglycaemia and no weight gain at week 30 74 \(72\) \<.0001 86 \(84\) \<.0001 19 ([18](#dom13082-bib-0018){ref-type="ref"}) Body weight reduction ≥5% 30 ([29](#dom13082-bib-0029){ref-type="ref"}) .0002 58 \(57\) \<.0001 7 ([7](#dom13082-bib-0007){ref-type="ref"}) ≥10% 7 ([7](#dom13082-bib-0007){ref-type="ref"}) .0141 19 ([19](#dom13082-bib-0019){ref-type="ref"}) \<.0001 0 \(0\) Abbreviations: BG, blood glucose; BMI, body mass index; BP, blood pressure; bpm, beats per minute; CI, confidence interval; ETD, estimated treatment difference; MMRM, mixed model for repeated measurements; s.d., standard deviation; s.e., standard error. Baseline is observed mean for the entire trial population. *P* values are 2‐sided testing the null hypothesis of no treatment difference. Changes from baseline (s.e.) and ETD are estimated values from MMRM. "On‐treatment without rescue medication" data from participants in the full analysis set, with the exception of pulse rate values, which are observed means using "on‐treatment" data from subjects in the safety analysis set. For treatment target endpoints, missing data are imputed from the MMRM and subsequently classified. At week 30, the ADA‐ and JDS‐defined target (HbA1c \<7.0%) was achieved by 84% and 95% of 0.5 and 1.0 mg semaglutide‐treated participants, respectively, vs 35% in the sitagliptin group (both *P* \< .0001). At week 30, the AACE target (HbA1c ≤6.5%) was achieved by 71% and 87% of 0.5 and 1.0 mg semaglutide‐treated participants, respectively, vs 16% in the sitagliptin group (both *P* \< .0001; Figure [2](#dom13082-fig-0002){ref-type="fig"}C,D and Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}). Similarly, the proportion of participants achieving HbA1c \<7.0% with no severe or BG‐confirmed symptomatic hypoglycaemia and no weight gain at week 30 was greater with semaglutide than with sitagliptin (72% and 84% vs 18%, respectively; *P* \< .0001 for both; Figure [2](#dom13082-fig-0002){ref-type="fig"}E and Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}). Reductions in mean FPG were significantly greater with semaglutide than with sitagliptin: 2.8 and 3.3 mmol/L with semaglutide 0.5 and 1.0 mg, respectively, vs 1.3 mmol/L with sitagliptin (ETD −1.47 mmol/L \[95% CI −1.78; −1.16\] and −1.99 mmol/L \[95% CI −2.30; −1.67\], respectively; both *P* \< .0001). Mean 7‐point SMPG values significantly decreased with both semaglutide doses vs sitagliptin (ETD −1.66 mmol/L \[95% CI −2.09; −1.23\] and −2.24 mmol/L \[95% CI −2.68; −1.80\], respectively; both *P* \< .0001). Similarly, there was a significant reduction in the SMPG postprandial increment with semaglutide vs sitagliptin (ETD −0.78 mmol/L \[95% CI −1.27; −0.29\] and −1.44 mmol/L \[95% CI −1.95; −0.94\], respectively; *P* = .0020 and *P* \< .0001). ### 3.3.2. Body weight {#dom13082-sec-0018} At week 30, body weight (baseline 69.3 kg) was reduced by 2.2 and 3.9 kg with semaglutide 0.5 and 1.0 mg, respectively, vs no reduction with sitagliptin (ETD −2.22 kg, 95% CI −3.02; −1.42, and −3.88 kg, 95% CI −4.70; −3.07; both *P* \< .0001 \[Figure [2](#dom13082-fig-0002){ref-type="fig"}F and G and Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}\]). Sensitivity analyses for change in body weight at week 30 supported the main analysis result (Figure [S4B](#dom13082-supitem-0001){ref-type="supplementary-material"}). The cumulative distribution of changes in body weight from baseline to week 30 indicated that greater proportions of participants treated with semaglutide 0.5 and 1.0 mg achieved reductions in weight compared with sitagliptin‐treated participants; most semaglutide‐treated participants achieved weight loss (Figure [S5B](#dom13082-supitem-0001){ref-type="supplementary-material"}). A weight loss of ≥5% was achieved by 29% and 57% of 0.5 and 1.0 mg semaglutide‐treated participants, respectively, vs 7% in the sitagliptin group (*P* \< .0002 and *P* \< .0001, respectively). A weight loss of ≥10% was achieved by 7% and 19% of 0.5 and 1.0 mg semaglutide‐treated participants, respectively, vs 0% in the sitagliptin group (*P* .0141 and *P* \< .0001, respectively) (Figure [2](#dom13082-fig-0002){ref-type="fig"}H and I, and Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}). ### 3.3.3. Other efficacy endpoints {#dom13082-sec-0019} Treatment with semaglutide 0.5 and 1.0 mg reduced BMI and waist circumference, compared with sitagliptin (*P* \< .0001 for both). In addition, decreases in BP were generally greater with semaglutide 1.0 mg than with sitagliptin. The difference in systolic BP with semaglutide 1.0 mg, vs sitagliptin, was significant (ETD −6.01 mmHg, 95% CI −9.16; −2.85; *P* = .0002); there was no significant difference in diastolic BP (Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}). Total cholesterol and LDL cholesterol were significantly reduced with both semaglutide doses, whereas VLDL cholesterol and triglycerides were significantly reduced with semaglutide 1.0 mg vs sitagliptin. There was no significant difference in change in HDL cholesterol or FFA levels between semaglutide and sitagliptin (Figure [S6](#dom13082-supitem-0001){ref-type="supplementary-material"}). 3.4. Supportive secondary safety endpoints {#dom13082-sec-0020} ------------------------------------------ No severe episodes of hypoglycaemia were reported in any treatment group. One hypoglycaemic episode, classified as "severe or BG‐confirmed symptomatic," was reported in the semaglutide 1.0 mg group (Table [S5](#dom13082-supitem-0001){ref-type="supplementary-material"}). This event was symptomatic and plasma glucose level was 2.6 mmol/L. No EAC‐confirmed pancreatitis events were reported. Mean levels of pancreatic enzymes (lipase and amylase) increased for both semaglutide doses vs sitagliptin. One participant treated with semaglutide 1.0 mg experienced cholelithiasis (Table [S5](#dom13082-supitem-0001){ref-type="supplementary-material"}), which did not lead to premature treatment discontinuation. There were 2 EAC‐confirmed cardiovascular events in the semaglutide 0.5 mg group (Table [S5](#dom13082-supitem-0001){ref-type="supplementary-material"}; 1 event of silent myocardial infarction; 1 of percutaneous revascularization). Pulse (baseline 71.9 bpm) increased in all 3 groups (ETD 3.41 bpm \[95% CI 1.22; 5.59\] and 4.94 bpm \[95% CI 2.69; 7.19\] for semaglutide 0.5 and 1.0 mg vs sitagliptin, respectively; *P* = .0024 and *P* \< .0001 \[Table [3](#dom13082-tbl-0003){ref-type="table-wrap"}\]). Overall, 2 EAC‐confirmed malignant neoplasms were reported; 1 with semaglutide 1.0 mg (malignant low‐grade bladder cancer tumour) and 1 with sitagliptin (malignant pancreatic carcinoma stage IV). The participant in the semaglutide 1.0 mg arm completed treatment; the sitagliptin‐treated participant discontinued treatment. No malignant neoplasms were reported with semaglutide 0.5 mg (Table [S5](#dom13082-supitem-0001){ref-type="supplementary-material"}). No EAC‐confirmed thyroid neoplasms or medullary thyroid carcinomas were reported. Calcitonin levels were similar between groups with no apparent change during the trial. Nervous system disorders were reported by 12 participants (11.7%) in the semaglutide 0.5 mg group, none in the semaglutide 1.0 mg group and 5 (4.9%) in the sitagliptin group (Table [S5](#dom13082-supitem-0001){ref-type="supplementary-material"}), and were composed of single unrelated AEs. One semaglutide 1.0 mg‐treated participant developed anti‐semaglutide antibodies; however, there was no cross‐reaction with endogenous GLP‐1. At follow‐up, the participant tested antibody‐negative and therefore no in vitro neutralizing effect was assessed. There were 4 reported events of diabetic retinopathy in the semaglutide 0.5 mg group, 2 events in the semaglutide 1.0 mg group, and 4 events in the sitagliptin group. There were no clinically relevant changes in other safety laboratory assessments, physical examinations or electrocardiograms. 4. DISCUSSION {#dom13082-sec-0021} ============= The results of the present trial showed that, in Japanese participants with T2D, more participants receiving semaglutide reported TEAEs than with sitagliptin. This was driven mainly by GI AEs, a well‐known side effect of GLP‐1 RAs. The discontinuation rate was low and similar between semaglutide 0.5 mg and sitagliptin, but was higher with semaglutide 1.0 mg, owing to a larger proportion of participants experiencing GI AEs. The slightly higher rate of constipation vs nausea and diarrhoea than reported in global trials may have been influenced by other factors, such as differences in diet, cultural differences in how constipation is defined or reported, or AE‐reporting bias among participants. The frequency of GI AEs diminished over time, as observed with other GLP‐1 RAs and other studies with semaglutide.[10](#dom13082-bib-0010){ref-type="ref"}, [11](#dom13082-bib-0011){ref-type="ref"}, [12](#dom13082-bib-0012){ref-type="ref"}, [13](#dom13082-bib-0013){ref-type="ref"}, [14](#dom13082-bib-0014){ref-type="ref"}, [15](#dom13082-bib-0015){ref-type="ref"}, [24](#dom13082-bib-0024){ref-type="ref"}, [25](#dom13082-bib-0025){ref-type="ref"} Dose escalation of GLP‐1 RAs has also previously been shown to partially ameliorate GI AEs, as reflected in the design of this trial.[26](#dom13082-bib-0026){ref-type="ref"} In general, the AE profile of semaglutide was similar to that of other GLP‐1 RAs[5](#dom13082-bib-0005){ref-type="ref"}; for example, in addition to GI AEs, a commonly reported TEAE was increased levels of pancreatic enzymes for both semaglutide doses vs sitagliptin (although no pancreatitis events were reported). Pancreatitis cases have previously been reported in exenatide‐treated patients and in other clinical development programmes for GLP‐1 RAs and other incretin‐based drugs,[27](#dom13082-bib-0027){ref-type="ref"}, [28](#dom13082-bib-0028){ref-type="ref"} although retrospective analyses suggest there is no increased risk of acute cases.[29](#dom13082-bib-0029){ref-type="ref"}, [30](#dom13082-bib-0030){ref-type="ref"}, [31](#dom13082-bib-0031){ref-type="ref"}, [32](#dom13082-bib-0032){ref-type="ref"} In addition, reduced appetite was frequently reported in semaglutide‐treated participants. This is a class effect, as GLP‐1 receptors are expressed in the hypothalamus, the region of the brain that regulates satiety and appetite. Effectively, this is thought to be the mechanism that enables GLP‐1 RAs to lower energy intake, thereby reducing body weight,[33](#dom13082-bib-0033){ref-type="ref"}, [34](#dom13082-bib-0034){ref-type="ref"} although to a varying degree. The increase in pulse rate reported with semaglutide treatment has also been previously reported in, for example, liraglutide‐ and exenatide‐treated participants,[24](#dom13082-bib-0024){ref-type="ref"}, [35](#dom13082-bib-0035){ref-type="ref"} although the potential mechanisms behind this haemodynamic effect are unclear.[36](#dom13082-bib-0036){ref-type="ref"} Compared with sitagliptin 100 mg monotherapy, both semaglutide 0.5 and 1.0 mg monotherapy significantly improved glycaemic control (change in HbA1c: −1.9%, −2.2% vs −0.7%, respectively) and reduced body weight (−2.2 kg, −3.9 kg vs 0.0 kg, respectively). These improvements were sustained over 30 weeks of treatment, and were supported by the sensitivity analyses. Furthermore, significantly greater proportions of participants receiving semaglutide achieved the ADA and JDS target of HbA1c \<7.0% vs sitagliptin (84% and 95% of 0.5 and 1.0 mg semaglutide‐treated participants, respectively, vs 35% in the sitagliptin group). Importantly, a higher proportion of participants treated with semaglutide than with sitagliptin achieved this target with no severe or BG‐confirmed symptomatic hypoglycaemia and no weight gain at week 30 (72% and 84% vs 18%, respectively). Semaglutide's efficacy relating to body weight was also evident compared with sitagliptin, with greater proportions of participants achieving body weight loss of ≥5% (29% and 57% vs 7%, respectively) and ≥10% (7% and 19% vs 0%, respectively). These findings are consistent with the global SUSTAIN trials, in which semaglutide was associated with clinically meaningful and superior reductions in HbA1c and body weight vs placebo (SUSTAIN 1 and SUSTAIN 5 \[add‐on to insulin\]),[10](#dom13082-bib-0010){ref-type="ref"}, [14](#dom13082-bib-0014){ref-type="ref"} sitagliptin (SUSTAIN 2)[11](#dom13082-bib-0011){ref-type="ref"} -- all multinational trials that included Japan, exenatide ER (SUSTAIN 3),[12](#dom13082-bib-0012){ref-type="ref"} and insulin glargine (SUSTAIN 4).[13](#dom13082-bib-0013){ref-type="ref"} The findings also align with trials of liraglutide in participants with T2D, where significant reductions in HbA1c and body weight were observed, either compared with placebo or a thiazolidinedione, all when added to a sulphonylurea[24](#dom13082-bib-0024){ref-type="ref"} ; or compared with placebo or a sulphonylurea, all on a background of metformin.[25](#dom13082-bib-0025){ref-type="ref"} In addition, similar results have been observed with other GLP‐1 RAs, such as albiglutide[37](#dom13082-bib-0037){ref-type="ref"} and dulaglutide,[38](#dom13082-bib-0038){ref-type="ref"} either compared with placebo or active comparators such as sitagliptin or insulin glargine, as monotherapy or add‐on therapy to other antidiabetic agents. In the present trial, semaglutide also resulted in marked improvements in cardiometabolic risk compared with sitagliptin, and included reductions in BMI, waist circumference and BP, and improvements in most lipid level profiles. In SUSTAIN 6, semaglutide was associated with a significant 26% reduction in cardiovascular risk (primary composite outcome of death from cardiovascular causes, non‐fatal myocardial infarction or non‐fatal stroke) compared with placebo in a standard‐of‐care setting.[15](#dom13082-bib-0015){ref-type="ref"} Baseline characteristics, although similar across the 3 treatment groups, differed slightly from the global SUSTAIN trials. In this trial population, baseline body weight and BMI were markedly lower than in participants from various ethnic groups in the global trials, reflecting an overall leaner population. Furthermore, more marked reductions in HbA1c were reported in this trial than in the SUSTAIN 1 to 5 trials.[10](#dom13082-bib-0010){ref-type="ref"}, [11](#dom13082-bib-0011){ref-type="ref"}, [12](#dom13082-bib-0012){ref-type="ref"}, [13](#dom13082-bib-0013){ref-type="ref"}, [14](#dom13082-bib-0014){ref-type="ref"} The reasons for this difference between the present trial and SUSTAIN 1 to 5 are unclear; however, these results are in line with a recent meta‐analysis in which GLP‐1 analogues were associated with greater HbA1c reductions and a higher proportion of participants achieving HbA1c target ≤7% in Asian (including Japanese) vs non‐Asian populations.[39](#dom13082-bib-0039){ref-type="ref"} Participants in this trial had a mean baseline diabetes duration of 8.0 years, similar to the global SUSTAIN 2 trial (6.6 years), but early initiation of treatment may be beneficial in Asian people to provide sustained glycaemic control and avoid long‐term diabetes complications. Although body weight was reduced with semaglutide in this trial, greater reductions were reported in the SUSTAIN 1 to 5 trials (3.5--4.3 kg with semaglutide 0.5 mg; 4.5--6.4 kg with semaglutide 1.0 mg).[10](#dom13082-bib-0010){ref-type="ref"}, [11](#dom13082-bib-0011){ref-type="ref"}, [12](#dom13082-bib-0012){ref-type="ref"}, [13](#dom13082-bib-0013){ref-type="ref"}, [14](#dom13082-bib-0014){ref-type="ref"} Both baseline body weight and BMI were lower in this trial population, by \~20 kg and 7 kg/m^2^, respectively; thus, the body weight reductions of 3.3% and 5.7% with semaglutide 0.5 and 1.0 mg, respectively, should be considered in this context. It might be expected that Japanese participants would lose less body mass overall than a heavier and/or obese cohort, as \~80% of weight lost with semaglutide treatment is fat mass.[40](#dom13082-bib-0040){ref-type="ref"} This, together with the influence of the Japanese diet, may have affected the overall change in body weight in this trial. These efficacy results might also be considered in the context that, in East Asia, T2D is characterized by significant β‐cell dysfunction with less adiposity and less insulin resistance than in Western populations. Consequently, incretin‐based drugs may show more efficacy in Asian populations, mostly because of amelioration of β‐cell dysfunction.[41](#dom13082-bib-0041){ref-type="ref"}, [42](#dom13082-bib-0042){ref-type="ref"} The open‐label design of this trial is a limitation, increasing the risk of bias. The relatively short trial duration is also a limitation, as the full potential of semaglutide with regard to durability, efficacy and safety may not have been observed over 30 weeks. In addition, approximately three‐quarters of the total trial population was male and, as such, may not accurately reflect the T2D population in Japan. Differential results between men and women with T2D, however, would not be expected. Furthermore, the authors acknowledge that GLP‐1 RAs are not usually considered for monotherapy; neither is sitagliptin the most frequently used monotherapeutic oral agent in Japan. Thus, in clinical practice, both semaglutide and sitagliptin would probably be given in combination with metformin, and the applicability of the results should be interpreted in this context, whereby the treatment regimens described may not reflect mainstream practice in the T2D population in Japan. Nevertheless, the SUSTAIN programme as a whole covered a broad spectrum of patients with T2D[10](#dom13082-bib-0010){ref-type="ref"}, [11](#dom13082-bib-0011){ref-type="ref"}, [12](#dom13082-bib-0012){ref-type="ref"}, [13](#dom13082-bib-0013){ref-type="ref"}, [14](#dom13082-bib-0014){ref-type="ref"}, [15](#dom13082-bib-0015){ref-type="ref"}, [16](#dom13082-bib-0016){ref-type="ref"}, [17](#dom13082-bib-0017){ref-type="ref"} ; for example, SUSTAIN 2, which included patients from Japan, investigated semaglutide vs sitagliptin as an add‐on to metformin.[11](#dom13082-bib-0011){ref-type="ref"} It should be noted that, in line with regulatory requirements, this trial was designed to assess these treatments as monotherapy, in order to investigate the safety and efficacy of each in isolation. Finally, the collection of body composition data would have been useful, to clarify whether the observed body weight loss with semaglutide was caused by loss of fat or lean body mass. A recently published study involving semaglutide treatment in obese participants has, however, demonstrated that semaglutide is associated with a 3‐fold greater loss of fat over lean body mass compared with placebo.[40](#dom13082-bib-0040){ref-type="ref"} Despite these limitations, the significant reductions in HbA1c and body weight observed in this trial indicate the potential for semaglutide as a treatment option in Japanese participants with T2D. This is important given that many other treatments are either weight‐neutral or associated with weight gain.[43](#dom13082-bib-0043){ref-type="ref"}, [44](#dom13082-bib-0044){ref-type="ref"}, [45](#dom13082-bib-0045){ref-type="ref"} In conclusion, in Japanese participants with T2D, more TEAEs were reported with semaglutide (0.5 and 1.0 mg) than with sitagliptin, but semaglutide was well tolerated, with a safety profile similar to that of other GLP‐1 RAs. Semaglutide had a greater effect on glycaemic control, body weight reduction and other efficacy variables vs sitagliptin 100 mg monotherapy. ORCID {#dom13082-sec-0024a} ===== *Yutaka Seino* [http://orcid.org/0000‐0002‐1099‐7989](http://orcid.org/0000-0002-1099-7989) Supporting information ====================== ###### **Figure S1.** Trial design. **Figure S2.** Cumulative incidence of premature discontinuation of treatment due to adverse events over time. **Figure S3.** Time course of constipation (A), nausea (B) and diarrhoea (C) from baseline to 35 weeks. **Figure S4.** Statistical sensitivity analyses for estimated treatment differences in HbA~1c~ (A) and body weight (B) at week 30. **Figure S5.** Cumulative distribution of change from baseline in HbA~1c~ (A) and body weight (B) to week 30. **Figure S6.** Estimated treatment ratios for lipid levels from baseline to 30 weeks. **Table S1.** Key inclusion and exclusion criteria. **Table S2.** Events types predefined for Event Adjudication Committee review. **Table S3.** Treatment‐emergent serious adverse events (by system organ class). **Table S4.** Most frequent treatment‐emergent adverse events (≥5% of subjects in any group). **Table S5.** Additional treatment‐emergent adverse events of interest. ###### Click here for additional data file. We thank all the participants, investigators and trial‐site staff who were involved in the conduct of the trial. We also thank Sayeh Tadayon and Hrvoje Vrazic (Novo Nordisk) for their review and input to the manuscript, and Sola Neunie, MSc, CMPP (AXON Communications) for medical writing and editorial assistance, who received compensation from Novo Nordisk. Conflict of interest {#dom13082-sec-0023} ==================== Y.S. has received honoraria for consulting and/or speakers bureau from Kao; Kyowa Hakko Kirin; Taisho Pharmaceutical; Becton, Dickinson and Company; Novo Nordisk; MSD; Intarcia Therapeutics; Johnson & Johnson; GlaxoSmithKline; Takeda Pharmaceutical Company; Sanofi; Taisho Toyama Pharmaceutical; Eli Lilly & Company; Mitsubishi Tanabe Pharma; Ono Pharmaceutical; Kowa; Astellas Pharma; Boehringer Ingelheim; AstraZeneca; Sumitomo Dainippon Pharma; Daiichi Sankyo; Terumo; Arkray; and clinical commissioned/joint research grants from Arkray Marketing; Kowa; Hayashibara; Nippon Boehringer Ingelheim; Eli Lilly & Company; Terumo; Taisho Toyama Pharmaceutical; MSD K.K.; Ono Pharmaceutical; Novo Nordisk; Sumitomo Dainippon Pharma; Arklay. Y. T. has received honoraria for speakers bureau from Astellas Pharma; AstraZeneca K.K.; Bayer Yakuhin; Daiichi Sankyo; Sumitomo Dainippon Pharma; Eli Lilly Japan K.K.; Kowa; MSD K.K.; Mitsubishi Tanabe Pharma; Nippon Boehringer Ingelheim; Novo Nordisk; Ono Pharmaceutical; Sanwa Kagaku Kenkyusho; Sanofi K.K.; Shionogi & Company; Taisho Toyama Pharmaceutical; Takeda Pharmaceutical Company; and grants from Astellas Pharma; AstraZeneca K.K.; Bayer Yakuhin; Daiichi Sankyo; Sumitomo Dainippon Pharma; Eli Lilly Japan K.K.; Kowa; MSD K.K.; Mitsubishi Tanabe Pharma; Nippon Boehringer Ingelheim; Novo Nordisk; Ono Pharmaceutical; Pfizer Japan; Sanwa Kagaku Kenkyusho; Sanofi K.K., Shionogi & Company; Takeda Pharmaceutical Company. T. O. is a lecturer and clinical study investigator for Novo Nordisk. D. Y. has received consulting and/or speaker fees from MSD K.K.; Novo Nordisk; Takeda Pharmaceutical Company; Taisho Toyama Pharmaceutical; and clinical commissioned/joint research grants from Nippon Boehringer Ingelheim; Eli Lilly & Company; Taisho Toyama Pharmaceutical; MSD K.K.; Ono Pharmaceutical; Novo Nordisk; Arklay; Takeda Pharmaceutical Company. NA has no conflicts of interest. T.N. is an employee of Novo Nordisk and holds shares in the company. J. Z. is an employee of Novo Nordisk and holds shares in the company. S. K. has received honoraria for speakers bureau from Astellas Pharma; AstraZeneca K.K.; Daiichi Sankyo; Sumitomo Dainippon Pharma; Eli Lilly Japan K.K.; Kissei Pharmaceutical; Mitsubishi Tanabe Pharma; Novo Nordisk; Novartis Pharma K.K.; Ono Pharmaceutical; Sanofi K.K.; Taisho Toyama Pharmaceutical; Takeda Pharmaceutical Company; and grants from WebMD Global LLC. Author contributions {#dom13082-sec-0024} ==================== **Y.S.:** study design, study conduct/data collection, writing the manuscript; **Y.T.:** study conduct/data collection, data analysis, writing the manuscript; **T.O.:** data collection; **D.Y.:** study conduct/data collection, data analysis, writing the manuscript; **N.A.:** study conduct/data collection, reviewing the manuscript; **T.N.:** study design, data analysis, writing the manuscript; **J.Z.:** study design, study conduct/data collection, data analysis, writing the manuscript; **S.K.:** study conduct/data collection, data analysis, writing the manuscript.

Background {#s1} ========== Diabetes is a multi-factorial disorder of the pancreas, in which the pancreas fails to perform its function to produce insulin hormone properly in the body. It involves multiple disorders like hyperglycemia, glycosuria and abnormal metabolism of lipids, carbohydrates and proteins \[[@R01],[@R02]\].This affects the human body at physiological, physical and social level. It has been known as the 3rd leading cause of death in humans along with other diseases such as cancer, cerebro-vascular and heart Hypoglycemic medication is helps to lower the blood sugar level in body or treat the other severe symptoms and complications of diabetes mellitus \[[@R02]\]. The side effects of these medications include extreme hypoglycemia, liver cell injury, lactic acidosis, digestive discomfort, permanent neurological deficit, headache, dizziness and even death \[[@R03],[@R04]\]. The basic challenge in curing diabetes is to maintain blood glucose level close to normal levels \[[@R05]-[@R10]\]. These therapies are used as mono therapy or in combination for optimal control of glycemia \[[@R11]-[@R14]\].As mentioned before that these drugs are normally expensive and come with side effects. These drugs have their limitations due to low pharmacokinetic properties, secondary failure rates and relative bad effects \[[@R15]-[@R21]\]. Thus, there is a need for efficient class of compounds to reduce the side effects. Molecular docking is a competent tool for novel micro molecule drugs discovery for targeting protein. This study has been carried out in order to identify effective, selective and efficient antidiabetic Lead compound and its analogues. Chalcone is a class of open-chain flavonoids that is not only biosynthesized by plants but also can be prepared synthetically. The simplest chalcone can be prepared by an aldol condensation between a benzaldehyde and an acetophenone in the presence of base \[[@R22]-[@R24]\]. Hydrazones of chalcones have shown a wide variety of pharmacological effects, including anti-inflammatory and anticancer activities \[[@R25]-[@R29]\].Despite the comprehensive biological studies on chalcones, reports on their anti-diabetic activity are limited \[[@R30]\]. Significant advances have been made in the past few years in the isolation and preparation of several hydrazones of chalcones derivatives. Methodology {#s2} =========== Chemistry: {#s2a} ---------- Thin layer chromatography (TLC) was used to examine the progress of the reaction. Open glass vessels were used to make a decision for the dissolving on outstanding softening mechanical assembly and were uncorrected. H1 and 13C atomic enticing reverberation (H1 proton magnetic resonance and 13CNMR) spectra were recorded on Bruker Avance II four hundred proton magnetic resonance spectroscope (400 MHz) at 298K, in correct deuterated dissoluble. Concoction move were accounted for as δ (ppm) with relation to tetra methyl silane (TMS) within allowable limit. Infrared spectra (IR) were recorded as KBr pellet on Shimadzu FT-IR spectroscope. Preparation of AC-CdO-TiO2 nonocomposite material by precipitation method: {#s2b} -------------------------------------------------------------------------- AC-CdO-TiO2 nanocomposite material was synthesized by precipitation method. Initially cadmium acetate dihydrate (0.4 M) were dissolved in anhydrous ethanol solution beaker A. 0.4m citric acid and tetra isopropyl orthotitanate were in ethanol is taken as another solution beaker B and Activated carbon (AC) were dissolved in anhydrous ethanol solution beaker C, the solution A and solution C is added to Solution beaker B and stirred well. Then to thise 2 drops NH4OH is added at room temperature under vigorous stirring until the precipitate was formed. The obtained precipitate was washed with water and ethanol. Then the precipitate was collected and dried in oven at 100oC for 12 hrs. The resulting powder was finally calcinated at 5000 C at 4 hrs. General procedure for the Synthesis of (E)-1-(4-(difluoromethoxy)-3-hydroxyphenyl)-3-phenylprop-2-en-1-one(3): {#s2c} -------------------------------------------------------------------------------------------------------------- 4-(difluoromethoxy)-3-hydroxybenzaldehyde 1 (0.02 mol) and 1- (benzo\[d\]\[1,3\]dioxol-5-yl)ethanone 2 (0.02 mol) were dissolved in 30 ml of alcohol. To this reaction mixture 40% NaOH (10 ml) and AC-CdO-TiO2 nanoparticles catalyst (0.003 g), in ethanol (5 mL) were added. TLC followed the progress of the reaction. After completion of the reaction, the mixture was filtered to remove the catalyst and the filtrate was taken in ether, washed with water and dried over anhydrous sodium sulfate. Removal of solvent gave the crude product which was recrystallized from methanol to obtain the pure compounds.m.p:960C; M.F: C17H12F2O5;M.W:334. General procedure for the Synthesis of 5-((1Z, 3Z)-3- (benzo\[d\]\[1,3\]dioxol-5-yl)-4-(substituted pyridin-2-yl)buta-1,3- dien-1-yl)-2-(difluoromethoxy)phenol (4a-e): {#s2d} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- (E)-1-(4-(difluoromethoxy)-3-hydroxyphenyl)-3-phenylprop-2-en-1- one 3 (0.01mol) and substituted aniline (0.01 mol) was dissolved in ethanol (20 ml). To this mixture AC-CdO-TiO2 nano particles was added and it was refluxed for 3 hrs. On cooling and dilution with ice-cold water, a solid mass separated out. It was re-crystallized from ethanol. Docking studies: {#s2e} ---------------- X-ray crystal structures of pig pancreatic alpha-amylase (PDB Id: 3L2M) were retrieved from the Protein Data Bank \[[@R31]\]. To put together the receptor for docking studies, co-crystallized ligand and water molecules were eliminated. At the same time polar hydrogen atoms and Kollman-united costs have been protected by the DNA Gyrase receptor. The essential pdb and pdbqt documents of ligands and Pig pancreatic alpha-amylase receptor were prepared for the AutoDock 4.2 software \[[@R32]\]. The usual docking protocol was applied in the AutoDock Vina in PyRx 0.8 software \[[@R33]\]. The docking results were analyzed using Discovery Studio 4.0 (Accelrys, Inc. San Diego, CA 92121, U.S.). Inhibition assay for α-Amylase activit: {#s2f} --------------------------------------- A stock solution of 10 mg/10 mL concentration was prepared using DMSO solvent. Activity of amylase was assayed with different concentrations (50, 100, 200 µg/mL) of sample with control and reagent solution without test sample was used as control. Starch solution (1% w/v) or (0.5% w/v) was prepared by stirring and boiling 0.5 g of soluble potato starch in 50 mL of deionized water for 15 minutes. The enzyme solution (1 unit/mL) was prepared by mixing 100 mg in 100 mL of 20 mM sodium phosphate buffer (pH 6.9). The color reagent was a solution containing 96 mM 3,5- dinitrosalicylic acid (DNSA) (20 mL), 5.31 M sodium potassium tartrate in 2 M NaOH (8 mL) and deionized water (12 mL). Acarbose was used as a standard at the concentration of 1mg/mL. 100 µl of test solution and 100 µL of enzyme solution were mixed in viols and incubated at 25°C for 30 min. To this mixture 100 µL of color reagent was added and the mixture was heated on water bath at 85°C for 15 min. Further, the reaction mixture was removed from water bath, cooled and absorbance value determined at 595 nm. Individual blanks were prepared for correcting the background absorbance. Control experiment was conducted in the same manner by replacing the drug sample with 1 mL DMSO. Inhibition percentage of α- amylase was calculated by the formula \[[@R34]\]. Enzyme activity was calculated and percentage of inhibition is ((Control-Test)/100) x 100. Results and Discussion: {#s3} ======================= [Table 1](#T1){ref-type="table"} show all the physical data like color, molecular formula, molecular weight, solubility, melting point, of synthesized compounds. The IR frequencies of compounds 4a-e is shown in [Table 2](#T2){ref-type="table"} in which the C=N stretching frequency appear at 1586-1667 cm-1. Aromatic (CH) stretching frequencies appear at 3084-3093cm-^1^ and stretching frequency observed at 1625-1666cm^-1^ C=O group present in the derivatives. The 1H NMR chemical shift values of compound (4a-e) given in [Table 3](#T3){ref-type="table"}. The singlet observed in the range 6.30-6.60ppm is due --CH~2~ methylene proton of 3\',4\'- methylenedioxy acetophenone moiety proton. The singlet observed at 7.41-7.49ppm is due --CH proton of --CHF~2~ moeity. The signals appearing 7.14-8.38ppm are obviously due to aromatic protons. The five chalconeimine derivatives (4a-e) shown in [Figure 1a](#F1){ref-type="fig"} were taken for docking studies. These compounds are synthesized and their structures have been determined by IR,^1^H and ^13^CNMR spectroscopy. In vitro α-amylase inhibition: {#s3a} ============================== All the synthesized compounds (4a-e) and standard drug were explored for their in vitro α- amylase inhibition studies at different concentrations (50, 100, 200µg/mL) as shown in the [Table 4](#T4){ref-type="table"}. All the compounds showed good % inhibition of α-amylase when compared with standard drug acarbose. Compound 4b and 4d were found to be more potent among all the synthesized compounds when explored at the concentration of 50µg/mL. Compound 4d shows 76.58% inhibition followed by 4b with 77.18% inhibition. There was a significant rise in % inhibition when concentration has been changed to 100µg/ml from 50µg/mL. Among all, 4b shows 81.35% inhibition followed by 4a which showed 85.23% inhibition at 100µg/ml. Inspired by the results obtained at 100µg/mL concentration, all the synthesized compounds were further screened for there in vitro α-amylase inhibition at 200µg/mL. All compounds exhibited a linear rise in % inhibition. Docking studies: {#s3b} ================ Interactions between inhibitors and active site of the target protein can be explored using molecular docking studies. The above results showed that all the synthesized molecules were stronger inhibitors of alpha-amylase as compared to acarbose. Therefore, for ascertaining the binding conformation and interactions responsible for the activity, docking simulation of compound 4a and 4d was performed against active site of pig pancreatic alpha-amylase (PDB ID: 3L2M). Ligands taken for the docking studies are shown in [Figure 1a](#F1){ref-type="fig"}. Pig pancreatic alpha-amylase protein is considered as target protein for this study. Its structure was taken from RCSB Protein Data Bank (PDB) with PDB ID: 3L2M as shown in [Figure 2](#F2){ref-type="fig"}. Target protein has their active sites where the compound shows maximum number of interaction with protein. The complete dataset was docked and found to bind at the same active site position. Amino acids are intimately involved in the binding ligand to protein and form a complex. The residue that is significant for binding interaction and thus comprising the binding pocket of target protein are shown in [Table 4](#T4){ref-type="table"}. Docking studies reveled that these amino acids present in the target proteins pocket involves in the binding interaction with the selected compounds. These complex structures reveal essential interactions between the inhibitor and the protein and these interactions are taken as the reference for the hydrazone derivative (4a-e). The co-crystallized ligand are forms hydrogen bond interaction with the residues GLY 309, GLN 302, ARG 346, ASP316, ARG 267) ([Figure 5](#F5){ref-type="fig"}) which are present within the ATP binding pocket. The ligand is also further stabilized by a number of hydrophobic contacts with the residues. The five hydrazone derivatives (4a-e) shown in [Figure 1a](#F1){ref-type="fig"} were taken for docking studies. These compounds are synthesized and their structures have been determined by IR, 1H and13CNMR spectroscopy. The docking studies clearly reveal that some of these compounds bind efficiently to the enzymes of pig pancreatic alphaamylase. Binding score of autodock 4.2 varies between -7.8 to -8.9 for compounds 3a-g tested for pig pancreatic alpha-amylase ([Table 5](#T5){ref-type="table"}) Out of the five hydrazone derivatives analyzed, compound 4b and 4d forms the best interaction with pig pancreatic alphaamylase. The compound 4a and 4d has the highest binding score of -8.9 and - 8.7. The fluorine, oxgen atom on hydrazone compound forms hydrogen bond with the hydrogen atom of ALA 198, ARG 195, and HIS 299 of pig pancreatic alpha-amylase ([Figure 3](#F3){ref-type="fig"} and [Table 6](#T6){ref-type="table"}). Compound 4d having a binding score of -8.9 makes hydrogen bonds with the active site residue ASP 300, GLU 233, LYS200 and ILE 235 of enzyme ([Figure 4](#F4){ref-type="fig"}).Re-docking of the inhibitor from the co-crystallized complex structure ([Figure 5](#F5){ref-type="fig"}) of pig pancreatic alphaamylase resulted in a binding score of -7.8, which is comparable to the scores found for compound 4b and 4d ([Table 5](#T5){ref-type="table"}). The re-docked conformation of co-crystallized ligand ([Figure 2](#F2){ref-type="fig"}) resembles the conformation of the hydrazone derivative (compound 4b and 4d respectively). Conclusion {#s4} ========== We describe the synthesis and evaluation of five hydrazone derivatives as α-amylase inhibitors. The structures of all synthesized compounds were confirmed by elemental and spectroscopic analysis (IR, 1H and 13C-NMR). The biological potential of synthesized compounds was investigated through in vitro α-amylase inhibition activity. The results showed that some of the synthesized compounds exhibited significant inhibitory activities. The compound 5-((1Z,3Z)-3-(benzo\[d\]\[1,3\]dioxol-5-yl)-3- ((4-chloropyridin-2-yl)imino)prop-1-en-1-yl)-2-(difluoromethoxy) phenol (4b) in 100 µg/mL concentration showed remarkable inhibition of 81.35%. Docking studies of compound 4a-e were performed against active site of pig pancreatic alpha amylase (PDB ID: 3L2M). It has been revealed from docking studies that the bonding interactions found between 4b and 4d Rwith pig pancreatic α-amylase are similar to those responsible for α-amylase inhibition by acarbose. **Edited by P Kangueane** **Citation:**Balu *et al.* Bioinformation 15(7):523-529 (2019) ###### Physical data of various synthesized compounds Compound Color Mol. Formula Mol. weight Solubility Melting point (°C) ---------- ------------- ---------------------------- ------------- ------------ -------------------- 4a Yellow C~22~H~15~C~l~F~2~N~2~O~4~ 444 Ethanol 157 4b Yellow C~22~H~15~C~l~F~2~N~2~O~4~ 444 Ethanol 133 4c Yellow C~22~H~15~C~l~F~2~N~2~O~4~ 444 Ethanol 148 4d Pale Yellow C~23~H~18~F~2~N~2~O~4~ 424 Ethanol 128 4e Pale Yellow C~23~H~18~F~2~N~2~O~4~ 424 Ethanol 118 ###### Data from IR spectra of chalconeimine derivatives (4a-e) Compounds FREQUENCY cm-1 ----------- ---------------- ------ ------ ------ ------ 4a 1666 1597 2966 1452 3089 4b 1645 1586 2924 1425 3084 4c 1645 1589 2924 1448 3084 4d 1625 1586 2926 1452 3093 4e \- 1667 2924 1450 3088 ###### Data from 1H NMR spectra of hyrazone derivatives (4a-e) Compounds 0 CHF~2~ Aromatic protons ----------- ------------------- ------------------- ---------------------------- 4a 6.30 (2H,singlet) 7.48 (1H,singlet) 7.43-8.36 (11H, multiplet) 4b 6.60 (2H,singlet) 7.47(1H,singlet) 7.47-7.95 (11H, multiplet) 4c 6.55 (2H,singlet) 7.49 (1H,singlet) 7.43-8.38 (11H, multiplet) 4d 6.58 (2H,singlet) 7.46 (1H,singlet) 7.28-7.96 (11H, multiplet) 4e 6.55 (2H,singlet) 7.41 (1H,singlet) 7.14-8.38 (11H, multiplet) ###### alpha-amylase inhibition activity of compounds 4a-e Compound Concentration (µg/mL) \% Inhibition ---------- ----------------------- --------------- 4a 50 70.84 100 85.23 200 86.84 4b 50 77.18 100 81.35 200 83.64 4c 50 54.82 100 68.58 200 73.34 4d 50 76.58 100 75.03 200 77.84 4e 50 50.12 100 69.87 200 71.93 Acarbose 50 56.69 100 63.85 200 69.78 ###### Binding energy of docked compounds (4a-e) Compound 4a 4b 4c 4d 4e Co-ligand ---------------- ------ ------ ------ ------ ------ ----------- Binding energy -8.9 -8.9 -8.3 -8.7 -8.5 -7.8 ###### Binding interactions of docked compounds Compound Type of interaction Between Distance Type of interaction Between Distance Type of interaction Between Distance ----------- --------------------- ------------------ ---------- --------------------- ----------------- ---------- --------------------- --------- ---------- 4a Hydrogen bon NH-0 (GLC 701) 3.04 Halogen F-0 (GLU 233) 2.99 pi-pi interaction TYR 62 4.57 H-0 (GLU 233) 2.53 4b Hydrogen bon NH-O (ALA 198) 2.72 Halogen F-OD1(ASP 197) 3.65 Alkyl Interaction VAL 163 5.23 o-NH (ARG 195) 2.86 Alkyl Interaction LEU 165 4.09 NH-F (HIS 299) 2.55 F-NE2(HIS299) 3.31 Alkyl Interaction ALA 198 5.35 4c Accptor -aceptor o-o2 (GLC 701) 2.83 Halogen F-OD2 (ASP 356) 3.15 Alkyl Interaction ILE 235 3.99 Charge -change N-OD2 (ASP 300) 5.53 Alkyl Interaction LYS 200 3.82 N-OD1 (ASP 191) 5.41 Pi-Alkyl HIS 201 4.82 N-OE1 (GLU233) 5.06 4d Hydrogen bon NH-oD2 (ASP 300) 2.44 Halogen F-CD(GLU 233) 3.62 Pi-Alkyl TYR 62 4.29 NH-OE1 (GLU 233) 2.28 F-O(GLU 233) 3.44 Pi-Alkyl TRP 58 4.57 NH-F (LYS200) 2.54 F-C (ILE 235) 3.52 Pi-Alkyl ALA 198 5.2 NH-F (ILE 235) 2.22 F-NE2((HIS 201) 3.69 Pi-Alkyl LEU 162 4.82 pi-pi interaction HIS 201 4.81 4e Hydrogen bon \-\-\-\-- Halogen F-NE2 (HIS 299) 3.2 Pi-Alkyl VAL 163 5.36 F-OD1(ASP 197) 3.45 Pi-Alkyl TRY 62 4.48 F-OE2(GLU 233) 3.14 F-OE1((GLU 233) 2.93 Co ligand Hydrogen bon H-0 (GLY 309) 2.28 Accptor -aceptor 0-0 (GLN 302) 2.98 \-\-\-\-- H-0 (GLN 302) 2.01 Donar -Donal H-H (ARG 346) 2.35 O-H (ARG 346) 2.35 H-OD1 (ASP316) 2.16 O-H (ARG 267) 2.33 {#F1} {#F2} {#F3} {#F4} {#F5}

###### Significance Voltage clamp fluorometry (VCF) is a powerful tool to track conformational rearrangement of the voltage sensor domain (VSD) of voltage-gated ion channels. The KCNQ1 channel is a voltage-gated potassium channel and its gating properties are largely dependent on which types of auxiliary KCNE proteins bind. Recent applications of VCF to KCNQ1 channel have started to reveal the mechanisms in which KCNE proteins modulate the VSD movement and the gating properties of the KCNQ1 channels. The heart is an organ for pumping blood through the entire body. The atria and ventricles alternately contract in each cardiac cycle and that is how the heart works as an efficient and rhythmic pump. These rhythmic activities are produced by electrical activity (i.e., action potentials), and each action potential is shaped by voltage-gated ion channels including Na^+^ channels, Ca^2+^ channels and K^+^ channels. Ventricular myocytes are depolarized by Na^+^ channels and Ca^2+^ channels, leading to contraction. Depolarization is typically maintained for 200--300 ms in the human heart. K^+^ channels subsequently repolarize the myocytes, leading to relaxation. Therefore, any defect of these ion channels or their associated proteins can cause prolongation (or shortening) of ventricular depolarization. As a result, the QT interval of an electrocardiogram is prolonged (or shortened) and is diagnosed as long (short) QT syndrome (LQTS, SQTS). LQTS (or SQTS) increases the risk of atrial and ventricular arrhythmias and is potentially lethal. Therefore, it is important that these ion channels work at proper timings in cardiac cycles. Structure and function of KCNQ1 channel and KCNE proteins ========================================================= The KCNQ1 channel is a member of the voltage-gated K^+^ channel family. As KCNQ1 is known as a causative gene for long QT syndrome (LQTS type 1 or LQT1), it is obviously important for shaping cardiac action potentials. KCNQ1 α subunit is a membrane protein having six transmembrane segments (S1--S6), and four α subunits form a single KCNQ1 channel (tetramer) like other voltage-gated K^+^ channels. The first four transmembrane segments (S1--S4) form a voltage-sensing domain (VSD) and the remaining two segments (S5--S6) form a pore domain (PD) ([Fig. 1a](#f1-16_121){ref-type="fig"}). The S4 segments (red helices in [Fig. 1a](#f1-16_121){ref-type="fig"}), which possess positively-charged amino acid residues, are known to move up and down voltage-dependently by sensing membrane potential, and the upward movement of the S4 segment leads to channel opening \[[@b1-16_121]\]. In addition, KCNQ1 is known to be modulated by auxiliary KCNE subunits. Five KCNE genes (KCNE1-5) have been identified in the human genome. Gating properties of KCNQ1 channel are largely dependent on which types of KCNE proteins bind. KCNE1, for example, binds to a KCNQ1 channel and slows its activation and deactivation kinetics by 100 times \[[@b2-16_121],[@b3-16_121]\]. KCNE1 also shifts the voltage-dependence (G-V) curve of KCNQ1 channel by 40 mV in a positive direction. In other words, KCNE1 makes the KCNQ1 channel much more difficult to be activated ([Fig. 1b](#f1-16_121){ref-type="fig"}). Since KCNE1 has also been identified as a causative gene for LQTS (LQT5), both KCNQ1 and KCNE1 subunits are indispensable for proper rhythmic activity of the heart. The stoichiometry of the KCNQ1-KCNE1 channel has been a matter of debate for quite a while. It has been reported that two KCNE1 subunits bind to a KCNQ1 tetramer (4:2 stoichiometry) \[[@b4-16_121]--[@b6-16_121]\]. We and other groups, on the other hand, showed that multiple stoichiometry (4:1\~4:4 stoichiometry) is possible at least *in vitro* and human embryonic stem cell derived cardiomyocytes \[[@b7-16_121]--[@b10-16_121]\]. Further study, especially *in vivo* experiments, may be necessary to determine which stoichiometries are predominant in the human heart. Modulation of KCNQ1 voltage sensor movement by KCNE1 ==================================================== It has also long been debated how KCNE1 changes the KCNQ1 gating. Initial reports suggested that KCNE1 directly interacts with the PD of KCNQ1 and prevents it from opening \[[@b11-16_121]--[@b14-16_121]\]. We and other groups showed that VSD movement (or state transition) was affected by co-expression of KCNE1 by utilizing the substituted cysteine accessibility method (SCAM) \[[@b15-16_121]--[@b17-16_121]\]. Those studies led to works focusing on the VSD movement of KCNQ1 channels, and works in which voltage-clamp fluorometry (VCF) was used to track the KCNQ1 VSD revealed the detailed mechanisms by which KCNE1 regulates KCNQ1 gating via the VSD \[[@b18-16_121]--[@b20-16_121]\]. In VCF experiments, a cysteine residue was introduced at the extracellular side of the S4 segment (S3--S4 linker) by site-directed mutagenesis. Fluorescent dye such as tetra-methylrhodamine maleimide or Alexa488 maleimide was then attached to the site of the introduced cysteine residue to track voltage-dependent conformational rearrangement by detecting change of fluorescence intensity under a voltage-clamp ([Fig. 2a](#f2-16_121){ref-type="fig"}) \[[@b21-16_121]--[@b23-16_121]\]. VCF experiments clearly showed that KCNQ1 VSD movement was greatly changed by KCNE1, though the change was far more complicated than expected \[[@b18-16_121],[@b20-16_121]\]. In KCNQ1, the fluorescence-voltage (F-V) relationship (representing the VSD movement) and the G-V relationship (representing the pore opening) are mostly overlapped ([Fig. 2b](#f2-16_121){ref-type="fig"} upper panel) \[[@b19-16_121]\]. This is because each individual VSD movement contributes sequentially to channel conductance in KCNQ1 channel \[[@b24-16_121],[@b25-16_121]\], not like other shaker-type potassium channels in which all four VSD must be in the up state before opening. On the other hand, KCNE1 shifts the G-V relationship in the positive direction, while the F-V relationship is split into two components, a steep component F1 and a shallow component F2 ([Fig. 2b](#f2-16_121){ref-type="fig"} bottom panel). F1 reflects a non-conducting state(s) from −160 to −50 mV and F2 nearly overlaps the G-V relationship. It is assumed that there are three states in KCNQ1 gating: closed state (S4 segments in the down state and the PD closed), intermediate state (S4 segments in the up state and the PD closed), and open state (S4 segments in the "full" up state and the PD open). The results of VCF studies indicated that KCNE1 may stabilize the intermediate state and that is why the F-V relationship is split into two ([Fig. 2c](#f2-16_121){ref-type="fig"}). Since the intermediate state is non-conducting, channel opening requires higher and longer depolarization. Conversely, the channel requires much larger hyperpolarization (−200 mV) to enter the "fully" down state \[[@b18-16_121],[@b20-16_121]\]. Since membrane potential of cardiomyocytes rarely reaches −200 mV, most of the KCNQ1 channels with KCNE1 subunits should be in the intermediate state. Steric hindrance between the VSD and PD of KCNQ1 induced by KCNE1 ================================================================= As discussed in the previous section, co-expression of KCNE1 dramatically changes the voltage-dependence of the KCNQ1 VSD and that is at least one of the mechanisms by which KCNE1 slows KCNQ1 gating kinetics. However, it still remains largely unknown how KCNE1 changes the voltage-dependence of KCNQ1. We previously made a series of cysteine mutants of the KCNQ1 S3--S4 linker and the first half of the S4 segment (G219C-Q234C) to identify the best site for SCAM study \[[@b15-16_121]\], and we noticed that the F232C mutant did not show the characteristic positive G-V shift induced by KCNE1. Phe232 (F232) is located next to the second positive charge (R231) of the S4 segment. We therefore made several other Phe232 mutants (mostly hydrophobic amino acid residues) to explore the role of Phe232 in the positive G-V shift induced by KCNE1. Remarkably, the G-V curve shifts induced by KCNE1 were dependent on the bulkiness of the amino acid side chain ([Fig. 3a](#f3-16_121){ref-type="fig"}). Interestingly, KCNQ1 mutants without KCNE1 did not show bulkiness-dependence in the G-V curves. Phe232 affects the G-V curve only in the presence of KCNE1 \[[@b26-16_121]\]. According to the closed state and open state of KCNQ1-KCNE1 structural models by Sanders's group \[[@b27-16_121]\], Phe232 on the S4 segment of the VSD faces the S5 segment of the PD of an adjacent subunit, and Phe232 seems to have to hurdle another bulky Phe279 on the S5 segment during the transition from the closed state to the open state. Therefore, we hypothesized that Phe232 collides with Phe279 during upward movement of the S4 segment ([Fig. 3b](#f3-16_121){ref-type="fig"}). It turned out that was probably the case because the size of the Phe279 side chain affected the G-V relationship as the size of the Phe232 side chain did. The larger the introduced side chain at Phe279 was, the greater was the shift in the G-V relationship in the positive direction and, again, the G-V relationship was shifted only in the presence of KCNE1 ([Fig. 3c](#f3-16_121){ref-type="fig"}). We concluded that the presence of KCNE1 created a steric hindrance between two phenylalanine residues of the S4 and S5 segments, and this could be the reason why activation of the KCNQ1-KCNE1 channel becomes more difficult than that of the KCNQ1 channel. We next applied VCF to reveal which gating step was mostly affected by steric hindrance of the two phenylalanine residues. By attaching Alexa 488 maleimide to G219C mutation on the S3--S4 linker, ionic currents and fluorescence intensity were simultaneously recorded from wild-type and F232A and F279A mutants of KCNQ1 co-expressed with KCNE1 ([Fig. 4a and b](#f4-16_121){ref-type="fig"}). In the wild-type KCNQ1-KCNE1 channel, an apparent delay of ionic currents (PD opening) after the increase of fluorescence (upward movement of the VSD) was observed. On the other hand, the delay was substantially diminished or almost disappeared in F232A or F279A mutants ([Fig. 4a](#f4-16_121){ref-type="fig"}). In addition, the ratio of a shallow (presumably F2) component was increased in the F-V relationships of the mutants, indicating the transition from the intermediate state to the open state requires less depolarization in the mutants ([Fig. 4b](#f4-16_121){ref-type="fig"}). Therefore, it was concluded that the steric hindrance mostly affects the gating transition from the intermediate state to the open state and prevents the channel from opening. Why does steric hindrance occur between Phe232 and Phe279 in the KCNQ1 channel only when KCNE1 is co-expressed? In movies of MD simulation of the voltage-gated K^+^ channel gating provided by Jensen *et al*. in 2012, four VSDs are moving around slightly and have a loose coupling with the PD \[[@b1-16_121]\]. We hence propose that the distance between the S4 and S5 may become smaller due to the interaction of KCNE1 with both the VSD and PD \[[@b28-16_121]\]. Only when KCNE1 interacts with the KCNQ1 channel, Phe232 on S4 and Phe279 on S5 become close enough to collide with each other during gating, and that would prevent the channel from entering the open state ([Fig. 4c](#f4-16_121){ref-type="fig"}). F279I, a novel mutation of KCNQ1, was recently reported and shown to be responsible for short QT syndrome by electrophysiology and simulation study \[[@b29-16_121]\]. F279I was also examined in our study. Isoleucine is slightly smaller than phenylalanine, and the G-V curve of the F279I mutant is therefore shifted in the negative direction (easier to be activated) by 10 mV ([Fig. 3c](#f3-16_121){ref-type="fig"}, blue curve). This 10-mV shift may be large enough to cause short QT syndrome, which is a result of upregulation of the potassium current and shortened cardiac action potential. Comparison to KCNE3 and other recent advances revealed by VCF ============================================================= KCNQ1 can also be modulated by other types of KCNE proteins. KCNE3 is one of them and is known to make the KCNQ1 channel constitutively open. The KCNQ1-KCNE3 complex is known to be expressed in epithelial cells of the intestine and trachea to assist Cl^−^ transport \[[@b30-16_121]--[@b32-16_121]\]. We and other groups have shown by mutagenesis and SCAM experiments that KCNE3 might lock the S4 segment in the up state and that is probably the mechanism by which the channel becomes constitutively open \[[@b15-16_121],[@b16-16_121]\]. Recent VCF experiments have directly confirmed that VSDs are stabilized in the up state in KCNQ1-KCNE3 channels \[[@b33-16_121]\]. The VSD is actually not locked in the up state: the G-V curve is shifted in the negative direction by 100 mV, and the VSD therefore requires large hyperpolarization (lower than −100 mV) to start entering the down state. This means that KCNQ1-KCNE3 channels virtually stay in the open state in the physiological range of membrane potentials. The KCNQ1 channel is known to require phosphatidylinositol-4,5-bisphosphate (PIP~2~) for activation \[[@b34-16_121]\]. Recent VCF experiments have revealed that the VSD still shows voltage-dependent movement without PIP~2~. However, the channel loses functional coupling between the VSD and PD in the absence of PIP~2~ and that is why KCNQ1 is unable to open the pore \[[@b35-16_121]\]. PIP~2~ may directly interact with the S4--S5 linker to control coupling of the VSD and PD of the KCNQ1 channel. Therefore, it is possible to functionally separate the VSD from the PD by abolishing PIP~2~. VCF experiments were performed in the absence of PIP~2~ to reveal which part(s) of the KCNQ1 channels was (were) affected by KCNE proteins \[[@b36-16_121]\]. The results suggested that KCNE3 functionally interacted only with the VSD and does not need an interaction with the PD to make the channel constitutively open. On the other hand, KCNE1 functionally interacted with both the VSD and PD. While the VSD is stabilized at the intermediate state, KCNE1 also directly affects the PD \[[@b36-16_121]\]. Future direction of KCNQ1 study =============================== It still remains unknown how KCNE1 (and KCNE3) interacts with KCNQ1 and modulates the gating. Recently, Sun and MacKinnon reported the first tetrameric KCNQ1 structure determined by cryo-EM ([Fig. 1a](#f1-16_121){ref-type="fig"}) \[[@b37-16_121]\]. Because the structure lacks PIP~2~, there is no coupling between the PD and VSD, and the VSD is therefore in the up state (at 0 mV) and the PD is closed. Because it is a homo tetrameric structure, it does not offer any information on how and where KCNE proteins interact. Structures of the KCNQ1-KCNE complex are, therefore, eagerly awaited. On the other hand, as I discussed in this short review, the KCNQ1 channel goes back and forth among many states to produce a voltage-dependent (or independent) potassium current with KCNE proteins. Electrophysiological experiments including current recording and VCF should remain powerful tools for studying state transitions of ion channels. Applications of VCF to other membrane proteins ============================================== Although VCF is most suitable for tracking voltage-dependent conformational change of voltage-gated ion channels, applications of VCF to other membrane proteins such as pumps and transporters have also been reported \[[@b38-16_121],[@b39-16_121]\]. For example, conformational changes of the Na^+^/K^+^-ATPase induced by extracellular potassium are detectable by VCF \[[@b38-16_121]\]. Thanks to the recent advancement of cryo-EM, structural information of membrane proteins has been explosively increasing. This simple yet powerful tool for tracking real-time structural changes will be applicable to more diverse membrane proteins by utilizing these new structural information. I thank Professor Yoshihiro Kubo (National Institute for Physiological Sciences) for his continuous support in conducting experiments during my time in Kubo lab. This work was supported by KAKENHI (24590292, 25136724, 17K08552). **Conflict of Interest** The author declares no conflict of interests. **Author contributions** K. N. wrote the manuscript. ![Structure and ionic currents of the KCNQ1 channel with KCNE1. (a) Top view (left) and side view (right) of the tetrameric KCNQ1 structures \[[@b37-16_121]\]. S1--S3, S4, S4--S5 linker, and S5--S6 are colored in green, red, orange and yellow, respectively. Dotted circles indicate putative KCNE1 sites. S4 segments (red helices) voltage-dependently move up and down as indicated by a cyan arrow in the right figure. (b) Representative currents and G-V curves of the KCNQ1 channel in *Xenopus* oocytes with or without KCNE1 co-expression. G/G~max~ indicates apparent open probability of the channel. Co-expression of KCNE1 slows the activation of the channel and shifts the G-V curve in the positive direction (The channel requires higher membrane potential to be activated.). Inset shows the voltage clamp protocol. The currents at 0 mV are colored in red for comparison. Figures and the data presented are adapted from \[[@b26-16_121],[@b40-16_121]\].](16_121f1){#f1-16_121} ![KCNE1 stabilizes the intermediate state of the KCNQ1 channel. (a) Preparation of fluorescence-labelled ion channels expressed in *Xenopus* oocytes for VCF experiments. Fluorescence changes are acquired by a photomultiplier tube (or a photodiode) while ionic currents are simultaneously recorded under two-electrode voltage clamp. (b) G-V (black; related to pore opening) and F-V (green; related to VSD movement) relationships of KCNQ1 alone and the KCNQ1-KCNE1 channel. The F-V relationship of KCNQ1-KCNE1 is split into two components, F1 and F2 \[[@b20-16_121]\]. (c) Model of three states of KCNQ1 channel gating without KCNE1 (upper row) and with KCNE1 (bottom row). Green circles indicate attached fluorophores (e.g., Alexa Fluor 488) for the VCF experiments. S4 segments (red) voltage-dependently move up and down (cyan arrows in upper row). KCNE1 stabilizes the intermediate state. Figures and the data presented are adapted from \[[@b26-16_121],[@b40-16_121]\].](16_121f2){#f2-16_121} ![Bulky amino acid residues on S4 and S5 prevent the KCNQ1-KCNE1 channel from opening. (a) Side chain dependence of G-V curves (left) and voltage of half activation (V~1/2~, right) in KCNQ1 F232 mutant series. Mutation of a larger amino acid residue tends to require higher membrane potential to be activated. (b) Locations of F232 on S4 (red) and F279 on S5 (yellow) in the cryo-EM structure of the KCNQ1 channel \[[@b37-16_121]\]. S4 moves up and down in the direction of the cyan arrow. (c) Side chain dependence of G-V curves (left) and voltage of half activation (V~1/2~, right) in KCNQ1 F279 mutant series. As observed in the F232 mutants, mutation of a larger amino acid residue tends to require higher membrane potential to be activated. Figures and the data presented are adapted from \[[@b26-16_121],[@b40-16_121]\].](16_121f3){#f3-16_121} ![F232 and F279 create a delay of channel opening after upward VSD movement. (a) Representative ionic currents (black traces) and fluorescence intensity (green traces) of KCNQ1 WT, F232A and F279A in the presence of KCNE1. Membrane potential was +40 mV from the holding potential of −90 mV. (b) F-V (green) and G-V relationships (black) of WT, F232A and F279A in the presence of KCNE1. (c) S4 segment (red bar) easily moves upward during depolarization in the absence of KCNE1. In the presence of KCNE1, the distance between S4 and S5 (yellow bar) becomes smaller and that may cause steric hindrance between the two phenylalanine residues. Figures and the data presented are adapted from \[[@b26-16_121],[@b40-16_121]\].](16_121f4){#f4-16_121}

Background ========== Muscle force is considered to be an important determinant for physical performance, activities of daily living, and work or sport performance \[[@B1]\]. Several processes such as aging, development of pathological symptoms, or injury may result in reduced muscle force. Muscle force can be quantified by several viable instruments. Precise measurements are feasible by employing hand-held dynamometers, which allows muscle force to be measured on a continuous scale. Several authors have demonstrated in various settings that hand-held dynamometry is reliable and the data valid for quantifying muscle force. They ascertained an intraclass correlation coefficient (ICC) of 0.8 or higher, indicating good sound reliability \[[@B2]-[@B11]\]. However, precise and reliable measurement outcomes are only meaningful if they can be compared with unaffected muscle groups or, more precisely, with reference values. For example, chronically ill patients may exhibit a bilateral decrease of muscle force. This signifies that the extent of the decline in a specific patient can only be quantified if measured muscle force values are compared with objective reference values. This emphasizes the relevance for the utilization of reference values with which to compare the outcomes of those measurements \[[@B8]\]. However reference values are employed in generally every type of physical examination and are often generated for a specific population. For example, for six minutes walking or pulmonary function tests, reference values are based on a population's particular origin, and outcomes demonstrate considerable mutual differences \[[@B10],[@B11]\]. Muscle force values, however, are utilized without any ethnic, geographic, or cultural background taken into consideration. Until now, reference values used in clinical practice and in research in the Netherlands are based on populations in the USA. The consideration is justified if reference values for the American population can be generalized to the Dutch population. However, geographical location and cultural backgrounds vary considerably and, therefore, this generalization may not be credible. The first objective of this study is to confirm the reliability of the use of hand-held dynamometers for isometric strength measurement; the second objective is to determine references values for a population of Dutch workers; the third objective is to compare these values with those of the USA population presented in studies by Bohannon and Andrews \[[@B3],[@B9]\]. Comparison between reference values for muscle force has not been previously performed. Methods ======= Subjects -------- The subjects have been employed in several fields, had miscellaneous physical workloads, and were recruited via local press from different localizations in the Netherlands. Inclusion criteria: Subjects had to meet the inclusion criteria including being between 20 to 60 years of age and working at least 20 hours per week. No absence from work due to illness for more than 2 weeks in the year prior to participation. Subjects were included after providing informed consent and signing a statement of good health after meeting the criteria of the Physical Activity Readiness Questionnaire \[[@B12],[@B13]\]. Exclusion criteria: Subjects were excluded if systolic and diastolic blood pressures exceeded 159 mmHg and 100 mmHg, respectively, as described by the WHO, to prevent cardiovascular injury \[[@B14]\]; were absent from work in the last year as a result of a musculoskeletal disorder; or presented co-morbidities relating to either the cardiovascular or respiratory systems or otherwise did not meet the inclusion criteria. The Medical Ethics Committee of the University Medical Center Groningen, the Netherlands, approved the study protocol. Measurement procedure --------------------- The subjects' gender, age, hand dominance, height, weight, physical activity, and Dictionary of Occupational Titles (DOT) level were recorded \[[@B15]\]. The DOT level describes the difficulty of comprehending, nature, tasks of specific types of work, or specified occupational titles. The DOT is meant to match job requirements and employees' functional abilities and consists of five categories: sedentary, light, medium, heavy, and very heavy. Maximal isometric voluntary contraction (MVC) was measured with a MicroFet 2 hand-held dynamometer (Hogan Health Industries, Inc. 8020 South 1300 West, West Jordan, USA). Three consecutive measurements were performed with one minute intervals between contractions. Isometric muscle force from elbow flexion and extension, knee flexion and extension, and shoulder abduction were measured. The protocol consisted of one contraction for every individual muscle in the following sequence; 1 elbow flexion, 2 elbow extension, 3 knee extension, 4 knee flexion, and 5 shoulder abduction. This sequence was performed three times. Observers were allowed to begin left or right according to their preference. Subjects were asked to gradually increase their muscle force to a maximum effort which would need to be sustained for three seconds. The 'break technique' was employed whereby the examiner overpowers the maximum effort of the patient, thereby producing a measurement of eccentric muscle force \[[@B16],[@B17]\]. The average muscle force of three repetitions was calculated to compensate for and minimize measurement errors. Subjects were assessed by third or fourth year physical therapy students from the Hanze University of Applied Sciences Groningen, the Netherlands. An experienced instructor trained the students prior to the tests. Students were instructed to perform the break technique in the following manner. First, they were instructed to 'break through' the subject's muscle force by countering the force employing a continuous, slow movement. Second, they were to maintain their position and the patient's position throughout the entire test. Observers provided standardized encouragement. In the event that the observer was unable to break through the patient's generated force, this was recorded in the administration form, and that result was omitted from the data analysis. Measurements were taken in a standardized and gravity neutral body position. Measurement positions are described in Table [1](#T1){ref-type="table"} (Figures [1](#F1){ref-type="fig"}, [2](#F2){ref-type="fig"} and [3](#F3){ref-type="fig"}). ###### Description of body positions during measurements **Muscle force/movement** **Joint/Limb position** **Localization HHD** **Position subject** **Fixation subject** **Position/fixation observer** --------------------------- ------------------------------------------------------------- -------------------------------------------- --------------------------------------- --------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------- **Elbow flexion** Neutral shoulder, elbow flexed 90°; upper arm against trunk Just proximal to styloid process of radius Lying supine By body weight; feet against wall Alongside the table and test subject, leaning backward **Elbow extension** Same as in flexion Just proximal to ulnar head Same as in flexion Same as in flexion Same as in flexion **Knee flexion** Hip and knee flexed 90° Just proximal to calcaneus Sitting on table By body weight and active fixation while gripping table In front of test subject; feet fixed onto table **Knee extension** Same as in flexion Just proximal to talis Same as in flexion By body weight and active fixation. In front of test subject; fixated by body weight, gripping table, and pushing forward; HHD fixation against upper leg **Shoulder abduction** 90° abduction in shoulder Just proximal to lateral epicondyle Sitting on examination table or chair Body weight Behind subject HHD, Hand-held dynamometer. {#F1} {#F2} {#F3} Statistical analyses -------------------- All data were analyzed with SPSS 14.0. To answer the primary objective of this study, reliability of the three repeated measurements, the intraclass correlation coefficients (ICC) two-way random effects model including lower and upper confidence limits (LCL and UCL), as well as the limits of agreement (LOA) were calculated \[[@B18]\]. Limits of agreement were collectively calculated for males and females and encompassing all age groups and for each pair of repeated measurements \[[@B19]\]. ICCs were interpreted as follows: ICC \< 0.25 is low reliability; 0.25 \< ICC \< 50 moderate reliability; 50 \< ICC \< 75 good reliability and ICC \> 0.75 is excellent reliability \[[@B19]-[@B21]\]. To address the second objective, reference values for muscle force were constructed by calculating means and standard deviations. Results are stratified by age groups and gender. Differences between males and females were analyzed utilizing independent t-tests. To investigate the degree to which muscle force is linearly related to four independent variables, i.e., gender, weight, height and age, a linear regression analysis was performed. Due to these regression equations and the explained variance, the r^2^ was calculated. To answer the third objective, comparisons between muscle force outcomes of the current study and two different studies were performed by comparing the means of the two other studies with our means and a 95% confidence interval (95%CI) \[[@B3],[@B9]\]. Results ======= Subjects -------- A sample of 462 healthy subjects (259 males and 203 females) was included in this study. The subject and group characteristics are presented in Table [2](#T2){ref-type="table"}. ###### Characteristics of the population stratified by age group and gender ------------ ----------- ---------- ------------ ---------- ---------- **Male** 20-59 20-29 30-39 40-49 50-59 Age 41.7(11) 25.2(3) 33.6(3) 44.9(3) 54.1(3) Height 182.1(8) 182.0(8) 181.4.6(8) 183.5(8) 181.2(7) Weight 80.8(12) 74.3(10) 80.6(13) 82.4(10) 82.4(14) BMI 24.4(4) 22.4(3) 24.4(4) 24.4(3) 25.1(4) DOT 1 18.9 10.0 22.0 25.4 15.4 DOT 2 44.3 33.3 24.3 46.0 60.0 DOT 3 22.9 43.3 34.1 12.7 16.9 DOT 4 13.9 13.3 19.5 15.9 7.7 **Female** 20-59 20-29 30-39 40-49 50-59 Age 40.2(10) 25.9(3) 34.8(3) 44.3(3) 53.6(3) Height 170.1(7) 172.5(6) 170.8(8) 170.0(7) 167.4(7) Weight 68.0(11) 68.1(13) 68.2(9) 68.0(11) 67.8(12) BMI 23.4(3) 22.9(4) 23.5(3) 23.4(3) 24.1(4) DOT 1 21.2 14.0 21.1 12.7 43.6 DOT 2 42.5 41.9 33.4 54.0 33.3   DOT 3 35.3 44.2 45.5 30.2 23.1 DOT 4 1 0 0 3.1 0 ------------ ----------- ---------- ------------ ---------- ---------- Age, Height, Weight and BMI expressed in mean(sd) and DOT is expressed in percentage of the population. DOT; Dictionary of Occupational Titles. DOT 1 = sedentary, DOT 2 = light, DOT 3 = medium, DOT 4 = heavy/very heavy. Reliability ----------- Correlations between the different measurements vary between 0.83 and 0.94, ICC values vary between 0.83 and 0.92 and are presented in Table [3](#T3){ref-type="table"}. Since the confidence intervals were small, it is relatively certain that the population values of the coefficients are similar to the estimated values. All ICC values were higher than 0.75, indicating good reproducibility for all ten muscle measurements \[[@B19],[@B21]\]. The limits of agreement varied between 37.0 and 117.8 Newton. Elbow extension left demonstrates a small 95%CI while knee extension right exhibits a large range of the 95%CI. ###### Correlation between the three measurements, intraclass correlation coefficient, limits of agreement, for three repeated measurements **Muscle force** **Corr1-2** **Corr1-3** **Corr2-3** **ICC** **LCL- UCL** **LOA1-2** **LOA1-3** **LOA2-3** ------------------- ------------- ------------- ------------- --------- -------------- ------------ ------------ ------------ Elbow flex. left 0.88 0.87 0.87 0.87 0.85 - 0.89 ± 59.4 ± 61.6 ±61.0 Elbow flex. right 0.85 0.86 0.89 0.87 0.85 - 0.89 ± 67.8 ± 66.8 ± 56.3 Elbow ext. left 0.89 0.85 0.88 0.88 0.86 - 0.89 ± 42.8 ± 50.3 ±44.1 Elbow ext. right 0.91 0.91 0.93 0.92 0.90 - 0.93 ± 37.0 ± 37.2 ±32.9 Knee flex. left 0.83 0.81 0.87 0.84 0.81 - 0.86 ± 75.9 ± 82.5 ±68.8 Knee flex. right 0.87 0.82 0.88 0.86 0.83 - 0.88 ± 69.8 ± 81.4 ±67.8 Knee ext. left 0.88 0.85 0.91 0.88 0.86 - 0.89 ± 104.0 ±117.8 ±94.3 Knee ext. right. 0.91 0.90 0.94 0.92 0.90 - 0.93 ± 96.3 ± 106.7 ± 80.2 Abduction left 0.83 0.80 0.87 0.83 0.80 - 0.86 ± 55.2 ± 59.5 ±48.4 Abduction right 0.85 0.87 0.92 0.88 0.85 - 0.90 ± 52.6 ± 46.5 ±37.3 Corr: Correlation, ICC: intraclass correlation coeficient, LCL: lower control limit, UCL: upper control limit, LOA: limits of agreement. Reference values ---------------- Tables [4](#T4){ref-type="table"} and [5](#T5){ref-type="table"} illustrate the mean muscle force values for reference values from elbow flexion and extension, knee flexion and extension, and shoulder abduction stratified by age groups, gender, and dominance. Regression equations and explained variance are presented in Table [6](#T6){ref-type="table"}. Regression equations were calculated with height, weight, age, and gender. The explained variance varies between 0.25 for knee extension right and 0.51 for elbow extension left. ###### Dominant and non-dominant muscle strength means (sd) per age group for males **Male**   **Dominant** **Non dominant** --------------------- ------- -------------- ------------------ ---------- ---------- Elbow flexion 20-29 48 281(48) 48 261(49)   30-39 51 273(50) 51 266(51) 40-49 70 271(59) 70 261(51) 50-59 59 259(52) 59 245(47) Elbow extension 20-29 48 186(38) 48 182(37)   30-39 51 183(40) 51 179(45) 40-49 70 185(46) 70 179(44) 50-59 59 181(37) 59 173(36) Knee flexion 20-29 48 267(57) 48 252(52)   30-39 51 262(60) 51 250(55) 40-49 68 274(77) 69 263(77) 50-59 59 242(57) 59 234(55) Knee extension 20-29 47 379(105) 47 371(112)   30-39 51 351(99) 51 341(101) 40-49 69 368(114) 70 341(107) 50-59 59 337(103) 57 335(102) Schoulder abduction 20-29 14 172(48) 14 173(35)   30-39 26 181(38) 26 176(40) 40-49 35 173(43) 35 177(40) 50-59 37 178(39) 39 177(43) ###### Dominant and non-dominant muscle strength means (sd) per age group for females **Female**   **Dominant** **Non dominant** --------------------- ------- -------------- ------------------ --------- --------- Elbow flexion 20-29 51 191(30) 51 183(30)   30-39 39 195(34) 39 186(35) 40-49 66 191(37) 66 186(37) 50-59 34 181(29) 34 166(22) Elbow extension 20-29 51 132(28) 51 131(28)   30-39 39 128(24) 39 125(26) 40-49 66 131(28) 66 125(29) 50-59 34 120(20) 34 118(27) Knee flexion 20-29 51 198(38) 51 191(37)   30-39 39 190(41) 39 188(35) 40-49 66 190(51) 67 183(52) 50-59 34 174(42) 34 169(45) Knee extension 20-29 51 261(80) 51 260(75)   30-39 38 273(87) 39 264(88) 40-49 66 262(127) 67 245(79) 50-59 34 244(66) 34 228(51) Schoulder abduction 20-29 14 115(19) 14 124(23)   30-39 22 116(26) 22 118(30) 40-49 41 119(28) 41 118(26) 50-59 15 114(22) 15 116(21) ###### Regression equations for calculation of reference values **Muscle force** **Regression equations** **R**^**2**^ -------------------------- ------------------------------------------------ -------------- Elbow flexion left -4.93 + 56.96\*S-0.64\*A + 0.89\*W + 0.89\*H 0.51 Elbow flexion right 10.67 + 57.47\*S-0.72\*A + 0.95\*W + 0.85\*H 0.49 Elbow extension left 23.85 + 36.56\*S-0.50\*A + 1.07\*W + 0.29\*H 0.44 Elbow extension right 80.39 + 41.56\*S-0.47\*A + 1.14\*W-0.06\*H 0.48 Knee flexion left 47.92 + 43.52\*S-0.60\*A + 1.36\*W + 0.40\*H 0.34 Knee flexion right 43.84 + 47.03\*S-0.71\*A + 1.33\*W + 0.50\*H 0.35 Knee extension left -204.36 + 43.69\*S-1.13\*A + 1.90\*W + 2.19\*H 0.31 Knee extension right -215.54 + 40.73\*S-0.82\*A + 2.0\*W + 2.22\*H 0.25 Shoulder abduction left -20.68 + 45.25\*S-0.04\*A + 0.64\*W + 0.56\*H 0.46 Shoulder abduction right 10.07 + 43.63-0.16\*A + 0.76\*W + 0.36\*H 0.43 A, age; W, weight; H, height; S, sex (1 for male, 0 for female). Comparison ---------- Mean muscle force values and the 95% CI from the current study and mean muscle force values from studies by Bohannon and Andrews are presented in Tables [7](#T7){ref-type="table"} and [8](#T8){ref-type="table"}. Comparison of Dutch mean muscle force values to those from Bohannon and Andrews \[[@B3],[@B9]\] revealed that an significant difference exists between reference muscle force values between different populations. ###### **Comparison between the present study and studies of Bohannon**\[[@B3]\]**and Andrews**\[[@B9]\]**for male** **Male** **Dominant** **Non dominant** ------------------------ --------------- ------------------ --------- --------------- --------- ---------- **Elbow flexion**               **Age group**               20-29 281 (267-295) 285 \- 261 (247-276) **279** \-   30-39 273 (259-287) 269 \- 266 (252-281) 281 \-   40-49 271 (258-286) 269 \- 261 (249-274) 270 \-   50-59 259 (246-272) **287** **292** 245 (232-257) **268** **272** **Elbow extension**               **Age group**               20-29 186 (175-197) **244** \- 182 (171-194) **245** **-**   30-39 185 (172-194) **214** \- 179 (167-192) **231** **-**   40-49 185 (174-196) **210** \- 179 (169-190) **214** \-   50-59 181 (171-190) **19**7 188 173 (164-182) **186** 178 **Knee extension**               **Age group**               20-29 379 (348-409) **575** \- 371 (339-404) **579** \-   30-39 351 (323-378) 573 \- 341 (312-369) **572** \-   40-49 368 (341-395) **583** \- 341 (315-366) **589** \-   50-59 337 (310-363) **471** 448 335 (308-362) **468** **-439** **Shoulder abduction**               **Age group**               20-29 172 (144-200) **258** \- 173 (152-193) **246** \-   30-39 181 (165-196) **249** \- 176 (159-192) **237** \-   40-49 173 (158-188) **246** \- 177 (163-191) **244** \-   50-59 178 (165-191) **240** 238 177 (163-191) **223** **-222** Bold printed numbers are values outside the 95%CI of this study. Age groups are given in decades in years, Muscle force is given in Newtons, LCL = lower control limit, UCL upper control limit, -- not available. ###### **Comparison between the present study and studies of Bohannon**\[[@B3]\]**and Andrews**\[[@B9]\]**for female** **Female** **Dominant** **Non dominant** ------------------------ --------------- ------------------ ----- --------------- --------- ----- **Elbow flexion**               **Age group**               20-29 191 (182-199) **155** \- 183 (175-192) **151** \-   30-39 195 (184-206) **164** \- 186 (175-198) **161** \-   40-49 191 (182-199) **151** \- 186 (176-195) **157** \-   50-59 181 (171-191) **155** 167 166 (158-174) **156** 160 **Elbow extension**               **Age group**               20-29 132 (124-139) **116** \- 131 (123-139) **115** \-   30-39 128 (121-135) **117** \- 125 (116-133) **119** \-   40-49 131 (124-137) **110** \- 125 (118-132) **112** \-   50-59 120 (113-127) **111** 108 118 (109-127) **107** 104 **Knee extension**             **Age group**               20-29 261 (234-288) **467** \- 260 (238-281) **466** \-   30-39 273 (244-302) **408** \- 264 (235-292) **411** \-   40-49 262 (231-293) **381** \- 245 (225-265) **363** \-   50-59 244 (221-267) **335** 298 230 (210-246) **319** 293 **Shoulder abduction**             **Age group**               20-29 115 (104-127) **153** \- 124 (110-137) **135** \-   30-39 116 (105-128) **139** \- 118 (104-131) **136** \-   40-49 119 (110-128) **139** \- 118 (109-126) **129** \-   50-59 114 (110-128) **137** 135 116 (104-128) **135** 124 Bold printed numbers are values outside the 95%CI of this study. Age groups are given in decades in years, Muscle force is given in Newtons, LCL = lower control limit, UCL upper control limit, -- not available. Comparison indicates that, for males, mean muscle force values of Bohannon and Andrews are greater than those of the current study except for elbow flexion of the dominant and non-dominant sides in which only the age group 50--59 years exhibits greater values. For females, mean muscle force values of Bohannon and Andrews were lower for elbow flexion and extension than those in the present study with the exception of elbow extension, non-dominant for age group 30 to 39 (Bohannon) and 50 to 59 (Andrews) years. Shoulder abduction and knee flexion and extension indicated greater values in the study of Bohannon and Andrews, except for shoulder abduction of the non-dominant side with age group 20 to 29 years. Discussion ========== Reliability of muscle force measurements with a hand held dynamometer is good to excellent. All ICC values exceeded the criterion of 0.80, indicating good reliability for all ten muscle measurements. These findings corroborate with those of Bohannon \[[@B3]\]. The LOA, however, varies substantially. Reference values for muscle force for the Dutch working population between 20 and 60 years of age are now made available. Reference values including age gender, weight and height can be calculated with regression analysis as independent predictors for muscle force. Comparison of the Dutch mean muscle force values to those published by Bohannon and Andrews revealed significant differences between reference values for muscle force values between the assessed populations. Comparison of reference values between populations have not been initiated previously. Muscle force measurements with a hand held dynamometer exhibit a good reliability as demonstrated by the ICC. The LOA, however, vary substantially. Muscle groups with a relatively low muscle force demonstrate a small range of the LOA while muscle with a greater muscle force exhibit a larger range of the LOA, indicating that measurements of stronger muscles are less precise. Though hand held dynamometers have shown to be a reliable and beneficial instrument for measuring muscle force, a hand held dynamometer may possess some practical limitations. In subjects with high Quadriceps muscle force, it might be impossible to perform a correct measurement. During our study, it was not possible to perform a correct measurement of the Quadriceps muscle in six subjects due to high muscle force as observers were not capable of performing a correct break procedure. As reliability and validity may be affected during these measurements, bias was likely present, which is the reason that these results were omitted from the analysis. The influence of exclusion of these data on reliability, regression formulas, and reference values is very limited due to the considerable sample size. Provided that observers were able to properly perform according to the protocol, the regressions formula for knee extension might be only slightly changed. In our opinion, a hand-held dynamometer is not suitable for measuring Quadriceps muscle force in stronger subjects. Reference values for muscle force for the Dutch working population between the ages of 20 and 60 years are now made available. Regression equations illustrate that gender, weight, and height are of major influence on muscle force. The effect of age, however, is limited. In several of the regression analyses, the effect of age was small, though significant, due to the considerable sample size. Regression analysis demonstrated that the effect of aging for subjects aged 20--60 years is larger for lower extremities than for upper extremities. These results are predominantly consistent with previously reported results \[[@B3],[@B9]\]. Bohannon and Andrews also reported that gender, age, height, and weight are predictors of muscle force and that age correlated significantly, though very limited, with muscle force. Comparison of the outcomes of our study to those earlier exhibited an important difference between reference values. The differences in upper extremity tests, however, were moderate in all cases, whereas most of the lower extremity differences were considerable. For instance, differences in muscle force greater than 100 Newton for knee extension may have clinical consequences as 100 Newton's may be up to 43 percent of the maximum knee extension force in the Dutch female population. The observed differences, however, exceed 100 Newton. This is all the more remarkable because, in our study, we employed the break method while, in the studies of Bohannon and Andrews, the make method is used. The break method may lead to higher levels in muscle strength measurements \[[@B16]\]. The observed differences in the lower extremity are relevant for clinical practice. It appears to be evident that these differences may probably cause unattainable and/or undesirable training goals to be set and may result in undesired side effects as these external reference values may be too high and, therefore, not suitable for the Dutch population. However, reference values formulated for the United States are, at this moment, utilized in clinical practice and research in the Netherlands. The results of our study demonstrate that reference values cannot simply be generalized to any country, geographical area, or population. Therefore, it is necessary to generate reference values for different countries or geographical areas. For other physiological tests such as the six minutes walking test reference values for specific geographic reference values are available and indicate considerable differences \[[@B10]\]. Although we did not assess cultural habits or demographic aspects of populations, it is likely that the outcomes of muscle force measurements may be influenced by several such factors. Psychological state or prior experiences related to exertion or physiological responses to exercise, exertion, or pain might have influenced the outcomes \[[@B22]\]. In addition, body composition and weight are related to muscle force as presented in the regression equations. Another potential explanation for the differences between our reference values and those previously reported by Andrews and Bohannon is the difference in time periods. The reference values of Andrews and Bohannon were determined in 1996 and ours in 2010. In approximately 20 years, some characteristics such as BMI may have changed which may affect references values equations. Study limitations ----------------- In our study, we only tested the employed working population between the ages of 20 and 60 years. Our study, therefore, only provides reference values and comparison for this group. Our study does not provide information regarding, for example, unemployed businessmen or housekeepers. Another limitation in our study is that observers were male and female. We did not register whether subjects were tested by male or female observers. The outcomes of measurements may be biased by the gender of the observer. Reliable muscle force measurements, appropriate and applicable reference values, and accurate knowledge of acquired muscle force in daily living facilitates formulating an effective and accurate rehabilitation process with clear and realistic goals and objective effects. Although reliable measurements of a person's muscle force are beneficial, no reliable procedures are currently available for translating isometric contractions or reference values, for that matter, into function. Functional tests probably provide an improved reflection of a subject's functional muscle force, capacity, or ability for activities of daily living or work. This probably indicates that the role of muscle force should be interpreted with caution and that other variables may also influence activities of daily living. Additional studies are needed to define the specific role and the amount of muscle force required in activities of daily living. Conclusions =========== Measuring muscle force by dynamometry is reliable and suitable for clinical practice. Substantial differences exist for reference muscle force values between different populations. Reference values are specific for different regions and cannot simply be generalized to other populations. Abbreviations ============= MRC: Medical research council; MVC: Maximal voluntary contraction. Competing interests =================== The authors declare that they have no competing interests. Authors' contributions ====================== R.S Performed the actual measurements and instructed other observers. Participated in in the design and coordination of the study. Made several intellectual contributions to the draft of the manuscript. W.K. Conceived and performed and coordinated the majority of the statistical analyses. Made several intellectual contributions to the design and draft of the manuscript. M.R. Made several intellectual contributions to the draft of the manuscript and participated in in the design and coordination of the study. C.S. Made several intellectual contributions to the design and draft of the manuscript. Participated in the statistical analyses. All authors read and approved the final manuscript. Pre-publication history ======================= The pre-publication history for this paper can be accessed here: <http://www.biomedcentral.com/2052-1847/6/10/prepub>

\section{\label{sec:Introduction}Introduction} While occupying myself with two articles not long ago \cite{Matsinos2022a,Matsinos2022b}, summarising the knowledge I have gained about the interaction between pions ($\pi$) and nucleons ($N$) at low energy (i.e., for pion laboratory kinetic energy $T \leq 100$ MeV), it occurred to me that the general procedure, which is routinely followed in the analyses of the ETH $\pi N$ project (as detailed in Section 3.1 of the latter report \cite{Matsinos2022b}), could be somewhat simplified. It is for this reason that I have decided to refrain from presenting this study as another update of the analysis of the low-energy $\pi N$ measurements (three such updates are available online as versions of Ref.~\cite{Matsinos2017a}), but rather start a new `thread' in the arXiv\textsuperscript{\textregistered} registration system. There are additional reasons in favour of the different placement of the material of this study. For the first time in several years, the database (DB) of the low-energy measurements will be enhanced with the addition of the differential cross sections (DCSs) of two $\pi^+ p$ experiments of the late 1970s \cite{Moinester1978,Blecher1979}. Similarly, the three-point dataset of Ref.~\cite{Ullmann1986}, corresponding to measurements of the DCS of the $\pi^- p$ charge-exchange (CX) reaction $\pi^- p \to \pi^0 n$ at $T \approx 49$ MeV, will also be appended to the $\pi^- p$ CX DB. Actually, these three datasets barely qualify for inclusion, in that the measurements had been acquired only for the sake of calibration of DCSs relevant to $\pi$-nucleus reactions. This fact alone may explain why these sixteen datapoints had appeared only in figures in the experimental reports, not in tabular form (which is one of the requirements for the inclusion of measurements in the DB of this project). Having said that, as these datapoints had been reported \emph{in some form}, it was eventually decided to accept the measurements as they are found in the SAID website \cite{SAID}. One additional datapoint undoubtedly qualifies for inclusion: two decades after the relevant experiments were conducted at the Paul Scherrer Institut (PSI), the Pionic Hydrogen Collaboration published in 2021 their final result for the total decay width $\Gamma_{1s}$ of the ground state in pionic hydrogen \cite{Hirtl2021}. Their value, which is compatible with the estimates of earlier experiments \cite{Sigg1995,Sigg1996,Schroeder2001} but is considerably more precise, will act as an important `anchor point' in the optimisation, enabling a more precise determination of the $\pi^- p$ CX scattering amplitude. Before setting forth one additional reason why this work be better categorised under a different link in the arXiv\textsuperscript{\textregistered} platform, a few words about the isospin invariance in the hadronic part of the $\pi N$ interaction~\footnote{In the following, `isospin invariance in the $\pi N$ interaction' will be used as the short form of `isospin invariance in the hadronic part of the $\pi N$ interaction'; it is known that the isospin invariance is broken in the electromagnetic (EM) interaction.} are in order. Provided that this theoretical constraint holds, only two (complex) scattering amplitudes enter the description of the three $\pi N$ reactions which can be subjected to experimental investigation at low energy, namely of the two elastic-scattering (ES) processes $\pi^\pm p \to \pi^\pm p$ and of the $\pi^- p$ CX reaction: the isospin $I=3/2$ amplitude ($f_3$ or $f^{3/2}$) and the $I=1/2$ amplitude ($f_1$ or $f^{1/2}$). The partial-wave amplitudes are usually denoted as $f_{2I,2J}$ or as $f^I_{l\pm}$, where $J$ stands for the total angular momentum of the $\pi N$ system and the quantum number $l$ identifies the orbital ($0$, $1$, $2$, $3$, \dots, for the $s$, $p$, $d$, $f$, \dots orbitals, respectively). The isospin invariance in the $\pi N$ interaction implies that the $\pi^+ p$ reaction is accounted for by $f_3 = f_{\pi^+ p}$, the $\pi^- p$ ES reaction by the linear combination $(f_3 + 2 f_1)/3 = f_{\pi^- p}$, and the $\pi^- p$ CX reaction by $\sqrt{2} (f_3-f_1)/3 = f_{\rm CX}$, see Appendix 1 of Ref.~\cite{Matsinos1997}. From these relations, the following expression (also known as `triangle identity') links together the scattering amplitudes of the physical processes $f_{\pi^+ p}$, $f_{\pi^- p}$, and $f_{\rm CX}$: \begin{equation} \label{eq:EQ001} f_{\pi^+ p} - f_{\pi^- p} = \sqrt{2} f_{\rm CX} \, \, \, . \end{equation} The violation of isospin invariance in the $\pi N$ interaction implies that Eq.~(\ref{eq:EQ001}) does not hold. Conversely, if Eq.~(\ref{eq:EQ001}) does not hold, then the isospin invariance is broken: given the isospin structure of the ES amplitudes, one would need (at least) one additional scattering amplitude (i.e., $f_3^\prime \coloneqq f_3 + \delta f_3$ and/or $f_1^\prime \coloneqq f_1 + \delta f_1$) to account for the $\pi^- p$ CX reaction. As a result, the test of the fulfilment of isospin invariance in the $\pi N$ interaction reduces to an evaluation of the amount by which the scattering amplitudes $f_{\pi^+ p}$, $f_{\pi^- p}$, and $f_{\rm CX}$ depart from the triangle identity (as well as to the assessment of the statistical significance entailed by that amount). Until now, two such tests had been implemented/used within the ETH $\pi N$ project. \begin{itemize} \item The first test rests upon the extraction of a phase-shift solution from the ES reactions and the determination (from that solution) of the scattering amplitude of the $\pi^- p$ CX reaction via Eq.~(\ref{eq:EQ001}). Having reconstructed the scattering amplitude $f_{\rm CX}$, predictions are generated for the various measurable quantities (observables), corresponding to the conditions (e.g., energy, scattering angle, etc.) at which the measurements of the $\pi^- p$ CX DB had been acquired. The comparison between these predictions and the experimental data enables the extraction of an estimate for the discrepancy between the predicted and the measured DCSs. This method does not require the extraction of the scattering amplitude $f^{\rm extr}_{\rm CX}$ from the data. \item More recently, a second test was established \cite{Matsinos2017a}, featuring two types of joint fits: used as input DB in the first type are the data of the ES reactions, whereas the DBs of the $\pi^+ p$ and $\pi^- p$ CX reactions are jointly analysed for the purposes of the second type of fits. In both cases, the $I=3/2$ partial-wave amplitudes are (predominantly) fixed from the $\pi^+ p$ reaction, leaving the determination of the $I=1/2$ partial-wave amplitudes to the relevant $\pi^- p$ reaction. The differences between the two phase-shift solutions (and those between sets of predictions emerging thereof) provide a measure of the violation of the triangle identity. \end{itemize} The more recent of these tests has been problematical for two reasons. First, the extraction of the scattering amplitude $f^{\rm extr}_{\rm CX}$ is not `clean', in that it also involves the measurements of another (i.e., in addition to those of the $\pi^- p$ CX reaction) $\pi N$ process. Second, the joint fits to the data of the $\pi^+ p$ and $\pi^- p$ CX reactions have never been satisfactory in the strict statistical sense, in that the fitted values of the scale factor $z$, see Eq.~(\ref{eq:EQA002}), have invariably exhibited a pronounced energy dependence: therefore, the fit results depart from the statistical expectation for an unbiased optimisation when the Arndt-Roper formula \cite{Arndt1972} is used as minimisation function. In fact, given that similar effects have never been observed in the joint analyses of the ES reactions, the energy dependence of the fitted values of the scale factor $z$ (obtained from the joint fits to the data of the $\pi^+ p$ and $\pi^- p$ CX reactions) was interpreted in former works as strong indication of the violation of isospin invariance on the part of the $\pi^- p$ CX reaction. While retaining the former test, this work puts forward a promising alternative for the latter, by introducing a method to extract the scattering amplitude $f^{\rm extr}_{\rm CX}$ from the $\pi^- p$ CX data without recourse to the measurements of another $\pi N$ process. The aforementioned modification proffers one additional advantage. One of the established indicators of the violation of isospin invariance is the symmetrised relative difference $R_2$: \begin{equation} \label{eq:EQ002} R_2 \coloneqq 2 \frac{\Re \left[ f_{\rm CX} - f^{\rm extr}_{\rm CX} \right]}{\Re \left[ f_{\rm CX} + f^{\rm extr}_{\rm CX} \right]} = 2 \frac{ \Re \left[ f_{\pi^+ p} - f_{\pi^- p} - \sqrt{2} f^{\rm extr}_{\rm CX} \right] }{ \Re \left[ f_{\pi^+ p} - f_{\pi^- p} + \sqrt{2} f^{\rm extr}_{\rm CX} \right] } \, \, \, , \end{equation} where the operator $\Re$ returns the real part of a complex number. The indicator $R_2$ of Eq.~(\ref{eq:EQ002}) is usually obtained (in other studies) at different energies and separately for the $s$ and $p$ waves. The direct extraction of the $\pi^- p$ CX scattering amplitude $f^{\rm extr}_{\rm CX}$, without the involvement of the measurements of another $\pi N$ process, will enable the reliable and unambiguous evaluation of $R_2$ within the ETH $\pi N$ project, throughout the low-energy region, separately for the $s$ and the $p$ waves. A similar test had been implemented/used in the distant past \cite{Matsinos1997}, but it had been replaced by a more formal statistical test, featuring the p-value of the reproduction of the low-energy $\pi^- p$ CX DB by the results of the phase-shift analysis (PSA) of the ES measurements. The structure of this study will remain as simple as possible. The following section addresses a few theoretical issues in some detail, as well as matters of the data analysis. Section \ref{sec:Results} provides all important results. The conclusions are found in the last section of this study, Section \ref{sec:Conclusions}. The appendices deal with technicalities, regarding the numerical minimisation and the reproduction of datasets on the basis of available phase-shift solutions. Those of the tables and the figures, which call for immediate inspection, will be placed in close proximity to the relevant text. Longer tables (e.g., those detailing the description or the reproduction of experimental data), as well as series of related figures (e.g., those containing the predictions of this study for the two $s$- and the four $p$-wave phase shifts), will be placed at the end of the preprint, before the appendices. The following notation is expected to facilitate the repetitive referencing to the DBs in this work. \begin{itemize} \item DB$_+$ for the $\pi^+ p$ DB; \item DB$_-$ for the $\pi^- p$ ES DB; \item DB$_0$ for the $\pi^- p$ CX DB; and \item DB$_\pm$ for the ES DB (combined DB$_+$ and DB$_-$). \end{itemize} In addition, the occasional prefix `t' (as, for instance, in the tDB$_+$) will denote a `truncated' DB, i.e., a DB after the removal of the outliers (i.e., of the measurements in the initial DB which do not tally well with the general behaviour of the bulk of the relevant data). All rest masses of particles and all $3$-momenta will be expressed in energy units. The $s$-wave scattering lengths (and the $\pi N$ scattering amplitudes) will generally be expressed in fm (and, in most cases, also in units of the reciprocal of the charged-pion rest mass ($m_c^{-1}$), which might be a more familiar unit to some readers); the $p$-wave scattering volumes will generally be given in fm$^3$ (and, occasionally, also in $m_c^{-3}$). If two uncertainties accompany a result, the first one will be statistical and the second systematic. Apart from the masses, the total decay widths, and the branching fractions of a few higher states (entering the Feynman diagrams of the ETH model, see Section \ref{sec:Models}), the physical constants will be fixed from the 2022 compilation of the Particle-Data Group (PDG) \cite{PDG2022}. Finally, DoF will stand for `degree of freedom' and NDF for `number of DoFs'. Distinction must made between the `NDF of a DB', representing the total number of measurements contained in that DB, reduced by the number of datasets (of that DB) which have lost their absolute normalisation (as a result of the application of the analysis procedure which will be set forth in Section \ref{sec:Procedure}), and the `NDF of/in a fit', which is equal to the NDF of the fitted DB, reduced by the number of free model parameters in that fit. \section{\label{sec:Method}Modelling, database, analysis procedure} \subsection{\label{sec:Models}On modelling the hadronic part of the $\pi N$ interaction} The essential details about the modelling options of the hadronic part of the $\pi N$ interaction can be found in Section 3.1 of Ref.~\cite{Matsinos2022a}. Used within the ETH $\pi N$ project are two such options. \begin{itemize} \item In the first phase of each analysis, the $s$- and $p$-wave $K$-matrix elements $K_{2I,2J}$ (or the reciprocal quantities) are parameterised by means of simple polynomials, suitable for low-energy $\pi N$ applications. Regarding that phase, the ETH parameterisation of Ref.~\cite{Matsinos2022b} has been used in all analyses since 1997 \cite{Fettes1997}. The primary task in this phase is the removal of the outliers from the DB, and the preparation of the input for the next (and, in Physics terms, more interesting) phase. \item In the second phase of each analysis, the same quantities are modelled by means of the partial-wave amplitudes of the ETH model \cite{Matsinos2022a,Matsinos2017a,Matsinos2014}. \end{itemize} The ETH model is an isospin-invariant hadron-exchange model, which obeys crossing symmetry~\footnote{The scattering amplitudes of the ES processes are linked via the interchange $s \leftrightarrow u$ in the invariant amplitudes $A_{\pm}(s,t,u)$ and $B_{\pm}(s,t,u)$, where $s$, $t$, and $u$ are the usual Mandelstam variables.}. The model is (predominantly) based on $f_0(500)$ and $\rho(770)$ $t$-channel exchanges, as well as on the $N$ and $\Delta(1232)$ $s$- and $u$-channel contributions, see Fig.~\ref{fig:GraphsETH}. The small effects of the well-established (four-star) $s$ and $p$ higher baryon resonances (HBRs) with masses up to $2$ GeV are also analytically included \cite{Matsinos2017a,Goudsmit1994}; the physical properties of these HBRs are currently fixed from Ref.~\cite{Matsinos2020a}. The derivative coupling of the scalar-isoscalar meson $f_0(500)$ (to the pion) was added in Ref.~\cite{Matsinos1997} for the sake of completeness. Important details about the historical development of the ETH model can be found in Ref.~\cite{Matsinos2017a}. \begin{figure} \begin{center} \includegraphics [width=15.5cm] {FeynmanGraphsETHZ.eps} \caption{\label{fig:GraphsETH}The dominant Feynman diagrams of the ETH model: scalar-isoscalar ($I=J=0$) and vector-isovector ($I=J=1$) $t$-channel graphs (upper part), and $N$ and $\Delta(1232)$ $s$- and $u$-channel graphs (lower part). Not shown in this figure, but also analytically included, are the small contributions from the well-established (four-star) $s$ and $p$ HBRs with masses up to $2$ GeV and known branching fractions to $\pi N$ decay modes \cite{Matsinos2020a}, as well as those from the $t$-channel exchanges of four (three scalar-isoscalar and one vector-isovector) mesons with masses up to $2$ GeV and known branching fractions to $\pi \pi$ decay modes \cite{Matsinos2020b}.} \vspace{0.5cm} \end{center} \end{figure} Before 2019, the $t$-channel contributions to the $s$ and $p$ partial-wave amplitudes of the ETH model were accounted for by the exchange of one scalar-isoscalar ($I^G(J^{PC})=0^+(0^{++})$ or $I=J=0$) meson (i.e., of the $f_0(500)$, simply named $\sigma$-meson in other works) and of one vector-isovector ($I^G(J^{PC})=1^+(1^{--})$ or $I=J=1$) meson (i.e., of the $\rho(770)$). On account of consistency, there is no reason to refrain from analytically including in the model the $t$-channel exchanges of all scalar-isoscalar and vector-isovector mesons with masses up to $2$ GeV and known branching fractions to $\pi \pi$ decay modes, considering that the corresponding effects of the $s$ and $p$ HBRs (in that mass range) to the $s$ and $u$ channels have been part of the ETH model for over twenty-five years \cite{Goudsmit1994}. The current version of the model includes four such Feynman diagrams, three associated with the exchange of scalar-isoscalar mesons and one with the exchange of the only vector-isovector meson above the $\rho(770)$ (and up to $2$ GeV) with known branching fraction to the $\pi^+ \pi^-$ decay mode; the physical properties of these states are currently fixed from Ref.~\cite{Matsinos2020b}. The effort notwithstanding, the impact of all these additions (to the dominant Feynman diagrams of the ETH model shown in Fig.~\ref{fig:GraphsETH}) on the analysis is (taking everything into consideration) nugatory. Information about the model parameters can be obtained from several sources, e.g., from Refs.~\cite{Matsinos2017a,Matsinos2014,Goudsmit1994}. Regarding the $f_0(500)$, the recommendation by the PDG is to make use of a Breit-Wigner mass between $400$ and $550$ MeV \cite{PDG2022}. In the current implementation, the joint fits of the ETH model to the tDB$_\pm$ are instead carried out at one hundred $m_\sigma$ values, randomly generated in normal distribution according to the recent result: $m_\sigma=497^{+28}_{-33}$ MeV \cite{Matsinos2020b}. All uncertainties in this work (in the estimates for the model parameters, for the $\pi N$ phase shifts, for the low-energy constants (LECs) of the $\pi N$ interaction, etc.) contain the effects of the $m_\sigma$ variation, as well as the Birge factor $\sqrt{\chi^2_{\rm min}/{\rm NDF}}$ (if exceeding $1$), which takes account of the quality of each fit \cite{Birge1932}. When a fit to the data is performed treating all eight parameters of the ETH model as free, it turns out that the quantities $G_\sigma$, $G_\rho$, and $x$ are strongly correlated. To reduce the correlations, the fits of the ETH model have been carried out (since a long time) using a pure pseudovector $\pi N$ coupling (i.e., using $x=0$). Therefore, each fit (at a fixed $m_\sigma$ value) involves the variation of the following seven parameters. \begin{itemize} \item Scalar-isoscalar $t$-channel Feynman diagram ($f_0(500)$ exchange): $G_\sigma$ and $\kappa_\sigma$; \item Vector-isovector $t$-channel Feynman diagram ($\rho(770)$ exchange): $G_\rho$ and $\kappa_\rho$; \item $N$ $s$- and $u$-channel Feynman diagrams: $g_{\pi N N}$; and \item $\Delta(1232)$ $s$- and $u$-channel Feynman diagrams: $g_{\pi N \Delta}$ and $Z$. \end{itemize} The $s$ and $p$ HBRs do not introduce any free parameters \cite{Matsinos2014}. The same applies to the $t$-channel contributions from the $f_0(980)$, $f_0(1500)$, $f_0(1710)$, and $\rho(1700)$, see the last version of Ref.~\cite{Matsinos2017a} for details. It must be mentioned that the low-energy $\pi N$ data could be fitted to with fewer model parameters. For instance, the coupling constant $g_{\pi N \Delta}$ could be fixed from the decay width of the $\Delta(1232)$ resonance. In addition, the derivative coupling $\kappa_\sigma$ could be set to $0$: since its inclusion in the mid 1990s, the fitted $\kappa_\sigma$ values have always (if my memory serves me) been compatible with $0$. Therefore, the low-energy $\pi N$ data could be fitted to with just five model parameters. Although this possibility might be explored in the future, the freedom and flexibility of the seven-parameter optimisation of the description of the tDB$_\pm$ will be retained at this time. Unlike the works of the recent past, an exclusive fit of the ETH model to the tDB$_0$ will be carried out in this work. There can be no doubt that the scalar-isoscalar $t$-channel Feynman diagram does not contribute to the $\pi^- p$ CX scattering amplitude, at least at the lowest (tree-level) order. Although this is an inevitable outcome (owing to the electrical neutrality of the $f_0(500)$), it has been corroborated by the results of the seven-parameter fit to the tDB$_0$: the resulting $\chi^2_{\rm min}$ value is about $335.44$ for $325$ DoFs, and the fitted values of $G_\sigma$ and $\kappa_\sigma$ came out close to $0$. Unfortunately however, the numerical evaluation of the Hessian matrix failed in this fit (the matrix contains negative diagonal elements and the MINUIT software library takes action to enforce positive-definiteness). As a result, one cannot accept the results of the seven-parameter fit of the ETH model to the tDB$_0$ as reliable. The easiest way to reduce the correlations among the model parameters (while retaining connection to the physical reality) is to set $G_\sigma$ to $0$ GeV$^{-2}$ (the value of $\kappa_\sigma$ is irrelevant in this case, hence it can also be set to $0$) and carry out the exclusive fit of the ETH model to the tDB$_0$ using five free parameters. Of course, the consequence of this choice is that the higher-order contributions (e.g., see Fig.~\ref{fig:LoopGraphs}) involving the scalar-isoscalar $t$-channel exchange, which may be thought of as entering the $\pi^- p$ CX scattering amplitude via the unitarisation procedure, are also explicitly suppressed. The benefit, however, outweighs the drawback: this action reduces the correlations, resulting in the $\chi^2_{\rm min}$ value of about $337.32$ for $327$ DoFs and correctly emerging Hessian matrix. In relation to this fit, one last remark is due. As mentioned in Section \ref{sec:Introduction}, the $\pi^- p$ CX scattering amplitude is proportional to the difference between the two isospin amplitudes $f_3$ and $f_1$. Only this difference can reliably be determined from the exclusive fit of the ETH model to the tDB$_0$. Exclusive fits of the ETH model to the data of any of the three low-energy $\pi N$ reactions cannot (reliably) determine \emph{both} isospin amplitudes, which explains why extracting a phase-shift solution from such fits does not make much sense. The same applies to the extraction of corresponding estimates for the spin-isospin scattering lengths/volumes: the exclusive fit of the ETH model to the tDB$_0$ can (reliably) determine only the LECs $b_1$, $c_1$, and $d_1$ in the usual low-energy expansion of the $\pi N$ scattering amplitude, see Eq.~(\ref{eq:EQ004}). \begin{figure} \begin{center} \includegraphics [width=15.5cm] {LoopDiagrams.eps} \caption{\label{fig:LoopGraphs}Examples of Feynman diagrams which, along with the scalar-isoscalar ($I=J=0$) $t$-channel graph of Fig.~\ref{fig:GraphsETH} (left graph, upper part), are suppressed in the exclusive fit of the ETH model to the tDB$_0$.} \vspace{0.5cm} \end{center} \end{figure} \subsubsection{\label{sec:IBFD}Isospin-breaking Feynman diagrams relevant to the $\pi N$ interaction} As it does not contain isospin-breaking Feynman diagrams, the ETH model is isospin-invariant by construction. There are two consequences of employing an isospin-invariant model in the partial-wave analyses (PWAs) of data which might contain isospin-breaking effects (such as the measurements of the $\pi^- p$ CX reaction presumably do \cite{Matsinos2022a,Matsinos2017a,Matsinos1997,Matsinos2006}). \begin{itemize} \item First, one part of the isospin-breaking effects are absorbed in the model parameters, which thus become effective. \item Second, the result of the presence of hadronic effects in the data, which have no direct counterpart in the modelling of the hadronic part of the $\pi N$ interaction, usually leads to the increase in the $\chi^2_{\rm min}$ value of the fits (in comparison with the typical values obtained from the description of data which contain no such effects). In this context, significant effects were observed in the past, whenever replacing the tDB$_-$ by tDB$_0$ in the joint fits of the ETH model \cite{Matsinos2022a,Matsinos2017a,Matsinos2006}. \end{itemize} At this moment, one question arises. What modifications/additions would be required so that the ETH model accommodate any isospin-breaking effects? To answer this question, one first ought to identify the potential sources of isospin-breaking effects in the $\pi N$ interaction in the form of Feynman diagrams, i.e., the mechanisms which would entail a departure of the scattering amplitudes $f_{\pi^+ p}$, $f_{\pi^- p}$, and $f_{\rm CX}$ from Eq.~(\ref{eq:EQ001}). In fact, two such possibilities have been documented since a long time: the first mechanism, the Quantum-Mechanical (QM) admixture of the $\rho^0$ and the $\omega(782)$ mesons (usually referred to as `$\rho^0 - \omega$ mixing' in the literature), was suspected of affecting the ES processes; the second, the QM admixture of the $\pi^0$ and the $\eta$ mesons (usually referred to as `$\pi^0 - \eta$ mixing' in the literature), could have an impact on the $\pi^- p$ CX reaction. As both the $\omega(782)$ and the $\eta$ mesons are singlets, the coupling of the former to the $\rho^0(770)$ and of the latter to the $\pi^0$ explicitly violate the isospin invariance in the $\pi N$ interaction. Regarding the QM admixture of the $\rho^0$ and the $\omega(782)$ mesons, the importance of the contributions from the Feynman diagram of Fig.~\ref{fig:RhoOmega} was assessed in Ref.~\cite{Matsinos2018} and found to be small, below the $1~\%$ level in the low-energy region. To summarise in one sentence, assuming the validity of the $t$ dependence of the effects of the $\rho^0 - \omega$ admixture of Ref.~\cite{Matsinos2018} (which was imported therein from external sources), it is unlikely that this mechanism could play an appreciable role in the low-energy $\pi N$ interaction. \begin{figure} \begin{center} \includegraphics [width=7.75cm] {IsospinBreakingRhoOmega.eps} \caption{\label{fig:RhoOmega}Feynman diagram involving the QM admixture of the $\rho^0$ and the $\omega(782)$ mesons, a potential mechanism for the violation of isospin invariance in the hadronic part of the $\pi N$ interaction in case of the ES reactions.} \vspace{0.5cm} \end{center} \end{figure} The QM admixture of the $\pi^0$ and the $\eta$ mesons was proposed as a potential source of isospin-breaking effects in the $\pi^- p$ CX reaction over four decades ago \cite{Cutkosky1979}. Given that only one Feynman diagram (the one shown in Fig.~\ref{fig:RhoOmega}) is involved in case of the ES reactions (at least at the lowest order), whereas all contributing Feynman diagrams are impacted on in case of the $\pi^- p$ CX reaction (see Fig.~\ref{fig:EtaPi0}), it seems to be realistic to anticipate that the isospin-breaking effects could leave a deeper imprint in the latter case. \begin{figure} \begin{center} \includegraphics [width=15.5cm] {IsospinBreakingEtaPi0.eps} \caption{\label{fig:EtaPi0}Feynman diagrams involving the QM admixture of the $\pi^0$ and the $\eta$ mesons, a potential mechanism for the violation of isospin invariance on the part of the $\pi^- p$ CX reaction \cite{Cutkosky1979}.} \vspace{0.5cm} \end{center} \end{figure} \subsection{\label{sec:Database}The updated low-energy DB of the ETH $\pi N$ project} The references to the low-energy $\pi N$ measurements of the DB of this project can be found in former papers, see Ref.~\cite{Matsinos2017a} and the works cited therein. Only those of the experimental reports, which attract particular attention in parts of this study, will be explicitly cited. As the word `dataset' might take on different meanings to different researchers (e.g., involving a choice of the experimental conditions which ought to remain stable/identical during the data acquisition), I shall start this section with an explanation of what the term implies in the context of the ETH $\pi N$ project. The properties of the incident beam and the (geometrical, physical, chemical) characteristics of the target were employed in the past, as the means to distinguish the results of experiments conducted at one place over a (short) period of time. However, datasets have appeared in experimental reports relevant to the $\pi N$ interaction, which not only involved different beam energies, but also contained measurements of different reactions. The requisite for accepting in the DB of this project a set of observations as comprising one dataset is that these observations share the same measurement of the absolute normalisation~\footnote{Of course, this is a necessary, not a sufficient, condition. Additional requirements may apply after the examination of the original experimental reports, in particular regarding the off-line processing of the raw experimental data.} (and, consequently, identical normalisation uncertainty). The five reasons for confining the analyses of this project to the low-energy region ($T \leq 100$ MeV) are laid out in Section 2.1 of Ref.~\cite{Matsinos2022a}. The condition for the acceptance of measurements in the DB is that they represent final results of a formal experimental activity, undertaken to the purpose of fulfilment of an accepted proposal for a new experiment, and have appeared (in a usable form) in peer-reviewed Physics journals, in the sixteen issues of the $\pi N$ Newsletter, or in (approved) dissertations. Measurements, which have found their way into the SAID DB as `private communications', without becoming broadly available to the community via the established scientific procedures, have been (and will continue to be) omitted. As mentioned in Section \ref{sec:Introduction}, seventeen new datapoints \cite{Moinester1978,Blecher1979,Ullmann1986,Hirtl2021} will be appended to the low-energy $\pi N$ DB in this study. There are two reasons why the old datasets of the first three works had not been used in the former analyses of this project. \begin{itemize} \item These datasets had only been included in three figures in Refs.~\cite{Moinester1978,Blecher1979,Ullmann1986}; they did not appear in tabular form in the original experimental reports. \item These datasets had been a by-product of the experimental investigation relevant to these papers, taken only for the sake of verification of the absolute normalisation of the DCSs which were of primary interest in those works. \end{itemize} Regarding the first dataset, the experimentalists remark \cite{Moinester1978}: ``An indication of the correctness of our absolute normalization is given by the comparison \dots between our measured values of the absolute differential cross-section for $\pi^+ p$ scattering with that measured by Bertin \etal~As is seen \dots the two datasets agree within the uncertainty on our data (about $\pm 10~\%$).'' In the light of the present-day knowledge of the reliability of the absolute normalisation of the Bertin \etal~data \cite{Bertin1976} (which nearly comprised the entire body of the available $\pi N$ measurements at low energy in the late 1970s), this remark sounds amusing. The authors of Ref.~\cite{Blecher1979} also compared their DCSs with those of (one of the datasets of) the Bertin \etal~paper, and remarked: ``Except at small angles and one extreme backward angle the agreement between the two datasets is good. At small angles the present data are in better agreement with the phase shift predictions, even though the latter were influenced by the Bertin \etal~data.'' Despite the fact that the three datasets of Refs.~\cite{Moinester1978,Blecher1979,Ullmann1986} have been accepted (albeit not without a mite of hesitation) in the DB, the ES datasets of Ref.~\cite{Bussey1973} hardly qualify for inclusion: the corresponding data had neither appeared in tabular form, nor had they been shown in a figure as genuine measurements, also containing the relevant EM contributions. Found in Table 1 of Ref.~\cite{Bussey1973} are only values of the DCS \emph{after} the removal of some EM effects (column 4: ``$d \sigma / d \Omega$ (nuclear)''), as well as some EM contributions (column 5: ``Coulomb correction which has been applied''). However, one remark in the text (``At small angles, one cannot reconstruct the value of $d \sigma / d \Omega$ including Coulomb effects at the quoted value of $\cos \theta$ by adding columns 4 and 5.'' \cite{Bussey1973}, pp.~370,376) suggests that one cannot retrieve the measured DCSs by simply adding the contributions listed in these two columns. Regarding these two three-point datasets, the remark in the SAID DB ``BU(73) 0 CERN BUSSEY, NPB58, 363(73), PC BUGG'' indicates that the original DCSs of Ref.~\cite{Bussey1973} had been received as private communication by Bugg (who had co-authored the paper in question). Unless the authors of Ref.~\cite{Bussey1973} publicise their results (preprint, published paper) or, at least, explain how the original DCSs can/have be/been reconstructed from the information available in Ref.~\cite{Bussey1973}, these datasets will not find their way into the analyses of this project (regardless of the smallness of the impact they could possibly have on the results of the optimisation). Also included in the DB are the four estimates for the two $\pi^- p$ scattering lengths $a_{\rm cc}$ and $a_{\rm c0}$ (pertaining to the $\pi^- p$ ES and CX reactions, respectively), obtained via the Deser formulae \cite{Deser1954,Trueman1961} from the PSI measurements of the strong-interaction shift $\epsilon_{1 s}$ \cite{Schroeder2001,Hennebach2014} and of the total decay width $\Gamma_{1s}$ \cite{Hirtl2021,Schroeder2001} of the ground state in pionic hydrogen, after the application of the EM corrections of Ref.~\cite{Oades2007}. The experimental result of Ref.~\cite{Hirtl2021} was formally published in 2021, and is included in the DB$_0$ for the first time. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:AccAndAc0}}}The current input values of the two $\pi^- p$ scattering lengths $a_{\rm cc}$ and $a_{\rm c0}$, obtained via the Deser formulae \cite{Deser1954,Trueman1961} from the PSI measurements of the strong-interaction shift $\epsilon_{1 s}$ \cite{Schroeder2001,Hennebach2014} and of the total decay width $\Gamma_{1s}$ \cite{Hirtl2021,Schroeder2001} of the ground state in pionic hydrogen, after the application of the EM corrections of Ref.~\cite{Oades2007}. Also quoted are the normalisation uncertainties $\delta z_j$ of these estimates (systematic uncertainties). These datapoints are not part of the SAID $\pi N$ DB. \vspace{0.25cm} \begin{center} \begin{tabular}{|l|c|c|} \hline Quantity & Ref.~\cite{Schroeder2001} & Ref.~\cite{Hennebach2014}\\ \hline $a_{\rm cc}$ (fm) & $0.12097(22)$ & $0.12098(12)$\\ $a_{\rm cc}$ ($m_c^{-1}$) & $0.08556(16)$ & $0.085571(86)$\\ $\delta z_j$ & $0.0082$ & $0.0067$\\ \hline \hline Quantity & Ref.~\cite{Schroeder2001} & Ref.~\cite{Hirtl2021}\\ \hline $a_{\rm c0}$ (fm) & $-0.1284(30)$ & $-0.1275(12)$\\ $a_{\rm c0}$ ($m_c^{-1}$) & $-0.0908(21)$ & $-0.09020(84)$\\ $\delta z_j$ & $0.022$ & $0.013$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} The composition of the low-energy $\pi N$ DB of the ETH $\pi N$ project in terms of number of entries (datapoints), arranged in datasets (following the definition given at the beginning of this section), for the usual low-energy $\pi N$ observables is shown in Table \ref{tab:DB}. \begin{itemize} \item The low-energy $\pi N$ DB at finite (non-zero) $T$ comprises measurements of the DCS, of the analysing power (AP), of the partial-total cross section (PTCS), and of the so-called `total-nuclear' cross section~\footnote{Being a popular observable in the 1970s, this quantity was obtained from measurements of the DCS after a part of the EM effects (i.e., the Coulomb peak and the Coulomb phase factors $\exp (\pm 2 i \Sigma_{l\pm})$, but not the distortions to the phase shifts and to the partial-wave amplitudes \cite{Tromborg1977}) were removed, and the resulting DCS, named (at those times) `nuclear', was integrated over the entire sphere.} (TNCS) for the ES reactions. One dataset of AP measurements contains data of both ES reactions, namely seven $\pi^+ p$ datapoints and three $\pi^- p$ ES datapoints \cite{Meier2004}. It must be borne in mind that the DCSs of the CHAOS Collaboration \cite{Denz2006}, which are not included in the DB$_\pm$ of this project, are omitted from Table \ref{tab:DB}: the two attempts to analyse these data within the context of this project a few years ago \cite{Matsinos2013a,Matsinos2015} were not satisfactory; excepting only one datapoint, the DCSs of the CHAOS Collaboration are included in the SAID $\pi N$ DB. \item In addition to the DCS and AP measurements, the DB$_0$ contains measurements of the total cross section (TCS). Furthermore, two experiments (conducted in the 1980s) measured the $\pi^- p$ CX DCS, but the experimentalists published the corresponding (fitted) values of the first three coefficients in the Legendre expansion (CLE) of their DCSs. \end{itemize} \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:DB}}}The breakdown of the low-energy DB of the ETH $\pi N$ project into reactions and measurable physical quantities. The entries represent the numbers of the datapoints and of the corresponding datasets in the DB. The data of this table have appeared in peer-reviewed Physics journals, in the sixteen issues of the $\pi N$ Newsletter, or in (approved) dissertations. The DCSs of the CHAOS Collaboration \cite{Denz2006} have been omitted from this table; the same applies to the nine measurements of the $\pi^- p$ PTCSs and TNCSs, see Ref.~\cite{Matsinos2017a}. \vspace{0.25cm} \begin{center} \begin{tabular}{|l|c|c|c|c|c|c|c|c|c|} \hline \multicolumn{10}{|c|}{Datapoints}\\ \hline Reaction & DCS & AP & PTCS & TNCS & TCS & CLE & $\epsilon_{1s}$ & $\Gamma_{1s}$ & Total\\ \hline $\pi^+ p$ & $404$ & $31$ & $24$ & $6$ & $-$ & $-$ & $-$ & $-$ & $465$\\ $\pi^- p$ ES & $251$ & $85$ & $-$ & $-$ & $-$ & $-$ & $2$ & $-$ & $338$\\ $\pi^- p$ CX & $297$ & $10$ & $-$ & $-$ & $10$ & $18$ & $-$ & $2$ & $337$\\ $\pi^+ p$ and $\pi^- p$ ES & $-$ & $10$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $10$\\ \hline Total & $952$ & $136$ & $24$ & $6$ & $10$ & $18$ & $2$ & $2$ & $1150$\\ \hline \hline \multicolumn{10}{|c|}{Datasets}\\ \hline Reaction & DCS & AP & PTCS & TNCS & TCS & CLE & $\epsilon_{1s}$ & $\Gamma_{1s}$ & Total\\ \hline $\pi^+ p$ & $35$ & $5$ & $19$ & $6$ & $-$ & $-$ & $-$ & $-$ & $65$\\ $\pi^- p$ ES & $27$ & $9$ & $-$ & $-$ & $-$ & $-$ & $2$ & $-$ & $38$\\ $\pi^- p$ CX & $36$ & $2$ & $-$ & $-$ & $10$ & $6$ & $-$ & $2$ & $56$\\ $\pi^+ p$ and $\pi^- p$ ES & $-$ & $1$ & $-$ & $-$ & $-$ & $-$ & $-$ & $-$ & $1$\\ \hline Total & $98$ & $17$ & $19$ & $6$ & $10$ & $6$ & $2$ & $2$ & $160$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} To my knowledge, the studies of this project are the only ones which include in the DB the $\pi^+ p$ PTCSs and TNCSs, as well as the $\pi^- p$ CX TCSs. It would have been controversial to also include in the DB$_-$ the nine measurements of the $\pi^- p$ PTCSs and TNCSs: an appreciable fraction ($75 - 80~\%$, depending on $T$) in these measurements originates from the $\pi^- p$ CX reaction, see Ref.~\cite{Matsinos2017a} for details. \subsection{\label{sec:Procedure}General procedure in the analyses carried out within the ETH $\pi N$ project} The general procedure, which had been followed in the past when new analyses were carried out within this project, has been laid out in Section 3.1 of Ref.~\cite{Matsinos2022b}. After one decade of application, this analysis procedure will be somewhat simplified, and (hopefully) become more straightforward and comprehensible to the (non-expert) reader. The course of analysis comprises two (largely automated) phases. As aforementioned, the first phase makes use of the ETH parameterisation of the $s$- and $p$-wave $K$-matrix elements \cite{Matsinos2022b}. This simple polynomial parameterisation provides a model-independent way of identifying the outliers, one which is devoid of theoretical constraints (other than the expected low-energy behaviour of the $K$-matrix elements). In addition, measurements are marked as outliers after a comparison with the same type of data: for instance, a decision on whether or not a measurement of the DB$_+$ is an outlier rests upon its proximity to the \emph{bulk} of the low-energy $\pi^+ p$ data. From now on, the first phase of any new analyses will comprise four steps, each one involving a different input DB. At each step, a loop \begin{equation*} A \rightleftharpoons B \, \, \, , \end{equation*} where \begin{itemize} \item[$A$] represents the operation `Fit to the DB' and \item[$B$] represents the operation `Remove from the DB the most discrepant outlier in the fit', \end{itemize} is set until all outliers are removed from the DB whose consistency is examined at that step. At the end of each cycle (one optimisation run) of each step, the p-values of the description of the datasets~\footnote{Each p-value is calculated from the contribution of the dataset to the overall $\chi^2_{\rm min}$ and the number of the active DoFs of that dataset, i.e., of the number of datapoints which currently comprise the dataset (any outliers, identified prior to the current cycle, are assumed permanently removed from the input). The absolute normalisation of each dataset is also subjected to testing (and removal, if that test fails, in which case the NDF of the dataset is the current number of its accepted datapoints reduced by one), see Ref.~\cite{Matsinos2017a} for details.}, which comprise the DB at that cycle, are compared in order that the worst-described dataset be identified. If the p-value, corresponding to the description of that dataset, is below a user-defined significance threshold $\mathrm{p}_{\rm min}$, then the worst-described entry of that dataset (corresponding to the largest contribution to the $(\chi^2_j)_{\rm min}$ value of that dataset) is removed from the DB (one outlier at a time) and the fit to the updated DB (i.e., to the former DB without the newly-marked outlier) is carried out. The loop is repeated until the p-values of all datasets in the DB, whose consistency is examined at that step, exceed $\mathrm{p}_{\rm min}$, in which case the analysis enters the next step. A few words about the choice of the significance threshold $\mathrm{p}_{\rm min}$ are in order. In the analyses which are carried out within this project, its default value is chosen to correspond to the frequency of occurrence of $2.5 \sigma$ effects in the normal distribution. This value is approximately equal to $1.24 \cdot 10^{-2}$, i.e., slightly exceeding $1.00 \cdot 10^{-2}$, which is the threshold regarded by most statisticians as the outset of statistical significance, see also Appendix \ref{App:AppC}. (To ensure the consistency of the analyses, the entire procedure used to be routinely repeated for $\mathrm{p}_{\rm min}$ values associated with $2$ and $3 \sigma$ effects in the normal distribution. Regarding the ZRH22 analysis, this (low-priority) test is pending.) After these explanations, it is time I laid out the four steps of the first phase and the two steps of the second phase of each analysis. \begin{enumerate} \item Exclusive fits to the DB$_+$ (starting from the initial low-energy DB$_+$ of Table \ref{tab:DB}) by variation of the three (one $s$-wave and two $p$-wave) $I=3/2$ partial-wave amplitudes (seven parameters in total). The $I=3/2$ partial-wave amplitudes are fixed from the final fit. \item Exclusive fits to the DB$_-$ (starting from the initial low-energy DB$_-$ of Table \ref{tab:DB}) by variation of the three $I=1/2$ partial-wave amplitudes (seven parameters in total). The final $I=3/2$ partial-wave amplitudes of step (1) are used (expectation values, no uncertainties). \item Exclusive fits to the DB$_0$ (starting from the initial low-energy DB$_0$ of Table \ref{tab:DB}) by variation of the three $I=1/2$ partial-wave amplitudes (seven parameters in total). The final $I=3/2$ partial-wave amplitudes of step (1) are (again) used (expectation values, no uncertainties). \item Joint fits to the tDB$_\pm$ by variation of all six (two $s$-wave and four $p$-wave) $I=3/2$ and $I=1/2$ partial-wave amplitudes (fourteen parameters in total). Any additional outliers from this step are removed only in the fits of step (5); in most cases, these fits produce no further outliers. \item Joint fits of the ETH model to the tDB$_\pm$. There is no identification of outliers at this step. \item Exclusive fit of the ETH model to the tDB$_0$. There is no identification of outliers at this step. \end{enumerate} As their results had never been reported/used, global fits to the data of all three low-energy $\pi N$ reactions will not be carried out henceforth. Joint fits to the combined tDB$_+$ and tDB$_0$, using the ETH parameterisation of the $s$- and $p$-wave K-matrix elements, will also not be carried out, given that the joint fits of the ETH model to the same data are now replaced by the exclusive fit of the ETH model to the tDB$_0$ (last step above). \section{\label{sec:Results}Results} To examine the possibility of biases in the analysis, all datasets (even those containing only one datapoint) must be accompanied by a normalisation uncertainty. As a result, realistic uncertainties must be assigned to the incomplete datasets, i.e., to those of the datasets whose normalisation uncertainty is unknown; the alternative would have been to simply exclude all such measurements. Normalisation uncertainties were assigned to $126$ of the $1150$ datapoints of the three initial DBs of Table \ref{tab:DB}. These data are: \begin{itemize} \item the $81$ $\pi^+ p$ BERTIN76 \cite{Bertin1976} and AULD79 \cite{Auld1979} DCSs: $8$ ($7+1$) datasets; \item the $9$ $\pi^+ p$ FRIEDMAN99 \cite{Friedman1999} PTCSs, as well as the $2$ CARTER71 \cite{Carter1971} and the $4$ PEDRONI78 \cite{Pedroni1978} TNCSs: in total, $12$ one- or two-point datasets; \item the $6$ $\pi^+ p$ and $5$ $\pi^- p$ ES SEVIOR89 \cite{Sevior1989} APs (see the last paragraph of this section): $2$ datasets; \item the $\pi^- p$ CX DUCLOS73 \cite{Duclos1973} DCSs: $3$ one-point datasets; \item the $\pi^- p$ CX SALOMON84 \cite{Salomon1984} DCSs: $2$ sets of the first $3$ coefficients in the Legendre expansion of the measured DCS (the DCSs were not reported); and \item the $\pi^- p$ CX BUGG71 \cite{Bugg1971} and BREITSCHOPF06 \cite{Breitschopf2006} TCSs: $10$ ($1+9$) one-point datasets. \end{itemize} A robust fit to the normalisation uncertainties ($T$ being the independent variable), which have been reported in the modern $\pi^+ p$ DCS experiments, using Huber's objective function (along with the default value $1.345$ for the tuning constant), led to the result: $\delta z_+ (T) = -0.532267 \cdot 10^{-3} T + 0.069248$, where $T$ is expressed in MeV. (The results of the robust fits, using Tukey's (bisquare) objective function, were nearly identical.) The $\pi^+ p$ BERTIN76 and AULD79 datasets were assigned the normalisation uncertainty of $2 \, \delta z_+ (T_j)$, where $T_j$ (in MeV) is obviously the $T$ value in each of these datasets. A robust fit (with the same objective function) to the normalisation uncertainties, reported in the $\pi^- p$ CX DCS experiments, led to the nearly flat result: $\delta z_0 (T)=-0.032710 \cdot 10^{-3} T + 0.061755$, where $T$ is again expressed in MeV. The $\pi^- p$ CX DUCLOS73 datasets were assigned the normalisation uncertainty of $2 \, \delta z_0 (T_j)$. In both cases, the use of generous uncertainties (i.e., double the fitted values at each $T=T_j$) was not meant as retribution (for the omission of the appropriate reporting of the quantity $\delta z_j$), but as a precaution: it is unlikely that due attention was paid to absolute-normalisation effects by all experimental groups in the 1970s; it is not even certain that the normalisation effects were generally recognised as potentially important sources of uncertainty (at those times). Normalisation uncertainties were assigned to the remaining incomplete datasets as follows. \begin{itemize} \item All PTCSs, TNCSs, and TCSs (FRIEDMAN99, CARTER71, PEDRONI78, BUGG71) were assigned the normalisation uncertainty of $6~\%$, i.e., double the reported uncertainty of the KRISS99 \cite{Kriss1999} PTCSs. \item The SALOMON84 measurements were assigned the normalisation uncertainty of $3.1~\%$, i.e., the normalisation uncertainty of the (similar, as well as contemporaneous) BAGHERI88 \cite{Bagheri1988} experiment. It is likely that some normalisation effects are already contained in the SALOMON84 data. \item The $\pi^- p$ CX BREITSCHOPF06 TCSs were assigned the normalisation uncertainty of $3~\%$; the experimentalists had already combined statistical and systematic effects in quadrature and reported only the total uncertainty. \end{itemize} Only one reported normalisation uncertainty was replaced: the two SEVIOR89 \cite{Sevior1989} AP datasets at $98$ MeV were assigned the normalisation uncertainty of $5~\%$, the maximal reported normalisation uncertainty in experiments which aimed at measuring the AP at low energy. On p.~2785 of Ref.~\cite{Sevior1989}, one reads: ``The uncertainty in the magnitude of the target polarization was $1.6~\%$.'' However, it is unclear from the paper whether the quoted value represents the total normalisation uncertainty in that experiment, and whether or not the uncertainties of the reported AP values (see their Table I) already contain such effects. Importantly, the target polarisation, the dominant source of normalisation uncertainty in the measurements of the AP, has been reported around $3~\%$ by all other experimental groups (which measured that quantity), even two decades after Ref.~\cite{Sevior1989} appeared. In fact, there is no other AP dataset in the three low-energy $\pi N$ DBs with normalisation uncertainty below $3~\%$. Having accepted the reported normalisation uncertainty of $1.6~\%$ for the SEVIOR89 datasets would have been unfair towards all other AP experiments at low energy. \subsection{\label{sec:Fits}Results of the fits} After the addition of the $d$- and $f$-wave contributions (taken from the phase-shift solution XP15 \cite{XP15} of the SAID group) to the $\pi N$ scattering amplitudes, obtained with either of the two modelling options of Section \ref{sec:Models}, as well as the inclusion of the EM contributions \cite{Oades2007,Gashi2001a,Gashi2001b}, the extraction of estimates for the usual low-energy $\pi N$ observables (following the long chain of equations of Section 2 of Ref.~\cite{Matsinos2006}), i.e., for the DCS, for the AP, etc., is achieved on the basis of a given vector of model-parameter values, at all permissible values of the relevant kinematical variable(s), e.g., \begin{itemize} \item of the energy $T$ and of the scattering angle $\theta$ in the centre-of-mass/centre-of-momentum (CM) coordinate system for the DCSs and APs; \item of the energy for the TNCSs and TCSs; \item of the energy and the laboratory-angle cut $\theta_0$ (half the aperture of the forward cone, whose apex coincides with the geometrical centre of the target, usually $20-30^\circ$) for the PTCSs, etc. \end{itemize} The contribution of each of the $N$ datasets of the input DB to the minimisation function is given by the Arndt-Roper formula \cite{Arndt1972}, see Eq.~(\ref{eq:EQA001}). The sum of these contributions, $\chi^2$, is a function of the parameters entering the modelling of the hadronic part of the $\pi N$ interaction. By variation of these parameters, the overall $\chi^2$ is minimised, resulting in $\chi^2_{\rm min}$. \subsubsection{\label{sec:Fits1}Step (1) of the procedure of Section \ref{sec:Procedure}} The criteria for the removal of entire datasets, on account of the number of outliers they contain, can be found in Ref.~\cite{Matsinos2017b}, p.~7. The probability of a single datapoint being an outlier, a quantity which is required in this evaluation and which is denoted by p in Ref.~\cite{Matsinos2017b}, was estimated herein to $23/472 \approx 4.87 \cdot 10^{-2}$ (at the default significance threshold $\mathrm{p}_{\rm min}$ of this project). This estimate was obtained after relaxing the condition for the removal of entire datasets and carrying out exclusive fits to the DB$_+$, which is known to contain the largest amount of outliers, until a consistent DB was obtained. Having evaluated p, it is straightforward to obtain the maximal number of outliers which a dataset can contain (which, of course, depends on the number of its initial datapoints~\footnote{As a result of the application of this procedure, a dataset with $11$ DoFs can have a maximum of $2$ outliers (at the default significance threshold $\mathrm{p}_{\rm min}$ of this project), whereas one with $20$ DoFs can have $3$.}). (If, for a dataset, the calculated maximal number of outliers did not exceed $2$, it was replaced by $2$.) The first exclusive fit to the DB$_+$ of $472$ datapoints (comprising the $465$ datapoints of the $65$ `genuine' $\pi^+ p$ datasets, as well as the $7$ $\pi^+ p$ APs of one of the datasets of Ref.~\cite{Meier2004} which contains measurements of both ES reactions, see Table \ref{tab:DB}) resulted in $\chi^2_{\rm min} \approx 956.96$ (for $465$ DoFs, as the fit involves seven parameters). By removing one DoF per cycle, as explained in Section \ref{sec:Procedure}, the tDB$_+$ of $433$ DoFs was obtained, see Table \ref{tab:ProgPIP}. Three datasets, identified as problematical already in 1997 \cite{Fettes1997}, stick out of the DB$_+$ in a rather dramatic manner: the BRACK90 dataset at $66.80$ MeV (with eleven datapoints), the BERTIN76 dataset at $67.40$ MeV (with ten datapoints), and the JORAM95 dataset at $32.70$ MeV (with seven datapoints). Of the remaining outliers, two relate to the absolute normalisation: the analysis suggested that two of the four BRACK86 datasets, which are accompanied by unrealistically small normalisation uncertainties ($1.2$ and $1.4~\%$, respectively), be freely floated~\footnote{In comparison with the bulk of the tDB$_+$, the absolute normalisation of these two datasets seems to have been underestimated by about $11~\%$ and $6~\%$, respectively, see also Table \ref{tab:DBPIP}!}. In summary, the removal of $39$ DoFs results in the reduction of the $\chi^2_{\rm min}$ value by about $392.10$, i.e., by about $10.1$ per removed DoF. The tDB$_+$ is presented in Table \ref{tab:DBPIP}. Apart from a few experimental details, the table also contains the contribution $(\chi^2_j)_{\rm min}$ of each dataset to the $\chi^2_{\rm min}$, the p-value associated with the quality of the description of each dataset in the final exclusive fit to the tDB$_+$, and the fitted value of the scale factor $z_j$. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:ProgPIP}}}The results of the examination of the consistency of the DB$_+$ via the application of the procedure of step (1) of Section \ref{sec:Procedure}. The quantities $T$ and $\theta$ stand for the pion laboratory kinetic energy and the CM scattering angle, respectively. If the $\theta$ entry is omitted, then the action applies to the entire dataset. The result `flagged' implies the removal (of the datapoint, of the dataset, or of the absolute normalisation of the dataset in question, as the case might be) at the following cycle (optimisation run). \vspace{0.25cm} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline $\chi^2_{\rm min}/{\rm NDF}$ & Identifier & $T$ (MeV) & $\theta$ (deg) & Result\\ \hline \hline $956.96/465$ & BERTIN76 & $67.40$ & $150.69$ & flagged\\ $920.88/464$ & BRACK90 & $66.80$ & $147.00$ & flagged\\ $896.84/463$ & BERTIN76 & $67.40$ & $133.31$ & flagged\\ $866.40/462$ & BRACK90 & $66.80$ & $47.60$ & flagged\\ $846.36/461$ & JORAM95 & $32.70$ & $131.28$ & flagged\\ $826.19/460$ & BRACK90 & $66.80$ & $59.00$ & flagged\\ & BRACK90 & $66.80$ & & flagged\\ $769.57/451$ & JORAM95 & $44.60$ & $30.74$ & flagged\\ $755.68/450$ & BRACK86 & $66.80$ & & absolute normalisation flagged\\ $720.34/449$ & BERTIN76 & $67.40$ & $142.09$ & flagged\\ & BERTIN76 & $67.40$ & & flagged\\ $682.69/441$ & JORAM95 & $44.60$ & $35.40$ & flagged\\ $666.35/440$ & JORAM95 & $32.70$ & $52.19$ & flagged\\ $656.74/439$ & JORAM95 & $32.20$ & $37.40$ & flagged\\ $638.66/438$ & JORAM95 & $45.10$ & $124.42$ & flagged\\ $631.41/437$ & JORAM95 & $44.60$ & $14.26$ & flagged\\ $623.45/436$ & JORAM95 & $32.70$ & $74.16$ & flagged\\ & JORAM95 & $32.70$ & & flagged\\ $604.87/431$ & BERTIN76 & $39.50$ & $75.05$ & flagged\\ $600.25/430$ & BRACK86 & $86.80$ & & absolute normalisation flagged\\ $586.83/429$ & JORAM95 & $45.10$ & $131.69$ & flagged\\ $578.96/428$ & BERTIN76 & $39.50$ & $85.81$ & flagged\\ $573.02/427$ & BERTIN76 & $95.90$ & $65.67$ & flagged\\ $564.86/426$ & & & &\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} The analysis of the fitted values of the scale factor $z$, obtained from the final exclusive fit to the tDB$_+$, is interesting for its own sake (in particular, after considering that most discrepancies in the low-energy $\pi N$ DB involve the $\pi^+ p$ reaction). A similar analysis of the corresponding values, as they had come out of the global fit of the SAID group to all $\pi N$ data in their phase-shift solution XP15 \cite{XP15}, had been carried out in Ref.~\cite{Matsinos2022a}, see Section 5.1 therein, leading to the conclusion that the normalisation uncertainties, which had been reported in the low-energy $\pi N$ experiments, had seriously been underestimated (on average), (perhaps) by as much as $45~\%$. As Table 2 therein reveals, this effect does not have the same impact on all three low-energy $\pi N$ reactions. Ideally, the distribution of the normalised (standardised) residuals \begin{equation} \label{eq:EQ003} \zeta_j \coloneqq \frac{z_j - 1}{\delta z_j} \end{equation} is the standard normal distribution $N(\mu=0,\sigma^2=1)$. This implies that the sum of the second terms on the right-hand side (rhs) of Eq.~(\ref{eq:EQA001}) \begin{equation*} \chi^2_{\rm sc} \coloneqq \sum_{j=1}^N \zeta_j^2 \, \, \, , \end{equation*} associated with the scaling contribution in each fit, follows the $\chi^2$ distribution with $N$ DoFs, where $N$ stands for the number of datasets which contribute to the scaling part of the $\chi^2_{\rm min}$; the datasets, which have lost their absolute normalisation, do not contribute. From the final exclusive fit to the tDB$_+$, $\chi^2_{\rm sc}$ came out equal to about $109.22$ for $61$ datasets, resulting in the p-value of $1.48 \cdot 10^{-4}$. The smallness of this result attests to the departure of the distribution of the quantity $\chi^2_{\rm sc}$ from the $\chi^2$ distribution, which (in turn) demonstrates the departure of the distribution of the standardised residuals $\zeta$ from the statistical expectation of the standard normal distribution $N(0,1)$. Due to two reasons, it is not straightforward to compare the result of Ref.~\cite{Matsinos2022a} with the one obtained in this work: to start with, the fitted values of the scale factor $z$ in Ref.~\cite{Matsinos2022a} emerge from a global fit to the $\pi N$ data \cite{XP15}, hence they are expected to also contain effects which originate from the theoretical constraints on the PWAs of the SAID group. On the contrary, the results of this work have been obtained from exclusive fits to the tDB$_+$ (and, consequently, are expected to be closer to the physical reality for the $\pi^+ p$ reaction). In addition, included in the SAID low-energy DB are the DCSs of the CHAOS Collaboration \cite{Denz2006}, which (as mentioned in Section \ref{sec:Database}) do not enter the DB of the ETH $\pi N$ project. In spite of these differences, some effort will next be undertaken towards a rudimentary comparison. To this end, one ought to remove first the $z_j$ values which are obtained from the measurements of the PTCS and TNCS, which (albeit part of this work) are not included in the SAID DB. One thus obtains the result: $\chi^2_{\rm sc} \approx 96.21$ for $36$ datasets, resulting in the p-value of $2.17 \cdot 10^{-7}$; the reduced $\chi^2_{\rm sc}$ value increases to about $2.67$. By absorbing this effect in a redefinition of the normalisation uncertainty $\delta z_j \to \delta z^\prime_j$, where $\delta z^\prime_j$ can be thought of as the \emph{true} normalisation uncertainty, i.e., the one which (had it been used in place of $\delta z_j$) would have led to the statistical expectation $\avg{\chi^2_{\rm sc}} = N$, one obtains: $(\delta z^\prime_j - \delta z_j) / \delta z^\prime_j \approx 38.8~\%$. This result supports the thesis of a serious underestimation (on average) of the reported values of the normalisation uncertainty in the $\pi^+ p$ experiments at low energy, and generally agrees well with the findings of Ref.~\cite{Matsinos2022a}. After focusing on the DCS measurements, one obtains from the fitted values of the scale factor $z$ a more significant result: $\chi^2_{\rm sc} \approx 90.32$ for $30$ datasets, resulting in the p-value of $5.88 \cdot 10^{-8}$, and $(\delta z^\prime_j - \delta z_j) / \delta z^\prime_j \approx 42.4~\%$. As this result is obtained from comparisons involving \emph{only} the $\pi^+ p$ data, with no theoretical constraints whatsoever (other than the expected energy dependence of the $I=3/2$ $K$-matrix elements at low energy), it lends momentum to the supposition that the reported normalisation uncertainties in the $\pi^+ p$ experiments, which measured the DCS at the three meson factories (LAMPF, PSI, and TRIUMF) for over three decades, have been overly optimistic, and ought to be corrected/adjusted to more realistic values, hopefully by the experimental groups which had been responsible for those measurements in the first place. If the term `irrefutable evidence' could be found in a Physics glossary, then this case would probably qualify for justifiable use. \subsubsection{\label{sec:Fits23}Steps (2) and (3) of the procedure of Section \ref{sec:Procedure}} Details about the optimisation procedure in case of the two $\pi^- p$ reactions, ES and CX, are provided in Tables \ref{tab:ProgPIMEL} and \ref{tab:ProgPIMCX}, respectively. In the former case, the removal of $8$ of the initial $341$ DoFs of the DB$_-$ results in the reduction of the $\chi^2_{\rm min}$ by $155.23$ (from about $523.26$ to about $368.03$), i.e., by about $19.4$ per removed DoF. As in the former PSAs of this project since 2006, the five-point BRACK90 dataset at $66.80$ MeV had to be removed. In case of the BD$_0$, the removal of the absolute normalisation of four (of the seven) FITZGERALD86 datasets, in fact of the lowest four in terms of the energy of the incoming beam, is noticeable. Although this failure might provide an acceptable argument for calling into question the validity of the absolute normalisation of all FITZGERALD86 datasets, the absolute normalisation of the remaining three datasets was retained. The removal of $5$ DoFs of the $330$ initial DoFs of the fit to the DB$_0$ results in the reduction of the $\chi^2_{\rm min}$ by $83.56$ (from about $393.56$ to about $310.00$), i.e., by about $16.7$ per removed DoF. The two tDBs, tDB$_-$ and tDB$_0$, are presented in Tables \ref{tab:DBPIMEL} and \ref{tab:DBPIMCX}, respectively. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:ProgPIMEL}}}The equivalent of Table \ref{tab:ProgPIP} when examining the consistency of the DB$_-$ (step (2) of Section \ref{sec:Procedure}). \vspace{0.25cm} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline $\chi^2_{\rm min}/{\rm NDF}$ & Identifier & $T$ (MeV) & $\theta$ (deg) & Result\\ \hline \hline $523.26/334$ & BRACK90 & $66.80$ & $70.00$ & flagged\\ $490.20/333$ & BRACK95 & $98.10$ & $36.70$ & flagged\\ $436.57/332$ & WIEDNER89 & $54.30$ & & absolute normalisation flagged\\ $413.63/331$ & BRACK90 & $66.80$ & $80.80$ & flagged\\ $394.54/330$ & WIEDNER89 & $54.30$ & $15.55$ & flagged\\ $379.50/329$ & BRACK90 & $66.80$ & $111.00$ & flagged\\ & BRACK90 & $66.80$ & & flagged\\ $368.03/326$ & & & &\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:ProgPIMCX}}}The equivalent of Table \ref{tab:ProgPIP} when examining the consistency of the DB$_0$ (step (3) of Section \ref{sec:Procedure}). \vspace{0.25cm} \begin{center} \begin{tabular}{|c|c|c|c|c|} \hline $\chi^2_{\rm min}/{\rm NDF}$ & Identifier & $T$ (MeV) & $\theta$ (deg) & Result\\ \hline \hline $393.56/330$ & FITZGERALD86 & $40.26$ & & absolute normalisation flagged\\ $369.54/329$ & FITZGERALD86 & $36.11$ & & absolute normalisation flagged\\ $346.64/328$ & FITZGERALD86 & $32.48$ & & absolute normalisation flagged\\ $329.71/327$ & BREITSCHOPF06 & $75.10$ & & flagged\\ $321.19/326$ & FITZGERALD86 & $47.93$ & & absolute normalisation flagged\\ $310.00/325$ & & & &\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} Importantly, the two analyses of the fitted values of the scale factor $z$, obtained from the final exclusive fits to the tDB$_-$ and to the tDB$_0$, revealed no significant effects in the determination of the normalisation uncertainties of the datasets for these two reactions. Therefore, the effects, which are seen in Table 2 of Ref.~\cite{Matsinos2022a} for the two low-energy $\pi^- p$ reactions, are (in all likelihood) attributable to the practice followed by the SAID group, namely to carry out \emph{global} fits to all $\pi N$ measurements. \subsubsection{\label{sec:Fits4}Step (4) of the procedure of Section \ref{sec:Procedure}} Prior to submitting the tDBs to further analysis, the joint fit of step (4) of Section \ref{sec:Procedure} was carried out, resulting in $\chi^2_{\rm min} \approx 920.46$ for $752$ DoFs and no additional outliers. Interestingly, the $\chi^2_{\rm min}$ value of this fit is just short of the sum of the $\chi^2_{\rm min}$ values of the two exclusive fits to the same data, as reported in the last two sections. This demonstrates that the two ES reactions essentially fix different partial-wave amplitudes: the $I=3/2$ amplitudes are (largely) determined from the $\pi^+ p$ data, whereas the measurements of the $\pi^- p$ ES reaction fix the $I=1/2$ amplitudes. To summarise, the tDB$_\pm$ and the tDB$_0$, comprising the datasets of Tables \ref{tab:DBPIP}-\ref{tab:DBPIMCX}, may be submitted to further analysis (using the ETH model). A total of $52$ entries (of the initial $1150$ DoFs of Table \ref{tab:DB}) were identified as outliers in this work, corresponding to about $4.5~\%$ of the data. \subsubsection{\label{sec:Fits5}Step (5) of the procedure of Section \ref{sec:Procedure}} The modelling of the $s$- and $p$-wave $K$-matrix elements by means of simple polynomials enables tests of the consistency of the DB and serves as an unbiased method for the identification of the outliers. However, neither does it provide insight into the underlying physical processes, nor can it easily incorporate the important theoretical constraint of crossing symmetry. To this end, the ETH model is employed at the second stage of each new analysis. The PSA of the tDB$_\pm$ encompasses the results of one hundred joint fits of the ETH model to the data. Each of these fits was carried out at one $m_\sigma$ value, randomly generated in normal distribution, see Section \ref{sec:Method}: in these fits, $\chi^2_{\rm min}$ varied between about $991.52$ and $1008.99$, in correlation with $m_\sigma$. For the median value of the $m_\sigma$ distribution in Ref.~\cite{Matsinos2020b}, which is equal to about $497$ MeV, one obtains $\chi^2_{\rm min} \approx 1002.36$ for $759$ DoFs in the fit. The scaling contribution to the $\chi^2_{\rm min}$ value is equal to about $149.12$ for the $97$ datasets ($=61+37-1$, as the set of AP measurements, which contains data of both ES reactions \cite{Meier2004}, is treated as one dataset in the joint fits to the tDB$_\pm$) whose absolute normalisation has not been removed during the examination of the consistency of the input DBs, as laid out at steps (1), (2), and (4) of Section \ref{sec:Procedure}. The reduced $\chi^2_{\rm min}$ value of the joint fit using the median $m_\sigma$ value is about $1.32$ and the ensuing p-value is small, about $6.07 \cdot 10^{-9}$. In the strict statistical sense, the fit is unacceptable, though (as the ratio $\chi^2_{\rm min}/{\rm NDF}$ comes out `close' to $1$) most physicists would rather consider it `reasonable' or even `fairly good'. Be that as it may, a significant fraction of these $\chi^2_{\rm min}$ values reflects the serious underestimation of the normalisation uncertainties of the $\pi^+ p$ datasets, see Section \ref{sec:Fits1}. At the end of the day, the analyst (who is resolute in extracting some information from the $\pi N$ measurements) is left with no alternative, but to accept the situation - however disagreeable, and move on by simply correcting the fitted uncertainties via the application of the Birge factor, which (in case of the joint fit using the median $m_\sigma$ value) comes out equal to about $1.15$. The increase in the $\chi^2_{\rm min}$ values between the final joint fit of step (4) and those obtained at this step is accounted for by two effects. First, the polynomial parameterisation of the two $s$- and the four $p$-wave $K$-matrix elements implies the use of fourteen parameters in total, whereas the joint fit of the ETH model to the same data uses seven. Without doubt, more room is left to the measurements (to accommodate themselves closer to the fitted values) in the former case. Second, imposed in case of the joint fits of the ETH model to the tDB$_\pm$ is the theoretical constraint of crossing symmetry, which the partial-wave amplitudes of the ETH model obey. The results of each joint fit for the model parameters, as well as the corresponding Hessian matrices (all uploaded as ancillary material), enable the extraction of estimates for the model parameters (Section \ref{sec:PSAETHPar}), for the $\pi N$ phase shifts and scattering amplitudes (Sections \ref{sec:PhaseShifts} and \ref{sec:TriangleIdentity}), for the LECs of the $\pi N$ interaction (Section \ref{sec:LECs}), and for the usual low-energy $\pi N$ observables. For the sake of brevity, each baseline solution (BLS), see Appendix \ref{App:AppC}, obtained from the PSA of the tDB$_\pm$, will be labelled BLS$_\pm$ in this study. \subsubsection{\label{sec:Fits6}Step (6) of the procedure of Section \ref{sec:Procedure}} As explained in Section \ref{sec:Models}, the contributions from the scalar-isoscalar $t$-channel Feynman diagram (left graph of the upper part of Fig.~\ref{fig:GraphsETH}) to the $\pi^- p$ CX scattering amplitude are suppressed in case of the exclusive fit of the ETH model to the tDB$_0$. This is simply achieved by setting $G_\sigma=0$ GeV$^{-2}$; in that case, the value of the model parameter $\kappa_\sigma$ is irrelevant. The $\chi^2_{\rm min}$ value of this fit comes out equal to about $337.32$ for $327$ DoFs, suggesting a satisfactory optimisation of the description of the tDB$_0$ (p-value $\approx 3.35 \cdot 10^{-1}$). The scaling contribution $\chi^2_{\rm sc}$ to $\chi^2_{\rm min}$ is equal to about $43.07$ for $51$ DoFs. Although there is no indication of any problems in this fit, it must be borne in mind that the model parameters are expected to be effective as (if one blames the violation of isospin invariance in the low-energy $\pi N$ interaction on the $\pi^- p$ CX reaction, e.g., via the QM admixture of the $\pi^0$ and the $\eta$ mesons \cite{Cutkosky1979}) they contain the effects of the Feynman diagrams of Fig.~\ref{fig:EtaPi0}. \subsubsection{\label{sec:PSAETHPar}Optimal values of the parameters of the ETH model} The optimal values of the seven parameters of the ETH model, obtained from the PSA of the tDB$_\pm$, as well as those of the five parameters, obtained from the exclusive fit to the tDB$_0$, are given in Table \ref{tab:ParETH}. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:ParETH}}}The optimal values of the seven parameters of the ETH model, obtained from the PSA of the tDB$_\pm$, as well as those corresponding to the exclusive fit to the tDB$_0$. To facilitate the comparison with other works, the estimate for the pseudoscalar coupling constant $g_{\pi N N}$ is converted into a result for the pseudovector coupling $f_{\pi N N}$, which (in case of the joint fits to the tDB$_\pm$) can be identified with the charged-pion coupling constant $f_c$, see Ref.~\cite{Matsinos2019}. As they presumably contain the effects of the Feynman diagrams of Fig.~\ref{fig:EtaPi0}, the parameters of the ETH model ought to be thought of as effective (and potentially unrealistic, as the case appears to be for the coupling constant $g_{\pi N N}$) in the five-parameter fit of the ETH model to the tDB$_0$. \vspace{0.25cm} \begin{center} \begin{tabular}{|l|c|c|} \hline & tDB$_\pm$ & tDB$_0$\\ \hline \hline $G_\sigma$ (GeV$^{-2}$) & $24.56(45)$ & $0$\\ $\kappa_\sigma$ & $0.066(45)$ & $-$\\ $G_\rho$ (GeV$^{-2}$) & $51.47(60)$ & $57.53(35)$\\ $\kappa_\rho$ & $1.32(37)$ & $2.72(38)$\\ $g_{\pi N N}$ & $13.18(13)$ & $14.28(15)$\\ $g_{\pi N \Delta}$ & $28.94(29)$ & $26.96(55)$\\ $Z$ & $-0.481(73)$ & $-0.26(11)$\\ \hline $f_{\pi N N}^2$ & $0.0764(15)$ & $0.0896(18)$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} The differences between the two sets of fitted results are significant for all model parameters, in particular for $G_\rho$ (the difference is equivalent to about a $8.7 \sigma$ effect in the normal distribution) and for $g_{\pi N N}$ (the difference is equivalent to about a $5.7 \sigma$ effect in the normal distribution). The fitted result for $\kappa_\sigma$ from the PSA of the tDB$_\pm$ is small and compatible with $0$. Both $\kappa_\rho$ values are small, well below the corresponding results extracted at the $\rho(770)$-meson pole via dispersion relations (for details, see Ref.~\cite{Matsinos2018} and the works cited therein). The fitted result for the coupling constant $\pi N \Delta$ from the optimisation of the description of the tDB$_\pm$ is (and has always been) in good agreement with the value of $28.93(39)$, extracted directly from the decay width of the $\Delta(1232)$ resonance, see footnote 10 of Ref.~\cite{Matsinos2014}. This agreement justifies the approach, which was put forward already in 1994 \cite{Goudsmit1994}, to determine each $\pi N R$ coupling constant (needed in order to fix the contributions from each HBR $R$ to the partial-wave amplitudes of the ETH model) from the partial decay width (to $\pi N$ decay modes) of each such resonance. As in the former PSAs in this project, the fitted result for the model parameter $Z$ (which is associated with the spin-$1/2$ admixture in the $\Delta(1232)$-resonance propagator) from the optimisation of the description of the tDB$_\pm$ comes out compatible with $Z=-1/2$, which had been one of the popular theoretical preferences in the remote past; albeit somewhat less negative, the corresponding fitted result from the exclusive fit to the tDB$_0$ is not incompatible with that preference. \subsection{\label{sec:Tests}Consistency checks} \subsubsection{\label{sec:ScaleFactors}Analysis of the fitted values of the scale factor $z$} Before extracting predictions from the two types of fits of the ETH model to the low-energy $\pi N$ measurements, the results of these fits must be subjected to a few consistency checks. To start with, when the Arndt-Roper formula \cite{Arndt1972} is used in the optimisation, the statistical expectation is that the datasets which are scaled `upwards' ($z_j<1$) balance (on average) those which are scaled `downwards' ($z_j>1$). Furthermore, the energy dependence of the fitted values of the scale factor $z$ must not be significant. It was demonstrated several times in the past that the fulfilment of these conditions should not only involve the entire set of the fitted values of the scale factor $z$ in each fit, but also those of subsets (random or selected in compliance with the basic principles of Sampling Theory) of each fitted DB. As a result, two additional (i.e., on top of the two tests, which involve the entire input tDBs in the two types of fits of the ETH model, i.e., the tDB$_\pm$ and the tDB$_0$) tests are suggested by the compartmental structure of the tDB$_\pm$: in short, it must be verified that the fitted values of the scale factor $z$, corresponding to the two distinct subsets of the tDB$_\pm$, i.e., to the tDB$_+$ and the tDB$_-$, are centred on $1$ and exhibit no significant energy dependence. For both the $\pi^+ p$ (see Fig.~\ref{fig:sfPIP}) and $\pi^- p$ ES (see Fig.~\ref{fig:sfPIMEL}) datasets, the $z_j$ values above and below $1$ roughly balance, and their energy dependence is insignificant, see also Table \ref{tab:FitsToZ}. The weighted linear least-squares fit ($T$ being the independent variable) to the fitted values of the scale factor $z$ for the $\pi^+ p$ reaction results in the intercept of $1.011(20)$ and the slope of $(-0.9 \pm 2.6) \cdot 10^{-4}$ MeV$^{-1}$. A similar fit to the scale factors of the $\pi^- p$ ES reaction results in the intercept of $1.0085(54)$ and the slope of $-1.51(93) \cdot 10^{-4}$ MeV$^{-1}$. (Both uncertainties are substantially smaller in the latter case because of the inclusion in the DB$_-$ of the two $a_{\rm cc}$ estimates from pionic hydrogen, which act as `anchor points' in the optimisation; it is unfortunate that there is no corresponding observable for the $\pi^+ p$ reaction, which could be conducive to a measurement at similar precision.) In both cases, the departure from the statistical expectation for an unbiased outcome of the optimisation (intercept $1$ and vanishing slope) is not significant (at the default significance threshold $\mathrm{p}_{\rm min}$ of this project). A weighted linear least-squares fit to the $z_j$ values, obtained from the exclusive fit of the ETH model to the tDB$_0$ (see Fig.~\ref{fig:sfPIMCX}), was also carried out. Although the departure from the ideal optimisation is not significant, some effects (in particular in the slope) are noticeable. However, these effects are due to a slight mismatch between \begin{itemize} \item the extrapolated result of the extracted (in the scattering region) $\pi^- p$ CX amplitude to the $\pi N$ threshold ($T = 0$ MeV) and \item the two $a_{\rm c0}$ values, obtained directly at the $\pi N$ threshold from the PSI measurements on pionic hydrogen: \end{itemize} after removing the two fitted values of the scale factor $z$, corresponding to the two $a_{\rm c0}$ results, the intercept of $1.005(16)$ and the slope of $(-1.9 \pm 2.6) \cdot 10^{-4}$ MeV$^{-1}$ were obtained from the weighted linear least-squares fit, as well as $\chi^2_{\rm min} \approx 31.01$ for $47$ DoFs. Unlike the joint analyses of the tDB$_+$ and the tDB$_0$ \cite{Matsinos2017b}, as well as the results of the global fits to all data by the SAID group \cite{Matsinos2022a,Matsinos2017a}, no pronounced effects in the energy dependence of the fitted values of the scale factor $z$ for the $\pi^+ p$ and $\pi^- p$ CX reactions have been observed in this study. In this respect, the exclusive fit of the ETH model to the tDB$_0$ seems to be satisfactory. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:FitsToZ}}}The fitted values of the parameters of the weighted linear least-squares fit to the data shown in Figs.~\ref{fig:sfPIP}-\ref{fig:sfPIMCX}, as well as the fitted uncertainties, corrected via the application of the Birge factor $\sqrt{\chi^2_{\rm min}/{\rm NDF}}$ (if exceeding $1$). Also quoted are the $\chi^2_{\rm min}$ values of each linear fit, along with the corresponding NDF. \vspace{0.25cm} \begin{center} \begin{tabular}{|l|c|c|c|} \hline Reaction & Intercept & Slope ($10^{-4}$ MeV$^{-1}$) & $\chi^2_{\rm min}$/NDF \\ \hline \hline \multicolumn{4}{|c|}{Results from the joint fits of the ETH model to the tDB$_\pm$}\\ \hline Both reactions & $1.0090 \pm 0.0057$ & $-1.05 \pm 0.86$ & $131.34/92$\\ $\pi^+ p$ & $1.011 \pm 0.020$ & $-0.9 \pm 2.6$ & $90.79/56$\\ $\pi^- p$ ES & $1.0085 \pm 0.0054$ & $-1.51 \pm 0.93$ & $39.17/34$\\ \hline \multicolumn{4}{|c|}{Results from the exclusive fit of the ETH model to the tDB$_0$}\\ \hline $\pi^- p$ CX & $1.017 \pm 0.010$ & $-3.8 \pm 1.8$ & $32.04/49$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} \subsubsection{\label{sec:Residuals}Analysis of the fitted values of the standardised residuals $\zeta_j$ of Eq.~(\ref{eq:EQ003})} When the Arndt-Roper formula \cite{Arndt1972} is used in the optimisation, it is expected that the fitted values of the standardised residuals $\zeta_j$ of Eq.~(\ref{eq:EQ003}) follow the standard normal distribution $N(0,1)$. In this section, the sets of the extracted $\zeta_j$ values will be tested for normality. A number of algorithms have been put forward for testing the normality of a given distribution (e.g., see Refs.~\cite{Shapiro2015,DAgostino1990} and the works cited therein), including some which are based on fits to histograms, the physicist's delight, which (given the obvious dependence of the results on the selected bin size, as well as the loss of information due to the data binning) are not even considered in most `power studies'. In this work, the normality of the $\zeta_j$ distributions will be tested by means of two well-established (and popular) statistical methods. \begin{itemize} \item The formal Shapiro-Wilk normality test, which emerges as the ultimate test for normality in many power studies (e.g., see Ref.~\cite{MohdRazali2011}), introduced in 1965 \cite{Shapiro1965} for small samples (in the first version of the test, a maximum of $50$ observations could be tested) and extended in a series of studies by Royston \cite{Royston1982-1995}, to enable its application to large samples (certainly up to $n=5000$ observations, perhaps to even larger samples). \item D'Agostino's (or the D'Agostino-Pearson) $K^2$ test, which was introduced in 1973 \cite{DAgostino1973} and appeared in its current form in 1990 \cite{DAgostino1990}. \end{itemize} There is only one commonality between these two tests, the obvious one: both result in an estimate for the p-value for the acceptance of the null hypothesis (that the underlying distribution is the normal distribution $N(\mu,\sigma^2)$). To extract a p-value for each input set of $n$ observations (sample data), the former test makes use of the $W$-statistic, which essentially represents an approximate measure of the linear correlation between two quantities plotted against one another: the ordered sample-data quantiles (on one axis) and the standard normal quantiles (on the other); this plot is generally known as `normal probability plot' (which is the Q-Q plot for the normal distribution). If the sample data comprises independent observations, which have been sampled from one normal distribution $N(\mu,\sigma^2)$, then the points on the normal probability plot are arranged along a straight line. The departure of this scatter plot from linearity is equivalent to the departure of the sample data from the normal distribution. The $W$-statistic is valued in $[0,1]$: the upper bound represents ideal normality (of the sample data), whereas $W=0$ the exact opposite. On the other hand, the D'Agostino-Pearson $K^2$ test compares the sample skewness and excess kurtosis (i.e., the difference between the evaluated kurtosis and $3$, which is the kurtosis of the normal distribution) to their expectation values (for the dimension $n$ of the sample-data array) and combines two relevant tests into one `omnibus' test, featuring the $K^2$-statistic: if the underlying distribution is the normal distribution $N(\mu,\sigma^2)$, then $K^2$ is $\chi^2$-distributed with $2$ DoFs. The authors of Ref.~\cite{DAgostino1990} explain: ``By \emph{omnibus}, we mean it [i.e., the test] is able to detect deviations from normality due to either skewness or kurtosis.'' The distribution of the $97$ standardised residuals $\zeta_j$ of the joint fits of the ETH model to the tDB$_\pm$ is somewhat negatively-skewed; the skewness of the distribution varies little across the one hundred joint fits, between about $-0.1332$ and $-0.1116$. The kurtosis of the distribution (which varies between about $2.5805$ and $2.6799$) is below the expectation value of $3 (n-1) / (n+1) \approx 2.938776$ \cite{Pearson1931}. However, these effects are insignificant at the default significance threshold $\mathrm{p}_{\rm min}$ of this project: the p-values of the Shapiro-Wilk normality test vary between about $2.84 \cdot 10^{-1}$ and $5.10 \cdot 10^{-1}$. The D'Agostino-Pearson $K^2$ test results in even higher p-values, between about $6.24 \cdot 10^{-1}$ and $7.95 \cdot 10^{-1}$. The distribution of the standardised residuals is slightly translated to the right (the arithmetic mean $\mu$ of the distribution varies little in the joint fits, between about $0.2107$ and $0.2143$), and it is broader than the standard normal distribution $N(0,1)$ (the variance $\sigma^2$ of the distribution varies between about $1.4967$ and $1.5747$ in the fits). The normal probability plot in case of the standardised residuals $\zeta_j$ of the joint fit of the ETH model to the tDB$_\pm$ for the median of the $m_\sigma$ distribution (see Section \ref{sec:Fits5}) is shown in Fig.~\ref{fig:NPPES}. \begin{figure} \begin{center} \includegraphics [width=15.5cm] {NormalProbabilityPlotES.eps} \caption{\label{fig:NPPES}The normal probability plot in case of the standardised residuals $\zeta_j$ of the joint fit of the ETH model to the tDB$_\pm$ for the median of the $m_\sigma$ distribution, see Section \ref{sec:Fits5}.} \vspace{0.5cm} \end{center} \end{figure} On the other hand, the distribution of the $51$ standardised residuals $\zeta_j$ of the exclusive fit of the ETH model to the tDB$_0$ is somewhat positively-skewed; the skewness of the distribution comes out equal to about $0.2132$. The kurtosis of the distribution (which is equal to about $2.3800$) is again below the expectation value of $3 (n-1) / (n+1) \approx 2.884615$ \cite{Pearson1931}. Once again, these effects are insignificant at the default significance threshold $\mathrm{p}_{\rm min}$ of this project: the p-value of the Shapiro-Wilk normality test comes out equal to $4.27 \cdot 10^{-1}$. The D'Agostino-Pearson $K^2$ test results in a higher p-value, about $5.12 \cdot 10^{-1}$. The distribution of the standardised residuals is slightly translated to the left (the arithmetic mean $\mu$ of the distribution is equal to about $-0.1517$), and it is slightly narrower than the standard normal distribution $N(0,1)$ (the variance $\sigma^2$ of the distribution comes out equal to $0.8215$). The normal probability plot in case of the standardised residuals $\zeta_j$ of the exclusive fit of the ETH model to the tDB$_0$ is shown in Fig.~\ref{fig:NPPCX}. \begin{figure} \begin{center} \includegraphics [width=15.5cm] {NormalProbabilityPlotCX.eps} \caption{\label{fig:NPPCX}The normal probability plot in case of the standardised residuals $\zeta_j$ of the exclusive fit of the ETH model to the tDB$_0$.} \vspace{0.5cm} \end{center} \end{figure} At the end of the day, there is no significant evidence of non-normality in the results of the tests of the fitted values of the standardised residuals $\zeta_j$. It must be borne in mind that the results of this section correspond to tests for \emph{normality}, i.e., whether or not the standardised residuals $\zeta_j$ have been sampled from one normal distribution $N(\mu,\sigma^2)$, without regard to the $\mu$ and $\sigma^2$ values. Ideally, these results follow the standard normal distribution $N(0,1)$, which - predominantly due to the effects in the tDB$_+$, addressed at the end of Section \ref{sec:Fits1} - can hardly be the case in the joint fits of the ETH model to the tDB$_\pm$. \subsection{\label{sec:Predictions}Predictions} As explained in Section \ref{sec:Models}, the analysis of the tDB$_0$ using the ETH model is technically not a PSA, in that the two isospin amplitudes $f_3$ and $f_1$ cannot be determined from that analysis reliably; only their difference can. Evidently, any predictions (for the various quantities entering the low-energy $\pi N$ interaction), obtained from such an analysis, make sense only if they relate to quantities which are associated with the $\pi^- p$ CX reaction: for instance, the isovector scattering length $b_1$, the isovector scattering volumes $c_1$ and $d_1$, and the $\pi^- p$ scattering amplitude $f^{\rm extr}_{\rm CX}$. To be able to carry out a PSA of the tDB$_0$, one must also involve data which can lead to the reliable determination of (at least) one of the two isospin amplitudes. This is why joint fits to the tDB$_+$ and the tDB$_0$ were pursued in the works of the recent past. Evidently, the PSA of the tDB$_0$ was forfeited in this study, in order to enable the `clean' determination of the $\pi^- p$ scattering amplitude $f^{\rm extr}_{\rm CX}$, i.e., without any involvement of measurements other than those contained in the tDB$_0$. \subsubsection{\label{sec:PhaseShifts}$\pi N$ phase shifts} The predictions for the $s$- and $p$-wave phase shifts, obtained from the PSA of the tDB$_\pm$, are given in Table \ref{tab:PSAPhSh}; they are also shown in Figs.~\ref{fig:S31}-\ref{fig:P11}, along with the XP15 solution \cite{XP15}, including (wherever available) the five single-energy values of that solution for $T \leq 100$ MeV. Although the XP15 phase shifts are not accompanied by uncertainties~\footnote{Due to the overconstrained fits, not even tentative/working uncertainties are published in the phase-shift solutions obtained from analyses using dispersion relations. Only the single-energy phase shifts of the XP15 solution (wherever available) are accompanied by realistic uncertainties.}, there can be no doubt that they do not match well the phase-shift solution of this work: the two sets of values remain well apart from one another in the low-energy region, converging only in the vicinity of the upper $T$ bound of this project. Of course, the differences between the two solutions are aggravated by the absence of (meaningful) uncertainties on the part of the XP15 solution. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:PSAPhSh}}}The values of the two $s$- and the four $p$-wave phase shifts (in degrees), obtained in this study from the PSA of the tDB$_\pm$. \vspace{0.25cm} \begin{center} \begin{tabular}{|c|c|c|c|c|c|c|} \hline $T$ (MeV) & $\delta_{0+}^{3/2}$ ($S_{31}$) & $\delta_{0+}^{1/2}$ ($S_{11}$) & $\delta_{1+}^{3/2}$ ($P_{33}$) & $\delta_{1-}^{3/2}$ ($P_{31}$) & $\delta_{1+}^{1/2}$ ($P_{13}$) & $\delta_{1-}^{1/2}$ ($P_{11}$)\\ \hline \hline $20$ & $-2.366(32)$ & $4.300(19)$ & $1.308(10)$ & $-0.2304(51)$ & $-0.1644(43)$ & $-0.3781(78)$\\ $25$ & $-2.769(34)$ & $4.784(21)$ & $1.855(14)$ & $-0.3174(72)$ & $-0.2232(61)$ & $-0.498(11)$\\ $30$ & $-3.168(36)$ & $5.212(24)$ & $2.478(17)$ & $-0.4115(96)$ & $-0.2852(80)$ & $-0.615(13)$\\ $35$ & $-3.567(37)$ & $5.596(26)$ & $3.177(20)$ & $-0.512(12)$ & $-0.349(10)$ & $-0.728(16)$\\ $40$ & $-3.969(38)$ & $5.944(29)$ & $3.953(22)$ & $-0.617(15)$ & $-0.415(13)$ & $-0.833(19)$\\ $45$ & $-4.374(38)$ & $6.259(31)$ & $4.810(25)$ & $-0.728(18)$ & $-0.483(15)$ & $-0.930(23)$\\ $50$ & $-4.783(39)$ & $6.548(34)$ & $5.751(26)$ & $-0.843(22)$ & $-0.550(18)$ & $-1.016(26)$\\ $55$ & $-5.197(39)$ & $6.811(38)$ & $6.783(28)$ & $-0.961(25)$ & $-0.619(21)$ & $-1.090(29)$\\ $60$ & $-5.615(40)$ & $7.052(41)$ & $7.910(29)$ & $-1.083(29)$ & $-0.687(24)$ & $-1.152(33)$\\ $65$ & $-6.038(42)$ & $7.272(45)$ & $9.139(30)$ & $-1.209(33)$ & $-0.755(27)$ & $-1.200(37)$\\ $70$ & $-6.465(44)$ & $7.473(49)$ & $10.480(32)$ & $-1.337(38)$ & $-0.823(30)$ & $-1.233(41)$\\ $75$ & $-6.898(48)$ & $7.656(53)$ & $11.941(35)$ & $-1.469(42)$ & $-0.891(34)$ & $-1.252(45)$\\ $80$ & $-7.334(52)$ & $7.821(58)$ & $13.532(41)$ & $-1.602(47)$ & $-0.958(38)$ & $-1.254(50)$\\ $85$ & $-7.776(58)$ & $7.970(62)$ & $15.264(50)$ & $-1.739(52)$ & $-1.024(42)$ & $-1.240(54)$\\ $90$ & $-8.221(65)$ & $8.103(67)$ & $17.150(62)$ & $-1.878(58)$ & $-1.090(46)$ & $-1.209(59)$\\ $95$ & $-8.670(74)$ & $8.221(72)$ & $19.203(77)$ & $-2.018(64)$ & $-1.154(51)$ & $-1.160(64)$\\ $100$ & $-9.123(83)$ & $8.324(77)$ & $21.437(96)$ & $-2.161(70)$ & $-1.218(55)$ & $-1.093(70)$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} \subsubsection{\label{sec:LECs}Low-energy constants of the $\pi N$ interaction} At low energy, the hadronic part of the $\pi N$ scattering amplitude $\mathscr{F} (\vec{q}^{\, \prime}, \vec{q})$ may safely be confined to $s$- and $p$-wave contributions, see Ref.~\cite{Ericson1988}, pp.~17--18. Introducing the isospin of the pion as $\vec{t}$ and that of the nucleon as $\vec{\tau}/2$ (and using natural units for the sake of brevity), $\mathscr{F} (\vec{q}^{\, \prime}, \vec{q})$ may be put in the form: \begin{equation} \label{eq:EQ004} \mathscr{F} (\vec{q}^{\, \prime}, \vec{q}) = b_0 + b_1 \, \vec{\tau} \cdot \vec{t} + \left( c_0 + c_1 \, \vec{\tau} \cdot \vec{t} \, \right) \vec{q}^{\, \prime} \cdot \vec{q} + i \left( d_0 + d_1 \, \vec{\tau} \cdot \vec{t} \, \right) \vec\sigma \cdot (\vec{q}^{\, \prime} \times \vec{q}) \, , \end{equation} where $\vec\sigma$ is (double) the spin of the nucleon; $\vec{q}$ and $\vec{q}^{\, \prime}$ are the CM $3$-momenta of the incoming and outgoing pions, respectively. The third term on the rhs of this equation is the no-spin-flip $p$-wave part of the $\pi N$ scattering amplitude, whereas the fourth (i.e., the last) term is the spin-flip part. Equation (\ref{eq:EQ004}) defines the isoscalar (subscript $0$) and isovector (subscript $1$) $s$-wave scattering lengths ($b_0$ and $b_1$) and $p$-wave scattering volumes ($c_0$, $c_1$, $d_0$, and $d_1$), which are related to the usual spin-isospin scattering lengths/volumes $a_{l\pm}^I$ according to the transformations, e.g., see Eqs.~(2.39) of Ref.~\cite{Ericson1988}. \begin{align} \label{eq:EQ005} b_0 & \coloneqq \left( 2 a_{0+}^{3/2} + a_{0+}^{1/2} \right) / 3\nonumber\\ b_1 & \coloneqq \left( a_{0+}^{3/2} - a_{0+}^{1/2} \right) / 3\nonumber\\ c_0 & \coloneqq \left( 4 a_{1+}^{3/2} + 2 a_{1-}^{3/2} + 2 a_{1+}^{1/2} + a_{1-}^{1/2} \right) / 3\nonumber\\ c_1 & \coloneqq \left( 2 a_{1+}^{3/2} + a_{1-}^{3/2} - 2 a_{1+}^{1/2} - a_{1-}^{1/2} \right) / 3\nonumber\\ d_0 & \coloneqq \left( - 2 a_{1+}^{3/2} + 2 a_{1-}^{3/2} - a_{1+}^{1/2} + a_{1-}^{1/2} \right) / 3\nonumber\\ d_1 & \coloneqq \left( -a_{1+}^{3/2} + a_{1-}^{3/2} + a_{1+}^{1/2} - a_{1-}^{1/2} \right) / 3 \end{align} Table \ref{tab:PSACnts} contains predictions for these quantities, obtained from the PSA of the tDB$_\pm$. A few remarks are in order. \begin{itemize} \item There is no doubt that the inclusion in the DB$_-$ (for the first time, in the ZRH19 PSA \cite{Matsinos2017a}) of the two scattering lengths $a_{\rm cc}$, extracted from the PSI measurements of the strong shift of the ground state in pionic hydrogen (see Table \ref{tab:AccAndAc0}), leads to solutions with \emph{enhanced isoscalar components} in the low-energy $\pi N$ amplitude. Prior to this inclusion, the isoscalar scattering length $b_0$ used to come out small, close to (and not incompatible with) $0$. There are two consequences of this `new' reality. First, the $\pi^+ p$ scattering length $a_{\pi^+ p} = b_0 + b_1$ tends to come out slightly less negative (in comparison with former PSAs). The second consequence concerns the estimates for the $\pi N$ $\Sigma$ term: although the prediction, obtained in this work on the basis of the ETH model ($\Sigma = 81.0 \pm 1.7$ MeV), is in good agreement with the estimate obtained by the SAID group in 2002 \cite{Pavan2002}, it significantly exceeds the result of Ref.~\cite{Matsinos2014}: the difference between the two values is (nearly entirely) due to the aforementioned `enhancement' of the DB$_-$ in 2019. (Incidentally, Olsson set forth a pioneering scheme for evaluating the $\pi N$ $\Sigma$ term over two decades ago, using only a few LECs of the low-energy $\pi N$ interaction \cite{Olsson2000}, see also Section 4.2.2 of Ref.~\cite{Matsinos2014}. Using Eq.~(63) therein, one obtains from the LECs of Table \ref{tab:PSACnts}: $\Sigma = 77.3(1.2)(1.7)$ MeV. The estimate of this work for the isoscalar effective range $C^+$, which is needed in the aforementioned evaluation, is $-0.1208(24)~m^{-3}_c$.) \item The estimate for $a_{\pi^- p}$, fully corrected (for the scale-factor effects of Section \ref{sec:ScaleFactors}), is in good agreement with the experimental results of Refs.~\cite{Schroeder2001,Hennebach2014}. Of course, this is hardly surprising, given that (owing to their unprecedented precision) the two `experimental' $a_{\rm cc}$ results act as `anchor points' in the optimisation, forcing the solution to `gravitate' towards their direction. \item The prediction for the scattering length of the $\pi^- p$ CX reaction (i.e., for the quantity $a_{CX}$ in Table \ref{tab:PSACnts}), obtained from the PSA of the tDB$_\pm$, is significantly less negative than the estimates of Table \ref{tab:AccAndAc0}, obtained from the PSI measurements of the total decay width of the ground state in pionic hydrogen \cite{Hirtl2021,Schroeder2001}. This result is neither new nor surprising: as a matter of fact, it is one manifestation of what is outlined in Ref.~\cite{Matsinos2022a} as the `low-energy $\pi N$ enigma'. \end{itemize} \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:PSACnts}}}Upper part: The isoscalar and isovector $s$-wave scattering lengths (in $m_c^{-1}$ and in fm) and $p$-wave scattering volumes (in $m_c^{-3}$ and in fm$^3$), obtained from the PSA of the tDB$_\pm$; these quantities are defined by Eq.~(\ref{eq:EQ004}). Middle part: the corresponding results for the spin-isospin quantities. Lower part: the predictions for the fully-corrected (for the scale-factor effects of Section \ref{sec:ScaleFactors}) scattering lengths which are associated with the three low-energy $\pi N$ processes (i.e., with the $\pi^+ p$, $\pi^- p$ ES, and $\pi^- p$ CX reactions). \vspace{0.25cm} \begin{center} \begin{tabular}{|c|c|c|} \hline Scattering length/volume & In (reciprocal) powers of $m_c$ & In powers of fm\\ \hline \hline \multicolumn{3}{|c|}{Isoscalar-isovector scattering lengths and volumes}\\ \hline $b_0$ & $0.00629(79)$ & $0.0089(11)$\\ $b_1$ & $-0.07851(63)$ & $-0.11099(90)$\\ $c_0$ & $0.2101(25)$ & $0.5936(71)$\\ $c_1$ & $0.1782(19)$ & $0.5035(54)$\\ $d_0$ & $-0.1885(20)$ & $-0.5326(56)$\\ $d_1$ & $-0.06919(86)$ & $-0.1955(24)$\\ \hline \multicolumn{3}{|c|}{Spin-isospin scattering lengths and volumes}\\ \hline $a_{0+}^{3/2}$ & $-0.0722(13)$ & $-0.1021(19)$\\ $a_{0+}^{1/2}$ & $0.16331(83)$ & $0.2309(12)$\\ $a_{1+}^{3/2}$ & $0.2153(22)$ & $0.6084(62)$\\ $a_{1-}^{3/2}$ & $-0.04236(87)$ & $-0.1197(24)$\\ $a_{1+}^{1/2}$ & $-0.03207(74)$ & $-0.0906(21)$\\ $a_{1-}^{1/2}$ & $-0.0822(16)$ & $-0.2322(45)$\\ \hline \multicolumn{3}{|c|}{Fully-corrected scattering lengths associated with the physical processes}\\ \hline $a_{\pi^+ p}$ & $-0.0727(15)$ & $-0.1027(22)$\\ $a_{\pi^- p}$ & $0.08535(69)$ & $0.12066(97)$\\ $a_{CX} \coloneqq \left( a_{\pi^+ p} - a_{\pi^- p} \right) / \sqrt{2}$ & $-0.1117(12)$ & $-0.1580(17)$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} Table \ref{tab:CXCnts} contains the predictions for the isovector scattering lengths and volumes, obtained from the exclusive fit of the ETH model to the tDB$_0$. The differences between the corresponding entries of Tables \ref{tab:PSACnts} and \ref{tab:CXCnts} for the isovector scattering lengths and volumes are significant in all cases. A comparison between the fully-corrected result $a_{CX}$ of Table \ref{tab:CXCnts} and the corresponding `experimental' entry of Table \ref{tab:AccAndAc0} suggests that the scattering data of the tDB$_0$ would rather favour a somewhat smaller value for the total decay width $\Gamma_{1s}$ of the ground state in pionic hydrogen; to be specific, a value in the vicinity of $790$ meV. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:CXCnts}}}The isovector $s$-wave scattering length (in $m_c^{-1}$ and in fm) and $p$-wave scattering volumes (in $m_c^{-3}$ and in fm$^3$), obtained from the exclusive fit of the ETH model to the tDB$_0$. The table also contains the fully-corrected (for the scale-factor effects of Section \ref{sec:ScaleFactors}) scattering length associated with the $\pi^- p$ CX reaction. \vspace{0.25cm} \begin{center} \begin{tabular}{|c|c|c|} \hline Scattering length/volume & In (reciprocal) powers of $m_c$ & In powers of fm\\ \hline \hline \multicolumn{3}{|c|}{Isovector scattering lengths and volumes}\\ \hline $b_1$ & $-0.08668(56)$ & $-0.12255(79)$\\ $c_1$ & $0.1975(23)$ & $0.5582(65)$\\ $d_1$ & $-0.0740(14)$ & $-0.2092(39)$\\ \hline \multicolumn{3}{|c|}{Fully-corrected scattering length of the $\pi^- p$ CX reaction}\\ \hline $a_{CX}$ & $-0.1235(12)$ & $-0.1746(16)$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} \subsubsection{\label{sec:ReprCX}Reproduction of the absolute normalisation of the DB$_0$ by the BLS$_\pm$} The methodology, which is relevant to this part, was put forward in Ref.~\cite{Matsinos2015}, and was first applied to the reproduction of the absolute normalisation of the DB$_0$ in the ZRH17 PSA \cite{Matsinos2017a}. One first determines a set of predictions from the PSA of the tDB$_\pm$, relating to the observables and to the corresponding values of the kinematical variables at which the actual measurements in the DB$_0$ had been acquired. Predictions and measured data are subsequently compared. The amount, at which each set of predicted values must be floated in order that it optimally reproduces the corresponding dataset of the DB$_0$, is determined; relevant to this part of the analysis are the free-floating scale factors $\hat{z}_j$ of Eq.~(\ref{eq:EQC007}). The extracted $\hat{z}_j$ values for the datasets, which are associated with measurements of the DCS (this includes the TCS, as well as the coefficients in the Legendre expansion of the DCS~\footnote{Being \emph{ratios} of cross sections, the APs might not be equally sensitive to the effect under investigation in this section, i.e., to the violation of isospin invariance.}), and their total uncertainties $\delta \hat{z}_j$, i.e., the purely statistical uncertainties \begin{equation*} \left( \delta \hat{z}_j \right)_{\rm stat} = \left( \sum_{i=1}^{N_j} w_{ij} \right) ^{-1/2} \end{equation*} (where the weights $w_{ij}$ are defined in Appendix \ref{App:AppC}) combined (quadratically) with the normalisation uncertainty $\delta z_j$ of the dataset, are given in Table \ref{tab:PSAReprCX} and, plotted separately for the DCS, for the TCS, and for the coefficients in the Legendre expansion of the DCS, in Fig.~\ref{fig:PSAReprCX}. Not included in the figure (but given in the table) are the results for the four FITZGERALD86 datasets which had been freely floated in the optimisation, as well as the $\hat{z}_j$ estimate for the BREITSCHOPF06 one-point dataset which had been removed at step (3) of the procedure of Section \ref{sec:Procedure}, see also Table \ref{tab:ProgPIMCX}. An interesting feature of Table \ref{tab:PSAReprCX} is that no dataset is labelled as having problematical shape; on the contrary, $21$ datasets are labelled as having problematical absolute normalisation. Inspection of Fig.~\ref{fig:PSAReprCX} leaves no doubt that the free-floating scale factors $\hat{z}_j$ of the datasets in the tDB$_0$ contain an appreciable amount of fluctuation. As the BLS$_\pm$ prediction is smooth, this fluctuation reflects the variation of the absolute normalisation of the datasets in the DB$_0$. For the sake of example, the $\hat{z}_j$ value for the FRLE{\v Z}98 dataset comes out equal to $1.405(99)$. This dataset lies in-between three datasets with considerably smaller $\hat{z}_j$ values, namely between the two ISENHOWER99 $20.60$-MeV datasets and the MEKTEROVI{\'C}09 $33.89$-MeV dataset. The results for the absolute normalisation of the two neighbouring datasets of DUCLOS73 ($22.60$ and $32.90$ MeV) and that of the JIA08 $34.37$-MeV dataset (both accompanied by large normalisation uncertainties), are compatible with the BLS$_\pm$. The exponential function \begin{equation} \label{eq:EQ006} \hat{z} = \alpha \exp(-\beta T) + 1 \end{equation} has been fitted to the $\hat{z}_j$ values of Fig.~\ref{fig:PSAReprCX}. The fitted values of the parameters $\alpha$ and $\beta$, corrected for the quality of the fit (resulting in $\chi^2_{\rm min} \approx 64.66$ for $47$ DoFs), are equal to $0.327(35)$ and $0.0225(30)$ MeV$^{-1}$, respectively. Similar results had been obtained in the former analyses of the recent past \cite{Matsinos2017a}, see Table \ref{tab:CXParametersOfExponential}, pointing to a significant discrepancy (between the DB$_0$ and the corresponding BLS$_\pm$) in the low-energy region. \vspace{0.5cm} \begin{table} {\bf \caption{\label{tab:CXParametersOfExponential}}}The fitted values and uncertainties of the parameters $\alpha$ and $\beta$ of the empirical exponential function of Eq.~(\ref{eq:EQ006}), which is used for modelling the energy dependence of the free-floating scale factors $\hat{z}_j$ of the $\pi^- p$ CX datasets, obtained from the reproduction of the DB$_0$ by the BLS$_\pm$. If the scale factors $\hat{z}_j$ had come out as predicted by the PSA of the tDB$_\pm$, then the parameter $\alpha$ would have vanished. \vspace{0.25cm} \begin{center} \begin{tabular}{|l|c|c|} \hline Year & $\alpha$ & $\beta$ (MeV$^{-1}$) \\ \hline \hline 2017 & $0.361(59)$ & $0.0221(39)$\\ 2019 & $0.359(70)$ & $0.0259(53)$\\ 2020 & $0.364(69)$ & $0.0263(52)$\\ This work & $0.327(35)$ & $0.0225(30)$\\ \hline \end{tabular} \end{center} \vspace{0.5cm} \end{table} Integrated between $0$ and $100$ MeV, this discrepancy between measured and predicted cross sections is equivalent to an effect at the level of $12.99(92)~\%$ or, naively converted into a relative difference (denoted as $r_2$ in Ref.~\cite{Matsinos2022a}) between the two $\pi^- p$ CX scattering amplitudes, of $6.30(43)~\%$. Although, owing to the pronounced energy dependence of the effect, these percentages should not to be taken seriously, they nevertheless corroborate the mathematically-sound comparison, based on Eq.~(\ref{eq:EQ002}) and presented in the following section. \subsubsection{\label{sec:TriangleIdentity}Test for the departure from the triangle identity} The objective in this section is to compare the real parts of two $\pi^- p$ CX scattering amplitudes: the first, $f_{\rm CX}$, represents the prediction obtained in this study from the PSA of the tDB$_\pm$; the second, $f^{\rm extr}_{\rm CX}$, is obtained from the exclusive fit of the ETH model to the tDB$_0$. According to Eq.~(3) of Ref.~\cite{Matsinos2022b}, the real part of each $\pi N$ partial-wave amplitude $f^{I}_{l\pm}$ and the corresponding phase shift $\delta^{I}_{l\pm}$ are linked via the expression: \begin{equation*} \Re [ f^{I}_{l\pm} ] = \frac{\sin \left( 2 \delta^{I}_{l\pm} \right)}{2 q} \, \, \, . \end{equation*} Expressed in terms of the two isospin amplitudes, the partial-wave amplitudes of the $\pi^- p$ CX reaction are given by the relation: \begin{equation*} (f_{CX})_{l\pm} = \frac{\sqrt{2}}{3} \left( f^{3/2}_{l\pm} - f^{1/2}_{l\pm} \right) \, \, \, . \end{equation*} For the $s$ wave, one obtains: \begin{equation} \label{eq:EQ008} \Re [ f_{CX} ]_s = \frac{\sqrt{2}}{6 q} \left( \sin \left( 2 \delta^{3/2}_{0+} \right) - \sin \left( 2 \delta^{1/2}_{0+} \right) \right) \, \, \, . \end{equation} Using Eqs.~(\ref{eq:EQ005}), the corresponding expressions for the no-spin-flip and for the spin-flip $p$-wave parts of the $\pi^- p$ CX scattering amplitude (respectively) read as: \begin{align} \label{eq:EQ009} \Re [ f_{CX} ]_{p; nsf} = \frac{\sqrt{2}}{6 q} & \Big( 2 \sin \big( 2 \delta^{3/2}_{1+} \big) + \sin \big( 2 \delta^{3/2}_{1-} \big)\nonumber\\ & - 2 \sin \big( 2 \delta^{1/2}_{1+} \big) - \sin \big( 2 \delta^{1/2}_{1-} \big) \Big) \, \, \, \text{and} \end{align} \begin{align} \label{eq:EQ010} \Re [ f_{CX} ]_{p; sf} = \frac{\sqrt{2}}{6 q} & \Big( \sin \big( 2 \delta^{3/2}_{1+} \big) - \sin \big( 2 \delta^{3/2}_{1-} \big)\nonumber\\ & - \sin \big( 2 \delta^{1/2}_{1+} \big) + \sin \big( 2 \delta^{1/2}_{1-} \big) \Big) \, \, \, . \end{align} The real parts of the amplitudes $f_{\rm CX}$ and $f^{\rm extr}_{\rm CX}$ are shown in Figs.~\ref{fig:TCXREs}-\ref{fig:TCXREpsf}, separately for the $s$ wave, and for the no-spin-flip and spin-flip $p$-wave parts. It was over twenty-five years ago when such a comparison was made, in the first report on the violation of isospin invariance in the $\pi N$ interaction at low energy \cite{Gibbs1995}, see Figs.~1-3 therein. Shown in those figures was the energy dependence of the real parts of the three partial-wave amplitudes in the energy domain between $T=30$ and $50$ MeV: their difference in the $s$ wave was found to be nearly constant, evaluated in Ref.~\cite{Gibbs1995} to $D \coloneqq \Re [ f^{\rm extr}_{CX} ]_s - \Re [ f_{CX} ]_s = -0.012(3)$ fm. Although the real part of the $s$-wave $\pi^- p$ CX scattering amplitude, obtained in this study from the PSA of the tDB$_\pm$, seems to exhibit a more pronounced energy dependence (in comparison with the corresponding result of Ref.~\cite{Gibbs1995}), the estimate of this work for the difference $D$ (an average over the energy domain of Ref.~\cite{Gibbs1995}) is nearly unchanged: $D \approx -0.013$ fm. A similar result was obtained in Ref.~\cite{Matsinos2022b} using (as modelling option of the hadronic part of the $\pi N$ interaction) the ETH parameterisation. As Fig.~\ref{fig:TCXREs} indicates, the energy dependence of the difference between the real parts of the two $s$-wave $\pi^- p$ CX scattering amplitudes is pronounced in the low-energy region, decreasing with increasing energy. In comparison with the result of Ref.~\cite{Matsinos2022b}, the convergence of the two solutions evidently occurs above the upper $T$ bound of this project (i.e., above $100$ MeV). One of the popular ways of quantifying the difference between the two aforementioned $\pi^- p$ CX scattering amplitudes employs the indicator $R_2$ of Eq.~(\ref{eq:EQ002}). The energy dependence of the quantity $R_2$ is shown in Figs.~\ref{fig:R2REs}-\ref{fig:R2REpsf}, separately for the $s$ wave, and for the no-spin-flip and spin-flip $p$-wave parts. \subsubsection{\label{sec:ReprCHAOSDCS}Reproduction of the DCSs of the CHAOS Collaboration by the BLS$_\pm$} This section examines the reproduction of the CHAOS DCSs \cite{Denz2006}, which amount to a prodigious $546$ datapoints, the largest (by far) contribution to the low-energy $\pi N$ DB from a single experimental campaign (thus far). After two exclusive analyses of these measurements a few years ago \cite{Matsinos2013a,Matsinos2015}, it was decided to exclude them from the DB of this project; although the subject was revisited in subsequent PSAs \cite{Matsinos2017a}, that decision remained unchanged. After following the recommendation by two members of the CHAOS Collaboration \cite{Denz2011} to avoid segmenting their data, the twelve original datasets (six for each ES reaction~\footnote{For both reactions, the CHAOS data contain two datasets at $43.30$ MeV: the second set (in each reaction) was obtained after the target was rotated; to distinguish between the two target configurations for the $43.30$-MeV datasets, the ones with rotated target are labelled `$43.30$(rot.)'.\label{ftn:FTN1}}) were submitted to the tests of the overall reproduction, of the reproduction of the shape, and of the reproduction of the absolute normalisation, as these tests are set forth at the end of Appendix \ref{App:AppC}. The following conclusions were drawn. \begin{itemize} \item None of the $\pi^+ p$ datasets can be accounted for. The best-reproduced dataset was the one at the lowest energy ($19.90$ MeV): the p-value of its reproduction came out equal to a mere $1.12 \cdot 10^{-6}$. \item In all cases, the inability to reproduce the $\pi^+ p$ datasets may be imputed to the \emph{shape} of the angular distribution of the DCS. (The absolute normalisation of the $32.0$-MeV dataset also emerges as problematical.) \item Apart from the $32.00$-MeV dataset, the shape of the angular distribution of the $\pi^- p$ ES measurements is also problematical, though in a less pronounced way than it is in case of the $\pi^+ p$ data. \item If one pays no heed to the problematical shape of the angular distribution of these datasets and proceeds with the determination of a representative, so to speak~\footnote{Without doubt, given the problematical shape of the DCSs in both cases, the averages of this paragraph ought to be taken with a grain of salt.}, absolute normalisation of the data (e.g., by averaging the free-floating scale factors obtained with Eq.~(\ref{eq:EQC007})), then the absolute normalisation of the $\pi^+ p$ datasets turns out to be between $5.9$ and $15.2~\%$ ($95~\%$ confidence interval) \emph{above} the bulk of the tDB$_+$ ($\avg{\hat{z}}=1.105(24)$), whereas that of the $\pi^- p$ ES datasets is (on average) between $1.9$ and $11.3~\%$ ($95~\%$ confidence interval) \emph{below} the absolute normalisation of the tDB$_-$ ($\avg{\hat{z}}=0.934(24)$); of course, the absolute normalisation of the tDB$_-$ is largely determined by the precise PSI measurements of $\epsilon_{1s}$ in pionic hydrogen (which lead to the scattering length $a_{\rm cc}$). Although one might be tempted to conclude that the absolute normalisation of the CHAOS $\pi^- p$ ES datasets does not match well the value of the scattering length $a_{\rm cc}$, as this quantity emerges from the experimental investigation at the $\pi N$ threshold, such a conclusion might be rash at this time~\footnote{Prior to the inclusion in the DB of this project of the two `experimental' values of the scattering length $a_{\rm cc}$, the extrapolation of the $\pi^- p$ ES amplitude from the scattering region to the $\pi N$ threshold invariably resulted in \emph{lower} $a_{\rm cc}$ values than those extracted from the experiments on pionic hydrogen. In this respect, the (representative) absolute normalisation of the CHAOS $\pi^- p$ ES datasets is not necessarily in conflict with that of the tDB$_-$; it only appears to be, simply due to the large impact which the precise measurements at the $\pi N$ threshold have on the optimisation, by `attracting' the fitted values of the $\pi^- p$ ES amplitude in their direction. One consequence of the inclusion of the two experimental $a_{\rm cc}$ results in the DB of this project is that the reproduction of the CHAOS $\pi^- p$ ES measurements by the BLS$_\pm$ deteriorated.}. \end{itemize} The reproduction of the CHAOS DCSs by the BLS$_\pm$ of this work is shown in Figs.~\ref{fig:PIPPEDENZ} and \ref{fig:PIMPEDENZ}. As a sum over the $(\chi^2_j)_{\rm min}$ contributions of Eq.~(\ref{eq:EQC006}), one obtains the (reproduction) $\chi^2$ values of $1194.10$ for $275$ $\pi^+ p$ datapoints and $730.36$ for $271$ $\pi^- p$ ES datapoints: the largeness of these results demonstrates beyond doubt that the CHAOS DCSs (original datasets) cannot be reproduced by the BLS$_\pm$ of this work (as well). In comparison with former works in this project (e.g., with the ZRH19 PSA), the quality of the reproduction of these datasets slightly deteriorated due to one additional reason: the predictions are now accompanied by somewhat smaller uncertainties, emanating from the narrower distribution from which the parameter $m_\sigma$ is currently sampled for the joint fits of the ETH model to the tDB$_\pm$. In Refs.~\cite{Matsinos2013a,Matsinos2015}, it was argued that the segmentation of the original datasets of the CHAOS Collaboration, into forward-, intermediate-, and backward-angle data, is dictated by the procedure which had led to the reported results in Ref.~\cite{Denz2006}. In fact, the CHAOS DCSs had been inserted into (and can be found in) the SAID DB in the form of segmented datasets. Of course, in comparison with the original ones, the segmented datasets are expected to be accounted for in a more satisfactory manner (regardless of the analysis method), the reason being that the segmented parts of each original dataset are floated independently of each other. As contribution of the CHAOS DCSs to the $\chi^2_{\rm min}$ of their fit, the SAID group report a total value of $1102.72$ for $545$ datapoints ($626.91/274 \approx 2.29$ for the $\pi^+ p$ and $475.81/271 \approx 1.76$ for the $\pi^- p$ ES datasets); one datapoint of the $\pi^+ p$ $37.10$-MeV dataset at intermediate angles has been removed from the input by the SAID analysts. Therefore, even the segmented datasets are poorly described by the WI08 solution~\footnote{Using the SAID Analysis Program \cite{SAID}, I was not able to obtain information about the description of the data on the basis of the more recent solution XP15. The selection of the `current' solution in the user interface brings up the details about the description which is obtained on the basis of the former solution WI08 \cite{Arndt2006}, the remaining two options involving results corresponding to two outdated Karlsruhe analyses of the 1980s. However, the chances are that the difference in the description of the CHAOS DCSs between the WI08 and the XP15 solutions is (at worst) negligible.}: when considering the description of the $29$ segmented datasets by the WI08 solution, $15$ p-values are below the default significance threshold $\mathrm{p}_{\rm min}$ of this project. The segmented datasets were also analysed in this study. The conclusion is that the following $\pi^+ p$ datasets cannot be reproduced: \begin{itemize} \item the observations at $19.90$, $25.80$, $32.00$, $43.30$, and $43.30$(rot.) MeV at forward angles (i.e., five of the six datasets involving measurements in the Coulomb peak) and \item the observations at $19.90$, $25.80$, $32.00$, and $43.30$(rot.) MeV at intermediate angles. \end{itemize} In three cases, the problems are due to problematical shape; in six, due to problematical absolute normalisation. From the $\pi^- p$ ES measurements, both $25.80$-MeV datasets, as well as the $37.10$- and $43.30$(rot.)-MeV datasets at intermediate/backward angles, are poorly reproduced (all due to problematical shape). Using the segmented datasets, the reproduction $\chi^2$ values were obtained (sums over the $(\chi^2_j)_{\rm min}$ contributions given in Eq.~(\ref{eq:EQC006})): $516.54$ for $275$ $\pi^+ p$ datapoints and $521.42$ for $271$ $\pi^- p$ ES datapoints. Table \ref{tab:PSAReprDENZ} provides the important details about the reproduction of the CHAOS DCSs, both of the original, as well as of the segmented datasets. It would be instructive to enquire further into the description by the WI08 solution of the nine $\pi^+ p$ (segmented) datasets, which cannot be reproduced by the BLS$_\pm$ of this work. The sum of the relevant $\chi^2$ values, directly copied from the output of the SAID Analysis Program, is equal to $385.75$. As $149$ datapoints are contained within these sets, the reduced $\chi^2$ value, corresponding to the description of these datasets by the WI08 solution, is equal to about $2.59$, hence poorer by comparison with the average description of the CHAOS DCSs by that solution. Further examination might be revealing as far as the identification of the source of the observed discrepancies is concerned. One of the possibilities is to analyse the results obtained in the three angular intervals of the segmented $\pi^+ p$ datasets (i.e., forward, intermediate, and backward). The reproduction of the $\pi^+ p$ data at backward angles by the BLS$_\pm$ of this work leads to $\chi^2_{\rm b} \approx 58.15$ for $N_{\rm b} = 62$ datapoints. This result is very satisfactory~\footnote{As the hadronic part of the $\pi N$ interaction dominates at backward angles (the direct EM contributions vanish at $\theta=\pi$ rad), this result suggests that the hadronic component of the $\pi N$ amplitudes, emerging from the DCSs of the CHAOS Collaboration, can be accounted for by the BLS$_\pm$ of this work!} and may be used as reference in the $F$-tests (see Ref.~\cite{Spiegel2013}, pp.~116-117), which will aim at assessing the quality of the reproduction of the datasets at forward and at intermediate angles \emph{relative to the reproduction at backward angles}. In the former case, $\chi^2_{\rm f} \approx 176.39$ for $N_{\rm f} = 48$ datapoints. Using the reduced $\chi^2$ value of the datasets at backward angles as reference, one obtains for the $F$-statistic: $F_{\rm f/b} = (\chi^2_{\rm f} / N_{\rm f}) / (\chi^2_{\rm b} / N_{\rm b}) \approx 3.92$ for $N_{\rm f}$ and $N_{\rm b}$ DoFs, resulting in the p-value of about $3.38 \cdot 10^{-7}$ and (thus) suggesting a \emph{substantial} deterioration in the reproduction of the datasets at forward angles (with respect to that of the measurements at backward ones); visual inspection of Fig.~\ref{fig:PIPPEDENZ} confirms this result. The corresponding p-value for the datasets at intermediate angles comes out equal to about $3.77 \cdot 10^{-3}$, hence indicative of a (milder) deterioration in the reproduction of the $\pi^+ p$ data as one shifts from backward to intermediate angles. Evidently, the quality of the reproduction of the CHAOS $\pi^+ p$ DCSs steadily deteriorates as one moves from $\theta=\pi$ to $0$ rad. It must be mentioned that the description of the $\pi^+ p$ data by the WI08 solution is equally poor (or, for those who see the glass half full, equally good) in all three angular intervals (forward, intermediate, and backward): their reduced $\chi^2$ values for the $\pi^+ p$ datasets range between $2.22$ and $2.31$, thus providing support for the scenario of a general underestimation of the uncertainties on the part of the CHAOS Collaboration regardless of the angular domain. In any case, the output of the SAID Analysis Program \cite{SAID} provides no evidence of significant deterioration of the description of the segmented datasets by the WI08 solution in any of the aforementioned angular intervals. This is an interesting difference between the BLS$_\pm$ of this work and the WI08 solution. The problematical nature of these DCSs may already be revealed by examining the experimental reports of the CHAOS Collaboration. For instance, upon inspection of the single-energy phase shifts obtained from these data (see Table 6.1 of Denz's dissertation and Fig.~4 of their main publication \cite{Denz2006}), one cannot but feel rather uneasy about the quoted values. Of course, it is true that single-energy phase-shift solutions cannot exhibit the smoothness of the results obtained when the energy dependence of the phase shifts is modelled by means of continuous functions. Having said that, the estimate for the phase shift $P_{31}$ ($+0.65^\circ$, no uncertainty has been quoted) in Table 6.1 of the dissertation, at $19.90$ MeV, is wrong by about $0.9^\circ$; the \emph{largest} of the $p$-wave phase shifts at low energy, $P_{33}$, is itself about $1^\circ$ at that energy! At $20$ MeV, the WI08 result \cite{Arndt2006} for $P_{31}$ ($-0.22^\circ$) generally agrees with the values reported in this programme over time, see also Table \ref{tab:PSAPhSh}. It is my thesis that this discrepancy alone ought to have sufficed in providing the motivation for the thorough re-examination of the CHAOS DCSs. Due to the reasons which have been laid out in this section, I cannot recommend the use of the CHAOS DCSs in analyses of the low-energy $\pi N$ data, in particular in sensitive ones, as (for the sake of example) are those which aim at evaluating the $\pi N$ $\Sigma$ term \cite{Matsinos2013b}. Unless these measurements are revised in the light of Refs.~\cite{Matsinos2017a,Matsinos2013a,Matsinos2015}, as well as of this work, I see no alternative to excluding the CHAOS DCSs from the PSAs of this project. It would be constructive if the exclusive description of the CHAOS DCSs were pursued by means of a different methodology. As the beam energy of these datasets is low, an investigation of their description within the framework of Chiral-Perturbation Theory, e.g., using the method of Ref.~\cite{Alarcon2013}, or following the method of Ref.~\cite{Hoferichter2016} should be possible. Alternatively, the reproduction of these datasets could be examined on the basis of available results, obtained from the remaining $\pi N$ data in the aforementioned schemes. I shall conclude this section with a few words about the inclusion of the CHAOS DCSs in the SAID DB. Provided that their number remains reasonably small, the methodology, which is followed in the analyses of the ETH $\pi N$ project, can handle the presence of outliers in the DBs, both in terms of the shape as well as of the absolute normalisation of the datasets. However, the breakdown point~\footnote{The breakdown point of an analysis is the fraction of discrepant data it can handle before it starts to produce erroneous results. The breakdown point is an increasing function of the level of robustness of the analysis method.} of the approach is expected to be low, presumably not exceeding about $20~\%$. Given the restriction of the DBs to low energy, the inclusion of the CHAOS DCSs in the DB$_\pm$ of this project cannot be considered `small' by any means; the addition of $546$ datapoints to a DB, which contains $813$ measurements, makes sense if the existing and the added parts are not in conflict. Figures \ref{fig:PIPPEDENZ} and \ref{fig:PIMPEDENZ} amply demonstrate that this is hardly the case. On the other hand, the SAID group could afford to add the entirety of the CHAOS DCSs to their DBs because they (i.e., their 2004 DBs) already contained tens of thousands of datapoints (namely all $\pi N$ measurements up to the few-GeV region, barring the measurements of the PTCS/TNCS/TCS and those at the $\pi N$ threshold); their sensitivity to the treatment of a few hundred unscreened measurements is certainly lower than it would have been in case of this project. \section{\label{sec:Conclusions}Conclusions and discussion} Carried out in this study is an improved analysis of the measurements of the three pion-nucleon ($\pi N$) reactions which can be subjected to experimental investigation at low energy (pion laboratory kinetic energy $T \leq 100$ MeV), i.e., of the two elastic-scattering (ES) processes $\pi^\pm p \to \pi^\pm p$ and of the $\pi^- p$ charge-exchange (CX) reaction $\pi^- p \to \pi^0 n$. There are a number of differences to the works of the recent past (which can be found online as versions of Ref.~\cite{Matsinos2017a}). The most important modification relates to the analysis procedure. The old analysis procedure, which had been applied to several works during the past decade, can be found in Section 3.1 of Ref.~\cite{Matsinos2022b}. The new analysis procedure, which is detailed in Section \ref{sec:Procedure} of this study, introduces a number of improvements, which (hopefully) make its application more straightforward and comprehensible to the (non-expert) reader. \begin{itemize} \item[a)] The joint fits of the ETH model to the measurements of the $\pi^+ p$ and $\pi^- p$ CX reactions, which had been carried out in order that a second phase-shift solution (one involving the $\pi^- p$ CX data) be reliably determined, will be replaced (from now on) by exclusive fits of that model to the low-energy database (DB) of the $\pi^- p$ CX reaction. This development became possible after suppressing (whenever the $\pi^- p$ CX measurements are submitted to the optimisation) the contributions to the partial-wave amplitudes of the ETH model which originate from the scalar-isoscalar $t$-channel Feynman diagram, see Fig.~\ref{fig:GraphsETH} (left graph, upper part).\\ The main solution, obtained with the ETH model, will continue to be the one associated with the phase-shift analysis (PSA) of the measurements of the ES reactions. The second solution will be obtained from the exclusive fit of the ETH model to the low-energy $\pi^- p$ CX data. As a result, the extraction of the $\pi^- p$ CX scattering amplitude from the experimental data of that reaction will be `clean', in the sense that it will no longer rely on the involvement of the experimental data of a second $\pi N$ process. As the two types of fits now involve input DBs which have no commonality, the two solutions, obtained from these fits, are \emph{independent} of one another. \item[b)] The modifications of case (a) above enabled the reliable evaluation of the popular measure of the violation of isospin invariance $R_2$ of Eq.~(\ref{eq:EQ002}), in the entire low-energy region, separately for the $s$ wave, as well as for the no-spin-flip and spin-flip $p$-wave parts of the $\pi N$ scattering amplitude, see Figs.~\ref{fig:TCXREs}-\ref{fig:R2REpsf}. \item[c)] For the first time in several years, the DB of the low-energy measurements has been enhanced in this work with the inclusion of sixteen datapoints, which had been available for a long time but had not been used thus far (on account of the reasons which can be found in Section \ref{sec:Database}), namely of the differential cross sections (DCSs) of two $\pi^+ p$ experiments of the late 1970s \cite{Moinester1978,Blecher1979}, as well as of the three-point dataset of Ref.~\cite{Ullmann1986}, which corresponds to measurements of the $\pi^- p$ CX DCS at $T \approx 49$ MeV. One additional datapoint, the scattering length $a_{\rm c0}$, as this entry emerges from the recent publication of the total decay width of the ground state in pionic hydrogen by the Pionic Hydrogen Collaboration \cite{Hirtl2021} - and the application of the electromagnetic (EM) corrections of Ref.~\cite{Oades2007}, was appended to the DB of the $\pi^- p$ CX reaction. \item[d)] Apart from the physical properties of a few higher baryon resonances \cite{Matsinos2020a}, as well as of three scalar-isoscalar and one vector-isovector mesons \cite{Matsinos2020b}, needed in order to determine the (inessential) contributions from these states to the partial-wave amplitudes of the ETH model, the values of the physical constants have been fixed in this analysis from the recent compilation of the Particle-Data Group \cite{PDG2022}. \end{itemize} After the application of the four steps of the first phase of the new analysis procedure, $52$ entries (of the $1150$ DoFs of the new initial DB of this project, see Table \ref{tab:DB}) were identified as outliers and were removed from the DBs, see Tables \ref{tab:ProgPIP}-\ref{tab:ProgPIMCX}. In this manner, three truncated DBs (tDBs, see the notation at the end of Section \ref{sec:Introduction}) were obtained, which were submitted to further analysis using the ETH model, see the two last steps of the analysis procedure described in Section \ref{sec:Procedure}. The optimisation of the description of the ES measurements (step (5) of Section \ref{sec:Procedure}), i.e., of the tDB$_\pm$, was achieved on the basis of the numerical minimisation provided by the MINUIT software package \cite{James} of the CERN library (FORTRAN version), using the Arndt-Roper formula \cite{Arndt1972} to obtain the contributions of the datasets to the overall $\chi^2$, see Appendix \ref{App:AppA}. The PSA of the tDB$_\pm$ comprises the results of one hundred joint fits to the same data, each fit being carried out at one value of the model parameter $m_\sigma$, which is associated with the effective range of the scalar-isoscalar part of the $\pi N$ interaction: the $m_\sigma$ values were randomly generated in normal distribution, in accordance with the result of Ref.~\cite{Matsinos2020b}. The optimal values of the parameters of the ETH model, taking all uncertainties into account (as well as the $m_\sigma$ variation), can be found (corrected for the quality of the fits via the application of the Birge factor \cite{Birge1932}) in Table \ref{tab:ParETH}. The most important of these parameters is the pseudoscalar $\pi N$ coupling constant $g_{\pi N N}$, which comes out equal to $13.18(13)$. Converted into a result for the pseudovector $\pi N$ coupling constant $f_{\pi N N}$, to be identified with the charged-pion coupling constant $f_c$ of Ref.~\cite{Matsinos2019}, this $g_{\pi N N}$ estimate translates into: \begin{equation*} f_{\pi N N}^2=0.0764(15) \, \, \, . \end{equation*} This result is in good agreement with the findings of Refs.~\cite{Matsinos2019,Reinert2021}, as well as with the values extracted in the analyses of the SAID group over time, see Ref.~\cite{Matsinos2019} for details. The results of the PSA of the tDB$_\pm$ lead to predictions for the $\pi N$ phase shifts (Section \ref{sec:PhaseShifts}, Figs.~\ref{fig:S31}-\ref{fig:P11}), for the low-energy constants (LECs) of the $\pi N$ interaction (Section \ref{sec:LECs}), etc. In addition, they can be used to derive estimates for all measurable quantities, for all three low-energy $\pi N$ reactions. Two such sets of predictions were obtained in this work: the first was used in the assessment of the quality of the reproduction of the absolute normalisation of the DB$_0$ (Section \ref{sec:ReprCX}, Table \ref{tab:PSAReprCX}, Fig.~\ref{fig:PSAReprCX}); the second was used in the assessment of the quality of the reproduction of the DCSs of the CHAOS Collaboration, which are not included in the DB of this project (Section \ref{sec:ReprCHAOSDCS}, Table \ref{tab:PSAReprDENZ}, Figs.~\ref{fig:PIPPEDENZ} and \ref{fig:PIMPEDENZ}). The interest in the former case relates to the tests of the triangle identity of Eq.~(\ref{eq:EQ001}), a relation which is fulfilled if the isospin invariance holds in the low-energy $\pi N$ interaction. The interest in the latter case relates to providing an explanation for the inability to account for the DCSs of the Chaos Collaboration within the ETH $\pi N$ project. The exclusive fit of the ETH model to the $\pi^- p$ CX data cannot determine both isospin amplitudes, $f_3$ and $f_1$; only their difference can reliably be determined from that fit. Consequently, the analysis of the tDB$_0$ using the ETH model is technically not a PSA, which implies that it cannot lead to a reliable phase-shift solution. From the exclusive fit of the ETH model to the tDB$_0$, only quantities associated with the $\pi^- p$ CX reaction can be estimated, e.g., the isovector scattering lengths and volumes (see end of Section \ref{sec:LECs}), as well as the scattering amplitude $f^{\rm extr}_{\rm CX}$. Cutkosky introduced in 1979 the Quantum-Mechanical (QM) admixture of the $\pi^0$ and the $\eta$ mesons as a potential mechanism for the violation of isospin invariance in the $\pi N$ interaction \cite{Cutkosky1979}; affected by this mechanism is the $\pi^- p$ CX reaction. The consequences of fitting an isospin-invariant model (such as the ETH model is) to measurements which presumably contain isospin-breaking effects have been touched upon at the beginning of Section \ref{sec:IBFD}. Provided that the QM admixture of the $\pi^0$ and the $\eta$ mesons is the dominant means by which the violation of isospin invariance occurs in the $\pi^- p$ CX reaction, the creation of a complete hadronic model of that reaction necessitates the addition of the contributions from the Feynman diagrams of Fig.~\ref{fig:EtaPi0} to those of Fig.~\ref{fig:GraphsETH}. As the five-parameter fit of the ETH model to the tDB$_0$ verges on the border of the realms of possibility, any additional parameters (i.e., introduced by the Feynman diagrams of Fig.~\ref{fig:EtaPi0}) ought to be fixed from external sources; unfortunately, the coupling constants and the vertex factors of the $\eta$ meson are not well known (if at all) at this time, see end of Section 4.1 of Ref.~\cite{Matsinos2022a} for comments on the coupling constant $g_{\eta N N}$. Therefore, one is left with only one option: to fit the ETH model, in its current isospin-invariant form, to the data and hope that the optimisation of the description of the tDB$_0$ will return a reasonable $\chi^2_{\rm min}$ for the NDF of the fit and no significant effects in the distribution of the fitted values of the scale factor $z$. This was the strategy in this study (step (6) of Section \ref{sec:Procedure}), and it seems to have been successful on both accounts. Having said that, the fitted values of the parameters of the ETH model, as they emerge from this exclusive fit, are bound to be effective, as they certainly contain contributions which are (currently) beyond the modelling of the hadronic part of the $\pi^- p$ CX reaction by the ETH model. The differences between the two sets of results (i.e., between those obtained from the PSA of the tDB$_\pm$ and the ones resulting from the exclusive fit of the ETH model to the tDB$_0$) are significant for all model parameters, see Table \ref{tab:ParETH}. In Section \ref{sec:ReprCX}, the reproduction of the absolute normalisation of the DB$_0$ was investigated on the basis of the results obtained from the PSA of the tDB$_\pm$. Figure \ref{fig:PSAReprCX} demonstrates that most of the $\pi^- p$ CX datasets have an absolute normalisation which exceeds the prediction obtained from the ES data. This result is not new; similar plots have been obtained in all analyses of the low-energy $\pi N$ data within the ETH $\pi N$ project during the past two decades, resulting in a significant overall discrepancy between the two $\pi^- p$ CX scattering amplitudes, i.e., between \begin{itemize} \item the amplitude $f_{\rm CX}$, which is reconstructed from the ES data, i.e., obtained from the PSA of the tDB$_\pm$ via Eq.~(\ref{eq:EQ001}), and \item the amplitude $f^{\rm extr}_{\rm CX}$, which is obtained from the exclusive fit of the ETH model to the tDB$_0$. \end{itemize} Integrated over the entire energy domain of this project, this discrepancy amounts to $r_2=6.30(43)~\%$; the corresponding results, obtained in the 2019 and 2020 PSAs of this project, were equal to $6.17(64)~\%$ and $6.19(63)~\%$, respectively. However, given the pronounced energy dependence of the effect, it must not be forgotten that the quantity $r_2$ serves as an overall measure of the discrepancy at low energy; unlike the symmetrised relative difference $R_2$ between the two amplitudes, see Eq.~(\ref{eq:EQ002}), neither does it convey information about the energy dependence of the effect, nor can it reveal the source of the effect (i.e., the partial wave which is predominantly affected). As aforementioned, the two amplitudes $f_{\rm CX}$ and $f^{\rm extr}_{\rm CX}$ emerge in this study from independent observations for the first time in years (in fact, since 1997 \cite{Matsinos1997}); therefore, the results of their comparison are expected to be reliable. Separate comparisons of the energy dependence of these amplitudes are shown: in Fig.~\ref{fig:TCXREs} for the $s$ wave; in Fig.~\ref{fig:TCXREpnsf} for the no-spin-flip $p$-wave part; and in Fig.~\ref{fig:TCXREpsf} for the spin-flip $p$-wave part. The quantity $R_2$ is also shown in Figs.~\ref{fig:R2REs}-\ref{fig:R2REpsf}. Inspection of these figures suggests that significant effects can be established in the $s$ wave and in the no-spin-flip $p$-wave part of the $\pi N$ scattering amplitude; less significant effects can be observed in the spin-flip $p$-wave part. These findings generally agree with the results of Refs.~\cite{Matsinos1997,Gibbs1995}: in itself, this is quite remarkable, given that both studies had been based on a small fraction of the DB$_0$ of this work. Aiming at providing an explanation for the departure of the quantity $R_2$ from $0$, three possibilities (or their combination) have been introduced and thoroughly discussed in Ref.~\cite{Matsinos2022a}, see Sections 5.1-5.3 therein; there is no need to repeat the arguments in this study. To summarise, the discrepancy may be attributable to: \begin{itemize} \item systematic effects in the absolute normalisation of the bulk of the low-energy $\pi N$ data, \item substantial residual contributions (i.e., at present not included) in the EM corrections (which are applied to the data in order that the hadronic quantities be extracted), and \item the violation of isospin invariance in the $\pi N$ interaction well beyond the expectations of Chiral-Perturbation Theory \cite{Hoferichter2010}. \end{itemize} I should like to remind the interested reader that predictions for the usual $\pi N$ observables (DCS, AP, PTCS, TNCS, and TCS) for the three low-energy $\pi N$ reactions are simple to obtain, free of charge, and available within one week of the request date. Unlike the results obtained from dispersion relations, the predictions obtained within the ETH $\pi N$ project are accompanied by uncertainties which reflect the statistical and systematic variation of the experimental data. \begin{ack} This research programme has been shaped to its current form after the long-term interaction with B.L.~Birbrair (deceased), A.~Gashi, P.F.A.~Goudsmit, A.B.~Gridnev, H.J.~Leisi, G.C.~Oades (deceased), G.~Rasche, and W.S.~Woolcock (deceased). I am deeply grateful to them for their contributions. I have additional reasons to be indebted to G.~Rasche: for our numerous discussions on issues regarding this research programme, many subjects in Hadronic Physics, and broader matters regarding the Philosophy of Science. The Feynman diagrams in this paper were created with the software package JaxoDraw \cite{Binosi2004,Binosi2009}, available from jaxodraw.sourceforge.net. The remaining figures were created with MATLAB\textsuperscript{\textregistered}~(The MathWorks, Inc., Natick, Massachusetts, United States). \end{ack}

1. Introduction =============== Sialyl-Tn (STn) is a carbohydrate antigen discovered as a cancer marker in the early 80s. Detected in virtually all epithelial cancers investigated, STn has known a 20 years long golden age as a research topic with two main focuses. The first focus, until about the mid 90s, was the value of STn as a marker for diagnosis and subsequently prognosis in cancer. The second focus was the targeting of STn by state-of-the-art immunotherapy strategies to treat cancers, notably breast cancer. On this topic, a very enthusiastic literature spanned from the mid 90s to 2005, the year when STn was somewhat crushed and buried under the failure of the Phase III clinical trial of the Theratope. Writing this review, we have tried to provide the reader with the most comprehensive vision of STn in cancer, from its molecular synthesis to its possibly underestimated usefulness as a therapeutic target. A substantial part of this manuscript is dedicated to the expression of STn across various types of cancers in order to discuss, in depth, the therapeutic strategies, past and future, targeting STn to treat cancer. 2. Sialyl-Tn Structure and Biosynthesis ======================================= STn (Neu5Acα2-6GalNAcα-*O*-Ser/Thr), also referred to as CD175s by the "cluster of differentiation" nomenclature, is the simplest sialylated mucin-type *O*-glycan. STn is a disaccharide formed of one residue of *N*-acetyl-galactosamine (GalNAc) alpha-*O*-linked to a serine or a threonine residue, and substituted by a sialic acid (Neu5Ac in human) on carbon 6. This sialylation prevents the formation of various core structures otherwise found in mucin-type *O*-glycans ([Figure 1](#biomolecules-02-00435-f001){ref-type="fig"}). {#biomolecules-02-00435-f001} Biosynthesis of *O*-glycans is a step-by-step process that occurs in the Golgi apparatus. *O*-glycans are synthesised by the sequential action of several glycosyltransferases, each transferring a monosaccharide \[e.g., GalNAc, Neu5Ac, galactose (Gal), *N*-acetyl-glucosamine (GlcNAc)\] from a donor nucleotide-sugar (e.g., UDP-GalNAc or CMP-Neu5Ac) to an acceptor that is the glycan being synthesized. Glycosyltransferases are membrane-bound enzymes of which the level of expression, substrate specificity and localisation in Golgi compartments are responsible for the pattern of *O*-glycans expressed in a given cell or carried by a given glycoprotein. Furthermore, because several distinct enzymes may compete for the same acceptor structure, their relative expression and activities are decisive for the structures expressed in each cell. Consequently, STn expression depends on the expression of polypeptide-GalNAc-transferase and sialyltransferase activities. Similarly, STn expression might be promoted by the weak expression or activity of enzymes competing with sialyltransferases that are the Gal-, GalNAc- and GlcNAc-transferases that synthesize *O*-glycan 'cores' ([Figure 1](#biomolecules-02-00435-f001){ref-type="fig"}). Polypeptide GalNAc-transferases constitute a family of 20 enzymes, all catalyzing the transfer of a GalNAc on the protein backbone. This diversity allows a finely tuned control of the initiation of *O*-glycosylation in a cell-specific and, possibly, a protein-specific manner. As the GalNAc is necessary for the transfer of sialic acid to create STn structure, one can say that the efficient GalNAc-transferase activity is crucial for STn expression. Logically, cells lacking these activities are therefore unable to express STn or any other GalNAc-based *O*-glycans, as observed in the testis for spermatogonia and Sertoli cells \[[@B1-biomolecules-02-00435]\]. However, to date, there is no demonstration that the quantity or the quality (e.g., clustering or specific sites) of *O*-GalNAcylation of proteins affects the subsequent sialylation of GalNAc residues. Sialylation of single *O*-GalNAc has been shown to be performed *in vitro* by two members of the sialyltransferase family, namely ST6GalNAc I and ST6GalNAc II \[[@B2-biomolecules-02-00435],[@B3-biomolecules-02-00435]\]. However, studies using cells transfected by either one of these enzymes have demonstrated that in a cellular context, only ST6GalNAc I is able to create STn structures as recognized by anti-STn antibodies \[[@B2-biomolecules-02-00435],[@B4-biomolecules-02-00435]\]. Furthermore, ST6GalNAc I over-expression was shown to correlate with STn expression in gastric and breast tumors confirming the crucial role of ST6GalNAc I in STn biosynthesis \[[@B4-biomolecules-02-00435],[@B5-biomolecules-02-00435]\]. Indeed, over-expression of ST6GalNAc I is able to compete with *O*-glycan cores biosynthesis as shown in MDA-MB-231 breast cancer cells where stable expression of ST6GalNAc I converted 22% of Core 1-based *O*-glycans carried by MUC1 mucin into STn \[[@B6-biomolecules-02-00435]\]. ST6GalNAc I activity can be eased by a weak activity of competing core-synthases. One such mechanism has been discovered by Cummings *et al*. who reported the existence of a chaperone protein for the Core 1 synthase (core 1 β1,3-galactosyltransferase) designated Cosmc ([co]{.ul}re 1 β3-Gal-T-[s]{.ul}pecific [m]{.ul}olecular [c]{.ul}haperone) \[[@B7-biomolecules-02-00435]\]. Mutations and loss of heterozygosity of the *COSMC* gene were described in STn-positive melanoma and colon cancer cell lines, as well as in tissue samples from two STn-positive cervical cancers \[[@B8-biomolecules-02-00435]\]. However, extensive studies of *COSMC* mutations in breast or colon cancers showed that these events were rare and could only account for some of the cases of STn expression in cancers \[[@B9-biomolecules-02-00435]\]. Thus, STn expression in cancer is most probably due to over-expression of ST6GalNAc I, with enhancing effects of increased synthesis of precursors (*i.e*., increased transfer of *O*-GalNAc) and decreased competition (*i.e*., decreased core-synthesis). 3. STn Immunodetection in Tissues and Tumors ============================================ 3.1. Anti-STn Antibodies ------------------------ Many anti-STn antibodies have been developed since the mid 80s (see [Table 1](#biomolecules-02-00435-t001){ref-type="table"} for the most used ones), each using different immunogens and displaying subtle differences in their affinity for STn antigen. These differences may account for conflicting results when comparing several studies reporting STn expression. biomolecules-02-00435-t001_Table 1 ###### Most used anti-STn monoclonal antibodies. Monoclonal antibody Immunogen Isotype Specificity Ref. --------------------- ----------------------------------------------- ------------ ---------------------------------------------------------------------- --------------------------------------------------------------- **B72.3** Membrane fraction of breast cancer metastasis Mouse IgG1 Clustered STn bound to serine. Cross reacts with Tn clusters \[[@B10-biomolecules-02-00435],[@B11-biomolecules-02-00435]\] **MLS102** LS 180 colonic cancer cells Mouse IgG Clustered STn, independently of the linkage to the peptide backbone. \[[@B12-biomolecules-02-00435],[@B13-biomolecules-02-00435]\] **TKH2** Ovine submaxillary mucin Mouse IgG Monomeric STn \[[@B14-biomolecules-02-00435],[@B15-biomolecules-02-00435]\] **HB-STn1** Ovine submaxillary mucin Mouse IgG1 \[[@B16-biomolecules-02-00435]\] 3.2. Expression in Normal Tissues --------------------------------- ### 3.2.1. Fetal Tissues {#sec3dot2dot1-biomolecules-02-00435} There are a few reports of STn detection in fetal organs such as esophagus and stomach \[[@B17-biomolecules-02-00435],[@B18-biomolecules-02-00435]\], pancreas \[[@B12-biomolecules-02-00435]\], colon (goblet cells) \[[@B17-biomolecules-02-00435],[@B19-biomolecules-02-00435]\], lung \[[@B18-biomolecules-02-00435]\], mammary gland \[[@B20-biomolecules-02-00435]\], and gonadal tissues from fetuses of both sexes \[[@B1-biomolecules-02-00435],[@B18-biomolecules-02-00435]\]. Furthermore, STn was detected in meconium and amniotic fluids \[[@B21-biomolecules-02-00435]\] suggesting a relation to digestive tract secretions. However, nothing is known about the biological role of STn during embryonic development. ### 3.2.2. Adult Healthy Tissues {#sec3dot2dot2-biomolecules-02-00435} STn expression was rarely systematically screened throughout normal adult tissues \[[@B14-biomolecules-02-00435],[@B17-biomolecules-02-00435],[@B22-biomolecules-02-00435]\]. However, many studies focusing on STn expression in cancer also reported staining performed on normal healthy tissues used as control. STn expression seems to be closely related to the upper digestive tract, since it was detected in salivary glands, esophagus and stomach. This expression is finely regulated and restricted to some specific cell types such as squamous cells in salivary glands and esophagus \[[@B17-biomolecules-02-00435],[@B22-biomolecules-02-00435],[@B23-biomolecules-02-00435]\], mucous acini in the submandibular gland and submucosae of the vestibular fold and soft palate \[[@B24-biomolecules-02-00435]\], and parietal and goblet cells in the stomach \[[@B25-biomolecules-02-00435],[@B26-biomolecules-02-00435],[@B27-biomolecules-02-00435]\]. In the colon, STn was visible in colonocytes only and only after the removal of *O*-acetyl groups by saponification \[[@B14-biomolecules-02-00435],[@B15-biomolecules-02-00435],[@B28-biomolecules-02-00435]\]. STn was also detected in bile ducts \[[@B17-biomolecules-02-00435],[@B22-biomolecules-02-00435]\] but never detected in normal hepatocytes \[[@B22-biomolecules-02-00435],[@B29-biomolecules-02-00435]\]. Likewise, STn antigen was never found in normal pancreas \[[@B10-biomolecules-02-00435],[@B14-biomolecules-02-00435],[@B17-biomolecules-02-00435],[@B22-biomolecules-02-00435],[@B30-biomolecules-02-00435],[@B31-biomolecules-02-00435],[@B32-biomolecules-02-00435]\]. Within the urogenital tract, STn was detected in uterine and cervix cells and always with a membranous apical staining \[[@B22-biomolecules-02-00435],[@B33-biomolecules-02-00435],[@B34-biomolecules-02-00435]\]. In contrast, normal ovarian epithelia were always found to be STn negative \[[@B35-biomolecules-02-00435],[@B36-biomolecules-02-00435],[@B37-biomolecules-02-00435]\] Some testis cells were occasionally found to be STn positive such as Leydig cells or interstitial cells \[[@B1-biomolecules-02-00435],[@B14-biomolecules-02-00435],[@B22-biomolecules-02-00435]\]. Finally, STn was found at the apical membrane of the ciliated cells of the lung \[[@B17-biomolecules-02-00435],[@B22-biomolecules-02-00435],[@B38-biomolecules-02-00435]\] and at the apical side of the ductal epithelium in normal breast tissue \[[@B39-biomolecules-02-00435],[@B40-biomolecules-02-00435],[@B41-biomolecules-02-00435]\]. Although quantified data are not properly available, overall, the authors reported STn expression in normal tissues to be rare and/or low compared to cancer tissues. Reports stating that STn expression is restricted to cell types specialized in secretion, apical surface of secreting cells or even content of lumen of secretory structure all suggest that sparse expression of STn in normal tissue relates to external fluids of the body. One may thus hypothesize that STn is unlikely to be naturally present in the bloodstream or the lymph. 3.3. STn Expression in Cancers ------------------------------ From the characterization of the first anti-STn monoclonal antibodies, it was rapidly discovered that STn antigen was over-expressed in cancer cells compared to the matching healthy cells; hence STn was described as an onco-fetal antigen. STn neo-expression or over-expression was reported in many epithelial cancers with highest frequencies in pancreas, colorectal and ovarian cancers ([Figure 2](#biomolecules-02-00435-f002){ref-type="fig"}). STn was therefore considered as a good tumor marker of carcinogenesis and potentially useful for diagnosis. However, we invite the reader to keep in mind that many parameters may account for the high variability in the frequencies reported by different authors. Amongst those factors are the distinct antibodies used, the protocol of staining, the method of scoring, the size of the cohort of the samples and the heterogeneity of cancer subtypes included in the respective studies. Some of these elements are included in [Figure 3](#biomolecules-02-00435-f003){ref-type="fig"} for a more comprehensive reading. {ref-type="fig"} for references). The red bar is the mean of the percentage of STn positive cases.](biomolecules-02-00435-g002){#biomolecules-02-00435-f002} ### 3.3.1. Early in Carcinogenesis STn was reported to be over-expressed in several epithelial benign lesions considered to be potential precursors of cancers, such as esophageal dysplastic squamous epithelia \[[@B23-biomolecules-02-00435]\], gastric intestinal metaplasia \[[@B42-biomolecules-02-00435]\], colonic moderate dysplasia \[[@B43-biomolecules-02-00435]\], lung atypical adematous hyperplasia \[[@B38-biomolecules-02-00435]\], or breast ductal hyperplasia and apocrine metaplasia \[[@B40-biomolecules-02-00435],[@B44-biomolecules-02-00435],[@B45-biomolecules-02-00435]\]. In these different cases, STn over-expression was primarily observed at the apical or luminal cell surface as well as in the corresponding healthy tissues. More unexpectedly, STn was also reported in benign lesions in pancreas \[[@B46-biomolecules-02-00435]\] and ovaries \[[@B35-biomolecules-02-00435],[@B36-biomolecules-02-00435],[@B37-biomolecules-02-00435]\], two tissues that are devoid of STn expression in the healthy state. Notably, STn staining was observed in pancreatic intraepithelial neoplasia stage III (PanIN3), the last histological grade relevant to benign tumor before the tumor becomes invasive, but not in earlier stages \[[@B46-biomolecules-02-00435],[@B47-biomolecules-02-00435]\]. All together, these observations suggest that STn over-expression occurs earlier in carcinogenesis in tissues that normally express the antigen than in the others. There is a body of evidence linking STn expression to inflammatory diseases of the stomach (gastritis) or the colon (ulcerative colitis and Crohn's colitis). In gastritis, STn was detected in 50--100% of the cases \[[@B48-biomolecules-02-00435],[@B49-biomolecules-02-00435]\] and it appeared to be independent on the presence of *Helicobacter pylori*, a common bacterial parasite of the stomach \[[@B48-biomolecules-02-00435]\]. In ulcerative colitis, de-*O*-acetylated STn was shown to be an independent marker of the dysplasia-carcinoma sequence \[[@B50-biomolecules-02-00435]\]. De-*O*-acetylated STn was also detected in 44% of the cases of Crohn's colitis, another inflammatory disease associated with colon cancer risks \[[@B51-biomolecules-02-00435]\]. These reports suggest that STn expression might be regulated by inflammatory signals, such as γ-interferon \[[@B51-biomolecules-02-00435]\], in pre-malignant lesions of these two organs. Whether or not such signals could trigger STn expression in other organs remains to be determined. ### 3.3.2. Cytological Types of Carcinomas Epithelial cancers, or carcinomas, are divided into two major types depending on the cytology of the cancer cells: adenocarcinomas and squamous cell carcinomas. Adenocarcinomas (AC) arise from glandular epithelia and often secrete gland related molecules (e.g., mucins). When these secretions are abundant, AC are classified as mucinous carcinomas. By contrast, squamous cell carcinomas (SCC) are formed of thin, flat and poorly secreting cells. Cancers arising in the pancreas, ovary, colon, breast, stomach and liver are more likely to be AC, whereas cancers of the cervix or esophagus are more often SCC. Non-small cell lung cancers may be either AC or SCC and arise from a common precursor cell located in the basal bronchial epithelium. The cellular pattern of STn expression is different according to the cancer cell morphology. Mucinous cancer cells are stained on the whole cell membrane, focally in the cytoplasm (perinuclear area, presumably Golgi) and sometimes in extracellular compartments \[[@B19-biomolecules-02-00435],[@B39-biomolecules-02-00435],[@B52-biomolecules-02-00435]\]. This indicates that STn is carried by both membrane bound and secreted glycoproteins such as mucins. The role of mucins as STn carriers in mucinous cancers has recently been confirmed using the Proximity Ligation Assay \[[@B53-biomolecules-02-00435]\]. Squamous cancer cells are mainly and intensively stained in the cytoplasm and sometimes on the cell membrane \[[@B23-biomolecules-02-00435],[@B52-biomolecules-02-00435],[@B54-biomolecules-02-00435]\]. This supports the idea that STn antigen is carried by a diversity of glycoproteins differently expressed and/or processed according to the cell type. Furthermore, it appears that STn is more often detected in AC than in SCC in both cervix \[[@B16-biomolecules-02-00435]\] and lung cancers \[[@B38-biomolecules-02-00435],[@B55-biomolecules-02-00435]\]. Because STn seems to be carried by different glycoproteins and in different compartments in AC and SCC, its putative biological function in cancer development might be different in these tumor types. ![STn frequency in various carcinomas. The diagram shows the percentage of STn positive cases detected in various carcinomas using different anti-STn mAbs. The average percentage of positive cases is indicated below the type of cancer. The numbers on the top of the bars indicate the number of samples used for each study. References are indicated at the base of each bar, including \[[@B56-biomolecules-02-00435],[@B57-biomolecules-02-00435],[@B58-biomolecules-02-00435],[@B59-biomolecules-02-00435],[@B60-biomolecules-02-00435],[@B61-biomolecules-02-00435],[@B62-biomolecules-02-00435]\] that are not quoted elsewhere in the text. SCC: Squamous cell carcinoma, AC: adenocarcinoma.](biomolecules-02-00435-g003){#biomolecules-02-00435-f003} ### 3.3.3. Cell Differentiation Loss of cell differentiation, often participating in a high histological grade classification, was reported to qualitatively modulate the STn expression. In colonic poorly differentiated cells, STn cytoplasm staining is more frequent than in well differentiated cells \[[@B19-biomolecules-02-00435],[@B63-biomolecules-02-00435]\]. There is also a redistribution of antigen expression to the whole cell membrane, including the basolateral surface \[[@B43-biomolecules-02-00435],[@B63-biomolecules-02-00435]\]. Loss of differentiation was reported to be associated with a decreased frequency of STn positive cases in pancreatic \[[@B46-biomolecules-02-00435]\] and colorectal cancers \[[@B43-biomolecules-02-00435]\]. Inversely, a tendency towards increased STn expression frequency in poorly differentiated cells was observed in liver and bladder cancers \[[@B29-biomolecules-02-00435]\]. Finally, the histological grade does not seem to affect STn expression in gastric \[[@B25-biomolecules-02-00435],[@B64-biomolecules-02-00435],[@B65-biomolecules-02-00435]\] or breast cancer \[[@B66-biomolecules-02-00435],[@B67-biomolecules-02-00435]\]. All these observations are subjected to differences in histological grading of each tumor type, which may reflect different biological events for the cancer cells, depending on the organ considered. However, using nasopharyngeal cell lines, Lin *et al*. recently showed that STn expression was associated with epithelial to mesenchymal transition \[[@B68-biomolecules-02-00435]\], a loss of differentiation that is an important milestone towards cancer metastasis. ### 3.3.4. Heterogeneity within Tumors STn expression is generally reported to be heterogenous in tumors with proportions of STn-positive cells ranging from 5% to 100% (rarely 100%). This phenomenon is consistent whatever the origin of the tumor: stomach \[[@B25-biomolecules-02-00435],[@B26-biomolecules-02-00435]\], colon \[[@B43-biomolecules-02-00435]\], ovary \[[@B52-biomolecules-02-00435],[@B69-biomolecules-02-00435]\], cervix \[[@B33-biomolecules-02-00435]\] and breast \[[@B39-biomolecules-02-00435],[@B70-biomolecules-02-00435],[@B71-biomolecules-02-00435]\]. Various clinical features were correlated with the degree of heterogeneity of STn expression. For example, Federici *et al.* reported that ovarian mucinous cancers were more likely to be uniformly stained than ovarian serous cancers \[[@B52-biomolecules-02-00435]\]. Lopez-Ferrer *et al*. also observed that the percentage of STn positive cells was higher in lung AC (34%) than in SCC (9%) \[[@B55-biomolecules-02-00435]\]. In gastric cancers, an increased percentage of STn positive cells was correlated to deeper invasion and advanced stage \[[@B25-biomolecules-02-00435]\]. Flucke *et al*. reported that patients with more than 35% of stained cells in their esophageal SCC (25% of cases) had a decreased overall survival, compared to the low expressing group (\<35% positive cells) \[[@B54-biomolecules-02-00435]\]. In ovarian cancers, Davidson *et al*. observed that STn expression was sometimes more intense at the invasive front of the tumor and that the percentage of STn positive cells was higher in effusions than in the matching primary tumors \[[@B69-biomolecules-02-00435]\]. We reported a similar pattern of expression at the periphery of the tumors in a model of breast cancer cells injected as xenograft in SCID mice \[[@B6-biomolecules-02-00435]\]. Because in this model the cells were all derived from a selected clonal population, the heterogeneity of STn was assumed to be related to the expression of the protein(s) that carried the glycan. In other words, STn expression could be regulated in the tumor, *via* the regulation of its carrier(s). All together, these observations suggest that STn expression could be correlated with the invasive and aggressive potential of epithelial cancer cells, when expressed at the right time and right place. ### 3.3.5. Association with Invasiveness For a majority of the authors, STn expression detected in tissue and/or sera samples of patients with gastric cancers was correlated with depth of invasion \[[@B25-biomolecules-02-00435],[@B42-biomolecules-02-00435],[@B72-biomolecules-02-00435],[@B73-biomolecules-02-00435],[@B74-biomolecules-02-00435],[@B75-biomolecules-02-00435],[@B76-biomolecules-02-00435]\], lymph vessel and venous invasion \[[@B65-biomolecules-02-00435],[@B72-biomolecules-02-00435],[@B73-biomolecules-02-00435]\] and peritoneal dissemination \[[@B25-biomolecules-02-00435],[@B73-biomolecules-02-00435]\]. Ikeda *et al*. reported that stromal STn detection was associated with peritoneal dissemination \[[@B77-biomolecules-02-00435]\]. In accordance, Ozaki *et al*. recently reported that STn expression increased peritoneal metastasis in a model of human gastric cell lines transplanted in nude mice \[[@B78-biomolecules-02-00435]\]. In samples from ovarian cancer patients, STn positive cells were more frequently observed at the invasion front of tumors and in peritoneal and pleural effusions, but less often in metastatic lesions than in primary tumors \[[@B69-biomolecules-02-00435],[@B79-biomolecules-02-00435]\]. These results suggest that in ovarian cancers, STn enhances the dissemination of cells, facilitating primary tumor/effusion transition, but does not improve the settlement of metastatic cells in distant organs. However, in colorectal cancers, STn expression was reported not to be correlated with depth of invasion \[[@B15-biomolecules-02-00435],[@B80-biomolecules-02-00435]\]. Similarly, Schmitt *et al*. have reported that breast ductal invasive carcinomas are less frequently STn positive than ductal carcinomas *in situ* \[[@B40-biomolecules-02-00435]\]. Thus, sparse data for colorectal and breast cancer suggest that the effect of STn expression on the invasiveness of tumor cells might be an organ-specific phenomenon. However, *in vitro* characterization of engineered breast cancer cell lines showed that STn expression induced a decreased adhesion, a decreased aggregation and an increased cell motility, all consistent with an increased invasiveness \[[@B6-biomolecules-02-00435],[@B81-biomolecules-02-00435]\]. Whether or not these observations are relevant at the disease level for breast or colorectal cancers would require further investigations. ### 3.3.6. Detection in Serum Detection of tumor markers in serum is a simple, non-invasive and sensitive method for diagnosis or post surgery follow-up of the patients. This is particularly useful for the care of patients with cancers in deep organs (stomach, colon, pancreas, biliary tract, ovaries or cervix), which are often asymptomatic at the earlier stage. The presence of STn antigen in serum is due to important *O*-glycoprotein secretion or to cell shedding from tumors into the bloodstream, both requiring a large amount of cancer cells in the primary tumor. This critical tumor mass is usually found in advanced cancers, which are more likely to be of poor prognosis. A high level of STn (cut-offs ranging from 38 to 50 U/ml) was detected in sera of patients to various degrees, depending on the cancer type: twenty-eight to eighty-six percent in gastric cancers \[[@B72-biomolecules-02-00435],[@B73-biomolecules-02-00435],[@B82-biomolecules-02-00435]\], 11%--28% in colorectal cancers \[[@B82-biomolecules-02-00435],[@B83-biomolecules-02-00435]\], 40%--55% in pancreatic cancers \[[@B82-biomolecules-02-00435],[@B84-biomolecules-02-00435]\], 25%--53% in biliary tract cancers \[[@B82-biomolecules-02-00435],[@B84-biomolecules-02-00435]\], 29%--69% in ovarian cancers \[[@B36-biomolecules-02-00435],[@B85-biomolecules-02-00435],[@B86-biomolecules-02-00435],[@B87-biomolecules-02-00435]\] and 15% in cervical cancers \[[@B36-biomolecules-02-00435]\]. Unsurprisingly, a high level of serum STn was significantly associated with tumor size, lymph node, liver metastasis and advanced stages in gastric cancers \[[@B72-biomolecules-02-00435],[@B73-biomolecules-02-00435],[@B76-biomolecules-02-00435]\]. Furthermore, high serum STn was found to be associated with a decreased overall survival of patients with gastric \[[@B72-biomolecules-02-00435],[@B88-biomolecules-02-00435],[@B89-biomolecules-02-00435]\], colorectal \[[@B83-biomolecules-02-00435],[@B90-biomolecules-02-00435]\] or ovarian \[[@B86-biomolecules-02-00435],[@B87-biomolecules-02-00435]\] cancers. Thus, STn detection in serum is more useful for prognosis than for diagnosis and is usually considered as a poor prognosis marker. ### 3.3.7. Prognosis Value of STn in Cancers Because of its putative role in cancer cell invasion and spreading, there is a consensus to say that STn expression is associated with an adverse outcome, such as lymph node or distant metastasis and decreased overall survival of the patients. However, this may not be true for all types of cancers. For example, STn was not associated with lymph node invasion or overall survival in cervical cancer \[[@B16-biomolecules-02-00435]\] or survival in lung cancer \[[@B91-biomolecules-02-00435]\]. In esophageal cancers, Flucke *et al*. reported an association with decreased survival, disagreeing with Ikeda *et al*. while they both reported an absence of correlation with lymph node metastasis or TNM staging \[[@B54-biomolecules-02-00435],[@B92-biomolecules-02-00435]\]. However, there are only few studies concerning these types of cancer, and extensive investigations might prove to be more insightful. Furthermore, there are no data available concerning the prognostic value of STn in pancreas or liver cancers. These gaps are probably due to the fact that STn expression is very frequent in pancreatic cancers (≈90%) and very rare in hepatic ones (≈10%) ([Figure 2](#biomolecules-02-00435-f002){ref-type="fig"}, [Figure 3](#biomolecules-02-00435-f003){ref-type="fig"} and [Figure 4](#biomolecules-02-00435-f004){ref-type="fig"}), making the design of an unbiased cohort difficult for statistical analysis. Nonetheless, as more studies are available we can draw an overall picture of the prognostic value of STn in gastric, colorectal, ovarian and breast cancers. {ref-type="sec"}. and [Section 3.2.2](#sec3dot2dot2-biomolecules-02-00435){ref-type="sec"}. In colorectal cancer, *O*-acetylated STn is considered to be a positive report since it requires the activity of ST6GalNAc I (see Chapter 2.). The forth line summarizes [Figure 2](#biomolecules-02-00435-f002){ref-type="fig"} and [Figure 3](#biomolecules-02-00435-f003){ref-type="fig"}. Bottom part: reports assessing the correlation between STn expression and clinical features of cancers, with a significant correlation found in pink and no correlation found in white. Size of the circles represents the number of published reports (diameter calculated as 1 + 0.1 unit per publication).](biomolecules-02-00435-g004){#biomolecules-02-00435-f004} In gastric cancers, numerous studies reported a positive association of STn expression (in serum or primary tumors) with lymph node metastasis \[[@B25-biomolecules-02-00435],[@B73-biomolecules-02-00435],[@B75-biomolecules-02-00435],[@B88-biomolecules-02-00435],[@B89-biomolecules-02-00435],[@B93-biomolecules-02-00435]\], peritoneal metastasis \[[@B25-biomolecules-02-00435],[@B73-biomolecules-02-00435],[@B77-biomolecules-02-00435],[@B88-biomolecules-02-00435]\] and liver metastasis \[[@B73-biomolecules-02-00435],[@B88-biomolecules-02-00435]\] with few contradicting reports \[[@B65-biomolecules-02-00435]\]. Logically, many authors also reported a correlation between STn expression and decreased survival \[[@B26-biomolecules-02-00435],[@B65-biomolecules-02-00435],[@B72-biomolecules-02-00435],[@B75-biomolecules-02-00435],[@B88-biomolecules-02-00435],[@B94-biomolecules-02-00435]\]. In colorectal cancers, few authors reported STn association with distant metastasis \[[@B95-biomolecules-02-00435]\]. Reflecting this weak relationship of STn to invasiveness, the association of STn with decreased survival was reported by some authors \[[@B27-biomolecules-02-00435],[@B80-biomolecules-02-00435],[@B83-biomolecules-02-00435],[@B90-biomolecules-02-00435],[@B96-biomolecules-02-00435]\] but contradicted by others \[[@B95-biomolecules-02-00435],[@B97-biomolecules-02-00435],[@B98-biomolecules-02-00435],[@B99-biomolecules-02-00435]\]. STn association with poor prognosis is clearer in ovarian cancers since all authors concurringly reported a decreased survival for patients expressing high levels of STn in tissues or serum \[[@B35-biomolecules-02-00435],[@B86-biomolecules-02-00435],[@B87-biomolecules-02-00435]\]. However, the way STn may affect survival is somewhat elusive, since STn may not be involved in the metastatic process of ovarian cancers \[[@B69-biomolecules-02-00435]\]. Finally, STn over-expression was reported to correlate with negative estrogen receptor (ER) status in breast cancer \[[@B67-biomolecules-02-00435],[@B100-biomolecules-02-00435]\], and ER-negative breast cancers are known to be more aggressive than ER-positive ones. However, one paper reported an association of STn with lymph node invasion \[[@B100-biomolecules-02-00435]\] while two others disagreed \[[@B66-biomolecules-02-00435],[@B67-biomolecules-02-00435]\]. Regarding survival, authors reported some association of STn with poor prognosis depending on various clinico-pathological parameters. For instance, Miles *et al*. reported a significant correlation of STn expression with decreased survival for patients with lymph node metastasis \[[@B71-biomolecules-02-00435]\]. Kinney *et al*. concluded that STn was an independent predictor of recurrences \[[@B70-biomolecules-02-00435]\]. Leivonen *et al*. reported a decreased survival of STn patients in the short-term (five years) but not in the long-term (\>15 years) \[[@B67-biomolecules-02-00435]\], whereas Imai *et al.* did not find any correlation at all \[[@B66-biomolecules-02-00435]\]. So, from a general point of view, it seems that STn expression is mostly associated with an adverse outcome. However, the numerous discrepancies and subtleties reviewed above suggest that if STn exerts any biological function promoting cancer development, it occurs through various mechanisms depending on each cancer type or sub-type. 4. Immunotherapy Targeting STn ============================== The exploitation of the patient's own immune system, i.e. immunotherapy, to control cancer growth rises as an attractive approach, offering the potential of enhancing the effects of conventional treatment such as post-surgery radiotherapy or chemotherapy, without substantially increasing toxicity. While boosting immune responses against cancer is obviously not sufficient to eradicate a solid tumor by itself, it might be crucial in the elimination of minimal residual disease or micrometastases following primary treatment or in the prevention of transformation from a *benign* to a precancerous *lesion*. There are several current immunotherapeutic approaches to treat cancer that have been approved for use in humans, such as cytokines (e.g., interleukin-2 \[[@B101-biomolecules-02-00435]\]), which are able to modulate the leukocyte function, and monoclonal antibodies (e.g., Cetuximab), which recognize tumor antigens and hence mount a specific immune response against cells expressing them. Vaccination or immunization is a promising immunotherapeutic approach that is being evaluated in a variety of different tumor types, targeting different antigens, in an attempt not only to boost anti-tumor immune responses, but also to develop immunologic memory, producing long-lived protection lasting for years or even decades. The side effects associated with immunization are mild compared with the conventional cytotoxic therapies. However, contrarily to microbial antigens, tumor-associated antigens may be considered self-antigens and the use of cancer vaccines may therefore easily result in enhanced activity of self-reactive lymphocytes and autoimmune responses. Thus, the designation of right target antigens is essential for successful tumor vaccination. The finding of the STn antigen as a good tumor marker and, in particular, the fact that normal adult epithelial cells seem to never expose STn epitope to the bloodstream and immune system, turns it into an interesting antigen for immunization. So far, few strategies have been developed as anti-STn vaccine, such as Theratope \[[@B102-biomolecules-02-00435],[@B103-biomolecules-02-00435],[@B104-biomolecules-02-00435],[@B105-biomolecules-02-00435],[@B106-biomolecules-02-00435]\] which has been pointed out for its use in clinical trials \[[@B107-biomolecules-02-00435]\]. 4.1. Theratope -------------- The Theratope vaccine consists of a synthetic construct of STn disaccharide conjugated to the Keyhole limpet hemocyanin (KLH) that has been designed by the biotech company Biomira (now Oncothyreon, Alberta, Canada) \[[@B108-biomolecules-02-00435]\]. KLH was chosen as carrier protein because of its apparent safety, while stimulating antibody production and T cell activation. In addition, KLH offers a great carrier capacity as one mole of KLH can be substituted with approximately 3000 moles of glycan hapten, which is crucial to better simulate the common high STn density on clusters of mucins found on tumor cells. ### 4.1.1. Theratope in Clinical Trials Theratope was initially designed for use in metastatic breast cancer. In 1996, the results of Phase-II studies in women with metastatic breast cancer were published \[[@B109-biomolecules-02-00435],[@B110-biomolecules-02-00435],[@B111-biomolecules-02-00435],[@B112-biomolecules-02-00435]\] and abundantly reviewed some years later \[[@B104-biomolecules-02-00435],[@B105-biomolecules-02-00435],[@B106-biomolecules-02-00435],[@B113-biomolecules-02-00435],[@B114-biomolecules-02-00435],[@B115-biomolecules-02-00435],[@B116-biomolecules-02-00435],[@B117-biomolecules-02-00435],[@B118-biomolecules-02-00435]\]. In these clinical trials, Theratope immunotherapy was well tolerated by vaccinated patients with minimal toxicity \[[@B110-biomolecules-02-00435],[@B111-biomolecules-02-00435]\]. Patients receiving Theratope had a significant improved survival by 12.1 months and developed anti-STn humoral immune responses \[[@B113-biomolecules-02-00435]\]. These data enlightened the relevance of this antigen as tumor-specific antigen, and its safety showed that the vaccine was unable to break immune tolerance to other self-antigens and was not generating adverse autoimmune responses. Patients with metastatic breast cancer receiving low-dose intravenous cyclophosphamide (inhibitor of suppressor T cells) before vaccinations showed longer survival and generated higher antibody titers than control patients \[[@B110-biomolecules-02-00435],[@B111-biomolecules-02-00435]\]. In a single-arm Phase II study, the Theratope was used in 95 women with metastatic breast cancer undergoing therapy with aromatase inhibitors or estrogen receptor antagonists \[[@B119-biomolecules-02-00435]\]. The primary objective of this trial was to determine tolerability and immune responses elicited by the vaccine in patients with hormone-sensitive breast cancer who did not require chemotherapy. Another combined therapy was tested on patients with advanced breast or ovarian cancers. Patients were treated with high-dose chemotherapy, followed by autologous stem cell rescue to restore the immune cells, and then vaccinated with Theratope. Interestingly, the patients developed T cell responses against cancer cells, as demonstrated by the lytic activity against STn positive cancer cell lines. The cancer vaccine was well tolerated in patients after autologous transplant and, while not highly significant, the trends in data supported the concept that the Theratope vaccine decreased the risk for relapse and death \[[@B109-biomolecules-02-00435],[@B120-biomolecules-02-00435]\]. Regarding the combination with chemotherapy, a Phase II study assessed the efficacy and safety of Theratope in patients with colorectal cancer receiving first-line chemotherapy \[[@B121-biomolecules-02-00435]\]. The results showed that patients were capable of mounting an immune response to the vaccine while receiving concurrent chemotherapy. Side effects, such as mild injection site reactions with ulcerations were reported. Nevertheless, this trial was the first describing the efficacy of using the Theratope vaccine in combination with chemotherapy. Interestingly, the ability of Theratope to stimulate an immune response was not affected by chemotherapy and almost all patients showed IgG responses \[[@B121-biomolecules-02-00435]\]. The encouraging results led to the fulfilment of a randomized, double-blind, Phase III trial across 126 centers in 10 countries, involving 1030 women with metastatic breast cancer. Nonetheless this trial failed to demonstrate that Theratope improved median time to disease progression or overall patient survival \[[@B107-biomolecules-02-00435]\]. The potential explanation for the failure of the Phase III clinical trial is that the patient population was not evaluated for STn expression prior to enrolment, possibly masking any benefit from the vaccine due to heterogeneous STn expression between patients. At the same time as it was concluded that the vaccine did not increase survival in patients with metastatic breast cancer disease, the analysis of a pre-stratified subset of patients who were receiving hormonal therapy showed a significant difference in median overall survival, with an increased survival by 6.5 months when treated with Theratope \[[@B122-biomolecules-02-00435]\]. This was in some way unanticipated since it had been reported that STn expression was either not correlated with hormone receptor positivity \[[@B39-biomolecules-02-00435],[@B66-biomolecules-02-00435]\] or was associated to hormonal receptor negativity \[[@B67-biomolecules-02-00435]\]. Moreover, STn is associated with loss of cell differentiation, while hormonal receptor expression is known to be correlated with a differentiated state of breast tumor cells \[[@B123-biomolecules-02-00435],[@B124-biomolecules-02-00435]\]. Thus, the reasons why Theratope seemed to be only efficient in this subset of patients remain unclear. While additional investigations are needed to clarify the discrepant results concerning the Theratope clinical trials, the overall results point to a relative efficacy and safety of this anti-STn immunotherapy. ### 4.1.2. Highlights on the Theratope Mechanism With the purpose of elucidating the mechanisms involved in the Theratope efficacy, some noteworthy studies have been performed \[[@B125-biomolecules-02-00435],[@B126-biomolecules-02-00435]\]. Braun *et al*. centered on the results from the Phase III clinical trial showing a prolonged survival in Theratope-vaccinated patients who were treated with concomitant hormone therapy, and they built the hypothesis that the immune response elicited in Theratope-treated patients could collaborate with the effects of hormonal therapy, improving anti-tumor responses. These authors proved that, by an undescribed mechanism, tumor cells lines when treated with aromatase inhibitors exhibit increased sensitivity to both monocyte-mediated and antibody-dependent cellular cytotoxicity. Therefore the hormone-based treatment may collaborate with anti-tumor antibodies (e.g., anti-STn antibodies induced by Theratope) to produce improved tumor control in patients \[[@B125-biomolecules-02-00435]\]. In another study, using a murine model of breast cancer, Burchell's group also clearly demonstrated that Theratope-induced tumor protection was dependent on the quantity of anti-STn antibodies raised by immunization \[[@B126-biomolecules-02-00435]\]. Interestingly, anti-STn antibodies were able to recognize a wide range of STn-carrying glycoproteins, such as osteopontin expressed by mammary carcinomas, suggesting that a response to multi-targets expressing STn was important to induce tumor protection. It is widely known that anti-tumor antibodies could delay tumor growth by antibody-dependent cellular cytotoxicity, inhibition of function or a combination of the two mechanisms. Pre-clinical and clinical studies all showed that immunization with Theratope usually induces STn-specific IgGs \[[@B107-biomolecules-02-00435],[@B118-biomolecules-02-00435],[@B126-biomolecules-02-00435]\], including the IgG2a subtype, known to mediate antibody-dependent cellular cytotoxicity in mouse models \[[@B127-biomolecules-02-00435]\]. While this could explain the observed delay in tumor growth in mice and the increased time to progression in patients \[[@B107-biomolecules-02-00435],[@B126-biomolecules-02-00435]\], little is known about the role of the Theratope-induced antibodies in the inhibition of tumor function. However, Blixt *et al*. recently reported that the detection of high titers of auto-antibodies directed against MUC1 cancer specific glycoforms, including MUC1-STn, in early stage breast cancer patients was associated with increased time to metastasis, supporting the protective role of anti-STn antibodies \[[@B128-biomolecules-02-00435]\]. 4.2. What Can Be Improved in Anti-STn Approaches? ------------------------------------------------- ### 4.2.1. Inducing Better Immune Responses Based on the data referred to above, it is widely assumed that if a vaccine could elicit a strong immune response towards STn, it can potentially exert increased protective effects in the host against cancer. However, there are significant immunological challenges to develop an active anti-tumor immunotherapy based on STn antigens, due to its low immunogenicity \[[@B129-biomolecules-02-00435],[@B130-biomolecules-02-00435]\]. Tumor immunology involves two main interrelated mechanisms, the cellular immune response and the humoral response. Cell-mediated immunity involves the activation of a variety of immune cells, including antigen presenting cells, such as dendritic cells (DCs) and macrophages, which uptake antigens and then present a small portion of peptide antigens (epitopes) to activate specific CD4^+^ helper T cells (Th) and CD8^+^ cytotoxic T cells \[[@B131-biomolecules-02-00435]\]. The T cell receptors (TCR) are restricted to recognizing antigenic peptides only when presented in the context of MHC, by the antigen presenting cells ([Figure 5](#biomolecules-02-00435-f005){ref-type="fig"}). DCs are pivotal due to their role in regulating both innate and adaptive immune responses, their migratory capacity, and by dictating whether the course of the immune responses will be tolerogenic or immunogenic \[[@B132-biomolecules-02-00435],[@B133-biomolecules-02-00435]\]. The humoral response, in contrast, is mediated by B cells that secrete antibodies, upon cross-linking of their receptors (B cell receptors) with specific epitopes. IgM is the first immunoglobulin isotype to appear, but due to its relatively low affinity to antigens and short time period, it provides only a quick, short-term protection. Only when B cells are able to receive stimulatory signals from Th cells (T cell-dependent B cell activation), the switching of antibody subtypes from IgM to high-affinity IgGs takes place ([Figure 5](#biomolecules-02-00435-f005){ref-type="fig"}). These high-affinity IgG antibodies can bind with the target cancer cells, marking them for destruction by either the complement or antibody-dependent cell-mediated cytotoxicity. The T cell-dependent B cell activation is essential for the generation of significant memory B cells and for long-lasting humoral responses \[[@B131-biomolecules-02-00435],[@B134-biomolecules-02-00435],[@B135-biomolecules-02-00435]\]. {#biomolecules-02-00435-f005} Carbohydrates when administered alone activate the B cells through a T cell-independent mechanism ([Figure 5](#biomolecules-02-00435-f005){ref-type="fig"}), as by themselves they cannot be presented by MHC and then recognized by T cells. Without additional help from Th cells, no high-affinity IgG antibodies can be generated. Yet, when carbohydrates are conjugated to a protein carrier, peptide presentation takes place, providing the activation of B cells through a T cell-dependent mechanism \[[@B136-biomolecules-02-00435],[@B137-biomolecules-02-00435]\]. Thus, in order to induce T cell-dependent B cell activation, the STn antigen has been covalently linked with an immunogenic carrier, such as KLH in Theratope, which contains protein epitopes that can be presented by antigen presenting cells and be subsequently recognizable by the Th cells. Apparently and given the reported increased titers of IgG1 and IgG2a but also of IgM, it may be hypothesized that both T cell-dependent and -independent B cell activation occur during Theratope immunization. The chemical conjugation of STn with other protein or peptide carriers may probably improve the switch to T cell-dependent activation, increasing the IgG/IgM ratios, thus leading to improved anti-tumor responses. Notably, in the beginning of 90s, partially desialylated ovine submaxillary mucins, which contained Tn and STn antigens were used to immunize patients with colorectal cancer \[[@B138-biomolecules-02-00435]\] and induced both IgM and IgG antibodies against STn, thus proving the immunogenicity of this type of conjugate. Remarkably, glycopeptides bearing multiple Tn epitopes also proved to be immunogenic, inducing anti-STn antibodies without the use of a carrier protein \[[@B139-biomolecules-02-00435]\]. However, improved titers were obtained with glycopeptides bearing Tn and/or STn that were conjugated with carriers such as KLH \[[@B140-biomolecules-02-00435],[@B141-biomolecules-02-00435]\] or tetanus toxoid \[[@B142-biomolecules-02-00435]\]. However, *in vivo* studies showed that the humoral responses induced by such peptides were still not sufficient to provide protection from a tumor challenge \[[@B126-biomolecules-02-00435]\]. Comparative studies evaluating the humoral responses that are elicited based on the type of STn carrier are still necessary. Interestingly, we have found that desialylation of DCs potentiates the immune responses they triggered, in particular their unique capacity to prime Th cells \[[@B143-biomolecules-02-00435]\]. Concordantly, cancer cells expressing STn antigen tend to inhibit DC maturation and hinder subsequent T cell activation, when compared with parental cells lines with absent or low expression of STn \[[@B144-biomolecules-02-00435]\] suggesting that STn-expressing cancer cells are prone to cause immune tolerance. Thus, to the immunological challenge of anti-STn immunization, one should include the identification of tools to fine tune the innate response and thus surpassing tolerogenic responses. Some approaches have actually been performed in that sense, which include the concomitant use of selected adjuvants and the use of Toll like receptor agonists such as CpG and BCG. However, many lessons are still to be learnt from vaccines against other tumor associated antigens, whose results have been recently reviewed \[[@B145-biomolecules-02-00435],[@B146-biomolecules-02-00435]\] and therefore are not included in this article. As referred to above, the antigen structure is crucial for tuning the immune responses during immunization and carbohydrate antigens, as the STn antigens have to be coupled to a relevant protein or peptide carrier to trigger long-lasting immune responses. The identification of the physiological STn-positive proteins present in cancer cells is another key piece to improve anti-STn immunization. On one hand, STn-positive proteins can be used as immunogen in vaccination protocols, expanding the immunization targets while providing means to foment T cell-dependent B cell activation. On the other hand, STn-positive proteins offer the possibility to target more specific pathological mechanisms. ### 4.2.2. Co-targeting STn and Relevant Glycoproteins So far the available information relies on human cell lines and of the various STn-positive proteins present, four have been fully confirmed: MUC1 and CD44 in human \[[@B6-biomolecules-02-00435],[@B78-biomolecules-02-00435]\], integrin β1 \[[@B147-biomolecules-02-00435]\] and osteopontin \[[@B126-biomolecules-02-00435]\] in mice. All of these molecules have described functions in adhesion and/or migration and chemotaxis, and it has been suggested that the modification of their STn content may influence their function and the mobility of the cancer cells \[[@B148-biomolecules-02-00435]\]. A comprehensive compilation of our current knowledge about the role of STn-positive glycoproteins and their potential effect on tumor behavior has been recently reviewed \[[@B148-biomolecules-02-00435]\]. While, the physiological role of STn on tumor cell behavior remains obscure, targeting these glycoproteins is more likely to interfere with specific mechanisms involved in tumor development. Moreover the targeting of multiple epitopes may be advantageous in overcoming the problem of tumor escape that has been documented in a number of clinical studies \[[@B149-biomolecules-02-00435]\]. There has been particular interest in MUC1 as an immunotherapeutic target. MUC1 are expressed on the luminal surface of epithelial cells and act in cell adhesion and signaling \[[@B150-biomolecules-02-00435]\]. However, on many carcinomas, MUC1 is over-expressed and found on all cell surfaces. Here, MUC1 also display aberrant *O*-glycosylation patterns including STn \[[@B151-biomolecules-02-00435]\], causing different B and T cell-specific antigenic epitopes to be exposed. A limited number of MUC1-based vaccines are now being used and evaluated in advanced clinical trials. PANVAC-VF and MVA-MUC1-IL2 are virus-based vaccines expressing MUC1 and T cell co-stimulatory molecules (B7.1, intracellular adhesion molecule-1 and leukocyte function-associated antigen-3) \[[@B152-biomolecules-02-00435]\] or interleukin-2 \[[@B153-biomolecules-02-00435]\], respectively. Stimuvax is a liposome-based vaccine composed of a synthetic MUC1 peptide, coupled with the adjuvant monophosphoryl lipid A, evaluated in NSCLC \[[@B154-biomolecules-02-00435]\]. ImMucin, which is being evaluated in multiple myeloma patients, consists of a synthetic peptide composed of the entire signal peptide domain of the MUC1 that is expressed only on tumor cells in association with MHC molecules, thereby ensuring specific anti-cancer activity \[[@B155-biomolecules-02-00435]\]. In these trials, there is a clear concern not only over targeting the MUC1 tumor antigens, but also over boosting either innate and/or adaptive immune responses. However, no attention has been given to the modifications that MUC1 may have related with STn or other *O*-glycosylation types. These variables would be important to consider since specific glycoforms of MUC1, bearing Tn, STn, sialyl-Lewis^a^ and/or sialyl-Lewis^x^ antigens have been detected in different types of cancer \[[@B156-biomolecules-02-00435],[@B157-biomolecules-02-00435],[@B158-biomolecules-02-00435],[@B159-biomolecules-02-00435]\]. Specific glycoforms of MUC1 or MUC1 glycopeptides distinctively modulate the immune response that is set up against tumors \[[@B160-biomolecules-02-00435]\]. This may be due to the fact that MUC1 glycopeptides can be presented to cytotoxic T cells, and glycans are integral parts of their TCR defined epitopes \[[@B160-biomolecules-02-00435]\]. However, densely glycosylated MUC1 glycopeptides are unable to be processed by antigen-presenting cells \[[@B161-biomolecules-02-00435]\], compromising antigen presentation and, consequently, T cell activation. By contrast, glycopeptides carrying the Tn or TF antigens induce carbohydrate-specific cytotoxic T cell response in mice \[[@B162-biomolecules-02-00435]\]. These observations reinforce the complex role of glycosylation in the modulation of cellular immune responses. Humoral responses to MUC1 have also been observed in carcinoma patients. In breast cancer patients, the presence of circulating antibodies against MUC1 at the time of cancer diagnosis has been correlated with a favorable disease outcome. Furthermore, modification of the MUC1 peptides with GalNAc (Tn antigen) leads to stronger antibody binding, probably due to the changes in conformational epitopes \[[@B128-biomolecules-02-00435],[@B163-biomolecules-02-00435],[@B164-biomolecules-02-00435]\]. The interpretation of all this data in the sense of fully understanding how glycosylation modifies the immune response against MUC1 or even other proteins would lead to the development of a MUC1-based cancer vaccine that consistently elicits relevant humoral and cellular immunity that has not yet been developed. Other approaches against MUC1, rather than molecule vaccines, are also in trials in patients with specific types of cancer. One is a novel cell-based vaccination with autologous DCs loaded with MUC1 or Tn-MUC1 peptides \[[@B155-biomolecules-02-00435]\] and the other approach is the use of therapeutic humanized anti-MUC1 antibodies, such as the PankoMab-GEX from Glycotope and the radiolabeled anti-MUC1 humanized antibody, 90Y-hPAM4 from Immunomedics. Both the use of DC-based vaccines and therapeutic antibodies are excellent approaches to be considered in the future for specific anti-STn therapy. For the anti-STn therapy to reach the level of application that anti-MUC1 has reached, its effect on immune responses should be fully understood. This should be combined with the knowledge of the role that the STn antigen plays in the function of the STn-expressing proteins, MUC1 and others and in overall tumor cell progression in order to identify potential synergistic solutions for treating STn-expressing carcinomas. ### 4.2.3. Combining STn with other Cancer Associated Carbohydrates Considering that beyond STn other tumor associated carbohydrate antigens are usually aberrantly expressed by cancer cells \[[@B165-biomolecules-02-00435]\], multi-carbohydrate antigens covering different carbohydrates found aberrantly expressed in cancer cells have been constructed to trigger multi-antigenic responses in patients. Glycopeptide dendrimers have therefore been developed to simultaneously present different carbohydrate antigens such as STn, Tn, T, Globo H, G~M2~, Lewis^y^, and MUC1-Tn antigens. Presently, the challenge of efficiently preparing carbohydrate clusters was solved at the level of chemical synthesis \[[@B166-biomolecules-02-00435]\]. These constructs and their application in pre-clinical and clinical trials have been reviewed elsewhere \[[@B167-biomolecules-02-00435],[@B168-biomolecules-02-00435]\]. It is anticipated that these constructs will foster new platforms for effective and selective delivery of anti-tumor therapeutics. ### 4.2.4. Whom to Treat with Anti-STn Vaccine? As referred to above, a potential explanation for the Theratope failure in the Phase III clinical trial in breast cancer patients was the fact that the patient population was not evaluated for STn expression prior to enrolment. Just over 50% of the breast cancer cases express STn ([Figure 2](#biomolecules-02-00435-f002){ref-type="fig"}, [Figure 3](#biomolecules-02-00435-f003){ref-type="fig"} and [Figure 4](#biomolecules-02-00435-f004){ref-type="fig"}). The lack of patient selection was therefore possibly masking any benefit from the vaccine due to heterogeneous STn expression between patients. At the same time, [Figure 2](#biomolecules-02-00435-f002){ref-type="fig"} and [Figure 4](#biomolecules-02-00435-f004){ref-type="fig"} show the percentages of cases expressing STn in different types of cancer and reveal different populations that may benefit from STn-targeting therapies, such as the already mentioned patients with breast cancer, colorectal cancer as well as patients with gastric, ovarian and pancreatic cancer. In fact, the STn neo-expression or over-expression was reported with highest frequencies in pancreas, colorectal and ovarian cancers. However, considering that the role of STn may differ from cancer to cancer and probably from individual to individual, it is likely that selective approaches should be developed and patients should not be treated exactly by the same STn-targeting therapy, but instead based on their personal tumor profile. On the other hand, STn immunotherapy has always been tested as a post-surgical adjuvant therapy in combination with hormonotherapy or chemotherapy, with the aim of improving survival by decreasing relapse, as discussed above. However, recent data in the literature suggest that anti-cancer vaccine might be more efficiently used in a prophylactic rather than in a therapeutic way \[[@B169-biomolecules-02-00435],[@B170-biomolecules-02-00435]\]. Interestingly, anti-STn antibodies were detected in healthy women who never developed cancer in the 25 to 30 years following the date of blood sampling \[[@B128-biomolecules-02-00435]\]. Assuming that STn presentation to the immune system has to result from carcinogenesis, the authors propose that these anti-STn antibodies might be functional in suppressing tumor development and progression. This concept is further supported by a recent study showing that a robust and specific auto-immune response against a cancer specific antigen is able to prevent the emergence of autochthonous ovarian tumors and control the growth of established ovarian malignancies \[[@B171-biomolecules-02-00435]\]. With these concepts in mind, anti-STn vaccine, such as Theratope, may become attractive again, now as a prophylactic anti-cancer vaccine, for several reasons. First, the protective effect of anti-STn has already been demonstrated *in vivo*. Second, the efficacy of the anti-STn vaccine to induce a proper immune response in humans without major risks of autoimmune responses in patients has been reported repeatedly. Third, there is a substantial body of evidence showing that STn is expressed in early carcinogenesis in all the epithelial cancers investigated ([Figure 4](#biomolecules-02-00435-f004){ref-type="fig"}) making these cancers good targets from the earliest stage of the tumors. At last, most of these cancers where STn is more frequently expressed, such as cancers of the digestive tract, pancreas, colon, lungs and ovaries, are also cancers whose incidence are associated with age, with a drastic increase from 45 years onwards ([Figure 6](#biomolecules-02-00435-f006){ref-type="fig"}). Thus the whole population above 45 years of age would constitute a suitable cohort to test the ability of anti-cancer vaccine to decrease epithelial cancer incidence. There are some prophylactic vaccines being tested for their ability to decrease cancer incidence by protecting the patients against potentially cancer-inducing viral infection. For example, anti-papilloma virus vaccines have been shown to protect the immunized population against cervical cancer \[[@B172-biomolecules-02-00435]\] However, it is noteworthy to know that prophylactic anti-cancer vaccines directly targeting a cancer-associated antigen have never been tested in clinical trials \[[@B169-biomolecules-02-00435]\]. For the reason listed above, we believe that an improved anti-STn vaccine should be a good candidate to be among the first to be tested in the future. {#biomolecules-02-00435-f006} 5. General Conclusion ===================== While additional investigations are needed to clarify the discrepant results concerning Theratope clinical trials, the pursuit of improved anti-STn immunotherapy remains an active area of investigation. The overall concept of anti-STn vaccines as any other anti-cancer vaccine is a promising idea. However, putting the theory into practice has proven to be challenging. The advances in vaccines using STn-expressing proteins, other carbohydrate antigens and DC-based vaccines associated with a robust knowledge of the STn physiological function and the identification of means of surpassing tolerogenic responses will warrant a selective and successful development of anti-STn therapy, possibly in a prophylactic setting. Other anti-STn therapeutic approaches, such as therapeutic antibodies that may prove particularly useful in targeting immune responses and also drug delivery are an expected open field. This work was supported by the University of Sciences and Technologies of Lille, the Association pour la Recherche sur le Cancer (Grant n° 7936 and 5023) et le comité de l'Aisne de La Ligue contre le Cancer.

\section{Introduction} In conventional non-orthogonal multiple access (NOMA) networks, spectrum sharing among multiple users is encouraged under the condition that users can be distinguished based on their channel conditions \cite{NOMAPIMRC,Nomading}. A recent work in \cite{8786203} proposed a new form of NOMA, termed OTFS-NOMA, which applies orthogonal time frequency space modulation (OTFS) to NOMA and yields an alternative method to distinguish users by their mobility profiles. In particular, OTFS-NOMA is motivated by the fact that conventional OTFS mainly relies on the delay-Doppler plane, where a high-mobility user's signals are converted from the time-frequency plane to the delay-Doppler plane, such that the user can experience time-invariant channel fading \cite{OTFS2, OTFS, 8424569,8516353,8503182}. Unlike conventional OTFS, OTFS-NOMA uses the bandwidth resources available in both the time-frequency plane and the delay-Doppler plane \cite{8786203, otfsnoma1,nomaotfs2}. As a result, in OTFS-NOMA, high-mobility users can be still served with time-invariant channels in the delay-Doppler plane, and the resources in the time-frequency plane can also been released to those low-mobility users, which improves the overall spectral efficiency. Such spectrum sharing among the users with different mobility profiles can be particularly important to 5G and beyond communication scenarios, where some users might be static, e.g., Internet of Things (IoT) sensors, and there might be some users which are moving at very high speeds, e.g., users in a high-speed train. In this paper, we focus on an OTFS-NOMA downlink transmission scenario, in which a high-mobility user and multiple low-mobility NOMA users share the spectrum and are served simultaneously by a base station. Unlike \cite{8786203}, we assume that the base station has multiple antennas and each user has a single antenna. The design of beamforming is considered in this paper, where our objective is to maximize the low-mobility NOMA users' data rates while guaranteeing that the high-mobility user's targeted data rate can be be met. Because of the users' heterogeneous mobility profiles, we assume that the low-mobility NOMA users' channel state information (CSI) is perfectly known by the base station, but there exist errors for the high-mobility user's CSI. In the presence of these CSI errors, a robust beamforming optimization problem is formulated and solved by applying successive convex approximation (SCA). We note that, in the case with perfect CSI, the formulated robust beamforming design problem can be degraded to a simplified form which facilitates the development of a more computationally efficient method based on semidefinite relaxation (SDR). Computer simulation results are provided to demonstrate that the proposed SCA robust beamforming scheme can efficiently utilize the spatial degrees of freedom and achieve a significant performance gain over the case with a randomly chosen beamformer. In the case with perfect CSI, the developed SDR method can realize a better performance than the SCA-based scheme, where both the proposed schemes outperform random beamforming. \section{ System Model } \label{section 2} This paper considers an OTFS-NOMA downlink scenario, with one base station communicating with $(M+1)$ users, denoted by $\text{U}_i$, $0\leq i\leq M$. The base station has $V$ antennas, and each user is equipped with a single antenna. As in \cite{8786203}, $\text{U}_0$ is assumed to be a high-mobility user and its $N M$ information bearing signals, denoted by $x_0[k,l]$, $k\in\{0, \cdots, N-1\}$, $l\in\{0,\cdots, M-1\}$, are placed in the delay-Doppler plane, where $M$ and $N$ denote the OTFS parameters and define how the delay-Doppler and time-frequency planes are partitioned, e.g., the time-frequency plane can be partitioned by sampling at intervals of $T$ and frequency spacing $\Delta f$. By using the inverse symplectic finite Fourier transform (ISFFT), $\text{U}_0$'s signals are converted from the delay-Doppler plane to the time-frequency plane \cite{OTFS}: \begin{align} \mathbf{x}_0^{\rm{TF}} =& \mathbf{F}^H_N\otimes \mathbf{F}_M \mathbf{x}_0 , \end{align} where $\mathbf{F}_n$ denotes an $n\times n$ discrete Fourier transform (DFT) matrix and $\mathbf{x}_0$ is an $NM\times 1$ vector collecting all $x_0[k,l]$. Denote $X_0[n,m]$ by the $(nN+m+1)$-th element of $\mathbf{x}_0^{\rm{TF}}$, where $0\leq n \leq N-1$ and $0\leq m\leq M-1$. Without using NOMA, i.e., OTFS with orthogonal multiple access (OTFS-OMA), only the high-mobility user, $\text{U}_0$, is served during $NT$ and $M\Delta f$, whereas other users cannot be admitted to these time slots or frequencies. The main motivation for using OTFS-NOMA is to create an opportunity for ensuring that the bandwidth resources, $NT$ and $M\Delta f$, can be shared between the high-mobility user and additional low-mobility users, by applying the principle of NOMA. Such spectrum sharing is particularly important if the high-mobility user has weak channel conditions or needs to be served with a small data rate only\footnote{We note that even if the high-mobility user's channel conditions are strong and this user wants to be served with a high data rate, channel uncertainties caused by the user's high mobility can still reduce the user's achievable data rate. As a result, it is spectrally inefficient to allow all the bandwidth resources to be solely occupied by the high-mobility user.}. Similar to \cite{8786203}, the $M$ NOMA users, $\text{U}_i$, $1\leq i \leq M$, are assumed to be low-mobility users, and their signals, denoted by $X_i[n,m]$, are placed directly in the time-frequency plane, and are superimposed with the high-mobility user's signals, $X_0[n,m]$. As in \cite{8786203}, we assume that the NOMA users' time-frequency signals are generated as follows: \begin{align}\label{indooruser2} X_i[n,m] =\left\{\begin{array}{ll} x_i(n) & \text{if}\quad m=i-1 \\ 0 &\text{otherwise}\end{array}\right., \end{align} where $x_i(n)$, for $1\leq i \leq M$ and $0\leq n\leq N-1$, are $\rm{U}_i$'s information bearing signals. At its $v$-th antenna, $1\leq v\leq V$, the base station superimposes $\text{U}_0$'s signals with the NOMA users' signals by using the beamforming coefficient, denoted by $w_v$, as follows~: \begin{align}\label{downlink1} X^v[n,m] =& w_v\sum^{M}_{i=0} X_i[n,m]. \end{align} We note that the use of power allocation among the users can further improve the performance of OTFS-NOMA, however, scaling the $(M+1)$ users' signals with different coefficients makes an explicit expression for the signal-to-interference-plus-noise ratio (SINR) difficult to obtain. Therefore, the design of beamforming is focused in this paper, where the joint design of beamforming and power allocation is a promising direction for future research but beyond the scope of this paper. Following steps similar to those in \cite{8786203,OTFS, OTFS2,8424569}, the received signal at $\text{U}_i$ in the time-frequency plane can be modelled as follows: \begin{align}\label{downlink1x} Y_i[n,m] =& \sum^{V}_{v=1}w_v H^v_i[n,m]X[n,m] +W_i[n,m], \end{align} where $W_i(n,m)$ is the white Gaussian noise in the time-frequency plane, and $H^v_i[n,m]$ denotes the time-frequency channel gain between the $v$-th antenna at the base station and $\text{U}_i$. The users' detection strategies are described in the following subsections. \vspace{-1em} \subsection{ Detecting the High-Mobility User's Signals} \label{section 4} Each user will first try to detect $\text{U}_0$'s signals in the delay-Doppler plane. In particular, the system model in the delay-Doppler plane at $\text{U}_i$' is given by \cite{8786203}: \begin{align}\label{yx model2} \mathbf{y}_i = \sum^V_{v=1}w_v\mathbf{H}^v_i \mathbf{x}_0 +\underset{\text{Interference and noise terms}}{\underbrace{ \sum^V_{v=1}w_v\sum^{M}_{q=1} \mathbf{H}_i^v {\mathbf{x}}_q +\mathbf{z}_i}}, \end{align} where $\mathbf{H}_i^v$ is an $NM\times NM$ block-circulant matrix, $\mathbf{y}_i$ denotes $\text{U}_i$'s $NM$ observations in the delay-Doppler plane, $\mathbf{x}_q$ denotes $\text{U}_q$'s signals in the delay-Doppler plane and $\mathbf{z}_i$ denotes the noise vector. In this paper, the use of a frequency-domain linear equalizer is considered\footnote{ We note that equalization is still carried out in the delay-Doppler plane, not in the time-frequency plane. The terminologies, frequency-domain equalizers, are used because the system model in \eqref{yx model2} can be regarded as a model in conventional single carrier cyclic prefix systems, to which various equalization techniques, termed frequency-domain equalizers, have been developed \cite{8786203}. In addition to the frequency-domain linear equalizer, we note that other types of equalizers can also be used. For example, one can use a frequency-domain decision feedback equalizer (FD-DFE). As shown in \cite{8786203}, the use of such more advanced equalizers can further improve the performance of OTFS-NOMA, but results in more computational complexity. }. In particular, by applying the detection matrix, $\left( \mathbf{F}_N^H \otimes \mathbf{F}_M\right)\left(\sum^{V}_{v=1}w_v\mathbf{D}_i^v \right)^{-1}\mathbf{F}_N \otimes \mathbf{F}_M^H $, to the observation vector $\mathbf{y}_i$, the received signals for OTFS-NOMA downlink transmission can be written as follows: \begin{align}\label{approach 1 system model 1} &\left( \mathbf{F}_N^H \otimes \mathbf{F}_M\right) \mathbf{D}_i^{-1}\mathbf{F}_N \otimes \mathbf{F}_M^H {\mathbf{y}}_i\\\nonumber = & \mathbf{x}_0 +\underset{\text{Interference and noise terms}}{\underbrace{ \ \sum^{M}_{q=1} {\mathbf{x}}_q + \left( \mathbf{F}_N^H \otimes \mathbf{F}_M\right) \mathbf{D}_i^{-1}\mathbf{F}_N \otimes \mathbf{F}_M^H{\mathbf{z}}_i}}, \end{align} where $\mathbf{D}_i=\sum^{V}_{v=1}w_v\mathbf{D}_i^v $, $\mathbf{D}_i^v $ is a diagonal matrix whose $(kM+l+1)$-th main diagonal element is given by \begin{align} D_{i}^{k,l,v}= \sum^{N-1}_{n=0}\sum^{M-1}_{m=0} a_{i,n}^{m,1,v} e^{j2\pi \frac{lm}{M}} e^{-j2\pi \frac{kn}{N}}, \end{align} for $0\leq k \leq N-1$, $0\leq l \leq M-1$, and $a^{m,1,v}_{i,n}$ is the element located in the $(nM+m+1)$-th row and the first column of $\mathbf{H}_i^v$. Therefore, the SINRs for detecting $x_0[k,l]$, $0\leq k \leq N-1$ and $0\leq l \leq M-1$, are identical and given by \begin{align}\label{sinrzf} \text{SINR}_{i}^{\text{I}} =& \frac{\rho }{\rho +\frac{1}{NM}\sum^{N-1}_{\tilde{k}=0}\sum^{M-1}_{\tilde{l}=0} \left|\sum^{V}_{v=1}w_vD_i^{\tilde{k},\tilde{l},v}\right|^{-2}}, \end{align} where $\rho$ denotes the transmit signal-to-noise ratio (SNR). \subsection{Detecting the Low-Mobility NOMA Users' Signals} \label{subsection 2} Assume that $\text{U}_0$'s signals can be decoded and removed successfully, which means that, in the time-frequency plane, the NOMA users observe the following: \begin{align} Y_i[n,m] =&\sum^{V}_{v=1}w_v H^v_i[n,m] x_{m+1}(n) +W_i[n,m] . \end{align} Therefore, the signals for $\rm{U}_i$ experience the following SNR: \begin{align}\label{snrx} \text{SNR}_{i}^{\rm{II}} = \rho \left|\sum^{V}_{v=1}w_v {D}_i^{0,i-1,v}\right|^2, \end{align} where ${D}_i^{0,i-1,v}$ is used since it is assumed that the low mobility NOMA users' channels are time invariant. Define $\mathbf{h}_{i,l}=\begin{bmatrix}{D}_i^{0,l,1}&\cdots &{D}_i^{0,l,V}\end{bmatrix}^T$, and $\mathbf{g}_{k,l}=\begin{bmatrix}{D}_0^{k,l,1}&\cdots &{D}_0^{k,l,V}\end{bmatrix}^T$. Because of the $\rm{U}_0$'s high mobility, it is assumed that the base station does not have the perfect knowledge of $\rm{U}_0$'s CSI, and this CSI uncertainty is modelled as follows \cite{5164911,6626661}: \begin{align} \mathbf{g}_{k,l} = \hat{\mathbf{g}}_{k,l} +\mathbf{e}_{k,l}, \end{align} where $\hat{\mathbf{g}}_{k,l}$ denotes the channel estimates available at the base station and $\mathbf{e}_{k,l}$ denotes the CSI errors. In particular, it is assumed that the CSI errors are bounded as follows: $\mathbf{e}_{k,l}^H\mathbf{C}_{k,l}\mathbf{e}_{k,l}\leq 1$. As in \cite{5164911,6626661}, we choose $\mathbf{C}_{k,l}=\sigma^{-2}\mathbf{I}_V$, where $\sigma$ denotes the parameter for indicating the accuracy of the channel estimates. In this paper, we focus on the robust beamforming design problem which can be formulated as follows: \begin{problem}\label{pb:1} \begin{alignat}{2} \rm{max.} &\qquad {\rm min.} \quad \left\{ \log(1+\text{SINR}_{i}^{\text{II}}), 1\leq i\leq M \right\}\label{obj:1} \\ \rm{s.t.} & \qquad \! \begin{aligned} \underset{\substack{ \mathbf{e}_{ {k}, {l}}^H \mathbf{e}_{ {k}, {l}}\leq \sigma^{2}\\ 0\leq k\leq N-1, 0\leq l \leq M-1 }}{\rm min.}& \log(1+\text{SINR}_{i}^{\text{I}})\geq R_0, \\&\qquad \qquad0\leq i \leq M\end{aligned}\label{st:1} \\ & \qquad \mathbf{w}^H\mathbf{w}\leq 1,\label{st:2} \end{alignat} \end{problem} where $\mathbf{w}$ denotes a $V\times1$ beamforming vector and $R_0$ denotes the high-mobility user's target data rate. The aim of the optimization problem formulated in \eqref{pb:1} is explained in the following. The objective function in \eqref{obj:1} is the minimum of the $M$ low-mobility NOMA users' data rates, i.e., the objective of the problem formulated in (P1) is to maximize the low-mobility NOMA users' data rates. \eqref{st:1} is to ensure that the first stage of successive interference cancellation (SIC) is successful at both the low-mobility and high-mobility users, and \eqref{st:2} is to ensure the power normalization. It is important to point out that the constraints shown in \eqref{st:1} serve two purposes. One is to ensure that the high-mobility user's data rate realized by the OTFS-NOMA transmission scheme is larger than this user's target data rate, $R_0$, i.e., the high-mobility user's target data rate can be guaranteed. The other is to ensure that the low-mobility NOMA users can successfully decode the high-mobility user's signals. Without these constraints, it is possible that one low-mobility user fails the first stage of SIC, which means that the rate $\log(1+\text{SINR}_{i}^{\text{II}})$ might not be achievable. \section{Low-Complexity Beamforming Design} In this section, we first focus on the case $\sigma\neq 0$, where we will convert the objective function into a convex form, then recast the robust beamforming design problem to an equivalent form without CSI errors $\mathbf{e}_{k,l}$, and finally apply SCA to obtain a low-complexity solution. In addition, the case with perfect CSI will also be studied, where an optimal solution based on SDR can be obtained for a special case. \subsection{Robust Beamforming Design} By using the expressions of the SINRs in \eqref{sinrzf} and the SNRs in \eqref{snrx} and with some algebraic manipulations, the robust beamforming design problem can be recasted as follows: \begin{problem}\label{pb:3} \begin{alignat}{2} \rm{max.} &\qquad {\rm min.} \quad \left\{ |\mathbf{w}^H\mathbf{h}_{i,i-1}|^2, 1\leq i\leq M \right\} \\ \rm{s.t.} & \qquad \!\begin{aligned} \sum^{N-1}_{ {k}=0}\sum^{M-1}_{ {l}=0} \underset{ \mathbf{e}_{ {k}, {l}}^H \mathbf{e}_{ {k}, {l}}\leq \sigma^{2}}{{\rm max.}} (\mathbf{w}^H(\hat{\mathbf{g}}_{ {k}, {l}}+\mathbf{e}_{ {k}, {l}})\\\times(\hat{\mathbf{g}}_{ {k}, {l}}+\mathbf{e}_{ {k}, {l}})^H\mathbf{w})^{-1} \leq \epsilon \end{aligned} \label{st 30} \\ & \qquad \sum^{M-1}_{ {l}=0} \frac{1}{\mathbf{w}^H\mathbf{h}_{ {i,l}}\mathbf{h}_{ {i,l}}^H\mathbf{w} }\leq \epsilon_1 , 1\leq i \leq M\label{st 31} \\ & \qquad \mathbf{w}^H\mathbf{w}\leq 1,\label{st 32} \end{alignat} \end{problem} $\eta=2^{R_0}-1$, $\epsilon=\rho NM(\eta^{-1}-1)$ and $\epsilon_1=\rho M(\eta^{-1}-1)$. In order to remove the minimization operator in the objective function, we recast \eqref{pb:3} as follows: \begin{problem}\label{pb:41} \begin{alignat}{2} \rm{max.} &\qquad z \\ \rm{s.t.} & \qquad \label{410} - |\mathbf{w}^H\mathbf{h}_{i,i-1}|^2\leq -z, 1\leq i\leq M \\ &\qquad \eqref{st 30}, \eqref{st 31}, \eqref{st 32}.\nonumber \end{alignat} \end{problem} Although the objective function in \eqref{pb:41} becomes an affine function, the constraint in \eqref{410} is not convex, since the left-hand side of the inequality is a concave function. Therefore, we further recast \eqref{pb:41} as follows: \begin{problem}\label{pb:42} \begin{alignat}{2} \rm{min.} &\qquad t \\ \rm{s.t.} & \qquad \label{411} \frac{1}{|\mathbf{w}^H\mathbf{h}_{i,i-1}|^2}\leq t, 1\leq i\leq M \\ &\qquad \eqref{st 30}, \eqref{st 31}, \eqref{st 32}.\nonumber \end{alignat} \end{problem} The left-hand side of the constraint in \eqref{411} is a convex function, as can be shown in the following. Define $\tilde{f}_{i,l}(\mathbf{w})= \frac{1}{ \mathbf{w}^H {\mathbf{h}}_{ {i}, {l}} {\mathbf{h}}_{ {i}, {l}}^H\mathbf{w}} $, where its first order derivative is given by \begin{align}\label{de1} &\bigtriangledown \tilde{f}_{i,l}(\mathbf{w})= \frac{-2 {\mathbf{h}}_{ {i}, {l}} {\mathbf{h}}_{ {i}, {l}}^H\mathbf{w} }{ \left(\mathbf{w}^H {\mathbf{h}}_{ {i}, {l}} {\mathbf{h}}_{ {i}, {l}}^H\mathbf{w}\right)^{2} } , \end{align} and its second order derivative is given by \begin{align}\nonumber &\bigtriangledown^2 \tilde{f}_{i,l}(\mathbf{w})= \frac{6 {\mathbf{h}}_{ {i}, {l}} {\mathbf{h}}_{ {i}, {l}} ^H } {\left(\mathbf{w}^H {\mathbf{h}}_{ {i}, {l}} {\mathbf{h}}_{ {i}, {l}}^H\mathbf{w}\right)^{2} } , \end{align} which is positive semidefinite. Therefore, the constraint in \eqref{411} is in a convex form. In oder to remove the CSI errors, $\mathbf{e}_{k,l}$, from the optimization problem, we first define the following feasibility problem: \begin{problem}\label{pb:x} \begin{alignat}{2} \rm{Find} &\quad \mathbf{e}_{k,l} \\ \rm{s.t.} & \quad \mathbf{w}\mathbf{w}^H \mathbf{e}_{ {k}, {l}} =- \mathbf{w}\mathbf{w}^H\hat{\mathbf{g}}_{ {k}, {l}} \\ &\quad \mathbf{e}_{ {k}, {l}}^H \mathbf{e}_{ {k}, {l}}\leq \sigma^{2} .\end{alignat} \end{problem} The following lemma is provided to simplify the optimization problem shown in \eqref{pb:42} and facilitate the application of SCA and SDR. \begin{lemma} The robust beamforming optimization problem in \eqref{pb:42} can be equivalently recast as follows: \begin{problem}\label{pb:4} \begin{alignat}{2} \rm{min.} &\quad t \\ \rm{s.t.} & \quad \frac{1}{ \mathbf{w}^H\mathbf{h}_{i,i-1}\mathbf{h}_{i,i-1}^H\mathbf{w}} \leq t, 1\leq i\leq M \label{stx400} \\ &\quad \sum^{N-1}_{ {k}=0}\sum^{M-1}_{ {l}=0} \frac{1}{ \left( \left(\mathbf{w}^H \hat{\mathbf{g}}_{ {k}, {l}}\hat{\mathbf{g}}_{ {k}, {l}}^H\mathbf{w}\right)^{\frac{1}{2}}-\left(\sigma^{2} \mathbf{w}^H\mathbf{w}\right)^{\frac{1}{2}} \right)^2 }\leq \epsilon \label{stx40} \\ & \quad \sum^{M-1}_{ {l}=0} \frac{1}{\mathbf{w}^H\mathbf{h}_{ {i,l}}\mathbf{h}_{ {i,l}}^H\mathbf{w} }\leq \epsilon_1 , 1\leq i \leq M \label{stx41} \\ & \quad \mathbf{w}^H\mathbf{w}\leq 1, \label{stx42} \end{alignat} \end{problem} if an optimal solution of \eqref{pb:4} can ensure that the problem shown in \eqref{pb:x} is infeasible for any $\hat{\mathbf{g}}_{k,l}$, $k\in \{0, N-1\}$ and $l\in \{0, \cdots, M-1\}$. Otherwise, the robust beamforming optimization problem in \eqref{pb:42} is infeasible. \end{lemma} \begin{proof} Please refer to the appendix. \end{proof} {\it Remark 1:} The fact that strong CSI errors results in the infeasibility situation can be explained in the following. With strong CSI errors, it is very likely to have \begin{align} \underset{ \mathbf{e}_{ {k}, {l}}^H \mathbf{e}_{ {k}, {l}}\leq \sigma^{2}}{{\rm min.}} \mathbf{w}^H(\hat{\mathbf{g}}_{ {k}, {l}}+\mathbf{e}_{ {k}, {l}}) (\hat{\mathbf{g}}_{ {k}, {l}}+\mathbf{e}_{ {k}, {l}})^H\mathbf{w} =0, \end{align} which leads to the situation that the constraint in \eqref{st 30} can never be satisfied, e.g., the problem is infeasible. We also note that an optimal solution obtained by solving \eqref{pb:4} is not necessarily an optimal solution for \eqref{pb:42}. Or in other words, only if an optimal solution of \eqref{pb:4} fails the feasibility check for \eqref{pb:x}, one can claim that this optimal solution of \eqref{pb:4} is also optimal for \eqref{pb:42}. {\it Remark 2:} The formulation in \eqref{pb:4} is general and can also be applicable to the case without CSI errors. In particular, by setting $\sigma=0$, one can easily verify that the formulation in \eqref{pb:4} is indeed applicable to the one without CSI errors. We note that problem \eqref{pb:4} is still not convex, mainly due to the non-convex constraint in \eqref{stx40}. In the following, SCA is applied to obtained a low-complexity suboptimal solution. Define $f_{k,l}(\mathbf{w})= \left( \left(\mathbf{w}^H \hat{\mathbf{g}}_{ {k}, {l}}\hat{\mathbf{g}}_{ {k}, {l}}^H\mathbf{w}\right)^{\frac{1}{2}}-\left(\sigma^{2} \mathbf{w}^H\mathbf{w}\right)^{\frac{1}{2}} \right)^{-2}$, where its first-order derivative is given by \begin{align}\label{de2} &\bigtriangledown f_{k,l}(\mathbf{w})=\\\nonumber &\frac{-2\left[ \left(\mathbf{w}^H \hat{\mathbf{g}}_{ {k}, {l}}\hat{\mathbf{g}}_{ {k}, {l}}^H\mathbf{w}\right)^{-\frac{1}{2}} \hat{\mathbf{g}}_{ {k}, {l}}\hat{\mathbf{g}}_{ {k}, {l}}^H\mathbf{w}-\sigma \left( \mathbf{w}^H\mathbf{w}\right)^{-\frac{1}{2}} \mathbf{w}\right]}{ \left( \left(\mathbf{w}^H \hat{\mathbf{g}}_{ {k}, {l}}\hat{\mathbf{g}}_{ {k}, {l}}^H\mathbf{w}\right)^{\frac{1}{2}}-\sigma\left( \mathbf{w}^H\mathbf{w}\right)^{\frac{1}{2}} \right)^3 } . \end{align} By applying the Taylor expansion to the constraints in \eqref{pb:4} at a feasible point $\mathbf{w}_0$ and also using $\bigtriangledown \tilde{f}_{i,l}(\mathbf{w})$ in \eqref{de1} and $\bigtriangledown f_{k,l}(\mathbf{w})$ in \eqref{de2}, the optimization problem in \eqref{pb:4} can be approximated as follows: \begin{problem}\label{pb:5x} \begin{alignat}{2} \rm{min.} &\quad t \\ \rm{s.t.} & \quad \!\begin{aligned} \tilde{f}_{i,i-1}(\mathbf{w_0}) +\left[\bigtriangledown \tilde{f}_{i,i-1}(\mathbf{w}_0)\right]^H&\left(\mathbf{w}-\mathbf{w}_0\right) \leq t,\\& 1\leq i\leq M\end{aligned} \\ &\quad \sum^{N-1}_{ {k}=0}\sum^{M-1}_{ {l}=0} \left( f_{k,l}(\mathbf{w_0}) +\left[\bigtriangledown f_{k,l}(\mathbf{w}_0)\right]^H\left(\mathbf{w}-\mathbf{w}_0\right) \right)\leq \epsilon \\ & \quad \!\begin{aligned} & \sum^{M-1}_{ {l}=0} \left( \tilde{f}_{i,l}(\mathbf{w_0}) +\left[\bigtriangledown \tilde{f}_{i,l}(\mathbf{w}_0)\right]^H\left(\mathbf{w}-\mathbf{w}_0\right) \right)\leq \epsilon_1 ,\\ &\qquad\qquad\qquad\qquad\qquad\qquad\qquad 1\leq i \leq M\end{aligned} \\ & \quad \mathbf{w}^H\mathbf{w}\leq 1, \end{alignat} \end{problem} to which a straightforward iterative SCA algorithm facilitated by convex optimization solvers, such as CVX, can be applied to find a solution \cite{Boyd}. {\it Remark 3:} We note that the problem in \eqref{pb:5x} is only an approximated form of the original problem in \eqref{pb:4}, which means that the solution obtained by SCA is only a suboptimal solution of \eqref{pb:4}. {\it Remark 4:} One challenge for the implementation of SCA is to find a feasible $\mathbf{w}_0$ which is required for the SCA initialization. However, given the large number of constraints in \eqref{pb:4}, it is difficult to find a feasible $\mathbf{w}_0$. For the simulations carried out for the paper, we simply use a randomly generated $\mathbf{w}_0$. With this randomly generated $\mathbf{w}_0$, the use of SCA can still yield a solution which outperforms the benchmarking scheme, as shown in the next section. \vspace{-1em} \subsection{Beamforming Design Without CSI Errors} When $\sigma=0$, the problem in \eqref{pb:4} is degraded to a simplified form without CSI errors, to which SCA can be still applicable. We note that, because of the simplified form, SDR also becomes applicable \cite{5447068}, where the advantage of using SDR is to avoid the iterations for updating $\mathbf{w}_0$ as in SCA. Therefore, the complexity of SDR can be much smaller than that of SCA. In addition, the use of SDR can also avoid the challenging issue about generating $\mathbf{w}_0$, as discussed in Remark 4. By applying the principle of SDR, the problem in \eqref{pb:4} can be recasted to the following equivalent form: \begin{problem}\label{pb:6x} \begin{alignat}{2} \rm{min.} &\quad t \\ \rm{s.t.} & \quad \frac{1}{ \rm{tr} \{\mathbf{W}\mathbf{H}_{i,i-1} \}} \leq t, 1\leq i\leq M \\ &\quad \sum^{N-1}_{ {k}=0}\sum^{M-1}_{ {l}=0} \frac{1}{ \rm{tr}\{ \mathbf{W} {\mathbf{G}}_{ {k}, {l}} \} }\leq \epsilon \\ & \quad \sum^{M-1}_{ {l}=0} \frac{1}{\rm{tr}\{\mathbf{W}\mathbf{H}_{ {i,l}}\} }\leq \epsilon_1 , 1\leq i \leq M \\ & \quad \rm{tr}\{\mathbf{W}\}\leq 1 \\ & \quad \mathbf{W}\succeq 0 \\ & \quad \rm{rank}\{\mathbf{W}\} = 1, \end{alignat} \end{problem} where $\mathbf{H}_{i,l}=\mathbf{h}_{i,l}\mathbf{h}_{i,l}^H$, ${\mathbf{G}}_{ {k}, {l}}={\mathbf{g}}_{ {k}, {l}} {\mathbf{g}}_{ {k}, {l}}^H$ and $\rm{tr}\{\cdot\}$ denotes the trace operator. The optimization problem in \eqref{pb:6x} is still not convex, mainly due to those constraints containing $ \frac{1}{ \rm{tr} \{\mathbf{W}\mathbf{H}_{k,l} \}}$ and $ \frac{1}{ \rm{tr} \{\mathbf{W}\mathbf{G}_{k,l} \}}$. By introducing auxiliary variables, the problem in \eqref{pb:6x} can be converted into the equivalent form shown in the following: \begin{problem}\label{pb:7x} \begin{alignat}{2} \rm{min.} &\quad t \\ \rm{s.t.} & \quad \frac{1}{ x_{i,i-1}} \leq t, 1\leq i\leq M \label{st 70} \\ &\quad \sum^{N-1}_{ {k}=0}\sum^{M-1}_{ {l}=0} \frac{1}{ y_{k,l} }\leq \epsilon \label{st 71} \\ & \quad \sum^{M-1}_{ {l}=0} \frac{1}{x_{i,l} }\leq \epsilon_1 , 1\leq i \leq M \\ &\quad \rm{tr}\{ \mathbf{W} {\mathbf{G}}_{ {k}, {l}} \}=y_{k,l}, 0\leq k \leq N-1, 0\leq l\leq M-1, \\ &\quad \rm{tr}\{\mathbf{W}\mathbf{H}_{ {i,l}}\}=x_{i,l}, 1\leq i \leq M, 0\leq l\leq M-1, \\ &\quad y_{k,l}\geq0, 0\leq k \leq N-1, 0\leq l\leq M-1, \\ &\quad x_{i,l}\geq 0, 1\leq i \leq M, 0\leq l\leq M-1, \\ & \quad \rm{tr}\{\mathbf{W}\}\leq 1 \\ & \quad \mathbf{W}\succeq 0 \\ & \quad \rm{rank}\{\mathbf{W}\} = 1. \end{alignat} \end{problem} We note that all the constraints are either affine or convex, except the rank-one constraint. Take the constraint \eqref{st 71} as an example. Provided $y_{k,l}\geq0$, $\frac{1}{ y_{k,l} }$ is convex, and hence the left-hand side of the inequality, \eqref{st 71}, is also convex since it is the sum of those convex functions, $\frac{1}{ y_{k,l} }$. By removing the rank-one constraint, the problem in \eqref{pb:7x} can be efficiently solved by using CVX. {\it Remark 5:} Simulation results indicate that the rank of the SDR solution obtained for the problem shown in \eqref{pb:7x} is larger than one, and therefore, the Gaussian randomization method is needed, which means that the obtained SDR solution is just a suboptimal solution \cite{5447068}. {\it Remark 6:} For a special case which is to maximize a single NOMA user's data rate, the problem in \eqref{pb:7x} can be simplified as follows: \begin{problem}\label{pb:8x} \begin{alignat}{2} \rm{min.} &\quad x_{1,0} \\ \rm{s.t.} & \quad \rm{P9c-P9j}.\nonumber \end{alignat} \end{problem} Simulation results show that the rank of the solution for the problem in \eqref{pb:8x} is always one, which means that the SDR solution is optimal for this single-user special case. \begin{figure}[t] \vspace{-2em} \begin{center}\subfigure[$V=4$]{\label{fig1a}\includegraphics[width=0.34\textwidth]{fig1.eps}} \subfigure[ $V=8$]{\label{fig1b}\includegraphics[width=0.34\textwidth]{fig2.eps}} \vspace{-0.5em} \end{center}\vspace{-1em} \caption{The performance of MISO-OTFS-NOMA with CSI errors. $M=N=8$ and $R_0=0.5$ BPCU. }\label{fig 1}\vspace{-2em} \end{figure} \section{Simulation Results} In this section, the performance of the developed beamforming schemes is studied by using computer simulations. A simple two-path channel model is used. In particular, for $\rm{U}_0$, the delay-Doppler indices for the two channel taps are $(0,0)$ and $(1,1)$, respectively. With the subchannel spacing $\Delta f = 2$ kHz, the maximal speed corresponding to the largest Doppler shift $ 250 $ Hz is $67.5$ km/h if the carrier frequency is $f_c=4$ GHz. There is no Doppler shift for the NOMA users. In Fig. \ref{fig 1}, the performance of the robust beamforming design is studied, where $R_{\min}={\rm min.} \left\{ \log(1+\text{SINR}_{i}^{\text{II}}), 1\leq i\leq M \right\}$ and the random beamforming scheme is used as a benchmarking scheme. As can be observed from the figure, the proposed low-complexity beamforming design can provide a significant performance gain over the random beamforming scheme. Furthermore, the performance of the proposed SCA scheme is improved by reducing $\sigma$, i.e., the channel estimates become more accurate. We note that the performance of the random beamforming scheme is also affected by the choices of $\sigma$, which can be explained in the following. For a randomly chosen $\mathbf{w}$, it is still possible for this $\mathbf{w}$ to make the problem shown in \eqref{pb:x} feasible, if $\sigma$ is large. If this event occurs, we set $R_{\min}=0$, since this event means that the high-mobility user's targeted data rate cannot be met, as discussed in Remark~1. Comparing Fig. \ref{fig1a} to Fig. \ref{fig1b}, one can also observe that the performance of the proposed beamforming design is improved by increasing $V$, whereas the performance of random beamforming stays the same. This is due to the fact that the proposed beamforming scheme can more effectively use the spatial degrees of freedom than random beamforming. Fig. \ref{fig 2} shows the performance of the proposed beamforming design for the case without CSI errors. Comparing Fig. \ref{fig 2} to Fig. \ref{fig 1}, one can observe that removing the effects of CSI errors improves the performance of the proposed scheme, which is expected. Recall that in the case without CSI errors, the SDR method is also applicable. As can be observed from Fig. \ref{fig2a}, the SDR method outperforms the SCA method, even though the rank of the SDR solution is not one and the Gaussian randomization method has to be applied. This performance loss of SCA might be due to the used Taylor expansion, since the problem shown in \eqref{pb:5x} is just an approximation to the original problem considered in \eqref{pb:4}. Fig. \ref{fig2b} shows that, for the single-user case, the performance of the proposed beamforming schemes achieve the same performance. It is worth pointing out that the rank of the SDR solution for this case is always one, which means that SDR solution is optimal. Therefore, Fig. \ref{fig2b} indicates that the SCA solution is also optimal for the single-user case. Furthermore, we note that both the SDR and SCA based schemes outperform the random beamforming scheme, as can be observed from the two subfigures in Fig. \ref{fig 2}. \begin{figure}[t] \vspace{-2em} \begin{center}\subfigure[Multi-User Case]{\label{fig2a}\includegraphics[width=0.34\textwidth]{fig3.eps}} \subfigure[Single-User Case]{\label{fig2b}\includegraphics[width=0.34\textwidth]{fig4.eps}} \vspace{-0.5em} \end{center}\vspace{-1em} \caption{The performance of MISO-OTFS-NOMA without channel estimation errors. $M=N=8$ and $R_0=0.5$ BPCU. }\label{fig 2}\vspace{-2em} \end{figure} \vspace{-0.5em} \section{Conclusions} In this paper, we have considered the design of beamforming for OTFS-NOMA downlink transmission. The objective of the considered beamforming design is to maximize the low-mobility NOMA users' data rates while guaranteeing that the high-mobility user's targeted data rate can be met. Both the cases with and without CSI errors have been considered, where low-complexity solutions were obtained by applying SCA and SDR. Simulation results have also been provided to show that the use of the proposed beamforming schemes can yield a significant performance gain over random beamforming. In this paper, we focused on the downlink transmission scenario, in which more than one high-mobility user can be accommodated. In particular, we can use the resource blocks in the delay-Doppler plane to serve multiple high-mobility users. The multiple high-mobility users' signals cause interference to each other in the time-frequency plane, but the use of equalizers can ensure that there is no interference in the delay-Doppler plane. As a result, the optimization problem formulated in this paper can be straightforwardly extended to the case with multiple high-mobility users, by adding more constraints about the high-mobility users' target data rates. However, in the uplink scenario with multiple high-mobility users, one high-mobility user's signals can cause interference to the other users' signals in both the time-frequency and delay-Doppler planes. The design for robust OTFS-NOMA transmission for such a challenging uplink scenario is an important topic for future research. \vspace{-0.5em}

\section{Introduction} \label{sec:Introduction} Spurred by the remarkable commercial success of cryptocurrencies, there have been many attempts to date to extend Blockchain as a decentralized management platform for general business transactions. In most application scenarios, such as product traceability~\cite{tse2017blockchain,lu2017adaptable,toyoda2017novel} and those involving smart contracts~\cite{christidis2016blockchains}, Blockchain has been used essentially as an immutable data storage whose goal is simply to manage identical replicas of data among distributed nodes. Although this is a meaningful first step, we argue that the true value of Blockchain lies in its potential for value co-creation by network participants (``agents'') through knowledge sharing~\cite{scekic2018blockchain,seebacher2017blockchain,kondrateva2021potential}. The platform should help the agents obtain extra insights that cannot be accessed when looking at their own data alone, disconnected from the others. We envision that the next generation of Blockchain will integrate \textit{collaborative learning} capabilities at the core. Inspired by the Nakamoto's original agenda~\cite{nakamoto2008bitcoin}, our collaborative learning platform features three main characteristics: 1) The entire learning procedure is done in a decentralized manner, i.e.,~without relying on the central authority. 2) The outcomes of the learning reflect specific situations of the individual agents, resulting in generally different models for each of the agents. 3) The data and the learned model of the agents are protected as a private property. The \textit{first} characteristic is to ensure \textit{democracy} in the platform. The agents voluntarily communicate with the others through given network infrastructure, but there is no such thing as the central server that collects a piece of data from the agents and perform, e.g.,~stochastic gradient descent to train a deep learning model. The \textit{second} characteristic is to ensure \textit{diversity} in the platform. Here, a `model' refers to a probability distribution in general that captures patterns hidden in the data, as implied in Fig.~\ref{fig:S7expander_model.pdf}. In the machine learning literature, this is commonly called \textit{multi-task learning} because the goal is to learn $S$ different models (`tasks') simultaneously, where $S$ is the number of agents participated in the network. The \textit{third} characteristic is to ensure the \textit{privacy} of the agents in the platform. Although they learn from the other agents and share their learnings in return in a certain way, their original data and the resulting model must be protected. \begin{figure} \centering \includegraphics[trim={3cm 0.5cm 4cm 0cm},clip,height=4.5cm]{./fig/S7expander_model.pdf} \caption{Illustration of decentralized multi-task learning illustrated on a 3-regular expander graph of $S=7$ nodes, where $p(\cdot \mid \bm{\Theta},\bm{\Pi}^s)$ denotes a statistical machine learning model for the agent $s$ with $\bm{\Theta}$ being common model parameters and $\bm{\Pi}^s$ being agent-specific model parameters. } \label{fig:S7expander_model.pdf} \end{figure} This paper proposes a new framework of decentralized collaborative learning with certain privacy guarantees. Given a specific parametric model agreed upon, the goal of the agents is to maximize the likelihood function of the entire system in a collaborative manner. Higher likelihood implies higher model fidelity, which will lead to better business outcomes through more accurate predictions. In our previous work~\cite{ide2019efficient}, we presented a decentralized multi-task learning protocol for a mixture of exponential family distributions, where the dynamic consensus algorithm~\cite{wu2005agreement,ren2005survey,olfati2007consensus} eliminates the need for the central server. We found that the spectral structure of network topology plays a critical role in the convergence of the algorithm, but the analysis was mainly on the cycle graph, in which an analytic form of the eigenspectrum is available. In this paper, we show that a family of graphs called the \textit{expander graph}~\cite{vadhan2012pseudorandomness,lubotzky2010discrete} can dramatically speedup global consensus. We also provide a systematic analysis on the probability of privacy breach in a couple of different scenarios. Decentralized collaborative learning is a new research field. We conclude the paper with some remarks on open problems. \section{Related work} One recent major trend in distributed learning is to distributedly train deep neural networks using SGD (stochastic gradient descent)~\cite{mohassel2017secureml,mcmahan2017communication,Konecny16NIPSworkshop,agarwal2018cpsgd,yang2019federated}, in which managing huge computational overhead has been an issue. Although most of the existing work falls into the category of distributed \textit{single}-task learning, some recent works point out that model biases and thus a lack of diversity can be a serious issue~\cite{xu2014distributed,Mohri19ICML}. A multi-task extension is discussed in~\cite{smith2017federated} without a data privacy context. A unique role of the exponential family in data privacy is pointed out in~\cite{bernstein2018differentially}. For privacy preservation in distributed learning, a common approach is to use differential privacy~\cite{xie2017privacy,heikkila2017differentially}, but for real-valued data, performance degradation in learning due to introduced noise is an open question, as discussed in~\cite{ding2018comparing}. Privacy preservation in a decentralized environment is a challenging task in general. Almost the only solution known so far is the use of homomorphic encryption and its variants~\cite{mohassel2017secureml,liu2020secure,ruan2017secure}, but its computational cost is known to be often prohibitive. \section{Multi-Task learning framework } \label{sec:framework} This section summarizes the problem setting from our previous work~\cite{ide2019efficient}. \subsection{Problem setup} \label{subsec:setting} As shown in Fig.~\ref{fig:S7expander_model.pdf}, there are $S \geq 3$ agents in the network. Each agent indexed by $a$ privately keeps its own data set \begin{align} \mathcal{D}^a \triangleq \{\bm{z}^{(1)},\ldots, \bm{z}^{(N^a)} \} \end{align} about a random variable $\bm{z}$. Here, $N^a$ is the number of samples of the $a$-th agent. As a general rule, we use the superscript such as $^{(n)}$ to represent the $n$-th instance of a random variable. The random variable $\bm{z}$ can be a pair of a feature vector and its label like $\bm{z}=(y,\bm{x})$ or simply a feature vector alone $\bm{z} = \bm{x}$. We assume $\bm{x}$ is \textit{real-valued} and noisy in general. The goal of the agents is to learn a predictive distribution for~$\bm{z}$. As illustrated in Fig.~\ref{fig:S7expander_model.pdf}, the distribution will have model parameters $\bm{\Theta}$ and $\bm{\Pi}^a$, where $\bm{\Theta}$ is parameters shared by all the agents and $\bm{\Pi}^a$ is agent-specific parameters. Both $\bm{\Theta}$ and $\bm{\Pi}^a$ are to be learned from the total data $\mathcal{D} \triangleq\{\mathcal{D}^1,\ldots, \mathcal{D}^S \}$. The question is how the agents leverage information from the other ones while keeping data privacy. As part of the network infrastructure, a set of bi-directional communication paths are given as an undirected graph, whose nodes are the agents and the edges are pairwise communication paths, as shown in the figure. As is the case in the IP (internet protocol) network of the Internet, the infrastructure is assumed to do basic bookkeeping jobs such as network routing and clock synchronization without any interest in the contents communicated. Network failures are unavoidable in real networks but we do not consider them for simplicity. We also assume a consortium-based network, where the agents have verified identities. The agents are \textit{honest but curious}, meaning that they do not lie about computed statistics but they always try to selfishly get as much information as possible from the other agents. \subsection{Mixture of exponential family} \label{subsec:probmodel} To capture the diversity among the agent, we employ a \textit{mixture} of the exponential family. In what follows, we focus on the case where $\bm{z}=\bm{x} \in \mathbb{R}^M$ with $M$ being the dimensionality of the variable and the learning task is unsupervised (a.k.a.~density estimation). Extension to the supervised setting can be done easily. Practical applications of this setting include failure detection of industrial robots, where all the $S$ agents are assumed to have the same set of physical sensors. See~\cite{ide2018collaborative} for more detail. Now the observation model of the $a$-th agent is given by: \begin{align}\label{eq:mixtureModel} p(\bm{x} \mid \bm{\Theta}, \bm{u}^a) &= \prod_{k=1}^K f(\bm{x} \mid {\bm \theta}_k)^{u_k^a} \\ \label{eq:exponentialFamilyDistribution} f(\bm{x} \mid \bm{\theta}_k) &= G(\bm{\theta}_k)H(\bm{x})\exp\{ \bm{\eta}( \bm{\theta}_k)^\top \bm{T}(\bm{x})\}, \end{align} where $K$ is the number of mixture components and $\bm{u}^a \triangleq (u^a_1,\ldots,u^a_K)^\top$ is the one-hot indicator variable representing cluster assignment. Notice that the dependency on $a$ in $\bm{u}^a$ represents diversity of the model. In $f(\bm{z} \mid \bm{\theta}_k)$, functional forms of $G, H$ (scalar function) and $\bm{\eta},\bm{T}$ (vector-valued function) are given by a specific choice in the exponential family~\cite{betulehmann2006theory}. We will give a Gaussian-based model as an example below. In the unsupervised scenario, the observation model in the form Eq.~\eqref{eq:mixtureModel} is almost always combined with a prior distribution of $\bm{u}^a$: \begin{align}\label{eq:prior_pi} p({\bm u}^a \mid {\bm \pi}^a) &= \mathrm{Cat}({\bm u}^a \mid {\bm \pi}^a) \triangleq \prod_{k=1}^K (\pi_k^a)^{u^a_k} \end{align} where $\mathrm{Cat}$ denotes the categorical distribution. The parameter $\bm{\pi}^a \triangleq (\pi^a_1,\ldots, \pi^a_K)$ can be viewed as the probability distribution over the $K$ clusters and satisfies $\sum_{k=1}^K \pi^a_k =1$. For stable numerical estimation, prior distributions are imposed also on $\bm{\Theta} \triangleq \{ \bm{\theta}_k\}$ and $\bm{\Pi} \triangleq \{ \bm{\pi}^a\}$ as \begin{align} p(\bm{\Theta}) &= \prod_{k=1}^K p(\bm{\theta}_k), \quad p({\bm \Pi}) = \prod_{a=1}^S p({\bm \pi}^a), \end{align} where we have used $p(\cdot)$ to generically represent potentially different probability distributions. Here we give an example when $f(\bm{z} \mid \bm{\theta}_k)$ is the Gaussian $\mathcal{N}(\bm{x} \mid \bm{\mu}_k,(\mathsf{\Lambda}_k)^{-1})$, where $\bm{\mu}_k$ is the mean and $\mathsf{\Lambda}_k$ the precision matrix. For $\bm{\theta}_k = \{ \bm{\mu}_k, \mathsf{\Lambda}_k \}$, one practically recommended choice for the prior distribution is the Gauss-Laplace distribution: \begin{gather}\label{eq:Gauss-Laplace} p({\bm \mu}_k,{\sf \Lambda}_k) \propto {\cal N}({\bm \mu}_k|{\bm m}_0,(\lambda_0{\sf \Lambda}_k)^{-1}) \exp\left( -\frac{\rho}{2} \| {\sf \Lambda}_k \|_1 \right) \end{gather} where $\rho, \lambda_0,\bm{m}_0$ are predefined constants and $\| \cdot \|_1$ is the $\ell_1$ norm. For $p({\bm \pi}^a)$, a common choice is the Dirichlet distribution $p(\bm{\pi}^a)\propto (\pi^a_1\cdots\pi^a_K)^\gamma$ with $\gamma \sim 1$ is a given constant. \subsection{Model estimation algorithm} \label{subsec:MAP} The probabilistic model presented above contains unknown model parameters $\bm{\Theta},\bm{\Pi}$ in addition to the latent variable $\bm{U} \triangleq \{ \bm{u}^1, \ldots, \bm{u}^S\}$. The standard strategy to learn the model in such a case is to maximize the log marginalized likelihood $L_0$ with respect to $\bm{\Theta},\bm{\Pi}$: \begin{align} \nonumber L_0&\triangleq \ln p({\bm \Pi})p(\bm{\Theta}) \\ \nonumber &+ \ln \left\{\sum_{\bm{U}} \prod_{a=1}^s\prod_{n=1}^{N^s}p(\bm{z}^{a(n)}| \bm{\Theta}, \bm{u}^{a(n)})p(\bm{u}^{a(n)} \mid \bm{\pi}^a) \right\}. \end{align} Unfortunately, this maximization problem is intractable even in the simple Gaussian case. We instead maximize the lower bound of $L_0$ derived by applying Jensen's inequality. We can derive a simple iterative algorithm to estimate the unknown model parameters $\bm{\Theta},\bm{\Pi}$ as well as the posterior distribution of the latent variable $\bm{U}$: \begin{align} Q(\bm{U}) = \prod_{a=1}^S\prod_{n=1}^{N^s}\prod_{k=1}^K (r^{a(n)}_k)^{u^{a(n)}_k}. \end{align} Here, we summarize the result presented in our previous work~\cite{ide2019efficient}. With an initialized $\{r^{a(n)}\}$ so $\sum_{k=1}^K r^{a(n)}_k = 1$, we repeat the following steps until convergence: \begin{itemize} \item In each $a \in \{1,\ldots,S\}$, with the latest $\{r_k^{a(n)}\}$, locally compute % \begin{align}\label{eq:localStats} N_k^a \triangleq \sum_{n=1}^{N^a} r^{a(n)}_k \;\;\mbox{and} \;\; \bm{T}^a_k \triangleq \sum_{n=1}^{N^a} r^{a(n)}_k \bm{T}(\bm{z}^{a(n)}). \end{align} % \item Among all $a=1,\ldots, S$, build a global consensus on \begin{align}\label{eq:aggregate_global_exponential_family} N_k \triangleq \sum_{a=1}^{S} N^{a}_k, \quad \mbox{and} \quad \bm{T}_k \triangleq \sum_{a=1}^{S} \bm{T}^a_k, \end{align} \item In each $a \in \{1,\ldots,S\}$, locally solve \begin{align} \label{eq:solve} \max_{\bm{\theta}_k} \left\{ \ln p(\bm{\theta}_k)+N_k \ln G(\bm{\theta}_k) + \bm{T}_k^\top\bm{\eta}(\bm{\theta}_k) \right\} \end{align} \item In each $a \in \{1,\ldots,S\}$, with the latest $\{\bm{\theta}_k\}$ and $ \{\bm{\pi}^a\}$ with $\pi^a_k = \frac{N_k^a +\gamma}{N^a +K\gamma}$, locally update \begin{align}\label{eq:QofZ} r_k^{a(n)}= \frac{ \pi_k^a f(\bm{z}^{a(n)} \mid \bm{\theta}_k) }{ \sum_{m=1}^K \pi_m^a f(\bm{z}^{a(n)} \mid \bm{\theta}_m) }. \end{align} \end{itemize} Here, we assumed that we have used the Dirichlet distribution $p(\bm{\pi}^a)\propto (\pi^a_1\cdots\pi^a_K)^\gamma$ for $p({\bm \Pi})$. Notice that agent-agent communication is involved only in the second step; All the other steps need only local computation that can be complete within each agent. The complexity per agent per iteration is $\mathcal{O}(N^a + M^3 + \ln S)$, assuming Eq.~\eqref{eq:solve} takes $M^3$ and the consensus step Eq.~\eqref{eq:aggregate_global_exponential_family} takes $\ln S$ (See Theorem~\ref{prop:expander}). \subsection{Gaussian example} \label{app:GaussianExample} To be concrete, we provide parameter updating equations for the Gaussian observation model with the Gauss-Laplace prior Eq.~\eqref{eq:Gauss-Laplace}. In this case, instead of $\bm{T}^a_k$ in Eq.~\eqref{eq:localStats}, we compute \begin{align} \bm{m}^a_k = \sum_{n=1}^{N^s}r_k^{a(n)}\bm{x}^{a(n)}, \;\;\; \mathsf{C}^a_k = \sum_{n=1}^{N^s}r_k^{a(n)}\bm{x}^{a(n)}{\bm{x}^{a(n)}}^\top. \end{align} for each $a\in\{1,\ldots, S\}$ and each mixture component $k\in\{1,\ldots,K\}$. Then, in the step of global consensus in Eq.~\eqref{eq:aggregate_global_exponential_family}, we compute aggregated values as \begin{align}\label{eq:Gaussian-consensus} \bar{\bm{m}}_k = \sum_{a=1}^S \bm{m}^a_k, \;\;\; \bar{\mathsf{C}}_k = \sum_{a=1}^S \mathsf{C}^a_k. \end{align} Finally, in the step of optimization in Eq.~\eqref{eq:solve}, we compute \begin{gather} \label{eq:Nk} \bm{\mu}_k = \frac{1}{\lambda_0 + N_k}\bar{\bm{m}}_k,\;\;\; \mathsf{\Sigma}_k = \frac{1}{N_k}\bar{\mathsf{C}}_k + \bm{\mu}_k \bm{\mu}_k^\top, \\ {\sf \Lambda}_k = \arg \max_{{\sf \Lambda}_k}\left\{ \ln \det{\sf \Lambda}_k - \mathrm{Tr}({\sf \Lambda}_k\mathsf{\Sigma}_k) -\frac{\rho}{N_k}\|{\sf \Lambda}_k \|_1 \right\}, \end{gather} where $\mathrm{Tr}$ is the matrix trace and $\det$ is the matrix determinant. This optimization problem is well-known in covariance selection and can be solved very efficiently with the graphical lasso algorithm~\cite{Friedman08glasso,hsieh2014quic}. To ensure that all the agents have the same $\{\bm{\mu}_k,{\sf \Lambda}_k\}$, they can run another global consensus step to register the average as the final outcome in each iteration round. \section{Aggregation via secret sharing}\label{sec:Aggregation} This section focuses on Eq.~\eqref{eq:aggregate_global_exponential_family}, i.e.,~how to securely aggregate the local statistics. Since aggregation can be done element-wise, without loss of generality, we consider the problem of computing the sum of scalars $\{ \xi_a \}$: \begin{align}\label{eq:averageConsensus} \bar{\xi} = \sum_{a=1}^S \xi_a = \bm{1}_S^\top\bm{\xi}(0), \end{align} where $\xi_a$'s are constants to be summed, $\bm{1}_S$ is the $S$-dimensional vector of ones, and we defined $\xi_a(0) = \xi_a$. Aggregation would be trivial if a trusted coordinator existed in the network. The question is how to compute the summation only through local communications and how to make it secure. \subsection{Aggregation through Markov transitions on graph}\label{subsec:dynamicalConsensus} Let $\mathsf{A} \in \{0,1\}^{S\times S}$ be the incidence matrix of the communication graph, where only connected nodes (or neighboring nodes) can communicate with each other. As illustrated in Fig.~\ref{fig:S7expander_model.pdf}, the graph is undirected but may have self-loops and multiple edges. Given $\mathsf{A}=[A_{a,j}]$, consider the following updates: \begin{align}\label{eq:Dynamics_elementwise} \xi_a(t+1) = \xi_a(t) + \epsilon \sum_{j=1}^S {A}_{a,j}[\xi_j(t) - \xi_a(t)], \end{align} where $t$ is the number of update rounds, and $\epsilon$ is a given parameter controlling convergence. All the $S$ nodes perform this update by communicating with their neighbors. In the matrix form, Eq.~\eqref{eq:Dynamics_elementwise} is written as \begin{align}\label{eq:Dynamics_matrix} \bm{\xi}(t+1) =\mathsf{W}_\epsilon \bm{\xi}(t) \;\;\; \mbox{with} \;\;\; \mathsf{W}_\epsilon \triangleq \mathsf{I}_S - \epsilon( \mathsf{D}-\mathsf{A}), \end{align} where $\mathsf{D}\triangleq\mathrm{diag}(d_1,\ldots,d_S)$ is the degree matrix with $d_s$ being the degree of the $s$-th node, $\mathsf{I}_S$ is the $S$-dimensional identity matrix, and $\bm{\xi}(t) \triangleq (\xi_1(t),\ldots,\xi_S(t))^\top$. The key idea of multi-agent coordination is to associate a stationary solution of the Markov transition defined by Eq.~\eqref{eq:Dynamics_matrix} with the process of consensus building. As can be easily verified, $\bm{u}_1 = \frac{1}{\sqrt{S}}\bm{1}_S$ is an $\ell_2$-normalized eigenvector of $\mathsf{W}_\epsilon$ whose eigenvalue is $\lambda_1=1$. If this is non-degenerated and the other absolute eigenvalues are less than one, Eq.~\eqref{eq:Dynamics_matrix} will converge to the stationary solution $\bm{\xi}^*$ \begin{align}\label{eq:achievingConv} \bm{\xi}^* = {\mathsf{W}_\epsilon }^\infty \bm{\xi}(0) \approx {\lambda_1}^\infty \bm{u}_1 \bm{u}_1^\top \bm{\xi}(0)= \frac{1}{S}\bar{\xi}\bm{1}_S \end{align} because only the largest eigenvalue survives in the spectral expansion after an infinite number of transitions~\cite{Strang1976}. This means that all of the agents end up having the same value of $\frac{1}{S}\bar{\xi}$, which is the aggregation we wanted. We call this approach the \textit{dynamical consensus algorithm}. One obvious limitation of this approach is that the connected peers can see the original value in the first iteration, which is a privacy breach. To address this issue, we propose two secret sharing approaches: One is Shamir's secret sharing~\cite{shamir1979share} and the other is random chunking we proposed originally in~\cite{ide2019efficient}. \subsection{Shamir's secret sharing}\label{subsec:Shamir} In Shamir's secret sharing scheme, each of the $S$ agents first generates random numbers $R^a_1, \ldots, R^a_{S-1}$ in a large enough integer domain to define an $(S-1)$-th order polynomial function. For the $a$-th agent, the polynomial is defined by \begin{align} {g}_a(n) = \xi_a + R^a_1 n +\cdots + R^a_{S-1} n^{S-1}.\label{eq:Shamir-what-is-sent} \end{align} Using this polynomial, each agent locally computes $S$ values $\{ {g}_s(1), \ldots, {g}_s(S) \}$. Then, by repeating the dynamical consensus algorithm $S$ times (i.e.,~by setting $\xi_s(0) = g_s(n)$ for $n=1,\ldots, S$ in Eq.~\eqref{eq:Dynamics_elementwise}), the agents build a consensus on $S$ aggregated values $\{ \bar{g}(1), \ldots, \bar{g}(S) \}$, where $\bar{g}(l) \triangleq \sum_{s=1}^{S}g_s(l)$. Since $R^a_1$ etc. are random numbers, raw data $\{\xi_a\}$ will not be revealed to the peers in the communication process. At this point, each agent has the $S$ input-output pairs $\{ (1,\bar{g}(1)),\ldots, (S,\bar{g}(S)) \}$ of the polynomial \begin{align} \bar{g}(n) = \bar{\xi} + \bar{R}_1 n +\cdots + \bar{R}_{S-1} n^{S-1} \end{align} at hand, where $\bar{R}_i \triangleq \sum_{a=1}^S R^a_{i}$. Notice that we have $\bar{\xi}$ on the r.h.s.~as the intercept. Since an $(S-1)$-th order polynomial is uniquely determined by distinctive $S$ points, each agent can uniquely identify the functional form of $\bar{g}(n)$. With Lagrange's interpolation formula, the agents obtain the intercept by \begin{align}\label{eq:LagrangeInterpolation} \bar{\xi} = \bar{g}(0) = \sum_{l=1}^S \bar{g}(l) \prod_{m\neq l} \frac{m}{(m-l)}. \end{align} Shamir's algorithm is secure in the sense that it satisfies rigorous cryptographic conditions~\cite{boneh2016graduate}. However, when $S$ is large, the product form of Eq.~\eqref{eq:LagrangeInterpolation} needs a special attention to avoid numerical issues. Also, the need for $S$ repeated consensus makes it computationally less efficient, as will be empirically shown in Section~\ref{sec:Experiments}. \subsection{Random chunking}\label{subsec:chunking} Shamir's algorithm is based on the idea of recovering the data from seemingly non-informative multiple signatures. Here it is interesting to see what happens if we use a datum itself as the signature. Specifically, each of the agents secretly splits a local datum $\xi_s$ into a few chunks $\{ \xi_s^{[h]}\}$ such that $\sum_h \xi_s^{[h]} = \xi_s$. Then, the agents run the dynamical consensus algorithm on each of the chunks to get $\{\bar{\xi}^{[1]}, \ldots, \bar{\xi}^{[N_\mathrm{C}]}\}$, where $N_\mathrm{C} \ll S$ is the predefined number of chunks. Since aggregation is an linear operation, we have \begin{align} \bar{\xi}^{[1]}+\cdots+\bar{\xi}^{[N_\mathrm{C}]} = \sum_{s=1}^S (\xi_s^{[1]}+\cdots +\xi_s^{[N^c]}) =\bar{\xi}. \end{align} This means that the total aggregation can be computed by performing chunk-wise aggregations without sharing the raw data $\{\xi_s\}$ explicitly. The procedure is summarized in Algorithm~\ref{algo:conChunk}. \begin{algorithm}[t] \caption{Dynamical consensus with random chunking} \label{algo:conChunk} \begin{algorithmic}[1] \STATE Input: $\epsilon$, $N_\mathrm{C}$. Initialize $\bar{\xi}=0$ \STATE Split $\xi_a$ into $N_\mathrm{C}$ chunks for $\forall a$. \FOR {$i_\mathrm{C} \leftarrow 1, N_\mathrm{C}$} \STATE Randomly generate $\mathsf{A}$. Confirm all the links are available for communication (re-generate $\mathsf{A}$ otherwise). \STATE Initialize $\xi_a(0)$ as $\xi_a^{[i_\mathrm{C}]}$ for $\forall a$. \REPEAT \STATE Each agent synchronously perform~\eqref{eq:Dynamics_elementwise}. \UNTIL{convergence} \STATE $\bar{\xi} \leftarrow \bar{\xi} + \bar{\xi}^{[i_\mathrm{C}]}$ \ENDFOR \end{algorithmic} \end{algorithm} Here, to prevent the agents from recovering the raw data by receiving all the $N_\mathrm{C}$ chunks, the chunk-wise aggregations must be made on different incidence matrices. There are two major solutions having different levels of intervention of the network infrastructure (see Section~\ref{subsec:setting}). For the higher intervention side, upon starting aggregation, the router may randomly pick $\mathsf{A}$ from a set of prepopulated incidence matrices, in which any pair of the graphs do not share the same edge. If the router always picks an $\mathsf{A}$ that is different from the previous aggregation round, $N_\mathrm{C}=2$ suffices. If we cannot expect too much from the infrastructure, the security guarantee becomes probabilistic. This is reminiscent of Bitcoin's probabilistic security guarantee, which makes it the \textit{probabilistic finality} protocol~\cite{xiao2020survey}. In the next section, we discuss the detail of probabilistic guarantee of the random chunking algorithm. \section{Privacy Breach Analysis}\label{sec:security} This section provides probabilistic guarantees of the random chunking algorithm under three major attack scenarios. \subsection{Independent agent scenario} As suggested in Algorithm~\ref{algo:conChunk}, one practical scenario is for the router to randomly assign the node name whenever starting aggregation, keeping the graph structure itself fixed. To study the risk of privacy breach, consider the event that the $j$-th node is the breach node to $s$, meaning that the node $j$ is able to fully recover $\xi_s$ by summing over $N_\mathrm{C}$ chunks it received. This happens when the $j$-th node stays as the neighbor over the $N_\mathrm{C}$ rounds, regardless of the other nodes. Since the $j$-th node gets chosen in each of the $N_\mathrm{C}$ chunking round with a probability \begin{align} \binom{S-2}{d_s-1}\binom{S-1}{d_s}^{-1} =\frac{d_s}{S-1}, \end{align} the probability that the $j$-th node is the breach node to $s$ regardless of the other nodes is given by $\left(\frac{d_s}{S-1}\right)^{N_\mathrm{C}}$. Since some of the other $S-2$ nodes may be the breach node as well, we have \begin{align}\label{eq:P_breach-byChance} p_\mathrm{breach}^s \leq \sum_{j \neq s}\left( \frac{d_s}{S-1} \right)^{N_\mathrm{C}} = (S-1)\left( \frac{d_s}{S-1} \right)^{N_\mathrm{C}} \end{align} by Boole's inequality (or the union bound). Therefore, we conclude that the probability of not having any breach node in the network is lower-bounded as \begin{align} p_\mathrm{secure} &\geq \prod_{s=1}^S\left\{1- (S-1) \left( \frac{d_s}{S-1} \right)^{N_\mathrm{C}}\right\}\nonumber\\ &\geq 1- S(S-1) \left( \frac{d_\mathrm{max}}{S-1} \right)^{N_\mathrm{C}},\label{eq:p_secure_bound3} \end{align} where the second inequality is due to Bernoulli's inequality. The second term of the bound can be made arbitrarily small for \textit{sparse graphs} $d_\mathrm{max} \ll S$ by appropriately choosing the value of $N_\mathrm{C}$. To summarize, we have proved the following theorem: \begin{theorem}\label{prop:breach} The dynamical consensus algorithm with random chunking has a privacy guarantee of Eq.~\eqref{eq:p_secure_bound3}. The probability of privacy breach can be made arbitrarily small. \end{theorem} \subsection{Colluded agent scenario}\label{app:collusion_prob} Another interesting question is whether the algorithm is secure under collusion. Although the risk of collusion is minimal in consortium-based networks, where the identity and the motivation of the agents are known to each other, we have the following guarantee: \begin{theorem}\label{prop:collusion} Let $N_\mathrm{L}$ be the number of colluded agents. The probability that the privacy of a non-colluded node (indexed by $s$) is compromised due to collusion is given by \begin{align} \label{eq:collusion_equality} p_\mathrm{c.breach}^s &= \left\{ 1 - \prod_{l=1}^{N_\mathrm{L}}\left(1 - \frac{d_s}{S-l}\right) \right\}^{N_\mathrm{C}}\\ &\leq \label{eq:collusion} \exp\left\{-N_\mathrm{C} \left( 1 - \frac{d_s}{S-N_\mathrm{L}}\right)^{N_\mathrm{L}}\right\} \end{align} for $N_\mathrm{L} < S-d_s$, and 1 otherwise. \end{theorem} \noindent (\textit{Proof}) For an $s$-th non-colluded node, privacy breach occurs when at least one colluded node receives the chunk in all the $N_\mathrm{C}$ chunking rounds. Therefore, we have \begin{align}\label{eq:pL1} p_\mathrm{c.breach}^s = (1 -p_\mathrm{L})^{N_\mathrm{C}}, \end{align} where $p_\mathrm{L}$ is the probability that no colluded nodes are chosen in the neighbor set. The probability $p_\mathrm{L}$ can be evaluated as \begin{align} p_\mathrm{L} &= \binom{S-1-N_\mathrm{L}}{d_s}\binom{S-1}{d_s}^{-1} \nonumber\\ &= \frac{ (S-d_s-1)(S-d_s-2)\cdots (S-d_s-N_\mathrm{L}) }{ (S-1)(S-2) \cdots (S-N_\mathrm{L}) },\label{eq:pL2} \end{align} which holds only for $N_\mathrm{L} \leq S-d_s -1$. If there are so many colluded nodes that $N_\mathrm{L} \geq S-d_s$ holds, it is impossible not to choose a colluded nodes $N_\mathrm{C}$ times in a row and thus $p_\mathrm{L}$ must be zero. From Eqs.~\eqref{eq:pL1} and~\eqref{eq:pL2}, we have the equality of Theorem~\ref{prop:collusion}: \begin{align}\label{eq:collusion-breach-equality} p_\mathrm{c.breach}^s &= \left\{ 1 - \prod_{l=1}^{N_\mathrm{L}} \left( 1 - \frac{d_s}{S-l} \right)\right\}^{N_\mathrm{C}}. \end{align} Next, let us derive the upper bound. By replacing the product in Eq.~\eqref{eq:collusion-breach-equality} with the power of the minimum term and apply the Bernoulli inequality, we have \begin{align} p_\mathrm{c.breach}^s &\le \left\{ 1 - \left( 1 - \frac{d_s}{S- N_\mathrm{L}} \right)^{N_\mathrm{L}}\right\}^{N_\mathrm{C}}\nonumber\\ &\leq \exp\left\{-N_\mathrm{C} \left( 1 - \frac{d_s}{S-N_\mathrm{L}}\right)^{N_\mathrm{L}}\right\}.\label{eq:c-breach-better}, \end{align} which completes the proof. $\square$ Figure~\ref{fig:CollusionBreach-bounds.pdf}~(a) visualizes the privacy breach probability due to collusion (Eq.~\eqref{eq:collusion}) as a function of $N_\mathrm{C}$ and $N_\mathrm{L}$. As expected, the probability is almost one if $N_\mathrm{L}\sim \frac{S}{2}$ and $N_\mathrm{C} \sim 1$. However, we also see that the risk can be made negligible by properly choosing $N_\mathrm{C}$, as claimed in Theorem~\ref{prop:collusion}. Also, Fig.~\ref{fig:CollusionBreach-bounds.pdf}~(b) shows comparison among the exact probability and its upper bound computed by Eq.~\eqref{eq:c-breach-better}. \begin{figure}[tb] \begin{center} \includegraphics[trim={0cm 0cm 0cm 0cm},clip,height=4.5cm]{./fig/CollusionBreachHeatMap.pdf} \includegraphics[={0cm 0cm 0cm 0cm},clip,width=4.2cm]{./fig/CollusionBreach-bounds.pdf} \end{center}\vspace{-0.4cm} \caption{Privacy breach probability due to collusion computed for $3$-regular graph of $S=100$. (a) Left panel: $p_\mathrm{c.breach}^s$ computed by Eq.~\eqref{eq:collusion_equality} with a gradation from red being 1 to white being 0. (b) Right panel: $p_\mathrm{c.breach}^s$ (solid lines) and its upper bound (Eq.~\eqref{eq:c-breach-better}; dashed lines) at a few $N_\mathrm{C}$ values. Best viewed in color. } \label{fig:CollusionBreach-bounds.pdf} \end{figure} The inequality enables us to reasonably set the number of chunks, $N_\mathrm{C}$, so that the privacy of all non-colluded nodes is guaranteed with a sufficiently high probability, say, at least $1 - \eta$ for a small $\eta$. This corresponds to $p_\mathrm{c.breach}^s \leq \eta$ and is translated into \begin{align} N_\mathrm{C} \geq | \ln \eta | \left( 1 - \frac{d_s}{S-N_\mathrm{L}} \right)^{- N_\mathrm{L}}. \end{align} As long as the graph is sparse ($d_s \ll S$) and $N_\mathrm{L}$ is expected to be much smaller than $S$, a useful rule-of-thumb will be $2| \ln \eta |$. In consortium-based networks, $N_\mathrm{L}$ should be much smaller than $S$. In that case, the use of graphs having $d_s \ll S$ is critical to guarantee data privacy under the risk of collusion. Again, this motivates us to consider a family of \textit{sparse graphs}, as discussed in the next section. \subsection{Eavesdropping scenario} \label{app:eavesdropping} Eavesdropping is one of the major attacks in communication network~\cite{karlof2003secure}. Although the risk of eavesdropping is limited in consortium-based networks, we can compute the probability of privacy breach in a similar way to the incidental privacy breach. The Shamir-based algorithm is safe even under eavesdropping and collusion because each agent sends only obfuscated values as Eq.~\eqref{eq:Shamir-what-is-sent} to the connected agents. In the random chunking algorithm, however, we again have a probabilistic guarantee. Consider the scenario that an eavesdropper picks a set of the edges and captures all the contents of communications. Let $N_\mathrm{E}$ be the maximum number of edges the eavesdropper can tap. Let $p^s_\mathrm{e.breach}$ be the probability that the $s$-th agent incurs a privacy breach under eavesdropping. A breach occurs when the eavesdropper receives all of the $N_\mathrm{C}$ chunks. This happens when the eavesdropper successfully tapped any of the $d_s$ edges from the $s$-th agent in every chunking round. Thus \begin{align}\nonumber p^s_\mathrm{e.breach} =& [1-\mathrm{Pr}(\mbox{No edges from $s$ are eavesdropped})]^{N_\mathrm{C}}, \\ \nonumber =& [ 1-\mathrm{Pr}(\mbox{All the $N_\mathrm{E}$ edges are chosen} \\ \nonumber &\quad \quad \quad \mbox{from the other $E-d_s$ edges})]^{N_\mathrm{C}}, \\ \nonumber &= \left\{ 1- \binom{E-d_s}{N_\mathrm{E}}\binom{E}{N_\mathrm{E}}^{-1} \right\}^{N_\mathrm{C}} \\ \label{eq:eavasdrop_breach_prob_equality} &= \left\{ 1 -\prod_{l=0}^{d_s -1} \left( 1 - \frac{N_\mathrm{E}}{E - l}\right) \right\}^{N_\mathrm{C}} \end{align} where $\mathrm{Pr}(\cdot)$ denotes the probability of the event specified by the argument, and $E \triangleq \sum_{i=1}^S d^i$ is the total number of edges of the graph. Using the Bernoulli's inequality as in Eq.~\eqref{eq:c-breach-better}, we have \begin{align} p^s_\mathrm{e.breach} & \leq \left\{ 1- \left(1 - \frac{N_\mathrm{E}}{E-d_s+1} \right)^{d_s} \right\}^{N_\mathrm{C}}\nonumber \\ \label{eq:eavasdrop_prob_bound} &\leq \exp\left\{ -N_\mathrm{C} \left(1 - \frac{N_\mathrm{E}}{E-d_s+1} \right)^{d_s} \right\}. \end{align} We summarize this result as bellow: \begin{theorem}\label{th:evasdropping} The probability of privacy breach due to eavesdropping in the dynamic consensus algorithm with random chunking is upper-bounded as Eq.~\eqref{eq:eavasdrop_prob_bound}. \end{theorem} Note that for $d$-regular graphs, $E=dS$. Since $S \gg d_s$ for sparse graphs, as long as the infected edge ratio $\frac{N_\mathrm{E}}{E}$ is expected to be much smaller than one, we can always make $p^s_\mathrm{e.breach}$ negligible by choosing a sufficiently large $N_\mathrm{C}$. To keep $p^s_\mathrm{e.breach}$ less than $\eta >0$, we need \begin{align} N_\mathrm{C} \geq |\ln \eta| \left(1 - \frac{N_\mathrm{E}}{E-d_s+1} \right)^{-d_s} . \end{align} For sparse graphs, unless a majority of edges are tapped, $N_\mathrm{C} \sim 2|\ln \eta|$ again is a useful rule-of-thumb. Figure~\ref{fig:eaves-line} illustrates $p^s_\mathrm{e.breach}$ as a function of $N_\mathrm{E}/E$ and $N_\mathrm{C}$ for a $3$-regular graph with $S=100$. As shown, even under massive eavesdropping where 20\% of the edges are tapped, $N_\mathrm{C}=6$ gives only about 1\% of breach risk. \begin{figure}[tb] \begin{center} \includegraphics[={0cm 0cm 0cm 0cm},clip,width=4.5cm]{./fig/eaves-heatmap.pdf} \includegraphics[={0cm 0cm 0cm 0cm},clip,width=4.2cm]{./fig/eaves-line-bound.pdf} \end{center}\vspace{-0.4cm} \caption{Privacy breach probability due to eavesdropping computed for $3$-regular graph of $S=100$. (a) Left panel: $p^s_\mathrm{e.breach}$ computed by Eq.~\eqref{eq:eavasdrop_breach_prob_equality} as a function of $N_\mathrm{C}$ and the infected edge ratio $N_\mathrm{E}/E$ with a gradation from red being 1 to white being 0. (b) Right panel: $p^s_\mathrm{e.breach}$ (solid lines) and its upper bound (Eq.~\eqref{eq:eavasdrop_prob_bound}; dashed line) at a few values of $N_\mathrm{C}$. Best viewed in color.} \label{fig:eaves-line} \end{figure} In addition to the above case, we may consider the case where an eavesdropper hijacks an agent and capture all the outgoing communications. In this case, the random chunking algorithm is no longer secure although the Shamir-based method is still safe. This issue may be handled by combining with a more sophisticated secret sharing scheme such as in~\cite{gupta2017privacy}. Exploring this topic would be an interesting future topic. \section{Network topology analysis} In the previous section, we pointed out that the graph sparsity plays an important role in privacy preservation. This section discusses another important aspect of the network topology: \textit{Spectral} structure. Specifically, we discuss properties of a specific type of graph called the \textit{expander graph}. \subsection{Estimating the number of iterations to converge} In the dynamic consensus algorithm discussed in Section~\ref{subsec:dynamicalConsensus}, the convergence speed is governed by the ratio between the first and second absolute largest eigenvalues. Let $\lambda_1=1, \lambda_2, \ldots, \lambda_S$ be the eigenvalues of $\mathsf{W}_\epsilon$ arranged in the decreasing order. For ease of exposition, let us assume that $\epsilon$ has been chosen so $\lambda_2$ is the absolute second largest as well, which is always possible. We are interested in the scalability of the algorithm in terms of the network size $S$. Let us first think about how the convergence rate is evaluated. Equation~\eqref{eq:achievingConv} shows that $$ \| \sqrt{S} { \mathsf{W}_\epsilon }^t \bm{\xi}(0) \|_2 \to \left\| \frac{1}{\sqrt{S}}\bm{1}_S \bar{\xi} \right\|_2 = |\bar{\xi}| $$ as $t \to \infty$, where $\|\cdot\|_2$ denotes the $\ell_2$ norm. Thus it makes sense to define the relative error at iteration round $t$ as \begin{align} e_\epsilon(t) \triangleq \frac{1}{|\bar{\xi}|}\sqrt{\| \sqrt{S} { \mathsf{W}_\epsilon }^t \bm{\xi}(0) \|_2^2 - \bar{\xi}^2 } \end{align} Let $\bm{v}_2$ be the $\ell_2$-normalized eigenvector corresponding to $\lambda_2$. As $t$ grows to infinity, we asymptotically have: \begin{align} e_\epsilon(t) \to \sqrt{S} (\lambda_{2})^t \times \frac{\bm{v}_2^\top \bm{\xi}(0)}{\bar{\xi}} \sim \sqrt{S}(\lambda_{2})^t \times \mathrm{O}(1). \end{align} Thus, we conclude that $\sqrt{S}(\lambda_{2})^t$ is a reasonable nondimensional metric for convergence. To achieve a relative error $\delta$, we need a number of iterations as $ t \sim \mathrm{O}\left( \frac{\ln (\sqrt{S}/\delta)}{ |\ln\lambda_{2}|} \right) $. Let us summarize this result in Theorem~\ref{prop:iterationOrder}: \begin{theorem}\label{prop:iterationOrder} In the dynamical consensus algorithm, the number of iterations $t$ to achieve a relative error $\delta$ is given by \begin{align}\label{eq:number_of_iterations_wrt_S2} t \sim \mathrm{O}\left( \frac{\ln (\sqrt{S}/\delta)}{ |\ln(1 - \Delta_\lambda)|} \right), \end{align} where $\Delta_\lambda \triangleq \lambda_1 - \lambda_2$. \end{theorem} In Eq.~\eqref{eq:number_of_iterations_wrt_S2}, we have used the fact that $\lambda_1 =1$ and thus $\lambda_2 = 1 -\Delta_\lambda $. In our previous work~\cite{ide2019efficient}, we showed that the cycle graph scales quadratically as $t \sim S^2 \ln S$. The question is whether we can improve this (relatively poor) scalability by using a different communication graph. The next subsection answers this question. \subsection{Expander graph}\label{subsec:expander} Intuitively, we expect that the more information the agents circulate, the faster convergence we will get. The complete graph is clearly an extreme case, where the agents share information most generously with the peers and thus data privacy is least protected. Let $\mu_1,\ldots,\mu_S$ be the eigenvalues of $\mathsf{A}$ in the decreasing order. In the complete graph, it is easy to verify that $\mu_1 = S-1$ and $\mu_2 = \cdots = \mu_S=-1$. Hence, we have $\Delta_\lambda = \epsilon(\mu_1-\mu_2) = \epsilon S$. The complete graph is a $d$-regular graph with $d=S-1$. By Theorem~\ref{prop:iterationOrder}, with a choice of $\epsilon \sim \mathrm{O}(1/d)$, the convergence speed measured by $t$ scales as $\ln S$, which is much better than the quadratic scalability of the cycle graph. Is there any \textit{sparse} graph that behaves like the complete graph when communicating with the other agents? This may sound like a ridiculous question, but surprisingly, there exists a class of sparse graph called the \textit{expander graph} that has the same logarithmic convergence. Formally, the expander graph is defined as a graph whose \textit{expansion constant} is lower-bounded, where \begin{align} \mbox{(expansion constant)}\triangleq \inf_{\mathcal{V}_1}\left\{ \frac{|\partial \mathcal{V}_1|}{\min\{|\mathcal{V}_1|, S-|\mathcal{V}_1|\}} \right\} \end{align} with $\mathcal{V}_1$ being an arbitrary subset of the graph nodes and $|\partial \mathcal{V}_1|$ is the number of outgoing edges from the subgraph~\cite{davidoff2003elementary}. Intuitively, the more the expansion constant is, the more ``talkative'' the nodes tend to be. Very interestingly, this purely geometric definition has a hidden connection to the graph spectrum, and, in our case, we have \begin{align}\label{eq:Delta_lambda_expander} \Delta_\lambda \geq \epsilon \frac{\alpha^2}{2d} \quad \mbox{($d$-regular expander graphs)} \end{align} by Cheeger's inequality~\cite{brouwer2011spectra}, where $\alpha$ is the lower bound of the expansion constant. Thus we again have the logarithmic scalability $t\sim \ln S$ with a choice of $\epsilon \sim \mathrm{O}(1/d)$. We now summarize the result as follows. \begin{theorem}\label{prop:expander} On $d$-regular expander graphs, $t \sim \mathrm{O}(\ln (\sqrt{S}/\delta))$ with a choice of $\epsilon \sim \mathrm{O}(1/d)$. \end{theorem} Construction of expander graphs is nontrivial. Fortunately, there are two ways available to obtain expander graphs for our purpose. The first one is a $3$-regular expander graph called the cycle with inverse chords~\cite{vadhan2012pseudorandomness}. As the name suggests, it starts with the cycle graph (2-regular graph), and connects each node $s$ to the nodes $j = s\pm 1, (s-1)(j-1) =1 \mod S$ for a prime $S$. See Figs.~\ref{fig:S7expander_model.pdf} and~\ref{fig:ExpanderExample} for a few examples. As mentioned before, the node indices need to be randomly shuffled prior to each round of the random chunking algorithm. Since the eigenspectrum is invariant to re-labeling the nodes, it does not affect the convergence of dynamic consensus. \begin{figure}[t] \begin{center} \includegraphics[trim={0cm 0cm 0cm 0cm},clip,height=2.5cm]{./fig/S11.pdf} \includegraphics[trim={0cm 0cm 0cm 0cm},clip,height=2.5cm]{./fig/S19.pdf} \includegraphics[trim={0cm 0cm 0cm 0cm},clip,height=2.5cm]{./fig/S31.pdf} \includegraphics[trim={0cm 0cm 0cm 0cm},clip,height=2.5cm]{./fig/S47.pdf} \includegraphics[trim={0.75in 2in 0.75in 2in},clip,height=2.5cm]{./fig/S79-.pdf} \includegraphics[trim={0.75in 2in 0.75in 2in},clip,height=2.5cm]{./fig/S97-.pdf} \end{center}\vspace{-0.4cm} \caption{Example of 3-regular expander graphs with $S=11,19,31$ (top row) and $47, 79, 97$ (bottom row) from left to right. Note that self-loops and double edges (thick lines) exist.} \label{fig:ExpanderExample} \end{figure} The other possible approach is to use a random graph. It is known that uniformly sampled $d$-regular random graphs approximate $\Delta_\lambda$ of the expander graph Eq.~\eqref{eq:Delta_lambda_expander} almost always perfectly~\cite{Friedman2004A}. To construct such a graph, we can leverage the configuration model in random graph theory~\cite{kim2006generating}. Specifically, each agent simply picks $d$ neighbors randomly and uniformly to form a $d$-regular graph. This approach is preferable in the sense that it does not require any intervention of the network router and does not assume network connection that is stable throughout the entire consensus process. On the other hand, however, the randomness in the network structure may lead to some unpredictability in the eigenspectrum, which is in contrast to the first option. To avoid extra randomness, we use the cycle with inverse chords in the empirical study in the next section. Evaluating and designing random graph construction algorithms in dynamic consensus is an interesting future research topic. \section{Experiments}\label{sec:Experiments} This section reports on experimental results of the proposed model. All the experiments were conducted locally on a laptop PC with a Core i7 Processor and 32 GB memory. \subsection{Number of iterations to converge} Figure~\ref{fig:experiments} compares between the expander graph (cycle with a random chord) and the 2-regular cycle graph (or the ring) on the number of iterations $t$ to achieve $\delta = 10^{-3}$ as defined in Theorem~\ref{prop:iterationOrder} and~\ref{prop:expander}. It is interesting to observe that just adding an extra edge to the ring by the rule $(s-1)(j-1) =1 \mod S$ drastically changes the convergence behavior. As mentioned before, the ring scales quadratically as $t \sim S^2 \ln S$. In contrast, $t$ grows very slowly in the expander graph, being consistent to the logarithmic scalability predicted by Theorem~\ref{prop:expander}. The original construction of the 3-regular expander graph is for $S$ that is a prime. One interesting question in practice is that the excellent convergence behavior is maintained for non-prime $S$'s. We extended the original construction by giving a self-loop whenever the equation $(s-1)(j-1)=1 \mod S$ does not have a solution, so that the graph is always $3$-regular. Fortunately, apart from the visible fluctuations, the figure suggests that the overall convergence behaviors are robust and we should be able to safely use the model even for non-primes. \begin{figure}[t] \begin{center} \includegraphics[trim={0cm 0cm 0cm 0.5cm},clip,height=6cm]{./fig/NumIteration2.pdf} \end{center}\vspace{-0.4cm} \caption{Comparison of the number of iterations $t$ for $\delta = 10^{-3}$. } \label{fig:experiments} \end{figure} \subsection{Shamir vs.~random chunking}\label{subsec:Shamir_vs_chunking} Since one iteration needs one matrix-vector multiplication (see Eq.~\eqref{eq:Dynamics_matrix}), actual computational time may have a different scalability from that of the number of iterations to converge. Specifically, one may expect an extra $S\bar{d}$ factor, where $\bar{d}$ is the mean degree. Figure~\ref{fig:ActualTime_Shamir-chunk_S7to53.pdf} compares actual computation times in different $N_\mathrm{C}$s. We randomly initialized $\xi_s(0)$ with the uniform distribution in $[-1,2]$. Convergence was declared when the relative error gets smaller than $10^{-5}$. The mean and standard deviation (s.d.) were computed by repeating computation by 10 times with a different random seed. The figure confirms the expected \textit{linear} dependency in the random chunking algorithm. Figure~\ref{fig:ActualTime_Shamir-chunk_S7to53.pdf} also compares the Shamir's secret sharing algorithm with the random chunking algorithm. Since the Shamir-based method needs to repeat the process of global consensus $S$ times, the actual computation time is expected to depend quadratically on $S$, which is consistent with the figure. Apart from $S \lesssim 10$, where the both methods are comparable, the Shamir-based method tended to have a larger variability than the random chunking algorithm, which might suggest numerical issues in the implementation. Finally, we comment on the origin of non-smoothness of the computational time in Fig.~\ref{fig:ActualTime_Shamir-chunk_S7to53.pdf}. In the figure, close inspection shows that the computational times are correlated: whenever the Shamir-based method gets much time, so does the random chunking method. This suggests that the source of the non-smoothness is in the spectral structure of the underlying network because the same dynamical consensus algorithm was shared by both. Although we have obtained a very solid result on the convergence behavior, how the spectral gap $\Delta_\lambda$ scales as $S$ increases is not straightforwardly predictable. Further investigation on this point is left to future work. \begin{figure}[tb] \begin{center} \includegraphics[trim={0cm 0cm 0cm 0cm},clip,height=4.28cm]{./fig/ActualTime_Shamir-chunk-3reg-Matrix3.pdf} \includegraphics[={0cm 0cm 0cm 0cm},clip,width=4.28cm]{./fig/ActualTime_Shamir-chunk_S7to53.pdf} \end{center}\vspace{-0.4cm} \caption{Actual computation time for aggregation on the $3$-regular expander graph. The right panel covers the range of $7 \leq S \leq 53$. } \label{fig:ActualTime_Shamir-chunk_S7to53.pdf} \end{figure} \subsection{Shamir vs.~homomorphic encryption} Table~\ref{table:BeatHE} compares computation time for a few small $S$'s between the proposed Shamir-based method and homomorphic encryption (HE)-based method of~\cite{ruan2017secure}. Both approaches have cryptographic security and share the same dynamical consensus algorithm. For a fair comparison, we ran both on the same expander graph and used the same experimental design as above. For HE, we used an implementation of the Paillier cryptosystem~\cite{homomorpheR}, which encrypts and decrypts every communication between the agents with a new key. From the table, and in the light of Fig.~\ref{fig:ActualTime_Shamir-chunk_S7to53.pdf}, we conclude that the proposed random chunking-based method is several orders of magnitude faster than the HE-based alternative. \begin{table}[bt] \caption{Comparison of actual computation times in aggregation [sec].}\label{table:BeatHE} \vspace{-1pt} \centering \begin{tabular}{c|cccc} \toprule & \multicolumn{2}{c}{Shamir} & \multicolumn{2}{c}{HE} \\ \cline{2-5} $S$ & mean & s.d. & mean & s.d. \\ \midrule 7 & 0.159 & 0.024 & 155 & 18.8 \\ 11 & 0.485 & 0.045 & 288 & 44.1 \\ 13 & 0.458 & 0.020 & 375 & 24.8 \\ 17 & 0.671 & 0.060 & 574 & 55.1 \\ 19 & 2.24 & 0.156 & 699 & 47.0 \\ \bottomrule \end{tabular} \vspace{-1pt} \end{table} \subsection{Remarks on synchronization issues} \label{app:smallS-Shamir-RC} In Section~\ref{subsec:Shamir_vs_chunking}, we mentioned that actual computational time of the random chunking algorithm on the expander graph should have a linear dependency on $S$. This statement implicitly assumes that computation is made sequentially. Although in theory it is true that the agents can perform updates in parallel, sequential execution should be a reasonable assumption for evaluating computational time. In real distributed environments, we always need to handle the issue of synchronization across the network. If the cost of local computation is negligible, network delays will almost always dominate the time required to move on to the next iteration. In that case, most of the agents would spend most of the time waiting for the last message to be delivered to one of the agents. On the other hand, if the computational cost is on average much higher than network delays, we will need to consider parallelization as a realistic option. This is actually the case in homomorphic encryption (HE)-based secure computation, as recently pointed out in~\cite{liu2020secure}. There may also be some room for improvement in the HE implementation itself, where we employed~\texttt{homomorpheR}~\cite{homomorpheR} with the key size of $1\,024$ to follow the protocol proposed in~\cite{ruan2017secure,RuanGaoWang2019}, although handling signed floating-point numbers can be a subtle issue (see, e.g.~\cite{cheon2017homomorphic}). The key exchange protocol can be simplified, too. Regarding Table~\ref{table:BeatHE}, it would be an interesting future research topic to compare the random chunking method with a highly optimized HE implementation in a realistic setting. \section{Concluding remarks}\label{sec:Conclusion} We have presented new research directions of Blockchain as a collaborative value co-creation platform rather than a mere immutable data storage. Our platform is designed to respect the values of democracy, diversity, and privacy in its collaborative learning process. As such an instance, we have proposed a multi-task federated learning framework combined with a global consensus-building algorithm. We discussed two topics in the global consensus-building process. The first topic was data privacy. We proposed two secure consensus-building approaches built upon the idea of secret sharing: The Shamir-based dynamic consensus, which has a cryptographic security guarantee, and the random chunking algorithm, which falls into the category of the probabilistic finality protocol. For the latter, we have provided the upper bound of privacy breach under three different attack scenarios. The second topic discussed was the issue of network design in global consensus. Based on the profound results in spectral graph theory, we showed that expander graphs dramatically improve the scalability of the algorithm. We conclude this paper by summarizing a few future research topics: \paragraph{Learning under network errors} We have assumed perfectly synchronized and stable communication among the agents. As suggested in Sections~\ref{subsec:expander} and~\ref{app:smallS-Shamir-RC}, extensions to include network errors is of primary importance to make our platform truly useful. \paragraph{Meta-agreement issues} Another potential issue in practice is how to agree on the learning task itself (choice of the algorithm, data dimensionality $M$, the definition of each dimension, etc.) and how to initiate peer-to-peer communication (who to communicate with). \paragraph{External data privacy} One interesting business scenario is to sell the learned model to external parties. This calls for a guarantee that is different from the \textit{internal} data privacy among the agents. It is known that the classical concept of differential privacy has issues in evaluating noisy real-valued variables~\cite{minami2016differential}. Establishing a method of evaluating the degree of information leak is important. \paragraph{Randomness in graph spectra} As discussed in Section~\ref{subsec:expander}, the origin of the non-smoothness shown in Fig.~\ref{fig:ActualTime_Shamir-chunk_S7to53.pdf} and the consensus algorithm based on random graphs are still an open question. \paragraph{Security analysis} The random chunking algorithm combined with the dynamic consensus algorithm appears to have more flexibility than traditional cryptographic methods. We need to study further the pros and cons of those methods. \paragraph{Use-cases} Finally, we need to develop practical use-cases where the decentralized architecture is truly useful. The lightweight probabilistic privacy guarantee seems suitable in internet-of-things (IoT) applications~\cite{ide2018collaborative}, but more study is needed. We hope that, just as Nakamoto's Bitcoin changed the landscape of financial transaction management, our decentralized collaborative learning platform as the next-generation Blockchain has the potential to change the way of doing business. \begin{small} \section*{Acknowledgement} T.I.~thank Dr.~Sachiko Yoshihama for her insightful suggestions. T.I.~is partially supported by the Department of Energy National Energy Technology Laboratory under Award Number DE-OE0000911. A part of this report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. \end{small} \bibliographystyle{IEEEtran}

Introduction {#sec1} ============ Membranous nephropathy (MN) is one of the leading causes of nephrotic syndrome in adults. Primary MN (PMN) is associated with a significant risk of developing end-stage renal disease, but clinical evolution is overall unpredictable.[@bib1] Until recently, proteinuria was the main biomarker to assess severity and choose therapeutic indications,[@bib2] and it is the primary criterion used to define remissions. Indeed, according to Kidney Disease: Improving Global Outcomes (KDIGO) guidelines, partial remission and complete remission are defined by proteinuria between 0.5 and 3.5 g per day (and less than 50% of the baseline value) and less than 0.5 g per day, respectively.[@bib3] However, proteinuria is an imperfect marker. In some patients, residual proteinuria may long persist in the absence of immunologic activity, because of tubulo-interstitial damage or secondary glomerular changes caused by prolonged disease, and conversely, in other patients with moderate levels of proteinuria, renal function may fast deteriorate.[@bib4] In 2009, Beck *et al.*[@bib5] found that 70% of adult patients with PMN produced autoantibodies against an antigen expressed on podocyte cells, the M-type phospholipase A2 receptor (PLA2R). There is accumulating evidence that high titers of PLA2R antibodies (PLA2R-Abs) are correlated with a lower risk of spontaneous or immunosuppressor-induced remission, a higher risk of nephrotic syndrome, and of progression to end-stage renal disease.[@bib6], [@bib7], [@bib8] Rituximab is effective in achieving immunological remission, which usually precedes clinical remission of proteinuria by several months.[@bib9]^,^[S1--S3](#appsec1){ref-type="sec"} Detection and measurement of PLA2R-Abs therefore play an important role in monitoring disease activity and treatment efficacy. However, there may be discrepancies between immunological and clinical parameters, and the KDIGO guidelines do not specify whether patients with clinical partial remission but persisting immunological activity should be re-treated with immunosuppressive agents despite the harmful effects of prolonged proteinuria and cardiovascular risk.[@bib3] We report here the case of a patient with persistent circulating PLA2R-Abs in the setting of prolonged partial remission and stable renal function, in whom we undertook a third line of immunosuppressive therapy. Case Presentation {#sec2} ================= A 31-year-old man, without any past medical history, developed a nephrotic syndrome in December 2003 (proteinuria 8.2 g/d, serum albumin 2 g/dl) without hematuria or renal failure (serum creatinine 0.9 mg/dl and estimated glomerular filtration rate \[Modification of Diet in Renal Disease formula\] 99 ml/min) ([Table 1](#tbl1){ref-type="table"}, [Figure 1](#fig1){ref-type="fig"}). A kidney biopsy revealed a stage-2 MN. Antiproteinuric treatment with ramipril and amiloride was started. A first line of immunosuppressive treatment with alternating monthly cycles of corticosteroids and chlorambucil was administered, starting in June 2004 for 6 months. This treatment was complicated by azoospermia and neutropenia and was not effective.Table 1Laboratory resultsCharacteristicsNovember 2013September 2014March 2015June 2016August 2016March 2017June 2017Creatinine (mg/dl) (μmol/l)1.31.11.51.41.61.41.6MDRD GFR (ml/min per m^2^)60634854465446Proteinuria (g/d)220.51.11.60.50.1Serum albumin (g/dl)3.23.03.03.02.93.84.0PLA2R-Ab (ELISA)11212412298194400[^1]Figure 1Evolution of creatinine, serum albumin, proteinuria, and anti-PLA2R-Ab. PLA2R-Abs are assessed by enzyme-linked immunosorbent assay (ELISA) and expressed as RU/ml. One year later, the patient still presented with nephrotic syndrome (creatinine 0.9 mg/dl, proteinuria 7 g/d, serum albumin 2.5 g/dl). According to KDIGO recommendations, he received a second line of immunosuppressive treatment with cyclosporine 5 mg/kg per day for 18 months.[S4](#appsec1){ref-type="sec"} The patient achieved partial remission of nephrotic syndrome within 3 months after initiating cyclosporine (creatinine 1.45 mg/dl, proteinuria 2.8 g/d, serum albumin 3.5 g/dl). Cyclosporine was progressively reduced and then withdrawn in August 2007. The patient remained in partial remission of nephrotic syndrome with stable kidney function for the following 10 years. PLA2R-Ab was assessed for the first time by enzyme-linked immunosorbent assay (EUROimmun, Lübeck, Germany) in November 2013, and was positive at 112 RU/ml ([Table 1](#tbl1){ref-type="table"}). Circulating PLA2R-Abs persisted for the next 3 years, PLA2R-Ab level was 194 RU/ml in August 2016. Then, contrary to KDIGO guidelines, despite the absence of nephrotic syndrome or rise of serum creatinine (serum albumin was 2.9g/dl and proteinuria was 1.6 g/d), a third line of immunosuppressive treatment was given with the aim to achieve complete clinical remission. The patient received 2 infusions of rituximab (375 mg/m^2^ per infusion) at 1-week interval. Because the patient developed leukopenia and infertility after chlorambucil treatment, we started with a relatively low dose of rituximab. Six months later, a complete clinical remission was achieved (serum albumin, 3.8 g/dl; proteinuria, 0.5 g/d), associated with partial immunological remission (PLA2R-Ab level dropped to 40 RU/ml). CD19+ B cells rose above 10 per mm^3^. Because of CD19+ B-cell repletion and the risk of relapse associated with persisting PLA2R-Abs,[@bib9] we decided to perform a second course of rituximab (1 g) in March 2017 with the aim of stopping the residual antibody production. Three months after this course, the patient maintained complete remission of nephrotic syndrome (albumin level 4.0 g/dl, proteinuria 100 mg/d) and PLA2R-Abs were for the first time undetectable. The patient remains in complete clinical and immunological remission 2 years later without any side effect of treatment. Discussion {#sec3} ========== PMN is an autoimmune disease caused by autoantibodies directed against podocyte antigens. The commonest antibody is directed against PLA2R. Indications of, and response to, therapy is traditionally assessed by using standard laboratory parameters, such as proteinuria, serum albumin, and serum creatinine levels. Based on KDIGO guidelines,[@bib3] our patient achieved partial remission in 2005, did not relapse, and thus did not present any indication for use of an additional immunosuppressive therapy during the 12 years of follow-up. Here we show that this "wait and see" attitude should be reconsidered based on persistence of immunological activity. The pathogenic role of PLA2R-Abs is not clearly established, but antibody levels correlate with disease activity ([Table 1](#tbl1){ref-type="table"}). A high level of antibodies is a risk factor for development of nephrotic syndrome and end-stage renal disease, for failure to treatment, and recurrence after kidney transplantation.[S5](#appsec1){ref-type="sec"} Antibody levels commonly decrease before spontaneous or induced remissions and rise before relapses. Their persistence usually predicts poor clinical outcome.[@bib6], [@bib7], [@bib8], [@bib9] Clinical parameters (proteinuria, albumin level, and glomerular filtration rate) of our patient remained stable for more than 10 years, but these biomarkers are clearly imperfect and may underestimate smoldering kidney damage induced by this immunological disease ([Table 2](#tbl2){ref-type="table"}). Thus, because of persistent high levels of PLA2R-Ab, we feared subclinical, progressive renal damage with risk of evolution to end-stage renal disease.Table 2Distinct teaching pointsPhospholipase A2 receptor antibody (PLA2R-Ab) levels correlate with disease activity.There may be discrepancies between immunological and clinical parameters.Immunosuppressive treatment of patients in partial clinical remission should be considered because of renal and cardiovascular risk.Rituximab can help achieve complete clinical and immunological remission in patients with partial remission.Detection of PLA2R-Ab by immunofluorescence test (IIFT) is more sensitive than enzyme-linked immunosorbent assay when using the threshold of 20 RU/ml recommended by the manufacturer.Reports suggest that the ultimate goal of the immunosuppressive treatment in PLA2R-Ab--positive patients should be complete disappearance of antibody by IIFT, but this should be confirmed by controlled prospective studies. For more than a decade, the anti-CD20 monoclonal antibody rituximab has been used in patients with MN, and data indicate that rituximab is an effective drug in promoting disease remission in patients with PMN and severe nephrotic syndrome.[@bib9]^,^[S2,S3,S6](#appsec1){ref-type="sec"} An important, but frequently neglected advantage of rituximab is its major impact on the quality of life of patients. Patients receiving rituximab benefit from improvement of symptoms while they are not affected by toxicity of treatment, contrary to those treated with steroids, alkylating agents, or calcineurin inhibitors.[S3--S7](#appsec1){ref-type="sec"} Thus, we decided to treat this patient with rituximab therapy after more than 10 years of sustained partial remission. De Vriese *et al.*[S8](#appsec1){ref-type="sec"} suggested a serologically based approach to manage patients with MN. They based the indication of treatment on classic parameters such as proteinuria and albumin level, but also on PLA2R-Ab titer. For the first time, the authors recommended to treat patients with a high titer of PLA2R-Ab (\>204 RU/ml) with immunosuppressive treatment, regardless of the level of proteinuria, and to evaluate therapy response on clinical remission and evolution of PLA2R-Ab titers. They also suggested that the goal of immunosuppressive treatment should be to decrease the level of antibodies by 90% after 6 months. Our case supports the PLA2R-Ab--driven therapy suggested by De Vries *et al.*,[S8](#appsec1){ref-type="sec"} which we applied to a patient with prolonged partial remission and immunological activity. We treated the patient with rituximab because of persistent high levels of PLA2R-Ab, even though antibody level remained less than 204 RU/ml. We agree with the approach of De Vries *et al.*[S8](#appsec1){ref-type="sec"} and we think that treatment endpoint should be disappearance of immunological activity ([Table 1](#tbl1){ref-type="table"}). Ruggenenti *et al.*[S1](#appsec1){ref-type="sec"} showed that the rate of clinical complete remission is higher when PLA2R disappear (*P* \< 0.0001). Therefore, in our patient, we aimed at complete disappearance of antibodies and we hoped that he would subsequently achieve sustained complete remission. Optimal dosing of rituximab is still controversial, although already in 2010, a pharmacological study in nephrotic patients with PMN showed that B-cell depletion was shorter and that CD19+ B cells recurred before 90 days, whereas depletion continued over 180 days in patients with rheumatoid arthritis.[S9](#appsec1){ref-type="sec"} Because our patient developed leukopenia and infertility after chlorambucil treatment, we started with a relatively low dose of rituximab and we reinfused him 6 months later at a PLA2R-Ab level of 40 RU/ml. Although we could not retrieve archival sera before 2013, one can reasonably anticipate that the 2 lines of treatment with the protocol of Ponticelli and Glasscock[@bib1] and cyclosporine decreased PLA2R-Ab titers, but the disease was still immunologically active during partial remission. De Vries *et al.*,[S8](#appsec1){ref-type="sec"} using the enzyme-linked immunosorbent assay test, suggested that the goal of immunosuppressive treatment should be to decrease the level of antibodies by 90% after 6 months. However, there is a gray zone between 2 RU/ml and 20 RU/ml in which a substantial number of samples remain positive by immunofluorescence. A 90% drop may fall into this gray zone. We agree that enzyme-linked immunosorbent assay should be used for treatment monitoring, but we prefer to use as treatment endpoint the disappearance of antibodies by immunofluorescence. It is likely that the forthcoming KDIGO guidelines will integrate serology in the therapeutic algorithm of patients with MN and thus provide guidance for more personalized medicine as illustrated by this observation. Remaining questions regard PLA2R-Ab thresholds, the place of epitope spreading, and that of genetic risk score. They will hopefully be solved by adaptive, serology-driven randomized controlled trials. Disclosure {#sec4} ========== All the authors declared no competing interests. Supplementary Material {#appsec1} ====================== Supplementary References This study was funded by grants from the European Research Council ERC-2012-ADG_20120314 (Grant Agreement 322947); from 7th Framework Programme of the European Community Contract 2012--305608 (European Consortium for High-Throughput Research in Rare Kidney Diseases). **Supplementary References.** Supplementary material is linked to the online version of the paper at [www.kireports.org](http://www.kireports.org){#intref0010}. [^1]: ELISA, enzyme-linked immunosorbent assay; GFR, glomerular filtration rate; MDRD, Modification of Diet in Renal Disease; PLA2R-Ab, phospholipase A2 receptor antibodies.

Introduction {#Sec1} ============ X-linked hereditary nephropathy (XLHN) in dogs leads to chronic kidney disease (CKD) because of a defect in type IV collagen in the glomerular basement membrane (GBM). In the XLHN dogs in this study, a naturally-occurring, 10-base-pair deletion in the *COL4A5* gene located on the X chromosome results in the inability to synthesize complete α5 chains^[@CR1]^. This alteration in the type IV collagen network compromises the structure and function of the GBM in both affected (hemizygous) males and, to a lesser extent, carrier (heterozygous) female dogs^[@CR2]^. XLHN in dogs is analogous to Alport syndrome (AS) in humans, as approximately 85% of people with AS have an X-linked mutation in *COL4A5* ^[@CR3]^. AS is characterized by juvenile-onset CKD, ocular abnormalities, and hearing loss in affected males^[@CR4]^. Thus far, only the renal abnormalities have been detected in XLHN dogs^[@CR2]^. In dogs with XLHN, juvenile-onset CKD manifests as persistent proteinuria of glomerular origin as early as 3--6 months of age, followed by decreasing glomerular filtration rate and worsening azotemia, typically leading to end-stage renal failure before 1 year of age^[@CR2],[@CR5]^. Although XLHN has been studied as an example of canine CKD caused by glomerular disease and as an animal model of human AS^[@CR2]^, the gene expression profile that affects progression has only been partially characterized. Furthermore, dogs with the same mutation causing XLHN display substantial variation in the rate of disease progression such that some dogs reach end-stage disease by 6 months of age and others at 12 months of age or later. Although varied times of onset and rates of progression are common among different types of mutations in people with AS^[@CR6]^, disease progression may also vary among members of an AS family with an identical mutation^[@CR7]^, as seen in dogs^[@CR8]^. While several studies have characterized gene expression in humans with CKD and in animal Alport models using microarrays^[@CR9],[@CR10]^ or PCR^[@CR8],[@CR9],[@CR11],[@CR12]^, studies that have incorporated high-throughput RNA sequencing (RNA-seq) with the objective of identifying differentially expressed genes (DEGs) and upstream regulators are lacking. Compared with traditional approaches in gene expression analysis, RNA-seq provides unprecedented flexibility in the discovery of DEGs^[@CR13]^ while preserving accuracy and strong correlation with PCR^[@CR14]--[@CR18]^, even considering fold change levels^[@CR19]^. The objective of this study was to compare the gene expression between dogs with rapid versus slow disease progression phenotypes at 3 stages of the disease. We conducted Gene Ontology (GO) and pathway analyses to characterize DEGs among sample groups at specific time points. Since all CKDs share common pathways that lead to end-stage kidney disease^[@CR20]^, the results help elucidate the molecular basis of CKD progression and thus may benefit canine patients and indicate potential therapeutic targets for AS patients. Results {#Sec2} ======= Histopathological evaluation of kidney biopsies {#Sec3} ----------------------------------------------- Figure [1](#Fig1){ref-type="fig"} presents representative cortical fields of the kidney biopsies and the mean interstitial fibrosis scores comparing the rapid versus slow groups. Clinical time points in affected dogs were defined as: T1 - onset of proteinuria (the earliest time point that clinical disease can be detected); T2 - onset of azotemia (serum creatinine ≥ 1.2 mg/dL); and T3 - end-stage disease (serum creatinine ≥ 5 mg/dL). At both T2 and T3, mean fibrosis and chronic inflammation scores were significantly higher in diseased dogs than in controls (Supplementary Tables [S1](#MOESM1){ref-type="media"} and [S2](#MOESM1){ref-type="media"}). The degree of fibrosis in the rapid group was more severe at T2 and T3 than that in the slow group, despite the two groups being clinically indistinguishable; however, statistical significance between the scores was reached only at T3 (Supplementary Table [S1](#MOESM1){ref-type="media"} and Fig. [S1](#MOESM1){ref-type="media"}). No statistically significant difference was observed for the chronic inflammation score between the rapid and slow groups (Supplementary Table [S2](#MOESM1){ref-type="media"} and Fig. [S2](#MOESM1){ref-type="media"}).Figure 1Kidney biopsies from representative dogs and mean interstitial fibrosis scores (score range: 0 \[normal\] to 3 \[severe\]) for each group of affected dogs at 3 time points (T1, T2, and T3). Fibrosis scores were based on evaluation of multiple 20x fields. \*Statistical significant. Scale bar: 100 μm. (Trichrome-stained). RNA-sequencing (RNA-seq) of dog transcriptomes {#Sec4} ---------------------------------------------- The average RNA yield from the 24 kidney biopsies was 86.2 ng/μL, and the average RNA integrity number (RIN) was 3.4 (Supplementary Table [S3](#MOESM1){ref-type="media"}). Because of the variable quality of RNA, the proper library preparation kit was used to compensate for the low-input samples according to the best practice for RNA with variable qualities^[@CR21]^ (see "RNA isolation and sequencing" section in Methods). After performing quality control, we obtained an average of over 30 million paired-end reads from each sample (n = 24). Overall, 91--96% of reads were mapped to the canine genome (CanFam 3.1) by HISAT2^[@CR22]^. Among them, 70--78% of reads were uniquely mapped (Supplementary Table [S3](#MOESM1){ref-type="media"} and Fig. [S3)](#MOESM1){ref-type="media"}. Based on the union setting of HTSeq^[@CR23]^, ambiguous reads that mapped to multiple genes were not included in our analysis. Principal component analysis (PCA) and hierarchical clustering analysis {#Sec5} ----------------------------------------------------------------------- We performed PCA at each time point to determine whether samples in each group clustered with each other or other groups. First, we used HTSeq^[@CR23]^ to count reads that uniquely aligned to one gene, and these data were then imported into DESeq2^[@CR24]^ to generate PCA plots. At T1, the PCA results demonstrated that most samples clustered together, regardless of the grouping (Fig. [2](#Fig2){ref-type="fig"}). Except for one dog, rapid and slow groups became separated from controls at T2. Although the slow group tended to be closer to the controls than the rapid group, there was no clear distinction between rapid and slow groups at T2 or T3. Furthermore, PCA scree plots confirmed that principal components 1 (PC1) and 2 (PC2) accounted for 76--90% of the total variation in gene expression at each time point (Supplementary Fig. [S4](#MOESM1){ref-type="media"}). To further investigate the time-dependent nature of the DEGs, we performed hierarchical clustering of the top 100 DEGs (i.e., those with the smallest q-values identified in the time course analysis in DESeq2). In agreement with the PCA plots, this analysis demonstrated clustering of almost all sample groups at T1 (Fig. [3a](#Fig3){ref-type="fig"}). At time points T2 and T3, the rapid and slow groups clustered together (Cluster 1 in Fig. [3a](#Fig3){ref-type="fig"}) and were distinctly separated from the control group for all but one T2 sample (Cluster 2 in Fig. [3a](#Fig3){ref-type="fig"}).Figure 2Principal component analysis (PCA) for all samples at 3 time points. Principal component 1 (PC1) and principal component 2 (PC2) were identified by variance stabilizing transformation in DESeq2 at the 3 time points. The percentage of variance indicates how much variance was explained by PC1 and PC2. (Red: control group; Green: rapid group; Blue: slow groups). Figure 3Hierarchical clustering analysis, heatmap, and gene expression. (**a**) Hierarchical clustering analysis and heatmap of the 100 genes with the smallest q-values in the time course analysis in DESeq2 (Column names: t1, t2, and t3 designate 3 clinical time points; slow, rapid, and control represent grouping; \_1, \_2, and \_3 are individual dogs in each group). (**b**) Trends of gene expression over time for the 10 genes with the smallest q-values (from left to right, top to bottom) (Red: control group; Green: rapid group; Blue: slow group). Differentially expressed genes (DEGs) {#Sec6} ------------------------------------- In average, 20,090 genes were mapped by at least one read in each of the kidney biopsy samples (Supplementary Table [S3](#MOESM1){ref-type="media"}). Overall, 1,947 DEGs with a q-value \< 0.05 were detected over the 3 time points in the time course analysis of DESeq2 (Supplementary Table [S4](#MOESM2){ref-type="media"}). We applied the plot counts function in DESeq2 to visualize the top 10 genes with the smallest q-values (Fig. [3b](#Fig3){ref-type="fig"}). While these genes were not differentially expressed at T1, group-specific changes were observed over time, and expression in the slow group was consistently closer to that in the control group for each gene at T2 (Fig. [3b](#Fig3){ref-type="fig"}). To achieve the primary goal of this study, we identified 70 DEGs between the rapid and slow groups among all time points (q-value \< 0.05) (Table [1](#Tab1){ref-type="table"} and Fig. [4b](#Fig4){ref-type="fig"}). In this comparison, T2 demonstrated the most DEGs, with 68 of the 70 DEGs unique to T2. Two DEGs were identified at T1: stearoyl-CoA desaturase 5 (SCD5) (fold change = −3.55, q-value = 1.1 × 10^−05^), which was also detected at T2, and thymidine kinase 1 (TK1) (fold change = 2.49, q-value = 0.02). At T3, no DEGs were identified when these 2 groups were compared.Table 1Overview of 70 significant DEGs comparing rapid and slow groups (q-value \< 0.05).Up-regulated DEGs in the rapid groupGene SymbolFull NameFold Changeq-valueCOL1A1collagen type I alpha 1 chain6.522.62E-20COL3A1collagen type III alpha 1 chain5.351.05E-17COL1A2collagen type I alpha 2 chain4.014.66E-13COL5A1collagen type V alpha 1 chain3.87.21E-05COL6A3collagen type VI alpha 3 chain3.277.08E-07COL6A1collagen type VI alpha 1 chain3.261.82E-05LOXlysyl oxidase3.131.42E-03COL6A2collagen type VI alpha 2 chain3.072.34E-04PAMR1peptidase domain containing associated with muscle regeneration 12.995.76E-03CDCA8cell division cycle associated 82.982.80E-02COL11A1collagen type I alpha 1 chain2.865.33E-03COL15A1collagen type XI alpha 1 chain2.826.24E-03C1QTNF6C1q and tumor necrosis factor related protein 62.814.98E-03FNDC1fibronectin type III domain containing 12.813.22E-02FN1fibronectin 12.84.10E-02CCDC80coiled-coil domain containing 802.727.98E-04MFSD7major facilitator superfamily domain containing 72.73.22E-02FBLN1fibulin 12.694.10E-02NID2nidogen 22.611.52E-07COL4A2collagen type IV alpha 2 chain2.61.42E-03FAM69Bfamily with sequence similarity 69, member B2.63.42E-02NDNnecdin2.594.39E-02COL4A1collagen type I alpha 1 chain2.595.76E-03HTR75-hydroxytryptamine (serotonin) receptor 7, adenylate cyclase-coupled2.582.60E-02PCOLCEprocollagen C-endopeptidase enhancer2.585.83E-03OLFML2Bolfactomedin like 2B2.566.24E-03TK1^a^Thymidine kinase 12.491.98E-02C15orf39chromosome 15 open reading frame 392.472.76E-02RCN3reticulocalbin 32.466.94E-04MFAP2microfibrillar associated protein 22.451.42E-03HSPG2perlecan2.451.20E-02PRSS35protease, serine 352.454.10E-02MMP2matrix metallopeptidase 22.434.48E-05FBN1fibrillin 12.373.57E-03CD248CD248 molecule2.362.72E-02FOXRED2FAD dependent oxidoreductase domain containing 22.351.11E-02GXYLT2glucoside xylosyltransferase 22.343.62E-02FSCN1fascin actin-bundling protein 12.341.20E-02ENSCAFG00000008741^b^novel gene2.291.18E-02ENSCAFG00000012963^b^novel gene2.281.23E-03BGNbiglycan2.191.85E-02FASFas (TNF receptor superfamily member 6)2.162.08E-02ADAMTS2ADAM metallopeptidase with thrombospondin type 1 motif 22.153.23E-02PXDNperoxidasin2.144.10E-02SPARCsecreted protein acidic and cysteine rich2.134.48E-05THBS1thrombospondin 12.121.26E-04KCPkielin/chordin-like protein2.124.19E-02LRP1LDL receptor related protein 12.117.98E-04CERCAMcerebral endothelial cell adhesion molecule2.114.39E-02ITGA5integrin subunit alpha 52.077.03E-03BMP1bone Morphogenetic Protein 11.886.74E-05FSTL1follistatin Like 11.721.25E-05PTGFRNprostaglandin F2 Receptor Inhibitor1.662.61E-05**Down-regulated DEGs in the rapid group**UGT1A6UDP glucuronosyltransferase family 1 member A6−4.771.23E-03NAT8N-acetyltransferase 8 (putative)−4.11.09E-04R3HDMLR3H domain containing like−4.021.42E-03LIX1limb and CNS expressed 1−3.837.21E-03PRLRprolactin receptor−3.734.64E-04SCD5^c^stearoyl-CoA desaturase 5−3.551.07E-05FMO2flavin containing monooxygenase 2−3.421.78E-02OAT3/SLC22A8solute carrier family 22 member 8−3.372.45E-02SIsucrase-isomaltase−3.251.16E-02ENSCAFG00000003760^b^novel gene−2.923.96E-02SLC26A4solute carrier family 26 member 4−2.797.14E-03ENSCAFG00000000799^b^novel gene−2.773.25E-02HEPACAM2HEPACAM family member 2−2.668.81E-03IDO2indoleamine 2,3-dioxygenase 2−2.583.42E-02PECRperoxisomal trans-2-enoyl-CoA reductase−2.576.24E-03MT-ND3mitochondrially encoded NADH dehydrogenase 3−2.163.49E-03ABCA4retinal-specific ATP-binding cassette transporter−1.911.08E-04^a^DEG identified only at T1. ^b^Genes are displyed with ensembl IDs if gene annotations are unavailable. ^c^DEG identified at both T1 and T2. Figure 4DEGs in different pairs of comparisons at the 3 time points (T1, T2, and T3). Comparing the rapid and slow groups, 70 DEGs were found. Comparing each affected group with controls, several thousand DEGs were identified, with most of the DEGs occurring at T3. For each pair, only genes with a q-value \< 0.05 were considered as DEGs. The total number of DEGs found at each time point appears in parentheses. We also compared the rapid and slow groups with the control group, both individually and combined as a single "affected" group (Fig. [4a,c, and d](#Fig4){ref-type="fig"}). In these comparisons, the number of DEGs increased with advancing disease, with the largest number of DEGs identified at T3. This phenomenon indicates that the DEGs are disease-dependent as they are more differentially expressed in the later time points (T2 and T3) than T1 (Fig. [4](#Fig4){ref-type="fig"}). A substantial overlap of DEGs was present when comparing rapid and slow groups with the control group at T2 and T3 (Fig. [5](#Fig5){ref-type="fig"}). The overlapping DEGs between the rapid and slow groups were more numerous at T3 (2,952 DEGs) than at T2 (190 DEGs), supporting that the two groups behave similarly at the end-stage disease, as expected based on Figs [2](#Fig2){ref-type="fig"}--[4](#Fig4){ref-type="fig"}. Furthermore, the number of DEGs (1,189 DEGs) identified in both T2 and T3 in the rapid group was higher than the number of DEGs (171 DEGs) identified in both T2 and T3 in the slow group. This supports the theory that rapidly-progressing dogs express end-stage DEGs at a young age. The complete lists of DEGs from the time course analysis and all pairs of comparisons appear in Supplementary Table [S4](#MOESM2){ref-type="media"}.Figure 5Overlapping DEGs in rapid and slow groups compared with control at T2 and T3. For each comparison, only genes with a q-value \< 0.05 were considered as DEGs. The total number of DEGs found at each time point appears in parentheses. Gene Ontology (GO) and pathway analysis of DEGs {#Sec7} ----------------------------------------------- To characterize the GO terms, including molecular functions, biological processes, cellular components, and functional pathways of DEGs, we conducted over-representation tests for all pairs of comparisons in PANTHER version 11.1 (released on October 24, 2016) (Fig. [6](#Fig6){ref-type="fig"} and Supplementary Table [S5](#MOESM3){ref-type="media"}). We used the GO-Slim PANTHER annotation data set, which represents phylogenetically inferred annotations^[@CR25]^.Figure 6Enriched pathways and GO term analyses for 10 selected comparisons. Enriched pathways, biological processes, and regulator analyses for all DEGs are presented based on T2 and T3 comparisons as well as a time course analysis incorporating all time points for all groups. The number of genes was normalized to allow comparisons between groups within the same pathway, and comparisons were color coded in pairs, with the darker color corresponding to the later time point or the rapid group. Overlaps of GO terms among comparisons were commonly seen in the current study. We focused on "biological process" since it is the most characterized GO term. Within this category, the "immune system process" family was upregulated in all 10 comparisons presented in Fig. [6](#Fig6){ref-type="fig"}, and the "immune response" family was upregulated in all except for the rapid versus slow comparison at T2 and the time course analysis. Both immune-related GO terms appeared to be more upregulated at T2 than T3. At T2, the "biological adhesion" family, especially the cell-cell adhesion subfamily, was expressed in the rapid group more than in the slow group (Supplementary Table [S5](#MOESM3){ref-type="media"}). The most common pathway represented by the DEGs within the various comparisons was the "integrin signaling pathway" (Fig. [6](#Fig6){ref-type="fig"}). This was the only pathway identified in the comparison between rapid and slow groups. It was also the top upregulated pathway at T2 comparing rapid and control groups. Interestingly, analyzing the overlapping DEGs between T2 and T3, the integrin signaling pathway was identified as the top upregulated pathway within the rapid group, but not within the slow group, as compared with control. The "T cell activation pathway\" was another frequently detected pathway in our study (Fig. [6](#Fig6){ref-type="fig"}). It was upregulated in all comparisons with control at T2 and in overlapping DEGs between T2 and T3 within both the rapid and the slow groups. Another upregulated pathway included the "inflammation mediated by chemokine and cytokine signaling pathway" that was identified only when comparing all affected dogs with controls at T2. Because only limited numbers of DEGs were discovered, no enriched pathways were identified at T1 with any comparison. To further explore the possible biological interaction between orthologous genes in the human, mouse, and rat, we performed Ingenuity Pathway Analysis (IPA) to discover the most prevalent pathways and upstream regulators within each comparison. The "hepatic fibrosis/hepatic stellate cell activation pathway" was identified as the top pathway in multiple comparisons (Supplementary Table [S6](#MOESM4){ref-type="media"}). Transforming growth factor beta 1 (TGF-β1) was the most activated upstream regulator when the rapid and slow groups were compared at T2 and in the time course analysis (Supplementary Table [S6](#MOESM4){ref-type="media"}). Immunohistochemistry (IHC) validation of inflammatory pathways {#Sec8} -------------------------------------------------------------- Lastly, we aimed to validate the overexpression of inflammatory pathways in kidney biopsies. We chose to identify the predominant lymphocyte subtype present (T versus B cell) since the identified inflammatory pathways are all closely related to the presence of T lymphocytes^[@CR26]^ and antibodies for CD3 (T cell) and CD20 (B cell) are validated for use in dogs^[@CR27],[@CR28]^. As shown in Fig. [7](#Fig7){ref-type="fig"}, we confirmed the lymphocyte infiltration in the affected dogs to be composed mostly of T cells rather than B cells. This result correlated with the previously assigned chronic inflammation scores based on the histopathological evaluation (Supplementary Table [S2](#MOESM1){ref-type="media"} and Fig. [S2](#MOESM1){ref-type="media"}) and validated our RNA-seq data.Figure 7Expression of CD3 and CD20 using immunohistochemistry (IHC) in kidney biopsies from representative dogs. In affected dogs, lymphocyte infiltration present at later time points consisted mostly of CD3-positive lymphocytes with few CD20-positive lymphocytes identified. Scale bar: 20 μm. Discussion {#Sec9} ========== Dogs with XLHN have been studied as both an example of progressive canine glomerular disease and an animal model of human AS, which has CKD as a major syndrome component^[@CR2]^. The genetic cause is well characterized; however, the gene expression and molecular pathways influencing disease progression are incompletely known. In particular, the variable rate of disease progression in dogs with the same mutation and within families affected by AS is intriguing^[@CR7]^. We, therefore, aimed to evaluate differential gene expression, overrepresented pathways, and upstream regulators by comparing RNA-seq data in dogs that displayed a rapid clinical progression of the disease to those with relatively slow disease progression. In this study, we examined serial biopsies from rapid and slow groups as well as healthy age-matched littermates. To understand the biological changes during the pathogenesis of CKD, we included the earliest time point at which clinical disease could be detected in these dogs (onset of proteinuria, T1), with the onset of azotemia (T2) and the advent of end-stage disease (T3). By performing renal biopsies when animals reached specific clinical markers of disease progression, we could compare the same clinical stage in the rapid and slow groups. We believe this type of approach provides more confidence in identifying DEGs that are involved in the rate of disease progression than the traditional method of using age-driven time points, as differences detected are likely to be the driving force rather than the consequence of disease progression. Our data demonstrate the dynamic changes in gene expression at different stages of the disease. It also supports that the biological processes and pathways of fibrosis/adhesion and inflammation are the likely driving forces producing different rates of disease progression in the rapid and slow groups. While it is known that serum creatinine correlates strongly with tubulointerstitial fibrosis^[@CR29]^, one of the most intriguing findings when comparing the rapid and slow groups was the increase in fibrosis observed in the rapid group at the same clinical stage of disease (T2 and T3; Fig. [1](#Fig1){ref-type="fig"}). This corresponds with many of the 70 DEGs identified using RNA-seq between the rapid and slow groups that are implicated in fibrosis, almost all of which were identified at T2. Several of these genes have been previously described as upregulated in XLHN dogs, Alport mice, and other kidney diseases^[@CR9],[@CR11],[@CR30],[@CR31]^. Among these, one of the upregulated genes, *CD248* (endosialin or tumor endothelial marker 1, *TEM1*), has been found to mediate the adhesion and migration of cells through ligand interaction with the upregulated collagen-related genes: collagen type I, collagen type IV, and fibronectin-1 (*FN1*)^[@CR32]^. CD248+ stromal cells bind extracellular matrix and have been implicated in kidney^[@CR33]^, liver^[@CR34]^, and lung fibrosis^[@CR35]^. In non-inflamed kidneys, CD248 is expressed by mesangial cells located in glomeruli. In fibrotic kidneys, CD248 is additionally expressed by myofibroblasts and stromal fibroblasts, and the increased expression is closely related to prognostic indicators, such as albuminuria and renal scarring^[@CR33]^. Of note, one of the downregulated genes in the rapid group, prolactin receptor (*PRLR*), decreased in association with the extent of interstitial collagen I deposition in kidney transplant rejection^[@CR36]^, suggesting that PRLR might be a protectant against renal fibrosis. In our study, the decrease in *PRLR* could be responsible for the more rapid development of fibrosis and consequential faster progression of disease in the rapid group. Involvement of inflammatory components in the progression of CKD is another major finding of this study. Several inflammatory genes involved in fibrotic changes, such as biglycan (*BGN*), kielin/chordin-like protein (*KCP*), and matrix metallopeptidase-2 (*MMP2*) were upregulated in the rapid versus slow groups. BGN plays a role in bone growth, muscle development and regeneration, and collagen fibril assembly in multiple tissues. *BGN* is upregulated in renal fibrosis^[@CR37]^, and BGN protein expression strongly correlated with chronic kidney progression in one study^[@CR30]^, which may suggest its role in regulating inflammation and innate immunity. KCP expression is stimulated by renal stress, and it enhances the antifibrotic function of BMP7 to attenuate the profibrotic stimulus of TGF-β and to suppress proinflammatory cytokines^[@CR38]^. In Alport mice, the administration of recombinant BMP7 reduces glomerular and interstitial fibrosis but also upregulates MMP2^[@CR39]^. This upregulation of MMP2 seems contradictory, as it is associated with renal fibrosis in several animal models, including XLHN dogs^[@CR11]^. However, the function of MMP2 is specific to the temporal context of fibrosis. At the prefibrotic phase, increased MMP2 induces epithelial to mesenchymal transition, tubular atrophy, and fibrosis. For established fibrosis, inducing MMP2 synthesis by BMP7 promotes proteolytic removal of accumulated extracellular matrix, which is thought to be a potential therapeutic strategy^[@CR40]^. Since both KCP and MMP2 are upregulated at T2, when fibrosis is already relatively well established in the rapid group, their downstream actions are likely skewed toward anti-fibrotic effects. Mechanisms other than fibrosis and inflammation can also play roles in the rapid progression of CKD. NAT8, which is almost exclusively expressed by tubular cells in the renal cortex^[@CR41]^, is a cysteine S-conjugate N-acetyltransferase that is responsible for glutathione-mediated detoxification of nephrotoxic substances. The downregulation of *NAT8* in the rapid group suggests a more severe loss of normal renal function in the presence of similar serum creatinine concentrations compared with the slow group. Another downregulated gene, organic anion transporter 3 (*OAT3*, also known as *SLC22A8*), is decreased in kidney biopsies from human CKD patients and in a nephrectomized rat model of CKD^[@CR42],[@CR43]^. Reduced protein expression of Oat3 is associated with decreased excretion of an endogenous uremic toxin; meanwhile, the accumulation of this uremic toxin further inhibits Oat3-mediated transportation, accelerating toxin accumulation in serum^[@CR42],[@CR43]^. Therefore, downregulation of *OAT3* in the rapid group may result in impaired urinary excretion that is not adequately represented by serum creatinine concentration. Only 2 DEGs, SCD5 and TK1, were differentially expressed in the rapid group as compared with the slow group at T1. SCD5 was downregulated at both T1 and T2 in the rapid compared with the slow group. And, it was downregulated in the rapid versus control comparison throughout all 3 time points. SCD5 is an isoform of stearoyl-CoA desaturase that is responsible for the formation of monounsaturated fatty acids. Although SCD5 has not previously been described in the CKD literature, it has been proposed as a novel regulator of neural cell proliferation and differentiation, likely through β-catenin-independent (non-canonical) Wnt pathways^[@CR44]^. In fibrotic kidneys, the canonical Wnt pathway induces myofibroblast differentiation, and the non-canonical Wnt pathway leads to cytoskeleton rearrangement, cell adhesion, and cell movement^[@CR45]^. Thymidine kinase 1 (*TK1*) is the only DEG that was exclusively upregulated at T1. Thymidine kinase is responsible for producing dTMP that is later incorporated into DNA. The cytoplasmic isoform of TK1 is cell cycle-dependent, as it substantially increases in the S phase of the cell cycle. Given that unregulated proliferation is the hallmark of neoplasia, TK1 is a valuable serum marker for breast cancer, non-Hodgkin's lymphoma, plasmacytoma, and lung cancer^[@CR46]^. Its upregulation in our study could indicate increased cell proliferation at T1 in the rapid group. However, further investigation of SCD5 and TK1 is needed to determine their roles in CKD progression. In addition to abovementioned genes, many genes that are rarely described in CKD progression were found differentially expressed in the rapid versus slow groups at T2 (e.g., *LOX, PAMR1, CDCA8, C1QTNF6, FNDC1, CCDC80, MFSD7, FMO2, SI, SLC26A4*). The protein product of the upregulated gene lysyl oxidase (*LOX*) is an extracellular enzyme that is essential for covalent cross-linking of collagen in irreversible extracellular matrix deposition^[@CR47]^. Despite reports of its upregulation in liver fibrosis^[@CR48],[@CR49]^ and cardiomyopathy^[@CR50]^, the upregulation of LOX in CKD has only been described in one study using a glomerulonephritis mouse model^[@CR51]^. Simtuzumab, a monoclonal antibody that inhibits one of the LOX family members, lysyl oxidase homologue 2 (LOXL2), has recently been a focus of research as a possible new treatment for lung, liver, and kidney fibrosis due to a similar pathogenesis^[@CR52]^. Another gene minimally described in nephrology is the downregulated sucrase-isomaltase (*SI*). SI is an α-glucosidase that commonly appears on the brush border of small intestinal enterocytes and is involved in glucose digestion. SI is also present in small amounts in non-intestinal cells such as blood leukocytes and kidney cells^[@CR53]^. Little is known about the function and significance of SI in the renal tubule. However, the decreased expression of SI could indicate renal damage. To characterize DEGs in functional groups, we performed GO terms and pathway analyses by comparing affected dog groups, both individually and collectively, with control dogs at each time point. GO terms showed that the functions of identified DEGs were associated with "biological adhesion," "immune system processes," and "immune response," representing a common mechanism of disease progression in the early stages of CKD^[@CR54]^. Consistent with the GO terms, the "integrin signaling pathway" was the most upregulated pathway, and the "T cell activation pathway" and the "chemokine and cytokine signaling pathway" were also upregulated in multiple comparisons. The exclusive early expression of the "integrin signaling pathway" in the rapid group and universal expression at a later time point could indicate that it is an essential pathway driving rapid progression of disease in these dogs. The "integrin signaling pathway" consists mainly of collagen and integrin genes. The integrin subunit alpha 2 gene (*ITGA2*) was increased in the rapid versus control group at both T2 and T3. The COL4A3^−/−^/ITGA2^−/−^ double knockout Alport mouse model has delayed renal fibrosis compared with COL4A3^−/−^/ITGA2^+/+^ Alport mice, which express significantly higher levels of MMP2, MMP9, MMP12, and TIMP1^[@CR55]^. Upregulation of *MMP2* is consistent with our Alport dog model, suggesting that the "integrin signaling pathway" is involved in matrix accumulation. To verify the pathway analysis results, we used IHC to identify the infiltrating lymphocyte population. The "T cell activation," "integrin signaling," and the "inflammation mediated by chemokine and cytokine signaling" pathways are all closely related to the presence of T lymphocytes^[@CR26]^. Consistent with this, T lymphocytes were identified as the predominant inflammatory cell population present in the affected dogs, as well as in a study of canine end-stage renal disease^[@CR56]^. Moreover, T cell infiltration inversely correlates with renal function at the time of renal biopsies in AS patients^[@CR57]^. Overall, IHC validated the results of GO terms and pathway analyses, which showed that inflammatory pathways and corresponding biological processes are altered during the progression of CKD. Ingenuity Pathway Analysis (IPA) allows for characterizing orthologous genes, and results identify possible mechanisms that have been validated in humans, mice, and rats. The top upregulated pathway identified in multiple comparisons was the "hepatic fibrosis/hepatic stellate cell activation pathway." The IPA identified enriched pathways based on the over-represented DEGs, and the principle is the same as that used in the GO terms analysis via Panther. There was an extensive overlap between the DEGs we identified and the genes involved in the "hepatic fibrosis/hepatic stellate cell activation pathway" in the IPA, including collagen genes, cytokine-related genes, matrix metalloproteinases, tissue inhibitor of metalloproteinase, and the TNF receptor superfamily. Therefore, the upregulation of this pathway in our study supports common mechanisms involved in hepatic and renal fibrosis. It could also indicate contributing genes beyond those identified using known canine gene pathways. The upstream regulator analysis of IPA identified the TGF-β group, especially TGF-β1, as the top upstream regulator, with both the highest activated z-score and the lowest p-value across multiple comparisons. Previous studies in canine^[@CR8]^ and murine^[@CR58]^ models of AS demonstrated the expression of TGF-β mRNA in kidney tissue. However, TGF-β was not differentially expressed in our comparisons, and IHC staining for TGF-β did not demonstrate appreciable differences in XLHN dogs compared to controls in a previous study^[@CR8]^. The IPA Upstream Regulator Analysis predicts the upstream regulator of gene expression changes based on the knowledge of this regulatory cascade in the literature compiled in the Ingenuity Knowledge Base. Thus, the identified upstream regulator may not be identified as a DEG despite its importance in gene regulation. A limiting factor of this study is that the kidney biopsies were immediately placed into RNALater to preserve RNA integrity, so microscopic evaluation could not be performed to determine whether the biopsy used for RNA isolation was representative of the cortex as a whole. Thus, one of the samples in the control group at T1 could have represented an area affected by a clinically insignificant insult. Another limiting factor is that expression data represents the mean expression by many cell types. Because the kidney has many cell types, all with different roles in CKD progression, it would be ideal to study gene expression changes in individual cells using laser-capture microdissection to further elucidate the progression of CKD. Last, extensive validation of the RNA-seq results was not performed; however, previous studies have shown RNA-seq to be a robust tool that highly correlates with qPCR results^[@CR14]--[@CR19]^. RNA-seq may even be more reliable than qPCR due to its higher sensitivity and lower probe bias^[@CR16]^. We did perform IHC to further characterize the inflammatory population, which supported the pathways identified through the RNA-seq results. In what appears to be the first RNA-seq study of a canine CKD model, we identified several previously described and novel genes and enriched pathways involved in the pathogenesis and development of CKD. The approach of acquiring biopsies at time points determined by the clinical stage of disease was an attempt to target causative gene expression starting at the clinical onset of proteinuria rather than secondary changes. Regardless of initial insult, CKD has common pathways that lead to end-stage kidney disease^[@CR20]^. Therefore, many genes found in the current study may serve as predictive or diagnostic biomarkers for early detection of CKD in dogs and people. They may also be potential targets for drug development for this condition. Methods {#Sec10} ======= Animals {#Sec11} ------- The dogs in this study were part of a colony with XLHN maintained at Texas A&M University^[@CR5]^. XLHN is caused by a 10-base deletion in the gene encoding the α5 chain of type IV collagen. Affected males develop juvenile-onset CKD that progresses to end-stage renal disease as previously described^[@CR5]^. Overall, 6 affected dogs and 2 unaffected littermates were studied. All dogs were raised according to standardized protocols, and no treatments were given to these dogs. All protocols were approved by the Texas A&M University Institutional Animal Care and Use Committee. Clinical phenotypes {#Sec12} ------------------- For this study, dogs were selected to represent both extremes in the speed of disease progression in this family of dogs (rapid versus slow progression). Clinical progression was determined by serial monitoring of serum and urine biomarkers of kidney disease, which allowed us to establish specific progression time points^[@CR29]^: T1 (onset of proteinuria: defined as the presence of microalbuminuria for 2 consecutive weeks (E.R.D. HealthScreen Canine Urine Test Strips, Loveland, CO, USA)); T2 (onset of azotemia: serum creatinine ≥ 1.2 mg/dL); and T3 (end-stage disease: serum creatinine ≥ 5 mg/dL). Rapidly-progressing (rapid) dogs (n = 3) reached each time point at an earlier age than slowly-progressing (slow) dogs (n = 3). On average, the rapid group reached T3 at 26.3 weeks of age (range: 26--27 weeks), while the slow group reached the last clinical time point (T3) at 49 weeks of age (range: 46--52 weeks) (Supplementary Table [S7](#MOESM5){ref-type="media"}). Tissue collection {#Sec13} ----------------- Kidney cortex was serially collected from each dog at the aforementioned 3 clinical time points (independent of age). Control dogs (n = 2) were biopsied to correspond with an affected littermate. All samples were collected by ultrasound-guided needle biopsy. This technique was appropriate for the current study as it is unlikely to induce changes that might be confused with those of CKD progression^[@CR59]^. Samples for pathology evaluation and immunohistochemistry were placed in formalin and embedded in paraffin. Samples for RNA sequencing were immediately placed in RNAlater Stabilization Solution (Life Technologies, Foster City, CA, USA) and stored at −80 °C until RNA isolation. Histopathological evaluation {#Sec14} ---------------------------- Paraffin-embedded samples were processed and stained as previously described^[@CR29]^. To determine the severity of interstitial fibrosis and chronic inflammation, a board-certified veterinary anatomic pathologist (REC) evaluated 5 or 20 randomly chosen 20x fields of renal cortex based on core size for each biopsy. For interstitial fibrosis, a score of 0 to 3 was assigned for each field based on the degree of tubulointerstitial architecture distortion caused by fibrosis: 0 - no fibrosis, 1 - fibrosis present but no distortion, 2 -- moderate distortion, and 3 -- severe distortion. For chronic inflammation, 0 -- no inflammatory cells, 1 -- scattered inflammatory cells, 2 -- aggregates of inflammatory cells that separate or replace tubules, and 3 -- diffusely distributed inflammatory cells. Statistical analysis comparing the average scores between groups was performed using bootstrap in R (version 3.2.4) to construct simultaneous 95% confidence intervals for all 3 pairwise comparisons of mean fibrosis and chronic inflammation scores at each of the latter 2 disease stages. RNA isolation and sequencing {#Sec15} ---------------------------- The MirVana miRNA Isolation Kit (Ambion, Austin, TX, USA) was used to isolate total RNA from homogenized kidney tissue according to the manufacturer's instructions. The library preparation, sequencing, and initial quality check were performed by the Texas A&M AgriLife Genomics and Bioinformatics Service (<http://www.txgen.tamu.edu/>). RNA integrity was assessed by the Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA). The average RIN was 3.4, and the average RNA yield was 86.2 ng/μL (Supplementary Table [S3](#MOESM1){ref-type="media"}). To compensate for the low-input samples, we use the TruSeq Stranded Total RNA Library Prep Kit with Ribo-zero Gold (Illumina, San Diego, CA, USA), based on the best practice for RNA with variable qualities^[@CR21]^ (to remove both cytoplasmic and mitochondrial rRNA) and its compatibility with canine samples. Samples were then sequenced using the Illumina Genome Analyzer (HiSeq. 2500v4 High Output). Raw sequencing data were submitted to the NCBI sequence read archive (SRA) (Accession: SRP101707; Samples: SAMN06560417, SAMN06560429-51; BioProject: PRJNA378728). Data analysis {#Sec16} ------------- FastQC (version 0.11.2) was used for quality control to ensure that the quality value was above Q30. The canine (*Canis lupus familiaris*) genome FASTA file (ftp://ftp.ensembl.org/pub/current_fasta/canis_familiaris/dna/Canis_familiaris.CanFam3.1.dna.toplevel.fa.gz) and gene annotation GTF file (CanFam 3.1 assembly; ftp://ftp.ensembl.org/pub/current_gtf/canis_familiaris/) were obtained from Ensembl. Although RNA-seq is a popular research tool, there is no gold standard for analyzing RNA-seq data. Among the available tools, we chose up-to-date open source tools for mapping, retrieving read counts, and differential analysis. We used HISAT2^[@CR22]^ (version 2.0.3-beta) to generate indexes and to map reads to the canine genome. For assembly, we chose SAMtools (version 1.2) and the "union" mode of HTSeq^[@CR23]^ (version 0.6.1), as the gene-level read counts could provide more flexibility in the differential expression analysis. Both HISAT2 and HTSeq analyses were conducted using the high performance research computing resources provided by Texas A&M University (<http://hprc.tamu.edu>) in the Linux operating system (version 2.6.32). Differential expression and statistical analysis were performed using DESeq2 (release 3.3) in R (version 3.2.4). DESeq2^[@CR24]^ was chosen as it is a popular parametric tool that provides a descriptive and continually updated user manual ([https://bioconductor.org/packages/release/bioc/vignettes/DESeq2/inst/doc/DESeq2.html](https://bioconductor.org/packages/release/bioc/vignettes/DESeq.%202/inst/doc/DESeq.%202.html)). DESeq2 internally corrects for library size, so it is important to provide un-normalized raw read counts as input. We used variance stabilizing transformation to account for differences in sequencing depth. P-values were adjusted for multiple testing using the Benjamini-Hochberg procedure^[@CR60]^. A false discovery rate adjusted p-value (i.e., q-value) \< 0.05 was set for the selection of DE genes. Gene ontology (GO), pathway, and upstream regulator analysis {#Sec17} ------------------------------------------------------------ Gene ontology and PANTHER pathway analyses were performed with the PANTHER Overrepresentation Test (released on July 15, 2016) in PANTHER^[@CR25]^ version 11.1 (<http://www.pantherdb.org/>, released on October 24, 2016). This program supports the canine genome. Also, QIAGEN's Ingenuity Pathway Analysis (IPA, QIAGEN Redwood City, [www.qiagen.com/ingenuity](http://www.qiagen.com/ingenuity)) was used to provide overrepresented orthologous genes in human, mouse, and rat databases and to identify orthologous pathways and upstream regulators in our data. PANTHER used the binomial test and Bonferroni correction for multiple testing, while IPA used the right-tailed Fisher Exact test and displayed z-scores to indicate whether a potential regulator was activated or inhibited. We used the default settings for statistical analysis in both the PANTHER pathway and IPA. In PANTHER, only pathways and GO terms with fold enrichment \> 0.2 were listed. In IPA, p-value \< 0.05 and fold change \> 2 were set as cutoff values. Immunohistochemistry (IHC) {#Sec18} -------------------------- Three-micrometer, formalin-fixed, paraffin-embedded renal cortex sections of both affected and control dogs were stained for CD3 (n = 24) and CD20 (n = 9). After deparaffinization, the sections were placed in citrate buffer (pH 6.0) for antigen retrieval, using a pressure cooker (Decloaking Chamber, Biocare Medical). Endogenous peroxidase activity and non-specific protein binding were blocked with 3% hydrogen peroxide and Sniper protein block (Biocare Medical), respectively. After blocking, the sections were incubated with primary antibodies CD3 (1:300 dilution; Dakocytomation, Carpinteria, CA) and CD20 (1:500 dilution; Thermo Scientific, Fremont, CA) for 1 hour at room temperature, then incubated with MACH 2 polymer for 30 minutes at room temperature. DAB (Vector Laboratories, Burlingame, CA) was used as the chromogen to demonstrate sites of antibody-antigen reaction. Mayer's hematoxylin was used for the counterstain. Photographs were obtained using a SPOT Insight 2Mp FW Color Mosaic Camera (Diagnostic Instruments, Inc., Sterling Heights, MI) and the SPOT software (version 5.2). Data Availability {#Sec19} ----------------- Raw sequencing reads generated from this study are deposited at the NCBI sequence read archive (SRA) under accession SRP101707 (Samples: SAMN06560417, SAMN06560429-51; BioProject: PRJNA378728). Electronic supplementary material ================================= {#Sec20} Supplementary File S1 Supplementary Table S4 Supplementary Table S5 Supplementary Table S6 Supplementary Table S7 **Electronic supplementary material** **Supplementary information** accompanies this paper at 10.1038/s41598-017-16603-y. **Publisher\'s note:** Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This work was funded by the National Institutes of Health (DK57676 and DK64273) and the Texas A&M Genomics Seed Grant Program. This study was presented in part at the 2016 American College of Veterinary Pathologists/American Society for Veterinary Clinical Pathology Annual Meeting in New Orleans, LA, December 2016. We would like to thank Dr. Ivan Ivanov for offering valuable suggestions for this study and providing RNA-seq analysis training through the Center for Translational Environmental Health Research (<http://www.ctehr.org>), Dr. Michael Dicken and all the personnel in the TAMU High Performance Research Computing (<http://hprc.tamu.edu>) for providing useful resources and instant help during data analysis, Dr. Michael Love for his rapid response in the Bioconductor online forum (<https://support.bioconductor.org/>), Dr. Andy Ambrus for the immunohistochemistry staining, Mia Aguilar and Mary Sanders for their technical assistance with RNA isolation and sample collection from dogs, respectively, and Cheng-An Richard Chang for guiding us through the first step of RNA-seq analysis. Special thanks to Dr. Barbara Gastel and Leah Poffenberger for their editing assistance on this manuscript. C.P.C. performed data analysis and prepared the manuscript; J.A.H. performed RNA isolations, revised the manuscript, and co-designed this project; R.E.C. evaluated the histopathology results and revised the manuscript; A.R.D. refined the statistical analyses; C.B.L. co-designed this project; G.E.L. collected samples and contributed to the study idea; and M.B.N. collected samples, designed and supervised this project, and revised the manuscript. All authors reviewed the manuscript. Competing Interests {#FPar1} =================== The authors declare that they have no competing interests.

The content published in Cureus is the result of clinical experience and/or research by independent individuals or organizations. Cureus is not responsible for the scientific accuracy or reliability of data or conclusions published herein. All content published within Cureus is intended only for educational, research and reference purposes. Additionally, articles published within Cureus should not be deemed a suitable substitute for the advice of a qualified health care professional. Do not disregard or avoid professional medical advice due to content published within Cureus. Introduction ============ Among the many responsibilities of a physician, the first and foremost is to his patients and their health and safety. A very important part of good medical practice is to maintain patient records. The General Medical Council recommends ensuring accurate, comprehensive and legible records are maintained for every patient by the surgeon \[[@REF1]\]. Accurate and detailed operation notes are of great importance in all surgical specialties not only for patient care but also for providing information for research, audit and medico-legal purposes \[[@REF2]\]. Operative notes are often presented in legal malpractice cases, and studies have demonstrated that up to 45% of operative notes are indefensible from a medico-legal standpoint. Incomplete and illegible notes are often a source of weakness in the defence of surgeons in courts \[[@REF3]\]. Clear and legible notes relating to all surgical procedures are therefore extremely important \[[@REF4]\]. Handwritten notes are still used worldwide; however, establishing their legibility could often be a major drawback \[[@REF5]\]. An audit conducted in Kuwait by Sweed et al. demonstrated that 20% of the orthopaedics operative notes they reviewed contained illegible parts, was incomplete and included confusing abbreviations \[[@REF6]\]. In order to improve our clinical practice, there is a need to adopt a standardised way to document operative notes, so that our records contain all the details necessary to give patients the best possible care. Although there are no known standardised guidelines in place in Sudan relating to operative notes, there are international guidelines that are in use and are well recognised, such as those set by the Royal College of Surgeons of England \[[@REF7]\]. Therefore, in this study, we assess the quality of operative notes against the standards set by Royal College of Surgeons of England, with a view to improve the quality of operative notes and ensure quality patient care. We also aim to test the applicability of these guidelines in the setting of a developing country with a high workload and limited medical technology. Materials and methods ===================== Our institute is a tertiary care hospital based in Khartoum, the capital of Sudan. A sample of 59 operation notes collected between March 1, 2019, to April 1, 2019, were randomly selected from the orthopaedic department at the Ribat University Hospital in Khartoum state, Sudan; they were retrospectively audited by three reviewers according to the Royal College of Surgeons of England Good Surgical Practice guidelines (Table [1](#TAB1){ref-type="table"}). The Operative notes were all based on the standard template used at the Ribat University Hospital. ###### Operation notes standard criteria as set by the Royal College of Surgeons of England ------------ Serial no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 ------------ The notes were audited in accordance with the Royal College of Surgeons of England Good Surgical Practice guidelines \[[@REF7]\], with additional items related to orthopaedic surgery making up 19 items in total: the diagnosis, elective/emergency, date and time of surgery, operating surgeon, assistants' names, anaesthetist\'s name, operative procedure, incision, operative diagnosis, operative findings, problems/complications, details of prosthesis/material used (including serial number), details of closure technique, anticipated blood loss, drain, tourniquet time, detailed post-op instructions, and the signature. Also, we audited who wrote the operation note, and whether it was written by the operating surgeon or not. After the descriptive analysis of the audit data using Microsoft Excel version 2010 (Microsoft Corporation, Redmond, WA), a memory aid was then placed in the operation theatre during May-June 2019, emphasising mainly the points with poor compliance (Figure [1](#FIG1){ref-type="fig"}). A re-audit was then performed for another 59 operation notes, with regard to the same items. Results ======= During the first half of the audit cycle, 59 operation notes of surgeries performed between March 1, 2019, to April 1, 2019, were retrospectively reviewed against the guidelines for the operation notes set by the Royal College of Surgeons of England in the recent 2014 edition; there was good compliance with respect to date documentation (86%), diagnosis (85%), name of operating surgeon (90%), assistants' names (86%), description of the operative procedure (98%), detailed post-operative instructions (98%) and the signature (75%). The audit highlighted five components with poor compliance: the time or length of the operation, the name of the anaesthetist, the incision type or shape, usage of drains and details of the prosthesis used such as the serial number or size. A memory aid was then designed and placed in the operation room for two months, emphasising on the poorly documented aspects (Figure [1](#FIG1){ref-type="fig"}). {#FIG1} In the re-audit phase, another 59 operative notes were reviewed of which four were emergency operations. An improvement was noted in documenting the information that had been poorly documented in the first audit. Time documentation improved from 37% to 56% of the audited notes, documenting the name of the anaesthetist improved slightly to 46% from 41%, and documenting the type of incision rose to 59% from 41%. In the operations where a prosthesis was used, the size and details of the prosthesis reported improved by more than half, from 15% to 34%. Likewise, reporting on the insertion of the drain in its right place on the operation sheet improved to 15% from 7% (Figure [2](#FIG2){ref-type="fig"}). {#FIG2} In the first audit, 20% of the operation notes were recorded by the operating surgeon, while only 14% were documented by the operating surgeon in the re-audit phase. Discussion ========== Meticulous record-keeping with regard to patient notes is becoming more important in medical practice, especially in a setting where complaints relating to medical negligence and malpractice cases are on the rise; but first and foremost, operative notes are an essential pillar in patient care as well as education and research. This audit examined one of the most important documents in medical practice in general and surgical care in particular. In our hospital, there is a single, standard operation sheet that is shared among all surgical specialties. In the audit we conducted, it seemed feasible to use a memory aid as a method to improve the quality of documentation in the operation notes. Din et al. used a similar approach in their audit of orthopaedic operation notes, which included elective operations as well as emergency surgeries; their use of memory aid improved the overall documentation of details from 90% to 97.1% of the notes \[[@REF8]\]. These results and the results from our audit prove that implementing a simple measure such as placing a memory aid in the operation theatre can affect patient care significantly in a positive way. Shayah et al. also used memory aids in the ear, nose and throat (ENT) theatres, and in their re-audit cycle, all the components of operation notes improved by 100% \[[@REF9]\]. On the other hand, Fionn et al. suggested the use of typed, orthopaedic-specific operation notes and tested them in their audit in comparison to handwritten notes \[[@REF10]\]. Their audit showed that typed notes had better legibility compared to the handwritten ones, and resulted in better documentation as they had added checks related to orthopaedic surgeries like tourniquet time and prosthesis details. Similarly, Andrew et al. used electronic, custom-made operation notes \[[@REF11]\]. However, they were procedure-specific, specifically related to hip arthroplasty. They demonstrated that the electronic template raised the general accuracy of operation notes from 58% to 92%, improved documentation as per all the specific Royal College of Surgeons parameters and reduced variations seen with handwritten notes. Nevertheless, the applicability of computerised or procedure-specific operation notes is very limited in many settings, even in developed countries, since many hospitals do not have a fully computerised system with regard to patient medical records. Ghani et al. used a smarter electronic operation note for orthopaedic surgeries in general, and this modern intervention demonstrated 100% accuracy and readability of the notes. Many other audits and studies have used similar methods like pre-typed form or checklists. All methods aim to improve the quality and documentation of operation notes in different surgical specialties, bearing in mind the importance of these records in patient care. In our view, the best method to improve the quality of operation notes is the use of electronic forms that take into account the differences between various surgical specialties; however, the applicability of this method in Sudan is difficult at present; nevertheless, it should be considered as a plan to be implemented in the near future as medical practice in general is becoming more digitalised. For the time being, we highly recommend modifying the standard operation notes to be tailored according to the surgical specialty in question; for example, adding headings related to tourniquet time and prostheses details in orthopaedic patient files will significantly improve the adherence by surgeons to these parameters. Conclusions =========== The proforma used in our hospital aligns well with the criteria described in the Royal College of Surgeons Good Surgical Practice guidelines released in 2014. Our implementation of a memory aid in the operation theatre helped to improve the reporting of some of the criteria; however, some components of the operation notes remained poorly filled in. We strongly recommend continuing with the implementation of audits in our clinical practice among trainees and departments in order to closely analyse our practices, especially with regard to patient safety and quality of care. The authors have declared that no competing interests exist. Consent was obtained by all participants in this study **Animal subjects:** All authors have confirmed that this study did not involve animal subjects or tissue.

\section{Introduction} \label{sec:intro} The accurate measurement of the mixing angle $\theta_{13}$ by reactor neutrino experiments~\cite{An:2015rpe}, with a small uncertainty comparable to that for $\theta_{12}$, has initiated a precision era for neutrino physics. In the standard three-family framework, the main remaining issues are the possible observation of leptonic CP violation, the determination of the ordering of neutrino masses and probing the Dirac or Majorana nature of neutrinos. Some hints currently exist in the latest data collected by NOvA and T2K which seem to point to maximal CP violation in the neutrino sector, but the statistical significance is still low~\cite{T2K-2017,NOvA-Seminar-FNAL-2017}. Likewise, a global fit to neutrino oscillation data seems to show a mild preference for a normal mass ordering (see for instance~\cite{nufit,Esteban:2016qun}), which needs to be confirmed as more data become available. At the same time, and in view of the precision of present and near future neutrino facilities, it is of key importance to verify if neutrinos have unexpected properties caused by New Physics (NP) beyond the standard three-family framework. In this work we study one of the possible windows to NP, the so-called quantum decoherence in neutrino oscillations, and update the existing bounds by analyzing IceCube and DeepCore data on atmospheric neutrinos. In particular, we are interested in a kind of decoherence effects in neutrino oscillations studied, for example, in \cite{Benatti:2000ph,Lisi:2000zt,Gago:2000qc,Gago:2002na,Morgan:2004vv,Hooper:2004xr,Fogli:2007tx,Farzan:2008zv} and, more recently, in \cite{Bakhti:2015dca,Gomes:2016ixi,Oliveira:2016asf,Coelho:2017zes,Coelho:2017byq,Carpio:2017nui}. These decoherence effects differ from the standard decoherence caused by the separation of wave packets (see e.g.~\cite{Giunti:1997wq, An:2016pvi}) and might arise, instead, from quantum gravity effects~\cite{Hawking:1976ra,Ellis:1983jz,Giddings:1988cx}. Throughout this work, for brevity, we will refer to such non-standard decoherence simply as decoherence. The authors of ref.~\cite{Lisi:2000zt} derived some of the strongest available constraints on neutrino decoherence in neutrino oscillations up to date, using atmospheric neutrino data from the Super-Kamiokande (SK) experiment~\cite{Fukuda:1998mi,Fukuda:1998tw,Fukuda:1998ub,Fukuda:1998ah}. Moreover, they considered the general case in which the decoherence parameters could depend on the neutrino energy via a power law, $\gamma=\gamma_{0}(E/\text{GeV})^n$, where $n = 0, -1, 2$. Nevertheless, these limits were obtained within a simplified two-family framework and without taking into account the matter effects in the neutrino propagation. Furthermore, only a reduced subset of SK data (taken, in fact, almost 20 years ago now) was analyzed~\cite{Fukuda:1998mi,Fukuda:1998tw,Fukuda:1998ub,Fukuda:1998ah}. In this work, we show that performing a three-flavour analysis which includes the matter effects is essential in order to correctly interpret such constraints. In particular, it is not obvious to which $\gamma_{ij}$ parameter the SK bounds derived in two families~\cite{Lisi:2000zt} would actually apply. We will show that it strongly depends on the neutrino mass ordering and on whether the sensitivity is dominated by the neutrino or antineutrino channels: for neutrinos the decoherence effects at high energies are mainly driven by $\gamma_{21}$ ($\gamma_{31}$) for normal (inverted) ordering, while in the antineutrino channel they are essentially controlled by $\gamma_{32}$ ($\gamma_{21 }$) for normal (inverted) ordering. Concerning the solar sector, the authors of ref.~\cite{Gomes:2016ixi} obtained strong constraints on $\gamma_{21}$ from an analysis of KamLAND data~\cite{Eguchi:2002dm, Abe:2008aa}, for $n=0,\pm1$.\footnote{It should be mentioned that, in~\cite{Fogli:2007tx}, very strong bounds on dissipative effects were derived from solar neutrino data, for $n=0,\pm1,\pm2$ and in a two-family approximation. However, such limits do not apply to the case in which only decoherence effects are included, as pointed out in~\cite{Oliveira:2014jsa,Gomes:2016ixi}. This will be further clarified in sec.~\ref{sec:adiabatic}.} Finally, the authors of ref.~\cite{deOliveira:2013dia} derived several bounds on the atmospheric decoherence parameters $ \gamma_{32}$ and $\gamma_{31}$ from an analysis of MINOS data~\cite{Adamson:2011fa, Adamson:2011ig, Adamson:2012rm}. Non-standard decoherence has been invoked several times in the literature in order to decrease the tension in the parameter space among different sets of neutrino oscillation data. For example, in refs.~\cite{Farzan:2008zv,Bakhti:2015dca} a solution to the LSND anomaly based on quantum decoherence, compatible with global neutrino oscillation data, was proposed. More recently, in~\cite{Coelho:2017zes} it was shown that the $\sim 2\sigma$ tension between T2K and NOvA on the measurement of the atmospheric mixing angle $\theta_{23}$ could be alleviated through the inclusion of decoherence effects in the atmospheric neutrino sector, namely, $\gamma_{23} = (2.3 \pm 1.1)\cdot 10^{-23}$ GeV. Such value of $\gamma_{23}$ would be close to the SK bound from ref.~\cite{Lisi:2000zt}, $\gamma < 3.5 \cdot 10^{-23}$ GeV ($90\%$ CL), but still allowed. This topic has recently brought the attention of a part of the community. In fact, several analyses of decoherence effects on present and future long-baseline neutrino oscillation experiments have been recently performed (albeit at the probability level only), see e.g. refs.~\cite{Oliveira:2016asf,Coelho:2017byq,Carpio:2017nui}. In this work we will show that the reference value for $\gamma_{23}$ considered in~\cite{Coelho:2017zes} is indeed already excluded by IceCube data (we note however that, according to the latest results reported by NOvA, the significance of the tension has been reduced to less than 1$\sigma$~\cite{NOvA-Seminar-FNAL-2017}). Moreover, we find that IceCube and DeepCore data are able to improve significantly over most of the constraints in past literature, both for solar and atmospheric decoherence parameters, in some cases by several orders of magnitude. The paper is structured as follows. In sec.~\ref{sec:formalism} we present the formalism and discuss the effects of decoherence on the oscillation probabilities. We first review the case of constant matter density profile, and then proceed to discuss the case of non-uniform matter. In particular we show that, within the adiabatic approximation, no significant bounds on the decoherence parameters can be extracted from solar neutrino data when the neutrino energy is assumed to be conserved. We then proceed to develop a formalism which permits a consistent treatment of the decoherence effects on neutrino propagation in non-uniform matter when the adiabaticity condition is not fulfilled, as is the case of atmospheric neutrino experiments. In sec.~\ref{sec:plots-probabilities} we apply this formalism to the computation of the relevant oscillation probabilities in the atmospheric neutrino case, discussing the main features arising in presence of decoherence. Section~\ref{sec:icecube} summarizes the main features of the IceCube and DeepCore experiments, the data sets considered in our analysis, and the details of our numerical simulations. Our results are then presented and discussed in sec.~\ref{sec:results}. Finally, we summarize and draw our conclusions in sec.~\ref{sec:conclusions}. \ref{app} and~\ref{sec:d5results} discuss technical details regarding some of the approximations used in our numerical calculations. \section{Quantum decoherence: Density matrix formalism} \label{sec:formalism} The evolution of the density matrix $\rho$ in the neutrino system can be described as \begin{equation} \frac{d\rho}{dt}=-i\left[H,\rho \right]-\mathcal{D}\left[\rho\right], \label{eq:1} \end{equation} where $H$ is the Hamiltonian of the neutrino system and the second term $\mathcal{D}\left[\rho\right]$ parameterizes the decoherence effects. In vacuum, the diagonal elements of the Hamiltonian are given by $h_i=m_i^2/(2E)$, where $m_i\ (i=1,2,3)$ are the masses of the three neutrinos and $E$ is the neutrino energy. Here $\rho$ is defined in the flavour basis, with matrix elements $\rho_{\alpha\beta}$. Throughout this work, we will use Greek indices for flavor ($\alpha,\beta=e,\mu,\tau$), and Latin indices for mass eigenstates ($i,j = 1,2,3$). A notable simplification of eq.~\eqref{eq:1} can be achieved via the following set of assumptions. First, assuming complete positivity, the decoherence term $\mathcal{D}\left[\rho\right]$ can be written in the so-called Lindblad form~\cite{Lindblad:1975ef,Banks:1983by} \begin{equation} \mathcal{D}\left[\rho\right]=\sum_m\left[\left\{\rho,D_m D_m^\dagger\right\}-2D_m\rho D_m^\dagger\right], \label{eq:Dm} \end{equation} where $D_m$ is a general complex matrix. Second, avoiding unitarity violation, which is equivalent to imposing the condition $d\,\text{Tr}[\rho]/dt=0$, requires $D_m$ to be Hermitian. Moreover, $D_m=D_m^\dagger$ implies that the entropy $S=\text{Tr}[\rho\,\ln\rho]$ increases with time. Finally, a key assumption is the average energy conservation of the neutrino system, which is satisfied when \begin{equation} [H,D_m]=0. \label{eq:[H,D]} \end{equation} In presence of matter effects, the Hamiltonian is diagonalized by the unitary mixing matrix\footnote{Note that we consider the standard definition for the relation between the mass and flavour eigenstates used in neutrino oscillations: for field operators, $\nu_\alpha = \sum_i U_{\alpha i}\nu_i$; for one-particle states, $|\nu_\alpha\rangle = \sum_i U^{*}_{\alpha i}|\nu_i\rangle$.} $\tilde{U}$ (throughout this paper, in our notation the presence of a tilde denotes that a quantity is affected by matter effects). Therefore, after imposing the energy conservation condition given by eq.~(\ref{eq:[H,D]}), we get \begin{equation} \begin{array}{rcl} H &=&\tilde{U} \, \text{diag} \left\{\tilde{h}_1,\tilde{h}_2,\tilde{h}_3\right\} \tilde{U}^\dagger\equiv \tilde{U} H_d \tilde{U}^\dagger, \label{eq:diag} \\ D_m &=&\tilde{U} \,\text{diag}\left\{d_m^1,d_m^2,d_m^3\right\} \tilde{U}^\dagger\equiv \tilde{U} D_m^d \tilde{U}^\dagger. \end{array} \end{equation} This implies that the average energy is conserved along the whole neutrino propagation (through vacuum and matter). This assumption is indeed crucial for our analysis. It is expected to be fulfilled in vacuum and in very good approximation in matter. While we assume that the quantum decoherence itself does not cause the violation of energy conservation, due to the standard neutrino interaction with matter, for large neutrino energies the energy is not exactly conserved in presence of matter due to a small energy transfer to the background fermions. Therefore, in this case, eq.~(\ref{eq:diag}) does not hold exactly. This issue has been analyzed in detail in \cite{Oliveira:2014jsa,Oliveira:2016asf,Carpio:2017nui}, where it was shown that in a more general framework in which energy conservation is not assumed, two types of effects in the neutrino oscillation probabilities can essentially be distinguished: pure decoherence effects which suppress the oscillating terms, and the so called relaxation effects which affect non-oscillating terms. In~\cite{Morgan:2004vv,Oliveira:2014jsa} it was shown that, for atmospheric neutrino oscillations, the relaxation effects which arise when the energy is not conserved are proportional to $\cos^22\theta_{23}$. This suppresses relaxation effects with respect to pure decoherence effects by at least two orders of magnitude, since $\cos^2 2\theta_{23}$ is currently constrained by experimental data at the level $\cos^2 2\theta_{23} < 0.034$ at 95\% CL~\cite{nufit,Esteban:2016qun}. We will thus focus on the analysis of pure decoherence effects in atmospheric neutrino oscillations assuming that the neutrino energy is conserved, and therefore eq.~(\ref{eq:diag}) satisfied, along the whole propagation. From a model-independent point of view, the $d_m^j$ are free parameters that could a priori depend on the matter effects. However, the most common assumption in the literature is to assume that the $d_m^j$ are independent of the matter density even in presence of matter effects.\footnote{This is the case for instance when decoherence is originated by quantum gravity.}. In order to be consistent with most previous studies and to compare the bounds obtained in our analysis with the constraints derived in previous publications, we will also assume that this is the case. Notice that this assumption does not imply that the matter effects are not relevant when neutrino propagation is affected by decoherence: it just implies that the $d_m^j$ are assumed to be constant during neutrino propagation through the Earth. \subsection{Neutrino propagation in uniform matter} \label{sec:uniform} Performing the following change of basis \begin{equation} \tilde{\rho}=\tilde{U}^\dagger\rho\tilde{U}, \end{equation} eq.~(\ref{eq:1}) can be rewritten as \begin{align} \frac{d\tilde{\rho}}{dt} &= -i\left[H_d,\tilde{\rho} \right]- \sum_m\left[\left\{\tilde{\rho},(D_m^d)^2\right\}-2D_m^d\,\tilde{\rho} \,D_m^{d}\right]\nonumber\\ &-\tilde{U}^\dagger\frac{d\tilde{U}}{dt}\tilde{\rho}-\tilde{\rho}\frac{d\tilde{U}^\dagger}{dt}\tilde{U}. \label{eq:mass} \end{align} If the matter profile is constant along the neutrino path, the system of equations becomes diagonal in $\tilde{\rho}_{ij}$ \begin{equation} \frac{d\tilde{\rho}_{ij}}{dt}=-\left[\gamma_{ij}-i\Delta \tilde{h}_{ij}\right]\tilde{\rho}_{ij}, \label{eq:cte} \end{equation} where we have defined \begin{equation} \gamma_{ij}\equiv \sum_m\left(d_m^i-d_m^j\right)^2=\gamma_{ji}>0\,;\;\;\;\;\;\;\; \Delta \tilde{h}_{ij}=\tilde{h}_i-\tilde{h}_j. \label{eq:gamma_dm} \end{equation} Therefore, the solution of eq.~(\ref{eq:1}) for constant matter is simply given by \begin{equation} \rho_{\alpha\beta}(t)= \left[\tilde{U}\tilde{\rho}(t)\tilde{U}^\dagger\right]_{\alpha\beta}, \label{sol1} \end{equation} with \begin{equation} \tilde{\rho}_{ij}(t)=\tilde{\rho}_{ij}(0)\, e^{-\left[\gamma_{ij}-i\Delta \tilde{h}_{ij}\right]t}, \label{sol2} \end{equation} where $\tilde{\rho}_{ij}(0)$ is determined by the initial conditions of the system. For instance, if the source produces only neutrinos of flavor $\alpha$, the initial conditions are given by \begin{equation} \tilde{\rho}_{ij}(0)=\tilde{U}^*_{\alpha i}\tilde{U}_{\alpha j}\, . \label{init_alpha} \end{equation} As a result, the oscillation probabilities in presence of decoherence (for a constant matter profile) read \begin{align} P_{\alpha\beta}& \equiv P(\nu_\alpha \to \nu_\beta) = \text{Tr}\left[\hat{\rho}^{(\alpha)}(t)\hat{\rho}^{(\beta)}(0)\right]\nonumber\\ &= \text{Tr} \left[\hat{\rho}^{(\alpha)}(t)|\nu_\beta\rangle\langle\nu_\beta|\right]=\langle\nu_\beta|\hat{\rho}^{(\alpha)}(t)|\nu_\beta\rangle \nonumber\\ &=\sum_{i,j}\tilde{U}_{\beta i}\tilde{U}^*_{\beta j}\, \tilde{\rho}_{ij}(t) \nonumber\\ &=\sum_{i,j}\tilde{U}^*_{\alpha i}\tilde{U}_{\beta i}\tilde{U}_{\alpha j}\tilde{U}^*_{\beta j} e^{-\left[\gamma_{ij}-i\Delta \tilde{h}_{ij}\right]t} \, , \end{align} where $\hat\rho$ denotes the density operator. Finally, after some manipulation the above equation can be rewritten in the more familiar form \begin{align} P_{\alpha\beta}&=\delta_{\alpha\beta}-2\sum_{i<j}{\rm Re}\left[\tilde{U}_{\alpha i}^*\tilde{U}_{\beta i} \tilde{U}_{\alpha j}\tilde{U}_{\beta j}^*\right]\left(1-e^{-\gamma_{ij}L}\cos\tilde{\Delta}_{ij}\right) \nonumber\\ &-2 \sum_{i<j}{\rm Im}\left[\tilde{U}_{\alpha i}^*\tilde{U}_{\beta i} \tilde{U}_{\alpha j}\tilde{U}_{\beta j}^*\right]e^{-\gamma_{ij}L}\sin\tilde{\Delta}_{ij}, \label{P_alpha_beta} \end{align} with \begin{equation} \label{gammaE} \tilde{\Delta}_{ij} \equiv\frac{\Delta\tilde{m}^2_{ij}L}{2E},\;\;\;\;\;\;\;\;\;\; \gamma_{ij}=\gamma_{ji}\equiv\gamma_{ij}^0\left(\frac{E}{\text{GeV}}\right)^n, \end{equation} % where $\Delta\tilde{m}^2_{ij} \equiv \tilde{m}^2_i - \tilde{m}^2_j$ are the effective mass squared differences of neutrinos in matter and we have used the approximation $L\approx t$, $L$ being the distance traveled by the neutrino as it propagates. Note that the power-law dependence on the neutrino energy given by eq.~(\ref{gammaE}) breaks Lorentz invariance, except for the case with $n=-1$ which gives similar effects to the neutrino decay (see e.g. \cite{GonzalezGarcia:2008ru}). However, the effect encoded in $\gamma_{ij}$ only suppresses the oscillatory terms in the oscillation probability while a neutrino decay would also affect the non-oscillatory terms. Therefore, in the framework considered in this work the total sum of the probabilities adds up to $1$, while this is not the case for neutrino decay. From eqs.~(\ref{P_alpha_beta}) and (\ref{gammaE}), one would expect to have a sizable effect in neutrino oscillations for $ \gamma_{ij} L \sim 1$. This condition gives an estimate of the values of $\gamma_{ij}$ for which an effect may be experimentally observable: \begin{equation} \label{gammaL} \gamma_{ij}^0 \sim 1.7 \cdot 10^{-19} \left(\frac{L}{\text{km}}\right)^{-1} \left(\frac{E}{\text{GeV}}\right)^{-n} \ \text{GeV}. \end{equation} Nevertheless, we would like to remark that fulfilling this condition is not enough to have sensitivity to decoherence effects, as we will discuss in the next subsection. Even though in our simulations we will numerically compute the exact oscillation probabilities, in order to understand qualitatively the impact of decoherence on the oscillation pattern it is useful to derive approximate analytical expressions. In this work, we will be focusing on the study of atmospheric neutrino oscillations, for which the oscillation channel $P_{\mu\mu}$ is most relevant. Recently, in~\cite{Denton:2016wmg,Denton:2018hal} approximated but very accurate analytical expressions for the standard oscillation probabilities in presence of constant matter density were derived. For the $\nu_\mu \to \nu_\mu$ oscillation channel including decoherence effects, using the same parametrization as in ref.~\cite{Denton:2018hal}, we find: \begin{align} \label{PmumuActe} P_{\mu\mu}&=1-A_{21} \left[1-e^{-\gamma_{21}L}\cos\tilde{\Delta}_{21}\right]-A_{32} \left[1-e^{-\gamma_{32}L}\cos\tilde{\Delta}_{32}\right]\nonumber\\ &-A_{31} \left[1-e^{-\gamma_{31}L}\cos\tilde{\Delta}_{31}\right], \end{align} where \begin{align} A_{ij} &\equiv A_{ij}(\theta_{23},\tilde{\theta}_{12},\tilde{\theta}_{13},\delta)\nonumber\\ &= 2|U_{\mu i}(\theta_{23},\tilde{\theta}_{12},\tilde{\theta}_{13},\delta)|^2|U_{\mu j}(\theta_{23},\tilde{\theta}_{12},\tilde{\theta}_{13},\delta)|^2, \end{align} and the effective mass splittings and mixing angles in matter can be expressed as~\cite{Denton:2018hal}: {\small \begin{align} \cos2\tilde{\theta}_{13}&=\frac{\cos2\theta_{13}-a/\Delta m^2_{ee}}{\sqrt{(\cos2\theta_{13}-a/\Delta m^2_{ee})^2+\sin^2 2\theta_{13}}}, \nonumber\\ \cos 2\tilde{\theta}_{12}&=\frac{\cos2\theta_{12}-a'/\Delta m^2_{21}}{\sqrt{(\cos2\theta_{12}-a'/\Delta m^2_{21})^2+\sin^2 2\theta_{12}\cos^2(\tilde{\theta}_{13}-\theta_{13})}}, \nonumber\\ \Delta\tilde{m}^2_{21}&=\Delta m^2_{21}\sqrt{(\cos2\theta_{12}-a'/\Delta m^2_{21})^2+\sin^2 2\theta_{12}\cos^2(\tilde{\theta}_{13}-\theta_{13})}, \nonumber\\ \Delta\tilde{m}^2_{31}&=\Delta m^2_{31}+(a-\frac{3}{2}a')+\frac{1}{2}(\Delta \tilde{m}^2_{21}-\Delta m^2_{21}), \nonumber\\ \Delta\tilde{m}^2_{32}&=\Delta\tilde{m}^2_{31}-\Delta\tilde{m}^2_{21}. \end{align}} Here, $a \equiv 2\sqrt{2}G_Fn_e E$, where $G_F$ is the Fermi constant and $n_e$ is the electron number density along the neutrino path, $\Delta m^2_{ee}\equiv \cos^2\theta_{12}\Delta m^2_{31}+\sin^2\theta_{12}\Delta m^2_{32}$, and $a'= a\cos^2\tilde{\theta}_{13}+\Delta m^2_{ee}\sin^2(\tilde{\theta}_{13}-\theta_{13})$. The corresponding probability for antineutrinos is obtained simply replacing $a \to -a$ and $\delta \to -\delta$, where $\delta$ denotes the Dirac CP phase. \subsection{Neutrino propagation in non-uniform matter: adiabatic regime} \label{sec:adiabatic} Equation~(\ref{P_alpha_beta}) applies for constant density profiles (which is a very good approximation in the case of long-baseline neutrino oscillation experiments such as T2K or NOvA), but if the matter density is not constant the analysis becomes more complicated. Nevertheless, when the adiabaticity condition ${d \tilde{U}}/{dt}\ll 1$ is fulfilled, as in the solar neutrino case, the solution of the evolution equations given by eqs.~(\ref{sol1}) and (\ref{sol2}) is still a good approximation. In such a case, the oscillation probability is given by \begin{align} P_{\alpha\beta} &= \langle\nu_\beta|\hat{\rho}^{(\alpha)}(t)|\nu_\beta\rangle\nonumber\\ &=\sum_{i,j}\tilde{\rho}^{(\alpha)}_{ij}(0) e^{-\left[\gamma_{ij}-i\Delta \tilde{h}_{ij}\right]t}\langle\nu_\beta|\tilde{\nu}_i^{eff}\rangle\langle\tilde{\nu}_j^{eff}|\nu_\beta\rangle, \end{align} where $\nu_i^{eff}$ denotes the effective mass eigenstates at time $t$. In the case of solar neutrinos, the initial flux of $\nu_e$ is produced in the solar core and the initial conditions are given by: \begin{equation} \tilde{\rho}^{(e)}_{ij}(0)=\tilde{U}_{e i}^{0*}\tilde{U}_{e j}^{0}, \end{equation} where $\tilde{U}^{0}$ denotes the effective mixing matrix at the production point. On the other hand, since the evolution is adiabatic, when the neutrinos come out from the Sun we have $|\tilde{\nu}_i^{eff}\rangle = |\nu_i\rangle$ and thus \begin{align} P_{e \beta}& \approx\sum_{i,j} \tilde{U}_{e i}^{0*}U_{\beta i}\tilde{U}^{0}_{e j}U_{\beta j}^* e^{-\left[\gamma_{ij}-i\Delta \tilde{h}_{ij}\right]t} \nonumber\\ &= \sum_{i} |\tilde{U}_{e i}^{0}|^2|U_{\beta i}|^2\nonumber \\ &+2\sum_{i<j}{\rm Re}\left[\tilde{U}_{e i}^{0*}U_{\beta i} \tilde{U}_{e j}^{0}U_{\beta j}^*\right]e^{-\gamma_{ij}t}\cos\tilde{\Delta}_{ij}\nonumber\\ &-2 \sum_{i<j}{\rm Im} \left[\tilde{U}_{e i}^{0*}U_{\beta i} \tilde{U}_{e j}^{0}U_{\beta j}^*\right]e^{-\gamma_{ij}t}\sin\tilde{\Delta}_{ij}. \end{align} Finally, for solar neutrinos observed at the Earth we obtain, after averaging over the oscillating phase: \begin{equation} P_{e \beta} \approx \sum_{i} |\tilde{U}_{e i}^{0}|^2|U_{\beta i}|^2, \end{equation} which coincides with the standard result. In other words, the decoherence effects encoded in $\gamma_{ij}$ can not be bounded by solar neutrino oscillation experiments. This is due to the standard loss of coherence in the propagation from the Sun to the Earth, which strongly suppresses the oscillating terms. In a similar fashion high-energy astrophysical neutrinos at IceCube are not sensitive to decoherence either, since these neutrinos are produced in distant astrophysical sources and thus the oscillations will have averaged out by the time they reach the detector. \subsection{Neutrino propagation in non-uniform matter: layers of constant density} \label{sec:layers} In the atmospheric neutrino case, the matter profile cannot be considered constant since the neutrinos propagate through the Earth crust, mantle and core, which have different densities. The adiabaticity condition is not fulfilled either. In this case, eq.~(\ref{eq:mass}) should be solved including the non-adiabatic terms, which give non-diagonal contributions. Even though this can be done numerically, we will show that dividing the matter profile into layers of constant density considerably simplifies the analysis and reduces the computational complexity of the problem. In particular, this is crucial in the case of atmospheric neutrino oscillation experiments, for which numerical studies are already computationally demanding even in the standard three-family scenario. Dividing the matter profile into layers of different constant densities has proved to be a very good approximation in the standard three-family scenario and, therefore, we expect the same level of accuracy in presence of decoherence. Since the matter is constant in each layer, the evolution equations can be solved for each layer $M$ as in sec.~\ref{sec:uniform}: \begin{align} \rho_{\alpha\beta}^M (t_M) & = \left[ \tilde{U}^{M}\tilde{\rho}^{M}(t_M)(\tilde{U}^M)^\dagger \right]_{\alpha\beta}, \nonumber\\ \tilde{\rho}^{M}_{ij}(t_M)&=\tilde{\rho}^{M}_{ij}(t_{M,0})\, e^{-\left[\gamma_{ij}-i\Delta \tilde{h}^{M}_{ij}\right]\Delta t_M} , \label{sol_layer} \end{align} where $\Delta t_M \equiv t_M- t_{M,0}$, and $t_{M,0}$ and $t_M$ denote the initial and final times for the propagation along layer $M$, respectively. Now the problem of computing the probability is just reduced to performing properly the matching among the evolution on the different layers. Let us first consider the simplest case of two layers $A$ and $B$. The oscillation probability when the neutrino exits the second layer (at time $t_B$) is given by \begin{equation} P_{\alpha\beta}= \langle\nu_\beta|\hat{\rho}^{(\alpha)}(t_B)|\nu_\beta\rangle =\sum_{i,j}\tilde{U}^{B}_{\beta i}\,\tilde{U}^{B*}_{\beta j}\, \tilde{\rho}^{B}_{ij}(t_{B,0})\, e^{-\left[\gamma_{ij}-i\Delta \tilde{h}^{B}_{ij}\right]\Delta t_{B} } \, . \label{Pab_layer1} \end{equation} The key point here is that the matching should be done between the solutions of eq.~(\ref{eq:1}) at the frontier between the two layers and in the flavor basis, as \begin{equation} \rho^{A}_{\alpha\beta}(t_A) = \rho^{B}_{\alpha\beta}(t_{B,0}). \label{matching} \end{equation} After imposing the matching condition, the elements of the density matrix in the second layer at $t_{B,0}$ can be written in the matter basis as: \begin{align} \tilde{\rho}^{B}_{ij}(t_{B,0}) &= \left[(\tilde{U}^{B})^\dagger\tilde{U}^{A}\tilde{\rho}^{A}(t_A)(\tilde{U}^{A})^\dagger \tilde{U}^{B}\right]_{ij} \nonumber\\ &= \tilde{U}^{B*}_{\delta i}\tilde{U}^{A}_{\delta l}\tilde{\rho}^{A}_{l n}(t_{A,0}) e^{-\left[\gamma_{ln}-i\Delta \tilde{h}^{A}_{ln}\right]\Delta t_A} \tilde{U}^{A*}_{\gamma n}\tilde{U}^{B}_{\gamma j} \nonumber\\ &= \tilde{U}^{B*}_{\delta i}\tilde{U}^{A}_{\delta l} \tilde{U}^{A*}_{\alpha l} \tilde{U}^{A}_{\alpha n} e^{-\left[\gamma_{ln}-i\Delta \tilde{h}^{A}_{ln}\right]\Delta t_A} \tilde{U}^{A*}_{\gamma n}\tilde{U}^{B}_{\gamma j}, \end{align} where we have considered that the initial flux is made of $\nu_\alpha$ as initial condition for the first layer. After substituting this result into eq.~(\ref{Pab_layer1}) we finally obtain \begin{align} P_{\alpha\beta}&= \sum_{\delta,\gamma,i,j,l,n} \tilde{U}^{B}_{\beta i}\tilde{U}^{B*}_{\delta i}\tilde{U}^{B}_{\gamma j}\tilde{U}^{B*}_{\beta j} e^{-\left[\gamma_{ij}-i\Delta \tilde{h}^{B}_{ij}\right]\Delta t_B} \nonumber\\ &\times \tilde{U}^{A}_{\delta l}\tilde{U}^{A*}_{\alpha l}\tilde{U}^{A}_{\alpha n}\tilde{U}^{A*}_{\gamma n} e^{-\left[\gamma_{ln}-i\Delta \tilde{h}^{A}_{ln}\right]\Delta t_A} \, . \end{align} It can be easily checked that, in the limit $\gamma_{ij}\rightarrow 0$, the standard oscillation probability is recovered. In the three-layer case, following the same procedure we find \begin{align} P_{\alpha\beta} & = \sum_{\delta,\gamma,\theta,\phi,i,j,l,n,m,k} \tilde{U}^{C}_{\beta i}\tilde{U}^{C*}_{\delta i}\tilde{U}^{C}_{\gamma j}\tilde{U}^{C*}_{\beta j} e^{-\left[\gamma_{ij}-i\Delta \tilde{h}^{C}_{ij}\right]\Delta t_C} \nonumber\\ & \times \tilde{U}^{B}_{\delta l}\tilde{U}^{B*}_{\theta l}\tilde{U}^{B}_{\phi n}\tilde{U}^{B*}_{\gamma n} e^{-\left[\gamma_{ln}-i\Delta\tilde{h}^{B}_{ln}\right]\Delta t_B} \nonumber\\ & \times \tilde{U}^{A}_{\theta m}\tilde{U}^{A*}_{\alpha m}\tilde{U}^{A}_{\alpha k}\tilde{U}^{A*}_{\phi k} e^{-\left[\gamma_{mk}-i\Delta \tilde{h}^{A}_{mk}\right]\Delta t_A}. \label{eq:P-3layers} \end{align} The procedure can be easily generalized to an arbitrary number of layers. Indeed, under the approximation $L \approx t$, and defining \begin{equation} \tilde{\mathcal{A}}^{M}_{\alpha\beta\gamma\delta} \equiv \sum_{i,j} \tilde{U}^{M}_{\alpha i}\tilde{U}^{M*}_{\beta i}\tilde{U}^{M}_{\gamma j}\tilde{U}^{M*}_{\delta j} e^{-\left[\gamma_{ij}-i(\Delta \tilde{m}^{M}_{ij})^2/2E\right] \Delta L_M}, \end{equation} the probabilities can be written in a more compact way as for two layers \begin{equation} P_{\alpha\beta} = \sum_{\delta,\gamma} \tilde{\mathcal{A}}^{B}_{\beta\delta\gamma\beta} \tilde{\mathcal{A}}^{A}_{\delta\alpha\alpha\gamma}, \end{equation} for three layers \begin{equation} P_{\alpha\beta} = \sum_{\delta,\gamma,\theta,\phi} \tilde{\mathcal{A}}^{C}_{\beta\delta\gamma\beta} \tilde{\mathcal{A}}^{B}_{\delta\theta\phi\gamma} \tilde{\mathcal{A}}^{A}_{\theta\alpha\alpha\phi}, \end{equation} for N layers \begin{equation} P_{\alpha\beta} = \sum_{\delta,\gamma,\theta,\phi,...,\xi,\omega,\varphi,\rho} \tilde{\mathcal{A}}^{N}_{\beta\delta\gamma\beta} \tilde{\mathcal{A}}^{N - 1}_{\delta\theta\phi\gamma} ...\; \tilde{\mathcal{A}}^{B}_{\xi\varphi\rho\omega} \tilde{\mathcal{A}}^{A}_{\varphi\alpha\alpha\rho} \phantom{mmmm}. \end{equation} \section{Atmospheric oscillation probabilities with decoherence} \label{sec:plots-probabilities} Atmospheric neutrino oscillations take place in a regime where matter effects are significant and can even dominate the oscillations. The relevance of matter effect increases with neutrino energy and is very different for neutrinos and antineutrinos, as the sign of the matter potential changes between the two cases. Matter effects also depend strongly on the neutrino mass ordering. In order to understand better the numerical results shown in this paper, it is useful to derive approximate expressions for the oscillations in the $\nu_\mu\to\nu_\mu$ and $\bar\nu_\mu\to\bar\nu_\mu$ channels in the presence of strong matter effects. From the results obtained in refs.~\cite{Denton:2016wmg, Denton:2018hal}, for neutrino energies $ E \gsim 15$~GeV matter effects drive the effective mixing angles in matter $\tilde\theta_{12}$ and $\tilde\theta_{13}$ to either 0 or $\pi/2$, depending on the channel (neutrino/antineutrino) and the mass ordering. It is easy to show that, in this regime, the oscillation probabilities in eq.~\eqref{PmumuActe} can be approximated as: \begin{equation} P^{\mathrm{NO}}_{\mu\mu} \approx 1 - \frac{1}{2}\sin^22\theta_{23}\left( 1 - e^{-\gamma_{21} L } \cos\tilde \Delta_{21} \right) \label{eq:NOnu} \end{equation} for neutrinos, and \begin{equation} P^{\mathrm{NO}}_{\bar\mu\bar\mu} \approx 1 - \frac{1}{2}\sin^22\theta_{23}\left( 1 - e^{-\gamma_{32} L } \cos\tilde \Delta_{32} \right) \label{eq:NOnubar} \end{equation} for antieneutrinos, assuming a normal ordering (NO). For inverted ordering (IO) we get instead \begin{equation} P^{\mathrm{IO}}_{\mu\mu} \approx 1 - \frac{1}{2}\sin^22\theta_{23}\left( 1 - e^{-\gamma_{31} L } \cos\tilde \Delta_{31} \right) \label{eq:IOnu} \end{equation} for neutrinos, \begin{equation} P^{\mathrm{IO}}_{\bar\mu\bar\mu} \approx 1 - \frac{1}{2}\sin^22\theta_{23}\left( 1 - e^{-\gamma_{21} L } \cos\tilde \Delta_{21} \right) \label{eq:IOnubar} \end{equation} for antineutrinos. In the limit $\gamma_{ij} = 0$, eqs.~\eqref{eq:NOnu}-\eqref{eq:IOnubar} reassemble the standard neutrino oscillation probabilities derived under the one-dominant mass-scale approximation~\cite{Pantaleone:1993di}. From eqs.~\eqref{eq:NOnu}-\eqref{eq:IOnubar} it is easy to see that the approximated oscillation probabilities for an inverted mass ordering can be obtained from the corresponding ones for normal mass ordering, just performing the following transformation: \begin{align} \gamma_{21}, \tilde{\Delta}_{21} & \to \gamma_{31}, \tilde{\Delta}_{31} , \label{eq:map1} \\ \gamma_{32}, \tilde{\Delta}_{32} & \to \gamma_{21}, \tilde{\Delta}_{21} . \label{eq:map2} \end{align} Moreover, note that since the three decoherence parameters and the three mass splittings are related (see eqs.~(\ref{eq:gamma_dm}) and (\ref{gammaE})), these two transformations automatically imply that \begin{equation} \label{eq:map3} \gamma_{31},\tilde{\Delta}_{31} \to \gamma_{32}, \tilde{\Delta}_{32}. \end{equation} Equations~\eqref{eq:NOnu}-\eqref{eq:IOnubar} illustrate why a proper consideration of the matter effects in the context of three families is of key importance in order to correctly interpret the bounds extracted within a simplified two-flavour approximation (as done in e.g. ref.~\cite{Lisi:2000zt}). According to our analytical results, which will be confirmed numerically below, the constraints obtained from SK in a two-family approximation cannot be simply applied to $\gamma_{31}$ or $\gamma_{32}$, contrary to the naive expectation. In fact, the interpretation of such limits depends strongly on the ordering of neutrino masses and on whether the sensitivity is dominated by the neutrino or antineutrino channels: for neutrinos the decoherence effects at high energies would be mainly driven by $\gamma_{21}$ ($\gamma_{31}$) for normal (inverted) ordering. Conversely, in the antineutrino channel decoherence effects are essentially controlled by $\gamma_{32}$ ($\gamma_{21 }$) for normal (inverted) ordering. Therefore, we conclude that in order to avoid any misinterpretation of the bounds from atmospheric neutrinos, a three-family approach including matter effects should be considered. \begin{figure*}[h!] \centering \includegraphics[width=.95\columnwidth]{figures/PmumuNH.pdf} \includegraphics[width=.95\columnwidth]{figures/PamuamuNH.pdf} \includegraphics[width=.95\columnwidth]{figures/PmumuIH.pdf} \includegraphics[width=.95\columnwidth]{figures/PamuamuIH.pdf} \caption{The $\nu_\mu \to \nu_\mu$ (top panels) and $\overline{\nu}_\mu \to \overline{\nu}_\mu$ (bottom panels) oscillation probabilities with ($n=0$) and without decoherence effects as a function of the neutrino energy. The probabilities have been computed for normal (left panels) and inverted (right panels) neutrino mass ordering, using a three-layer model for the Earth matter density profile, and correspond to the case in which the neutrinos cross the center of the Earth. In this figure, in the cases where decoherence effects are included we have set the parameters listed in the legend to the same constant value $\gamma = 2.3 \cdot 10^{-23}$~GeV. } \label{fig:prob_mumu} \end{figure*} Figure~\ref{fig:prob_mumu} shows the numerically obtained $\nu_\mu \to \nu_\mu$ (top panels) and $\overline{\nu}_\mu \to \overline{\nu}_\mu$ (bottom panels) oscillation probabilities for NO (left panels) and IO (right panels), with and without decoherence, as a function of the neutrino energy for a three-layer model (details on the accuracy of our three-layer model and the specific parameters used in our simulations can be found in~\ref{app}). For the sake of simplicity, in this section we focus on the case $n=0$, where the $\gamma_{ij}$ do not depend on the neutrino energy (the results for different values of $n$ show a similar qualitative behavior). The standard oscillation parameters have been fixed to the best-fit values given in~\cite{nufit,Esteban:2016qun}. Figure~\ref{fig:prob_mumu} clearly shows how the decoherence tends to damp the oscillatory behavior, in qualitative agreement with eq.~(\ref{PmumuActe}) and the corresponding approximate expressions given by eqs.~\eqref{eq:NOnu}-\eqref{eq:IOnubar}. Note that eq.~\eqref{PmumuActe} has been obtained under several approximations and, in particular, considering only one layer with constant matter density. However, we should stress that in our simulations the computation of the probability has been done numerically, considering a three-layer matter profile (see~\ref{app} for details). Since the three $\gamma_{ij}$ are not completely independent from one another (see eq.~(\ref{eq:gamma_dm})), in view of eqs.~\eqref{eq:NOnu}-\eqref{eq:IOnubar} and in order to simplify the analysis, hereafter we will distinguish three different representative cases, where the decoherence effects are dominated by just one parameter: \begin{enumerate}[label=(\Alph*)] \item Atmospheric limit: $\gamma_{21}=0$ ($\gamma_{32} = \gamma_{31}$), \item Solar limit I: $\gamma_{32}=0$ ($\gamma_{21} = \gamma_{31}$), \item Solar limit II: $\gamma_{31}=0$ ($\gamma_{21} = \gamma_{32}$). \end{enumerate} \noindent In~\ref{sec:d5results}, we will show that the bounds derived in these limits correspond to the most conservative bounds that can be extracted in the general case. As a reference value for the decoherence parameters, in this section we have considered $\gamma = 2.3 \cdot 10^{-23}$~GeV, for each of the three limiting cases listed above. The results in fig.~\ref{fig:prob_mumu} show that, for neutrinos with a NO (top left panel), the impact of decoherence is essentially controlled by $\gamma_{21}$, in good agreement with eq.~\eqref{eq:NOnu}: no significant effects are seen in the atmospheric limit (A), while a similar impact is obtained in the solar limits I (B) and II (C). Conversely, for IO (top right panel) the effects are dominated by $\gamma_{31}$ instead: no effect is observed for the solar limit II (C), while in scenarios (A) and (B) the effect is very similar. This can be qualitatively understood from the approximate probability derived in eq.~\eqref{eq:IOnu}, which only depends on the decoherence parameter $\gamma_{31}$. On the other hand, in the antineutrino case for NO (bottom left panel) no observable decoherence effects take place in case (B), while cases (A) and (C) show a similar behavior, in agreement with eq.~\eqref{eq:NOnubar}. Conversely, for IO (bottom right panel) decoherence effects are essentially controlled by $\gamma_{21}$ as shown in eq.~\eqref{eq:IOnubar}: therefore, no significant effects are observed in case (A) while a similar impact is obtained for case (B) and (C). \begin{figure*}[ht!] \centering \includegraphics[width=1.65\columnwidth]{figures/GridNH.png} \caption{ Oscillograms for the neutrino oscillation probability $P_{\mu\mu}$, assuming normal mass ordering. The top left panel corresponds to the case of no decoherence $\gamma_{ij} = 0$ whereas the rest of the panels correspond to the three limiting cases mentioned in the text: (A) $\gamma_{32} = \gamma_{31} $ (top right), (B) $\gamma_{31} = \gamma_{21} $ (bottom left) and (C) $\gamma_{32} = \gamma_{21} $ (bottom right). In all cases, the size of the decoherence parameters that are turned on is set to a constant value, $\gamma = 2.3 \cdot 10^{-23}$~GeV.} \label{fig:OscillogramNH} \end{figure*} Moreover, it should be pointed out that the transformations listed in Eqs.~\eqref{eq:map1}-\eqref{eq:map3} automatically imply the following equivalence for the results obtained in the three limiting cases listed above: \begin{align} ({\rm A})^{\rm NO} &\longleftrightarrow ({\rm C})^{\rm IO}, \nonumber\\ ({\rm B})^{\rm NO} &\longleftrightarrow ({\rm A})^{\rm IO}, \label{eq:map} \\ ({\rm C})^{\rm NO} &\longleftrightarrow ({\rm B})^{\rm IO}. \nonumber \end{align} This is also confirmed at the numerical level, as it can be easily seen by comparing the different lines shown in the left (NO) and right (IO) panels in fig.~\ref{fig:prob_mumu} for the three limiting cases. It is also remarkable that, for both normal and inverted mass orderings, even when the standard oscillations turn off (at very high energies), there is still a large effect on the probability due to decoherence effects, that could potentially be tested with neutrino telescopes like IceCube. In particular, for $E\gsim 200$ GeV one can approximate $\cos\tilde{\Delta}_{ij}\approx 1$, $\forall i,j $. Therefore, in the standard case (with $\gamma_{ij}=0$) the last three terms in eq.~(\ref{PmumuActe}) approximately vanish, leading to $P_{\mu\mu}\approx1$. However, in presence of decoherence those terms will not vanish completely, as $e^{-\gamma_{ij} L}\cos\tilde{\Delta_{ij}} \neq 1$. This leads to a depletion of $P_{\mu\mu}$, which is no longer equal to 1 in this case. The size of the effect will of course depend on the baseline of the experiment. Since at high energies the oscillation probability does no longer depend on the neutrino energy, at oscillation experiments with a fixed baseline the effect may be hindered by the presence of any systematic errors affecting the normalization of the signal event rates. However, at atmospheric experiments this effect can be disentangled from a simple normalization error by comparing the event rates at different nadir angles. The dependence of the neutrino probabilities with the zenith angle $\theta_z$ is illustrated in fig.~\ref{fig:OscillogramNH}, assuming a normal mass ordering and fixing the standard oscillation parameters to the best-fit values given in~\cite{nufit,Esteban:2016qun}. The results are shown as a neutrino oscillogram (see for instance \cite{Akhmedov:2006hb}), which represents the oscillation probability in the $P_{\mu\mu}$ channel as a function of neutrino energy and zenith angle $\theta_z$ (which can be related to the distance traveled by the neutrino). Figure~\ref{fig:OscillogramNH} shows the oscillation probability $P_{\mu\mu}$ in the three limiting cases described above, comparing it to the results in the standard scenario ($\gamma_{ij} = 0$). As expected, the effects depend on the direction of the incoming neutrino and they are more relevant in the region $-1\lesssim \cos\theta_z \lesssim -0.4$, this is, for very long baselines. This was to be expected, since the decoherence effects are driven by $e^{-\gamma_{ij}L}$. In addition, the dependence of the oscillation probability with the zenith angle at very high energies ($ E > 100$~GeV) is clearly visible in the bottom panels of fig.~\ref{fig:OscillogramNH}. As we will show in sec.~\ref{sec:results}, this will lead to an impressive sensitivity for the IceCube setup. Finally, note that the results for inverted ordering show similar features to those in fig.~\ref{fig:OscillogramNH}, once the mapping in eq.~(\ref{eq:map}) is applied, and are therefore not shown here. \section{IceCube/DeepCore simulation details and data set } \label{sec:icecube} The IceCube neutrino telescope, located at the South Pole, is composed of 5160 DOMs (Digital Optical Module) deployed between 1450~m and 2450~m below the ice surface along 86 vertical strings~\cite{Aartsen:2016nxy}. In the inner core of the detector, a subset of these DOMs were deployed deeper than 1750~m and closer to each other than in the rest of IceCube. This subset of strings is called DeepCore. Due to the shorter distance between its DOMs, the neutrino energy threshold in DeepCore ($\sim 5$~GeV) is lower than in IceCube ($\sim 100$~GeV). This allows DeepCore to observe neutrino events in the energy region where atmospheric oscillations take place, see fig.~\ref{fig:prob_mumu}, whereas IceCube only observes high-energy atmospheric neutrino events. As outlined in sec.~\ref{sec:formalism}, for high energy astrophysical neutrinos the effect of non-standard decoherence in the probability would be completely erased by the time they reach the detector. Therefore, in this work we will focus on the observation of atmospheric neutrino events at both IceCube and DeepCore, in the energy range $\sim10$ GeV to $\sim1$ PeV. In particular, we have used the three-year DeepCore data on atmospheric neutrinos with energies between $\sim10$ GeV and $\sim1$ TeV, taken between May 2011 and April 2014~\cite{Aartsen:2014yll}, and the one-year IceCube data taken between 2011-2012~\cite{TheIceCube:2016oqi,Jones:2015,Arguelles:2015}, corresponding to neutrinos with energies between 200~GeV and 1~PeV. At IceCube and DeepCore, events are divided according to their topology into ``tracks'' and ``cascades''~\cite{YanezGarza:2014jia}. Tracks are produced by the Cherenkov radiation of muons propagating in the ice. In atmospheric neutrino experiments, muons are typically produced by two main mechanisms: (1) via charged-current (CC) interactions of $\nu_{\mu}$ with nuclei in the detector, and (2) as decay products of mesons (typically pions and kaons) originated when cosmic rays hit the atmosphere. Conversely, cascades are created in CC interactions of $\nu_{e}$ or $\nu_\tau$\footnote{Technically, a CC $\nu_{\tau}$ event could be distinguished from a $\nu_e$ CC event, e.g., by the observation of two separates cascades connected by a track from the $\tau$ propagation~\cite{Aartsen:2015dlt}. However, for atmospheric neutrino energies the distance between the cascades cannot be resolved by the DOMs at IceCube/DeepCore, leaving in the detector a signal similar to a single cascade. }: in this case, the rapid energy loss of electrons as they move through the ice is the origin of an electromagnetic shower. At IceCube/DeepCore, cascades are also observed as the product of hadronic showers generated in neutral-current (NC) interactions for neutrinos of all flavors. Our analysis considers only track-like events observed at both IceCube and DeepCore although, as we will see, some small contamination from cascade events can be expected (especially at low energies). \subsection{IceCube simulation details} For IceCube, the observed event rates are provided in a grid of $10 \times 21$ bins~\cite{Jones:2015}, using 10 bins for the reconstructed energy (logarithmically spaced, ranging from 400 GeV to 20 TeV), and 21 bins for the reconstructed neutrino direction (linearly spaced, between $\cos\theta_z^{rec} = -1.02$ and $\cos\theta_{z}^{rec} = 0.24$). The muon energy is reconstructed with an energy resolution $\sigma_{\text{log}_{10}(E_{\mu}/\text{GeV})}\sim 0.5$~\cite{TheIceCube:2016oqi}, while the zenith angle resolution is in the range $\sigma_{\cos\theta_{z}}\in [0.005,0.015]$, depending on the scattering muon angle. \begin{figure*}[h!] \centering \includegraphics[width=1.85\columnwidth]{figures/GridIC.pdf} \caption{ Event distributions obtained for IceCube in our numerical simulations as a function of the reconstructed value of the cosine of the zenith angle, for neutrinos in different reconstructed energy ranges. The lines have been obtained assuming a normal mass ordering, for the following values for the decoherence parameters: $\gamma_{21} = \gamma_{31} = 2.3 \cdot 10^{-23}$ GeV (solid blue line), $\gamma_{21} = \gamma_{31} = 10^{-22}$ GeV (dot-dashed green line) and without decoherence (dashed red line). The observed data points~\cite{Jones:2015} are represented by the black dots, and the error bars indicate the statistical uncertainties for each bin. } \label{fig:event-distributionsIC} \end{figure*} The number of events in each bin is computed as: \begin{align} N_{i} (E^{rec}, \theta_z^{rec})= & \sum_{\pm} \int dE \; d\cos\theta_z \, \phi^\text{atm}_{\mu,\pm}(E,\theta_z ) P_{\mu\mu}^\pm(E,\theta_z )\nonumber \label{eq:events-IC} \\ & A^\text{eff}_{i, \pm, \mu}(E,\theta_z ,E^{rec},\theta_z^{rec}) e^{-X(\theta_z)\sigma^{\pm}(E)}, \end{align} where $E,\theta_z$ denote the true values of energy and zenith angle, while $E^{rec},\theta_z^{rec}$ refer to their reconstructed quantities. Here, $\phi^\text{atm}_{\mu,\pm}$ is the atmospheric flux for muon neutrinos~(+) and anti-neutrinos~(-), $P_{\mu\mu}^\pm(E,\theta_z)$ is the neutrino/antineutrino oscillation probability given by eq.~(\ref{eq:P-3layers}), and $A^\text{eff}_{i,\pm,\mu}(E,\theta_z)$ is the effective area encoding the detector response in neutrino energy and direction (which relates true and reconstructed variables), the interaction cross section and a normalization constant, and has been integrated over the whole data taking period. In our IceCube simulations, we have used the unpropagated atmospheric flux (HondaGaisser) provided by the collaboration~\cite{TheIceCube:2016oqi,IceCubeweb}, and for the effective area we have used the nominal detector response from refs.~\cite{TheIceCube:2016oqi,IceCubeweb}. On the other hand, the exponential factor takes into account the absorption of the neutrino flux by the Earth, which increases with the neutrino energy. Here, $X(\theta_z)$ is the column density along the neutrino path and $\sigma^{\pm}(E)$ is the total inclusive cross section for $\nu_{\mu}$ or $\bar\nu_\mu$. Note that in eq.~\eqref{eq:events-IC} no contamination from cascade events is considered since the mis-identification rate is expected to be negligible at these energies~\cite{Weaver:2015}. Similarly, the number of atmospheric muons that pass the selection cuts can also be neglected, given the extremely good angular resolution at these energies~\cite{TheIceCube:2016oqi}. Figure~\ref{fig:event-distributionsIC} shows the expected number of events for IceCube from our numerical simulations including decoherence, for $\gamma_{21} = \gamma_{31} = 2.3 \cdot 10^{-23}$ GeV (solid blue lines) and $\gamma_{21} = \gamma_{31} = 10^{-22}$ GeV (dot-dashed green lines), as a function of $\cos \theta_z^{rec}$, for events in different reconstructed energy ranges. For simplicity, we have considered the $n=0$ case (that is, $\gamma_{ij}$ independent of the neutrino energy). The expected result without decoherence is also shown for comparison (dashed red lines), while the observed data~\cite{Jones:2015} are shown by the black dots. For the analysis of the IceCube data we have performed a Poissonian log-likelihood analysis doing a simultaneous fit on the following parameters: $\Delta m^2_{32},\, \theta_{23}$ and $\gamma_{ij}$. The rest of the oscillation parameters have been kept fixed to their current best-fit values from ref.~\cite{nufit, Esteban:2016qun}. The most relevant systematic errors used in the fit are summarized in tab.~\ref{tab:sys-IC}, and have been taken from refs.~\cite{TheIceCube:2016oqi,Arguelles:2015,IceCubeweb}. For each systematic uncertainty a pull term is added to the $\chi^2$ following the values listed in the table, except in the cases indicated as ``Free'' (when the corresponding nuisance parameter is allowed to float freely in the fit). \begin{table} \begin{center} \caption{The most relevant systematic errors used in our analysis of IceCube data, taken from refs.~\cite{TheIceCube:2016oqi,Arguelles:2015,IceCubeweb}.\label{tab:sys-IC} } \begin{tabular}{l c } \hline\noalign{\smallskip} Source of uncertainty & Value \\ \noalign{\smallskip}\hline\noalign{\smallskip} Flux - normalization & Free \\ Flux - $\pi/K$ ratio & 10\% \\ Flux - energy dependence as $(E/E_0)^\eta $ & $\Delta \eta = 0.05$ \\ Flux - $\bar{\nu}/\nu$ & 2.5\%\\ DOM efficiency & 5\% \\ Photon scattering & 10\% \\ Photon absorption & 10\% \\ \noalign{\smallskip}\hline \end{tabular} \end{center} \end{table} \subsection{DeepCore simulation details} In the case of DeepCore, the observed event rates~\cite{Aartsen:2014yll} are provided in a grid of $8\times 8$ bins, using 8 bins for the reconstructed neutrino energy and 8 bins for the reconstructed neutrino direction. The energy resolution $\sigma_{E/\text{GeV}}$ is in the range of 30\%-20\% while the zenith angle resolution improves with the energy, from $\sigma_{\theta_{z}} = 12^{\circ}$ at $E_{\nu} = 10$~GeV to $\sigma_{\theta_{z}} = 5^{\circ}$ at $E_{\nu} = 40$~GeV~\cite{Aartsen:2014yll}. \begin{figure*}[ht!] \centering \includegraphics[width=1.78\columnwidth]{figures/GridDC.pdf} \caption{ Event distributions obtained for DeepCore in our numerical simulations as a function of the reconstructed values of the cosine of the zenith angle, for neutrinos in different reconstructed energy ranges. The lines have been obtained assuming a normal mass ordering, for the following values for the decoherence parameters: $\gamma_{21} = \gamma_{31} = 2.3 \cdot 10^{-23}$ GeV (solid blue line), $\gamma_{21} = \gamma_{31} = 10^{-22}$ GeV (dot-dashed green line) and without decoherence (dashed red line). The observed data points~\cite{Aartsen:2014yll} are represented by the black dots, and the error bars indicate the statistical uncertainties for each bin. } \label{fig:event-distributionsDC} \end{figure*} In each bin, the number of events is computed as \begin{align} N_{i} (E^{rec},\theta_z^{rec}) &= \sum_{\pm,\alpha,\beta} \int dE\, d\cos\theta_{z} \, \phi^\text{atm}_{\alpha,\pm}(E,\theta_{z}) \, P_{\alpha\beta}^\pm(E,\theta_{z})\nonumber\\ & A^\text{eff}_{i, \pm, \beta}(E,\theta_{z}, E^{rec},\theta_z^{rec}) \nonumber \\ &+N_{i,\mu} (E^{rec}, \theta_z^{rec}). \label{eq:events-DC} \end{align} Unlike for IceCube, at DeepCore muon tracks can be produced from $\nu_\mu \to \nu_\mu$ and $\nu_e \to \nu_\mu$ events\footnote{The flux from $\nu_\tau$ can be considered negligible at these energies.}. Moreover, the track-like event distributions at DeepCore will also receive partial contributions from cascades which are mis-identified as tracks: hence the sum over $\beta = e, \mu,\tau$ in eq.~\eqref{eq:events-DC}. Therefore, here $\phi^\text{atm}_{\alpha,\pm}$ stands for the atmospheric flux for neutrinos/antineutrinos of flavor $\alpha$ (where we have used the fluxes from ref.~\cite{Honda:2015fha}), and $ P_{\alpha\beta}^{\pm}$ refers to the neutrino/antineutrino oscillation probability in the channel $\nu_\alpha \to \nu_\beta$ for neutrinos (+) (or $ \bar\nu_\alpha \to\bar\nu_\beta$, for antineutrinos (-)). The rejection efficiencies for the contamination are included in the detector response function $A^\text{eff}_{i, \pm, \beta}$, which now depends on the flavor $\beta$ of the interacting neutrino. Finally, an estimate of the atmospheric muons that overcome the selection criteria (taken from refs.~\cite{Aartsen:2014yll,IceCubeweb}) is also added for each bin in reconstructed variables, $N_{i,\mu}$. Figure~\ref{fig:event-distributionsDC} shows the expected number of events for DeepCore obtained from our numerical simulations including decoherence, for $\gamma_{21} = \gamma_{31} = 2.3 \cdot 10^{-23}$ GeV (solid blue lines) and $\gamma_{21} = \gamma_{31} = 10^{-22}$ GeV (dot-dashed green lines), as a function of $\cos \theta_z^{rec}$, for events in different reconstructed energy ranges. For simplicity, we have considered the $n=0$ case (that is, $\gamma_{ij}$ independent of the neutrino energy). The expected result without decoherence is also shown for comparison (dashed red lines), while the observed data~\cite{Aartsen:2014yll} are shown by the black dots. In this work a Gaussian maximum likelihood is used to analyze the DeepCore data, performing a simultaneous fit on the following parameters: $\Delta m^2_{32},\, \theta_{23}$ and $\gamma_{ij}$. The rest of the oscillation parameters have been kept fixed to their current best-fit values from refs.~\cite{nufit, Esteban:2016qun}. The systematics used in the fit are those associated with the flux, the detector response and the atmospheric muons given in ref.~\cite{Aartsen:2014yll} and are summarized in tab.~\ref{tab:sys-DC}. For each systematic uncertainty a pull term is added to the $\chi^2$ following the values listed in the table, except in the cases indicated as ``Free'' (when the corresponding nuisance parameter is allowed to float freely in the fit). We have checked that our analysis reproduces the confidence regions in the $\Delta m_{32}^2 - \theta_{23}$ plane obtained by the DeepCore collaboration in ref.~\cite{Aartsen:2014yll} to a very good level of accuracy. \begin{table} \begin{center} \caption{Systematic errors used in our analysis of DeepCore data, taken from refs.~\cite{Aartsen:2014yll,YanezGarza:2014jia}. \label{tab:sys-DC} } \begin{tabular}{l c } \hline\noalign{\smallskip} Source of uncertainty & Value \\ \noalign{\smallskip}\hline\noalign{\smallskip} Flux - normalization & Free \\ Flux - energy dependence as $(E/E_0)^\eta $ & $\Delta \eta = 0.05$ \\ Flux - $(\nu_{e}+\bar{\nu}_e)/(\nu_{\mu}+\bar{\nu}_{\mu})$ ratio &20\%\\ Background - normalization & Free\\ DOM efficiency & 10\% \\ Optical properties of the ice & 1\%\\ \noalign{\smallskip}\hline \end{tabular} \end{center} \end{table} Finally, it should be noted that our fit does not include the latest atmospheric data recently published by the DeepCore collaboration~\cite{Aartsen:2017nmd}. The new analysis uses a different data set (from April 2012 to May 2015) and a new implementation of systematic errors, which lead to smaller confidence regions in the $\Delta m^2_{32} - \theta_{23}$ plane. However, the detector response parameters and systematic errors used in the latest release have not been published yet. In view of the better results obtained for the standard three-family oscillation scenario, a similar improvement is to be expected if the analysis performed in this work were to be repeated using the latest DeepCore data. \section{Results} \label{sec:results} Following the procedure described in sec.~\ref{sec:icecube} we have obtained the $\chi^2$ for every point in the parameter space. Marginalizing over the relevant mixing and mass parameters, namely, $\Delta m^2_{32}$ and $\theta_{23}$, the sensitivity of the data to $\gamma_{ij}$ parameters is determined by evaluating the $\sqrt{\Delta\chi^2}$, with $\Delta\chi^2 \equiv \chi^2 - \chi^2_\text{min}$, where $\chi^2_{\text{min}}$ is the value at the global minimum. In this section we will only show the results obtained for NO, since we have checked that extremely similar results are obtained for IO after applying the mapping given in eq.~(\ref{eq:map}). Nevertheless, in sec.~\ref{sec:conclusions} we will also provide the 95\% confidence level (CL) bounds obtained in our numerical analysis for the IO case. The bounds obtained are in very good agreement with the mapping given in eq.~(\ref{eq:map}). \begin{figure*}[ht!] \centering \includegraphics[scale = 0.3]{figures/GridXi.pdf} \caption{ Values of the $\sqrt{\Delta \chi^2}$ as a function of the decoherence parameter for the Atmospheric limit (red), Solar limit I (green) and Solar limit II (blue) defined in sec.~\ref{sec:plots-probabilities}. The results obtained from our analysis of IceCube (DeepCore) data are denoted by the solid (dashed) lines. The three panels have been obtained for NO, assuming a different dependence on the neutrino energy: $n=0$ (top panel), $n=1$ (middle panel) and $n=2$ (bottom panel). The shaded regions are disfavored by previous analysis of SK~\cite{Lisi:2000zt} and KamLAND~\cite{Gomes:2016ixi} data, see text for details. The horizontal black line indicates the value of the $\sqrt{\Delta \chi^2}$ corresponding to 95\% CL for 1 degree of freedom. } \label{fig:chi} \end{figure*} Figure~\ref{fig:chi} shows the obtained $\sqrt{\Delta \chi^2}$ as a function of $\gamma_0$ for the three limiting cases defined in sec.~\ref{sec:plots-probabilities}: (A) atmospheric limit, $\gamma_0 = \gamma_{32}^0=\gamma_{31}^0$ (red curve); (B) solar limit I, $\gamma_0 = \gamma_{21}^0=\gamma_{31}^0$ (green curve); and (C) solar limit II, $\gamma_0 = \gamma_{21}^0=\gamma_{32}^0$ (blue curve). In all cases, the solid (dashed) lines correspond to the results obtained from our analysis of the IceCube (DeepCore) data, and each panel shows the results obtained assuming a different energy dependence for the decoherence parameters, see eq.~\eqref{gammaE}: $n=0$ (top panel), $n=1$ (middle panel) and $n=2$ (bottom panel). The shaded regions are disfavored by previous analysis of SK~\cite{Lisi:2000zt} (90\% CL) and KamLAND~\cite{Gomes:2016ixi} data (95\% CL). As explained in sec.~\ref{sec:plots-probabilities}, the KamLAND constraints derived in~\cite{Gomes:2016ixi} apply to $\gamma_{12}^0$ (solar limits) while it is not clear to which $\gamma_{ij}$ the bounds from SK obtained in~\cite{Lisi:2000zt} would apply, since this depends on the true neutrino mass ordering (which is yet unknown). Note that the size of the atmosphere has been neglected in our calculations (see~\ref{app} for details). This is a good approximation for small values of the decoherence parameters, but it starts to fail if the decoherence effects are large enough to affect neutrinos with $\cos\theta_z > 0$. Therefore, in the case of IceCube we have shown our results only in the region where this approximation holds. In the case of DeepCore, due to the smaller energies considered, our approximation has no impact on the results even for large values of the decoherence parameters. Therefore, the approximation has only an impact on the IceCube results in a region of the decoherence parameter space which is already ruled out either by DeepCore or other experiments. Figure~\ref{fig:chi} shows that for both DeepCore and IceCube the best sensitivity is achieved for the solar limits (B) and (C) while the weakest limit is obtained in the atmospheric limit (A). In particular, the strongest bound is obtained for (C). This is in agreement with the behaviour of the oscillation probability in presence of strong matter effects, discussed in sec.~\ref{sec:plots-probabilities}. On one hand, as shown in sec.~\ref{sec:plots-probabilities}, for NO the decoherence effects are mainly driven by $\gamma_{21}$ in the neutrino channel and $\gamma_{32}$ in the antineutrino channel. On the other hand, the number of antineutrino events is going to be suppressed with respect to the neutrino case, due to the smaller cross section and flux. Hence, the best sensitivity is expected for case (C), where $\gamma_0 =\gamma_{21}^0=\gamma_{32}^0$, since both neutrinos and antineutrinos are sensitive to decoherence effects. Conversely, in case (B), where $\gamma_0 = \gamma_{21}^0=\gamma_{31}^0$, only neutrinos are sensitive to decoherence effects, and therefore some sensitivity is lost with respect to the results for case (C). Finally, in case (A), with $\gamma_0 = \gamma_{32}^0=\gamma_{31}^0$, the bounds come mainly from the impact of decoherence on the antineutrino event rates and, since these are much smaller than in the neutrino case, the obtained bounds are much weaker when compared to the results obtained in case (B). Figure~\ref{fig:chi} shows a flat asymptotic feature of the DeepCore's $\sqrt{\Delta\chi^2}$ for large values of $\gamma_0$, where the sensitivity becomes independent of $\gamma_0$. Conversely, for IceCube there is a decrease in sensitivity for values of $\gamma$ above a certain range: for example, for $n=0$ the best sensitivity is achieved for $\gamma_0 \sim \mathcal{O}(10^{-22})$~GeV while it decreases for higher values. This behaviour can be understood as follows. For the neutrino energies observed at IceCube (above 100~GeV) the oscillation phases do not develop and the probabilities do not depend on the energy ($\cos\tilde{\Delta}_{ij}\approx 1$ in eq.~(\ref{PmumuActe})). Therefore, at IceCube the sensitivity to the decoherence effects comes from the observation of a non-standard behaviour of the number of events with the zenith angle alone. Naively, eq.~(\ref{gammaL}) gives the values of $L$ and $\gamma$ that yield a large effect. Considering $n=0$, for example, where there is a one-to-one relation between the two, we get that for $\gamma_0 \sim 10^{-22}$~GeV the effect will be maximal for distances of the order $L\sim \mathcal{O}(10^{3})$~km. This is the typical distance traveled by atmospheric neutrinos crossing the Earth and therefore the sensitivity of IceCube is maximized in this range. Conversely, for larger (smaller) values of $\gamma_{0}$, only neutrinos coming from the most horizontal (vertical) directions are affected, leading to a reduced impact on the $\chi^2$. \begin{figure}[t!] \centering \includegraphics[width=0.98\columnwidth]{figures/BoundsHE+LE95CL.pdf} \caption{ 95\% CL bounds on the decoherence parameters $\gamma_0$, for NO, as a (discrete) function of the power-law index $n$ for the Atmospheric limit (red), Solar limit I (green) and Solar limit II (blue). The solid circles (triangles) correspond to the DeepCore (IceCube) analysis.} \label{fig:Bounds-vs-n} \end{figure} From the comparison between the different panels in fig.~\ref{fig:chi} we can see that the limits change considerably with the value of $n$, which parametrizes the energy dependence of the decoherence parameters (see eq.~\eqref{gammaE}). In particular, we observe in fig.~\ref{fig:chi} that the sensitivity improves with $n$ and that, as it is increased, the results for IceCube improve much faster (compared to DeepCore) due to the much higher neutrino energies considered in this case. The behaviour of the sensitivities with the value of $n$ is better appreciated in fig.~\ref{fig:Bounds-vs-n}, where we show the bounds obtained at 95\% CL (for 1 degree of freedom) as a (discrete) function of the power-law index $n$, for $n = -2, -1, 0, 1$ and 2. The DeepCore bounds are represented by solid circles while the IceCube constraints are given by the solid triangles. The results seem to follow the linear relation \begin{equation} \label{eq:gammavsn} \ln(\gamma_0 / \text{GeV})={\rm constant} - n \ln (E_0/\text{GeV}), \end{equation} where $E_0 \simeq$ 2.5 TeV (30 GeV) for IceCube (DeepCore). This can be understood as follows. Decoherence effects enter the oscillation probabilities only through the factor $\gamma L = \gamma_0 (E/\text{GeV})^n L$, for any value of $n$. Naively, we expect that the sensitivity limit is obtained for $\gamma L \sim \mathcal{O}(1)$ (although the precise value will eventually depend on the neutrino mass ordering, on the particular $\gamma_{ij}$ which drives the sensitivity, and on the data set considered). Taking the logarithm of $\gamma_0 (E/\text{GeV})^n L$ = constant, we reproduce eq.~(\ref{eq:gammavsn}). At first approximation, the value of $E_0$ in eq.~\eqref{eq:gammavsn} can be estimated as the average energy of the IceCube and DeepCore event distributions, $\langle E \rangle$, as \begin{equation} \langle E \rangle \equiv \frac{\displaystyle \int \frac{dN}{dE}E dE}{\displaystyle \int\frac{dN}{dE} dE}\ , \label{eq:average} \end{equation} where $dN/dE$ is the event number distribution. This leads to $\langle E \rangle\simeq 4$ TeV ($40$ GeV) for IceCube (DeepCore), which are in the right ballpark although somewhat different from the values of $E_0$ giving the best fit to the data shown in fig.~\ref{fig:Bounds-vs-n}. Nevertheless, we find these to be in reasonable agreement, given our naive estimation of $E_0$ as the mean energy for each experiment. \section{Summary and Conclusions} \label{sec:conclusions} \begin{table*}[ht!] \small \centering \tabcolsep=0.11cm \renewcommand{\arraystretch}{1.3} \begin{tabular}{| c | c | ccccc |} \hline \multirow{8}{*}{\begin{sideways} Normal Ordering \phantom{aaaaaaaa} \end{sideways}} & & $n=-2$ & $n=-1$ & $n=0$ & $n=1$ & $n=2$\cr \hline \hline & \textbf{IceCube (this work)} & & & & &\cr \cline{2-7} & Atmospheric ($\gamma_{31}=\gamma_{32}$) &$2.8\cdot10^{-18}$ & $4.2\cdot10^{-21}$ & $\mathbf{4.0\cdot10^{-24}}$ & $\mathbf{1.0\cdot10^{-27}}$ & $\mathbf{1.0\cdot10^{-31}}$\cr & Solar I ($\gamma_{31}=\gamma_{21}$) & $6.8\cdot10^{-19}$ & $1.2\cdot10^{-21}$ & $\mathbf{1.3\cdot10^{-24}}$ & $\mathbf{3.5\cdot10^{-28}}$& $\mathbf{1.9\cdot10^{-32}}$ \cr & Solar II ($\gamma_{32}=\gamma_{21}$) & $5.2\cdot10^{-19}$ & $9.2\cdot10^{-22}$ & $9.7\cdot10^{-25}$ & $2.4\cdot10^{-28}$ & $9.0\cdot10^{-33}$ \cr \cline{2-7} & \textbf{DeepCore (this work)} & & & & &\cr \cline{2-7} & Atmospheric ($\gamma_{31}=\gamma_{32}$) & $\mathbf{4.3\cdot10^{-20}}$ & $2.0\cdot10^{-21}$ & $8.2\cdot10^{-23}$ & $3.0\cdot10^{-24}$ & $1.1\cdot10^{-25}$\cr & Solar I ($\gamma_{31}=\gamma_{21}$) & $ \mathbf{1.2\cdot 10^{-20}}$ & $ 5.4\cdot10^{-22}$ & $ 2.1\cdot 10^{-23}$ & $6.6\cdot10^{-25}$& $2.0\cdot10^{-26}$ \cr & Solar II ($\gamma_{32}=\gamma_{21}$) & $7.5\cdot 10^{-21}$ & $3.5\cdot 10^{-22}$ & $ 1.4\cdot 10^{-23}$ & $4.2\cdot 10^{-25}$ & $ 1.1\cdot 10^{-26}$ \\ \hline \hline \multirow{7}{*}{\begin{sideways} Inverted Ordering \phantom{aaaaa} \end{sideways}} &\textbf{IceCube (this work)} & & & & & \cr \cline{2-7} & Atmospheric ($\gamma_{31}=\gamma_{32}$) & $6.8\cdot10^{-19}$ & $1.2\cdot10^{-21}$ & $\mathbf{1.3\cdot10^{-24}}$ & $\mathbf{3.5\cdot10^{-28}}$& $\mathbf{1.9\cdot10^{-32}}$ \cr & Solar I ($\gamma_{31}=\gamma_{21}$) & $5.2\cdot10^{-19}$ & $9.2\cdot10^{-22}$ & $9.8\cdot10^{-25}$ & $2.4\cdot10^{-28}$ & $9.0\cdot10^{-33}$ \cr & Solar II ($\gamma_{32}=\gamma_{21}$) &$2.8\cdot10^{-18}$ & $4.2\cdot10^{-21}$ & $\mathbf{4.1\cdot10^{-24}}$ & $\mathbf{1.0\cdot10^{-27}}$ & $\mathbf{1.0\cdot10^{-31}}$\cr \cline{2-7} &\textbf{DeepCore (this work)} & & & & &\cr \cline{2-7} & Atmospheric ($\gamma_{31}=\gamma_{32}$) & $\mathbf{1.4\cdot 10^{-20}}$ & $ 5.8\cdot10^{-22}$ & $ 2.2\cdot 10^{-23}$ & $7.5\cdot10^{-25}$& $2.3\cdot10^{-26}$ \cr & Solar I ($\gamma_{31}=\gamma_{21}$) & $8.3\cdot 10^{-21}$ & $3.6\cdot 10^{-22}$ & $ 1.4\cdot 10^{-23}$ & $4.7\cdot 10^{-25}$ & $ 1.3\cdot 10^{-26}$ \cr & Solar II ($\gamma_{32}=\gamma_{21}$) & $\mathbf{5.0\cdot10^{-20}}$ & $2.3\cdot10^{-21}$ & $9.4\cdot10^{-23}$ & $3.3\cdot10^{-24}$ & $1.2\cdot10^{-25}$\\ \hline \hline & \textbf{Previous Bounds} & & & & & \\ \hline & SK (two families)~\cite{Lisi:2000zt} & &$2.4\cdot 10^{-21}$ & $4.2\cdot10^{-23}$ & & $1.1\cdot10^{-27}$\\ & MINOS ($\gamma_{31},\gamma_{32}$)~\cite{deOliveira:2013dia} & & $\mathbf{2.5 \cdot 10^{-22}}$ &$1.1\cdot 10^{-22}$ & $2\cdot10^{-24}$ & \\ & KamLAND ($\gamma_{21}$)~\cite{Gomes:2016ixi} & & $\mathbf{3.7\cdot 10^{-24}}$ & $6.8\cdot 10^{-22}$ & $1.5\cdot10^{-19}$& \\ \hline \end{tabular} \caption{DeepCore/IceCube bounds on $\gamma_{ij}^0$ in GeV ($\gamma_{ij}=\gamma_{ij}^0(E/\text{GeV})^n$), at the $95\%$ CL (1 degree of freedom), for both normal and inverted ordering as indicated. Previous constraints are also provided for comparison, and the dominant limit in each case is highlighted in bold face (notice that we considered the most conservative bound from the two solar limits).} \label{tab:summary} \end{table*} In this work, we have derived strong limits on non-standard neutrino decoherence parameters in both the solar and atmospheric sectors from the analysis of IceCube and DeepCore atmospheric neutrino data. Our analysis includes matter effects in a consistent manner within a three-family oscillation framework, unlike most past literature on this topic. In sec.~\ref{sec:formalism} we have developed a general formalism, dividing the matter profile into layers of constant density, which permits to study decoherence effects in neutrino oscillations affected by matter effects in a non-adiabatic regime. Our analysis shows that the matter effects are extremely relevant for atmospheric neutrino oscillations and their importance in order to correctly interpret the two-family limits obtained previously in the literature, as outlined in sec.~\ref{sec:plots-probabilities}. We have found that the sensitivity to decoherence effects depends strongly on the neutrino mass ordering and on whether the sensitivity is dominated by the neutrino or antineutrino event rates. For neutrinos, the decoherence effects at high energies are mainly driven by $\gamma_{21}$ ($\gamma_{31}$) for normal (inverted) ordering, while in the antineutrino case they are essentially controlled by $\gamma_{32}$ ($\gamma_{21}$) for normal (inverted) ordering. This means that, considering a three-family framework including matter effects is essential when decoherence effects in atmospheric neutrino oscillations are studied. Our results are summarized in tab.~\ref{tab:summary}, together with the most relevant bounds present in the literature. Table~\ref{tab:summary} provides the 95\% CL bounds extracted from our analysis of DeepCore and IceCube atmospheric neutrino data, for both normal and inverted ordering, and for the three limiting cases considered in this work: (A) atmospheric limit ($\gamma_{21}=0$), (B) solar limit I ($\gamma_{32}=0$) and (C) solar limit II ($\gamma_{31}=0$). In~\ref{sec:d5results} we show that the bounds derived in these limits correspond to the most conservative results that can be extracted in the general case. In this work, we considered a general dependence of the decoherence parameters with the energy, as $\gamma_{ij}=\gamma_{ij}^0\left(E/\text{GeV}\right)^n$ with $n=\pm2,0,\pm2$. Our results improve over previous bounds for most of the cases studied, with the exception of the $n=-1$ case. For $n=-1$, KamLAND gives the dominant bound on $\gamma_{21}$ while MINOS gives the strongest constraints on $\gamma_{31}$ and $\gamma_{32}$\footnote{Reactor experiments as Double Chooz, Daya Bay or RENO are also expected to give a competitive bound in the atmospheric sector, as it was shown in~\cite{An:2016pvi} for Daya Bay in the standard decoherence case.} (indeed, both KamLAND and MINOS are also expected to give the strongest bound for $n=-2$, although to the best of our knowledge no analysis has been performed for this case yet). We have found that DeepCore considerably improves the present bounds in the solar sector ($\gamma_{21}$) for $n=0,1,2$ and gives a constraint in the atmospheric sector comparable to the SK limit, although a factor $2$ weaker, in the $n=0$ case. Our results show that, for $n=0$ (which is the case most commonly considered in the literature), IceCube improves the bound on $\gamma_{31}$ and $\gamma_{32}$ in (more than) one order of magnitude with respect to the SK constraint, obtained in a simplified two-family approximation, and by more than one order (almost two orders) of magnitude for NO (IO) with respect to the KamLAND constraint on $\gamma_{21}$. In particular, we find that the reference value for $\gamma_{23}$ considered in ref.~\cite{Coelho:2017zes} to explain the small tension previously reported between NOvA and SK data is indeed already excluded by IceCube data. Regarding the cases with $n = 1,2$, we find that the sensitivity of IceCube is particularly strong. For instance, IceCube improves the bound from KamLAND on $\gamma_{21}$ by almost 9 (8) orders of magnitude for $n=1$ and NO (IO), while for $n=2$ the bound on $\gamma_{31}$ and $\gamma_{32}$ is improved in 4 (5) orders of magnitude with respect to the SK limit for NO (IO). \begin{acknowledgements} We thank M.~C.~Gonz\'alez-Garc\'ia, M.~Maltoni and J.~Salvado for useful discussions. JLP, IMS and HN thank the hospitality of the Fermilab Theoretical Physics Department where this work was initiated. HN also thanks the hospitality of the CERN Theoretical Physics Department where the final part of this work was done. HN was supported by the Brazilian funding agency, CNPq (Conselho Nacional de Desenvolvimento Cient\i{\i}fico e Tecnol\'ogico), and by Fermilab Neutrino Physics Center. IMS acknowledges support from the Spanish grant FPA2015-65929-P (MINECO/FEDER, UE) and the Spanish Research Agency (``Agencia Estatal de Investigacion'') grants IFT ``Centro de Excelencia Severo Ochoa'' SEV2012-0249 and SEV-2016-0597. This work was partially supported by the European projects H2020-MSCA-ITN-2015-674896-ELUSIVES and 690575-InvisiblesPlus-H2020-MSCA-RISE-2015.This manuscript has been authored by Fermi Research Alliance, LLC under Contract No. DE-AC02-07CH11359 with the U.S. Department of Energy, Office of Science, Office of High Energy Physics. The publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes. \end{acknowledgements}

INTRODUCTION ============ The degree to which an amino (or nucleic) acid position is evolutionarily conserved is strongly dependent on its structural and functional importance. Thus, conservation analysis of positions among members from the same family can often reveal the importance of each position for the protein (or nucleic acid)'s structure or function. ConSurf ([@B1],[@B2]) and ConSeq ([@B3]) are web servers for calculating the evolutionary rate of each position of the protein and for identifying structurally and functionally important regions within proteins. The degree of conservation of each position is the inverse of the site's evolutionary rate; rapidly evolving positions are variable while slowly evolving positions are conserved. In ConSurf, the evolutionary rate is estimated based on the evolutionary relatedness between the protein and its homologues and considering the similarity between amino acids as reflected in the substitutions matrix ([@B4],[@B5]). One of the advantages of ConSurf in comparison to other methods is the accurate computation of the evolutionary rate by using either an empirical Bayesian method or a maximum likelihood (ML) method ([@B5]). The differences between the two methods are explained in detail in reference ([@B4]). The strength of those methods is that they explicitly account for the stochastic process underlying the evolution of the analyzed sequences, and that they rely on the phylogeny of the sequences. Thus, they can correctly discriminate between conservation due to short evolutionary time and genuine sequence conservation. In addition, the Bayesian based method provides reliability estimates for the site-specific conservation scores. METHODS ======= A short description of the methodology is provided below. More detailed description is available at <http://consurf.tau.ac.il/>, under 'OVERVIEW', 'QUICK HELP' and 'FAQ'. ConSurf protocol ---------------- A flowchart of the ConSurf web server is shown in [Figure 1](#F1){ref-type="fig"} and detailed below. The sequence is extracted from the 3D structure (if given).Homologous sequences are collected using a BLAST (or PSI-BLAST) ([@B6],[@B7]) search against a selected database. The user may specify criteria for defining homologues. The user can also manually select the desired sequences from the BLAST results.The sequences are clustered and highly similar sequences are removed using CD-HIT ([@B8]).A multiple sequence alignment (MSA) of the homologous sequences is constructed using MAFFT, PRANK, T-COFFEE, MUSCLE or CLUSTALW.A phylogenetic tree is reconstructed based on the MSA, using the neighbor-joining algorithm as implemented in the Rate4Site program ([@B4],[@B5]).Position-specific conservation scores are computed using the empirical Bayesian or ML algorithms ([@B4],[@B5]).The continuous conservation scores are divided into a discrete scale of nine grades for visualization, from the most variable positions (grade 1) colored turquoise, through intermediately conserved positions (grade 5) colored white, to the most conserved positions (grade 9) colored maroon.The conservation scores are projected onto the protein/nucleotide sequence and on the MSA. Figure 1.A flowchart of ConSurf protocol. Outputs ------- If a protein 3D structure is provided: The nine-color conservation scores are projected onto the 3D structure of the query protein and the colored protein structure is shown by FirstGlance in Jmol (<http://firstglance.jmol.org>).Scripts for visualizing the protein colored with ConSurf scores are generated for PyMol **(**<http://www.pymol.org>; [@B9]), Chimera ([@B10]), Jmol (<http://www.jmol.org/>; [@B11]) and RasMol ([@B12]). For all cases, ConSurf creates the following outputs: The sequence and MSA colored by ConSurf conservation scores.A text file that summarizes for each position the normalized score calculated, the assigned color, the reliability estimation (for the Bayesian method) and the amino acids/nucleotides observed in the respective MSA column.The sequences selected for the MSA and the MSA constructed (unless those files were uploaded by the user).A file with the frequency of each amino acid/nucleotide observed in each column of the MSA.The evolutionary tree, which was calculated by the server or uploaded by the user, is shown using an interactive Java applet written for that purpose. For proteins in which the 3D structure was not provided by the user, an up-to-date version of the Protein Data Bank ([@B13]) is searched for relevant homologues. If a structure of at least one homologous protein is available, the user may map the conservation scores on the structure. This option should ease the procedure for the non-expert users, who may be unfamiliar with the 3D structure homologue. This option can also be useful for analyzing proteins that share the same sequence but differ in their 3D structure (for example, two structures solved in different conformations or with different ligands). As an example we provide the main output of a ConSurf run for the N-terminal region of the GAL4 transcription factor in yeast (PDB ID: 3COQ, chain A and B) in complex with its DNA recognition site ([Figure 2](#F2){ref-type="fig"}). The analysis revealed, as expected, that the functional regions of this protein are highly conserved. For example, all the cysteines that form the Zn(2)-C6 DNA binding domain (CYS11, CYS14, CYS21, CYS28, CYS31, CYS38; [@B14]) were assigned the highest conservation scores. Likewise, PRO26, which is known to be central for DNA binding ([@B15]) is also highly conserved according to our analysis. In addition, other amino acid residues, which are in contact with the DNA (i.e. GLN9, LYS17, LYS18, LYS20, ARG15, LYS23; [@B16]) are relatively conserved. Figure 2.A ConSurf analysis for the GAL4 transcription factor and its DNA binding site. The 3D structure of the N-terminal region of the GAL4 transcription factor in yeast bound to the DNA is presented using a space-filled model. The amino-acids and the nucleotides are colored by their conservation grades using the color-coding bar, with turquoise-through-maroon indicating variable-through-conserved. Positions, for which the inferred conservation level was assigned with low confidence, are marked with light yellow. The figure reveals that the functionally important regions on both the DNA and the protein are highly conserved. The run was carried out using PDB code 3COQ and the figure was generated using the PyMol ([@B10]) script output by ConSurf. ConSurf was also applied to nucleic acid sequences from yeast, which are the known binding sites of GAL4 and their adjacent neighborhood ([Figure 2](#F2){ref-type="fig"}). As anticipated, the analysis revealed that the consensus pattern CGG-N~11~-CCG typical to GAL4 binding site is highly conserved. An extended full ConSurf analysis of this example is available in the 'GALLERY' section on the ConSurf web site. NEW ADDITIONS AND IMPROVEMENTS IN ConSurf 2010 ============================================== Analyzing nucleic acid sequences -------------------------------- Despite increasing interest in the non-coding fraction of transcriptomes, the number, the level of conservation, and functions, if any, of many non-protein-coding transcripts remain to be discovered. However, it has already been shown that many of the non-coding sequences are connected to regulatory processes. The new version of ConSurf offers estimations of the evolutionary rate for each position of nucleic acid sequences in the same manner used for amino acid residues. For that purpose, four evolutionary models were implemented in the Rate4Site program: (i) the Juke and Cantor 69 model (JC69), which assumes equal base frequencies and equal substitution rates ([@B17]). (ii) The Tamura 92 model that uses only one parameter, which captures variation in G-C content ([@B18]). (iii) The HKY85 model, which distinguishes between transitions and transversions and allows unequal base frequencies ([@B19]). (iv) The General Time Reversible (GTR) model, which is the most general time-reversible model. The GTR parameters consist of an equilibrium base frequency vector, giving the frequency at which each base occurs at each site, and the rate matrix ([@B20]). When enough data (i.e. sequences) are available, the GTR model is superior over the more simplified Tamura 92 model. However, the Tamura 92 model is recommended in cases in which the data are not sufficient for reliable estimation of the model parameters and thus it is the default option for analyzing nucleic acid sequences in ConSurf. Improved substitution matrix for protein sequences -------------------------------------------------- The LG substitution matrix, which incorporates variability of evolutionary rates across sites in the matrix estimation was shown to outperform other substitutions matrices for proteins ([@B21]). The LG matrix was added to Rate4Site and is offered in the new version of ConSurf in addition to the previous substitution models: JTT ([@B22]), Dayhoff ([@B23]), WAG ([@B24]), mtREV ([@B25]) and cpREV ([@B26]). Improved selection of homologous proteins ----------------------------------------- The accuracy of conservation scores is directly influenced by the amount and quality of sequence data available in the MSA and the relatedness between the homologous sequences themselves and the sequence of interest. For example, using homologous sequences with different functions might blur the signal. One of the important changes in the new version of ConSurf is the addition of a clear and intuitive interface that helps controlling which of the sequences are included in the analysis. These improvements include: A variety of sequence databases. The server offers the user the option to search for relevant sequences in several automatically updated sequence databases including: (i) SWISS-PROT (default) ([@B27]); (ii) A filtered version of the uniprot database ([@B28]); (iii) uniprot ([@B29]) (iv) UniRef90 in which redundant sequences were removed at level of 90% identity ([@B30]); (v) the NCBI non-redundant (nr) database.Manual selection of sequences for the analysis. After searching for homologous sequences, the user can manually select the relevant sequences to be included in the analysis using a simple form that provides all the relevant data for the sequences found and links to external web resources.Removing redundant sequences. The user can specify the level of redundant sequences for removal. The sequences found are clustered by their level of identity using CD-HIT ([@B8]) and the cutoff specified by the user (default level is 95% identity). Only one sequence (the longest) from each cluster is used for the analysis.Automatic removal of remote homologues. The user can control the level of sequence identity for which a hit sequence is still considered a homologue. Filtration according to the sequence identity between the sequences found and the sequence of interest enables the user to filter out sequences that share significant alignment with the protein of interest, however, might have different function or structure. The default level is set to 35% identity, which is the upper bound of the 'twilight zone' for protein structures ([@B31]).Better alignments. The user can choose to align the sequences using one of the following leading alignment algorithms: MAFFT ([@B32]), T-COFFEE (EXPRESSO mode) ([@B33]), PRANK ([@B34]) MUSCLE ([@B35]) and CLUSTALW ([@B36]). The EXPRESSO mode of T-COFFEE uses structural information (if available) and structural alignment methods to construct structure-based MSA. MAFFT and PRANK were shown to be among the leading sequence alignment algorithms ([@B34],[@B37]). MAFFT-LINSi is much faster than PRANK and thus was chosen to be the default alignment algorithm in ConSurf. Improved user interface ----------------------- In this new version of ConSurf, we put great emphasis on the user interface. ConSurf now presents an easier and more intuitive step-by-step interface, while still offering the user great flexibility during the process as described above. Each step is accompanied by built-in detailed help. IMPLEMENTATION ============== The new version of the ConSurf web server runs on a Linux cluster of 2.6GHz AMD Opteron processors, equipped with 4 GB RAM per quad-core node. The server runs with up to date versions of the supported MSA programs, and regularly updated databases. Running time depends on the dataset size (number and length of sequences) and the server load. The ConSurf server is implemented in PHP and Perl using the support of BioPerl modules ([@B38]). Rate4Site is implemented in C++ ([@B4]). For proteins with available 3D structure the conservation scores are projected on the structure and visualized using version 1.44 of FirstGlance in Jmol. CONCLUSIONS =========== ConSurf and ConSeq have an established reputation in the identification of functional regions in proteins using evolutionary information. In addition, these methods are a focal point that facilitates the development of more useful tools in our group and in other groups. For example, they are the basis for the development of the PatchFinder tool for the automatic detection of clusters of highly conserved amino acids ([@B39]), and the detection of DNA-binding proteins ([@B40]). Along with the massive growth of sequence and structure databases we believe that this new version of the ConSurf server will be highly useful to a growing number of molecular biology researchers and allow them to perform complex analyses using sophisticated algorithms accurately, easily and comprehensively. FUNDING ======= BLOOMNET ERA-PG; Israeli Science Foundation (878/09 to T.P.). Funding for open access charge: BLOOMNET ERA-PG. *Conflict of interest statement*. None declared. The authors are grateful to Nimrod Rubinstein, Adi Doron-Faigenboim, Eyal Privman, Itay Mayrose, Fabian Glaser, Maya Schushan, Guy Nimrod, Ofir Goldenberg, Yana Gofman, Uri Zonens, Gilad Wainreb and Matan Kalman for technical help, useful comments and helpful discussions. [^1]: The authors wish it to be known that, in their opinion the first two authors should be regarded as joint First Authors

Introduction {#s1} ============ *Entamoeba histolytica* (*Eh*), the causative agent of amebiasis, is an intestinal protozoan parasite that colonizes the intestinal lumen asymptomatically (known as noninvasive disease) in approximately 90% of cases. However, in 10% of individuals, this asymptomatic relationship breaks down and the parasite breaches the innate mucosal barrier and invades the underlying lamina propria resulting in 100,000 death per year globally (Stanley, [@B60]). Parasite cysts are acquired through the ingestion of contaminated food and water mostly in areas of poor sanitation (Ralston and Petri, [@B45]). Birth cohort studies done in an urban slum in Dhaka, Bangladesh found that approximately 50% of infants are infected in the first year of life, with repeated infections that are connected with malnourishment and stunting (Mondal et al., [@B37]; Korpe et al., [@B29]). The exact reasons why *Eh* occasionally invade the mucosal epithelium and what host-parasite factors are involved in parasite invasion are not clear. The outcome of invasive *Eh* infection is variable; it can result in amebic diarrhea, amebic colitis, and/or dissemination of the parasites through the portal circulation to cause liver abscess (Ralston and Petri, [@B46]). *Eh* invasion induces a robust pro-inflammatory response and host tissue destruction that exacerbates disease (Moonah et al., [@B38]; Mortimer et al., [@B41]). Currently, there is no effective vaccine for this disease; however nitroimidazoles (such as metronidazole) are effective to treat this tissue dwelling parasites. Metronidazole treatment can cause toxic side effects and enhance the probability of developing drug resistant (Petri, [@B42]; Ralston and Petri, [@B46]). Our host immune system sets up a series of defensive responses against the parasite. However, continued morbidity and mortality point out that this parasite is capable of escaping host defense responses to maintain its own survival (Moonah et al., [@B38]). Thus, an understanding of the human immune response to the parasite and the strategies used by the parasite to evade host defense will deeply improve the development of effective immunotherapies. In this review, we will focus on the host immune responses against *Eh* and the complex strategies the parasite uses to evade host immunity. Host immunity against *E. histolytica* {#s2} ====================================== Innate and adaptive immune response ----------------------------------- For any ingested parasite, immunity begins from the stomach acid as it has the ability to kill acid-sensitive microorganisms. However, the mucus barrier of the intestine serves as the first protective layer that prevents *Eh* from making contact with the underlying intestinal epithelial cells (IECs; Moonah et al., [@B38]). There are three major virulent factors of *Eh* known to mediate pathogenicity: (1) galactose/N-acetylgalactosamine-inhibitable lectin (Gal-lectin) is responsible for binding colonic mucin in colonization and host cell adhesion in disease pathogenesis, (2) pore forming peptide amoebapore used for host cell killing and, (3) cysteine proteases that lyse host extracellular matrix (Campos-Rodríguezp and Jarillo-Luna, [@B6]) and stimulate pro-inflammatory responses. The mucus barrier in the colon is composed of MUC2 mucin, which is a glycoprotein secreted from goblet cells. Mucin binds with high affinity to *Eh* Gal-lectin allowing the parasite to colonize the gut and at the same time it acts as a physical barrier to inhibit parasite invasion of the underlying epithelium (Chadee et al., [@B7]; Moonah et al., [@B38]). When *Eh* overcomes innate host defenses and contact IECs they produce a variety of pro-inflammatory mediators/chemokines including interleukin-1 (IL-1β), interleukin-8 (IL-8), and TNF-α. Several of these mediators trigger the recruitment of immune cells including neutrophils and macrophages to the site of parasite invasion (Yu and Chadee, [@B68]). The main amebicidal activity of neutrophils is the release of reactive oxygen species (ROS; Guerrant et al., [@B17]; Denis and Chadee, [@B10]). In macrophages *Eh* Gal-lectin up-regulates the mRNA expression of different cytokines. Gal-lectin induces pattern recognition receptor (PRRs) such as TLR-2 and TLR-4 mRNA expression in macrophages which is controlled by nuclear factor NF-κB and MAPK pathway. Recognition of parasite molecules by surface PRRs are crucial for the up-regulation of pro-inflammatory cytokine expression via NF-κB (Kammanadiminti et al., [@B27]). Macrophages activated with cytokines such as IFN-γ or TNF-α kills *Eh* trophozoites *in vitro* by producing nitric oxide (NO) from L-arginine mediated by inducible nitric oxide synthase (iNOS) (Lin et al., [@B35]). NO is critical for macrophage-mediated killing as iNOS deficient mice are more vulnerable to amebic liver abscess (ALA) and hepatocytic apoptosis (Seydel et al., [@B56]). Interferon gamma (IFN-γ) is the major cytokine that activates neutrophils and macrophages to exert amebicidal activity. Higher levels of IFN-γ are related to a lower incidence of *Eh* infection (Denis and Chadee, [@B10]; Ghadirian and Denis, [@B15]; Haque et al., [@B22]). Humoral immune responses against *Eh* are well characterized and it was been found that 81--100% of invasive amebiasis patients develop circulatory antibodies within 7 days of infection (Kaur et al., [@B28]). A prospective cohort study of pre-school children in Dhaka, Bangladesh, showed that mucosal IgA antibodies against the carbohydrate recognition domain (CRD) of the Gal-lectin heavy chain provided protection against *Eh* infection and disease (Haque et al., [@B20], [@B21]). In contrast, serum anti-lectin IgG was not associated with protection but mainly with the frequency of new infection (Haque et al., [@B20]). Higher levels of anti-lectin IgG was found in ALA and intestinal amebiasis as compared to asymptomatic patients (Kaur et al., [@B28]). These findings indicate that systematic anti-lectin antibodies are not involved in direct protection against amebiasis. Inflammasome activation by *E. histolytica* ------------------------------------------- *Eh* imposes damage through the adherence to host cells, which plays a critical role in killing or ingesting host target (Mortimer and Chadee, [@B39]). Thus, amebae adherence to host cell is one of the major characteristics of *Eh* pathogenicity. At present, the pattern recognition receptors (PRRs) that bind *Eh* Gal-lectin are not known. We recently identified that the inflammasome pathway is only activated on contact with live *Eh* and distinguishes between different physical forms of *Eh* (Mortimer et al., [@B40]). Inflammasome is a cytosolic multiprotein complex, which acts as a sensor for pathogens and cellular damage. This multimeric complex consists of an inflammasome sensor molecule (NOD-like receptor), the adaptor protein ASC and caspase-1. Activation of inflammasome leads to rapid and robust secretion of IL-1β, IL-18, IL-1α, FGF-2, IP-10 (Mortimer et al., [@B40]). Interestingly, when *Eh* activates the inflammasome, it does not trigger caspase-1 dependent cell death (known as pyroptosis) (Mortimer et al., [@B40]). It is unclear if inflammasome-activated macrophages are amebicidal and whether it plays other protective roles in amebic infection. On the other hand, if inflammasome activation triggers cell death in macrophages (pyroptotic cell death), *Eh* can use it as an advantage to limit immune elimination that can become detrimental to host defense. Immune evasion mechanisms of *E. histolytica* {#s3} ============================================= *Eh* has a two-phase life cycle: it can survive as an infective cyst in the environment or it can be found as trophozoites, the feeding and tissue dwelling stage in the human colon. After excystation in the colon, *Eh* trophozoites usually establishes harmless colonization where the parasites reside in the gut lumen and feed on enteric bacteria by phagocytosis (Voigt et al., [@B64]; Wilson et al., [@B67]). However, for unknown reasons trophozoites can become invasive, where parasite virulence factors allow it to degrade colonic mucin and other innate epithelial barrier functions (Wilson et al., [@B67]). Host immune responses, both innate and adaptive, are robust against invasive *Eh* but still this parasite is able to survive by developing immune evasion strategies. In particular, *Eh* cysteine proteases can cleave MUC2 mucin abrogating its protective functions allowing the parasite to breach the mucus layer and attach to the underlying epithelial cells (Lidell et al., [@B33]). Intestinal antimicrobial peptides are also an important component of host innate immune defense. Even though human LL-37 and murine CRAMP (cathelin-related antimicrobial peptide) cathelicidins are induced by *Eh* trophozoites both at the mRNA and protein level in IECs, *Eh* cysteine proteases can cleave these antimicrobial peptides (Cobo et al., [@B8]). Thus, *Eh* is resistant to both intact and cleaved antimicrobial cathelicidins in the intestine (Cobo et al., [@B8]). After amebic invasion, neutrophils are the earliest infiltrating cells but virulent *Eh* are effective in killing, lysing, and phagocytosing neutrophils. *In vitro*, one trophozoite was shown to kill approximately 3000 neutrophils (Guerrant et al., [@B17]; Guo et al., [@B18]). There are several conflicting mechanisms by which ameba interfere with neutrophil functions. *Eh* can disrupt NADPH oxidase activities and inhibit the respiratory burst of neutrophils to avoid oxidative stress. *Eh* iron-containing superoxide dismutase and NADPH:flavin oxidoreductase (Elnekave et al., [@B13]) are able to detoxify ROS by forming H~2~O~2~ (Bruchhaus et al., [@B5]; Sim et al., [@B59]). *Eh* trophozoites can protect themselves from neutrophil reactive oxygen properties with a 29-kDa surface protein, peroxiredoxin that has potent antioxidant activity (Davis et al., [@B9]). Studies have shown (Sim et al., [@B59]) that *Eh* can induce host cells (neutrophil) apoptosis through the activation of ERK1/2 by the generation of NADPH oxidase-derived ROS. Macrophages, another effector cell present during amebic infection also show suppressed cell mediated immunity due to *Eh*-induced strategic immune modulation. *Eh* trophozoites inhibit respiratory burst (ROS: H~2~O~2~, O^2−^, OH^−^) and NO production by macrophages (Lin et al., [@B34]; Wang et al., [@B66]). NOS substrate L-arginine is competitively converted to L-ornithine by ameba arginase that limit NO production by macrophages (Elnekave et al., [@B13]). When *Eh* and macrophages are exposed to each other, ameba produces the immunoregulatory molecule prostaglandin E2 (PGE~2~), synthesized by a cyclooxygenase (COX)-like enzyme by the parasite (Dey et al., [@B11]). Coupling through EP2/4 receptors, PGE~2~ increases cyclic adenosine monophosphate (cAMP) levels in macrophages that inhibits Th1 cytokine release, NADPH-mediated oxidative burst, and NO synthesis through the protein kinase C (PKC) pathway (Wang and Chadee, [@B65]). Another immunosuppressive pentapeptide, monocyte locomotion inhibitory factor (MLIF) produced by *Eh* showed anti-inflammatory activities by inhibiting NO production (Rico et al., [@B51]). The complement system of the host is able to prevent trophozoite dissemination into the extra intestinal space. Activated complement forms the membrane attack complex (MAC) that can potentially lyse the parasite. *Eh* resists complement activation by the Gal-lectin which have sequence resemblance and antigenic cross reactivity with the MAC-inhibitory protein CD59 and thus inhibit MAC-mediated lysis (Braga et al., [@B4]). Cysteine proteases can also cleave complement components (Reed et al., [@B50]). The potent pro-inflammatory activities of the complement component C3a and C5a are degraded by *Eh* secreted extracellular cysteine proteases (Zambrano-Villa et al., [@B69]). Secretory IgA and serum IgG mediate adaptive immunity against *Eh* and ameba can degrade these immunoglobulins *in vitro*. *Eh* extracellular cysteine proteases play a key role in the disruption of host adaptive defenses. For successful invasion, *Eh* secreted and membrane-bound cysteine proteases cleave extracellular matrix proteins, fibronectin, and laminin and avoid host defenses by cleaving gut lumen sIgA and circulatory IgG (Que and Reed, [@B43]; Zambrano-Villa et al., [@B69]). *E. histolytica*-induced cell death: Immune evasion strategy ------------------------------------------------------------ *Eh* uses different strategies to evade host immune defense but one striking mechanism is the induction of host cell death. The term "histolytica" refers to *Eh* ability to destroy host tissues by potent cytotoxicity/cell killing activity toward different host cells including neutrophils, macrophages, T-lymphocytes; though the exact mechanism of host cell killing is not clear. *Eh* has several cytotoxic effector molecules (Table [1](#T1){ref-type="table"}) but how ameba deliver this deadly action is not clear. *Eh* can induce host cell apoptosis, phagocytosis and amebic trogocytosis; the latter involving a recently described mechanism of ameba-induced host cell killing (Ralston et al., [@B47]). ###### **Potential cytotoxic effector molecules identified in ***Entamoeba histolytica*****. ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ***Eh* effectors molecules** **Identified effect on host cell** **References** ------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------- Amebapores Pore forming proteins Leippe et al., [@B31] All three types induce pore formation in synthetic liposomes Andrä et al., [@B1] Amebapore A Active at low pH 5.2. May play a role in host cell killing before ingestion Andrä et al., [@B1] Amebapore B Amebapore C *In vitro*, purified amebapores showed bactericidal activity against Gram-positive bacteria at nanomolar concentration\ Leippe et al., [@B30]; Andrä et al., [@B1] Purified amebapores are cytotoxic to Jurkat or U937 cells at micromolar concentration Cysteine proteinases (CPs) At least 50 CP genes are encoded and some of them are secretory Tillack et al., [@B62] Proteinases act on a variety of host substrates such as mucin, villin, laminin, collagen, proteoglycan, and extracellular matrix (ECM). It plays a role in pathogenesis by cleaving MUC mucin and ECM degradation Li et al., [@B32]; Lidell et al., [@B33] *In vivo*, overexpression of *Eh*CP5 (*Eh*CP5) increases liver abscess formation compared to wild-type controls. Other CPs like *Eh*CP1 or *Eh*CP2 overexpression had no effect Hellberg et al., [@B23]; Tillack et al., [@B63] Membrane proteins: Gal/GalNAc lectin associated 35kDa light subunit of the Gal/GalNAc lectin and surface-localized thiol-dependent peroxidase Ankri et al., [@B2]; Sen et al., [@B55] Antisense inhibition of both of these prevents cell killing and deceased liver abscess Ankri et al., [@B2]; Sen et al., [@B55] *Eh*STIRPs (*Eh* serine, threonine and isoleucine rice proteins) Silenced by dsRNA resulted in defects in both adhesion and cytotoxicity MacFarlane and Singh, [@B36] *Eh*TMKB1-9 (*Eh* transmembrane kinase B1-9) Antisense inhibition of this protein showed defects in both adhesion and host cell killing Shrimal et al., [@B58] KERP1 (Lysine and glutamic acid rice protein 1) Parasite membrane protein binds to host cell membrane Santi-Rocca et al., [@B53] Antisense inhibition of this protein failed to decrease mRNA but due to affinity for host cell membranes, it might have role in cytotoxicity ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ### Apoptotic cell death Host cell killing is usually a stepwise process mediated by parasite adherence to the target cell, elevation of intracellular calcium level, dephosphorylation of host proteins which all contribute to cell death via activation of caspase-3 (Ralston and Petri, [@B46]). Caspase-3 activation is the signature event of apoptosis. The first step of adherence is mediated by the parasite surface Gal-lectin to host cell carbohydrate determinants containing Gal and/or GalNAc residues (Figure [1A](#F1){ref-type="fig"}). Gal-lectin mediated adherence to target cell is a prerequisite for parasite cytotoxicity as the addition of excess Gal or GalNAc monomers inhibit *Eh* adherence and target cell killing (Ravdin and Guerrant, [@B48]; Saffer and Petri, [@B52]). Adherence to the target cell induces calcium flux which also contribute to the cell killing proved by the inhibition of calcium channel or by using calcium chelators (Ravdin et al., [@B49]; Ralston and Petri, [@B46]). Amebic cytotoxicity both *in vitro* and *in vivo* occurs via the caspase-3 dependent apoptotic cell death pathway in Jurkat cells after contact with *Eh* (Figures [1A,C](#F1){ref-type="fig"}) (Seydel and Stanley, [@B57]; Ralston and Petri, [@B46]). Caspase-3 knockout mice are resistant to amebiasis and a pharmacological inhibitor of caspase-3, Ac-DEVD-CHO (N-acetyl-Asp-Glu-Val-Asp-aldehyde) reduces parasite cytotoxicity to host cells (Huston et al., [@B25]; Becker et al., [@B3]). Studies illustrate that *Eh*-induced apoptosis is independent of caspase-8 or caspase-9 (Huston et al., [@B25]). Another study showed that *Eh* killing of hepatocyte and immune cell was not dependent or mediated by the classical Fas/Fas ligand or TNFα receptor pathway (Seydel and Stanley, [@B57]). Mice lacking Fas or producing a mutated non-functional Fas protein developed amebic ALA similar to wild type mice. Similarly, TNF receptor knockout mice and wild type mice develop comparable ALA (Seydel and Stanley, [@B57]). *Eh* induces a non-classical pathway of apoptosis that may have an important role in pathogenesis (Seydel and Stanley, [@B57]; Huston et al., [@B25]). Apoptotic cell death is immunologically silent, thus induction of host cell apoptosis by *Eh* tricks the host cell to kill itself without evoking an inflammatory response to avoid being detected by other immune cells. The importance of apoptosis in regards to amebic virulence is brought to light by studies of the leptin-signaling pathway. The hormone leptin is linked to malnutrition (signals satiety) and regulates the immune response to infection through the Th1 inflammatory response and by preventing apoptosis (Wilson et al., [@B67]). Leptin signaling provides protection from mucosal destruction and experiments in mice showed the anti-apoptotic role of leptin in gut epithelia (Guo et al., [@B19]). Polymorphism (even a single amino acid substitution) in the leptin receptor was found to be associated with increased *Eh* infection susceptibility (Duggal et al., [@B12]). {#F1} ### Phagocytosis by *E. histolytica* The fate of *Eh* induced apoptotic host cells (Huston et al., [@B24]) and subsequent phagocytosis may play an important role in the host-parasite relationship in amebiasis. Amebae induced phagocytosis of erythrocyte is one of the possible distinctive feature of *Eh* from the commensal *E. dispar* (González-Ruiz et al., [@B16]). In multicellular organisms, phagocytosis is the last step of the apoptotic pathway to clear dead cells without provoking an inflammatory response by the toxic components of dead cells (Figure [1D](#F1){ref-type="fig"}; Savill and Fadok, [@B54]; Fadok et al., [@B14]). Similarly, apoptotic host cells are phagocytosed by *Eh* with the help of exposed phosphatidylserine (Huston et al., [@B24]). As *Eh* rapidly clears apoptotically killed host cells by phagocytosis, this limits the spillage of toxic intracellular contents from killed cells. Through this strategy *Eh* restrains host inflammatory responses and build up prolonged infection. *Eh*-induced phagocytosis is important for pathogenicity but very little is known about the amebic receptors and the corresponding ligands that they bind to. Adherence with host cells by parasite Gal-lectin is critical for cell death but blocking of adherence does not prevent phagocytosis (Teixeira et al., [@B61]). This suggests the sequential exposure of new ligands on dying cells as well as the recruitment of new receptors on *Eh* in addition to the Gal-lectin following host cell killing (Teixeira et al., [@B61]). From an evolutionary concept, *Eh* phagocytoses bacteria for nutrient acquisition; therefore amebae preferentially recognize and phagocytose apoptotic cells that have surface similarities with bacteria (Teixeira et al., [@B61]). ### Ameba trogocytosis Recently, a new mechanism of *Eh* killing came to light. Using live cell imaging technology it was discovered that following host cell attachment, *Eh* trophozoites ingest separate parts (bites) of host cells which was termed "amebic trogocytosis" (Figure [1B](#F1){ref-type="fig"}; Ralston et al., [@B47]). This was a very rapid process as within 1 min of attachment amebic trogocytosis is initiated. Due to biting off and ingestion of separate pieces of host cells, intracellular calcium levels were elevated and this triggered ultimate cell death, evidenced by the loss of cell membrane integrity (Ralston et al., [@B47]). Interestingly, when trophozoites were incubated with either live or pre-killed host cells, only live cells were seen to be trogocytosed by ameba and pre-killed cells were ingested whole (phagocytosed; Ralston et al., [@B47]; Ralston, [@B44]). The cell surface characteristics of pre-killed host cells might be different from the directly killed cells and *Eh* might use this surface difference to determine the type of ingestion (Ralston, [@B44]). After ingestion of bites, amebae detach from the host cell, and the ingested cell eventually dies. It was speculated that this process contributes to amebic invasion in the colon as *Eh* can also trogocytose mucosal epithelial cells. Though amebic trogocytosis is a rapid process, this depends on specific conditions like physiological temperature, amoebic actin rearrangements, Gal-lectin, EhC2PK (*Eh* C2-domain-containing protein kinase), and PI3K (phosphoinositide 3-kinase) signaling (Ralston et al., [@B47]). Interference with any of these protein was shown to reduce *Eh* trogocytosis and subsequent decrease in host cell death (Ralston, [@B44]). It should be noted that trogocytosis also occurs in multicellular organism and different immune cell types but this trogocytosis does not trigger cell death. The exact reason for this distinction is not clear but trogocytosis in multicellular organisms mainly involves the exchange of cell membrane fragments; whereas *Eh* trogocytosis contains target cell cytoplasm and sometimes organelles (Joly and Hudrisier, [@B26]; Ralston et al., [@B47]). The concept of trogocytosis raises an important question, whether amebic trogocytosis is different from phagocytosis. The signaling proteins involved in amebic trogocytosis also play important roles during *Eh* phagocytosis but amebic trogocytosis is predominant in living cells (Ralston, [@B44]). Some factors might take part in the distinction of trogocytosis and phagocytosis like target cell deformability, target cell viability, target cell size etc. (Ralston, [@B44]). However, the specific signaling pathways that define amebic trogocytosis and phagocytosis in *Eh* are not identified; the relationship or differences between these two processes are not well understood. Conclusion {#s4} ========== *Eh* is an enteric dwelling protozoan parasite that causes significant morbidity and mortality in developing countries. This parasite can develop a harmless colonization in the colon and for unknown reasons it can become a pathogenic phenotype. With the pathogenic phenotype, *Eh* disrupts innate mucosal barriers and penetrates the underlying lamina propria where the parasite develops potent cytotoxic activity and extensive tissue destruction. It is still not clear how or what factors induce this pathogenic phenotype. Human and parasite genetics along with environmental factors might have a role as it has been found that not all children are equally susceptible to infection. Malnutrition extensively increases disease susceptibility. Gut microbiome also influence *Eh* infection susceptibility. Both host innate and adaptive immune response take part in the elimination of invasive *Eh*. The host immune system builds up a rapid inflammatory response by the secretion of cytokines/chemokines, recruitment of immune cells (neutrophils, macrophages), and the activation of inflammasome to control invasive parasites. This parasite also develops multiple strategies to subvert host immune responses and to promote its own survival. *Eh* induces host cell killing primarily by apoptosis, which is a non-inflammatory cell death mechanism. Induction of apoptotic cell death is an active and stepwise process. After inducing cell death, the parasite also clears the corpse by phagocytosis to inhibit further inflammatory responses. Another mechanism is amebic trogocytosis where *Eh* bites live cells very rapidly and induces cell death. The proper characterization of proteins, receptor/ligand interaction involved in parasite adherence, cell killing, phagocytosis, and amebic trogocytosis will provide promise of future vaccine candidates. Author contributions {#s5} ==================== SB, JQ, and KC wrote and edit the manuscript. Funding {#s6} ======= This work was supported by a grant from the Natural Sciences and Engineering Research Council of Canada (NSERC). SB and JQ is supported by scholarships from NSERC CREATE. Conflict of interest statement ------------------------------ The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. [^1]: Edited by: Alexandre Morrot, Federal University of Rio de Janeiro, Brazil [^2]: Reviewed by: Alberto A. Iglesias, Instituto de Agrobiotecnología del Litoral (UNL-CONICET), Argentina; Thomas Dandekar, University of Würzburg, Germany; Fabricio Montalvão Ferreira, Universidade Federal do Rio de Janeiro, Brazil [^3]: This article was submitted to Microbial Immunology, a section of the journal Frontiers in Microbiology

\section{Introduction } We study the solvability of the following fractional elliptic problem \begin{equation}\label{problem} \left\{ \begin{array}{rcll} (-\Delta)^su&=&|\nabla u|^q + \omega \quad \text{in }\mathbb{R}^n\\ u & > &0 \quad \text{in } \mathbb{R}^{n}\\ \lim_{|x|\to \infty}u(x) & =& 0, \end{array} \right. \end{equation}where $\frac{1}{2}<s<1$, $n>2s$ and $(-\Delta)^s$ is the classic fractional Laplacian operator of order $2s$. Here $\omega$ will be a non-negative Radon measure with compact support in $\mathbb{R}^n$. We consider the super-critical case $$q>p^*=\dfrac{n}{n-2s+1},\quad p+q=pq.$$This assumption on $q$ is motivated from the fact that in the local-case, as we shall detail below, no solutions exist for sub-critical $q$ unless $\omega \equiv 0$. For $W^{1, q}(\mathbb{R}^{n})$-solutions, a similar conclusion is obtained in our framework (we refer the reader to Remark \ref{rmk converse} for details). Moreover, the super-critical case allows us to obtain basic estimates on potentials as will be clarify in the proof of the main results. We highlight that non-local type operators arise naturally in continuum mechanics, image processing, crystal dislocation, Non-linear Dynamics (Geophysical Flows), phase transition phenomena, population dynamics, non-local optimal control and game theory (\cite{BCF}, \cite{BCF12}, \cite{BV}, \cite{C}, \cite{CV10}, \cite{CV}, \cite{DFV}, \cite{DPV}, \cite{GO}, \cite{L00} and the references therein). Indeed, models like \eqref{eq} may be understood as a Kardar-Parisi-Zhang stationary problem (models of growing interfaces) driving by fractional diffusion (see \cite{KPZ} for the model in the local setting and \cite{AyP} in the nonlocal stage). In the works \cite{MK} and \cite{MK1} the description of anomalous diffusion via fractional dynamics is investigated and various fractional partial differential equations are derived from L\'evy random walk models, extending Brownian motion models in a natural way. Finally, fractional type operators are also encompassed in mathematical modeling of financial markets, since L\'{e}vy type processes with jumps take place as more accurate models of stock pricing (cf. \cite{A04} and \cite{CT} for some illustrative examples). In this work we provide a sufficient condition for existence of global solutions to \eqref{eq} based in a relation between $\omega$ and a fractional capacity. We also derive a representation formula for the solution $u$ in terms of Riesz potentials and, as a result, we obtain pointwise and norm estimates and behaviour at infinity of $u$. Finally we demonstrate a necessary condition for the existence of solutions to the problem \eqref{problem}. We show that if problem \eqref{problem} has a positive $W^{1,q}(\mathbb{R}^n)$-solution, then $\omega$ does not charge sets of Riesz capacity zero, which means $\omega$ will necessarily be absolutely continuous with respect to the corresponding Riesz capacity. This result gives us a partial reciprocal of the main theorem. We refer the reader to the next section for further details. The current approach of the problem has been inspired by the enlightening results from \cite{Han}. In that paper, the authors provide a criteria for existence of solutions to equations of the form \begin{equation}\label{local} -\Delta u = |\nabla u|^{q} + \omega \end{equation}in $\mathbb{R}^{n}$. They also prove the same characterization in bounded domains, but in this case $q\geq 2$ is needed for sufficient conditions. The criteria of solvability is given explicitly in terms of pointwise behaviour of the corresponding Riesz potentials as well as in geometric capacitary terms. We also note that in a parallel theory of equations of the form $$-\Delta u=|u|^{q}+ \omega$$the role played by the Riesz capacity of order $(2, q')$ is analogous to the results in \cite{Han} (see for instance \cite{AP} and \cite{KV}). As a consequence of the results in \cite{Han} and the potential estimates from \cite{MV} (see also \cite[Section 7.2]{AH} and \cite[Section 11.5]{Ma}) it follows that no solution exists to \eqref{local} when $1 < q \leq n/(n-1)$ unless $\omega \equiv 0$. A different approach to characterize existence of solutions to local elliptic equations in terms of capacities was stated in \cite{BGO}. There, it was proved that a measure is absolutely continuous with respect to $(1, p)$ Riesz capacity if and only if it belongs to $L^{1}(\Omega)+ W^{-1, p'}(\Omega)$. This characterization allows to prove that a solution exists if and only if the measure data is absolutely continuous with respect to the Riesz capacity. Extensions of these results may be found, for instance, in \cite{BGO2}. We point out that the problem of existence of solutions to fractional elliptic problems with first-order terms like \eqref{eq} in bounded domains (with boundary data) has been considered in \cite{CV1}, \cite{CV} and \cite{AyP}. We highlight that our approach is also comparable to \cite[Section 5]{AyP} where a sufficient condition in terms of fractional capacity is also obtained for bounded domains and highly integrable sources. For fractional diffusion problems with non-homogeneous boundary conditions we refer to \cite{CH}. The paper is organized as follows. In Section \ref{preliminaries}, we provide the basic notation and definitions and also state the main result of the paper. Consequences of the main result, such as representation formula for the solution, pointwise and norm estimates and behaviour at infinite, are also provided. In Section \ref{proof} we give full details on the proof of Theorem \ref{Theo1}. Finally, Section \ref{converse sec} is dedicated to show the necessary condition for existence of solutions to problem \eqref{problem} and also discuss the converse of the main theorem which is currently an open problem. \section{Preliminaries and main results }\label{preliminaries} The Fourier definition of the fractional Laplace operator $(-\Delta)^s$ is given by \begin{equation}\label{fracLaplacianFou} (-\Delta)^su(x)=\mathcal{F}^{-1}(|\xi|^{2s}\mathcal{F}(u)(\xi))(x) \end{equation}$\xi\in\mathbb{R}^n$ and $u\in \mathcal{S}(\mathbb{R}^n)$, where $\mathcal{S}(\mathbb{R}^n)$ is the Schwartz class of smooth real-valued rapidly decreasing functions. We recall the following integral formulation of $(-\Delta)^{s}$ for functions in $\mathcal{S}(\mathbb{R}^{n})$ $$(-\Delta)^{s}u(x) :=a(n, s) \text{P. V.}\int_{\mathbb{R}^{n}}\frac{u(x)-u(y)}{|x-y|^{n+2s}}dy$$where $$a(n, s)=\frac{2^{2s}s\Gamma\left(\frac{n}{2}+s \right) }{\pi^{\frac{n}{2}}\Gamma(1-s)}$$is a normalization constant to recover \eqref{fracLaplacianFou}. We refer to \cite{Land} and \cite{S} for extensions to H\"{o}lder function spaces. Following the approach in \cite{S}, we consider the following spaces $$\mathcal{S}_s(\mathbb{R}^n):=\left\lbrace f\in \mathcal{C}^\infty(\mathbb{R}^n): \forall\beta\in\mathbb{N}_0^n, \sup_{x\in\mathbb{R}^n}(1+|x|^{n+2s})|D^\beta f(x)|<\infty\right\rbrace$$with the family of seminorms $$[f]_{\mathcal{S}_s(\mathbb{R}^n)}^\beta := \sup_{x\in\mathbb{R}^n}(1+|x|^{n+2s})|D^\beta f(x)|$$and the weighted Lebesgue space $$L_s(\mathbb{R}^n):=\left\lbrace u\in L^1_{\textnormal{loc}}(\mathbb{R}^n) : \int_{\mathbb{R}^n}\frac{|u(x)|}{1+|x|^{n+2s}}\;dx<\infty \right\rbrace.$$ We denote by $\mathcal{S}_s'$ the topological dual of $\mathcal{S}_s$. It is easy to see that $L_s(\mathbb{R}^n)\subset \mathcal{S}_s'(\mathbb{R}^n)$ and if $\varphi\in\mathcal{S}(\mathbb{R}^n)$, then $(-\Delta)^s\varphi\in \mathcal{S}_s(\mathbb{R}^n)$ (see \cite{Buc}). Then, for $u\in L_s(\mathbb{R}^n)$, we can define $(-\Delta)^su$ in sense of tempered distributions as $$\left\langle (-\Delta)^su, \varphi\right\rangle= \left\langle u,(-\Delta)^s\varphi\right\rangle= \int_{\mathbb{R}^n}u(x)(-\Delta)^s\varphi(x)\;dx$$for $\varphi\in \mathcal{S}(\mathbb{R}^n)$. (See \cite{Mar}). We shall consider weak solution of \eqref{eq} in the following sense. We start with the case without first-order terms. \begin{definition}\label{Defsolomega} Consider the equation \begin{equation}\label{eqomega} (-\Delta)^s u = \omega. \end{equation}Then, $u\in L_s(\mathbb{R}^n)$ is a weak (distributional) solution of \eqref{eqomega} if \begin{equation*} \int_{\mathbb{R}^n}u(x)(-\Delta)^s\varphi(x)\,dx = \int_{\mathbb{R}^n}\varphi(x)\,d\omega(x) \end{equation*}for all $\varphi\in \mathcal{S}(\mathbb{R}^n)$. If, instead of $\omega$ we have an integrable function $f$, we replace $d\omega(x)$ by $f(x)dx$. \end{definition} \begin{definition}\label{Defsol} We say that $u\in L_s(\mathbb{R}^n)\cap W_{\textnormal{loc}}^{1,q}(\mathbb{R}^n)$ is a weak solution of the equation \eqref{eq} if for all $\varphi\in S(\mathbb{R}^n)$, $|\nabla u|^{q}\varphi \in L^{1}_{\text{loc}}(\mathbb{R}^{n})$ and \begin{equation*} \int_{\mathbb{R}^n}u(x)(-\Delta)^s\varphi(x)\;dx = \int_{\mathbb{R}^n}|\nabla u(x)|^q\varphi(x)\;dx + \int_{\mathbb{R}^n}\varphi(y)\;d\omega(y). \end{equation*} \end{definition} Next we define the Riesz potential $\mathcal{I}_{\alpha}=(-\Delta)^{-\frac{\alpha}{2}}$ on $\mathbb{R}^n$ of order $\alpha$ for $0<\alpha<n$ as follows $$\mathcal{I}_{\alpha}(g)(x)=c(n,\alpha)\int_{\mathbb{R}^n}\frac{g(y)}{|x-y|^{n-\alpha}}\;dy$$for $g\in L^1_{\textnormal{loc}}(\mathbb{R}^n)$ such that $$\int_{|y|\geq 1}\frac{|g(y)|}{|y|^{n-\alpha}}\;dy<\infty.$$ The constant $c(n,\alpha)$ is defined by $$c(n,\alpha)=\pi^{-\frac{n}{2}}2^{-\alpha}\Gamma\left( \frac{n-\alpha}{2}\right) \Gamma\left( \frac{\alpha}{2}\right)^{-1}.$$For a non-negative Radon measure $\omega$, we define the Riesz potential $$\mathcal{I}_\alpha(\omega)(x)=c(n,\alpha)\int_{\mathbb{R}^n}\frac{1}{|x-y|^{n-\alpha}}\;d\omega(y).$$The kernel $$I_\alpha(x)=\frac{c(n,\alpha)}{|x|^{n-\alpha}}$$is called the Riesz kernel. In this way, we see that $$\mathcal{I}_\alpha (g)(x)=(I_\alpha*g)(x)$$and $$\mathcal{I}_\alpha(\omega)(x)=(I_\alpha*\omega)(x).$$ In order to state our main result, we first give the definition of Riesz capacity that will be employed along the work. \begin{definition} For $0<\alpha<n$ and $1<q<\infty$, the Riesz capacity $\textnormal{cap}_{\alpha,q}(E)$ of a measurable set $E\subset\mathbb{R}^n$ is defined by \begin{equation}\label{cap} \textnormal{cap}_{\alpha,q}(E)=\inf\left\lbrace \|u\|_{L^q(\mathbb{R}^n)}^q:\mathcal{I}_\alpha(u)\geq \mathcal{X}_E,\;u\in L^q_+(\mathbb{R}^n)\right\rbrace. \end{equation} \end{definition} The main result of this work is the following theorem. \begin{theorem}\label{Theo1} Let $\frac{1}{2}<s<1$, $q>p^*$. Suppose that there exists $C=C(n, q, s)> 0$, such that \begin{equation}\label{cap1} \omega(E)\leq C\textnormal{cap}_{2s-1,q'}(E) \end{equation}for all compact sets $E\subset\mathbb{R}^n$, then the equation \begin{equation}\label{eq} (-\Delta)^su=|\nabla u|^q + \omega \quad \text{in }\mathbb{R}^n \end{equation} has a non-negative weak solution $u\in W_{\textnormal{loc}}^{1,q}(\mathbb{R}^n)$. Moreover, $u$ has the representation \begin{equation}\label{represent} u = \mathcal{I}_{2s}(|\nabla u|^q+\omega) \end{equation} \end{theorem} \ We point out that \eqref{cap1} is equivalent, by Theorem 2.1 from \cite{MV}, to \begin{equation}\label{ineqpot} \mathcal{I}_{2s-1}\left( \left[ \mathcal{I}_{2s-1}(\omega)\right] ^q\right)\leq C_1 \mathcal{I}_{2s-1}(\omega), \quad \text{a. e. in }\mathbb{R}^{n} \end{equation}for some $C_1>0$ depending on $n$, $q$ and $s$. This relation, involving the non-lineal potential $\mathcal{I}_{2s-1}\left( \left[ \mathcal{I}_{2s-1}(\omega)\right] ^q\right)$, will be frequently used in the proof of Theorem \ref{Theo1}. The following result constitutes a partial converse of Theorem \ref{Theo1}. For the feasibility of the $W^{1, q}(\mathbb{R}^{n})$-regularity of solutions, we refer the reader to Proposition \ref{Coro}. \begin{theorem}\label{partial converse} Suppose that \eqref{eq} has a solution $u \in W^{1, q}(\mathbb{R}^{n})$. Then the measure $\omega$ does not charge sets of Riesz capacity zero. \end{theorem}\begin{remark}\label{rmk converse} In other words, Theorem \ref{partial converse} says that a necessary condition for existence of solutions is that $\omega$ is absolutely continuous with respect to the Riesz capacity. A direct consequence of this fact is that no global solution $u$ in $ W^{1, q}(\mathbb{R}^{n})$ exists if $q \leq p^{*}$ (see \cite[Propostition 2.6.1]{AH}).\end{remark} In the rest of this section, we shall provide some consequences of Theorem \ref{Theo1}. \begin{proposition}\label{Coro} The solution from Theorem \ref{Theo1} satisfies \begin{equation}\label{est 1 u} \mathcal{I}_{2s}(\omega) \leq u \leq C\mathcal{I}_{2s}(\omega) \end{equation} and \begin{equation}\label{est grad u} |\nabla u|\leq C\mathcal{I}_{2s-1}(\omega). \end{equation}As a result, if $q> n/(n-2s)$, we have $u \in W^{1, q}(\mathbb{R}^{n}).$ \end{proposition}\begin{proof} Observe that the estimate \eqref{est 1 u} is obtained as follows. From \eqref{estimate 2} we have \begin{equation}\label{estimate 23} u\leq C\left( \mathcal{I}_{2s}\omega + \mathcal{I}_{2s}\left[ \mathcal{I}_{2s-1}\omega\right]^q \right) \end{equation}Now, applying $\mathcal{I}_1$ in both sides of \eqref{ineqpot} yields $$\mathcal{I}_{2s}\left( \left[ \mathcal{I}_{2s-1}(\omega)\right] ^q\right)\leq C \mathcal{I}_{2s}(\omega)$$Plugging this inequality into \eqref{estimate 23} gives \eqref{est 1 u}. The lower bound in \eqref{est 1 u} is a consequence of \eqref{represent}. The estimate for the gradient \eqref{est grad u} follows from \eqref{to represent}. Next, we prove the final assertion. Observe $\mathcal{I}_{2s}(\omega)\in L^{q}(\mathbb{R}^{n})$ for $q> n/(n-2s)$. Indeed, for any $R>0$ so that $\textnormal{supp}(\omega) \subset B_R$ we have \begin{equation}\label{for the pf} \int_{B_R}\mathcal{I}_{2s}(\omega)^{q}dx \leq C(n,s,q,R)\int_{B_R}\mathcal{I}_{2s-1}(\omega)^{q}dx < \infty \qquad \text{by Lemma } \ref{regriesz}, \end{equation} and, on the other hand, for large $R$ and $q> n/(n-2s)$ it follows by Minkowski's inequality \begin{equation*} \begin{split} \int_{\mathbb{R}^{n}\setminus B_R}\mathcal{I}_{2s}(\omega)^{q}dx &\leq C\left( \int_{\text{supp }\omega}\left[\int_{\mathbb{R}^{n}\setminus B_R}\frac{dx}{|x-y|^{q(n-2s)}}\right]^{1/q}d\omega(y)\right) ^q\\ & \leq C\left( \int_{\text{supp }\omega}\left[\int_{\mathbb{R}^{n}\setminus B_R}\frac{dx}{|x|^{q(n-2s)}}\right]^{1/q}d\omega(y)\right) ^q< \infty \end{split} \end{equation*} Consequently by \eqref{est 1 u} $u \in L^{q}(\mathbb{R}^{n})$. Moreover, by Remark \ref{regI2s-1w} and \eqref{est grad u}, it follows that $\nabla u \in L^{q}(\mathbb{R}^{n})$. Hence $u \in W^{1, q}(\mathbb{R}^{n}).$\end{proof} The next corollary shows that $u$ from \eqref{represent} is actually a solution of \eqref{problem}. \begin{corollary}We have the following for the solution $u$ from Theorem \ref{Theo1}: \begin{enumerate} \item[(i)] $u$ vanishes at infinite $$\lim_{|x|\to \infty}u(x)=0;$$ \item[(ii)] $u$ is positive everywhere. \end{enumerate} \end{corollary} \begin{proof} Take $R >0$ so that $\textnormal{supp}(\omega) \subset B_R.$ Hence for $|x| > R$ we have by \eqref{est 1 u} $$u(x) \leq C \mathcal{I}_{2s}(\omega) = C\int_{B_R}\frac{d \omega(y)}{|x-y|^{n-2s}} \leq C\frac{\omega(B_R)}{\left( |x|-R\right)^{n-2s}}.$$This proves $(i)$. For $(ii)$, observe that for $x \in \mathbb{R}^{n}$ and $r>0$ so that $\textnormal{supp}(\omega) \subset B_r(x)$ the lower bound in \eqref{est 1 u} implies $$u(x) \geq c(n, 2s)\int_{B_r(x)} \frac{d\omega(y)}{|x-y|^{n-2s}}\geq c(n, 2s)r^{2s-n}\omega(\textnormal{supp}(\omega))>0.$$ \end{proof} Observe that from Theorem \ref{Theo1}, $u \in W^{1, q}_{\text{loc}}(\mathbb{R}^{n})$ and thus $u$ is finite almost everywhere. The following consequence of Proposition \ref{Coro} gives finiteness everywhere for $q > 2s$.\begin{corollary}Suppose that $q> 2s$. Then, $u < \infty$ everywhere in $\mathbb{R}^{n}$. \end{corollary}\begin{proof} By \eqref{est 1 u} and the Cavalieri's representation of Riesz potentials (see for instance Section 2.2 in \cite{Ponce}), we have for $x \in \mathbb{R}^{n}$ \begin{equation}\label{finiteness u} u(x)\leq C \mathcal{I}_{2s}(\omega)(x) = C \int_0^{\infty}\frac{\omega(B_r(x))}{r^{n-2s+1}}dr= C \int_0^{R(x)}\frac{\omega(B_r(x))}{r^{n-2s+1}}dr, \end{equation}where $R(x)>0$ satisfies supp $\omega \subset B(x, R(x))$. We now prove that \begin{equation}\label{ineqomegacap} \omega(B_r(x))\leq C_2(n,s,q)r^{n-(2s-1) p}. \end{equation}For $0 <\alpha < n$, take $$g=\frac{2^{n-\alpha}}{c(n,\alpha)\omega_n r^\alpha}\mathcal{X}_{B_r(x)}\in L^p_+(\mathbb{R}^n).$$Now, let $z\in B_r(x)$, then \begin{equation*} \begin{split} \mathcal{I}_\alpha g(z)&=c(n,\alpha)\int_{\mathbb{R}^n}\frac{1}{|z-y|^{n-\alpha}}\frac{2^{n-\alpha}}{c(n,\alpha)\omega_n r^\alpha}\mathcal{X}_{B_r(x)}(y)\;dy \\& = \frac{2^{n-\alpha}}{\omega_n r^\alpha}\int_{B_r(x)}\frac{1}{|z-y|^{n-\alpha}}\;dy\\& = 1. \end{split} \end{equation*}Hence $\mathcal{I}_\alpha g\geq \mathcal{X}_{B_r(x)}$. Therefore, \begin{equation*} \begin{split} \textnormal{cap}_{\alpha,p}(B_r(x))&\leq \|g\|_{L^p(\mathbb{R}^n)}^p\\& = \int_{\mathbb{R}^n}\left( \frac{2^{n-\alpha}}{c(n,\alpha)\omega_n r^\alpha}\right)^p\mathcal{X}_B(y)^p\;dy \\ & = \left( \frac{2^{n-\alpha}}{c(n,\alpha)}\right)^p \omega_n^{1-p}r^{n-\alpha p}. \end{split} \end{equation*}Thus \begin{equation*} \textnormal{cap}_{\alpha,p}(B)\leq Cr^{n-\alpha p}, \quad C=\left( \frac{2^{n-\alpha}}{c(n,\alpha)}\right)^p \omega_n^{1-p}. \end{equation*}Letting $\alpha = 2s$ and recalling \eqref{cap1}, we obtain \eqref{ineqomegacap}. Plugging this into \eqref{finiteness u} gives that $u(x)$ is finite (observe that $q > 2s$ implies $p \in (1, 2s/(2s-1)$ and so the last integral in \eqref{finiteness u} is finite). \end{proof} The following results show how the regularity of the source is transferred into the regularity of the solution. We start with measure data and then we analyse the regularity of solutions for integrable sources. To state estimates of the solution in terms of measures, we first recall the definition of the Marcinkiewicz spaces. \begin{definition}Let $\Omega \subset \mathbb{R}^{n}$ be a domain and $\mu$ be a positive Borel measure in $\Omega$. For $\kappa > 1$, $\kappa' = \kappa /(\kappa -1)$ we define the Marcinkiewicz space $M^{\kappa}(\Omega, d \mu)$ of exponent $\kappa$ or weak $L^{\kappa}$-space, as $$M^{\kappa}(\Omega, d \mu):= \left\lbrace v \in L^{1}_{\text{loc}}(\Omega, d \mu): \|v\|_{M^{\kappa}(\Omega, d \mu)}< \infty \right\rbrace$$where $$\|v\|_{M^{\kappa}(\Omega, d \mu)}:=\inf\left\lbrace c \in [0, \infty]: \int_E|v|d \mu \leq c \left(\int_E d\mu \right)^{\frac{1}{\kappa'}}, \text{ for all Borel }E \subset \Omega \right\rbrace.$$ \end{definition}The Marcinkiewicz type estimate of the solution $u$ is the following (recall $p^{*}=n/(n-2s+1)$). \begin{proposition}There exists $C=C(n, s)>0$ so that \begin{equation} \|u\|_{M^{n/(n-2s)}(\mathbb{R}^{n}, d\mu)}+ \|\nabla u\|_{M^{p^{*}}(\mathbb{R}^{n}, d\mu)} \leq C \|\omega\|_{\mathcal{M}^{b}(\mathbb{R}^{n})}, \end{equation}where $$d\mu(x):= \dfrac{dx}{1+|x|^{n+2s}}$$and $$\|\omega\|_{\mathcal{M}^{b}(\mathbb{R}^{n})}= \omega(\mathbb{R}^{n})$$is a norm in the space of bounded Radon measures $\mathcal{M}^{b}(\mathbb{R}^{n})$ of $\mathbb{R}^{n}$. \end{proposition} \begin{proof}We follow closely the proof of \cite[Proposition 2.2]{CV1}. We prove the estimates for $\nabla u$. By similar arguments, the control for $\|u\|_{M^{n/(n-2s)}(\mathbb{R}^{n}, d\mu)}$ is obtained. Observe that, in view of \eqref{est grad u}, it is enough to prove that there is $C>0$ with \begin{equation}\label{est riesz 2s-1} \|\mathcal{I}_{2s-1}(\omega)\|_{M^{n/(n-2s+1)}(\mathbb{R}^{n}, d\mu)} \leq C \|\omega\|_{\mathcal{M}^{b}(\mathbb{R}^{n})}. \end{equation}For $y \in \mathbb{R}^{n}$ and $\lambda >0$, define $$A_{\lambda}(y):=\left\lbrace x\in \mathbb{R}^n\setminus \left\lbrace y \right\rbrace: I_{2s-1}(x-y)=\frac{c(n, 2s)}{|x-y|^{n-2s+1}} > \lambda \right\rbrace, \quad m_\lambda(y):=\int_{A_\lambda(y)}\frac{1}{1+|x|^{n+2s}}dx.$$Since $$A_\lambda \subset B_r(y),$$with $$r=\left[\dfrac{c(n, 2s)}{\lambda}\right]^{1/(n-2s+1)},$$ we have for some $C>0$ $$m_\lambda(y) \leq C \lambda^{-p^{*}}.$$Let now $E$ be a Borel set. Then $$\int_E I_{2s-1}(x-y)d\mu(x) \leq \lambda \int_E d \mu(x)+ \int_{A_\lambda(y)}I_{2s-1}(x-y)d\mu(x)$$and $$\int_{A_\lambda(y)}I_{2s-1}(x-y)d\mu(x) = \lambda m_\lambda(y)+ \int_\lambda^{\infty}m_s(y)ds \leq C \lambda^{1-p^{*}},$$for some $C>0$. Hence $$\int_E I_{2s-1}(x-y)d\mu(x) \leq \lambda \int_E d \mu(x)+ C \lambda^{1-p^{*}}.$$Choosing $\lambda = (\int_E d \mu)^{-1/p^{*}}$, we obtain $$\int_E I_{2s-1}(x-y)d\mu(x) \leq C\left(\int_E d\mu \right)^{\frac{p^{*}-1}{p^{*}}},$$for a universal constant $C>0$ and all $y \in \mathbb{R}^{n}.$ As a result $$\int_E \mathcal{I}_{2s-1}(\omega)(x)d\mu(x) \leq C \|\omega \|_{\mathcal{M}^{b}(\mathbb{R}^{n})}\left( \int_E d\mu\right)^{\frac{p^{*}-1}{p^{*}}}.$$This ends the proof of the proposition. \end{proof} In the next proposition, we provide a complete scheme of Lebesgue, Sobolev and H\"{o}lder regularity of $u$ in terms of the regularity of the source. \begin{proposition} Let $\omega= f dx$, where $f\in L^{m}_{+}(\mathbb{R}^{n})$ vanishes outside a compact set. Then for the solution $u$ from Theorem \ref{Theo1} we get: \begin{itemize} \item[(i)] if $m=1$, then there is $C=C(s, n)>0$ so that for any $\lambda> 0$ $$|\left\lbrace x \in \mathbb{R}^{n}: u(x) > \lambda\right\rbrace| \leq C\left(\dfrac{\|f\|_{L^{1}(\mathbb{R}^{n})}}{\lambda}\right)^{m*}, \qquad m^{*}= 1-\frac{2s}{n};$$ \item[(ii)] if $m >1 $ and $2sm < n$, there is a constant $C=C(r, s,n, m)>0$ so that $$\|u\|_{L^{r}(\mathbb{R}^{n})} \leq C\|f\|_{L^{m}(\mathbb{R}^{n})}, \quad \text{ for all }r \in \left( \frac{n}{n-2s}, m^{*}\right]$$where $$m^{*}= \frac{nm}{n-2sm};$$\item[(iii)] for $m >1 $ and $(2s-1)m < n$, for some $C=C(r, s,n, m)>0$ we have $$\|\nabla u\|_{L^{r}(\mathbb{R}^{n})} \leq C\|f\|_{L^{m}(\mathbb{R}^{n})}, \quad \text{ for all }r \in \left( \frac{n}{n-2s+1}, m^{*}\right]$$where $$m^{*}= \frac{nm}{n-(2s-1)m};$$\item[(iv)] if $m=n/(2s)$, then $$u \in L^{r}(\mathbb{R}^{n})\quad \text{ for all }r \in \left( \frac{n}{n-2s}, \infty\right);$$ \item[(v)]in the case $m=n/(2s-1)$, $$u \in W^{1, r}(\mathbb{R}^{n})\quad \text{ for all }r \in \left( \frac{n}{n-2s+1}, \infty\right)$$ \item[(vi)] if $m > n/(2s-1)$, then $$\nabla u \in \mathcal{C}^{0, \gamma}(\mathbb{R}^{n}), \quad\gamma = 2s-1 - \frac{n}{m}.$$ \end{itemize} \end{proposition}\begin{proof} The estimates $(i)-(iii)$ are direct consequences of Proposition \ref{Coro} and well-known $L^{p}$ embeddings of the Riesz potential (see for instance \cite[Chapter V]{St} and \cite[Chapter 3]{AH}). Also, $(iv)$ and $(v)$ follow from $(ii)$ and $(iii)$, respectively, together with the assumption that $f$ has compact support. Finally, for $(vi)$ we first observe that from \eqref{represent} $$\nabla u = \mathcal{I}_{2s-1}(|\nabla u|^{q}+ f) \quad a.e. \,\,\mathbb{R}^{n}.$$Moreover, $\nabla u \in L^{qm}(\mathbb{R}^{n})$ since by $(v)$, $u \in W^{1, r}(\mathbb{R}^{n})$ for all $r > p^{*}$ and $qm > p^{*}$. Therefore, appealing to \cite[Theorem 2.2, Sec. 4.2]{Mi}, there is a constant $M>0$ so that $$|\nabla u(x)-\nabla u(y)| \leq M|x-y|^{2s-1 - n/m}\||\nabla u|^{q}+ f\|_{L^{m}(\mathbb{R}^{n})}\,\,\,\, a.e. \text{ in }\mathbb{R}^{n}.$$ \end{proof} By \cite[Theorem 3.2.1]{AH}, we also may obtain the following exponential summability for $u$ and $\nabla u$ which account, for instance, to local integrability: \begin{itemize} \item if $m=n/2s$ and supp $f \subset B_R$, there is a constant $A=A(n, m)$ such that $$\int_{B_R}\exp\left(A_0 u^{m'}\right) dx \leq A R^{n}, \quad A_0=\frac{n}{ c(n, 2s)^{m'}\omega_{n-1}};$$ \item in the case $m=n/(2s-1)$ and supp $f \subset B_R$, we have for the same $A$ as before that $$\int_{B_R}\exp\left(A_0 |\nabla u|^{m'}\right) dx \leq A R^{n}, \quad A_0=\frac{n}{ c(n, 2s-1)^{m'}\omega_{n-1}}.$$ \end{itemize} \begin{remark}The interested reader may compare the above regularity results to the related findings for bounded domains and no first-order terms presented in \cite[Theorem 15-16]{LP} and \cite[Lemma 2.15]{AyP}. \end{remark} \section{Proof of Theorem \ref{Theo1}}\label{proof} The structure of the proof consists of the following steps \begin{itemize} \item[(I)] the starting point will be to consider $u_0=\mathcal{I}_{2s}(\omega)$ and prove that $$(-\Delta)^{s}u_0=\omega$$in the sense of Definition \ref{Defsolomega}. This will be done in Lemma \ref{regI2sw} and Proposition \ref{SolRiesz}. \item[(II)] The next step is to consider first-order terms. Indeed, we show in Lemmas \ref{ineqgrad}-\ref{regriesz}, and Propositions \ref{propp}-\ref{solu0} that $v= \mathcal{I}_{2s}(|\nabla u_0|^{q})$ solves $$(-\Delta)^{s}v=|\nabla u_0|^{q}.$$ \item[(III)] The final step, developed at the end of the section, is to define by recursion the sequence $$u_{k+1}= \mathcal{I}_{2s}(\omega)+ \mathcal{I}_{2s}(|\nabla u_k|^{q}), \quad k\geq 0,$$and prove that $u_k$ converges in the right topology to a solution of \eqref{eq}. \end{itemize} \subsection{Step (I)} \begin{lemma}\label{regI2sw} Let $\omega$ be a nonnegative Radon measure with compact support in $\mathbb{R}^n$. Then, for $n>2s$, $\mathcal{I}_{2s}(\omega)\in L_s(\mathbb{R}^n).$ \end{lemma} \begin{proof} First note that $\mathcal{I}_{2s}(\omega)\in L^1_{\textnormal{loc}}(\mathbb{R}^n)$. Let $R>0$ such that $\textnormal{supp}(\omega)\subset B_R$. For simplicity take $R=1$. Then, \begin{equation*} \begin{split} \int_{\mathbb{R}^n}\frac{|\mathcal{I}_{2s}(\omega)(x)|}{1+|x|^{n+2s}}\; dx & = \int_{B_1}\left( \int_{\mathbb{R}^n}\frac{c(n,2s)}{(1+|x|^{n+2s})|x-y|^{n-2s}}\;dx\right)d\omega(y). \end{split} \end{equation*}Now we split the last integral as follows (we omit the constant for simplicity) \begin{equation*} \begin{split} \int_{B_1}\left( \int_{\mathbb{R}^n}\frac{1}{(1+|x|^{n+2s})|x-y|^{n-2s}}\;dx\right)d\omega(y) &=\int_{B_1}\left( \int_{B_2}\frac{1}{(1+|x|^{n+2s})|x-y|^{n-2s}}\;dx\right)d\omega(y) \\&+ \int_{B_1}\left( \int_{\mathbb{R}^n\setminus B_2}\frac{1}{(1+|x|^{n+2s})|x-y|^{n-2s}}\;dx\right)d\omega(y). \end{split} \end{equation*}Then, for the first integral we have \begin{equation*} \begin{split} \int_{B_1}\left(\int_{B_2}\frac{1}{(1+|x|^{n+2s})|x-y|^{n-2s}}\;dx\right)d\omega(y)&\leq \int_{B_1}\left( \int_{B_2}\frac{1}{|x-y|^{n-2s}}\;dx\right)d\omega(y)\\& \leq \int_{B_1}\left( \int_{B_{2+|y|}}|z|^{2s-n}\;dz\right)d\omega(y)\\ & = \omega_n \int_{B_1}\frac{(2+|y|)^{2s}}{2s}\;d\omega(y) \\ & <\infty. \end{split} \end{equation*}For the second integral we have \begin{equation*} \begin{split} \int_{B_1}\left( \int_{\mathbb{R}^n\setminus B_2}\frac{1}{(1+|x|^{n+2s})|x-y|^{n-2s}}\,dx\right)d\omega(y)& \leq \int_{B_1}\left(\int_{\mathbb{R}^n\setminus B_2}(|x|-|y|)^{2s-n}|x|^{-n-2s}\,dx \right)d\omega(y)\\& <\infty \end{split} \end{equation*} \end{proof} \begin{remark}\label{regI2s-1w} \rm{Reproducing the proof of Lemma \ref{regI2sw} with $2s-1$ instead of $2s$, one obtains $\mathcal{I}_{2s-1}(\omega) \in L_s(\mathbb{R}^{n})$ as well.} \end{remark} For the proof of Proposition \ref{SolRiesz} is necessary the following lemma (Proposition 2.4 of \cite{Buc}). \begin{lemma}\label{LemmaBucur} Let $n>2s$ and $f\in L^1(\mathbb{R}^n)\cap \mathcal{C}(\mathbb{R}^n)$ with $\mathcal{F}^{-1}(f)\in S_s(\mathbb{R}^n)$. Then \begin{equation}\label{eqprop2.4} \int_{\mathbb{R}^n}I_{2s}(x)\mathcal{F}^{-1}(f)(x)\;dx=\int_{\mathbb{R}^n}|x|^{-2s}f(x)\;dx. \end{equation} \end{lemma} \begin{remark}\label{rm Buc} A closer look at the proof of \cite[Proposition 2.4]{Buc} reveals that the assumption $\mathcal{F}^{-1}(f)\in S_s(\mathbb{R}^n)$ may be replaced by \begin{enumerate}[i.] \item $\int_{\mathbb{R}^n}I_{2s}(x)\mathcal{F}^{-1}(f)(x)\,dx<\infty$ \rm{and} \item $\mathcal{F}^{-1}(f)\in L^2(\mathbb{R}^n)$. \end{enumerate}We will use this observation in Proposition \ref{solu0}. \end{remark} \bigskip \begin{proposition}\label{SolRiesz} For $n>2s$, $\mathcal{I}_{2s}(\omega)$ is a weak solution of \begin{equation}\label{eq2} (-\Delta)^s u = \omega \quad \text{in }\mathbb{R}^n. \end{equation} \end{proposition} \begin{proof} We will prove that Definition \ref{Defsolomega} is satisfied. We already know that $\mathcal{I}_{2s}(\omega)\in L_s(\mathbb{R}^n)$ by Lemma \ref{regI2sw}. Let $\varphi_0\in\mathcal{S}(\mathbb{R}^n)$. Put $$\varphi(x)=|x|^{2s}\mathcal{F}(\varphi_0)(x)$$ and take $\psi$ such that $$\varphi(x)=\mathcal{F}(\psi)(-x).$$Then, $\psi(x)=\mathcal{F}^{-1}(\varphi)(-x)$. Thus, since $\varphi_0\in \mathcal{S}(\mathbb{R}^n)$, $\mathcal{F}^{-1}(\varphi)\in \mathcal{S}_s(\mathbb{R}^n)$ and therefore $\psi\in \mathcal{S}_s(\mathbb{R}^n)$. In what follows, we shall employ Lemma \ref{LemmaBucur}. We first prove that $\psi*\omega\in\mathcal{S}_s(\mathbb{R}^n)$. Since $D^\alpha \psi\in\mathcal{S}_s(\mathbb{R}^n)$ and $D^\alpha(\psi*\omega)=(D^\alpha \psi)*\omega$ we just need to verify that \begin{equation}\label{cotapsiconvomega} |\psi*\omega(x)|\leq C(1+|x|^{n+2s})^{-1}\qquad\qquad x\in\mathbb{R}^n. \end{equation} Take $R>0$ such that $\textnormal{supp}(\omega)\subset B_R$. \begin{equation*} \begin{split} |\psi*\omega(x)|& \leq \int_{B_R}|\psi(x-y)|\,d\omega(y)\leq C\int_{B_R}(1+|x-y|^{n+2s})^{-1}\,d\omega(y) \\& = C(1+|x|^{n+2s})^{-1}\int_{B_R}\frac{1+|x|^{n+2s}}{1+|x-y|^{n+2s}}\,d\omega(y). \end{split} \end{equation*}Now, observe that $$\frac{|x|^{n+2s}}{1+(|x|-R)^{n+2s}}\to 1\qquad\qquad \textnormal{for}\;|x|\to\infty.$$Then, there is $r>0$ such that \begin{equation}\label{lim int} \left|\frac{|x|^{n+2s}}{1+(|x|-R)^{n+2s}}-1 \right|<1 \end{equation}for $|x|>r$. Let $r_0>\max\{2R,r\}$. Then, for $x\in B_{r_0}$ $$\int_{B_R}\frac{1+|x|^{n+2s}}{1+|x-y|^{n+2s}}\,d\omega(y)\leq (1+r_0^{n+2s})\omega(B_R)<\infty$$and for $x\in\mathbb{R}^n\setminus B_{r_0}$, \begin{equation*} \begin{split} \int_{B_R}\frac{1+|x|^{n+2s}}{1+|x-y|^{n+2s}}\,d\omega(y)& = \int_{B_R}\frac{1}{1+|x-y|^{n+2s}}\,d\omega(y) + \int_{B_R}\frac{|x|^{n+2s}}{1+|x-y|^{n+2s}}\,d\omega(y) \\& \leq \omega(B_R) + \int_{B_R}\frac{|x|^{n+2s}}{1+|x-y|^{n+2s}}\,d\omega(y)\\&<\infty \end{split} \end{equation*}by \eqref{lim int} and the choice of $r_0$. Thus, we have proved \eqref{cotapsiconvomega} and we conclude that $\psi*\omega\in \mathcal{S}_s(\mathbb{R}^n)$. Moreover, observe that the integrability of $\psi*\omega$ implies $\mathcal{F}(\psi*\omega)\in\mathcal{C}^{\infty}(\mathbb{R}^n).$ To verify the integrability of $\mathcal{F}(\psi*\omega)$ note that it easily follows $$\mathcal{F}(\psi*\omega)(x) = \mathcal{F}(\psi)(x)\mathcal{F}(\omega)(x) \qquad \text{ for all }x.$$ Recalling $\mathcal{F}(\psi)(x)=\varphi(-x)=|x|^{2s}\mathcal{F}(\varphi_0)(-x)$, we have \begin{equation*} \begin{split} \int_{\mathbb{R}^n}|\mathcal{F}(\psi*\omega)(x)|\,dx & = \int_{\mathbb{R}^n}|\mathcal{F}(\psi)(x)||\mathcal{F}(\omega)(x)|\,dx \\& \leq\omega(B_R)\int_{\mathbb{R}^n}|\mathcal{F}(\psi)(x)|\,dx \\& = \omega(B_R)\int_{\mathbb{R}^n}|x|^{2s}|\mathcal{F}(\varphi_0)(-x)|\,dx \\& = \omega(B_R)\left[\int_{B_1}|x|^{2s}|\mathcal{F}(\varphi_0)(-x)|\,dx + \int_{\mathbb{R}^n\setminus B_1}|x|^{2s}|\mathcal{F}(\varphi_0)(-x)|\,dx\right] \\ & \leq C+ \int_{\mathbb{R}^n\setminus B_1}|x|^{2s}(1+|x|)^{-n-2}\,dx < \infty. \end{split} \end{equation*} We can apply now Lemma \ref{LemmaBucur} to $f=\mathcal{F}(\psi*\omega)$ to obtain \begin{equation}\label{1 eqq} \begin{split} \int_{\mathbb{R}^n}I_{2s}(x)(\psi*\omega)(x)\;dx &=\int_{\mathbb{R}^n}I_{2s}(x)\mathcal{F}^{-1}(\mathcal{F}(\psi*\omega))(x)\;dx\\& = \int_{\mathbb{R}^n}|x|^{-2s}\mathcal{F}(\omega)(x)\mathcal{F}(\psi)(x)\;dx. \end{split} \end{equation}Now, \begin{equation}\label{eq 1q} \begin{split} \int_{\mathbb{R}^n}I_{2s}(x)(\psi*\omega)(x)\;dx &= \int_{\mathbb{R}^n}\int_{\mathbb{R}^n}I_{2s}(x)\psi(x-y)\,d\omega(y)\,dx \\& = \int_{\mathbb{R}^n}\left[\int_{\mathbb{R}^n} I_{2s}(x)\mathcal{F}^{-1}(\varphi)(y-x)\,dx\right]\,d\omega(y). \end{split} \end{equation}We make the change of variable $z=x-y$ in the last integral and obtain \begin{equation}\label{eq 1} \begin{split} \int_{\mathbb{R}^n}\left[ \int_{\mathbb{R}^n}I_{2s}(y+z)\mathcal{F}^{-1}(\varphi)(-z)\,dz\right] \,d\omega(y) & = \int_{\mathbb{R}^n}\mathcal{F}^{-1}(\varphi)(-z)\left[ \int_{\mathbb{R}^n}I_{2s}(y+z)\,d\omega(y)\right]\,dz \\& = \int_{\mathbb{R}^n}\mathcal{I}_{2s}(\omega)(-z)\mathcal{F}^{-1}(\varphi)(-z)\,dz \\ & = -\int_{\mathbb{R}^n}\mathcal{I}_{2s}(\omega)(x)\mathcal{F}^{-1}(\varphi)(x)\,dx. \end{split} \end{equation}Thus from \eqref{1 eqq}, \eqref{eq 1q} and \eqref{eq 1} we get \begin{equation}\label{eqq12} \int_{\mathbb{R}^n}\mathcal{I}_{2s}(\omega)(x)\mathcal{F}^{-1}(\varphi)(x)\,dx =- \int_{\mathbb{R}^n}|x|^{-2s}\mathcal{F}(\omega)(x)\mathcal{F}(\psi)(x)\;dx =- \int_{\mathbb{R}^n}|x|^{-2s}\mathcal{F}(\omega)(x)\varphi(-x)\;dx. \end{equation} Then \begin{equation*} \begin{split} \int_{\mathbb{R}^n}\mathcal{I}_{2s}(\omega)(x)(-\Delta)^s(\varphi_0)(x)\,dx & = -\int_{\mathbb{R}^n}|x|^{-2s}\mathcal{F}(\omega)(x)\varphi(-x)\,dx \qquad \text{by }\eqref{eqq12}\\& = \int_{\mathbb{R}^n}|x|^{-2s}\mathcal{F}(\omega)(-x)\varphi(x)\,dx \\& = \int_{\mathbb{R}^n}|x|^{-2s}\mathcal{F}^{-1}(\omega)(x)|x|^{2s}\mathcal{F}(\varphi_0)(x)\,dx \\& = \int_{\mathbb{R}^n}\left[ \int_{\mathbb{R}^n}e^{2\pi ix\cdot y}\,d\omega(y)\right] \mathcal{F}(\varphi_0)(x)\,dx \\& = \int_{\mathbb{R}^n}\left[ \int_{\mathbb{R}^n}e^{2\pi ix\cdot y}\mathcal{F}(\varphi_0)(x)\,dx\right] d\omega(y)\\& = \int_{\mathbb{R}^n}\varphi_0(y)\,d\omega(y). \end{split} \end{equation*} \end{proof} \subsection{Step (II)} \begin{lemma}\label{ineqgrad} Let $u_0=\mathcal{I}_{2s}(\omega)$. There exists $C_0=C_0(n, s)$ such that \begin{equation}\label{eqgrad} |\nabla u_0|\leq C_0\mathcal{I}_{2s-1}(\omega). \end{equation} \end{lemma \begin{proof} We will show that \begin{equation}\label{deriv} \frac{\partial u_0}{\partial x_i}(x)=c(n,2s)\int_{\mathbb{R}^n}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\;d\omega(y)\qquad i=1,...,n. \end{equation}in the weak sense. Let $\varphi\in \mathcal{C}_0^\infty(\mathbb{R}^n)$ and let $R>0$ such that $\textnormal{supp}(\varphi)\subset B_R$. We want to show \begin{equation}\label{weakderiv} \int_{\mathbb{R}^n}u_0(x)\frac{\partial\varphi}{\partial x_i}(x)\;dx = -\int_{\mathbb{R}^n}\left[c(n,2s)\int_{\mathbb{R}^n}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\;d\omega(y) \right]\varphi(x)\;dx. \end{equation}Now, by Fubini's Theorem we have that \begin{equation*} \begin{split} & \int_{\mathbb{R}^n}\left[ \int_{\textnormal{supp}(\omega)}\frac{c(n,2s)}{|x-y|^{n-2s}}\frac{\partial\varphi}{\partial x_i}(x)\;d\omega(y)\right]dx \\ & \qquad \qquad = \int_{\textnormal{supp}(\omega)}\left[ \int_{B_R}\frac{c(n,2s)}{|x-y|^{n-2s}}\frac{\partial\varphi}{\partial x_i}(x)\;dx\right]d\omega(y) \end{split} \end{equation*}and \begin{equation*} \begin{split} &-\int_{\mathbb{R}^n}\left[ c(n,2s)\int_{\textnormal{supp}(\omega)}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)d\omega(y)\right]dx \\ & \qquad \qquad = -c(n,2s)\int_{\textnormal{supp}(\omega)}\left[ \int_{B_R}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)dx\right]d\omega(y). \end{split} \end{equation*}For $r>0$ small and $y\in\textnormal{supp}(\omega)$ fixed, integration by parts gives \begin{equation}\label{int by parts} \int_{B_R\setminus B_r(y)}\frac{1}{|x-y|^{n-2s}}\frac{\partial \varphi}{\partial x_i}(x)\;dx = \int_{\partial B_r(y)}\frac{\varphi(x)\eta_i}{|x-y|^{n-2s}}\;dS_x - \int_{B_R\setminus B_r(y)}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)\;dx. \end{equation}Here we have used the fact that $\varphi=0$ in $\partial B_R$. The vector $\eta=(\eta_1,...,\eta_n)$ is the exterior normal unit vector to $\partial B_r(y)$, so $$\eta_i=\frac{(x_i-y_i)}{|x-y|}.$$ Now, for all $y\in\textnormal{supp}(\omega)$ \begin{equation*} \begin{split} \left|\int_{\partial B_r(y)}\frac{1}{|x-y|^{n-2s}}\varphi(x)\eta_i\;dS_x \right|=&\left| \int_{\partial B_r(y)}\frac{(x_i-y_i)}{|x-y|^{n-2s+1}}\varphi(x)\;dS_x\right| \\&\leq \int_{\partial B_r(y)}\frac{1}{|x-y|^{n-2s}}|\varphi(x)|\;dS_x \\ & = r^{2s-n}\int_{\partial B_r(y)}|\varphi(x)|\;dS_x \\& = r^{2s-1}r^{1-n}\int_{\partial B_r(y)}|\varphi(x)|\;dS_x. \end{split} \end{equation*}Observe $$\lim_{r\to 0}r^{1-n}\int_{\partial B_r(y)}|\varphi(x)|\;dS_x = \omega_n|\varphi(y)|\leq C$$for all $y\in\textnormal{supp}(\omega).$ So $$\int_{\partial B_r(y)}\frac{1}{|x-y|^{n-2s}}\varphi(x)\eta_i \;dS_x= o(1)$$for $r\to 0$ uniformly on $y\in\textnormal{supp}(\omega).$ Regarding the last term in \eqref{int by parts} we have $$\int_{B_R\setminus B_r(y)}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)\;dx \leq C \int_{B_R}\frac{1}{|x-y|^{n-2s+1}}dx$$and taking $R_0>0$ big enough we obtain that \begin{equation*} \begin{split} \int_{\textnormal{supp}(\omega)}\int_{B_{R_0}}\frac{1}{|x-y|^{n-2s+1}}\;dx\;d\omega(y)& = \int_{\textnormal{supp}(\omega)}\int_{0}^{R_0}\omega_n r^{2s-n-1}r^{n-1}\;dr\;d\omega(y)\\& \leq C R_0^{2s-1}<\infty. \end{split} \end{equation*} \noindent Hence, by Lebesgue dominated convergence theorem $$\lim_{r\to 0}\int_{\textnormal{supp}(\omega)}\int_{B_R}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)\mathcal{X}_{B_R\setminus B_r(y)}\;dx\;d\omega(y)= \int_{\textnormal{supp}(\omega)}\int_{B_R}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)\;dx\;d\omega(y).$$Therefore $$\lim_{r\to 0}\left[ \int_{\textnormal{supp}(\omega)}\int_{B_R\setminus B_r(y)}\frac{1}{|x-y|^{n-2s}}\frac{\partial\varphi}{\partial x_i}(x)\;dx\;d\omega(y)\right]= -\int_{\textnormal{supp}(\omega)}\int_{B_R}\frac{(x_i-y_i)}{|x-y|^{n-2s+2}}\varphi(x)\;dx\;d\omega(y).$$ Similarly $$\lim_{r\to 0}\int_{\textnormal{supp}(\omega)}\int_{B_R\setminus B_r(y)}\frac{1}{|x-y|^{n-2s}}\frac{\partial \varphi}{\partial x_i}\;dx\;d\omega(y) = \int_{\textnormal{supp}(\omega)}\int_{B_R}\frac{1}{|x-y|^{n-2s}}\frac{\partial\varphi}{\partial x_i}\;dx\;d\omega(y).$$Then, we have proved \eqref{weakderiv} and therefore we have $$\nabla u_0(x) = c(n,2s)\int_{\mathbb{R}^n}\frac{(x-y)}{|x-y|^{n-2s+2}}\;d\omega(y)$$in the weak sense. \end{proof} \begin{comment} \begin{proposition}\label{sufcond} Suppose there exists a constant $C_1= C_1(n, s)$ such that \eqref{ineqpot} is satisfied. If \begin{equation}\label{sufcondeq} \int_{|y|>R_0}\frac{|\nabla u_0(y)|^q}{|y|^{n-2s}}\;dy<\infty \end{equation}for some $R_0>0$, then $\mathcal{I}_{2s}(|\nabla u_0|^q)\in L^1_{\textnormal{loc}}(\mathbb{R}^n)$. \end{proposition} \begin{proof} First of all, by Theorem 1.1 from \cite{Han} and the assumption \eqref{ineqpot} we have that \begin{equation}\label{regpot} \mathcal{I}_{2s-1}(\omega)\in L^q_{\textnormal{loc}}(\mathbb{R}^n). \end{equation} Hence, by Lemma \ref{ineqgrad}, \begin{equation}\label{eq4} |\nabla u_0|\in L^q_{\textnormal{loc}}(\mathbb{R}^n). \end{equation} Now, suppose there is $R_0>0$ such that \eqref{sufcondeq} holds. Since for all $r>0$ $$\int_{|x|<r}\frac{dx}{|x-y|^{n-2s}}=O\left( \frac{1}{|y|^{n-2s}}\right),\qquad\qquad|y|\to\infty$$ there exist $M,t_0>0$ such that $$\int_{|x|<r}\frac{dx}{|x-y|^{n-2s}}\leq M\frac{1}{|y|^{n-2s}}$$for $|y|\geq t_0$. Now let $r>0$ and take $r_0=\max\{t_0,R_0\}$. Then, \begin{equation*} \begin{split} \int_{|x|<r}\mathcal{I}_{2s}(|\nabla u_0|^q)(x)dx & = \int_{|x|<r}\left( \int_{\mathbb{R}^n}\frac{|\nabla u_0(y)|^q}{|x-y|^{n-2s}}\;dy\right)dx \\& = \int_{|x|<r}\left( \int_{|y|<r_0}\frac{|\nabla u_0(y)|^q}{|x-y|^{n-2s}}\;dy\right)dx + \int_{|x|<r}\left(\int_{|y|>r_0}\frac{|\nabla u_0(y)|^q}{|x-y|^{n-2s}}\;dy \right) dx\\& = (I)+(II). \end{split} \end{equation*}Now, for $(II)$, we use \eqref{sufcondeq} and that $r_0=\max\{t_0,R_0\}$ to obtain \begin{equation*} \begin{split} \int_{|x|<r}\left(\int_{|y|>r_0}\frac{|\nabla u_0(y)|^q}{|x-y|^{n-2s}}\;dy \right) dx& = \int_{|y|>r_0}|\nabla u_0(y)|^q\left(\int_{|x|<r}\frac{1}{|x-y|^{n-2s}}\;dx \right)dy \\& \leq M\int_{|y|>r_0}\frac{|\nabla u_0(y)|^q}{|y|^{n-2s}}\;dy \\& \leq M\int_{|y|>R_0}\frac{|\nabla u_0(y)|^q}{|y|^{n-2s}}\;dy\\& <\infty. \end{split} \end{equation*}On the other hand, since $|\nabla u_0|\in L^q_{\textnormal{loc}}(\mathbb{R}^n)$, we have \begin{equation*} \begin{split} \int_{|x|<r}\left( \int_{|y|<r_0}\frac{|\nabla u_0(y)|^q}{|x-y|^{n-2s}}\;dy\right)dx & = \int_{|y|<r_0}|\nabla u_0(y)|^q\left( \int_{|x|<r}\frac{1}{|x-y|^{n-2s}}\;dx\right)dy \\& = \int_{|y|<r_0}|\nabla u_0(y)|^q\left( \int_{|z|<r+r_0}\frac{1}{|z|^{n-2s}}\;dz\right) dy \\& = \left(\frac{\omega_n(r+r_0)^{2s}}{2s} \right)\int_{|y|<r_0}|\nabla u_0(y)|^q \;dy\\& <\infty \end{split} \end{equation*} \end{proof} \end{comment} \begin{lemma}\label{regriesz} For $n>2s$ and $q>p^*$, if hypothesis \eqref{ineqpot} is satisfied with a constant $C_1=C_1(n, q, s)$, then $\mathcal{I}_{2s-1}(\omega)\in L^q(\mathbb{R}^n)$. \end{lemma} \begin{proof} First of all, by Theorem 2.1 from \cite{MV} and the assumption \eqref{ineqpot} we have that \begin{equation}\label{rieszLqloc} \mathcal{I}_{2s-1}(\omega)\in L^q_{\textnormal{loc}}(\mathbb{R}^n). \end{equation} Let $R>0$ such that $\textnormal{supp}(\omega)\subset B_R$. We write \begin{equation*} \int_{\mathbb{R}^n}|\mathcal{I}_{2s-1}(\omega)(x)|^q\,dx = \int_{B_{2R}}|\mathcal{I}_{2s-1}(\omega)(x)|^q\,dx + \int_{\mathbb{R}^n\setminus B_{2R}}|\mathcal{I}_{2s-1}(\omega)(x)|^q\,dx. \end{equation*}Then, by \eqref{rieszLqloc}, \begin{equation*} \int_{B_{2R}}|\mathcal{I}_{2s-1}(\omega)(x)|^q\,dx <\infty. \end{equation*}On the other hand, using Minkowsky's inequality \begin{equation*} \int_{\mathbb{R}^n\setminus B_{2R}}|\mathcal{I}_{2s-1}(\omega)(x)|^q\,dx \leq C(n,2s-1)^q\left[\int_{B_R}\left(\int_{\mathbb{R}^n\setminus B_{2R}}\frac{1}{|x-y|^{q(n-2s+1)}}\,dx \right) ^{\frac{1}{q}}\,d\omega(y) \right] ^q \end{equation*} Now, since $q>\frac{n}{n-2s+1}$ \begin{equation*} \int_{\mathbb{R}^n\setminus B_{2R}}\frac{1}{|x-y|^{q(n-2s+1)}} \,dx \leq \frac{R^{n+(2s-n-1)q}}{(n-2s+1)q -n} \end{equation*}and so the integral \begin{equation*} \left[\int_{B_R}\left( \int_{\mathbb{R}^n\setminus B_{2R}} \frac{1}{|x-y|^{q(n-2s+1)}} \,dx \right)^{1/q}d\omega(y) \right]^q \end{equation*}is finite. \end{proof} \begin{remark} \rm{ Observe that by Lemma \ref{ineqgrad}} and Lemma \ref{regriesz}, $|\nabla u_0|\in L^q(\mathbb{R}^n)$. \end{remark} \begin{proposition}\label{propp} Under the same conditions of Lemma \ref{regriesz}, $\mathcal{I}_{2s}(|\nabla u_0|^q)\in L_s(\mathbb{R}^n)$. \end{proposition} \begin{proof} Using the inequality \eqref{eqgrad}, we just need to verify that $\mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)\in L_s(\mathbb{R}^n)$. Hence \begin{equation*} \begin{split} \mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)(x) & = \int_{\mathbb{R}^n}\frac{c(n,2s)}{|x-y|^{n-2s}}\left[ \mathcal{I}_{2s-1}(\omega)\right]^q(y)\,dy \\& = c(n,2s) \left(\int_{B_1(x)}\frac{\left[ \mathcal{I}_{2s-1}(\omega)\right]^q(y)}{|x-y|^{n-2s}}\,dy + \int_{\mathbb{R}^n\setminus B_1(x)}\frac{\left[ \mathcal{I}_{2s-1}(\omega)\right]^q(y)}{|x-y|^{n-2s}}\,dy\right) \\& \leq c(n,2s) \left(\int_{B_1(x)}\frac{\left[ \mathcal{I}_{2s-1}(\omega)\right]^q(y)}{|x-y|^{n-2s+1}}\,dy + \int_{\mathbb{R}^n\setminus B_1(x)}\left[ \mathcal{I}_{2s-1}(\omega)\right]^q(y)\,dy\right) \\& = C(n,s)\left[ \mathcal{I}_{2s-1}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)(x) + \|\mathcal{I}_{2s-1}(\omega)\|^q_{L^q(\mathbb{R}^n)}\right] \\& \leq C(q,n,s)\left[ \mathcal{I}_{2s-1}(\omega)(x) + \|\mathcal{I}_{2s-1}(\omega)\|^q_{L^q(\mathbb{R}^n)}\right] . \end{split} \end{equation*}Thus, using Remark \ref{regI2s-1w} we conclude the proof. \end{proof} \begin{proposition}\label{solu0} For $n>2s$ and $q>\frac{n}{n-2s+1}$, if there exists $C_1=C_1(n, q, s)>0$ such that \eqref{ineqpot} is satisfied, then $\mathcal{I}_{2s}(|\nabla u_0|^q)$ is a weak solution of \begin{equation}\label{eq3} (-\Delta)^sv=|\nabla u_0|^q \quad \text{in }\mathbb{R}^{n}. \end{equation} \end{proposition} \begin{proof} Let $\varphi_0\in \mathcal{S}(\mathbb{R}^n)$. We proceed as in Proposition \ref{SolRiesz}, naming $\varphi(x)=|x|^{2s}\mathcal{F}(\varphi_0)(x)$ and $\psi(x)=\mathcal{F}(\varphi)(-x)$. Recall that $\psi\in\mathcal{S}_s(\mathbb{R}^n)$. We will apply Lemma \ref{LemmaBucur} as in the proof of Lemma \ref{SolRiesz} to obtain the desired result. According to Remark \ref{rm Buc}, we just need to verify the following statements \ \begin{enumerate}[i.] \item $\mathcal{F}(\psi*|\nabla u_0|^q)\in L^1(\mathbb{R}^n)\cap\mathcal{C}(\mathbb{R}^n)$, \ \item $\int_{\mathbb{R}^n}I_{2s}(x)(\psi*|\nabla u_0|^q)(x)\,dx<\infty$, and \ \item $\psi*|\nabla u_0|^q\in L^2(\mathbb{R}^n)$. \end{enumerate} \ First of all, $\mathcal{F}(\psi*|\nabla u_0|^q)\in L^1(\mathbb{R}^n)\cap\mathcal{C}(\mathbb{R}^n)$. The continuity follows from the fact that $\psi*|\nabla u_0|^q\in L^1(\mathbb{R}^n)$ since $\psi, |\nabla u_0|^q \in L^1(\mathbb{R}^n)$. The integrability can be checked easily using that $|\nabla u_0|^q\in L^1(\mathbb{R}^n)$. Thus, we will concentrate on ii. and iii. ii. \begin{equation*} \begin{split} \int_{\mathbb{R}^n}|x|^{2s-n}(\psi*|\nabla u_0|^q)(x)\,dx &= \int_{\mathbb{R}^n}\frac{1}{|x|^{n-2s}}\left(\int_{\mathbb{R}^n}\psi(x-y)|\nabla u_0(y)|^q\,dy \right)\,dx \\& = \int_{\mathbb{R}^n}|\nabla u_0(y)|^q\left( \int_{\mathbb{R}^n}\frac{\psi(x-y)}{|x|^{n-2s}}\,dx\right)\,dy \\& = \int_{\mathbb{R}^n}|\nabla u_0(y)|^q\left(\int_{B_1}\frac{\psi(x-y)}{|x|^{n-2s}}\,dx + \int_{\mathbb{R}^n\setminus B_1}\frac{\psi(x-y)}{|x|^{n-2s}}\,dx \right)\,dy. \end{split} \end{equation*}Now, since $\psi\in\mathcal{S}_s(\mathbb{R}^n)$, \begin{equation*} \int_{B_1}\frac{\psi(x-y)}{|x|^{n-2s}}\,dx\leq C \int_{B_1}\frac{1}{|x|^{n-2s}}\,dx = \frac{C\omega_n}{2s}. \end{equation*}On the other hand, \begin{equation*} \int_{\mathbb{R}^n\setminus B_1}\frac{\psi(x-y)}{|x|^{n-2s}}\,dx\leq \|\psi\|_{L^1(\mathbb{R}^n)}. \end{equation*}Then, \begin{equation*} \int_{\mathbb{R}^n}|x|^{2s-n}(\psi*|\nabla u_0|^q)(x)\,dx\leq C(n,s,\psi)\|\nabla u_0\|_{L^q(\mathbb{R}^n)}<\infty. \end{equation*} iii. Follows by Young's inequality. \end{proof} \begin{comment} \begin{proof} We need to verify the hypothesis of Theorem 0.13 from \cite{Land}. First we proof that $\mathcal{F}(|\nabla u_0|^q)\in L^1_{\textnormal{loc}}(\mathbb{R}^n)$. Let $K\subset\mathbb{R}^n$ be compact. By Lemma \ref{ineqgrad}, $$\int_K\mathcal{F}(|\nabla u_0|^q)(x)\;dx \leq C\int_K \mathcal{F}(\mathcal{I}_{2s-1}(\omega)^q)(x)\;dx.$$ Now take $R>0$ such that $\textnormal{supp}(\omega)\subset B_R$. Then, \begin{equation*} \begin{split} \int_K \mathcal{F}(\mathcal{I}_{2s-1}(\omega)^q)(x)\;dx& = \int_K\left( \int_{\mathbb{R}^n}\mathcal{I}_{2s-1}(\omega)(y)^qe^{-2\pi x\cdot y}\;dy\right)dx \\&\leq \int_K\left( \int_{B_{2R}}\mathcal{I}_{2s-1}(\omega)(y)^q\;dy\right)dx + \int_K\left(\int_{\mathbb{R}^n\setminus B_{2R}}\mathcal{I}_{2s-1}(\omega)(y)^q\;dy \right)dx. \end{split} \end{equation*}Now, the first integral is bounded by $|K|\|\mathcal{I}_{2s-1}(\omega)\|_{L^q(B_{2R})}^q$ beacuse of \eqref{regpot}. Here, $|K|$ is the Lebesgue measure of the compact set $K$. For the second integral we first use Minckowscky inequality to obtain \begin{equation*} \begin{split} \int_{\mathbb{R}^n\setminus B_{2R}}\mathcal{I}_{2s-1}(\omega)(y)^q\;dy&= \int_{\mathbb{R}^n\setminus B_{2R}}\left( \int_{B_R}\frac{1}{|z-y|^{n-2s+1}}\;d\omega(z)\right)^q dy \\& \leq \left[\int_{B_R}\left( \int_{\mathbb{R}^n\setminus B_{2R}}\left( \frac{1}{|z-y|^{n-2s+1}}\right)^q dy \right)^{1/q}d\omega(z) \right]^q. \end{split} \end{equation*}(For simplicity, here we have omitted the constant from the Riesz potential). Now, using the fact that $q>\frac{n}{n-2s+1}$, we see that \begin{equation*} \begin{split} \int_{\mathbb{R}^n\setminus B_{2R}}\left( \frac{1}{|z-y|^{n-2s+1}}\right)^q dy & = \int_{\mathbb{R}^n\setminus(B_{2R-|z|})}\frac{1}{|t|^{(n-2s+1)q}}\;dt \\& = \int_{2R-|z|}^{\infty}r^{n-1}r^{(2s-n-1)q}\omega_n\;dr \\& = \lim_{\rho\to\infty}\left( \frac{\rho^{n+(2s-n-1)q}-(2R-|z|)^{n+(2s-n-1)q}}{n+(2s-n-1)q}\right) \\& = \frac{(2R-|z|)^{n+(2s-n-1)q}}{(n-2s+1)q-n}. \end{split} \end{equation*}(Since $q>\frac{n}{n-2s+1}$, $n+(2s-n-1)q<0$ and $(n-2s+1)q-n>0$.) Finally, since $2R-|z|>R$ for $|z|<R$, \begin{equation*} \begin{split} \left[\int_{B_R}\left( \int_{\mathbb{R}^n\setminus B_{2R}}\left( \frac{1}{|z-y|^{n-2s+1}}\right)^q dy \right)^{1/q}d\omega(z) \right]^q&\leq \left[\int_{B_R}\left( \frac{(2R-|z|)^{n+(2s-n-1)q}}{(n-2s+1)q-n}\right) ^{\frac{1}{q}}d\omega(z)\right]^q\\& \leq \frac{1}{(n-2s+1)q-n}\left[ \int_{B_R}\left( R^{n+(2s-n-1)q}\right) ^{\frac{1}{q}}d\omega(z)\right] ^q\\&\leq \frac{\left( R^{n+(2s-n-1)q}\right) ^{\frac{1}{q}}}{(n-2s+1)q-n}\omega(B_R)\\& \leq \frac{\left( R^{n+(2s-n-1)q}\right) ^{\frac{1}{q}}}{(n-2s+1)q-n}C_2R^{n-(2s-1)q'}. \end{split} \end{equation*}The only thing remains to prove is that $\mathcal{I}_{2s}(|\nabla u_0|^q)\in L^1_{\textnormal{loc}}(\mathbb{R}^n)$. For this purpose observe that \begin{equation*} \begin{split} \int_{|y|>2R}\frac{|\nabla u_0(y)|^q}{|y|^{n-2s}}\;dy&\leq C \int_{|y|>2R}\frac{\mathcal{I}_{2s-1}(\omega)^q}{|y|^{n-2s}}\;dy \\& = C\int_{|y|>2R}\frac{1}{|y|^{n-2s}}\left(\int_{B_R}\frac{1}{|x-y|^{n-2s+1}}\;d\omega(x) \right)^q dy \\& \leq C \left[ \int_{B_R}\left( \int_{|y|>2R}\frac{1}{|y|^{n-2s}|x-y|^{(n-2s+1)q}}\;dy\right) ^{\frac{1}{q}}d\omega(x)\right]^q. \end{split} \end{equation*}Now, using the fact that $|y|>2R$ and taking $z=x-y$ we obtain \begin{equation*} \begin{split} \int_{|y|>2R}\frac{1}{|y|^{n-2s}|x-y|^{(n-2s+1)q}}\;dy&\leq \frac{1}{(2R)^{n-2s}}\int_{|y|>2R}\frac{1}{|x-y|^{(n-2s+1)q}}\;dy \\&= \frac{1}{(2R)^{n-2s}}\int_{|z|>R}\frac{1}{|z|^{(n-2s+1)q}}\;dz \\& = \frac{\omega_nR^{(2s-n-1)q+2s}}{2^{n-2s}((n-2s+1)q-n)}\\& <\infty. \end{split} \end{equation*}Thus, $$\int_{|y|>2R}\frac{|\nabla u_0(y)|^q}{|y|^{n-2s}}\;dy<\infty.$$ Then, by Proposition \ref{sufcond}, we can conclude that $\mathcal{I}_{2s}(|\nabla u_0|^q)\in L^1_{\textnormal{loc}}(\mathbb{R}^n)$. Finally we are abble to apply Theorem 0.13 from \cite{Land} and obtain \eqref{eq3} in the weak sense with $v=\mathcal{I}_{2s}(|\nabla u_0|^q)$. \end{proof} \end{comment} \subsection{Step (III)} \begin{proof}[Proof of Theorem \ref{Theo1}] We begin constructing a sequence of functions $u_{k+1}$ as follows. Let $u_0$ be as in Lemma \ref{ineqgrad} and define by recursion \begin{equation}\label{seq} u_{k+1}=\mathcal{I}_{2s}(|\nabla u_k|^q)+\mathcal{I}_{2s}(\omega). \end{equation}Then by propositions \ref{SolRiesz} and \ref{solu0}, \begin{equation} (-\Delta)^su_{k+1}=|\nabla u_k|^q+\omega \quad \text{in }\mathbb{R}^{n}. \end{equation}We claim that \begin{equation}\label{ineq1} |\nabla u_k|\leq C_2\mathcal{I}_{2s-1}(\omega), \end{equation} \begin{equation}\label{ineq2} |\nabla u_{k+1}-\nabla u_k|\leq C_3\delta^k\mathcal{I}_{2s-1}(\omega) \end{equation}for some $0<\delta<1$. Let us begin with \eqref{ineq1} proceding by induction. Beacuse of Lemma \ref{ineqgrad}, \eqref{ineq1} holds for $k=0$. Suppose that \begin{equation}\label{induc} |\nabla u_k|\leq a_k\mathcal{I}_{2s-1}(\omega). \end{equation}Then, we see that \begin{equation*} \begin{split} |\nabla u_{k+1}|& \leq |\nabla(\mathcal{I}_{2s}(|\nabla u_k|^q))|+|\nabla(\mathcal{I}_{2s}(\omega))|\\&\leq C_0\left( \mathcal{I}_{2s-1}(|\nabla u_k|^q)+\mathcal{I}_{2s-1}(\omega)\right) \\& \leq C_0 \left(\mathcal{I}_{2s-1}(a_k^q\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)+\mathcal{I}_{2s-1}(\omega)\right) \quad \text{by \eqref{induc}}\\&\leq C_0\left(a_k^qC_1\mathcal{I}_{2s-1}(\omega)+\mathcal{I}_{2s-1}(\omega) \right)\quad \text{by \eqref{ineqpot}}\\& = C_0(a_k^qC_1+1)\mathcal{I}_{2s-1}(\omega). \end{split} \end{equation*}Then $$|\nabla u_{k+1}|\leq a_{k+1}\mathcal{I}_{2s-1}(\omega)$$with $a_{k+1}=C_0(a_k^qC_1+1)$. Then, if $C_1\leq (q')^{1-q}q^{-1}C_0^{-q}$ we see that $$\lim_{k\to\infty}a_k=a\leq C_0q'$$where $a$ is a root of the equation $x=C_0(x^qC_1+1)$. Hence \eqref{ineq1} holds with $C_2=C_0q'$. Now we prove \eqref{ineq2}. Assume \eqref{ineqpot} holds with $C_1\leq (q')^{1-q}q^{-1}C_0^{-q}$ so that \eqref{ineq1} is satisfied with $C_2=C_0q'$, where $C_0$ is the constant of \eqref{eqgrad}. Now, \begin{equation*} u_1-u_0 = \mathcal{I}_{2s}(|\nabla u_0|^q). \end{equation*}Then, \begin{equation*} \begin{split} |\nabla u_1 - \nabla u_0| & \leq C_0\mathcal{I}_{2s-1}(|\nabla u_0|^q) \\& \leq C_0 C_2^q \mathcal{I}_{2s-1}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)\\& \leq C_0C_2^qC_1\mathcal{I}_{2s-1}(\omega). \end{split} \end{equation*}Therefore $$|\nabla u_1 - \nabla u_0|\leq b_0 \mathcal{I}_{2s-1}(\omega)$$with $b_0=C_0C_2^qC_1$. Analogously, $$u_{k+1}-u_k = \mathcal{I}_{2s}(|\nabla u_k|^q-|\nabla u_{k-1}|^q)$$and \begin{equation}\label{eqdifgrad} |\nabla u_{k+1}-\nabla u_k|\leq C_0\mathcal{I}_{2s-1}(|\nabla u_k|^q-|\nabla u_{k-1}|^q). \end{equation}Using the inequality $|r^q-s^q|\leq q|r-s|\max\{r,s\}^{q-1}$ with $r=|\nabla u_k|$ and $s=|\nabla u_{k-1}|$ we obtain \begin{equation}\label{eqdifgrad2} \begin{split} \left| |\nabla u_k|^q-|\nabla u_{k-1}|^q\right|&\leq q\left| |\nabla u_k|-|\nabla u_{k-1}|\right|\max\{|\nabla u_k|,|\nabla u_{k-1}|\}^{q-1}\\& \leq C_2^{q-1}q|\nabla u_k-\nabla u_{k-1}|\left[ \mathcal{I}_{2s-1}(\omega)\right] ^{q-1}. \end{split} \end{equation}Thus by \eqref{eqdifgrad} and \eqref{eqdifgrad2} \begin{equation}\label{ineqgrad3} |\nabla u_{k+1}-\nabla u_k|\leq C_0C_2^{q-1}q\mathcal{I}_{2s-1}(|\nabla u_k-\nabla u_{k-1}|\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q). \end{equation}Now suppose $|\nabla u_k-\nabla u_{k-1}|\leq b_k\mathcal{I}_{2s-1}(\omega)$. Then, by \eqref{ineqgrad3} \begin{equation*} \begin{split} |\nabla u_{k+1}-\nabla u_k|&\leq C_0C_2^{q-1}q\mathcal{I}_{2s-1}(b_k\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)\\& \leq C_0C_2^{q-1}qb_kC_1\mathcal{I}_{2s-1}(\omega). \end{split} \end{equation*}Then, arguing by induction we see that $$|\nabla u_{k+1}-\nabla u_k|\leq b_{k+1}\mathcal{I}_{2s-1}(\omega)$$with $b_{k+1}\leq C_0C_2^{q-1}qC_1b_k$. Thus, $$b_{k+1}\leq \left( C_0C_2^{q-1}qC_1\right)^{k+1}b_0$$with $b_0=C_0C_2^qC_1$. Taking $C_1$ such that $\delta=C_0C_2^{q-1}qC_1<1$, we obtain \eqref{ineq2} with $C_3=b_0$. Now we claim that \begin{equation}\label{ineq3} |u_{k+1}-u_k|\leq C_4\delta^k\mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q) \end{equation}with $C_4>0$ and $0<\delta<1$ depending only on $q$, $n$ and $s$. Indeed, by \eqref{eqdifgrad2} and \eqref{ineq2} \begin{equation*} \begin{split} |u_{k+1}-u_k|& \leq \mathcal{I}_{2s}\left(\left| |\nabla u_k|^q-|\nabla u_{k-1}|^q\right| \right)\\&\leq C_2^{q-1}q\mathcal{I}_{2s}(|\nabla u_k-\nabla u_{k-1}|\left[ \mathcal{I}_{2s-1}(\omega)\right] ^{q-1})\\&\leq C_2^{q-1}qC_3\delta^k\mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q). \end{split} \end{equation*}Then \eqref{ineq3} holds for $C_4=C_2^{q-1}qC_3$. Now suppose \eqref{ineqpot} holds for $C_1$ small enough so that \eqref{ineq1} and \eqref{ineq2} are satisfied. Let $$u(x)=u_0(x)+\sum_{k=0}^{\infty}(u_{k+1}(x)-u_k(x))$$with $u_k$ defined by \eqref{seq}. By \eqref{ineq3}, $$|u_{k+1}(x)-u_k(x)|\leq C_4\delta^k\mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right]^q).$$Hence, $u(x)=\lim_{k\to\infty}u_k(x)$ and $u\in L_s(\mathbb{R}^n)$ since \begin{equation}\label{estimate 2} |u|\leq C\left( \mathcal{I}_{2s}\omega + \mathcal{I}_{2s}\left[ \mathcal{I}_{2s-1}\omega\right]^q \right) \end{equation}On the other side, by \eqref{ineq2} $$|\nabla u_{k+1}-\nabla u_k|\leq C_3\delta^k\mathcal{I}_{2s-1}(\omega),$$hence \begin{equation}\label{to represent} |\nabla u|\leq C\mathcal{I}_{2s-1}(\omega). \end{equation} Then, by Lemma \ref{regriesz} $|\nabla u|\in L^q(\mathbb{R}^n)$. Therefore, $u\in W_{\textnormal{loc}}^{1,q}(\mathbb{R}^n)\cap L_s(\mathbb{R}^n)$ (observe that $u \in L_{\text{loc}}^{q}(\mathbb{R}^{n})$ by \eqref{for the pf}). We now check that Definition \ref{Defsol} holds. Let $\varphi\in S(\mathbb{R}^n)$. We have that $\nabla u(x)=\lim_{k\to\infty}\nabla u_k(x)=\nabla u_0(x)+\sum_{k=0}^{\infty}(\nabla u_{k+1}(x)-\nabla u_k(x))$ a.e. in $\mathbb{R}^n$. Also, $|\nabla u_k|^q\leq C\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q\in L^1(\mathbb{R}^n)$ for all $k$. Then, letting $k\to\infty$ $$\int_{\mathbb{R}^n}\varphi |\nabla u_k|^q\;dx \rightarrow\int_{\mathbb{R}^n}\varphi |\nabla u|^q\;dx.$$ On the other side observe that, for all k $$|u_k|\leq C\left[ \mathcal{I}_{2s}(\omega) + \mathcal{I}_{2s}\left( \left[\mathcal{I}_{2s-1}(\omega) \right]^q \right) \right].$$Thus, since $\varphi\in\mathcal{S}(\mathbb{R}^n)$ and $\mathcal{I}_{2s}(\omega)$, $\mathcal{I}_{2s}\left( \left[ \mathcal{I}_{2s-1}(\omega)\right]^q\right) \in L_s(\mathbb{R}^n)$, we can apply Lebesgue dominated convergence theorem and obtain $$\int_{\mathbb{R}^n}u_{k+1}(-\Delta)^s\varphi\;dx \to \int_{\mathbb{R}^n}u(-\Delta)^s\varphi\;dx$$for $k\to\infty$. Now, since $(-\Delta)^su_{k+1}=|\nabla u_k|^q + \omega$ in $\mathbb{R}^n$, by Definition \ref{Defsol} $$\int_{\mathbb{R}^n}u_{k+1}(-\Delta)^s\varphi\;dx = \int_{\mathbb{R}^n}\varphi |\nabla u_k|^q\;dx +\int_{\mathbb{R}^n}\varphi\;d\omega.$$ Consequently, letting $k\to \infty$ in the preceding equality we get the desired conclusion. Finally, we prove the representation \eqref{represent}. By \eqref{seq} is enough to show that, as $k\to\infty$, $\mathcal{I}_{2s}(|\nabla u_k|^q)\to \mathcal{I}_{2s}(|\nabla u|^q)$. First, by definition of Riesz potential we have \begin{equation*} |\mathcal{I}_{2s}(|\nabla u_k|^q)-\mathcal{I}_{2s}(|\nabla u|^q)| \leq c(n,2s)\int_{\mathbb{R}^n}\frac{||\nabla u_k|^q-|\nabla u|^q|}{|x-y|^{n-2s}}\,dy \end{equation*} Now, by \eqref{ineq1} and \eqref{to represent} we get \begin{equation*} \frac{||\nabla u_k|^q-|\nabla u|^q|}{|x-y|^{n-2s}}\leq \frac{q\left|\nabla u_k-\nabla u \right|(\max\{|\nabla u_k|,|\nabla u|\})^{q-1}}{|x-y|^{n-2s}}\leq \frac{qC\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q}{|x-y|^{n-2s}} \end{equation*} In Proposition 2.12, we proved \begin{equation*} \mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)(x)\leq C(q,n,s)\left[ \mathcal{I}_{2s-1}(\omega)(x) + \|\mathcal{I}_{2s-1}(\omega)\|^q_{L^q(\mathbb{R}^n)}\right] \end{equation*}in $\mathbb{R}^n$. Therefore, $$\int_{\mathbb{R}^{n}}\frac{qC\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q}{|x-y|^{n-2s}}dy= C(n,q,s)\mathcal{I}_{2s}(\left[ \mathcal{I}_{2s-1}(\omega)\right] ^q)(x)$$ is finite, so we can apply Lebesgue's Theorem to obtain \begin{equation*} \lim_{k\to\infty}|\mathcal{I}_{2s}(|\nabla u_k|^q)-\mathcal{I}_{2s}(|\nabla u|^q)| = c(n,2s)\int_{\mathbb{R}^n}\lim_{k\to\infty}\frac{||\nabla u_k|^q-|\nabla u|^q|}{|x-y|^{n-2s}}\,dy = 0 \end{equation*} Hence, we get \eqref{represent}. \end{proof} \section{Proof of Theorem \ref{partial converse}}\label{converse sec} Suppose that \eqref{eq} has a solution $u \in W^{1, q}(\mathbb{R}^{n})$. Let $u_k \in \mathcal{C}^{\infty}_0(\mathbb{R}^{n})$ be a sequence converging to $\widetilde{u}$ in $W^{1, q}(\mathbb{R}^{n})$. Then for non-negative $\varphi \in \mathcal{C}^{\infty}_0(\mathbb{R}^{n})$ with $\varphi \geq \chi_E$ we have \begin{equation} \begin{split} \omega(E)& \leq \int_{\mathbb{R}^{n}}\varphi d\omega \\ & \leq \int_{\mathbb{R}^{n}}u(-\Delta)^{s}\varphi\quad \text{since $u$ is a solution}\\ & = \int_{\mathbb{R}^{n}}u_k(-\Delta)^{s}\varphi + o(1) \\ & = \int_{\mathbb{R}^{n}}(-\Delta)^{1/2}u_k(-\Delta)^{(2s-1)/2}\varphi+ o(1) \quad \text{ by Lemma 2.2 in \cite{CV1}}. \end{split} \end{equation}By continuity of the operator $(-\Delta)^{1/2}: W^{1, q}(\mathbb{R}^{n}) \to L^{q}(\mathbb{R}^{n})$ (see for instance \cite[Theorem 2.1]{DK}) and H\"{o}lder's inequality, we derive \begin{equation}\label{converse 1} \begin{split} \omega(E) & \leq \int_{\mathbb{R}^{n}}(-\Delta)^{1/2}u(-\Delta)^{(2s-1)/2}\varphi \\ & \leq \|(-\Delta)^{1/2}u\|_{L^{q}(\mathbb{R}^{n})} \|(-\Delta)^{(2s-1)/2}\varphi|\|_{L^{p}(\mathbb{R}^{n})} \\ & \leq C\|\nabla u\|_{L^{q}(\mathbb{R}^{n})} \| \varphi\|_{\mathcal{L}^{p}_{2s-1}(\mathbb{R}^{n})}, \end{split} \end{equation}where for $r \in (1, \infty)$ and $\alpha >0$, $\mathcal{L}_{\alpha}^{r}(\mathbb{R}^{n})$ denotes the Bessel potential space defined as $$\mathcal{L}_{\alpha}^{r}(\mathbb{R}^{n}):= \left\lbrace v \in L^{r}(\mathbb{R}^{n}): (I-\Delta)^{\alpha/2}v \in L^{r}(\mathbb{R}^{n}) \right\rbrace$$with the norm $$\| v\|_{\mathcal{L}_{\alpha}^{r}(\mathbb{R}^{n})}= \|v\|_{L^{r}(\mathbb{R}^{n})}+ \|(-\Delta)^{\alpha/2}v\|_{L^{r}(\mathbb{R}^{n})}.$$Since $\omega$ has compact support and the Riesz capacity of $E$ is zero, we obtain from \cite[Proposition 5.1.4 (b)]{AH} that the Bessel capacity of $E$, denoted by $ C_{2s-1, p}(E)$, is also zero. Then there is a sequence of non-negative and smooth functions $\varphi_k \in \mathcal{L}_{2s-1}^{p}(\mathbb{R}^{n})$ such that $$\varphi_k \geq \chi_E, \quad \varphi_k \to 0 \quad \text{in }\mathcal{L}_{2s-1}^{p}(\mathbb{R}^{n}).$$Applying \eqref{converse 1} to the sequence $\varphi_k$ and taking $k \to \infty$, we conclude $$\omega(E) =0.$$This ends the proof. \begin{remark} We now comment on the necessity of \eqref{cap1} for existence of solutions. We have not proved the converse of Theorem \ref{Theo1}, but we exhibit below how some arguments may be applied to find a relation between the measure $\omega$, the Riesz capacity and the solution $u$. We consider $\varphi$ as before and we take $\varphi^{p}$ as a test function to derive \begin{equation} \begin{split} \omega(E)& \leq \int_{\mathbb{R}^{n}}u_k(-\Delta)^{s}\varphi^{p}-\int_{\mathbb{R}^{n}}|\nabla u|^{q}\varphi^{p} + o(1) \\ & = \int_{\mathbb{R}^{n}}(-\Delta)^{1/2}u_k(-\Delta)^{(2s-1)/2}\varphi^{p}-\int_{\mathbb{R}^{n}}|\nabla u|^{q}\varphi^{p} + o(1) \end{split} \end{equation}Letting $k \to \infty$ and by Young's inequality, we get \begin{equation}\label{converse 11} \begin{split} \omega(E) & \leq \int_{\mathbb{R}^{n}}(-\Delta)^{1/2}u(-\Delta)^{(2s-1)/2}\varphi^{p}-\int_{\mathbb{R}^{n}}|\nabla u|^{q}\varphi^{p} \\ & \leq C_p\int_{\mathbb{R}^{n}}|(-\Delta)^{1/2}u| \varphi^{p-1}(-\Delta)^{(2s-1)/2}\varphi-\int_{\mathbb{R}^{n}}|\nabla u|^{q}\varphi^{p} \\ & \leq \int_{\mathbb{R}^{n}} |(-\Delta)^{1/2}u|^{q}\varphi^{p} + C_p \int_{\mathbb{R}^{n}}|(-\Delta)^{(2s-1)/2}\varphi|^{p} -\int_{\mathbb{R}^{n}}|\nabla u|^{q}\varphi^{p}\\ & = \int_{\mathbb{R}^{n}}\left( |(-\Delta)^{1/2}u|^{q}- |\nabla u|^{q}\right)\varphi^{p} + C_p \| \varphi\|^{p}_{\mathcal{L}_{2s-1}^{p}(\mathbb{R}^{n})}. \end{split} \end{equation} Taking the \'infimum over $\varphi$ in \eqref{converse 11}, we derive \begin{equation*} \omega(E) \leq \inf_{\varphi}\int_{\mathbb{R}^{n}}\left(|(-\Delta)^{1/2}u|^{q}- |\nabla u|^{q}\right)\varphi^{p} + C_p C_{2s-1, p}(E). \end{equation*}Since $\omega$ has compact support, we derive from \cite[Proposition 5.1.4 (b)]{AH} that there is a constant $C> 0$ so that \begin{equation}\label{est with inf} \omega(E) \leq \inf_{\varphi}\int_{\mathbb{R}^{n}}\left(|(-\Delta)^{1/2}u|^{q}- |\nabla u|^{q}\right)\varphi^{p} + C \text{cap}_{2s-1, p}(E). \end{equation}In the local case, exposed in \cite{Han}, the first term in \eqref{est with inf} does not appear and hence \eqref{cap1} is also necessary for existence. In the current scenario, we have not obtained the vanishing of that term. Indeed, adapting directly the proof of \cite[Lemma 2.1]{Han} to our case does not seem to be straightforward and the problem is left as open. \end{remark}

Introduction ============ Safety and regulatory concerns about silver nanoparticles (AgNPs) are growing due to their extensive use in consumer products and increasing potential of exposure to human and the environment \[[@b1-eht-29-e2014021]\]. According to the survey of Nanotechnology Consumer Products Inventory (<http://www.nanotechproject.org>), AgNPs are the most widely used engineered nanomaterials (ENMs). Approximately 25% of commercial products released onto the market contain AgNPs. The embryonic zebrafish assay is an appealing *in vivo* model in (eco)toxicology \[[@b2-eht-29-e2014021]\]. Zebrafish embryos are optically transparent, allowing for easy observations of phenotypic responses in the internal organs including the brain, jaw, eye, heart, yolk sac, notochord, trunk, and tail. Furthermore, molecular techniques, both established and in development, can be applied to assess toxicity of chemicals. More importantly, toxicological outcomes obtained from zebrafish may be extrapolated to human biology due to their homology and relevance to the physiological responses of mammals \[[@b3-eht-29-e2014021],[@b4-eht-29-e2014021]\]. The embryonic zebrafish provides considerable experimental flexibility and versatility in nanotoxicology \[[@b5-eht-29-e2014021]\]. To investigate toxic mode of action of AgNPs, in this study, we used zebrafish embryos that had developed normally and exposed them for 5 days in a test medium of low ionic strength where agglomeration was excluded \[[@b6-eht-29-e2014021]\]. We hypothesized that an attempt to control NP agglomeration (i.e., monodispersion of AgNPs) in the test media would enable us to determine the size- and surface coating-dependent toxicity of AgNPs accurately and reliably. In this study, we exposed dechorionated embryos to four different types of AgNPs to minimize chorion interference, because the chorion plays a role in protecting the embryo from external influences, including chemical exposure. Although the removal of the chorion is helpful in elucidating the mechanisms of toxicity of ENMs, questions concerning interference of the chorion in AgNP toxicity remain, because ecological toxicity testing has, in general, been performed in the presence of the chorion. Although no comprehensive investigation of chorion permeability or sensitivity to chemicals has yet been reported, the chorion is suspected to be a barrier to the entry of NPs into zebrafish embryos \[[@b7-eht-29-e2014021],[@b8-eht-29-e2014021]\]. The chorion possesses canals, the pore size of which is approximately 0.6-0.7 μm, larger than the size of the NP. Lee et al. \[[@b7-eht-29-e2014021]\] first visualized AgNP penetration through chorion canals. However, although the pore size is known to be larger than the NP size, in particular, the effect of the chorion on NP transport and subsequent biological toxicity may be complicated when NPs are agglomerated. The presence of the chorion is a critical concern on varied developmental toxicity of zebrafish embryos exposed to various formulations of AgNPs to understand the effect of size and surface coating \[[@b3-eht-29-e2014021],[@b9-eht-29-e2014021]\]. Consequently, the systematic examination of the changes in the biological responses of zebrafish embryos exposed to NPs in the presence or absence of the chorion is necessary. The objective of this study was to investigate how the size- and surface coating-dependent toxicity of AgNPs varied according to the presence or absence of the chorion in an embryonic zebrafish assay. We used 20 and 110 nm sized AgNPs, and AgNPs were coated with either polypyrrolidone (PVP) or citrate. The citrate and PVP are the most widespread and most commonly used stabilizer in AgNP synthesis. In order to determine the permeability differential of the AgNPs, we exposed normal and dechorionated embryos to four different AgNPs in a standard zebrafish embryo medium (EM) and 62.5 μM calcium chloride (CaCl~2~) solution. We observed that AgNPs were monodispersed in CaCl~2~ without agglomeration, and the exposure of zebrafish embryos in CaCl~2~ to AgNPs was useful for the observation of the variations in biological toxicity in the presence and absence of the chorion. This can be compared further to EM in which AgNPs were agglomerated. Materials and Methods ===================== Silver Nanoparticles and Characterization in Test Media ------------------------------------------------------- AgNPs synthesized by NanoComposix (San Diego, CA, USA) were used. The AgNPs were 20 nm PVP-coated AgNPs (20Ag-NPs-P), 20 nm citrate-coated AgNPs (20AgNPs-C), 110 nm PVP-coated AgNPs (110AgNPs-P) and 110 nm citrate-coated AgNPs (110AgNPs-C). This set of AgNPs was one of the ENMs selected by the Nanotechnology Health Implications Research Consortium, an interdisciplinary program of the National Institute of Environmental Health Sciences, which addresses the health and safety concerns of AgNPs. We followed the guidelines for sample handling and storage recommended by the Nanotechnology Characterization Laboratory (NCL). The NCL provided the physicochemical properties of AgNPs at a concentration of 100 mg/L suspended with 2 mM NaCl. The z-average and zeta potentials of four AgNPs were measured using dynamic light scattering (Malvern Zetasizer Nano ZS Instrument, Southborough, MA, USA) with a back scattering detector (173°). The core diameter was measured by transmission electron microscopy, also provided by the NCL (NCL-NIEHS201111A). In this study, we analyzed and compared the z-average of AgNPs in EM and CaCl~2~ by using a ZetaPALS (Brookhaven Instruments, Holtsville, NY, USA) after the preparation of 10-fold diluted aliquot from 1,000 mg/L of stock solution. Optimization of 62.5 μM Calcium Chloride ---------------------------------------- Because zebrafish embryos can develop normally in low ionic strength media \[[@b10-eht-29-e2014021]\], we tested media of various ionic strengths and found that incubation of embryos in 1 mM CaCl~2~ did not induce developmental morbidity or mortality at 120 hours post fertilization (hpf) \[[@b6-eht-29-e2014021]\]. To ascertain the minimum concentration of CaCl~2~, we carried out a titration, and determined that 62.5 μM CaCl~2~ was low sufficient for normal development until 120 hpf and nonagglomeration of AgNPs. Zebrafish Maintenance and Embryonic Zebrafish Assay --------------------------------------------------- Adult zebrafish (*Danio rerio*) were raised at Oregon State University's Sinnhuber Aquatic Research Laboratory in a water flow-through system that maintained a constant temperature of 28˚C under a 14:10 hours light-dark photoperiod. Embryos collected and staged \[[@b11-eht-29-e2014021]\] were from the tropical 5D zebrafish line and were cultured in buffered EM (pH 7±0.2) using reverse osmosis water. The standard EM consisted of 15 mM NaCl, 0.5 mM KCl, 1 mM CaCl~2~, 1 mM MgSO~4~, 0.15 mM KH~2~PO~4~, 0.05 mM Na~2~HPO~4~ and 0.7 mM NaHCO~3~ \[[@b12-eht-29-e2014021]\]. Healthy staged embryos were loaded into 96-well plates filled with 90 μL of test medium and 10 μL of the AgNP suspension. To achieve the final concentrations of 0.8, 4, 20, 10, and 50 mg/ L, after a 10×concentrated, serial diluted concentration of Ag- NPs was prepared, 10 μL of the AgNP suspension was mixed into the respective wells. Exposure to AgNPs was initialized at 6 hpf. Eight embryos were exposed to each treatment with four replicates. We evaluated a total of 22 endpoints including mortality and malformation at 24 and 120 hpf. The percentage of malformation was calculated by the number of embryos affected per living embryos at each treatment. Screened malformations were: yolk sac edema; axis, eye, snout, jaw, and otic (ear) defects; pericardial edema; brain, somite, and pectoral and caudal fin defects; pigmentation abnormality; circulation and trunk defects; swim bladder abnormality; notochord defects, and touch response abnormality. Using pronase (Sigma-Aldrich, St Louis, MO, USA), we removed the chorion in an automated dechorinator at 4 hpf. Briefly, after introducing pronase, the plate was shaken for 5 minutes, washed and then the shaking procedure was repeated. While dechorionated embryos were placed into each well using an automated embryo placement systems, normal embryos (i.e., undechorionated embryos) were transferred into a 96-well plate by hand. The ability to produce healthy embryos was confirmed in ae previous study \[[@b13-eht-29-e2014021]\]. No detectable mortality or morbidity of dechorionated embryos in the control group was observed over 5 days in this study. Results ======= Characterization of Silver Nanoparticles ---------------------------------------- We determined the z-average of each AgNPs in EM and CaCl~2~, and compared them to those provided by the NCL ([Figure 1](#f1-eht-29-e2014021){ref-type="fig"}). We noted that the stock solution of 1,000 mg/L AgNPs provided by NanoComposix was suspended in 2 mM NaCl, where the NCL measured the reference of the z-average. Depending on the AgNPs, the z-average increased dramatically up to micro-size in EM. The z-average of 20AgNPs-P, 20AgNPs-C, 110AgNPs-P, and 110AgNPs-C increased to 69.7±2.7, 3078.4 ±400.7, 691.8±40.3, and 1400.5±91.2 nm (mean±standard deviation), respectively. Although it appeared that the citrate surface coating caused more agglomeration, it is not clear why the diameter of the 20AgNPs-C was larger than that of the 110AgNPs-C. Further, the z-average of AgNPs in CaCl~2~ were totally different from those in EM. All AgNPs were stably dispersed (i.e., no increas in z-average), the z-average measured did not differ significantly from those provided by the NCL ([Figure 1](#f1-eht-29-e2014021){ref-type="fig"}). The z-average of 20AgNPs-P, 20AgNPs-C, 110AgNPs-P, and 110AgNPs-C were determined to be 30.5±1.2, 26.1±0.8, 123.2±0.9, and 100.4±2.1 nm, respectively. The suspension of 20AgNPs and 110AgNPs exhibited a bright and slightly darker yellow color, respectively, but their color turned to black when transferred into EM, indicating agglomeration, whereas the color did not change in CaCl~2~. These z-average corroborate UV-vis spectra analysis (data not shown). The absorbance of AgNPs was maintained over 5 days in CaCl~2~ while a drop in absorbance was observed in EM. After 5 days, the absorbance of 20AgNPs-P was attenuated and that of 20AgNPs-C, 110AgNPs-P, and 110AgNPs-C disappeared completely in EM, indicating NP agglomeration. The NCL report stated that all AgNPs exhibited negative zeta-potentials (NCL-NIEHS201111A). Because there is no consistent correlation between zeta-potential and either z-average or UV-vis absorbance spectra in 2 mM NaCl where agglomeration is controlled, zeta-potentials were not measured in EM and CaCl~2~ used in this study. The electrostatic repulsion might not be dominant in maintaining AgNP stability in suspension in comparison with the interfacial forces \[[@b14-eht-29-e2014021]\]. Taken together, AgNPs were dispersed stably in CaCl~2~. We hypothesized that the exposure of nonagglomerated AgNPs helped us to investigate the effect of the chorion on AgNPs with variable size and surface coating. Embryonic Toxicity of Silver Nanoparticles ------------------------------------------ The [Figures 2](#f2-eht-29-e2014021){ref-type="fig"} and [3](#f3-eht-29-e2014021){ref-type="fig"} show the total percentage of mortality and malformation of embryos exposed to four AgNPs in the presence and absence of the chorion in EM and CaCl~2~. A total of 18 endpoints evaluated at 120 hpf was used for the toxicity differential of each treatment. The black color indicates the mortality at 120 hpf over the range of AgNP exposure concentrations (0, 0.08, 0.4, 2, 1,0 and 50 mg/L). The percent incidence of malformation indicates the percentage of surviving embryos that exhibited any of the 17 morphological malformations described above. Embryonic toxicity to AgNPs suspended in EM was significantly lower than those suspended in CaCl~2~, probably due to the low bioavailability that resulted from strong agglomeration, which was expected from the increase in the zaverage and the disappearance of UV-vis absorbance in EM. A previous study showed that non-agglomerated AgNPs accumulated to a greater degree in embryos \[[@b6-eht-29-e2014021]\]. Typically, embryonic toxicity in the absence of the chorion was higher than in the presence of the chorion. Embryos that were counted as malformed in the presence of the chorion showed only one malformation, for example, mainly yolk sac edema. This indicates that the chorion definitely serves as a barrier to AgNP exposure and contact with the embryos. A quantitative study of the uptake of AgNPs according to the presence and absence of the chorion is needed. [Figure 4](#f4-eht-29-e2014021){ref-type="fig"} illustrates examples of embryos exposed to each AgNPs for 24 and 120 hpf. The greatest embryonic toxicity was observed at the highest concentration of 50 mg/L in both test media, regardless of the presence of the chorion. Below concentrations of 50 mg/L, a very low incidence of mortality and malformation was observed. In the presence of the chorion, embryonic death or decay without hatching was typical type of mortality observed at 50 mg/L. Below concentrations of 50 mg/L, no malformations were caused by any of the AgNPs in the hatched larvae. In the absence of the chorion, 20AgNPs-P and 20AgNPs-C caused some mortality at 10 mg/L, and approximately 70% of the embryos showed malformation of the yolk sac edema in 110AgNPs-P at 10 mg/L. Malformations observed in the absence of the chorion were jaw and eye abnormalities, pericardial and yolk sac edema, and snout and circulation defects. Discussion ========== Based on the percentage of mortality and malformation, we ranked the embryonic toxicity of AgNPs in EM in the presence of the chorion as: 20AgNPs-P\>20AgNPs-C\>110AgNPs-C ≈ 110AgNPs-P and in the absence of the chorion as 20AgNPs-C\>20AgNPs-P\>110AgNPs-C\>110AgNPs-P. The incidence of mortality was considered to be more toxic than malformation. The presence of the chorion changed the order of toxicity with respect to the surface coating rather than the size. It is not clear, however, that 20AgNPs-C with a larger z-average in EM were more toxic than 110 nm AgNPs, regardless of the presence of the chorion. In CaCl~2~, the embryonic toxicity of AgNPs was ranked as 20AgNPs-P\>20AgNPs-C\>110AgNPs-P ≈ 110Ag-NPs-C and 20AgNPs-P \>20AgNPs-C \>110AgNPs-P\>110AgNPs-C in the presence and absence of the chorion, respectively. We noted that in the environment in which particle agglomeration (i.e., in CaCl~2~) and chorion interference (i.e., in the absence of the chorion) were controlled, PVP was more toxic than the citrate coating, regardless of the particle size, as shown above. In CaCl~2~, where agglomeration was controlled, a change in the order of toxicity also occurred, predominantly regarding the surface coating in the presence of the chorion. Consequently, the presence of the chorion did influence the interaction with the chemical surface coating material, rather than the physical size effect in the embryonic toxicity of AgNPs, regardless of the test medium. A further study comparing the effect of chorion on AgNPs with sizes between 20 and 110 nm AgNPs is needed. To our knowledge, this is the first report of a systematic comparison of ENM toxicity with respect to chorion permeability. As demonstrated, the size- and surface coating-dependent toxicity of AgNPs was controlled by the function of embryonic permeability. Embryonic permeability was affected by particle agglomeration, and primarily caused a differential in toxicity dependent on the surface coating. The embryonic zebrafish assay is a well-established model not only for rapid and cost-efficient toxicity screening \[[@b15-eht-29-e2014021],[@b16-eht-29-e2014021]\], but also for the identification of molecular mechanisms \[[@b17-eht-29-e2014021]\] in assessing ENM toxicity. However, knlowedge with regard to the permeability limitations and secondary phenotypic responses to the chorion related to the hatching process in the presence of the chorion \[[@b13-eht-29-e2014021]\] is lacking in nanotoxicology. Therefore, further experimental investigations of the effect of the chorion should be carefully considered in assessing the toxicity of ENMs and other chemicals in the embryonic zebrafish assay. This study was supported by the Research Program funded by Seoul National University of Science and Technology. The authors have no conflicts of interest with material presented in this paper. {#f1-eht-29-e2014021} {#f2-eht-29-e2014021} {#f3-eht-29-e2014021} {#f4-eht-29-e2014021}

INTRODUCTION ============ Tropical forests have long been recognized as one of the largest pools of biodiversity ([@R1]). Global patterns of empirical abundance distributions show that tropical forests vary in their absolute number of species but display surprising similarities in the distribution of individuals across species ([@R2]--[@R4]). For practical reasons, biodiversity is typically measured or monitored at fine spatial scales. However, important drivers of ecological change tend to act at large scales ([@R5], [@R6]). Conservation issues, for example, apply to diversity at global, national, or regional scales. Extrapolating species richness from the local to the whole-forest scale is not straightforward. A vast number of different biodiversity estimators have been developed under different statistical sampling frameworks ([@R7]--[@R11]), but most of them have been designed for local/regional-scale extrapolations, and they tend to be sensitive to the spatial distribution of trees ([@R12]--[@R14]), sample coverage, and sampling methods ([@R15]). A common statistical tool used to describe the commonness and rarity of species in an ecological community is the relative species abundance distribution (SAD or RSA), which is a list of species present within a region along with the number of individuals per species ([@R16], [@R17]). Typically, the SAD is measured at local scales (for example, in quadrats or transects; see [Fig. 1](#F1){ref-type="fig"}), in which the identities of the individuals living in the area are known. The sampled SAD can be fit to a given functional form at that scale. However, that form may change at different spatial scales, thus hindering analytical treatment ([@R18]). Nonparametric approaches have also been proposed in the literature to infer species richness. Instead of assuming a specific functional form for the SAD and fitting data to arrive at the parameters, these methods are based on the intuitive idea that it is only the rare species that carry information on the undetected species in a sample. A successful example is the method introduced by Chao *et al*. ([@R15], [@R19], [@R20]), which takes into account only the number of singletons and doubletons (species with just one or two individuals) observed at the sample scale to infer the species richness of the whole forest. {ref-type="table"} and section S6 for a more detailed description of the data set).](1701438-F1){#F1} Recently, a semianalytical method to upscale species richness based on a log series (LS) for the SAD has been proposed (section S1 and [Fig. 2](#F2){ref-type="fig"}) ([@R21]--[@R25]). The LS distribution was obtained by Fisher et al. ([@R26]) as the limiting form of a negative binomial (NB) probability distribution (that is, the probability of observing *n* individuals when sampling from a population belonging to different species), excluding zero observations (no information on the number of missed species is available) and assuming that the distribution of individuals is known and simple (that is, Eulerian form). The LS distribution is often used to describe SAD patterns in ecological communities, including tropical tree communities. The robustness of the upscaling method relies on the stability property of Fisher's α \[approximately reflecting the number of observed singleton species ([@R26])\], which ought not to depend on the forest sample size and is given by$$\frac{\mathit{N}_{\mathit{p}}}{\alpha} = (\mathit{e}^{\mathit{S}_{\mathit{p}}/\alpha} - 1)$$where *N*~*p*~ and *S*~*p*~ are the total number of individuals and species, respectively, when sampling a fraction *p* of the forest (*N*~1~ = *N* and *S*~1~ = *S* corresponds to the total number of individuals and species when sampling the whole forest). Therefore, the LS method is composed of three main steps: (i) Fisher's α is calculated, assuming that the species have an LS distribution (see Materials and Methods) and using the observed species *S*~*p*~ and number of trees *N*~*p*~ as input. (ii) The total number of stems *N* for the whole area of interest is extrapolated \[This is not a trivial task, and there is no consensus on the best methods to implement it. Generally, constant average stem density is assumed ([@R24], [@R25]).\]. (iii) The number of species at the largest scale is estimated using the formula *S* = α ln (1 + *N*/α) ([@R26]). This method has been used to estimate the species richness of the Amazonia ([@R24]) and that of global tropical forests ([@R25]). For the latter case, Slik *et al*. ([@R25]) noted that when merging forests in different tropical regions, the value of Fisher's α shows an asymptotic behavior for large areas, as if it is converging to its asymptote for each region. From this limiting value, it is then possible to infer the total species richness of the different tropical regions. {ref-type="disp-formula"} and [15](#E15){ref-type="disp-formula"}, we predict the biodiversity *S*~pred~ of the whole forest.](1701438-F2){#F2} On the basis of theoretical and computational analysis as well as using the data from 15 tropical forests located all over the globe, we show that the LS method suffers from important limitations (see section S2, figs. S1 and S2, and table S1). Often, the SAD---especially at large scales or with increasing sampling effort ([@R27])---displays an interior mode ([@R14]), which an LS cannot capture. The Fisher's LS is not flexible enough ([@R18]) to describe different SAD patterns ([@R14], [@R17], [@R28]--[@R32]) found in tropical forests. Here, we present a more general analytical framework to extrapolate species richness from local to whole-forest scales. This framework, derived from first principles on the basis of biological processes, is based on the fact that the functional form of any given SAD can be approximated to any degree of accuracy with a linear combination of NB distributions (see Materials and Methods), as long as the population sizes are smaller than some fixed, but otherwise arbitrary, threshold, as suggested by Nachbin's theorem (see section S3) ([@R33], [@R34]). We will show that our method outperforms previously proposed methods and that the LS method turns out to be a special case of our framework. RESULTS ======= Theoretical framework --------------------- The NB distribution arises naturally as the steady-state SAD of an ecosystem that undergoes simple birth-and-death dynamics, with an effective birth rate accounting for the effects of immigration events and/or intraspecific interactions ([@R2], [@R14]), and under the neutral hypothesis that individuals are demographically identical (see Materials and Methods) ([@R31]). This distribution is able to adequately fit the SADs of diverse ecosystems, such as tropical forests and coral reefs ([@R14], [@R31]). In particular, [Eq. 2](#E2){ref-type="disp-formula"} below is the steady-state solution of the master equation governed by birth and death rates (see Materials and Methods). The continuum version of the NB (that is, the γ distribution) is also the stationary state of a model that captures the temporal turnover of species ([@R35]), an important aspect of tropical tree dynamics ([@R36]). A single NB SAD is given by$$\mathcal{P}(\mathit{n}|\mathit{r},\xi) = \frac{1}{1 - {(1 - \xi)}^{\mathit{r}}}\begin{pmatrix} {\mathit{n} + \mathit{r} - 1} \\ \mathit{n} \\ \end{pmatrix}\xi^{\mathit{n}}{(1 - \xi)}^{\mathit{r}}$$which is normalized so that $\sum_{\mathit{n} = 1}^{\infty}\mathcal{P}(\mathit{n}|\mathit{r},\xi) = 1$, where *r* \> 0 and 0 ≤ ξ \< 1 are the parameters accounting for immigration or intraspecific interactions and the ratio between the birth and death rates, respectively (see Materials and Methods). Fisher's LS is obtained as the *r* → 0 limit of [Eq. 2](#E2){ref-type="disp-formula"}. Owing to partial sampling, the empirical SAD of a small sample of a forest will likely show a monotonic decreasing behavior, because these samples contain many rare species with just a few individuals. However, a relatively larger sample may exhibit an internal mode, because relatively rare species are not found as the sampling effort increases (this happens, for example, if the SAD at the whole-forest scale is well described by a log-normal). Both situations are well captured by the NB distribution, whose functional form can accommodate both shapes, depending on the value of its different parameters ([Fig. 3](#F3){ref-type="fig"}). When extrapolating to larger spatial scales (upscaling), a single NB distribution ([Eq. 2](#E2){ref-type="disp-formula"}) retains the same value of the parameter *r*---so we say that *r* is scale-invariant---whereas the parameter ξ depends on the sampling scale (see Materials and Methods). The same holds true for a linear combination of NB distributions with different values of *r* and the same ξ (see Materials and Methods). ![Versatility of the NB distribution.\ The NB distribution is a two-parameter distribution that shows self-similarity and can display both monotonic LS-like behavior (in the limit *r*→0, the NB tends to the LS distribution) and a unimodal shape, as a function of the scaling parameter **ξ**. The red curve represents the analytical threshold separating these two cases. The SAD, especially at large scales or with increasing sampling effort ([@R27]), often displays an interior mode that cannot be captured by the LS distribution but can be described by the NB. The NB distribution naturally arises as the steady-state SAD of an ecosystem undergoing generalized dynamics of birth, death, speciation, and migration processes (see Materials and Methods). Finally, any discrete probability distribution, such as the SAD, can be approximated to any degree of accuracy by a suitable linear combination of NBs that retains the self-similarity feature (see Materials and Methods). An example is shown of how the parameter **ξ** of the NB increases as the area of the forest doubles. Starting from **ξ** = 0.36, as the area doubles, the **ξ** value moves upward to the value corresponding to the successive (dashed) horizontal line in the upward direction.](1701438-F3){#F3} We formulate our analytical framework on the basis of the following two steps: (i) Sample a fraction *p*\* of the whole forest and then obtain the vector, $\mathit{n}_{\mathit{p}*} = \{\mathit{n}_{1},\mathit{n}_{2},...,\mathit{n}_{\mathit{S}^{*}}\}$, of the abundances of the *S*\* sampled species. (ii) Use a linear combination of a suitable number of NBs with the same ${\hat{\xi}}_{\mathit{p}*}$ and different values of *r* to fit the empirical SAD at the desired degree of accuracy. This method is guaranteed to be effective according to Nachbin's theorem (see section S3 and figs. S3 and S4) ([@R33], [@R34]). The NB does not change its functional form when sampling different fractions of areas---that is, distribution form invariance under different sampling efforts---although the parameters of the distribution do change. More precisely, the NB at different scales has the same *r* parameters, but different ξ, which is a function of the scale (see Materials and Methods). Thus, we obtain an analytical expression of the upscaled SAD at scale *p* from the data at scale *p*\* in terms of the equation ${\hat{\xi}}_{\mathit{p}} = \mathit{U}(\mathit{p},\mathit{p}*|{\hat{\xi}}_{\mathit{p}*})$, defining ${\hat{\xi}}_{\mathit{p}}$ in terms of *p*, *p*\*, and ${\hat{\xi}}_{\mathit{p}*}$ (see Materials and Methods). Using the SAD at scale *p*\*, a maximum likelihood method is used to estimate the parameters of the SAD, and the upscaling equations (see Materials and Methods) are used to predict the species richness of the entire forest, that is, *p* = 1. In particular, we found that the total number of species *S* at the largest scale (*p* = 1) is related to the number of species at scale *p*, *S*~*p*~, by the following relation (see section S1 for detailed calculations) $$\mathit{S} = \mathit{S}_{\mathit{p}}\frac{1 - {(1 - \xi)}^{\mathit{r}}}{1 - {(1 - \xi_{\mathit{p}})}^{\mathit{r}}}$$where ξ~*p*~ and *r* are the NB-fitted parameters of the SAD at scale *p*. As noted above, *r* is scale-invariant and hence independent of *p*, whereas the parameter ξ at the largest scale, *p* = 1, is given by$$\xi = \frac{\xi_{\mathit{p}}}{\mathit{p} + (1 - \mathit{p})\xi_{\mathit{p}}}$$ The framework resembles the renormalization group technique in critical phenomena in which the behavior of a system at different scales is described in terms of equations for the model parameters, similarly to what has been suggested here ([@R37]). By using our framework (that we denote as the NB framework in the following sections), we were able to generate accurate and robust predictions for computer-generated forests and for 15 empirical tropical forests ([Fig. 1](#F1){ref-type="fig"} and [Table 1](#T1){ref-type="table"}). ###### Predicting the biodiversity in tropical forests. Predicted total number of species, *S*~pred~, at the whole-forest scale (corresponding to *p* = 1) for each of the 15 tropical forests in our database. Predictions are determined by using information on the sampled scale *p*\* (fourth column), where we observe *N*\* trees belonging to *S*\* species (second and third columns). In the fifth column, we show the predictions obtained by using the NB framework with a single NB for fitting the sampled SAD. SEs were computed by propagating the errors in the fitting parameters of the SAD (obtained by the bootstrapping method) and of *S*\*. The latter has been determined as follows: For each data set, we created the corresponding predicted forest at the scale *p* = 1 by generating *S*~pred~ numbers distributed according to an NB with parameters (*r*, ξ). We then sampled the *p*% of the list of individuals, as in the original data. The last two columns show the predictions of the LS and Chao methods. ------------------------------------------------------------------------------------------------------- **Forest** ***S*\*** ***N*\*** ***p*\*%** ***S*~pred~**\ ***S*~pred~**\ ***S*~pred~**\ **(NB)** **(LS)** **(Chao)** --------------- ----------- ----------- ------------ ---------------- ---------------- ---------------- Amazonia 4962 553949 0.00016 13602 ± 711 14984 5561 Barro\ 301 222602 3.20513 366 ± 15 419 315 Colorado Bukit\ 340 14974 0.00169 471 ± 40 1020 346 Barisan Bwindi 128 18490 0.01813 163 ± 15 288 129 Caxiuana 386 32701 0.01818 437 ± 14 915 386 Cocha Cashu 489 16640 0.00035 731 ± 63 1674 501 Korup 226 17427 0.00473 282 ± 23 591 226 Manaus 946 38933 0.06000 1016 ± 14 2242 956 Nouabalé-\ 110 7196 0.00143 125 ± 8 316 110 Ndoki Pasoh Forest\ 927 310520 0.35714 1193 ± 36 1590 1049 Reserve Ranomafana 269 34580 0.01463 336 ± 22 620 269 Udzungwa 109 18447 0.00302 146 ± 20 269 114 Volcan Barva 392 44439 0.02025 448 ± 16 895 395 Yanachaga 209 2041 0.00372 802 ± 211 802 259 Yasuni 481 13817 0.61100 565 ± 20 974 484 ------------------------------------------------------------------------------------------------------- Test on in silico forests ------------------------- We first compared the results of our method applied to a computer-generated forest. In this in silico experiment, we fixed the number of species (*S* = 5000) and their abundance distribution a priori and then generated the forest accordingly. Species abundances were extracted from a log-normal SAD of mean, μ = 5, and SD, σ = 1, and the individual trees were located according to a modified Thomas process (see fig. S5 and section S4) ([@R38], [@R39]) with two distinct clustering coefficients (high and low clustering). The log-normal SAD, originally proposed by Preston ([@R40]), has been used to fit the SAD of several tropical forests ([@R14], [@R41]), whereas Thomas cluster models have reproduced empirical species-area curves with high fidelity ([@R12], [@R42]). We then sampled nonoverlapping 1-unit plots at randomly chosen locations covering only a small fraction, *p*\* = 5 %, of the area and attempted to predict *S* using only this partial information. We performed the estimation of the total species richness of the computer-generated forest by using a single NB distribution or a linear combination of two NB distributions, the LS method and the Chao estimator, based on sampling without replacement (see section S5 and table S3). For both clustering regimes, the prediction of the number of species using the NB framework with just one NB was already very good (error \< 2%; high clustering, *S*~pred~ = 5095; low clustering, *S*~pred~ = 5067). The linear combination of two NBs increased the accuracy of the prediction at the whole-forest scale *p* = 1 (with two parameters, we obtained the following values: error \< 0.2 %; high clustering, *S*~pred~ = 4995; low clustering, *S*~pred~ = 5011). Chao's method gave results comparable to those with one NB (error \< 2%; high clustering, *S*~pred~ = 4938; low clustering, *S*~pred~ = 4931) while underestimating the true number of species instead of overestimating it. In contrast, the LS method strongly overestimated the number of species (error \> 56 %; high clustering, *S*~pred~ = 7838; low clustering, *S*~pred~ = 9036). We thus found that although the original forest had a log-normal SAD entangled with spatial correlations, a single NB or a linear combination of two NBs led to surprisingly good predictions and systematically outperformed the LS method; this result was also true for a computer-generated forest with an NB SAD and when a different sampling method was performed, consisting of collecting data within a unique spatial window covering the same percentage of the whole forest area (section S4 and table S2). Finally, we compared the results for an in silico LS forest. As expected, in this case, the LS method performed very well, predicting a species richness of 4930 against the true value of 5000 (error \~ 1.3%). The very same result was obtained by using the NB method. The best fit of the SAD with an NB led to an *r* parameter very close to zero (*r* \~ 10^−5^), so that the NB distribution was effectively converging to an LS. In contrast, the Chao method underestimated the number of species giving a prediction of 3878 (error \~ 22%). Previous results have shown ([@R43]) that the Chao estimator for upscaling species richness based on sampling with replacement perform poorly in hyperdiverse communities with many rare species. Here, we found that the very same result holds for the estimator based on sampling without replacement, an assumption consistent with the way empirical forests are sampled. Test on empirical data ---------------------- To test the accuracy of our method on more realistic distributions of trees (for example, habitat heterogeneity, species spatial distributions, etc.), we used subsamples taken from empirical forest data (see section S6 and table S6) and predicted the number of species at the corresponding largest empirically observable scale. That is, we extracted a fraction *p* of the data and applied our framework to infer the number of species at the scale *p*\*. Moreover, we compared our results to those obtained with other methods to upscale species richness and abundances, previously proposed in the literature (see tables S3 to S5 and fig. S6) ([@R19]--[@R21], [@R25]). We found that our method outperforms that of Chao and Chiu ([@R19], [@R20])---which typically overestimates the forest species richness---for Amazonia, Pasoh, and Yasuni ([Fig. 4](#F4){ref-type="fig"}). For the remaining forests, the NB method performed better than the LS method, which overestimates the number of species at *p*\*, and it was comparable to Chao's (see section S5 for a detailed discussion). However, we remark that the accuracy in Chao's predictions is due to the fact that, when sampling these forests at small scales, we found a low number of singleton and doubleton species. Therefore, Chao and Chiu ([@R19], [@R20]) conservatively gave the number of species at the observation scale itself as output, that is, $\mathit{S} \approx \mathit{S}_{\mathit{p}^{*}}$ (see section S5). This limitation is evident in [Fig. 4](#F4){ref-type="fig"}, which shows the tropical forest species area relationship (SAR), that is, the number of observable species as a function of the fraction of the sampled area *a*, (*p*\* ≤ *a* ≤ 1). Whereas LS and NB show the expected qualitative behavior, the method of Chao saturates almost immediately at *a* ≈ *p*\*, which is clearly an artifact of the method. The same results were obtained when using Chao's estimator based on sampling with replacement ([@R43]). {#F4} Biodiversity upscaling of tropical forest data ---------------------------------------------- After testing our model on controlled computer-generated data and real forest subsamples, we applied our framework to predict the species richness and abundances of tropical forest data. Because of the good agreement between the predictions made with a single NB for the artificial forests, we chose to work again with a single NB. Such a form can be derived from basic ecological processes ([@R14], [@R31]), and it also permits an exact analytical treatment of the upscaling protocol. Although in few cases, using more than one NB improves the accuracy of the predictions, in general, it increases the likelihood that the empirical data are overfit at the sampled scale. Therefore, through the NB method, we attempted to predict the species richness at the whole-forest scale (*p* = 1) for each of the 15 tropical forests around the equatorial zone, and we compared our predictions with those of previous results based on the LS distribution ([@R24], [@R25]) and with that obtained with the method of Chao. We found that the LS method systematically led to higher estimates of the number of rare species and consequently of the forest species richness at the largest scale (see [Table 1](#T1){ref-type="table"}). Only for the Yanachaga Chemillén National Park, the two estimates with NB and LS were essentially the same. The discrepancies in the estimates increased to approximately 10% for Amazonia and Barro Colorado Island (BCI), reached 30 to 40% for Pasoh and Bukit Barisan and ranged between 72 and 152% for the remaining 10 forests. In contrast, Chao's method predicted a much smaller number of species at the whole-forest scale. The errors in our estimates are also given in [Table 1](#T1){ref-type="table"}. Our framework is also able to give a quantitative estimate of the sampling effort (*p*~pred~ %; first column in [Table 2](#T2){ref-type="table"}) needed to achieve species richness predictions with error bars below approximately 5% (this percentage was arbitrarily chosen as an illustration, and our approach can be straightforwardly used for any other percentage of error). These estimates have been obtained through Monte Carlo simulations, which test the self-consistency of the NB method and allow us to infer these critical sampling thresholds (see section S7 and figs. S7 and S8). We found that for some forests (BCI, Caxiuana, Manaus, Volcan Barva, and Yasuni), the present sampling effort may be sufficiently informative and representative to characterize the biodiversity of the whole forest. In contrast, we propose an estimate of the further sampling required for all the other forests ([Table 1](#T1){ref-type="table"}). Amazonia, for example, would need approximately twice the current amount of sampling; Cocha and Nouabalé would need approximately 10 times; and Bwindi, Udzungwa, and Yanachaga would need several hundred times the current sampling (see the third column of [Table 2](#T2){ref-type="table"} showing the ratio between the predicted needed sampling and the actual one). ###### Sampling targets for forest percentage cover. Using our results on upscaled forest species richness, it is possible to estimate the percentage *p*~pred~% of the forest that must be sampled to achieve an estimation error of approximately 5% with a certainty of 95%. We derived these values by creating the predicted forest at the whole-forest scale (we generated *S*~pred~ numbers according to an NB with parameters *r* and **ξ**) and sampled it at increasingly larger scales until the desired accuracy in the estimation of the global species richness was reached (see section S7 for more details). The last column indicates how much extra sampling is needed (if the number is greater than 1) to reach 5% precision. **Forest** ***p*~pred~%** ***p*~pred~/*p*\*** ---------------------- ---------------- --------------------- Amazonia 0.0003 1.875 Barro Colorado 3 1 Bukit Barisan 0.05 18 Bwindi 5 386 Caxiuana 0.01 0.55 Cocha Cashu 0.003 8.57 Korup 0.02 1.06 Manaus 0.02 0.17 Nouabalé-Ndoki 0.015 10.5 Pasoh forest reserve 0.5 1.4 Ranomafana 0.1 6.84 Udzungwa 1.5 497 Volcan Barva 0.02 0.25 Yanachaga 1 269 Yasuni 0.3 0.49 We also estimated the number of hyper-rare species, defined as species with fewer than 1000 individuals, and the number of hyperdominant species, defined as the most abundant species contributing approximately 50% to the total number of individuals of the forest (see [Table 3](#T3){ref-type="table"}) ([@R24]). ###### Fisher's paradox. Hyper-rare species \[defined as species with fewer than 1000 individuals ([@R24], [@R25])\] and hyperdominant species (the most abundant species, accounting for ≈ 50 % of the total number of individuals) percentages were predicted in the whole area of each tropical forest obtained by applying both the NB and LS methods. We found that by using our NB method, the number of hyper-rare species in most of the forests was drastically reduced with respect to the LS method, thus suggesting that the extremely high value of hyper-rare species predicted in previous studies ([@R24], [@R25]) is an artifact of the LS method. Nevertheless, we found that the hyper-rarity phenomenon is a genuine emergent pattern in tropical forests. **Forest** **Hyper-rare (%)** **Hyperdominant (%)** ----------------------------------- -------------------- ----------------------- ------ ----- Amazonia 33 37 2.2 2.0 Barro Colorado Nature Monument 47 60 5.5 4.8 Bukit Barisan 22 46 7.9 1.9 Bwindi Impenetrable Forest 15 48 7.4 3.5 Caxiuana 6 49 10.3 3.2 Cocha Cashu Manu National Park 7 41 8.4 2.5 Korup National Park 9 51 9.3 3.1 Manaus 6 59 14.5 2.8 Nouabalé-Ndoki 4 43 11.2 2.4 Pasoh Forest Reserve 34 55 6.5 3.1 Ranomafana 12 49 7.5 2.7 Udzungwa Mountain National Park 12 48 6.3 3.0 Volcan Barva 8 52 10.5 2.5 Yanachaga Chemillén National Park 54 56 3.0 2.7 Yasuni National Park 39 74 11.6 4.4 DISCUSSION ========== Our analysis shows that hyper-rarity, as also suggested by previous works ([@R24], [@R25]), is a recurrent pattern in large-scale tropical forests, which may suggest that these tropical forests are biodiversity hot spots (see also discussion below) ([@R44]). Focusing on Amazonia, we predict that roughly 4500 Amazon tree species are hyper-rare. If they could be found and identified, then this would automatically qualify them for inclusion in the International Union for Conservation of Nature's Red List of Threatened Species. The NB upscaling for the entire Amazon forest predicts that half the total number of trees belong to just 300 hyperdominant species, whereas 33% of the 13,602 tree species are hyper-rare. In this way, ecologists would have an estimate of how many Amazon tree species face the most severe threats of extinction. These rare species in the Amazon forest (and our planet's biodiversity) are like dark matter in cosmology, which accounts for much of the universe. Nevertheless, in most of the forests, we obtained a smaller number of hyper-rare species and a higher number of hyperdominant ones with respect to previous estimates ([@R24], [@R25]). This result is in agreement with the tests we performed both in silico and on empirical forest data. We believe that this is due to the fact that the asymptotic value of Fisher's α in the LS method is strongly biased when a very small fraction of the forest is sampled (typically \< 1%) (section S2). As well as being a crucial and practical measure of fragile biodiversity in conservation ecology, hyper-rarity is also an important theoretically intriguing and open question that goes under the name of the "Fisher paradox" ([@R43], [@R45]). We still do not know why there is such a huge separation of population size scales between rare and hyperdominant species. Our framework provides a possible interpretation for this phenomenon and suggests that hyper-rarity could be a manifestation of criticality in tropical forests ([@R37], [@R46]). The parameters of the NB distributions that provided the best predictions of the upscaled species richness in tropical forests fall within a tiny region of parameter space: 0 \< *r* \< 0.7 and ξ ≈ 1. This result is surprising, because there are neither theoretical nor biological reasons why tropical forests should have their parameters localized within such a narrow region, especially when considering that they are in completely different geographical regions with differing evolutionary histories. However, a closer examination of the form of the NB distribution reveals that the relative fluctuation of abundances, that is, $\sqrt{\langle{(\mathit{n} - \langle\mathit{n}\rangle)}^{2}\rangle}/\langle\mathit{n}\rangle$, diverges as ξ → 1 and *r* → 0 (see [Fig. 5](#F5){ref-type="fig"} and section S8). Thus, parameter values in the vicinity of this region allow an ecosystem to have the highest heterogeneity in its abundance distribution. The points shown in [Fig. 5](#F5){ref-type="fig"} correspond to the parameter values obtained for the 15 forests. A physical system, such as water and vapor, in the vicinity of its critical point, is characterized by density fluctuations that become very large, with droplets of water and bubbles of gas of all sizes thoroughly interspersed, and the system appears the same at different scales (that is, it is self-similar) ([@R37]). This scale invariance confers to the system an acute sensitivity to certain types of external perturbations or disturbances whose effects are realized at long distances. The observed large abundance fluctuations suggest that tropical forests may be critical systems and may be relatively reactive to disturbances ([@R47], [@R48]) and able to adapt optimally to new external conditions/constraints. Under a given set of environmental conditions, only a few species are best at exploiting the limited available resources ([@R49]). Because of environmental fluctuations, these conditions may not continue to remain advantageous for the existing very few abundant species. However, a large pool of species may serve as a reservoir of new opportunities and responses and as a buffer against newly changed conditions ([@R49]). According to this view, hyper-rarity is essential for an ecosystem to maintain its functions and react promptly to changes: Rare species may provide the key to an ecosystem's future ([@R50]). {ref-type="table"} at the whole-forest scale, and the red line is the line of equation *y* = 1. All values are located above this line, thus indicating that the relative fluctuations in abundance are considerable for all the forests. (**B**) Contour plot of the relative fluctuation of abundances for an NB SAD **$\mathit{F}(\xi\mathit{r}) = \sqrt{\langle{(\mathit{n} - \langle\mathit{n}\rangle)}^{2}\rangle}/\langle\mathit{n}\rangle$.** The black dots represent the pair (*r*, log \[1 − ξ\]), where *r* and ξ are the predicted parameters for each forest of our data set after upscaling at the whole-forest scale. These dots are all located in the region of the parameter space around which the function *F*(ξ, *r*) diverges, that is, ξ ≈ 1 and 0 \< *r* \< 0.7.](1701438-F5){#F5} To summarize, we have presented a theoretical framework to upscale species richness and abundances in tropical forests from a limited number of samples. The advantage of our method mainly relies on two properties. First, it is flexible. The NB, depending on the value of its parameters, may display either an LS-like behavior or an interior mode, and it is therefore able to describe different SAD shapes. Thus, we can use the same functional form to reproduce different ecosystems' SAD, as those observed in our data set. In contrast, an LS SAD predicts a very specific form for the SAD that is not flexible enough to describe any SAD with an interior mode. Furthermore, our approach, relying on an appropriate linear combination of NBs, can basically accommodate any type of complex SAD functional form. Second, the NB (or a combination of them), besides being flexible, is also self-similar under different sampling intensities. This is the key feature that allows us to obtain an easy analytical formula to upscale the SAD from the sample scale to any arbitrary one. In the study of Harte *et al*. ([@R21]), despite the flexibility of the approach, the upscaling can be performed only by numerically solving a pair of analytical equations. In the study of Zillio and He ([@R51]), they proposed an iterative method for estimating the species richness and the abundance distribution. Again, this method is flexible, but no analytical treatment can be performed. Finally, in our framework, we only need the fraction of the sampled area with respect to the whole forest, whereas in other approaches, additional information on the upscaled forest is required \[for example, the number of individuals of the most abundant species ([@R52])\]. These two properties allow our method to be applied on statistical upscaling problems beyond forest ecology. A possible application is, for example, in the field of metagenomics. Using recently developed DNA sequencing machines, it is possible to obtain the total genomic DNA directly from a macro fauna or flora environmental sample (that is, a macrobiome). This metagenomic (gene of genes) approach, together with taxonomic classification algorithms ([@R53]), allows a characterization of the biodiversity of the samples (typically prokaryotes). However, SAD curves built in this way describe the biodiversity only very locally (the scale of the given environmental sample). Nevertheless, by assuming well-mixed communities and finding an appropriate combination of NBs fitting the observed SAD, we can use our framework to upscale the microbiome SAD to a larger scale (for example, the whole gut), as would be measured if it were possible to survey the entire environment. It can also be applied to immunology for finding the number of T cell receptor clonotypes in a human body. These examples show the promising generality of our approach and open the possibility of new applications of the upscaling framework to other taxa or type of systems. MATERIALS AND METHODS ===================== Upscaling NBs ------------- Here, we chose the NB distribution in [Eq. 2](#E2){ref-type="disp-formula"} as the SAD. Apart from its simplicity and versatility, we chose this form for our analysis for four reasons: \(1\) Any discrete probability distribution, such as the SAD, can be approximated to any degree of accuracy by a suitable linear combination of NBs (see section S3 for some examples and discussion). We made the parsimonious choice of a single NB function because it suffices to approximately describe the available tropical forest data, as discussed in the Results and Discussion. \(2\) The NB distribution arises naturally as the steady-state SAD of an ecosystem with sufficiently weak interspecies interactions and undergoing generalized dynamics of birth, death, speciation, and immigration to and emigration from the surrounding metacommunity (see "Stochastic model leading to an NB SAD"). \(3\) In the limit of *r* → 0, the NB becomes the well-known Fisher's LS, which has been widely used to describe the patterns of abundance in ecological communities. Of course, because of the flexibility of choosing *r* to be nonzero, the NB distribution is always more versatile than the LS. The SAD, especially at large scales or with increasing sampling effort ([@R27]), often displays an interior mode that cannot be captured by an LS distribution. To assess whether the increased reliability of the NB method with respect to the LS method is only due to the introduction of the additional parameter *r*, we used the Akaike information criterion, which shows that the NB is the preferred model for all tropical forests in our data set except one for which *r* is very close to zero. 4\. Finally and importantly, if one chooses two contiguous patches with NB as SADs characterized by the same parameters *r* and ξ ≡ ξ~1/2~ and combines the two, then remarkably, the resulting larger patch is also characterized by an NB distribution with the same scale-invariant value of *r* and a new scale-dependent parameter, ξ, given by the analytical expression in [Eq. 4](#E4){ref-type="disp-formula"} below with *p* = 1/2. This special form-invariant property of the NB distribution, albeit with a scale-dependent parameter, makes it particularly well suited for our extrapolation studies. When upscaling, we are interested in the SAD and in the total number of species, *S*, at the scale of the whole forest area *A*. We denote *P*(*n*\|1) as the probability that a species has exactly *n* individuals at the whole-forest scale (here, 1 refers to the whole forest). Note that *P*(*n*\|1) is defined only for *n* ≥ 1, because *S* is the total number of species actually present in the forest, thus each having at least one individual. We assumed that the SAD has the functional form of an NB, $\mathcal{P}(\mathit{n}|\mathit{r},\xi)$, for nonzero populations, with parameters (*r*, ξ) (*r* is known as the clustering coefficient), that is$$\begin{matrix} {\mathit{P}(\mathit{n}|1) = \mathit{c}(\mathit{r},\xi)\mathcal{P}(\mathit{n}|\mathit{r},\xi)~~~~~\text{with}} \\ {\mathcal{P}(\mathit{n}|\mathit{r},\xi) = \begin{pmatrix} {\mathit{n} + \mathit{r} - 1} \\ \mathit{n} \\ \end{pmatrix}\xi^{\mathit{n}}{(1 - \xi)}^{\mathit{r}},~~~~~\mathit{c}(\mathit{r},\xi) = \frac{1}{1 - {(1 - \xi)}^{\mathit{r}}}} \\ \end{matrix}$$where *c* is the normalization constant. The constant *c* was determined by imposing $\sum_{\mathit{n} = 1}^{\infty}\mathit{P}(\mathit{n}|1) = 1$, where the sum starts from *n* = 1, because species with zero abundance at the scale of the whole forest will be also absent in the subplots. Note that $\mathcal{P}(\mathit{n}|\mathit{r},\xi)$ was normalized for *n* ≥ 0. In the subplots, there is a nonzero probability to find species, which are present in the whole forest, with *n* = 0 individuals, and thus it accounts for the number of missing species in the subplots. Let us now consider a subsample of area *a* of the whole forest and define *p* = *a*/*A* as the scale of the sample, which is the fraction of the sampled forest. The goal is to compute the SAD in the subsample. We assumed that the subsample SAD was not affected by spatial correlations due to both interspecific and intraspecific interactions. This hypothesis is well satisfied using in silico--generated forests with various degrees of spatial correlations (see section S4). Under this hypothesis, the conditional probability that a species has *k* individuals in the smaller area, *a* = *pA*, given that it has total abundance *n* in the whole region of area *A* is given by the binomial distribution$$\mathcal{P}_{\text{binom}}(\mathit{k}|\mathit{n},\mathit{p}) = \begin{pmatrix} \mathit{n} \\ \mathit{k} \\ \end{pmatrix}\mathit{p}^{\mathit{k}}{(1 - \mathit{p})}^{\mathit{n} - \mathit{k}}~~~~~\mathit{k} = 0,\ldots,\mathit{n}$$and $\mathcal{P}_{\text{binom}}(\mathit{k}|\mathit{n},\mathit{p}) = 0$ if *k* \> *n*. Now, we want to prove that the subsample SAD, *P*(*k*\|*p*), is again an NB for *k* ≥ 1, with the rescaled parameter ξ and the same *r*. It can be shown that the probability, $\mathcal{P}_{\text{sub}}(\mathit{k}|\mathit{p})$, to find a species with population *k* ≥ 0 in the subplot of area *a* = *pA* is$$\mathcal{P}_{\text{sub}}(\mathit{k}|\mathit{p}) = \mathit{c}(\mathit{r},\xi) \cdot \mathcal{P}(\mathit{k}|\mathit{r},{\hat{\xi}}_{\mathit{p}})~~~~~\mathit{k} \geq 1$$$$\mathcal{P}_{\text{sub}}(0|\mathit{p}) = 1 - \sum\limits_{\mathit{k} \geq 1}\mathcal{P}_{\text{sub}}(\mathit{k}|\mathit{p})~~~~~\mathit{k} = 0$$where$${\hat{\xi}}_{\mathit{p}} = \frac{\mathit{p}\xi}{1 - \xi(1 - \mathit{p})}$$The method uses only the information that we can infer from a subsample at some scale *p*\*. Therefore, we only have information on the abundances of the *S*\*(≤*S*) species present in the surveyed area. By denoting with *S*\*(*k*), the number of species of abundance *k* at scale *p*\*, we obtained$$\frac{\mathit{S}^{*}(\mathit{k})}{\mathit{S}^{*}} \equiv \mathit{P}(\mathit{k}|\mathit{p}^{*}) = \frac{\mathcal{P}_{\text{sub}}(\mathit{k}|\mathit{p}^{*})}{\sum_{\mathit{k}^{\prime} \geq 1}\mathcal{P}_{\text{sub}}(\mathit{k}^{\prime}|\mathit{p}^{*})} = \frac{\mathcal{P}(\mathit{k}|\mathit{r},{\hat{\xi}}_{\mathit{p}^{*}})}{\sum_{\mathit{k}^{\prime} \geq 1}\mathcal{P}(\mathit{k}^{\prime}|\mathit{r},{\hat{\xi}}_{\mathit{p}^{*}})} = \mathit{c}(\mathit{r}{\hat{\xi}}_{\mathit{p}^{*}}) \cdot \mathcal{P}(\mathit{k}|\mathit{r},{\hat{\xi}}_{\mathit{p}^{*}})~~~~~\mathit{k} \geq 1$$which, from [Eq. 5](#E5){ref-type="disp-formula"}, is an NB normalized for *k* ≥ 1, whereas $\mathcal{P}(\mathit{k}|\mathit{r},{\hat{\xi}}_{\mathit{p}^{*}})$ is normalized for *k* ≥ 0. We therefore obtained the key result that starting with an NB distribution for the SAD at the whole-forest scale, the SAD at smaller scales is also distributed according to an NB with the same clustering coefficient *r* and a rescaled parameter ${\hat{\xi}}_{\mathit{p}^{*}}$ depending on both ξ and *p*\*. A SAD with the property of having the same functional form at different scales is said to be form-invariant. By fitting the SAD of the data at the scale *p*\*, we can thus find both the parameters *r* and ${\hat{\xi}}_{\mathit{p}^{*}}$ and, by inverting [Eq. 9](#E9){ref-type="disp-formula"}, we can obtain ξ$$\xi = \frac{{\hat{\xi}}_{\mathit{p}^{*}}}{\mathit{p}^{*} + {\hat{\xi}}_{\mathit{p}^{*}}(1 - \mathit{p}^{*})}$$ Using [Eq. 9](#E9){ref-type="disp-formula"} to eliminate ξ from the last equation, one obtains the following relation for the parameter ξ at the two scales *p* and *p*\* referred in the Results$${\hat{\xi}}_{\mathit{p}} = \frac{\mathit{p}{\hat{\xi}}_{\mathit{p}^{*}}}{\mathit{p}^{*} + {\hat{\xi}}_{\mathit{p}^{*}}(\mathit{p} - \mathit{p}^{*})} \equiv \mathit{U}(\mathit{p},\mathit{p}^{*}|{\hat{\xi}}_{\mathit{p}^{*}})$$from which, of course, one can recover both [Eqs. 9](#E9){ref-type="disp-formula"} and [11](#E11){ref-type="disp-formula"}, where $\xi \equiv {\hat{\xi}}_{\mathit{p} = 1}$. We now wish to determine the relation between the total number of species at the whole scale *p* = 1, *S*, with the total number of species surveyed at scale *p*, *S*~*p*~. Referring to the scale *p*\*, in the following equation, we also used the notation $\mathit{S}^{*} \equiv \mathit{S}_{\mathit{p}^{*}}$. This can be simply obtained by observing that$$\mathcal{P}_{\text{sub}}(\mathit{k} = 0|\mathit{p}^{*}) = (\mathit{S} - \mathit{S}^{*})/\mathit{S}$$$$\mathcal{P}_{\text{sub}}(\mathit{k}|\mathit{p}^{*}) = \mathit{S}^{*}(\mathit{k})/\mathit{S}$$ Using [Eq. 8](#E8){ref-type="disp-formula"}, we finally found that the prediction for the total number of species in the whole forest, in terms of the data on the surveyed subplot, is given by$$\mathit{S} = \frac{\mathit{S}^{*}}{1 - \mathcal{P}_{\text{sub}}(\mathit{k} = 0|\mathit{p}^{*})} = \mathit{S}^{*}\frac{1 - {(1 - \xi)}^{\mathit{r}}}{1 - {(1 - {\hat{\xi}}_{\mathit{p}^{*}})}^{\mathit{r}}}$$where ξ is given by [Eq. 11](#E11){ref-type="disp-formula"}. Our framework holds exactly when species are spatially uncorrelated. However, our in silico experiments indicated that the framework is robust even in the presence of spatial correlations and for different sampling methods (section S4). Stochastic model leading to an NB SAD ------------------------------------- As explained in the Introductions, the NB distribution can be derived from first principles on the basis of biological processes. Let $\mathcal{P}_{\mathit{n},\mathit{s}}(\mathit{t})$ be the probability that, at time *t*, species *s* has exactly *n* individuals, where *s* ∈ {1, ..., *S*}. We assumed that the population dynamics of each species is governed by two terms, *b*~*n*,*s*~ and *d*~*n*,*s*~, which are the birth and death rates, respectively, for species *s* with *n* individuals. The master equation regulating the evolution of $\mathcal{P}_{\mathit{n},\mathit{s}}(\mathit{t})$ for *n* ≥ 0 is then$$\frac{\partial}{\partial\mathit{t}}\mathcal{P}_{\mathit{n},\mathit{s}}(\mathit{t}) = \mathcal{P}_{\mathit{n} - 1,\mathit{s}}(\mathit{t})\mathit{b}_{\mathit{n} - 1,\mathit{s}} + \mathcal{P}_{\mathit{n} + 1,\mathit{s}}(\mathit{t})\mathit{d}_{\mathit{n} + 1,\mathit{s}} - \mathcal{P}_{\mathit{n},\mathit{s}}(\mathit{t})\mathit{b}_{\mathit{n},\mathit{s}} - \mathcal{P}_{\mathit{n},\mathit{s}}(\mathit{t})\mathit{d}_{\mathit{n},\mathit{s}}$$ The above equation is also valid for *n* = 0 and *n* = 1 if we set *b*~−1,*s*~ = *d*~0,*s*~ = 0. The steady-state solution is$$\mathcal{P}_{\mathit{n},\mathit{s}} = \mathit{c}_{\mathit{s}}\prod\limits_{\mathit{i} = 0}^{\mathit{n} - 1}\frac{\mathit{b}_{\mathit{i},\mathit{s}}}{\mathit{d}_{\mathit{i} + 1,\mathit{s}}}$$ The term *c*~*s*~ is a normalization factor found by imposing $\sum_{\mathit{n} = 0}^{\infty}\mathcal{P}_{\mathit{n},\mathit{s}} = 1$. Let us assume that the birth term in the above equation depends on a density-independent term, *b*~*s*~, which is the per-capita birth rate, and on the term *r*~*s*~, which takes into account immigration events or intraspecific interactions$$\mathit{b}_{\mathit{n},\mathit{s}} = \mathit{b}_{\mathit{s}}(\mathit{n} + \mathit{r}_{\mathit{s}})$$ Analogously, let us suppose that the death term depends on a density-independent term, *d*~*s*~, which is the per-capita death rate$$\mathit{d}_{\mathit{n},\mathit{s}} = \mathit{d}_{\mathit{s}}\mathit{n}$$ These suppositions are reasonable in ecology. By substituting in [Eq. 16](#E16){ref-type="disp-formula"} and setting ξ~*s*~ = *b*~*s*~/*d*~*s*~, we obtained$$\mathcal{P}_{\mathit{n},\mathit{s}} = \mathit{c}_{\mathit{s}}\begin{pmatrix} {\mathit{n} + \mathit{r}_{\mathit{s}} - 1} \\ \mathit{n} \\ \end{pmatrix}\xi_{\mathit{s}}^{\mathit{n}}$$ The normalization constant can be easily found by imposing$$1 = \sum\limits_{\mathit{n} = 0}^{\infty}\mathcal{P}_{\mathit{n},\mathit{s}} = \mathit{c}_{\mathit{s}}\sum\limits_{\mathit{n} = 0}^{\infty}\begin{pmatrix} {\mathit{n} + \mathit{r}_{\mathit{s}} - 1} \\ \mathit{n} \\ \end{pmatrix}\xi_{\mathit{s}}^{\mathit{n}} = \mathit{c}_{\mathit{s}}{(1 - \xi_{\mathit{s}})}^{- \mathit{r}_{\mathit{s}}}$$ Therefore, the probability that the *s*th species has *n* individuals at equilibrium is given by an NB with parameters (*r*~*s*~, ξ~*s*~)$$\mathcal{P}_{\mathit{n},\mathit{s}} = \begin{pmatrix} {\mathit{n} + \mathit{r}_{\mathit{s}} - 1} \\ \mathit{n} \\ \end{pmatrix}\xi_{\mathit{s}}^{\mathit{n}}{(1 - \xi_{\mathit{s}})}^{\mathit{r}_{\mathit{s}}}$$Under the neutral hypothesis, in which all species are considered to be equivalent, we can remove the species index *s* from the above equation, thus obtaining a negative binomially distributed SAD for the ecosystem under study. Supplementary Material ====================== ###### http://advances.sciencemag.org/cgi/content/full/3/10/e1701438/DC1 We are indebted to S. P. Hubbell for his insightful comments. We also thank the Center of Tropical Research Science for providing the empirical data of the BCI and Pasoh forests. The BCI forest dynamics research project was founded by S. P. Hubbell and R. B. Foster and is now managed by R. Condit, S. Lao, and R. Perez under the Center for Tropical Forest Science and the Smithsonian Tropical Research in Panama. It has been funded by numerous organizations, principally the U.S. NSF, and hundreds of field workers have contributed. All other data in this publication were provided by the Tropical Ecology Assessment and Monitoring Network, a collaboration between Conservation International, the Missouri Botanical Garden, the Smithsonian Institution, and the Wildlife Conservation Society, and partially funded by these institutions, the Gordon and Betty Moore Foundation, and other donors. **Funding:** S.A. would like to thank the Isaac Newton Institute for Mathematical Sciences, Cambridge, for the support and hospitality during the program "Stochastic Dynamical Systems in Biology: Numerical Methods and Applications," where work on this paper was partially undertaken. This work was supported by Engineering and Physical Sciences Research Council grant no. EP/K032208/1. The authors acknowledge the Department of Mathematics "Tullio Levi-Civita" of the University of Padova and the Department of Applied Maths of the University of Leeds for financial support. **Author contributions:** S.A., J.R.B., A.M., and S.S. designed the work. A.T. and S.S. performed the numerical simulations and the data analysis and prepared the figures. S.A., A.T., M. Formentin, M. Favretti, I.V., and A.M. performed the analytical calculations. All the authors contributed to other aspects of the paper and the writing of the manuscript. **Competing interests:** The authors declare that they have no competing interests. **Data and materials availability:** We use a global-scale compilation of 1248 local sites collected over 15 forests around the planet on different tropical field stations of the equatorial zone. All data are publicly available. See the Supplementary Materials for web links and a more detailed description of the data set. For numerical simulations performed in this study, we used standard commands and programming tools in R/Mathematica. A ready-to use code in R for predicting the upscaled SAD and total species richness and abundances is available at <https://github.com/annatovo> or in the web page of our group ([www.pd.infn.it/\~maritan](http://www.pd.infn.it/~maritan)). Requests for materials should be addressed to S.S. (<suweis@pd.infn.it>), to M. Formentin (<marco.formentin@unipd.it>), to J.R.B. (<banavar@uoregon.edu>), or to A.M. (<maritan@pd.infn.it>). Supplementary material for this article is available at <http://advances.sciencemag.org/cgi/content/full/3/10/e1701438/DC1> section S1. Upscaling biodiversity section S2. Limitation of the LS methods section S3. Flexibility of NB distribution section S4. Test on computer-simulated forests section S5. Comparison with other popular estimators section S6. Data set section S7. Self-consistency and estimation of the critical *p*\*: How much remains to be sampled? section S8. RSA parameters maximize relative fluctuation in abundances fig. S1. Assuming that the global RSA is distributed according to an NB, we can compute the probability that a species comprises a single individual at the scale *p* by using eq. S31. fig. S2. Fisher's α for three different rainforests: Amazonia, Barro Colorado Nature Monument, and Caxiuana. fig. S3. Fit of an RSA consisting of a combination of an LS and a log-normal distribution. fig. S4. We have generated synthetic data from a combination of discrete distributions (a binomial distribution of parameters *r* = 40 and ξ = 0.8, a geometric distribution of parameter μ = 0.15, and a Poisson distribution with parameter λ = 15) and fit these data with one, three, and six NBs, respectively. fig. S5. Robustness of the method. fig. S6. Comparison between biodiversity estimators for Amazonia and BCI forests. fig. S7. Self-consistency test of our framework. fig. S8. Plot, in logarithmic scale, of the percentage *p*~pred~% that one ought to sample to have a precision estimate of around 5% for the predicted percentage of hyper-rare species, that is, species with fewer than 1000 individuals at the global scale. table S1. Predicted number of singletons in the whole area of each tropical forest obtained by applying our method (NB method). table S2. Prediction of the total number of species obtained by applying both NB and LS methods to the forest generated according to an NB and distributed in 8900 × 8900 units according to two different modified Thomas processes with the same density of clusters ρ = 6 × 10^−5^ and different clump sizes σ = 15 and 200. table S3. Summary table of the most popular biodiversity estimators. table S4. Comparison between NB, LS, Chao~wor~, and the Harte methods on empirical data. table S5. Comparison between the NB, LS, Chao~wor~, and Harte methods on BCI empirical data. table S6. Number of species and singletons in 15 forests in our data set with the percentage *p*% (last column of the table) of surveyed area. Reference ([@R54]) [^1]: These authors contributed equally to this work.

1. Prevalence and Character of Olfactory Loss in PD {#sec1} =================================================== According to a recent study by Politis et al. \[[@B1]\], olfactory loss belongs to the top-five most prevalent motor and nonmotor symptoms in early stage PD patients that have affected their quality of life. Only pain is referred to as a more prevalent troublesome nonmotor problem in this patient group. In line with this result, virtually all studies performed since the 1970s have shown olfactory disturbances in PD patients. Published data on the prevalence of olfactory dysfunction in PD range from 45% and 49% in the pioneering studies of Ansari and Johnson \[[@B2]\], and Ward et al. \[[@B3]\], respectively, up to 74% in the work of Hawkes et al. \[[@B4]\], or as high as 90% in a study published by Doty et al. \[[@B5]\]. In our recent multicentre study \[[@B6]\] using a comprehensive testing method (see chapter 2) in a large sample of PD patients (*n* = 400) from 3 independent populations, the prevalence of olfactory dysfunction in people with PD was greater than previously reported with regard to norms obtained in healthy young subjects. More than 96% of PD patients were found to present with olfactory dysfunction. When using age-dependent normative criteria, 74.5% of this study population was diagnosed with olfactory loss ([Figure 1](#fig1){ref-type="fig"}). Furthermore, more than 80% of PD patients with smell loss were functionally anosmic or severely hyposmic regardless of the olfactory test being used for diagnosis. Only very few patients present with accompanying parosmia, or phantosmia. Our data also confirmed numerous previous studies with regard to the lack of olfactory improvement after therapy with dopaminergic agents \[[@B5], [@B11]\] and the missing correlation between olfactory loss and both duration of disease \[[@B4], [@B5], [@B7]\] and the clinical severity of PD as measured by means of the Hoehn and Yahr scale and the UPDRS (compare \[[@B8]\])---although some studies found a correlation between the severity of PD and certain measures of olfactory function, namely, latencies of olfactory event-related potentials \[[@B9]\] or results from an odor discrimination task \[[@B10]\]. With regard to olfactory function, we did not find major differences between subtypes of PD, namely, tremor-dominant PD, akinetic-rigid PD, and mixed-type PD. While this confirms previous observations in a small sample size of 37 patients \[[@B11]\] ([Figure 2](#fig2){ref-type="fig"}), the present findings are in contrast to reports by Stern and colleagues \[[@B7]\] who reported significantly better odor identification scores in patients with tremor-predominant PD than in cases with postural instability-gait disorder-predominant PD. While differences between studies may be due to the type of olfactory test used, sample size, normative data, and age distribution (which varied between these investigations), available data allow the conclusion that olfactory dysfunction is a highly reliable symptom of the disease. This concurs with the results of a case-control study on 90 PD patients and healthy controls by Bohnen et al. \[[@B12]\] who found that the accuracy of smell testing in PD diagnosis outweighs the accuracy of motor test batteries, and also other nonmotor tests of, for example, depression and anxiety. 2. Testing Methods {#sec2} ================== As the olfactory loss in PD has a general character, all three olfactory qualities (threshold, discrimination, identification) are involved. Therefore, different subtests of olfactory function may reflect smell loss in PD patients and may be used as single measurement. Only a comprehensive approach, however, allows a precise evaluation of olfactory function, that is, that it would be best to perform all 3 subtests to obtain a maximum of reliable information. Psychophysical assessment of olfactory function is based on the presentation of odors and the recording of the subjects\' response. Advantage of this "low-tech" approach include the speed of testing, allowing for rapid screening of olfactory function \[[@B13], [@B14]\]. While screening tests only differentiate between normal and pathologic states, more extensive tests allow for a reliable discrimination between anosmic, hyposmic, and normosmic subjects, respectively. Good tests have to be based (1) on normative data acquired and validated in (2) large samples (e.g., \[[@B15], [@B16]\]). Many tests are based on a forced choice verbal identification of odors while others also include results from odor discrimination and odor threshold measurements (comprehensive approach; [Figure 3](#fig3){ref-type="fig"}). Most tests are based on the identification of odors. In odor identification tasks, an odorant is presented at a suprathreshold concentration and subjects are required to identify the odor from a list of descriptors. This forced-choice procedure controls the subjects\' response bias. A major problem of odor identification is, however, that it strongly relies on the verbal abilities of the subject. Consequently, on average, this enables female subjects to outperform men \[[@B17]\]. In addition, odor identification tests have a strong cultural precondition as not all odors are known equally well in various cultural groups. The concept embedded in threshold tests is that a subject is repeatedly exposed to ascending and descending concentrations of the same odorant and is required to identify the least detectable concentration for this individual odor (\[[@B18]\]; see also \[[@B19]\]). Other measures assessing olfactory loss may include investigation of the patient\'s quality of life, for example, the "Questionnaire for Olfactory Dysfunction" \[[@B20]\], or the recording of olfactory event-related potentials (for review, see \[[@B21]\]). In the near future, immunohistochemical, volumetric, and functional neuroimaging studies of the olfactory sytem might become relevant for PD diagnosis. There is still little information about PD-specific changes of the olfactory epithelium and their diagnostic use. In a recent study, we compared bioptic material from PD patients for histological/histochemical changes with that from patients with olfactory dysfunction due to other reasons \[[@B51]\]. However, we found no specific changes in the nasal mucosa of PD patients. Further, it could be assumed that loss of olfactory bulb volume would be a reliable finding in PD which, to some degree, might be helpful in differential and early diagnosis of PD. Results from a recent study \[[@B52]\] indicated that there is little if any difference in OB volume between PD patients with anosmia/hyposmia and healthy, normosmic controls. In another effort to identify brain structures responsible for smell loss in PD, functional magnetic resonance imaging (fMRI) was used to investigate brain activity related to olfactory processing in PD. Overall, the results of these studies \[[@B53], [@B54]\] indicate that neuronal activity in the amygdala and hippocampus is reduced in PD patients which may specifically impact on olfactory sensitivity. In addition, neuronal activity in components of corticostriatal loops appears to be upregulated indicating compensatory processes involving the dopaminergic system. 3. Olfactory Dysfunction as a Prodromal Symptom {#sec3} =============================================== Support for the existence of a prodromal phase comes from imaging, neuropathology, and various clinical or epidemiological surveys. The best evidence that derives from large prospective studies relates to disorders affecting olfaction, the enteric nervous system, and depression \[[@B22]--[@B24]\]. Estimates for the duration of the prodrome range from 2 to 50 years depending on the symptom, duration of followup, accuracy of diagnosis, and individual variation. PD patients frequently report reduction in their sense of smell that occurs a few years prior to the onset of motor symptoms. However, patients\' unawareness of smell deficits may account for the inconsistent results described in retrospective surveys. In a small study \[[@B25]\], upon questioning prior to olfactory testing, 9 out of 37 patients (24%) indicated an awareness of a decrease of olfactory function which actually preceded their diagnosis of PD. A plethora of evidence from recent studies supports the view that deficits in the sense of smell may precede clinical motor symptoms by years. A study by Ponsen et al. \[[@B26]\] on 361 asymptomatic relatives of PD patients selected 40 relatives with the lowest olfactory performance. Within 2 years of followup, 10% of these first-degree relatives of PD patients with significant olfactory loss developed clinical PD. In a followup study, five years from baseline testing \[[@B27]\], five relatives had developed clinical PD as defined by the United Kingdom Parkinson\'s Disease Society Brain Bank Diagnostic Criteria for Parkinson\'s Disease. Initial clinical (motor) symptoms appeared 9 to 52 months (median 15 months) after baseline testing. Poorer performance on each of three olfactory tasks was associated with an increased risk of developing PD within 5 years. In 2007 \[[@B28]\], we published data on a clinical followup of a previous investigation \[[@B29]\], in which 30 patients diagnosed with idiopathic olfactory loss participated. Four years from baseline, 7% (*n* = 2) of the individuals with idiopathic olfactory loss who were available for followup examination (*n* = 24) had newly developed clinical PD symptoms. Altogether, 13% (*n* = 4) of the patients presented with PD-relevant abnormalities of the motor system. The results indicated that unexplained olfactory loss may be associated with an increased risk of developing PD-relevant motor symptoms. This is in accord with the results of a large longitudinal study by Ross and colleagues \[[@B30]\]. They assessed olfactory function in 2267 elderly men in the Honolulu Heart Program and found an association between smell loss and future development of PD. They came to the conclusion that impaired olfaction can predate PD by at least 4 years and may be a useful screening tool to detect those at high risk for development of PD in later life. However, this relationship appears to weaken beyond the 4-year period. Along with quantitative smell loss, such as hyposmia, idiopathic phantosmia has also been suggested to herald PD. A number of case reports could show that some patients have experienced phantosmia very early in the course of the disease \[[@B31]--[@B33]\]. According to a recent study by Landis et al. \[[@B34]\], however, idiopathic phantosmia as an early sign of PD remains probably a rather exceptional presentation whereas the overwhelming majority of people with idiopathic phantosmia will not develop PD. Recent data on olfactory loss as a PD symptom that is present at the earliest stages of the disease are compatible with predictions made on the basis of neuropathological investigations. Braak et al. \[[@B35]\] describe involvement of olfactory pathways and lower brainstem before nigrostriatal pathways are affected which might cause early nonmotor symptoms. Huisman et al. \[[@B36]\] found an increase of (inhibitory) dopaminergic neurons in the olfactory bulb in PD patients. They interpreted their finding within the context of a possible compensatory mechanism in response to the loss of dopaminergic neurons in the basal ganglia. This concurs with their observation that dopaminergic neurogenesis in the glomerular layer tripled after nigrostriatal lesioning and, consistent with this finding, the total number of tyrosine hydroxylase- (TH-) positive cells increased \[[@B37]\]. However, results of a followup study \[[@B38]\] indicated a gender-related change, that is, that the number of dopaminergic cells in the olfactory bulbs of both male and female Parkinson\'s patients equals that of healthy males of the same age group. Authors, therefore, concluded that the hyposmia in Parkinson\'s disease patients cannot simply be ascribed to dopamine in the olfactory bulb. Regardless the small number of prospective studies in this field, olfactory loss should be considered a promising contribution to the early diagnosis of PD. For instance, the current Parkinson\'s associated risk syndrome (PARS) study \[[@B39]\] will advance our understanding of early PD presentation. 4. Olfaction in Differential Diagnosis {#sec4} ====================================== Numerous studies suggest that olfactory disturbances in PD may have diagnostic utility for the differentiation of PD from other movement disorders. Wenning et al. \[[@B40]\] presented data suggesting that olfactory function is differentially impaired in distinct Parkinsonian syndromes. They reported a preserved or mildly impaired olfactory function to be more likely for atypical parkinsonism such as multiple system atrophy, progressive supranuclear palsy, or corticobasal degeneration whereas markedly pronounced olfactory loss appeared to suggest PD. Similar to the results of Wenning et al., in a study on 50 Parkinsonian patients \[[@B25]\], we also found evidence for olfactory loss in MSA, but little or no olfactory loss in (the few investigated) patients with PSP and CBD. With regard to the differentiation between MSA and PD ([Figure 3](#fig3){ref-type="fig"}) at a cutoff of a TDI score (combined results for odor thresholds, odor discrimination, and odor identification; see also \[[@B15]\]) of 19.5, psychophysical testing had a sensitivity of 78% and a specificity of 100%. When the cutoff TDI score was increased to 24.8, sensitivity in this sample was 100% while specificity fell to 63%. This moderate specificity seems to be the limiting parameter for diagnostic purposes. A recent American Academy of Neurology practice parameter on the diagnosis and prognosis of PD concluded that olfactory testing "should be considered" to differentiate PD from PSP and CBD but not from MSA \[[@B41]\]. Furthermore, Liberini et al. \[[@B42]\] reported a significant olfactory impairment in Lewy body disease (LBD) which does not allow differentiation from PD. In a sample of 116 patients with mild LBD, mild Alzheimer\'s disease, mild cognitive impairment, and controls, Williams et al. \[[@B43]\] describe even more marked olfactory impairment in patients with mild dementia with Lewy bodies than present in those with mild Alzheimer\'s disease. This lends significance to the role of Lewy body pathology in olfactory dysfunction \[[@B44]\] which would be in line with the observation that patients with nondegenerative causes of parkinsonism such as vascular parkinsonism \[[@B45]\] present with preserved smell function. There is also evidence for less olfactory disturbance in familial parkinsonism. In PARK2, the olfactory sense is relatively well preserved whereas PARK1 subjects are mildly hyposmic. Recent data \[[@B46]\] suggest that PARK 8 individuals present with impaired olfactory identification whilst asymptomatic carriers show normal olfactory performance. In secondary parkinsonism, study results also indicate a relationship between Parkinsonian symptoms and olfactory dysfunction. We found an association between medication-induced parkinsonism and olfactory dysfunction in patients with psychotic depression treated with D2-blocking neuroleptic drugs \[[@B47]\]. Here, the severity of motor symptoms was positively correlated with the degree of olfactory dysfunction which might indicate patients with a latent basal ganglia dysfunction. Similar to the results seen in drug-induced parkinsonism, data from a recent study reveal that Wilson\'s disease patients with neurological symptoms show a significant olfactory dysfunction compared to hepatic-type patients \[[@B48]\]. Individuals who are more severely neurologically affected also present with more pronounced olfactory deficits. Based on these observations, olfactory testing should not be considered to differentiate PD from these specific conditions. However, olfactory testing has been shown to be important in cases where patients present with Parkinsonian features but with preserved olfaction. Here, it appears valid to question a diagnosis of PD. 5. Conclusions {#sec5} ============== Recent data suggest that inexpensive olfactory probes improve the diagnostic process in patients with PD. In contrast to imaging procedures, olfactory testing is quick and easy to perform. Validated tests can be used as reliable diagnostic tools even in nonspecialized centers. Deeb et al. \[[@B49]\] found that a basic smell test is as sensitive as a dopamine transporter scan. According to this study, the sensitivities of the University of Pennsylvania Smell Identification Test \[[@B50]\] and DaTSCAN are high at 86% and 92%, respectively. Although DaTSCAN is superior for "localization," a smell test is considerably "cheaper," and neither is disease specific. Consequently, structured and validated tests of olfactory function should be a mandatory part of the early and differential diagnosis of PD. Our experience suggests that it only takes little time to follow up patients with a diagnosis of idiopathic smell loss neurologically as an essential part of their regularly scheduled visit to the Smell and Taste Clinic which is a time- and expense-efficient process well warranted. Such a comprehensive multidisciplinary approach might enable the physician to detect slight motor abnormalities in an at-risk population as early as possible. This may also give rise to clinical studies which allow administration of neuroprotective substances in individuals with, for example, unexplained smell loss. Up till now, therapeutic studies with neuroprotective agents in hyposmic PD patients are currently underway and may help us to evolve preventive strategies for PD in future. ![Olfactory function of the total number of 400 PD patients. Results are shown as a composite TDI score (sum of odor threshold, odor discrimination, and odor identification score) adjusted to age-related norms \[[@B6]\].](PD2011-450939.001){#fig1} ![Olfactory function in PD subtypes (TD: tremor-dominant, AR: akinetic-rigid, e: equivalent type) and multiple system atrophy (MSA) \[[@B11]\].](PD2011-450939.002){#fig2} {#fig3} [^1]: Academic Editor: Irena Rektorova

Carolina Vila Chã Vaz Saleiro and Joana Maria da Silva Barbosa Guimarães Ribeiro contributed equally to this work. {#joa312382-ltr-0001} {#joa312382-sec-0001} Institution of timely treatment for electrical storm is crucial but challenging. Beta‐blockers and amiodarone are often first line as they are effective in suppressing ventricular arrythmias (VAs) and have a favorable safety profile. However, in cases of Brugada Syndrome (BrS) and early‐repolarization syndromes, these drugs will be largely ineffective, and pathology‐guided treatment is required. We present a case that highlights the importance of early recognition of the precipitating mechanisms of electrical storm. A 66‐year‐old man was referred for percutaneous intervention (PCI) due to stable angina with positive functional imaging test. Previous medical history included hypertension, dyslipidemia, and overweight and he was medicated accordingly. There was no family history of cardiac disease or sudden death. Angiography showed three‐vessel coronary artery disease: the left anterior descending artery (LAD) had a critical distal lesion, which was treated with PCI, and an intermediate ostial lesion that was deferred; the circumflex artery (LCx) was diffusely diseased; and the right coronary artery (RCA) had two significant lesions---90% proximal and 75% distal stenosis that were treated with PCI. The procedure was successful and uneventful. The patient was admitted to the cardiology ward for surveillance. Ten hours later, he reported mild atypical chest pain. Electrocardiogram (ECG) showed de novo saddleback ST‐segment elevation from V1 to V3 consistent with a type‐2 Brugada pattern (BrP). The heart rate ranged between 50 and 60 beats/minute, similar to ambulatory heart rate. Twenty hours post‐PCI, the patient developed multiple episodes of polymorphic ventricular tachycardia and fibrillation requiring defibrillation (six shocks in total, over a period of 3 hours). No fever or ionic disturbances were documented. Echocardiography showed good biventricular function, with regional wall motion abnormalities (LCx and RCA territories). The ECG after the second shock revealed a type‐1 BrP. Two primary causes for the VAs were considered: acute ischemia and BrS. Initially, ischemia was thought to be more likely and esmolol and amiodarone were administered. As this approach was ineffective, treatment was switched to isoproterenol (1 ug/min perfusion) and overdriving ventricular pacing at the right ventricle, at 80 beats/minute, as BrS was presumed to be more likely. The last episode of VA occurred a few minutes after isoproterenol was initiated. Emergent coronary angiography excluded acute ischemia. VAs were successfully suppressed by these measures. Three days later, angioplasty of the ostial LAD was performed and the temporary pacemaker was removed due to sustained electrical stability. Isoproterenol was progressively withdrawn, and oral quinidine 200 mg four times a day was initiated (progressively tapered to a final dose of 200 mg twice daily which was maintained at hospital discharge). Magnetic resonance showed subendocardial late gadolinium enhancement in the lateral and posterior walls (42 mm), consistent with ischemic cardiomyopathy. The patient received a cardioverter defibrillator and was discharged home 12 days after the initial PCI. Genetic screening for the most common mutations associated with BrS (CACNA1C, CACNB2b, SCN10A, and SCN5A genes) was negative. After 1‐year follow‐up, the patient remains free of cardiovascular events, under oral quinidine. No significant arrhythmias were detected on routine ICD evaluations and resting ECGs shows no BrP. BrS has been associated with mutations in inward Na^+^ (*I* ~Na~) and L‐type Ca^2+^ channels that decrease phase‐0 rapid depolarization and mutations in the transient outward K^+^ (*I* ~to~) channel that increase phase‐1 rapid repolarization.[^1^](#joa312382-bib-0001){ref-type="ref"} Although the mechanisms are incompletely understood, functional re‐entry is commonly accepted as the basis for arrhythmogenesis.[^1^](#joa312382-bib-0001){ref-type="ref"} *I* ~to~ expression is greater in the right ventricle (RV) outflow tract and decreases from the epicardium to the endocardium, generating a transmural gradient that, in pathological conditions such as decreased phase‐0 rapid repolarization or gain‐of‐function of *I~to~* channels, provides a substrate for reentrant arrhythmias.[^1^](#joa312382-bib-0001){ref-type="ref"} Because of the particularities of its treatment, timely diagnosis of BrS is mandatory. Once this diagnosis was assumed, proper treatment was initiated with a favorable response. The efficacy of isoprosterenol lies in the fact that it augments L‐type calcium channels, preventing loss of action potential dome. Overdrive pacing may play a role in the termination of VAs, by inducing an extrastimuli and stopping the reentrant circuit. Moreover, *I* ~to~ currents are less prominent at faster heart rates; thus, by increasing the basal heart rate, rapid pacing may have a role in decreasing the risk of arrhythmia generation. Quinidine seems to have good long‐term effectiveness in preventing VAs due to its *I* ~to~ blocking properties. We also questioned the underlying trigger of the VAs. While Brugada phenocopies have been reported during RCA angiography and similarities between the generation of ischemic ST‐segment elevation and BrP have been proposed.[^2^](#joa312382-bib-0002){ref-type="ref"} This is, to the best of our knowledge, the first report of a de novo BrP unmasked after PCI associated with the clinical manifestation of VAs. Acute ischemia of the conus branch can induce BrP and precipitate VAs.[^3^](#joa312382-bib-0003){ref-type="ref"} However, the fact the clinical manifestations developed several hours after intervention and the coronary angiogram remained unchanged weight against the theory that acute ischemia was the trigger. The mechanisms by which the ST‐segment changes vary over time and the arrhythmias that are triggered are also poorly understood. In this case, there seems to be a direct myocardial insult. Moreover, the patient reported mild chest pain after PCI of a proximal RCA lesion, suggesting reperfusion injury and justifying the involvement of the RV. Swelling‐activated chloride channels (*I* ~Cl,swell~) have been reported to be activated during ischemia/reperfusion, generating an outward Cl^−^, that has been suggested to decrease phase‐0 inward sodium currents (*I* ~Na~).[^4^](#joa312382-bib-0004){ref-type="ref"} We hypothesize that in this case, the activation of *I* ~Cl,swell~ (further) decreased phase‐0 *I* ~Na~, triggering the VAs. It remains unclear, however, whether ischemia--reperfusion injury would be enough to justify this clinical presentation, particularly outside the context of an acute coronary syndrome, or if a pre‐existing channelopathy could have been exacerbated after PCI. Interestingly, a recent case series suggested that CAD‐related but non‐ischemia‐driven drug‐refractory VAs may have a similar background to BrS (and other arrhythmogenic syndromes with a structurally normal heart) and display a remarkable response to quinidine, although in this study there is no mention of a BrP preceding the VAs.[^5^](#joa312382-bib-0005){ref-type="ref"} This case highlights the importance of early recognition of the precipitating mechanisms of electrical storm and raises the possibility that reperfusion injury may precipitate VAs in patients with BrS. CONFLICT OF INTEREST {#joa312382-sec-0002} -------------------- The authors have no conflict of interest to declare.

\section{Introduction} \subsection{Main results and Outline}\label{introduction-main-results-section} Let $(G,X)$ be a Shimura datum of Hodge type, i.e.~it is equipped with an embedding $(G,X)\hookrightarrow (\operatorname{GSp}(V,\psi),S^{\pm})$, where $V$ is a $\mathbb{Q}$-vector space equipped with a symplectic pairing $\psi$. The embedding of Shimura data induces an embedding of Shimura varieties $\operatorname{Sh}_K(G,X)\hookrightarrow\operatorname{Sh}_{K'}(\operatorname{GSp},S^{\pm})$, where $K'\subset\operatorname{GSp}(\mathbb{A}_f)$. The moduli interpretation of the Siegel modular variety $\operatorname{Sh}_{K'}(\operatorname{GSp},S^{\pm})$ naturally gives rise to an integral model $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. We consider the integral model $\mathscr{S}_K(G,X)$ of $\operatorname{Sh}_K(G,X)$ with hyperspecial (resp.~parahoric) level structure, as constructed in \cite{Kisin-integral-model} (resp.~\cite{Kisin-Pappas}), which is initially defined as the normalization of the closure of $\operatorname{Sh}_K(G,X)$ inside $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. In this article, we show that this construction can be simplified, in that the normalization step is redundant, and that $\mathscr{S}_K(G,X)$ is simply the closure of $\operatorname{Sh}_K(G,X)$ inside $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. Our main theorem is the following, which is independent of the choice of symplectic embeddings. \begin{theorem}\label{Main-Theorem-intro} For $K\subset G(\mathbb{A}_f)$ small enough and hyperspecial, there exists some $K'\subset\operatorname{GSp}(\mathbb{A}_f)$, such that we have a closed embedding (``the Hodge embedding'') \[\mathscr{S}_K(G,X)\hookrightarrow\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})\] Consequentially, the normalization step $\mathscr{S}_K(G,X)\xrightarrow{\nu}\mathscr{S}_K^-(G,X)$ is redundant as the closure $\mathscr{S}_K^-(G,X)$ is already smooth, and the integral model $\mathscr{S}_K(G,X)$ has a moduli interpretation inherited from that of $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. \end{theorem} In particular, the Hodge morphism is a closed embedding in the PEL case, where we consider integral models constructed in \cite{Kottwitz} (resp.~\cite{Rapoport-Zink}). To see the result in that case (see \cite{PEL-embedding} for details), recall that the Hodge morphism is given by forgetting the $\mathcal{O}_B$-action on an abelian scheme $\mathcal{A}$, where $B$ is a semisimple $\mathbb{Q}$-algebra attached to the PEL moduli problem. Let $T^{(p)}(\mathcal{A})$ be the prime-to-$p$ Tate module. For a point on $\mathscr{S}_K(G,X)$, the corresponding level structure $\eta:V\otimes \mathbb{A}_f^p\xrightarrow{\sim}T^{(p)}(\mathcal{A})\otimes\mathbb{A}_f^p$ is compatible with the $\mathcal{O}_B$-actions. This shows that the $\mathcal{O}_B$-action on $\mathcal{A}$ is already determined by $\eta$, and hence that the Hodge morphism is an embedding. Strictly speaking, this argument only applies when we let the level structure away from $p$ go to zero, but it is not hard to deduce Theorem \ref{Main-Theorem-intro} from this. In the general Hodge type case, the mod $p$ points of the integral model $\mathscr{S}_K(G,X)$ can be interpreted as abelian varieties equipped with certain ``mod $p$ Hodge cycles'', which come from reduction mod $p$ of Hodge cycles in characteristic zero. We denote the mod $p$ Hodge cycle at a mod $p$ point $x\in\mathscr{S}_K(G,X)$ by a tuple $(s_{\alpha,\ell,x},s_{\alpha,\operatorname{cris},x})$, which is determined by either its $\ell$-adic \'etale component or its cristalline component (by Proposition \ref{final-num-cris-triviality-motivated}). This is analogous to the case of Hodge cycles in characteristic $0$, which are determined by either their \'etale components or their de Rham components. More specifically, let $\mathscr{S}_K^-(G,X)$ be the closure of $\operatorname{Sh}_K(G,X)$ in $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. By a criterion in \cite{Kisin-mod-p-points} (resp.~\cite{Rong-mod-p}), two mod $p$ points $x,x'\in\mathscr{S}_K(G,X)(k)$ that have the same image in $\mathscr{S}_K^-(G,X)(k)$ are equal if and only if $s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'}$. Therefore, to show that the normalization morphism is an isomorphism, it reduces to proving the following statement on cohomological tensors: \begin{prop}\label{key-tensor-implication} $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}\Longrightarrow s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'}$. \end{prop} By a CM lifting result on $\mathscr{S}_K(G,X)$ due to \cite{Kisin-mod-p-points} (resp.~\cite{Rong-mod-p}), these cohomological tensors lift, up to $G$-isogenies, to Hodge cycles on CM abelian varieties. A key observation is that when two mod $p$ points $x,x'\in \mathscr{S}_K(G,X)(k)$ map to the same image in $\mathscr{S}_K^-(G,X)(k)$, they can be CM-lifted using the same torus, whose cocharacter induces the filtration on the Dieudonn\'e modules $\mathbb{D}(\mathcal{A}_x)=\mathbb{D}(\mathcal{A}_{x'})$ which then identifies the filtrations on the Dieudonn\'e modules associated to CM-liftable mod $p$ points, giving rise to an isogeny \textit{in characteristic zero} between the two CM lifts. This observation allows us to match up the mod $p$ cristalline tensors using the input from $\ell$-adic \'etale tensors, precisely due to the rationality of Hodge cycles in characteristic zero and the existence of an isogeny lift in characteristic zero. It is worth pointing out that, in the case where the aforementioned cohomological tensors are algebraic--for example, at points where the Hodge conjecture is true--the family of Hodge cycles (tensors) $s_{\alpha}$ that naturally lives over the Hodge type integral model $\mathscr{S}_K(G,X)$ becomes a flat family of algebraic cycles over $\mathscr{S}_K(G,X)$. In this case, $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}$ implies that the two algebraic cycles corresponding to the two $\ell$-adic cycles are $\ell$-adic cohomologically equivalent, hence numerically equivalent, and we only need to show that they are also cristalline-cohomologically equivalent. Recall that the Grothendieck Standard Conjecture D says that numerical equivalence and cohomological equivalence agree for algebraic cycles. The proof of \ref{key-tensor-implication} thus follows from a cristalline realisation of this Standard Conjecture D, for points on the integral model of Hodge type and their associated cristalline tensors, which are mod $p$ Hodge cycles. Our result essentially establishes, unconditionally, rationality for mod $p$ Hodge cycles that live on mod $p$ points of Hodge type Shimura varieties. Finally, we state the following two analogues of our Theorem \ref{Main-Theorem-intro}. Firstly, in the case of parahoric integral models constructed in \cite{Kisin-Pappas}, we impose mild technical assumptions from \cite[6.18]{Rong-mod-p}. We certainly expect this technical assumption from \textit{loc.cit.} to be eventually unnecessary in our Theorem below. \begin{theorem}\label{parahoric-analogue-intro} Let $G_{\mathbb{Q}_p}$ be residually split, and $K$ a parahoric level structure. There exists a closed embedding $\mathscr{S}_K(G,X)\hookrightarrow\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$ of integral models, for some suitable $K'$. \\ In particular, the normalization step in the construction of $\mathscr{S}_K(G,X)$ is redundant as the closure $\mathscr{S}_K^-(G,X)$ is already normal. \end{theorem} The second analogue concerns toroidal compactifications of integral models of Hodge type constructed in \cite{Keerthi-compactification} (the PEL cases were constructed earlier in \cite{Lan-thesis}). Combining our main theorem \ref{Main-Theorem-intro} with an analysis from \cite{Lan-immersion} on the boundary components of toroidal compactifications, one immediately obtains the following result. \begin{Coro}\label{toroidal-intro} Let $(G,X)$ be a Shimura datum of Hodge type. For each $K\subset G(\mathbb{A}_f)$ sufficiently small\footnote{For parahoric level, assume the same condition as in Theorem \ref{parahoric-analogue-intro}.}, there exist collections $\Sigma$ and $\Sigma'$ of cone decompositions, and $K'\subset\operatorname{GSp}(\mathbb{A}_f)$, such that we have a closed embedding of toroidal compactifications of integral models \[\mathscr{S}_K^{\Sigma}(G,X)\hookrightarrow\mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})\] extending the Hodge embedding of integral models. \\ In particular, the normalization step is redundant, and $\mathscr{S}_K^{\Sigma}(G,X)$ can be constructed by simply taking the closure of $\operatorname{Sh}_K(G,X)$ inside $\mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})$. \end{Coro} The construction of smooth (resp.~normal) integral models of Shimura varieties plays an important part in the Langlands program. For a more detailed historical exposition, see \cite{Kisin-integral-model, Kisin-mod-p-points}. On the other hand, such results as Theorems \ref{Main-Theorem-intro}, \ref{parahoric-analogue-intro} and \ref{toroidal-intro} have been useful in various other aspects of arithmetic geometry, e.g.~in the construction of $p$-adic $L$-functions using Euler systems techniques, in the arithmetic intersection theory of special cycles on Shimura varieties and their integral models as in the Kudla-Rapoport program, arithmetic Gan-Gross-Prasad Program etc. \subsection{Organization} In Chapter \ref{Hodge-type-Chapter}, we recall the basic theory of Hodge type integral models as in \cite{Kisin-integral-model,Kisin-mod-p-points} (resp.~\cite{Kisin-Pappas,Rong-mod-p}), and explain how the question of whether the normalization morphism is an isomorphism essentially reduces to proving the key proposition \ref{key-tensor-implication}. In section \ref{CM-lifting-integral-models}, we review CM lifting results on the integral model $\mathscr{S}_K(G,X)$, and prove a few lemmas specific to our setting. In section \ref{finishing-up-the-proof-section}, we prove the key result \ref{key-tensor-implication} (see \ref{final-num-cris-triviality-motivated}), which leads to our Main Theorems \ref{Main-Theorem-intro} and \ref{parahoric-analogue-intro} (see Corollary \ref{coro-normalization-isomorphism}). We give the toroidal compactification version in section \ref{toroidal-cpct-section}. \textbf{Acknowledgments.} I would like to thank my advisor Mark Kisin for suggesting this problem to me, and for his continued encouragement. I would also like to thank Ben Howard for an observation that led to a key improvement in the arguments, and Keerthi Madapusi Pera for helpful conversations. The author is supported by Harvard University graduate student fellowships. \section{Integral models of Hodge type}\label{Hodge-type-Chapter} This section reviews the theory of integral models of Hodge type from \cite{Kisin-integral-model,Kisin-mod-p-points}. Note that the $p=2$ case is addressed in \cite{2-adic-integral-model}, henceforward we will not emphasize the difference between the $p>2$ and the $p=2$ cases in our expositions. Our expositions will mainly focus on the hyperspecial case, and will mention the parahoric analogues from \cite{Kisin-Pappas,Rong-mod-p} whenever necessary for our proofs. \subsection{Setup and notations} \begin{numberedparagraph}\label{Hodge-type-notations-1st} We fix a $\mathbb{Q}$-vector space $V$ with a perfect alternating pairing $\psi$. For any $\mathbb{Q}$-algebra $R$, denote by $V_R=V\otimes_{\mathbb{Q}}R$. Let $\operatorname{GSp}=\operatorname{GSp}(V,\psi)$ be the corresponding group of symplectic similitudes, and let $S^{\pm}$ be the Siegel double space. We fix an embedding of Shimura data $i:(G,X)\hookrightarrow(\operatorname{GSp},S^{\pm})$ and assume that $K_p\subset G(\mathbb{Q}_p)$ is hyperspecial, i.e. $K_p=G_{\mathbb{Z}_{(p)}}(\mathbb{Z}_p)$ for some reductive group $G_{\mathbb{Z}_{(p)}}$ over $\mathbb{Z}_{(p)}$ with generic fibre $G$. By \cite[2.3.1, 2.3.2]{Kisin-integral-model} the embedding $i$ of Shimura data is induced by an embedding $G_{\mathbb{Z}_{(p)}}\hookrightarrow \operatorname{GL}(V_{\mathbb{Z}_{(p)}})$ for some $\mathbb{Z}_{(p)}$-lattice $V_{\mathbb{Z}_{(p)}}\subset V$. By Zarhin's trick, up to replacing $V_{\mathbb{Z}_{(p)}}$ by $\operatorname{Hom}_{\mathbb{Z}_{(p)}}(V_{\mathbb{Z}_{(p)}},V_{\mathbb{Z}_{(p)}})^4$, we can assume that $\psi$ also induces a perfect pairing on $V_{\mathbb{Z}_{(p)}}$ which we again denote by $\psi$. For any $\mathbb{Z}_{(p)}$-algebra $R$, we denote $V_R=V_{\mathbb{Z}_{(p)}}\otimes_{\mathbb{Z}_{(p)}}R$. \\ Take $K_p'=\operatorname{GSp}(V_{\mathbb{Z}_{(p)}})(\mathbb{Z}_p)\subset\operatorname{GSp}(\mathbb{Q}_p)$. For each compact open $K^p\subset G(\mathbb{A}_f^p)$, by \cite[2.1.2]{Kisin-integral-model}, there exists a compact open $K'^p\subset\operatorname{GSp}(\mathbb{A}_f^p)$ such that $K'^p\supset K^p$ and that the embedding $i$ of Shimura data induces an embedding \[\operatorname{Sh}_K(G,X)\hookrightarrow\operatorname{Sh}_{K'}(\operatorname{GSp},S^{\pm})\] of $E$-schemes, where $K'=K'_pK'^p$ and $K=K_pK^p$ and $E=E(G,X)$ is the reflex field. \end{numberedparagraph} \begin{numberedparagraph} Let $\mathcal{B}$ be an abelian scheme over a $\mathbb{Z}_{(p)}$-scheme $T$, and we define the ``prime-to-$p$ Tate module'' to be the \'etale local system on $T$ given by $\widehat{V}^p(\mathcal{B})=\underset{p\nmid n}{\varprojlim}\mathcal{B}[n]$. We denote ``the rational Tate module away from $p$'' by $\widehat{V}^p(\mathcal{B})_{\mathbb{Q}}=\widehat{V}^p(\mathcal{B})\otimes_{\mathbb{Z}}\mathbb{Q}$. We work in the localized category of the category of abelian schemes over $T$, where the morphisms are given by $\operatorname{Hom}$ groups in the usual category tensored with $\mathbb{Z}_{(p)}$. We call an object in this category an \textit{abelian scheme up to prime to $p$ isogeny}. An isomorphism in this category will be called a $p'$-quasi-isogeny. Let $\mathcal{A}$ be an abelian scheme up to prime to $p$ isogeny, and let $\mathcal{A}^*$ be the dual abelian scheme. By a \textit{weak polarization} we mean an equivalence class of $p'$-quasi-isogenies $\lambda: \mathcal{A}\xrightarrow{\sim}\mathcal{A}^*$ such that some multiple of $\lambda$ is a polarization, and two weak polarizations are equivalent if they differ by an element of $\mathbb{Z}_{(p)}^{\times}$. \end{numberedparagraph} \begin{numberedparagraph}\label{moduli-integral-model} For an abelian scheme up to prime to $p$ isogeny with a weak polarization, i.e. a pair $(\mathcal{A},\lambda)$, we denote by $\underline{\operatorname{Isom}}(V_{\mathbb{A}_f^p},\widehat{V}^p(\mathcal{A})_{\mathbb{Q}})$ the \'etale sheaf on $T$ consisting of isomorphisms $V_{\mathbb{A}_f^p}\xrightarrow{\sim}\widehat{V}^p(\mathcal{A})_{\mathbb{Q}}$ which are compatible with the pairings induced by $\psi$ and $\lambda$ up to an $\mathbb{A}_f^{p\times}$-scalar. We define a $K'^p$-level structure on $(\mathcal{A},\lambda)$ to be a section \begin{equation}\label{K'-p-level-structure} \epsilon_{K'}^p\in \Gamma(T,\underline{\operatorname{Isom}}(V_{\mathbb{A}_f^p},\widehat{V}^p(\mathcal{A})_{\mathbb{Q}})/K'^p) \end{equation} We consider the following functor from the category $Sch_{/\mathbb{Z}_{(p)}}$ of $\mathbb{Z}_{(p)}$-schemes to the category of sets \begin{align*} Sch_{/\mathbb{Z}_{(p)}}&\to Sets\\ T&\mapsto \{\text{Isomorphism classes of triples }(\mathcal{A}/T, \lambda, \epsilon_{K'}^p)\} \end{align*} For $K'^p$ sufficiently small, the above functor is representable by a smooth $\mathbb{Z}_{(p)}$-scheme, which we denote by $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})_{\mathbb{Z}_{(p)}}$, whose $\mathbb{Q}$-fibre is precisely the Siegel modular variety $\operatorname{Sh}_{K'}(\operatorname{GSp},S^{\pm})$. Let $\mathcal{O}:=\mathcal{O}_E$ denote the ring of integers of the reflex field $E=E(G,X)$. Let $v|p$ be a prime of $E$ lying over $p$, and denote by $\mathcal{O}_{(v)}$ the localization at $v$. We denote by $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})_{\mathcal{O}_{(v)}}$ the base change of $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})_{\mathbb{Z}_{(p)}}$ to $\mathcal{O}_{(v)}$. Denote by $\mathscr{S}_K^{-}(G,X)$ the closure of $\operatorname{Sh}_K(G,X)$ in $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})_{\mathcal{O}_{(v)}}$, and by $\mathscr{S}_K(G,X)$ the normalization of $\mathscr{S}_K^-(G,X)$. We call the normalization $\mathscr{S}_K(G,X)$ the \textit{integral model of Hodge type} for the Hodge type Shimura datum $(G,X)$. In the following, we shall denote by $\nu:\mathscr{S}_K(G,X)\to\mathscr{S}_K^-(G,X)$ the normalization morphism. \end{numberedparagraph} \subsection{Cohomological tensors over the integral model}\label{section-cohomological-tensors} \begin{numberedparagraph}\label{tensor-stablizer} Recall from \ref{Hodge-type-notations-1st} that the embedding of Shimura data $i$ is induced by a closed embedding $G_{\mathbb{Z}_{(p)}}\hookrightarrow\operatorname{GL}(V_{\mathbb{Z}_{(p)}})$ for some $\mathbb{Z}_{(p)}$-lattice $V_{\mathbb{Z}_{(p)}}\subset V$. By \cite[Proposition 1.3.2.]{Kisin-integral-model}, there exists a finite collection of tensors $(s_{\alpha})\subset V_{\mathbb{Z}_{(p)}}^{\otimes}$ such that the subgroup $G_{\mathbb{Z}_{(p)}}\subset \operatorname{GL}(V_{\mathbb{Z}_{(p)}})$ is the scheme-theoretic stabilizer of these tensors $(s_{\alpha})$. The moduli interpretation of $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$ as the moduli space of polarized abelian schemes gives us a universal abelian scheme $\mathscr{A}\to \mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. Therefore, we can pullback this universal abelian scheme to an abelian scheme \begin{equation}\label{abelian-scheme-over-integral-model} h:\mathscr{A}\to\mathscr{S}_K(G,X) \end{equation} which, by abuse of notation, we still denote as $\mathscr{A}$. Let $V_B=R^1h_*^{\operatorname{an}}\mathbb{Z}_{(p)}$ and $s_{\alpha}$ have Betti realisations $s_{\alpha,B}\in V_B^{\otimes}$. Consider the first relative de Rham cohomology $\mathcal{V}=R^1h_*\Omega^{\bullet}$ of $\mathscr{A}$. The $s_{\alpha}$ can be viewed as parallel sections of the complex analytic vector bundle associated to $\mathcal{V}^{\otimes}$, which lie in the $\operatorname{Fil}^0$ part of the Hodge filtration. By \cite[Propositions 2.2.2, 2.3.9.]{Kisin-integral-model}, these sections have de Rham realisations $s_{\alpha,\operatorname{dR}}\in\mathcal{V}^{\otimes}$ defined over $\mathcal{O}_{(v)}$. On the other hand, for a prime $\ell\neq p$, consider the \'etale local system $\mathcal{V}_{\ell}$ with $\mathbb{Q}_{\ell}$-coefficients on $\mathscr{S}_K(G,X)$ given by $\mathcal{V}_{\ell}:=R^1{h_{\operatorname{\acute{e}t}}}_*\mathbb{Q}_{\ell}$, where $h_{\operatorname{\acute{e}t}}$ is the map on \'etale sites induced from $h$. The tensors $s_{\alpha}$ have $\ell$-adic \'etale realisations $s_{\alpha,\ell}\in\mathcal{V}_{\ell}^{\otimes}$ which descend to $\mathcal{O}_{(v)}$. Moreover, for an $\mathcal{O}_{(v)}$-scheme $T$ and $x\in\mathscr{S}_K(G,X)(T)$, we define $s_{\alpha,\ell,x}$ to be the pullback of the section $s_{\alpha,\ell}$ (defined over $\mathcal{O}_{(v)}$) to $T$. We denote by $\mathcal{A}_x$ the pullback of $\mathscr{A}$ to $x$, thus $s_{\alpha,\ell,x}\in H_{\operatorname{\acute{e}t}}^1(\mathcal{A}_{x,\overline{\kappa}},\mathbb{Q}_{\ell})^{\otimes}$, where $\kappa:=\kappa(x)$. At the prime $p$, we consider the local system $\mathcal{V}_p:=R^1{h_{\eta\operatorname{\acute{e}t}}}_*\mathbb{Z}_p$, where $h_{\eta}$ the generic fibre of $h$, and the tensors $s_{\alpha}$ have $p$-adic \'etale realisations $s_{\alpha,p}\in \mathcal{V}_{p}^{\otimes}$ which descend to $E$. Likewise, for an $E$-scheme $T$ and $x\in \mathscr{S}_K(G,X)(T)$, we define $s_{\alpha,p,x}$ to be the pullback of $s_{\alpha,p}$ to $T$. Thus $s_{\alpha,p,x}\in H_{\operatorname{\acute{e}t}}^1(\mathcal{A}_{x,\overline{\kappa}},\mathbb{Z}_p)^{\otimes}$ where $\kappa:=\kappa(x)$. \end{numberedparagraph} \begin{numberedparagraph}\label{promoting-section-paragraph} For a point $x\in \operatorname{Sh}_K(G,X)(T)\hookrightarrow\operatorname{Sh}_{K'}(\operatorname{GSp},S^{\pm})(T)$, then the image of $x$ corresponds to a triple $(\mathcal{A}_x/T,\lambda,\epsilon_{K'}^p)$, where $\mathcal{A}_x$ is an abelian scheme up to prime-to-$p$ isogeny over $T$ and $\epsilon_{K'}^p$ is a section as defined in \ref{K'-p-level-structure}. For any finite index subgroup $K'_1$ such that $K\subset K_1'\subset K'$, since we have the embedding \[\operatorname{Sh}_K(G,X)\hookrightarrow\operatorname{Sh}_{K_1'}(\operatorname{GSp},S^{\pm})\] (the image of) $x$ also gives rise to a point of $K_1'$-level structure, i.e. a section $\epsilon_{K_1'}^p\in \Gamma(T,\underline{\operatorname{Isom}}(V_{\mathbb{A}_f^p},\widehat{V}^p(\mathcal{A}_x)_{\mathbb{Q}})/{K_1'}^p)$. We define \begin{equation}\label{promoted-section} \epsilon_K^p:=\underset{\substack{K_1' \text{s.t.} \\ K\subset K_1'\subset K'}}{\varprojlim}\epsilon_{K_1'}^p\in \Gamma(T,\underline{\operatorname{Isom}}(V_{\mathbb{A}_f^p},\widehat{V}^p(\mathcal{A}_x)_{\mathbb{Q}})/K) \end{equation} Therefore, given a section $\epsilon_{K'}^p$, we can ``promote'' it to a section $\epsilon_K^p$ as defined in \ref{promoted-section} above, for $K\subset K'$. \end{numberedparagraph} \begin{remark}\label{promoted-level-structure-ell-adic} The ``promoted'' level structures are realized in terms of the $\ell$-adic tensors, in the sense that $\epsilon_K^p\colon s_{\alpha}\mapsto (s_{\alpha,\ell})_{\ell\neq p}$. \end{remark} \begin{numberedparagraph}\label{fields-notations-integral-model} We fix an algebraic closure $\overline{\mathbb{Q}}$ of $\mathbb{Q}$. For each place $v$ of $\mathbb{Q}$, we also fix an algebraic closure $\overline{\mathbb{Q}}_v$ of $\mathbb{Q}_v$ and embeddings $\overline{\mathbb{Q}}\hookrightarrow\overline{\mathbb{Q}}_v$. Let $\overline{\mathbb{F}}_p$ be the residue field of $\overline{\mathbb{Q}}_p$. Denote $L:=\operatorname{Frac} W(\overline{\mathbb{F}}_p)$ and $\mathbb{Q}_p^{\operatorname{unr}}\subset L$ the subfield of elements algebraic over $\mathbb{Q}_p$. We choose a fixed algebraic closure $\overline{L}$ of $L$, and an embedding $\overline{\mathbb{Q}}_p\hookrightarrow \overline{L}$ of $\mathbb{Q}_p^{\operatorname{unr}}$-algebras. Recall the notation $E:=E(G,X)$ for the reflex field of $(G,X)$. Let $E_p$ be the completion of $E$ at the prime corresponding to $\overline{\mathbb{Q}}\hookrightarrow\overline{\mathbb{Q}}_p$, and $k\subset\overline{\mathbb{F}}_p$ be a subfield containing the residue field $k_E$ of $E_p$. Let $W:=W(k)$ and $K\subset\overline{L}$ be a finite, totally ramified extension of $W[1/p]$. Take $x\in \mathscr{S}_K(G,X)(k)$ and let $\Tilde{x}\in \operatorname{Sh}_K(G,X)(K)$ be a point that specializes to $x$. Let $\overline{K}$ be the algebraic closure of $K$ in $\overline{L}$. As before, $\mathcal{A}_{\Tilde{x}}$ is the pullback of $\mathscr{A}$ to $\Tilde{x}$, and $\mathcal{A}_{\Tilde{x},\overline{K}}$ the geometric fibre of $\mathscr{A}$ over $\Tilde{x}$. Thus we have $p$-adic \'etale tensors $s_{\alpha,p,\Tilde{x}}\in H_{\operatorname{\acute{e}t}}^1(\mathcal{A}_{\Tilde{x},\overline{K}},\mathbb{Z}_p)^{\otimes}$, which are $\operatorname{Gal}(\overline{K}/K)$-invariant. \end{numberedparagraph} We will need the following result. \begin{lem}\label{Kisin-mod-p-Prop-1.3.7.}(\cite[Proposition 1.3.7.]{Kisin-mod-p-points}) (1) Under the $p$-adic comparison isomorphism \[H_{\operatorname{\acute{e}t}}^1(\mathcal{A}_{\widetilde{x},\overline{K}},\mathbb{Z}_p)\otimes_{\mathbb{Z}_p}B_{\operatorname{cris}}\xrightarrow{\sim}H_{\operatorname{cris}}^1(\mathcal{A}_x/W)\otimes_WB_{\operatorname{cris}}\] the $s_{\alpha,p,\Tilde{x}}$ map to $\varphi$-invariant tensors $s_{\alpha,\operatorname{cris},\Tilde{x}}\in \operatorname{Fil}^0(H_{\operatorname{cris}}^1(\mathcal{A}_x/W)^{\otimes})$. \\ (2) There is a $W$-linear isomorphism \[H_{\operatorname{\acute{e}t}}^1(\mathcal{A}_{\Tilde{x},\overline{K}},\mathbb{Z}_p)\otimes_{\mathbb{Z}_p}W\xrightarrow{\sim}H_{\operatorname{cris}}^1(\mathcal{A}_x/W)\] taking $s_{\alpha,p,\Tilde{x}}$ to $s_{\alpha,\operatorname{cris},\Tilde{x}}$. In particular, the $s_{\alpha,\operatorname{cris},\Tilde{x}}$ define a reductive group scheme $G_W\subset \operatorname{GL}(H_{\operatorname{cris}}^1(\mathcal{A}_x/W))$ which is isomorphic to $G_{\mathbb{Z}_{(p)}}\otimes_{\mathbb{Z}_{(p)}}W$.\\ (3) The filtration on $H_{\operatorname{cris}}(\mathcal{A}_x/W)\otimes_Wk$ is given by $\mu_0^{-1}$ where $\mu_0$ is a $G_W$-valued cocharacter conjugate to the Hodge cocharacter $\mu_h$ for $h\in X$. \end{lem} \begin{numberedparagraph}\label{normalization-same-image-point} Let $x,\widetilde{x}$ be defined as above. Suppose we also have another point $\widetilde{x}'\in \mathscr{S}_K(G,X)(K)$ which specializes to $x'\in \mathscr{S}_K(G,X)(k)$. Suppose moreover that both $x,x'\in\mathscr{S}_K(G,X)(k)$ map to the same image $\overline{x}\in\mathscr{S}_K^-(G,X)(k)$ under the normalization map $\nu$. We shall make use of the following property about the irreducible components of $\mathscr{S}_K^-(G,X)$. In the hyperspecial case, we have: \begin{lem}\label{Kisin-Prop-2.3.5}(\cite[Proposition 2.3.5]{Kisin-integral-model}) Let $x\in\mathscr{S}_K^-(G,X)$ be a closed point with characteristic $p$ residue field. Denote by $\widehat{U}_x$ the completion of $\mathscr{S}_K^-(G,X)$ at $x$. Then the irreducible components of $\widehat{U}_x$ are formally smooth over $\mathcal{O}_{(v)}$. \end{lem} In the parahoric case, the irreducible components are normal. \begin{lem}\label{irreducible-components-parahoric}(\cite[Prop.4.2.2]{Kisin-Pappas}) Let $\widehat{U}_{\overline{x}}$ be the completion of $\mathscr{S}_K^-(G,X)_{\mathcal{O}_{E^{\operatorname{unr}}}}$ at $\overline{x}$. Then the irreducible component of $\widehat{U}_{\overline{x}}$ containing $x$ is isomorphic to $\widehat{M}_{G,y}^{\mathrm{loc}}$ (which are normal by \cite{Pappas-Zhu}) as formal schemes over $\mathcal{O}_{E^{\operatorname{unr}}}$. \end{lem} In \ref{Kisin-Prop-2.3.5} and \ref{irreducible-components-parahoric}, the term ``irreducible components'' of the formal scheme $\hat{U}_x$ refers to the irreducible components of the rigid analytic space attached to $\hat{U}_x$. See \cite[$\mathsection$ 7]{deJong-formal-rigid} for details. An immediate corollary of the proof for \ref{Kisin-Prop-2.3.5} (resp.~\ref{irreducible-components-parahoric}) is the following ``cristalline criterion,'' which forms the basis for the proof of main theorem \ref{Main-Theorem-intro}. \begin{lem}\label{Kisin-1.3.11}(hyperspecial level \cite[Prop 1.3.9., Corollary 1.3.11]{Kisin-mod-p-points}; parahoric level \cite[Corollary 6.3]{Rong-mod-p})\\ (a) $s_{\alpha,\operatorname{cris},\Tilde{x}}$ depends only on $x$ and not on $\Tilde{x}$. (Thus we will write $s_{\alpha,\operatorname{cris},x}$ in place of $s_{\alpha,\operatorname{cris},\Tilde{x}}$ from now on.)\\ (b) Let $x,x'\in\mathscr{S}_K(G,X)(k)$ be two points having the same image in $\mathscr{S}_K^-(G,X)(k)$. Then $x=x'$ if and only if $s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'}$. \end{lem} We will also need the following result, which goes into the proof of the CM lifting theorem \ref{Kisin-CM-lifting}. \begin{lem}\label{Kisin-1.1.19}\cite[1.1.19]{Kisin-mod-p-points} Let $\mu:\mathbb{G}_m\to G$ be a cocharacter, defined over $K$ and conjugate to $\mu_0$. Suppose that $\mu^{-1}$ induces an admissible filtration on $\mathbb{D}(\mathscr{G})_K:=\mathbb{D}(\mathscr{G})(W)\otimes_WK$. Then there exists a finite extension $K'/K$ with residue field $k'$, a $p$-divisible group $\widetilde{\mathscr{G}}'$ over $\mathcal{O}_{K'}$, and a quasi-isogeny $\theta:\mathscr{G}\to\mathscr{G}'$ where $\mathscr{G}'=\widetilde{\mathscr{G}}'\otimes_{\mathcal{O}_{K'}}k'$ such that \begin{enumerate} \item $\theta$ identifies the filtration on $\mathbb{D}(\mathscr{G}')_K$ corresponding to $\widetilde{\mathscr{G}}'$ and the filtration on $\mathbb{D}(\mathscr{G})_K$ induced by $\mu^{-1}$. \item $\theta$ identifies $\mathbb{D}(\mathscr{G}')$ with $g\cdot \mathbb{D}(\mathscr{G})$ for some $g\in G(L)$ with \[g^{-1}b\sigma(g)\in G(\mathcal{O}_L)p^{v_0}G(\mathcal{O}_L).\] \item Viewing $s_{\alpha,\operatorname{cris}}\in \mathbb{D}(\mathscr{G}')^{\otimes}$ via $\theta$, the deformation $\widetilde{\mathscr{G}}'$ of $\mathscr{G}'$ is $gG_{W(k')}g^{-1}$-adapted, where $gG_{W(k')}g^{-1}$ is the stabilizer of $s_{\alpha,\operatorname{cris}}\in\mathbb{D}(\mathscr{G}')^{\otimes}$. \end{enumerate} \end{lem} The parahoric analogue of this lemma can be extracted from \cite[$\mathsection$3]{Kisin-Pappas}. \end{numberedparagraph} \subsection{A few more remarks on tensors}\label{remark-on-tensors-section} \begin{numberedparagraph} We continue to use the notations as in the previous sections. To emphasize the fact that $\mathscr{S}_K^-(G,X)$ depends on a choice of $K'\subset\operatorname{GSp}(\mathbb{A}_f)$, we denote it, briefly for now, by $\mathscr{S}_{K,K'}^-(G,X)$ instead. First we establish the following lemma, see also Corollary \ref{coro-normalization-isomorphism}. \begin{lem}\label{same-image-forall-K'} One of the following two situations always holds: either \begin{enumerate} \item there exists a sufficiently small $K'\subset\operatorname{GSp}(\mathbb{A}_f)$ such that $\mathscr{S}_K(G,X)\cong \mathscr{S}_{K,K'}^-(G,X)$; or \item there exist two points $x,x'\in\mathscr{S}_K(G,X)(k)$ which have the same image $\overline{x}_{K'}\in\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$ for all $K'$ containing $K$. \end{enumerate} \end{lem} \begin{proof} For each $K'$ containing $K$, let $U_{K'}\subset\mathscr{S}_{K,K'}^-(G,X)$ be the largest open set of points where the normalization $\nu$ is an isomorphism. Let $Z_{K'}$ be its complement, which shrinks when $K'$ shrinks. Since $\mathscr{S}_{K,K'}^-(G,X)$ is Noetherian, the decreasing sequence $\{Z_{K'}\}_{K'}$ stabilizes at some small enough $K_{*}'$. There are then two possibilities: (1) if $Z_{K_*'}=\varnothing$, then $\mathscr{S}_K(G,X)\cong \mathscr{S}^-_{K,K_*'}(G,X)$; (2) if $Z_{K_*'}\neq\varnothing$, then we can take a closed point $\overline{x}\in Z_{K_*'}$, hence $\overline{x}\in Z_{K'}$ for any $K'$ containing $K$, such that $\nu$ is not an isomorphism at $\overline{x}$. Therefore, when we are not in situation (1), there exist two points $x,x'\in\mathscr{S}_K(G,X)$ that map to $\overline{x}\in\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$ for all $K'$ containing $K$. \end{proof} Therefore, to prove Theorem \ref{Main-Theorem-intro}, it suffices to treat only case (2) in \ref{same-image-forall-K'}. We now show that case (2) reduces to an equality of the $\ell$-adic tensors $s_{\alpha,\ell}$. Let $\widetilde{x}\in \mathscr{S}_K(G,X)(K)$ be a characteristic $0$ point that specializes to $x\in\mathscr{S}_K(G,X)(k)$; likewise, let $\widetilde{x}'\in \mathscr{S}_K(G,X)(K)$ be a point that specializes to $x'\in\mathscr{S}_K(G,X)(k)$. \begin{lem}\label{ell-adic-tensor-equality} Fix a level $K\subset G(\mathbb{A}_f)$. If $x,x'\in\mathscr{S}_K(G,X)(k)$ map to the same image point $\overline{x}_{K'}\in\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})(k)$ for all $K'$ containing $K$, then $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}$. \end{lem} \begin{proof} For each $K'$ containing $K$, by \ref{promoting-section-paragraph} we have sections $\epsilon_{K',\widetilde{x}}$ and $\epsilon_{K',\widetilde{x}'}$ which promote to sections $\epsilon_{K,\widetilde{x}}^p$ and $\epsilon_{K,\widetilde{x}'}^p$ respectively. Since $x$ and $x'$ map to the same image point $\overline{x}_{K'}\in\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$, by \ref{promoted-level-structure-ell-adic}, we have $s_{\alpha,\ell,x}\equiv s_{\alpha,\ell,x'}\mod K'$. Therefore, $s_{\alpha,\ell,x}\equiv s_{\alpha,\ell,x'}\mod\bigcap\limits_{K\subset K'}K'=K$. Therefore, $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}$ since they are both invariant under $K$ by construction. \end{proof} \begin{remark} If $x,x'\in\mathscr{S}_K(G,X)(k)$ map to the same image point $\overline{x}\in\mathscr{S}_K^-(G,X)(k)$ under the normalization map $\nu$, and $\widetilde{x},\widetilde{x}'\in\mathscr{S}_K(G,X)(K)$ are points that specialize to $x,x'$ respectively. Then $\mathcal{A}_{x}=\mathcal{A}_{\overline{x}}=\mathcal{A}_{x'}$; moreover, $\mathcal{A}_{\widetilde{x}}$ reduce mod $p$ to $\mathcal{A}_x$, and $\mathcal{A}_{\widetilde{x}'}$ reduce mod $p$ to $\mathcal{A}_{x'}$, i.e. $\mathcal{A}_{\widetilde{x}}$ and $\mathcal{A}_{\widetilde{x}'}$ have the same mod $p$ reduction. \end{remark} Now by Lemmas \ref{Kisin-1.3.11}, \ref{same-image-forall-K'} and \ref{ell-adic-tensor-equality}, to show that the normalization morphism is an isomorphism, it suffices to show that: (i.e. Proposition \ref{key-tensor-implication}) \begin{equation*} s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}\Longrightarrow s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'} \end{equation*} \begin{remark} If we knew that both the Hodge conjecture for abelian varieties and the Grothendieck standard conjecture D for abelian varieties in characteristic $p$ are true, then our result follows trivially. \end{remark} \end{numberedparagraph} \section{Proof of Main Theorem} \subsection{CM lifting on the integral model}\label{CM-lifting-integral-models} \begin{numberedparagraph} First we recall some CM lifting results on the integral model $\mathscr{S}_K(G,X)$ from \cite[$\mathsection$2]{Kisin-mod-p-points}, whose notations we adopt. Let $v$ be a miniscule cocharacter. Consider the affine Deligne-Lusztig variety \[X_v(b)=\{g\in G(L)/G(\mathcal{O}_L): g^{-1}b\sigma(g)\in G(\mathcal{O}_L)p^vG(\mathcal{O}_L)\}.\] For an arbitrary point $x\in\mathscr{S}_K(G,X)(k)$, consider the abelian variety $\mathcal{A}_x$ and its $p$-divisible group $\mathscr{G}_x$. Consider the Dieudonne module $\mathbb{D}(\mathscr{G}_x):=\mathbb{D}(\mathscr{G}_x)(\mathcal{O}_L)$. Since $\mathbb{D}(\mathscr{G}_x)$ is a Dieudonn\'e module, $v$ acting on $\mathbb{D}(\mathscr{G}_x)$ has non-negative weights and induces a minuscule cocharacter of $\operatorname{GL}(\mathbb{D}(\mathscr{G}_x))$. If $g\in X_v(b)$, then $g\cdot \mathbb{D}(\mathscr{G}_x)$ is stable under Frobenius and satisfies the axioms of a Dieudonn\'e module. Thus $g\cdot\mathbb{D}(\mathscr{G}_x)$ corresponds to the Dieudonn\'e module of a $p$-divisible group $\mathscr{G}_{gx}$, which is naturally equipped with a quasi-isogeny $\mathscr{G}_x\to\mathscr{G}_{gx}$, corresponding to the natural isomorphism $g\cdot \mathbb{D}(\mathscr{G}_x)\otimes_{\mathbb{Z}_p}\mathbb{Q}_p\xrightarrow{\sim}\mathbb{D}(\mathscr{G}_x)\otimes_{\mathbb{Z}_p}\mathbb{Q}_p$. Note that by the definition of $G$-invariant tensors, we have \begin{equation} s_{\alpha,\operatorname{cris},x}=g(s_{\alpha,\operatorname{cris},x})\in (g\mathbb{D}(\mathscr{G}_x))^{\otimes}=\mathbb{D}(\mathscr{G}_{gx})^{\otimes}. \end{equation} Denote by $\mathcal{A}_{gx}$ the abelian variety corresponding to $\mathscr{G}_{gx}$. It is isogenous to $\mathcal{A}_x$ by construction, and comes equipped with a canonical $K'$-level structure $\epsilon_{K',g\cdot x}$ induced from the level structure on $\mathcal{A}_x$, and a weak polarization $\lambda_{g\cdot x}$ induced from the weak polarization on $\mathcal{A}_{x}$ since $G\subset \operatorname{GSp}$. We define a map \begin{align}\label{ADLV-map-to-Siegel} X_v(b)&\to \mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})(\overline{\mathbb{F}}_p)\\ g&\mapsto (\mathcal{A}_{gx},\lambda_{gx}) \end{align} \begin{lem}\label{lifting-isogeny-hyperspecial}\cite[Prop.1.4.4.]{Kisin-mod-p-points} There is a unique lifting of \ref{ADLV-map-to-Siegel} to a map \begin{equation}\label{iota-x-1.4.4} \iota_x:X_v(b)\to\mathscr{S}_K(G,X)(\overline{\mathbb{F}}_p) \end{equation} such that $s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},\iota_x(g)}\in \mathbb{D}(\mathscr{G}_{gx})^{\otimes}$. \end{lem} Note that the uniqueness here simply follows from Lemma \ref{Kisin-1.3.11}(b). Moreover, \ref{iota-x-1.4.4} extends to a $\langle\Phi\rangle\times Z_G(\mathbb{Q}_p)\times G(\mathbb{A}_f^p)$-equivariant map \begin{equation}\label{} \iota_x:X_v(b)\times G(\mathbb{A}_f^p)\to\mathscr{S}_{K_p}(G,X)(\overline{\mathbb{F}}_p) \end{equation} We call the image of $\iota_x$ the \textit{$G$-isogeny class} of $x$. Let $\delta\in G(K_0)$ be as in \cite[2.1.2]{Kisin-mod-p-points}, where $K_0=W(k)[1/p]$. More specifically, we fix an isomorphism \[\mathbb{D}(\mathscr{G}_x)\xrightarrow{\sim}V_{\mathbb{Z}_{(p)}}^*\otimes_{\mathbb{Z}_{(p)}}W\] which takes $s_{\alpha,\operatorname{cris},x}$ to $s_{\alpha}$. Thus we have identifications \[G_{\mathbb{Z}_{(p)}}\otimes_{\mathbb{Z}_{(p)}}K_0=G_{K_0}\xrightarrow{\sim}G(s_{\alpha,\operatorname{cris},x})\subset \operatorname{GL}(\mathbb{D}(\mathscr{G}_x)_{K_0})\xrightarrow{\sim}\operatorname{GL}(H_{\operatorname{cris}}^1(\mathcal{A}_x/W)_{K_0}),\] where $G(s_{\alpha,\operatorname{cris},x})\subset\operatorname{GL}(\mathbb{D}(\mathscr{G}_x)_{K_0})$ denotes the subgroup defined by $s_{\alpha,\operatorname{cris},x}$. Therefore, the Frobenius on $\mathbb{D}(\mathscr{G}_x)$ has the form $\delta\sigma$ with $\delta\in G(K_0)$, where $\sigma$ is the absolute Frobenius on $W(\overline{\mathbb{F}}_p)$. Note that the element $\delta$ is independent of the choices made in its construction, up to $\sigma$-conjugacy by elements of $G(W)$. \end{numberedparagraph} \begin{numberedparagraph} As in \cite[$\mathsection$ 2.1.2]{Kisin-mod-p-points}, attached to a point $x\in\mathscr{S}_K(G,X)(k)$, let $\gamma_{\ell}\in G(\mathbb{Q}_{\ell})$ be the geometric $q$-Frobenius in $\operatorname{Gal}(\overline{\mathbb{F}}_p/k)$ acting on $H_{\operatorname{\acute{e}t}}^1(\mathcal{A}_{\overline{x}},\mathbb{Q}_{\ell})$ and $I_{\ell/k}=I_{\ell,x/k}\subset G_{\mathbb{Q}_{\ell}}$ the centralizer of $\gamma_{\ell}$. Let $I_{\ell,n}$ be the centralizer of $\gamma_{\ell}^n$ in $G_{\mathbb{Q}_{\ell}}$, which forms a decreasing sequence in $G_{\mathbb{Q}_{\ell}}$ and stabilizes to what we shall denote as $I_{\ell}=I_{\ell,x}$ for $n$ sufficiently large. Likewise, we can define $I_{p/k}=I_{p,x/k}$ as the algebraic group over $\mathbb{Q}_p$ whose $R$-points for a $\mathbb{Q}_p$-algebra $R$ is given by \begin{equation}\label{defining-Ipk} I_{p/k}(R):=\{\alpha\in G(W\otimes_{\mathbb{Z}_p}R):\delta\sigma(\alpha)=\alpha\delta\}. \end{equation} For each positive integer $n$, let $k_n\subset\overline{\mathbb{F}}_p$ be the degree-$n$ extension of $k$ and we define $I_{p,n}$ by replacing $W$ with $W(k_n)$ in \ref{defining-Ipk}. For $n$ sufficiently large, the increasing sequence $I_{p,n}$ (as subgroups inside the group $J_{\delta}$ defined in \cite[1.2.12]{Kisin-mod-p-points}) stabilizes and we denote this group by $I_p:=I_{p,x}$. We write \begin{equation}\label{defining-Ix} I_x\subset\operatorname{Aut}_{\mathbb{Q}}(\mathcal{A}_{x}\otimes_k\overline{\mathbb{F}}_p)=\operatorname{Aut}_{\mathbb{Q}}(\mathcal{A}_{\overline{x}}\otimes_k\overline{\mathbb{F}}_p) \end{equation} for the subgroup whose points $I_x(R)$, for a $\mathbb{Q}$-algebra $R$, consist of those elements of $\operatorname{Aut}_{\mathbb{Q}}(\mathcal{A}_x\otimes_k\overline{\mathbb{F}}_p)(R)$ fixing the tensors $s_{\alpha,\operatorname{cris},x}$ and $s_{\alpha,\ell,x}$ for all $\ell\neq p$. Similarly, attached to the point $x'\in\mathscr{S}_K(G,X)(k)$, we define $I_{x'}$ analogously. Let $\varphi$ be the Frobenius. The following Lemma is essentially the analogous result to \cite[Main Theorem]{Tate-endo-ab-var-fin-fields} for the Hodge cycles $s_{\alpha}$. \begin{lem}\label{rank-I-equals-rank-Iell}\cite[2.1.7]{Kisin-mod-p-points} For some prime $\ell\neq p$, $I_{\mathbb{Q}_{\ell}}=I\otimes_{\mathbb{Q}}\mathbb{Q}_{\ell}$ contains the connected component of the identity in $I_{\ell}$. In particular, $\operatorname{rank} I=\operatorname{rank} I_{\ell}=\operatorname{rank} G$. \end{lem} The following CM lifting Theorem (and its proof) is crucial to our proofs. Thus we also sketch its proof. \begin{lem}\label{Kisin-CM-lifting} (\cite[Theorem 2.2.3.]{Kisin-mod-p-points}) The isogeny class $\iota_x(X_v(\delta)\times G(\mathbb{A}_f^p))$ contains a point which is the reduction of a special point on $\operatorname{Sh}_K(G,X)$. \end{lem} \begin{proof} (Sketch) We take a maximal torus $T\subset I_p$ defined over $\mathbb{Q}_p$. Since $I$ and $I_p$ have the same rank, we can assume that $T$ is induced by a maximal torus $T$ in the $\mathbb{Q}$-group $I$, and the induced action of $T$ on $\mathbb{D}(\mathscr{G}_{\widetilde{x}})_K$ respects filtrations, where $\mathscr{G}_{\widetilde{x}}$ is a $G_W$-adapted deformation such that the filtration on $\mathbb{D}(\mathscr{G}_{\widetilde{x}})_K$ is given by $\mu_T^{-1}$ (by Lemma \ref{Kisin-1.1.19}). One can then check that $T$ is in fact a maximal torus in $G$ and that $\widetilde{x}$ is a special point because the Mumford-Tate group commutes with $T$ and is hence a subgroup of $T$. \end{proof} \end{numberedparagraph} In the parahoric case, the analogue of Lemma \ref{lifting-isogeny-hyperspecial} is given in \cite[Prop.6.5]{Rong-mod-p} (from whence we inherited the assumption on $G_{\mathbb{Q}_p}$), where the isogeny classes are parametrized by a certain $X(\sigma\{\mu_y\},b)$, which is a certain union of affine Deligne-Lusztig varieties over certain $\mu$-admissible set. The parahoric version of the CM lifting theorem \ref{Kisin-CM-lifting} is given in \cite[Theorem 9.4]{Rong-mod-p} under the same assumption on $G_{\mathbb{Q}_p}$ as Prop.~6.5 \textit{loc.cit.} \subsection{Finishing up the proof}\label{finishing-up-the-proof-section} \begin{numberedparagraph}\label{last-section-recall-CM-lifting} Recall the setting from \ref{remark-on-tensors-section} that we start with two mod $p$ points $x,x'\in\mathscr{S}_K(G,X)(k)$ on the normalized integral model that map to the same image $\overline{x}\in\mathscr{S}_K^-(G,X)(k)$ for all $K'\supset K$. In particular, we have $\mathcal{A}_{x}=\mathcal{A}_{x'}=\mathcal{A}_{\overline{x}}$ by pulling back the abelian scheme $\mathscr{A}\to\mathscr{S}_K(G,X)$ to the point $x$ or $x'$. Consider the isogeny class $\iota_{x}(X_v(\delta)\times G(\mathbb{A}_f^p))$. By \ref{Kisin-CM-lifting}, there exists a point $y:=\iota_{x}(g)\in \iota_{x}(X_v(\delta)\times G(\mathbb{A}_f^p))$, for some $g\in X_v(\delta)\times G(\mathbb{A}_f^p)$, such that $\mathcal{A}_y$ is $G$-isogenous to $\mathcal{A}_x$ and such that $s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},y}\in\mathbb{D}(\mathscr{G}_{gx})^{\otimes}$, and such that $y$ is the reduction of a special point $\widetilde{y}\in \operatorname{Sh}_K(G,X)(K')$. \begin{lem}\label{Ix=Ix'} Let $x,x'\in\mathscr{S}_K(G,X)(k)$ be as in \ref{last-section-recall-CM-lifting} and $I_x, I_{x'}$ as defined in \ref{defining-Ix}. We have $I_x=I_{x'}$ as subgroups of $\operatorname{Aut}_{\mathbb{Q}}(\mathcal{A}_{\overline{x}}\otimes_k\overline{\mathbb{F}}_p)$. \end{lem} \begin{proof} By Lemma \ref{rank-I-equals-rank-Iell}, we have \[I_x\otimes_{\mathbb{Q}}\mathbb{Q}_{\ell}=\{g\in\operatorname{Aut}_{\mathbb{Q}_{\ell}}(H^1_{\operatorname{\acute{e}t}}(\mathcal{A}_{x,\overline{k}},\mathbb{Q}_{\ell}))\big| gs_{\alpha,\ell,x}=s_{\alpha,\ell,x}\text{ for all }\ell\neq p, g\varphi=\varphi g \},\] and likewise \[I_{x'}\otimes_{\mathbb{Q}}\mathbb{Q}_{\ell}=\{g\in\operatorname{Aut}_{\mathbb{Q}_{\ell}}(H^1_{\operatorname{\acute{e}t}}(\mathcal{A}_{x',\overline{k}},\mathbb{Q}_{\ell}))\big| gs_{\alpha,\ell,x'}=s_{\alpha,\ell,x'}\text{ for all }\ell\neq p, g\varphi=\varphi g\}.\] Since $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}$ for all $\ell\neq p$ by Lemma \ref{ell-adic-tensor-equality}, we have $I_{x}\otimes_{\mathbb{Q}}\mathbb{Q}_{\ell}=I_{x'}\otimes_{\mathbb{Q}}\mathbb{Q}_{\ell}$, thus we have $I_x=I_{x'}\subset\operatorname{Aut}_{\mathbb{Q}}(\mathcal{A}_{\overline{x}}\otimes_{\mathbb{Q}}\overline{\mathbb{F}}_p)$. \end{proof} \end{numberedparagraph} \begin{numberedparagraph} By the proof of Lemma \ref{Kisin-CM-lifting}, to make such a CM lift $\widetilde{y}$ of point $y$, one finds a maximal torus $T_x$ in $I_x$. The induced action of $T_x$ on $\mathbb{D}(\mathscr{G}_{\widetilde{y}})_{K'}$ respects filtrations, and thus the action of $T_x$ on $\mathcal{A}_y$ lifts to $\mathcal{A}_{\widetilde{y}}$. \begin{lem} There exists a special point lift $\widetilde{y}'$ of $x'$ (up to isogeny), such that there is an isogeny $\mathcal{A}_{\widetilde{y}}\to \mathcal{A}_{\widetilde{y}'}$. \end{lem} \begin{proof} Since $I_x=I_{x'}$ by Lemma \ref{Ix=Ix'}, we can take the same torus $T:=T_x=T_{x'}\subset I_x=I_{x'}$ to construct the CM liftings for $x$ and $x'$, i.e.~$\mathcal{A}_{\widetilde{y}}$ and $\mathcal{A}_{\widetilde{y}'}$. By Lemma \ref{Kisin-1.1.19}, the filtration on $\mathbb{D}(\mathscr{G}_y)$ corresponding to the deformation $\mathscr{G}_{\widetilde{y}}$ is identified via the isogeny $\mathscr{G}_y\to \mathscr{G}_x$ (induced by $g\in X_v(\delta)$) with the filtration on $\mathbb{D}(\mathscr{G}_x)=\mathbb{D}(\mathscr{G}_{x'})$ induced by $\mu_{T_x}=\mu_{T_{x'}}$, and this filtration is then identified with that on $\mathbb{D}(\mathscr{G}_{y'})$ induced by the deformation $\mathscr{G}_{\widetilde{y}'}$ (of $\mathscr{G}_{y'}$). In particular, the filtration induced by $\mathcal{A}_{\widetilde{y}}$ on $\mathbb{D}(\mathscr{G}_y)$ is identified with the filtration induced by $\mathcal{A}_{\widetilde{y}'}$ on $\mathbb{D}(\mathscr{G}_{\widetilde{y}'})$, thus we obtain a priori a map between $p$-divisible groups $\mathscr{G}_{\widetilde{y}}\to \mathscr{G}_{\widetilde{y}'}$ whose reduction mod $p$ is precisely the isogeny $\mathscr{G}_y\to\mathscr{G}_{y'}$ given by the composition $\mathscr{G}_y\to\mathscr{G}_x=\mathscr{G}_{x'}\to \mathscr{G}_{y'}$ of $G$-isogenies. (Note that $\mathscr{G}_y\to\mathscr{G}_{y'}$ is not \textit{a priori} a $G$-isogeny.) Since we have a map between the mod $p$ abelian varieties $\mathcal{A}_y\to \mathcal{A}_{y'}$ via composing the $G$-isogenies to and from $\mathcal{A}_x=\mathcal{A}_{x'}$, and a lifting of this map for $p$-divisible groups $\mathscr{G}_{\widetilde{y}}\to \mathscr{G}_{\widetilde{y}'}$, by Serre-Tate theory, we then obtain an isogeny between abelian varieties $\mathcal{A}_{\widetilde{y}}\to \mathcal{A}_{\widetilde{y}'}$. \end{proof} \begin{lem}\label{mapping-tildey-to-tildey'-Hodge-cycles} The isogeny (but not \textit{a priori} $G$-isogeny) $\mathcal{A}_{\widetilde{y}}\to\mathcal{A}_{\widetilde{y}'}$ sends the Hodge cycle $(s_{\alpha,\ell,\widetilde{y}},s_{\alpha,\operatorname{dR},\widetilde{y}})$ to the Hodge cycle $(s_{\alpha,\ell,\widetilde{y}'},s_{\alpha,\operatorname{dR},\widetilde{y}'})$. \end{lem} \begin{proof} Under the $G$-isogeny $\mathcal{A}_y\to \mathcal{A}_x$, the tensors $s_{\alpha,\ell,y}$ get sent to $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}$, which then get sent to $s_{\alpha,\ell,y'}$ under the $G$-isogeny $\mathcal{A}_{x'}\to\mathcal{A}_{y'}$. In particular, the isogeny (but not \textit{a priori} a $G$-isogeny) $\mathcal{A}_y\to\mathcal{A}_{y'}$, which factors through $\mathcal{A}_x=\mathcal{A}_{x'}$, sends $s_{\alpha,\ell,y}$ to $s_{\alpha,\ell,y'}$. Under the specialization isomorphism $H^*_{\operatorname{\acute{e}t}}(\mathcal{A}_{y,\overline{k}},\mathbb{Q}_{\ell})\to H^*_{\operatorname{\acute{e}t}}(\mathcal{A}_{\widetilde{y},\overline{K}},\mathbb{Q}_{\ell})$, the isogeny $\mathcal{A}_{\widetilde{y}}\to\mathcal{A}_{\widetilde{y}'}$ sends $s_{\alpha,\ell,\widetilde{y}}$ to $s_{\alpha,\ell,\widetilde{y}'}$. Since $\mathcal{A}_{\widetilde{y}}\to\mathcal{A}_{\widetilde{y}'}$ is an isogeny in characteristic zero and Hodge cycles in characteristic zero are determined by either its $\ell$-adic \'etale or de Rham components (as they both come from the Betti realizations), the isogeny $\mathcal{A}_{\widetilde{y}}\to\mathcal{A}_{\widetilde{y}'}$ (again not \textit{a priori} known to be a $G$-isogeny) sends the Hodge cycle $(s_{\alpha,\ell,\widetilde{y}},s_{\alpha,\operatorname{dR},\widetilde{y}})$ to $(s_{\alpha,\ell,\widetilde{y}'},s_{\alpha,\operatorname{dR},\widetilde{y}'})$. \end{proof} \begin{Coro}\label{y-to-y'-cris} The isogeny (again not \textit{a priori} $G$-isogeny) $\mathcal{A}_y\to \mathcal{A}_{y'}$ sends $s_{\alpha,\operatorname{cris},y}$ to $s_{\alpha,\operatorname{cris},y'}$. \end{Coro} \begin{proof} By Lemma \ref{mapping-tildey-to-tildey'-Hodge-cycles}, the mod $p$ reduction $\mathcal{A}_y\to\mathcal{A}_{y'}$ of $\mathcal{A}_{\widetilde{y}}\to \mathcal{A}_{\widetilde{y}'}$ sends the cristalline realizations $s_{\alpha,\operatorname{cris},y}$ to $s_{\alpha,\operatorname{cris},y'}$, via the specialization isomorphism $H^*_{\operatorname{dR}}(\mathcal{A}_{\widetilde{y}})\to H^*_{\operatorname{cris}}(\mathcal{A}_y/W)\otimes_WK$. \end{proof} Now we are ready to prove the key proposition. \begin{prop}\label{final-num-cris-triviality-motivated} $s_{\alpha,\ell,x}=s_{\alpha,\ell,x'}\Longrightarrow s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'}$ \end{prop} \begin{proof} Combining Lemmas \ref{Ix=Ix'} through \ref{y-to-y'-cris}, the isogeny $\mathcal{A}_y\to\mathcal{A}_{y'}$ sends $s_{\alpha,\operatorname{cris},y}$ to $s_{\alpha\operatorname{cris},y'}$. On the other hand, this isogeny factors through the $G$-isogeny $\mathcal{A}_y\to \mathcal{A}_x$ which sends $s_{\alpha,\operatorname{cris},y}\mapsto s_{\alpha,\operatorname{cris},x}$ and $G$-isogeny which sends $s_{\alpha,\operatorname{cris},x'}\mapsto s_{\alpha,\operatorname{cris},y'}$, thus we must also have $s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'}$ (otherwise the image of $s_{\alpha,\operatorname{cris},y}$ via this composition of isogenies $\mathcal{A}_y\to\mathcal{A}_x=\mathcal{A}_{x'}\to \mathcal{A}_{y'}$ would get sent to something other than $s_{\alpha,\operatorname{cris},y'}$, causing a contradiction). \end{proof} \end{numberedparagraph} \begin{remark} The Proposition \ref{final-num-cris-triviality-motivated} essentially suggests that the mod $p$ points of the integral model of Hodge type can be interpreted as abelian varieties equipped with a well-defined notion of ``mod $p$ Hodge cycles,'' written as tuples $(s_{\alpha,\ell,x},s_{\alpha,\operatorname{cris},x})$, which are determined by either their $\ell$-adic \'etale components or their cristalline components. This is analogous to the case with absolute Hodge cycles in characteristic $0$, which are determined by either their \'etale components or their de Rham components. \end{remark} \begin{Coro}\label{coro-normalization-isomorphism} The normalization morphism $\nu$ is an isomorphism. In particular, $\mathscr{S}_{K}(G,X)$ admits a closed embedding into $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$. \end{Coro} \begin{proof} Resume the setting in \ref{remark-on-tensors-section}. For any two mod $p$ points $x,x'\in\mathscr{S}_K(G,X)(k)$ that map to the same image in $\mathscr{S}_K^-(G,X)(k)$ for all $K'$ containing $K$, by Lemma \ref{ell-adic-tensor-equality} and Proposition \ref{final-num-cris-triviality-motivated}, we have $s_{\alpha,\operatorname{cris},x}=s_{\alpha,\operatorname{cris},x'}$. Therefore, by Lemma \ref{Kisin-1.3.11}, we have $x=x'$. Therefore, the normalization morphism $\nu$ is injective on $k$-points for any $k\subset\overline{\mathbb{F}}_p$ that contains the residue field $k_E$ of $E_p$ (\ref{fields-notations-integral-model}). On the other hand, since the source and target of $\nu$ also have the same generic fibre $\operatorname{Sh}_K(G,X)$, this implies that $\mathscr{S}_K^-(G,X)$ is unibranch. Therefore, by definition ([EGA4, Chapter IV (6.15.1)]), the local ring $U_x$ of $\mathscr{S}_K^-(G,X)$ at $x$ is unibranch. Thus by \cite[Tag0C2E]{stacks-project}, the complete local ring $\hat{U}_x$ only has one irreducible component. By Lemma \ref{Kisin-Prop-2.3.5} (resp.~Lemma \ref{irreducible-components-parahoric}), each irreducible component of the complete local ring $\hat{U}_x$ is formally smooth (resp.~normal), thus each $\hat{U}_x$ is formally smooth (resp.~normal) over $\mathcal{O}_{(v)}$. Therefore, $\mathscr{S}_K^-(G,X)$ is also smooth (resp.~normal) over $\mathcal{O}_{(v)}$. In particular, the normalization morphism $\nu$ is an isomorphism of schemes, and that the scheme-theoretic closure of $\operatorname{Sh}_K(G,X)$ inside $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})_{\mathcal{O}_{(v)}}$ is already smooth (resp.~normal). \end{proof} \begin{remark} Note that in particular Corollary \ref{coro-normalization-isomorphism} does not depend on the choice of a symplectic embedding $(G,X)\hookrightarrow(\operatorname{GSp},S^{\pm})$. \end{remark} \begin{numberedparagraph} In particular, in the construction of the parahoric integral model $\mathscr{S}_{K^{\circ}}(G,X)$ in \cite[$\mathsection 4.3.6$]{Kisin-Pappas}, we can simply define it as the normalization of $\mathscr{S}_K^-(G,X)$ in $\operatorname{Sh}_{K^{\circ}}(G,X)$. This reduces the number of normalization steps by one. We briefly recall the notations, and refer the reader to \textit{loc. cit.} for the details. We fix a point $x\in\mathcal{B}(G,\mathbb{Q}_p)$, i.e. the Bruhat-Tits building of $G$, and let $\mathcal{G}=\mathcal{G}_x$ be the corresponding smooth $\mathbb{Z}_p$-group scheme whose generic fibre is $G$, and such that $\mathcal{G}^0$ is a parahoric group scheme. We write $K_p^{\circ}=\mathcal{G}^{\circ}(\mathbb{Z}_p)$ and $K^{\circ}=K_p^{\circ}K^p$. The parahoric integral model $\mathscr{S}_{K^{\circ}}(G,X)$ with parahoric level structure $K^{\circ}$ is a priori defined as the normalization of $\mathscr{S}_K(G,X)$ in $\operatorname{Sh}_{K^{\circ}}(G,X)$, but can now be simplified as simply the normalization of $\mathscr{S}_K^-(G,X)$ in $\operatorname{Sh}_{K^{\circ}}(G,X)$, as above. \end{numberedparagraph} \begin{numberedparagraph} The fact that the mod $p$ Hodge cycles (tensors) are determined by either its $\ell$-adic or its cristalline component (Proposition \ref{final-num-cris-triviality-motivated}) can be thought of as a ``rationality'' statement. As a historical remark, we remind the reader of the following rationality conjecture of Deligne's, which is a weakened form of the Hodge conjecture. \begin{conj} \textit{(Deligne)} Suppose two abelian varieties $A_1$ and $A_2$ defined over $\overline{\mathbb{Q}}$ have the same reduction $A$ over $\overline{\mathbb{F}}_p$, and suppose there are given Hodge cycles $\xi_1$ and $\xi_2$ on $A_1$ and $A_2$, respectively. Then the intersection number of the reductions $\overline{\xi}_{1,\operatorname{\acute{e}t}}$ and $\overline{\xi}_{2,\operatorname{\acute{e}t}}$ is rational. \end{conj} In our specific situation for points on $\mathscr{S}_K(G,X)$, given $\mathcal{A}_{\widetilde{x}}$ and $\mathcal{A}_{\widetilde{x}'}$ equipped with Hodge cycles $(s_{\alpha,\widetilde{x}})$ and $(s_{\alpha,\widetilde{x}'})$ respectively. Suppose $s_{\alpha,\widetilde{x}}$ and $s_{\alpha',\widetilde{x}'}$ have \textit{complementary degrees} viewed as cohomology classes on powers of $\mathcal{A}_{\widetilde{x}}$ and $\mathcal{A}_{\widetilde{x}'}$ respectively. Then Deligne's rationality conjecture is automatic in this case: since $s_{\alpha,\widetilde{x}}$ deforms to a Hodge cycle on $\mathcal{A}_{\widetilde{x}'}$ (and likewise $s_{\alpha',\widetilde{x}'}$ deforms to a Hodge cycle on $\mathcal{A}_{\widetilde{x}}$), the intersection number $s_{\alpha,\ell,x}\cup s_{\alpha',\ell,x'}$ is rational by rationality in characteristic zero. The same is true at $p$, i.e. $s_{\alpha,\operatorname{cris},x}\cup s_{\alpha,\operatorname{cris},x'}$ is rational. \end{numberedparagraph} \subsection{Toroidal compactifications of integral models}\label{toroidal-cpct-section} We briefly mention one application of our main theorem to the toroidal compactifications of integral models of Hodge type constructed in \cite{Keerthi-compactification}. The main input is an analysis on the boundary components from \cite{Lan-immersion}, where the level structure is taken to be \textit{neat}, i.e. sufficiently small. We remark that while Lan's result is conditional on the existence of an embedding on the open part, our result is unconditional (since we proved the embedding on the open part). \begin{numberedparagraph} We adopt the notations \textit{loc. cit.}. Let $\mathscr{S}_K^{\Sigma}(G,X)=\coprod \mathcal{Z}_{[\sigma]}$ be a stratification into locally closed subschemes, where $\sigma\in \Sigma_{\mathcal{Z}}^+$. Let $\mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})=\coprod \mathcal{Z}'_{[\tau]}$ be a stratification, where $\tau\in {\Sigma^{'+}_{\mathcal{Z}'}}$. Note that the Hodge morphism $\mathscr{S}_K(G,X)\to\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$ extends uniquely to a morphism $\mathscr{S}_K^{\Sigma}(G,X)\to \mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})$ between their toroidal compactifications. In particular, the following diagram commutes: \begin{equation}\label{compactification-functoriality} \begin{tikzcd}\mathscr{S}_K^{\Sigma}(G,X)\arrow[]{r}{}&\mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})\\ \mathscr{S}_K(G,X)\arrow[hook]{u}{}\arrow[]{r}{}&\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})\arrow[hook]{u}{}\end{tikzcd} \end{equation} \end{numberedparagraph} \begin{Coro}\label{toroidal-cpct-corollary} Let $(G,X)$ be a Shimura datum of Hodge type. For each $K\subset G(\mathbb{A}_f)$ sufficiently small\footnote{When the level structure is parahoric, one needs to also impose the assumptions as in \ref{parahoric-analogue-intro}}, there exist collections $\Sigma$ and $\Sigma'$ of cone decompositions, and $K'\subset\operatorname{GSp}(\mathbb{A}_f)$, such that we have a closed embedding \begin{equation}\label{toroidal-cpct-embedding} \mathscr{S}_K^{\Sigma}(G,X)\hookrightarrow\mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm}) \end{equation} extending the Hodge embedding of integral models. \\ In particular, the normalization step is redundant, and $\mathscr{S}_K^{\Sigma}(G,X)$ can be constructed by simply taking the closure of $\operatorname{Sh}_K(G,X)$ inside $\mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})$. \end{Coro} \begin{proof} The result follows immediately by combining \cite[Thm 2.2]{Lan-immersion} with our main Theorem \ref{Main-Theorem-intro}. For the reader's convenience, we sketch the argument. Let $x'\in \mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})$ be any point that lies in the stratum $\mathcal{Z}'_{[\tau]}$. Then \'etale locally at $x'$, \ref{compactification-functoriality} becomes \begin{equation}\label{compactification-functoriality-local} \begin{tikzcd} \coprod\limits_j \big(E_{\mathcal{Z},j}(\sigma_j)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j}\big)\arrow[]{r}{}& E_{\mathcal{Z}'}(\tau)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'}\\ \coprod\limits_j \big(E_{\mathcal{Z},j}\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j}\big)\arrow[]{r}{}\arrow[hook]{u}{}&E_{\mathcal{Z}'}\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'}\arrow[hook]{u}{} \end{tikzcd} \end{equation} More precisely, there exists an \'etale neighborhood $\overline{U}'\to \mathscr{S}_{K'}^{\Sigma'}(\operatorname{GSp},S^{\pm})$ of $x'$ and an \'etale morphism $\overline{U}'\to E_{\mathcal{Z}'}(\tau)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'}$ which pullback via the Hodge morphism to \'etale morphisms $\overline{U}\to\mathscr{S}_K^{\Sigma}(G,X)$ and $\overline{U}\to \coprod\limits_j (E_{\mathcal{Z},j}(\sigma_j)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j})$. By \cite[Prop 4.9]{Lan-immersion}, there exist ``strictly compatible collections'' (see \cite[Definition 4.6]{Lan-immersion}) $\Sigma$ and $\Sigma'$ of cone decompositions with respect to the Hodge morphism $\Phi$. By Theorem \ref{Main-Theorem-intro} (resp.~Theorem \ref{parahoric-analogue-intro}), the Hodge morphism \[\Phi: \mathscr{S}_K(G,X)\to\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})\] is a closed embedding. Therefore, diagram \ref{compactification-functoriality-local} gives a closed embedding \begin{equation}\label{lower-horizontal-compactification-functoriality-local} \coprod\limits_j (E_{\mathcal{Z},j}\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j})\hookrightarrow E_{\mathcal{Z}'}\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'} \end{equation} over certain \'etale neighborhood $\overline{U}'$ of $x'$ (see \cite[3.11]{Lan-immersion}). This then implies that \begin{equation}\label{embedding-on-C-toroidal} C_{\mathcal{Z},j}\to C_{\mathcal{Z}'} \end{equation} are closed embeddings over the image of $\overline{U}'$ in $C_{\mathcal{Z}'}$ for all $j$. On the other hand, one can check directly from the construction, as in \cite[Lemma 4.3]{Lan-immersion}, that \begin{equation}\label{embedding-on-Esigma-toroidal} E_{\mathcal{Z},j}(\sigma_j)\hookrightarrow E_{\mathcal{Z}'}(\tau) \end{equation} is a closed embedding. Therefore for each $j$, by \ref{embedding-on-C-toroidal} and \ref{embedding-on-Esigma-toroidal} the map \[E_{\mathcal{Z},j}(\sigma_j)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j}\hookrightarrow E_{\mathcal{Z}'}(\tau)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'}\] is a closed embedding over the image of $\overline{U}'$. To show that \begin{equation}\label{upper-horizontal-compactification-functoriality-local} \coprod\limits_j\big(E_{\mathcal{Z},j}(\sigma_j)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j}\big)\to E_{\mathcal{Z}'}(\tau)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'} \end{equation} is a closed embedding, it then suffices to show that any point $x'$ in the image of \ref{upper-horizontal-compactification-functoriality-local} can only come from one of term indexed by $j$ on the left-hand-side. We show this by contradiction. Suppose there are $j_1\neq j_2$ and points $y_1$ (resp. $y_2$) of $E_{\mathcal{Z},j_1}(\sigma_{j_1})\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j_1}$ (resp. $E_{\mathcal{Z},j_2}(\sigma_{j_2})\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z},j_2}$) that map to the point $x'$ of $E_{\mathcal{Z}'}(\tau)\times_{\operatorname{Spec}\mathbb{Z}}C_{\mathcal{Z}'}$ under \ref{upper-horizontal-compactification-functoriality-local}. Then $x', y_1$ and $y_2$ have the same image $z'\in C_{\mathcal{Z}'}$. On the other hand, $z'$ is also in the image of closed embeddings $C_{\mathcal{Z},j_1}\hookrightarrow C_{\mathcal{Z}'}$ and $C_{\mathcal{Z},j_2}\hookrightarrow C_{\mathcal{Z}'}$. By considering the pullbacks to $z'$ of the closed embeddings $ E_{\mathcal{Z},j_1}(\sigma_{j_1})\to E_{\mathcal{Z}'}(\tau)$ and $E_{\mathcal{Z},j_1}(\sigma_{j_1})\to E_{\mathcal{Z}'}(\tau)$, we obtain closed embeddings $\alpha_{j_1}: E_{\mathcal{Z},j_1}(\sigma_{j_1})_{z'}\to E_{\mathcal{Z}'}(\tau)_{z'}$ and $\alpha_{j_2}: E_{\mathcal{Z},j_1}(\sigma_{j_1})_{z'}\to E_{\mathcal{Z}'}(\tau)_{z'}$ which have overlapping images. One then argues as in \cite[4.4, 4.8]{Lan-immersion} to conclude that $(E_{\mathcal{Z},j_1})_{z'}\to (E_{\mathcal{Z}'})_{z'}$ and $(E_{\mathcal{Z},j_2})_{z'}\to (E_{\mathcal{Z}'})_{z'}$ would be closed embeddings with overlapping images, but then this contradicts the Hodge embedding in Theorem \ref{Main-Theorem-intro} (resp.~Theorem \ref{parahoric-analogue-intro}), or more specifically its local consequence \ref{lower-horizontal-compactification-functoriality-local}. \end{proof} \begin{remark} Note that our current proof for Corollary \ref{toroidal-cpct-corollary} only gives us the \textit{existence} of cone decompositions $\Sigma,\Sigma'$ that produce closed embeddings of the form \ref{toroidal-cpct-embedding}. Our statement does not imply that for any cone decomposition $\Sigma'$ on $\mathscr{S}_{K'}(\operatorname{GSp},S^{\pm})$, there is an induced $\Sigma$ on $\mathscr{S}_K(G,X)$ such that $(\Sigma,\Sigma')$ gives the desired closed embedding \ref{toroidal-cpct-embedding}. It is unclear to the author at the moment whether \ref{toroidal-cpct-embedding} can be constructed for any $\Sigma'$. \end{remark} \bibliographystyle{amsalpha}

1. Introduction =============== Felids are important in the epidemiology of *Toxoplasma gondii* infection because they are the only hosts that can shed environmentally resistant oocysts ([@bib7]). Although antibodies to *T. gondii* are widely prevalent in many species of felids, including the domestic cat, clinical toxoplasmosis is rare in wild felids ([@bib7], [@bib9]). However, toxoplasmosis causes severe illness in Pallas cats (*Felis manul*), and the parasite can be transmitted transplacentally from asymptomatic mothers to kittens ([@bib10], [@bib1], [@bib7]). Recently, a similar phenomenon was described for Sand cats (*Felis margarita*) from the United Arab Emirates (UAE) ([@bib13]). The Sand cat is a small-sized desert felid weighing 2--3 kg. It is found in sand and stone deserts ranging from the north of Africa to Asia, with the Arabian Peninsula as its centre of distribution. It is well adapted to living in arid areas and in areas where temperature changes are extreme with temperatures ranging from 0 to 58 °C. The Sand cat is mainly nocturnal, spending the days in a shallow burrow or under vegetation. Their prey consists of small rodents and occasional birds and reptiles. It may be able to exist without drinking free water ([@bib17]). Sand cats are held and bred in captivity at institutes worldwide. In captivity, Sand cats are fed a variety of diets from small whole chicks and rodents to commercially prepared feline diets. Some vegetables are added to the diets and water is kept available ([@bib17]). At the Breeding Centre for Endangered Arabian Wildlife (BCEAW), the Sand cats are kept in large outside enclosures with an inner den. The enclosures are only accessed by the keepers and the breeding animals are not on display to the public. The cats are kept individually and only paired for part of the year to avoid having offspring during the extremely hot summer months. We recently reported severe congenitally acquired toxoplasmosis in Sand cats at the BCEAW ([@bib13]). In the present paper we report additional cases of fatal toxoplasmosis in Sand cats in UAE and the State of Qatar, and genetically characterize *T. gondii* isolates from these cats for the first time from this host. We also document prevalence of *T. gondii* antibodies in several other species of animals from the UAE and Qatar. 2. Materials and methods ======================== 2.1. Samples from BCEAW ----------------------- ### 2.1.1. Sand cats with clinical toxoplasmosis ([Table 1](#tbl1){ref-type="table"}) #### 2.1.1.1. FM016 and FM017 On 2 March 2006 two kittens (FM016 and FM017) were born in captivity to a breeding pair that had given birth to another *T. gondii* infected litter the previous year ([@bib13]). The kittens were weighed and examined twice a week in an attempt to detect disease in the early stages. They were developing normally until day 26, when FM016 died of acute toxoplasmosis ([@bib13]). The surviving kitten, FM017, had a MAT titer of \>3200 and was medicated with clindamycin (Dalacin C™, Pharmacia, Puurs, Belgium) for 15 weeks at a dosage of 20 mg/kg given orally with a syringe. Once the kitten started eating solid food the clindamycin powder was disguised in liquid fat on the food. The animal developed no clinical signs during this period but at the age of 1 year however, it was observed to walk into objects and a severe bilateral mydriasis was seen. Direct and indirect pupil reflexes were negative and ophthalmological examination revealed a partial degeneration of the retina. Blood samples excluded infections with feline immunodeficiency virus (FIV) and feline leukemia virus (FeLV) or *Bartonella* sp. Further differentials for retinal degeneration included taurine deficiency and feline infectious peritonitis (FIP) but no further indications could support these etiologies. Over several months the animal showed an increasing muscle atrophy of the hind legs and became ataxic. A general examination under anesthesia did not reveal any obvious cause for the problems seen. Liver values AST and ALT were slightly increased as well as bile acids (AST 78.9 U/l, ALT 144.5 U/l, and bile acids 35.98 μmol/l). The animal was treated several times with vitamin B complex (B.Complex-C, Bio-Pharmachemie Co. Ltd., Vietnam) and anabolica (Ilium Boldebal-H, Troy laboratories PTY Limited, Australia) injections and de-wormed regularly. At the age of 3 years, she became incontinent. Hematuria and high alkalinity was seen and many crystals were found in the urine sediment. After 2 weeks of treatment with marbofloxacine injected and later given orally at a dosage of 2 mg/kg (Marbocyl 2% and Marbocyl 20 mg tablets, Vetoquinol, B.V.'s-Hertogenbosch, The Netherlands) there was no improvement and since the ataxia and general condition of the animal had already deteriorated for such a long time, the cat was euthanized. Samples of frozen liver and lungs were sent to the Animal Parasitic Diseases Laboratory (APDL), Beltsville, MD for further evaluation. Table 1Data on Sand cats with clinical toxoplasmosis.Source (ID)Age (years)Date diedMAT titer[a](#tbl1fn1){ref-type="table-fn"}Main lesions*T. gondii* demonstrationHist.IHCPCRBioassayStrain designationBCEAW FM002152-15-2010\>3200NoneNeg.Neg+PositiveTgSandcatUAE1 FM01734-28-2009\>3200Mild encephalitis+++Not doneTgSandcatUAE2 FM01926-15-20091600Pneumonia and hepatitis+++NegativeTgSandcatUAE3

AWWP 365771-14-2010\>3200Mesenteric lymph node necrosis+++Not doneTgSandcatQA1[^1] #### 2.1.1.2. FM019 and FM020 On 12 March 2009 two new Sand cats, a 1-year-old male FM019 and a 2-year-old female FM020, were acquired from a nearby collection where they had been born in captivity. The animals were quarantined at the BCEAW for 6 weeks during which they were observed for signs of illness and examined once under anesthesia. Hematologic and biochemical evaluations were done at that time but no abnormal blood values were seen, and viral testing for FeLV and FIV (SensPERT kits, Combined FeLV Antigen/FIV antibody test, Vet All Laboratories, Korea) were negative. They were also seronegative for *T. gondii* antibodies (see serology section) on blood samples drawn on 4 April 2009. After their quarantine period, the animals were moved to a display enclosure in the zoo. In this enclosure Gordon\'s wildcats of which several had antibodies to *T. gondii* ([@bib14]) had previously been housed. Shortly after it was noticed that the female FM020 was not eating well, but this was attributed to the rather low temperatures in the cage, which is regulated with air-conditioning. The animal received a general examination and blood was collected under sedation with medetomidine (0.05 mg/kg, Domitor, Orion Pharma, Finland). Liver enzymes ALT and AST were slightly increased (AST 110.2 U/l and ALT 144.2 U/l) as well as urea (19 mmol/l) but with a normal creatinine (60 μmol/l). All other hematology and biochemistry values tested were in normal range. No further abnormalities could be detected during examination and her weight was as expected. Radiographs of the thorax and abdomen did not reveal any abnormalities either. Apart from some subcutaneous fluid during the procedure, no other treatment was indicated and since it was thought that the low temperature might have been the reason for the inappetance, only environmental changes were made. This cat is now apparently healthy. A blood sample was obtained from cat FM020 on 15 July 2009 for *T. gondii* antibodies. Four weeks later her male cage mate FM019 looked depressed and ate less than normal. He was hand caught to perform an examination but during normal restraint and handling to take radiographs he developed a severe dyspnoea, collapsed and died on 15 June 2009. A necropsy was performed on the same day. #### 2.1.1.3. FM002 Cat FM002 was imported from the USA in 1998 at the age of 3 years. In 2006 a blood sample was taken and the *T. gondii* antibody titer was determined with a latex agglutination test at the University of Glasgow, Scotland, using a commercial kit (Eiken, Tokyo, Japan), which showed a positive titer of \>512. Testing for FIV, FeLV and FeCoV were negative at that time. In 2007 he started to show signs of ataxia in the hind legs and a slight head tilt. A full examination under sedation revealed no obvious abnormalities but hematology results revealed a severe increase in globulins (80 g/l) and proteinuria. Because of previous problems in the Sand cats with toxoplasmosis infection, the animal was started on clindamycin (Dalacin C™, Pharmacia, Puurs, Belgium) at 20 mg/kg twice daily for 32 days. Clinical signs improved slowly but the animal never completely recovered and remained slightly ataxic. Behavioural changes indicated that he also might have had impaired vision and a full ophthalmological examination was performed which revealed an inactive chorioretinitis with partial degeneration of the retina. In February 2010 the condition of this cat deteriorated quickly. He was anorexic, had diarrhoea and had severe difficulties walking. A full examination revealed a severely extended bladder without obstruction and general muscle atrophy. A urine examination revealed proteinuria, but no bacterial growth. Blood samples revealed an increased urea (14.7 mmol/l) with a normal creatinine (95 μmol/l). X-rays showed some consolidated areas in the frontal lung lobes, round shaped kidneys and a severe scoliosis of the thoracic vertebrae with bone demineralization. Because of the poor condition of the animal, and its old age, the cat was euthanized. ### 2.1.2. Serum samples from captive animals Serum samples were obtained from Sand cats and other animals at BCEAW ([Table 2](#tbl2){ref-type="table"} ).Table 2Antibodies to *T. gondii* in sera of animals from BCEAW, United Arab Emirates.SpeciesNo. testedNo. positiveMAT titer (no. of animals)Felids Gordon' wild cat (*Felis sylvestris gordoni*)55100, \>3200 (4) Arabian leopard (*Panthera pardus nimr)*7650 (2), 100, 200, 800 (2) Cheetah (*Acinonyx jubatus soemmerringii*)343150, 100, 400, 800 (5), 1600 (4), \>3200 (15) Caracal (*Caracal caracal schmitsi*)65100 (2), 200 (2), 3200 African caracal (*Caracal caracal algira*)11200 Feral domestic cat (*Felis catus*)41\>1600

Foxes Sand fox (*Vulpes rueppellii*)33800, \>3200 (2) Red fox (*Vulpes vulpes arabica*)[a](#tbl2fn1){ref-type="table-fn"}15950, 100 (3), 200 (3), 800, 3200 Blanford\'s fox (*Vulpes cana*)[b](#tbl2fn2){ref-type="table-fn"}148All \>3200

Others Brandt\'s hedgehog (*Paraechinus hypomelas*)213200 Long eared hedgehog (*Hemiechinus auritus*)63\>3200 Ethiopian hedgehog (*Paraechinus aethiopicus*)21400 Wolf (*Canis lupus arabs*)[c](#tbl2fn3){ref-type="table-fn"}8550 (3), 200, 1600 Jackal (*Canis aureus*)8625, 50 (2), 100, 100, 200

Striped hyaena (*Hyaena hyaena*)6325, 50, 100Small spotted genet (*Genetta genetta*)11800White tailed mongoose (*Ichneumia albicauda*)22800 (2)Indian grey mongoose (*Herpestes edwardsii*)113200Hamadryas baboon (*Papio hamadryas*)11400[^2][^3][^4] 2.2. Samples from Al Wabra Wildlife Preservation, Doha, Qatar (AWWP) -------------------------------------------------------------------- AWWP is home to 27 Sand cats and other felids. The management of this facility ([@bib2]) is similar to that of BCEAW. ### 2.2.1. Clinical toxoplasmosis in a Sand cat (3657) A 7-year-old male Sand cat died of acute illness. This cat was born at AWWP, and had been vaccinated yearly with feline panleukopaenia virus, feline rhinotracheitis virus, feline calcivirus (Felocell CVR, Pfizer GmbH, Berlin, Germany), and rabies virus (Rabdomun, Essex Tierarznei, Muenchen, Germany) ([@bib2]). On 11 January 2010 this cat was examined because of anorexia and weakness. Its mucous membranes were pale and it showed labored breathing. A blood sample was obtained for biochemical and hematological tests. The cat died after 3 days of treatment with subcutaneous injections of 16.8/4.2 mg/kg amoxicillin--clavulanacid (Synulox RTU 140/30 mg/ml, Pfizer GmbH, Berlin, Germany) and 0.6 mg/kg meloxicam (Metacam 5 mg/ml, Boehringer Ingelheim, Ingelheim, Germany) and in addition 150 ml subcutaneous fluids (Electrolytloesung, Albrecht GmbH, Aulendorf, Germany). Initially formalin-fixed tissues were submitted to a commercial laboratory (Institut für Tierpathologie, Ludwig-Maximilians-Universitaet Muenchen, 80539 Muenchen, Veterinaerstr. 13, Germany) for histological examination. After the diagnosis of toxoplasmosis, formalin-fixed tissues and frozen liver and lung of the Sand cat 3657 were sent to the Animal Parasitic Diseases Laboratory (APDL), Beltsville, MD for further evaluation. ### 2.2.2. Serum samples from captive animals Serum samples were obtained from Sand cats and other animals at AWWP ([Table 3](#tbl3){ref-type="table"} ).Table 3Antibodies to *T. gondii* in sera of animals from AWWP, Qatar.SpeciesNo. testedNo. positiveMAT titer (no. of animals)Arabian Sand cat (*Felis margarita harrisoni*)2014200 (5), 400 (2), 800 (1), \>3200 (6)African wild cat (*Felis sylvestris gordoni*)111600Cheetah (*Acinonyx jubatus soemmerringii*)5550 (1), 200 (1),400 (1), 800 (1), 3200 (2)King cheetah (*Acinonyx jubatus rex*)11\>3200 2.3. Necropsy and histological examination ------------------------------------------ Samples of all major organs were fixed in 10% buffered formalin. Paraffin-embedded tissues were sectioned at 5 μm and examined after staining with hematoxylin and eosin (H and E). 2.4. Serological examination for *T. gondii* -------------------------------------------- Antibodies to *T. gondii* were determined using the modified agglutination test (MAT) performed at APDL as described ([@bib4]). A cut-off titer of 1:25 is considered indicative of *T. gondii* infection. Serum samples from animals at BCEAW examined for *T. gondii* antibodies are listed in [Table 2](#tbl2){ref-type="table"}. Unless indicated, the sera were obtained in 2009. Additionally, sera from 11 rodents (8 mice and 3 rats) were tested at 1:25 serum dilution. Serum samples from animals at AWWP examined for antibodies to *T. gondii* are listed in [Table 3](#tbl3){ref-type="table"}. 2.5. Attempted isolation of *T. gondii* --------------------------------------- Unfixed liver and lungs of FM019 and FM002 were shipped by air from UAE to the APDL for *T*. *gondii* isolation. Seven and four days elapsed between collection of samples from cat FM019 and FM002 and their receipt at APDL, respectively. Portions of lungs and liver from cat FM019 were homogenized in aqueous 0. 85% NaCl (saline), and inoculated subcutaneously in to 8 Swiss Webster (SW) mice. The inoculated mice were observed for 81 days and examined for *T. gondii* infection ([@bib7]). Tissues of cat FM002 were bioassayed in SW and gamma interferon gene knock out (KO) mice. Tissues were homogenized in saline, and the homogenate was divided into two equal portions (A and B). Portion A was digested in acid pepsin for 60 min at 37 °C, centrifuged, neutralized with sodium bicarbonate, and after adding antibiotics, inoculated subcutaneously in to five SW and two KO mice as described ([@bib7]). Portion B was not digested, centrifuged, and the sediment suspended in antibiotic saline containing 1000 units of penicillin and 100 μg of streptomycin per ml of saline ([@bib7]). The mice inoculated with cat tissues were examined for evidence of *T. gondii* infection. 2.6. Genetic characterization of *T. gondii* -------------------------------------------- *T. gondii* DNA was extracted from feline tissues or infected cell culture using DNeasy kit (Qiagen) and genotyped using the genetic markers SAG1, SAG2, SAG3, BTUB, GRA6, c22-8, c29-2, L358, PK1, and Apico ([@bib19], [@bib20]) ([Table 4](#tbl4){ref-type="table"} ).Table 4Genetic characterization of *T. gondii* from Sand cats from UAE and Qatar.Strain IDGenotypesGenetic markersSAG1(5′ + 3′) SAG2alt.SAG2SAG3BTUBGRA6c22-8c29-2L358PK1ApicoGTIIIIIIIIIIIIPTGIIIIIIIIIIIIIIIIIIIIIIIICTGIIIII or IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIMASAtypicalu-1IIIIIIIIIIIIu-1IIIIIITgCgCa1AtypicalIIIIIIIIIIIIIIu-1Iu-2ITgCtBr5AtypicalIIIIIIIIIIIIIIIIIIIu-1ITgCtBr64AtypicalIIu-1IIIIIIIIIu-1IIIIIIIITgRsCr1Atypicalu-1IIIIIIIIIIu-2IIIIIINegative control----------------------TgSandcatUAE3 (FM019)Atypicalu-1IIIIIIIIIIIIIIIIIIIu-2ITgSandcatUAE1 (FM002)Atypicalu-1IIIIIIIIIIIIIIIIIIIu-2ITgSandcatUAE2 (FM017)AtypicalII or IIIINDNDIIIINDIIIIINDITgSandcatQA1 (3657)IIII or IIIIIIIIIIIIIIIIIIIIIII[^5] 2.7. Immunohistochemical (IHC) examination ------------------------------------------ Immunohistochemistry was performed on paraffin-embedded sections at APDL using reagents and methods described previously ([@bib11]). 3. Results ========== 3.1. Samples from BCEAW ----------------------- ### 3.1.1. Sand cats with clinical toxoplasmosis #### 3.1.1.1. FM017 The lungs were edematous, and had multiple small nodules. The liver was congested. Microscopically, numerous tissue cysts were found in sections of brain ([Fig. 1](#fig1){ref-type="fig"} ). There was a single focus of perivascular infiltration of mononuclear cells and a focus of gliosis. Large sized *T. gondii* tissue cysts were seen in sections of heart, skeletal muscle, and ocular muscle but lesions were not seen ([Fig. 2](#fig1){ref-type="fig"}). A single focus of vasculitis and mononuclear cell infiltration was seen in the retina. Lesions or *T. gondii* were not found in sections of kidneys, adrenal, lung, liver, stomach, small intestine, and urinary bladder.Figs. 1--4Lesions and *T. gondii* in tissues of Sand cats. (1) Intracellular tissue cyst (arrow) in section of the brain of cat FM017. H and E stain. (2) Tissue cyst in section of the myocardium of cat FM017. H and E stain. (3) Tachyzoites in impression smear of lung of cat FM019. Giemsa stain. Note one tachyzoite (a) with a nucleus, two tachyzoites with nucleus dividing into two (b), and an intracellular tachyzoite (c). (4) Tachyzoites (arrows, all red areas) causing necrosis of hepatocytes in the liver of cat FM019. Section stained with *T. gondii* polyclonal rabbit antibody. Scale bar = 10 μm. DNA from tissues of this cat was typed at several of the 10 genetic loci and partial data was obtained. The data indicates this strain maybe of an atypical genotype ([Table 4](#tbl4){ref-type="table"}). #### 3.1.1.2. FM019 Grossly, the lungs had edema, focal consolidation, and multiple small white nodules were observed on the lung surface. There was serohemorragic fluid in the thorax and the liver was jaundiced. Tachyzoites were seen in impression smears of the lungs ([Fig. 3](#fig1){ref-type="fig"}). Bacteria were not isolated from culture of lung, liver and thorax fluid on standard culturing. Histologically, this cat had severe multifocal hepatitis, and interstitial pneumonitis associated with numerous tachyzoites ([Fig. 4](#fig1){ref-type="fig"}). Focal myocarditis and nephritis were associated with few tachyzoites. Tissue cysts were not seen. The diagnosis was confirmed by IHC. Antibodies to *T. gondii* were not detected in 1:25 serum dilution of blood sample collected on 9 April 2009 but blood sample collected directly from the heart at postmortem had a MAT titer of ≥1600. The mice inoculated with tissues of cat FM019 remained seronegative and tissue cysts were not found in their brains when killed 81 days post-inoculation. DNA from tissues of this cat was typed at the 10 genetic loci and was found to be an atypical strain ([Table 4](#tbl4){ref-type="table"}). This genotype was previously reported in *T. gondii* isolates of dogs from Sri Lanka ([@bib5]). #### 3.1.1.3. FM020 The cause of illness in cat FM020 was not determined. The cat seroconverted at the same time as FM019; the MAT titer on 9 April was \<1:25 but it had a high MAT titer (1:3200) on 15 July 2009. #### 3.1.1.4. FM002 On postmortem several large cysts were observed in both kidneys and multiple pinpoint white nodules were seen. Roundworms were found in the duodenum. Immunohistochemically, lesions and protozoa were not demonstrable in histological sections ([Table 1](#tbl1){ref-type="table"}). One of the two KO mice inoculated with tissues of FM002 died of acute toxoplasmosis on day 18 post-inoculation (p.i.) and numerous tachyzoites were found in smears of its lungs. *T. gondii* was successfully cultivated in CV1 cells inoculated with homogenate of lung of the infected KO mouse. This isolate was designated TgSandcatUAE1. Neither antibodies to *T. gondii* nor tissue cysts were found in the other KO mouse killed 47 days p.i. The SW mice inoculated with pepsin digest of tissues of FM002 remained asymptomatic. Antibodies to *T. gondii* (MAT \> 1:200) were found in three of the five mice but tissue cysts were not found in their brains 47 days p.i.; however, all four KO mice inoculated with pooled brains from seropositive SW mice died of acute toxoplasmosis on days 14--19. All SW mice inoculated with undigested tissues of the Sand cat died of bacterial infection 3 days p.i. and were discarded. DNA from tissues of this cat was typed at the 10 genetic loci and was found to be an atypical strain, identical to FM019 ([Table 4](#tbl4){ref-type="table"}). ### 3.1.2. Serological prevalence *T. gondii* antibodies were found in several species of animals tested, notably in 49 of 57 felids ([Table 1](#tbl1){ref-type="table"}). None of the 11 rodent (8 mice and 3 rats) sera from the breeding colony kept for feeding carnivores at the BCEAW had antibodies to *T. gondii*. 3.2. Samples from AWWP ---------------------- ### 3.2.1. Clinical toxoplasmosis in Sand cat 3657 Blood collected on the first day of illness indicated highly increased liver values AST (1000 U/l) and total bilirubin (44 μmol/l). Postmortem gross lesions were seen in the lungs which were edematous with multiple small white nodules in all lung lobes. Similar small white nodules were found in the pancreas and omentum. The mesenteric lymph nodes were enlarged and a tapeworm was found in the small intestine. Histologically, there were multifocal area of necrosis and inflammation in adrenal glands, subacute interstitial pancreatitis, interstitial pneumonia, abdominal adipose tissue necrosis, hepatitis, and mild myocarditis. There was gastritis characterized by multifocal necrosis of glands, mixed cellular infiltration of the lamina propria, cellular debris within gland lumina. Many tachyzoites were seen in the lamina propria and in glandular epithelial cells ([Fig. 5](#fig2){ref-type="fig"} ). Two loops of the colon showed mild enteritis associated with tachyzoites. Sections of small intestine and brain were not fixed for histology. There was almost total necrosis of mesenteric lymph nodes with enormous numbers of tachyzoites. *T. gondii* tissue cysts were not identified in any organ. All specimens were negative for feline coronavirus clone FIPV3-70 tested at the laboratory in Germany.Fig. 5Gastritis in Sand cat 3657 from AWWP. Immunohistochemical staining with *T. gondii* rabbit polyclonal antibodies. (A) Note foci of necrosis (arrows) and tachyzoites (all red bodies) in gastric glands and the lamina propria. Scale bar = 100 μm. (B) Several groups (arrows) and individual (arrowheads) tachyzoites in epithelial cells of a gastric gland. Scale bar = 10 μm. DNA from this cat was typed at the 10 genetic loci and was found to be a Type II strain with a Type I allele at locus Apico ([Table 4](#tbl4){ref-type="table"}). This Type II strain is highly prevalent worldwide. ### 3.2.2. Serological prevalence *T. gondii* antibodies were found in 14 of 20 Sand cats, and 7 of 7 other wild felids ([Table 3](#tbl3){ref-type="table"}). 4. Discussion ============= In the present study, two Sand cats (FM019 and 3657) died of acute toxoplasmosis at separate establishments. Cat FM019, born in captivity at BCEAW, and appeared clinically normal until 18 months of age. Its MAT titer was 1:1600 at the time of death, but was seronegative to *T. gondii* 2 months before its demise. Sand cat 3657 was also born in captivity at AWWP. The diagnosis of toxoplasmosis was confirmed immunohistochemically, by serology and by PCR. The acute nature of lesions indicated that toxoplasmosis was acquired recently, and only tachyzoites (not tissue cysts) were seen in lesions. In Sand cat 3657, lesions in the mesenteric lymph nodes and stomach are noteworthy. The finding of gastritis was unusual as this is rarely seen in domestic cats (*Felis domesticus*) naturally infected with *T. gondii* ([@bib7]) and has not been reported in wild felids. Why some animals develop clinical toxoplasmosis whereas most remain subclinical is unknown. Recently, attention has been focused on the genetic makeup of the strains in pathogenesis of clinical toxoplasmosis and atypical strains were found to cause clinical toxoplasmosis in immunocompetent humans ([@bib3], [@bib8]; see [@bib7]). Currently there is no information on the genetic characterization of *T. gondii* isolates from the Middle East or the Sand cat. In the present study, the strain of *T. gondii* from UAE was found to be atypical. This genotype has been reported in dogs from Sri Lanka ([@bib5]). The isolate from Qatar is a common Type II strain that is frequently identified worldwide. The cause of ataxia and retinal degeneration in FM017 was not determined. The presence of numerous tissue cysts in the brain and few lesions might have contributed to illness. It is known that clindamycin acts on tachyzoites but has no effect on tissue cysts ([@bib7]). Prevention of *T. gondii* infection in zoo animals is a major problem, especially in highly susceptible species, such as Pallas and Sand cats, New World primates, and marsupials. Cats, like other hosts of *T. gondii* can become infected transplacentally, by ingesting food and water contaminated with oocysts, and by the ingestion of infected tissues ([@bib7]). In domestic cats the ingestion of infected tissues is more efficient because even a few bradyzoites are infectious for cats whereas few oocysts are not, and transplacental transmission is considered infrequent ([@bib7]). However, Pallas and Sand cats are different because unlike humans, domestic cats and sheep, chronically infected asymptomatic dams transmit infection to their offspring ([@bib7]). The diet of all small carnivores at the BCEAW previously consisted of imported frozen buffalo and chicken meat, locally bought fresh camel and donkey meat, frozen day old chicks, and white mice and rats, pigeons, guinea fowl and quail that have been bred at the BCEAW. After the neonatal deaths in Sand cats in 2006 were confirmed as being caused by toxoplasmosis ([@bib13]), the Sand cat diet was adapted to include only meat that has been frozen at −20 °C for at least 1 week or fresh mice and rats bred at the BCEAW. Limited sampling of rats and mice sera fed to cats at BCEAW did not have any evidence of *T. gondii* infection. Although a serious effort is made to keep the enclosures pest free, wild rodents, birds and reptiles can still gain access and feral cats are sometimes found roaming within the BCEAW. The diet of the Sand cats in Qatar consists of white mice and rats bred at AWWP. In addition they are fed with wild pigeons which have been caught and slaughtered on the premises and then frozen for different periods (ranging from 1 day to 1 month). Pest control programs are limiting the number of wild mice and rats in the area of the enclosures, but are not eradicating all. Although no method can be regarded as 100% effective, most *T. gondii* tissue cysts will be destroyed when meat is frozen and thawed as well as by heating up to 67 °C or processing the meat (see [@bib7]). Heating or processing the meat is not considered as an option at the BCEAW for nutritional and behavioural enrichment reasons. Results of the present study indicate that several other species of animals at BCEAW were exposed to *T. gondii*. However, clinical toxoplasmosis at the BCEAW has been documented only in Sand cats ([@bib13], and present study), in a Blanford\'s fox ([@bib6]) and in a Sandfox ([@bib15]) and now in Sand cat at AWWP. In the present study, *T. gondii* antibodies were found in several species of carnivores in addition to Sand cats, Sand foxes and Blanford\'s foxes previously reported ([@bib6], [@bib15]). We are not aware of previous reports of serological prevalence in the long eared hedgehog (*Hemiechinus auritus*), white tailed mongoose (*Ichneumia albicauda*), Indian grey mongoose (*Herpestes edwardsii*), Hamadryas baboon (*Papio hamadryas*) and the Arabian wolf (*Canis lupus arabs*). [@bib18] reported *T. gondii* antibodies in 59.1% of 22 Egyptian mongoose (*Herpestis ichneumon*) from Spain. [@bib14] found MAT antibodies in 86.1% of 36 Gordon\'s wild cat from the BCEAW. In the present study, *T. gondii* antibodies were found in 5 of 5 Gordon\'s wild cat. Whether Sand cat, Gordon\'s wild cat, caracal, or leopard can excrete *T. gondii* oocysts is unknown ([@bib7]), whereas it is known that cheetahs can excrete *T. gondii* oocysts ([@bib12], [@bib16]). There are several Gordon\'s wild cats resident at the BCEAW. They were housed in the same pen 2 months before the Sand cats FM019 and FM020 were introduced in the enclosure. Oocysts excreted by these cats might have been present in the pen where the Sand cats were housed and could be the source of infection for Sand cats at BCEAW. It is of interest to see that the *T. gondii* strain identified in Sand cats from UAE has an atypical genotype that is different from the clonal Type I, II and III strains. This genotype was previously reported in *T. gondii* isolates of dogs from Sri Lanka ([@bib5]), indicating it is widely spread in Asia and the Arabian Peninsula. In addition, the strain identified from Sand cat from Qatar belongs to the Type II lineage that is also widely spread at the global scale. The high seropositivity of *T. gondii* in captive wild animals tested in this study, in addition with the lack of epidemiological studies on this parasite in the Persian Gulf region, makes it necessary to carry out further studies to investigate the diversity and transmission of *T. gondii* in animal and populations in the region. [^1]: Near death or postmortem. [^2]: Sampled in 2008, three of seropositive foxes were wild caught. [^3]: Three sampled in 2008. [^4]: One sampled in 2008. [^5]: ND, no data available.

Group A rotaviruses (RVAs) are a frequent cause of diarrhea in children. The RVA genome consists of 11 dsRNA segments that encode 6 structural (VP1--VP4, VP6, VP7) and 6 non-structural (NSP1-- NSP6) proteins ([@R1]). The basis of a new classification system is phylogenetic analysis of 11 RVA genome segments, although binary classification is still used; genotyping is based on the coding genes for VP7 (G) and VP4 (P) ([@R2]). Vaccination is considered effective in reducing the consequences of RVA. Two vaccines, Rotarix (Glaxo SmithKline, Brentford, UK) and RotaTeq (Merck & Co., Whitehouse Station, NJ, USA), are licensed in several countries. Both vaccines demonstrated broad protection against the most common RVA genotypes ([@R3]). In Brazil, Rotarix, a monovalent attenuated human rotavirus vaccine for infants 6--24 weeks of age, was introduced in the National Immunization Programs in March, 2006. The vaccine is delivered in 2 doses, [\>]{.ul}4 weeks apart. In 2009, when vaccine coverage achieved 85.9%, reduction in hospitalization associated with diarrhea (17%) and in related mortality rates (22%) was observed ([@R4]). This study aimed to assess phylogenetic relationships between G1P\[8\] RVA obtained from vaccinated and unvaccinated children hospitalized for acute gastroenteritis. The Study ========= During 2008--2010, fecal samples were collected from 3,852 children; 702 specimens (18.2%) had RVA-positive ELISA results; 27 of those (3.8%) were characterized as G1P\[8\] by using reverse transcription-PCR. Eighteen of the 27 specimens were from vaccinated children ([Table 1](#T1){ref-type="table"}). Study methods were approved by Fiocruz Ethical Committee (No. 311/06). ###### Clinical features of 27 patients infected with group A rotaviruses (G1P\[8\]), Brazil 2008--2010\* Strain State Age (mo)† Rotarix doses (mo/yr) Time elapsed between vaccination and hospitalization Clinical symptoms -------------- ------- ----------- ----------------------- ------------------------------------------------------ ------------------- ---- ----- ----- ----- ES15221--08 ES 42 -- -- -- -- -- -- -- SE15901--08 SE 2 09/2008 -- 7 d 3 No No Yes SE 16536--09 SE 2 -- -- -- 2 No Yes Yes SE 16537--09 SE 4 03/2009 -- 2 mo 6 No Yes Yes SE 16779--09 SE 41 03/2006 05/2006 37 mo 1 Yes Yes Yes SE 16782--09 SE 19 02/2008 04/2008 15 mo 3 No No y SE 16800--09 SE 7 02/2009 -- 5 mo 3 -- No No SE 16803--09 SE 10 11/2008 01/2009 6 mo 1 No Yes Yes SE 16894--09 SE 20 02/2008 05/2008 15 mo 2 No Yes No SE 16897--09 SE 15 08/2008 10/2008 10 mo 1 No No Yes SE 16898--09 SE 4 06/2009 -- 2 mo 1 No No Yes SE 16977--09 SE 9 02/2009 04/2009 4 mo 13 No Yes Yes SE 16978--09 SE 10 01/2009 03/2009 6 mo 2 No Yes No SE 17120--09 SE 22 02/2008 05/2008 17 mo 1 No No Yes SE 17122--09 SE 9 07/2008 09/2008 5 mo 2 -- Yes No SE 17123--09 SE 8 -- -- -- 3 No Yes Yes SE 17241--09 SE 7 06/2008 08/2008 3 mo 2 No Yes Yes PE17887--09 PE 6 -- -- -- -- No Yes Yes PE 17888--09 PE 10 11/2008 -- 8 mo -- No Yes Yes PE 17890--09 PE 20 02/2008 05/2008 18 mo -- No No Yes PE 17891--09 PE 12 -- -- -- -- No Yes Yes MA18999--10 MA 18 -- -- -- -- -- -- -- MA 19006--10 MA 2 08/10 -- 4 d 1 Yes No No MA 19013--10 MA 10 -- -- -- -- -- -- -- MA 19015--10 MA 48 -- -- -- -- -- -- -- MA 19030--10 MA 32 03/2008 -- 29 mo 7 Yes No No BA19391--10 BA 2 -- -- -- 2 -- No No \*Prefixes represent origins of circulating strains in Brazil: ES, Espirito Santo; --, no data; SE, Sergipe; PE, Pernambuco; MA, Maranhão; BA, Bahia.
†Age of child at time of fecal sample collection.
‡Number of days child had diarrhea before fecal sample collection. RVA detection, genotyping, and sequencing were performed ([@R5]). Sequences were deposited in GenBank under accession numbers: JQ926436-JQ926600 and JX683535-JX683664. Sequences were compared with those in strains obtained from GenBank (including Rotarix JX943604.2--JX943614.2). Strains analyzed belonged to the Wa-like genogroup (genotype 1); 26 strains showed a G1--P\[8\]--I1--R1--C1--M1--A1--N1--T1--E1--H1 genome constellation. One sample, collected from a child vaccinated with 1 dose during 2010 in Maranhão (MA) state (MA19030--10), contained the G1--P\[8\]--Ix--R1--Cx--M1--A1--N1--T3--E1--H1 constellation. Nucleotide (nt) identity values between circulating strains in Brazil and the Rotarix strain ranged 76%--100% ([Table 2](#T2){ref-type="table"}). Three samples showed 100% nt identity with the Rotarix strain in [\>]{.ul}1 gene. The SE15901--08 strain, collected on day 7 after the first dose, showed 100% nt identity with all Rotarix strain gene segments and could represent vaccine shedding. Vaccine antigen excretion detected by ELISA achieved 80%, declining to 18%--24% when collected at day 30; 11%--16% of children shed the virus at day 45 ([@R6]). Shedding is not associated with increased gastroenteritis-like symptoms. Widespread use of Rotarix might reveal adverse reactions not observed in clinical trials, emphasizing the need for global surveillance ([@R7]). ###### Nucleotide similarity values when comparing the Brazilian G1P\[8\] samples with the Rotarix strain gene segments and genotype constellation of selected human group A rotaviruses\* Protein encoded by genes† ------------- ----------------------- --------------------------- ------------- --------- --------- --------- --------- --------- --------- --------- ---------- --------- SE15901--09 Vaccine G1
100% P\[8\]
100% I1
100% R1
100% C1
100% M1
100% A1
100% N1
100% T1
100% E1
100% H1
100% MA19006--10 Vaccine G1
100% P\[8\]
99% I1
99% R1
99% C1
98% M1
99% A1
98% N1
100% T1
100% E1
99% H1
98% BA19391--10 Vaccine G1
100% P\[8\]
99% I1
100% R1
99% C1
98% M1
100% A1
100% N1
100% T1
100% E1
99,6% H1
100% ES15221--08 Wa-like (ES) G1
94% P\[8\]
90% I1
90% R1
100% C1
90% M1
93% A1
92% N1
91% T1
97% E1
92% H1
92% SE16800--09 Wa-like (SE) G1
94% P\[8\]
92% I1
89% R1
97% C1
90% M1
96% A1
92% N1
88% T1
98% E1
98% H1
92% PE17888--09 Wa-like (PE) G1
94% P\[8\]
92% I1
90% R1
94% C1
91% M1
93% A1
93% N1
90% T1
97% E1
94% H1
93% MA19013--10 Wa-like (MA) G1
95% P\[8\]
93% I1
92% R1
95% C1
92% M1
95% A1
94% N1
92% T1
97% E1
94% H1
93% MA19030--10 Wa/AU-1
reassortant G1
93% P\[8\]
91% X R1
93% X M1
94% A1
93% N1
89% T3
76% E1
92% H1
93% Wa Wa-like G1 P\[8\] I1 R1 C1 M1 A1 N1 T1 E1 H1 DS-1 DS-1-like G2 P\[4\] I2 R2 C2 M2 A2 N2 T2 E2 H2 AK26 Wa/DS-1 reassortant G2 P\[4\] I2 R2 C2 M2 A2 N1 T2 E2 H2 6809 Wa/DS-1 reassortant G8 P\[6\] I2 R1 C1 M1 A1 N1 T1 E1 H1 Matlab13 Wa/DS-1 reassortant G12 P\[6 I1 R1 C1 M1 A1 T2 T1 E1 H1 Mani-253 Wa/DS-1 reassortant G4 P\[4\] I1 R1 C1 M2 A8 N1 T1 E1 H1 AU-1 AU-1-like G3 P\[9\] I3 R3 C3 M3 A3 N3 T3 E3 H3 T152 AU-1-like G12 P\[9\] I3 R3 C3 M3 A12 N3 T3 E3 H6 K8 Wa/AU-1 reassortant G1 P\[9\] I1 R3 C3 M3 A1 N1 T3 E3 H3 Mani-265 DS-1/AU-1 reassortant G10 P\[6\] I2 R2 C2 M2 A3 N2 T7 E2 H2 \*Wa-like, DS-1-like, AU-like genotypes are shown in green, red, and orange, respectively, and the P\[6\] genotype is shown in blue; X, indeterminate genotypes; --, no data.
†One sample for each group was chosen as representative and the average was shown.
‡Prefixes represent origins of circulating strains in Brazil: SE, Sergipe; BA, Bahia; MA, Maranhão; ES, Espirito Santo; PE, Pernambuco. In phylogenetic analysis of 11 genes, 10 genes clustered into 5 clades. In the remaining gene, NSP5, the circulating strains clustered into 4 groups ([Figure 1](#F1){ref-type="fig"}, panel E). ![Phylogenetic analysis of nucleotide sequence of A) nonstructural protein 1 \[NSP1\], B) NSP2, C), NSP3, D) NSP4, and E) NSP5 encoding genes of G1P\[8\] RVA samples collected in Brazil and described in this study. Filled circles indicate samples from non-vaccinated children and filled triangles vaccinated children. The Rotarix strain is indicated by a filled diamond. Bootstrap values \>70%, estimated with 2,000 pseudoreplicate datasets, are indicated at each node. Only the genotype in which the strains under investigation cluster is shown completely; 1 prototype from another genotype has been added as an outgroup. Scale bars represent 0.5 substitutions per nucleotide site.](12-1407-F1){#F1} Strain ES15221--08, detected in an unvaccinated child, is genetically distinct and differs in origin from other G1P\[8\] circulating strains ([Figures 1](#F1){ref-type="fig"},[2](#F2){ref-type="fig"}). The VP1 gene (648 bp, [Figure 2](#F2){ref-type="fig"}, panel A) in this sample showed 100% nt identity with the Rotarix strain. Strain MA19006--10 was genetically similar to the strain; however, the NSP5 segment ([Figure 1](#F1){ref-type="fig"}, panel E) was closely related to a G1P\[8\] strain from Australia (JF490152). These 2 strains appear to have been generated by reassortment with this vaccine strain. ![Phylogenetic analysis of nucleotide sequence of (A) VP1, (B) VP2, (C), VP3, (D) VP4, (E) VP6 and (F) VP7 genes of G1P\[8\] RVA samples collected in Brazil and described in this study. Filled circles indicate samples from non-vaccinated children and filled triangles vaccinated children. The Rotarix vaccine strain is indicated by a filled diamond. Bootstrap values above 70%, estimated with 2,000 pseudoreplicate datasets, are indicated at each node. Only the genotype in which the strains under investigation cluster is shown completely and one prototype from another genotype has been added as outgroup. Scale bars represent 0.5 substitutions per nucleotide site~.~](12-1407-F2){#F2} Strain MA19030--10, detected in a child after 1 vaccine dose, was closely related to the MA group, but the NSP2 segment differed in origin from other MA samples ([Figure 1](#F1){ref-type="fig"}, panel B) because it clustered with other Wa-like strains. The NSP3 gene belonged to genotype 3 because it was 99.2% (nt) similar to the AU-1 prototype strain (DQ490535.1). These results suggested reassortment events between Wa-like and AU-1 like co--circulating strains. NSP2 and NSP3 genes from sample MA19030--10 and NSP5 genes from sample MA19006--10 differed from each other and from those of their respective clusters. Samples MA19006--10 and MA19030--10 are genetically distinct and might have distinct evolutionary histories. Studies that included this genogrouping system were performed to prove the existence of inter-genogroup reassortment between human RVA genogroups or human and animal genogroups. The existence and effectiveness of heterogeneous genome constellations remains unclear, probably because it is caused by mechanisms that create protein sets that work better when kept together ([@R8]). Phylogenetic analysis of the VP8\* (aa 1--247) portion of VP4 encoding gene showed circulating strains ([Figure 2](#F2){ref-type="fig"}, panel D) clustered into 2 lineages (P\[8\]-3 and P\[8\]-1). The alignment of the deduced aa sequences showed potential trypsin cleavage sites at arginine 240 and 246 conserved in all samples. All circulating strains contained 91.3%--100% identical aa residues to the Rotarix strain in VP8\* antigenic epitopes; no changes were observed on epitopes 8--2 and 8--4. Two VP7 lineages (G1-I and G1-II) were identified. Samples closely related to the Rotarix strain belonged to G1-II; remaining strains belonged to G1-I. Comparing the regions defined as antigenic epitopes (7--1 and 7--2) for VP7 protein, we found [\>]{.ul}2 epitopes (aa 94, 123, 148, 217) were not conserved among circulating strains compared with the Rotarix strain. The following amino acids were conserved in the non-VP7/non-VP4 segments of either the Rotarix strain or other strains analyzed: cysteine residues involved in disulfide bonds at positions 6, 8, 85, and 285 in NSP2 protein; the leucine, isoleucine, aspartic acid, methionine and glutamine at positions 275, 281, 288, and 295 in the NSP3 protein; the N-linked glycosylation sites at positions 8 and 18, and cysteine residues at positions 63 and 71 in NSP4 protein; and serine residues at positions 153, 155, 163, and 165 in NSP5 protein. Conclusions =========== This report characterizes the complete genome of G1P\[8\] strains in Brazil. Phylogenetic analysis showed that sequences clustered consistently with the region of sample collection. Three strains circulating in Brazil were closely related to those in Rotarix; 1 of the 3 was 100% identical to Rotarix, likely representing shedding of this vaccine. Sequence analysis confirmed the presence of the Rotarix VP1--derived segment in 1 sample (ES15221--08), indicating an unreported reassortment event between the vaccine and a community strain. The backbone of MA19030--10 sample was Wa-like, but the NSP3 segment exhibited a T3 genotype that was described for the AU-1 genogroup. This finding suggests that this strain derived its NSP3 gene from an AU-1-like strain through reassortment. Changes in antigenic regions of VP4 and VP7 proteins have been associated with mutated RVA strains that spread ([@R9],[@R10]). Comparison of the aa sequences of VP7 and VP8\* of strains circulating in Brazil and the monovalent vaccine strain demonstrated that VP7 and VP8\* of the circulating strains showed similar antigenic regions to those of the vaccine. No differences between strains from vaccinated and unvaccinated children were observed. Considering the segmented RVA genome and that the Rotarix vaccine is an attenuated RVA human strain, it is expected that reassortants will arise and circulate among humans. The effects of such events are not known. This study described strains that originated from reassortment events between the Rotarix vaccine strain and strains detected in vaccinated and unvaccinated children. Improvement of RVA surveillance programs that include full genome sequencing analysis will strengthen the understanding of how vaccines will affect the RVAs circulating among humans, and how those events could affect the use of live vaccines, the frequency of RVA intra- and intergenogroup reassortment events under natural conditions, and the stability of RVA generated by such events. *Suggested citation for this article*: (Rose TL, Silva MFM, Gomez MM, Resque HR, Ichihara MYT, Volotão EM, et al. Evidence of Vaccine-related Reassortment of Rotavirus, Brazil, 2008--2010. Emerg Infect Dis. 2013 Nov \[*date cited*\]. <http://dx.doi.org/10.3201/eid1911.121407> These authors contributed equally to this article. The authors thank the PDTIS DNA Sequence Platform (RPT01A) and LATER/BioManguinhos, FIOCRUZ, Rio de Janeiro, Brazil; and the Secretary of Public Health State of Sergipe, Maranhão, Espírito Santo, Bahia and Pernambuco. We also thank Alexandre Madi Fialho and Rosane Maria Santos de Assis for technical assistance. This research was supported by federal funds from the Brazilian Federal Agency for Support and Evaluation of Graduate Education (CAPES), the National Council for Scientific and Technological Development (CNPq), project PAPES VI/Fiocruz - CNPq, Oswaldo Cruz Institute (PROEP -- Fiocruz - CNPq), the General Coordination of Public Health Laboratories -- Secretary of Health Surveillance (CGLAB/SVS), project CAPES-MERCOSUL PPCP 023/2011, and Carlos Chagas Filho Foundation for Research Support of Rio de Janeiro State (FAPERJ). T.L.R. receives a post-doctoral scholarship from CNPq. Dr Rose is a post-doctoral student at the Laboratory of Comparative and Environmental Virology, Oswaldo Cruz Institute, Fiocruz, Brazil. Her research interests include RVA molecular epidemiology.

\section{Introduction} \label{sec:intro} Frequency domain transfer function identification is a widely-used engineering tool. Commonly this identification task is formulated as a nonlinear least squares (NLS) problem (\cite{ljung_system_1999,pintelon_system_2012}). \cite{sanathanan_transfer_1963} iteration (SK) is a popular approach to solve the NLS problem by fixed-point iterations, where each iteration is a linear least squares (LLS) problem. However, traditional implementations of SK iterations with monomial basis functions suffer from numerical issues, which can be an accuracy bottleneck. \cite{gustavsen_rational_1999} introduced the \emph{vector fitting} (VF) algorithm. VF algorithm is in the SK-iteration framework, but it utilizes barycentric representation (instead of monomial basis) that is updated at each iteration step. This approach has improved numerical properties (\cite{drmac_quadrature-based_2015, drmac_vector_2015}) and many variations followed. Orthonormal rational basis functions are utilized to further improve numerics (\cite{deschrijver_orthonormal_2007}). This is called \emph{orthonormal vector fitting} (OVF). See \cite{bultheel_orthogonal_2003} for data dependent orthogonal bases that yield optimal conditioning for SK-iterations. It is known that SK-iterations may not find a local minimum of the nonlinear problem when the frequency response measurements contain noise. This issue may necessitate hard relocation of the identified poles (\cite{grivet-talocia_improving_2006}), and use of nonlinear optimization approaches. A recent alternative is the frequency-domain instrumental variable (IV) iterations (\cite{van_herpen_optimally_2014}). The fixed-points of the linear IV iterations coincide with the stationary points of the original nonlinear problem. Hence if convergence is observed, a local minimum is found. IV iterations have worse condition numbers than SK iterations. \cite{van_herpen_optimally_2014} introduced bi-orthonormal bases that yield optimal conditioning for IV iterations to address this issue. This paper considers continuous-time transfer function estimation from frequency response measurements using SK and IV iterations with orthonormal rational basis functions (OVF). These functions lead to zero-pole-gain models, but it is still possible to perform grey-box transfer function estimation. We also consider numerical improvements to the OVF algorithm. These improvements are helpful when data-dependent orthogonal basis functions are not utilized for the sake of algorithmic simplicity. Our contributions are: \begin{itemize} \item An efficient method to enforce constraints on transfer function coefficients within the SK and IV iterations with any choice of basis functions. This allows incorporating prior knowledge into the estimation algorithm such as relative degree, integrators, and affine relationships between coefficients. \item A simple data scaling algorithm, and a domain mapping that maps the continuous domain (imaginary axis) to discrete (unit disk). These improve the numerics of underlying least squares problems. \item A minimal state-space realization for the orthonormal rational basis functions on the unit disk. \end{itemize} The organization of the paper is as follows. Section~\ref{sec:problem} presents the problem formulation and current approaches. Section~\ref{sec:main} presents the contributions. Section~\ref{sec:numerical} illustrates the algorithm performance on two examples. Concluding remarks are in Section~\ref{sec:conclusion}. The resulting algorithm is implemented in the \verb"tfest" function (when using frequency-domain data) in System Identification Toolbox (Release 2016b, for use with MATLAB). \section{Problem Statement and Current Approaches} \label{sec:problem} Assume that the frequency response measurements of a MIMO system $H(s)$ are available as $H_1,\ldots,H_l\in\mathbb{C}^{p \times m}$ at frequency points $s_i=j\omega_i$ for $i=1,\ldots,l$. The goal is to estimate a MIMO system $H_r(s)$ by minimizing the cost: \begin{equation} \label{eq:LSCost} J=\sum_{i=1}^l \|W(j\omega_i)\left(H_r(j\omega_i)-H_i\right)\|_F^2 \end{equation} where the degree of $H_r(s)$ is less than $r\times \min(m,p)$. $W(s)$ is an optional weight. $H_r(s)=N(s)/d(s)$ is a rational function with $d(s):\mathbb{C}^1\to\mathbb{C}^1$ and $N(s):\mathbb{C}^1\to\mathbb{C}^{p \times m}$. \cite{sanathanan_transfer_1963} iteration (SK) is a well-known method to estimate $H_r(s)$ iteratively by utilizing the fact that the problem \eqref{eq:LSCost} is linear for a fixed denominator. First, rewrite Equation \eqref{eq:LSCost} exactly: \begin{equation} J = \sum_{i=1}^l \frac{|W(j\omega_i)|^2}{|d(j\omega_i)|^2} \left\| N(j\omega_i)-d(j\omega_i)H_i \right\|_F^2 \end{equation} then iterate: \begingroup\makeatletter\def\f@size{9}\check@mathfonts \begin{equation} \label{eq:LSCostIter} J^{(k)} = \sum_{i=1}^l \frac{|W(j\omega_i)|^2}{|d^{(k-1)}(j\omega_i)|^2} \left\| N^{(k)}(j\omega_i)-d^{(k)}(j\omega_i)H_i \right\|_F^2 \end{equation} \endgroup Given $d^{(k-1)}(s)$ from the previous iteration, computation of the coefficients of $N^{(k)}$ and $d^{(k)}$ is a linear least squares problem. Two common choices for the initial denominator are $d^{(0)}(s)=1$ or $d^{(0)}(s)$ as a polynomial with lightly damped roots linearly or logarithmically spaced across $\omega_1$ and $\omega_l$. The problem in Eq. \eqref{eq:LSCostIter} has a trivial zero solution $d(s)=0$ and $N(s)=0$ that yields $J^{(k)}=0$. Two approaches to avoid the trivial solution are constraining the first denominator parameter to 1 or enforcing $\sum_{i=1}^l Re(d(j\omega_i))=1$ (\cite{gustavsen_improving_2006}). The key property of the SK fixed-point iterations is as follows: if there is no noise, and the estimated system order is equal to or greater than the true order, then the fixed-points of the SK iterations and the stationary points of the nonlinear problem coincide. Under these assumptions and barring potential numerical issues, the linear least squares problem in Equation \eqref{eq:LSCostIter} recovers the true system with $J=0$ in one step, regardless of the initial choice $d^{(0)}(s)$. Considering these assumptions, potential issues include: \begin{enumerate} \item Numerics: Large condition numbers that may arise in LLS matrices may prevent finding accurate solutions. The choice of basis functions for $N(s)$ and $d(s)$ is critical factor. \item Fixed-point iterations: When there is measurement noise or data cannot be captured by linear models, the fixed-points of SK iterations do not coincide with the stationary points of the nonlinear problem. The convergence of SK iterations in this case is an open problem. Even if the iterations converge, the result is not necessarily a local optima. IV iterations is an alternative approach whose fixed-points coincide with the stationary points of the nonlinear problem. \item Incorrect model order: This leads to the same issues in (2), but has the partial remedy of being a tunable parameter at estimation time. Model order estimation in SK and IV iteration framework is an open problem, but some promising results do exist (\cite{grivet-talocia_improving_2006, drmac_vector_2015}). \end{enumerate} The solution found for the NLS problem in Equation \eqref{eq:LSCost} depends on all the above factors. Any of these factors can constitute an accuracy bottleneck depending on the specific data. Sections \ref{sec:numericsAndVF} and \ref{sec:IVIterations} provide details of potential issues with numerics as well as SK and IV fixed-point iterations, respectively. \subsection{Basis Functions and Vector Fitting} \label{sec:numericsAndVF} The condition numbers of the least squares matrices that arise in SK iterations depend on the basis function choice and the weight $\frac{W(s)}{d^{(k-1)}(s)}$. The basis function choice involves algorithmic complexity versus accuracy trade-off. Some common basis functions and a brief discussion of this trade-off are presented here. A generic representation for the rational function $H_r(s)$ is: \begin{equation} \label{eq:baryOVF} H_r^{(k)}=\frac{N^{(k)}(s)}{d^{(k)}(s)}=\frac{\sum_{\nu=0}^r N_\nu^{(k)} B_\nu^{(k)}(s)}{\sum_{\nu=0}^r d_\nu^{(k)} B_\nu^{(k)}(s)} \end{equation} where $B_\nu^{(k)}(s):\mathbb{C}^1\to\mathbb{C}^1$ is the $\nu^{th}$ basis function utilized in iteration k. A simple implementation with monomials $B_\nu(s)=s^\nu$ leads to \begin{equation} H_r^{(k)}=\frac{N^{(k)}(s)}{d^{(k)}(s)}=\frac{N_0^{(k)}+\sum_{\nu=1}^{r}N_\nu^{(k)}\,s^\nu}{d_0^{(k)}+\sum_{\nu=1}^{r}d_\nu^{(k)}\,s^\nu}\end{equation} Monomial basis leads to large condition numbers. The vector fitting approach (\cite{gustavsen_rational_1999}) improves the numerics on two fronts. First, the $N^{(k)}(s)$ and $d^{(k)}(s)$ are expressed in barycentric form with $B^{(k)}_\nu(s)=\frac{1}{s-\lambda^{(k)}_\nu}$ where $\lambda^{(k)}_\nu\in\mathbb{C}^1$ is the $\nu^{th}$ distinct interpolation point: \begin{equation} \label{eq:baryVF} H_r^{(k)}=\frac{N^{(k)}(s)}{d^{(k)}(s)}=\frac{N_0^{(k)}+\sum_{\nu=1}^{r} \frac{N_\nu^{(k)}}{s-\lambda_\nu^{(k)}}}{d_0^{(k)}+\sum_{\nu=1}^{r} \frac{d_\nu^{(k)}}{s-\lambda_\nu^{(k)}}} \end{equation} This yields improved condition numbers over monomials. Second, $\lambda^{(k)}_\nu$ are chosen as the zeros of the identified $d^{(k-1)}(s)$ in the previous iteration step. This elegant choice eliminates the need for the weight $\frac{1}{d^{(k-1)}(s)}$, which typically worsens the condition numbers iterations evolve. These two ideas in the VF algorithm lead to better numerical stability (\cite{drmac_vector_2015}). The VF algorithm is not much more complicated than using monomial basis functions, yet it is significantly better regarding numerics. This makes it a good candidate for many applications. However, problems do exist where this method is not sufficient. One example is systems with closely packed or repeated poles. To see this, note that the VF algorithm uses the identified poles as the interpolation points $\lambda^{(k)}$ for barycentric representation. Repeated $\lambda^{(k)}$ values yield linearly dependent basis functions. Two further basis function candidates for handling such cases are highlighted here. One is orthonormal rational polynomials obtained by analytical Gram-Schmidt orthogonalization of barycentric basis functions (\cite{akcay_orthonormal_1999,ninness_basisfcns_1997}). These typically yield better condition numbers, and do not suffer from numerical issues with close or repeated poles (except when there are multiple integrators). The VF algorithm with such bases is called OVF. Orthonormal rational polynomials on the unit disk are utilized in \verb"tfest", and explained further in Section \ref{sec:ovfSSRealization}. Second option is data dependent orthogonal bases. See \cite{bultheel_orthogonal_2003}) for matrices observed in SK algorithm, and \cite{van_herpen_optimally_2014} for matrices in the IV algorithm. These data dependent orthogonal bases ensure that the least squares matrix have a condition number 1. The drawback is the higher computational complexity of the construction of these bases compared to the VF and OVF approaches. This is a future improvement direction for the \verb"tfest" implementation. \subsection{Instrumental Variable Iterations} \label{sec:IVIterations} In the presence of noise, nonlinearities or wrong model order choice, the fixed-points of SK iterations do not coincide with the stationary points of the nonlinear estimation problem. \cite{whitfield_asymptotic_1987} shows that even when the SK iterations converge to a solution, the solution is not necessarily a local optima. An alternative is the so-called instrumental-variable (IV) iterations. IV iterations have the property that its fixed-points coincide with the stationary points of the nonlinear least squares problem, even in presence of the aforementioned conditions (\cite{van_herpen_optimally_2014}). There is no convergence guarantee, but if convergence is observed, the result is a local optima. IV iterations aim to find the stationary points (i.e. where derivative with respect to parameters ($N_\nu^{(k)}$, $d_\nu^{(k)}$) are zero) of the nonlinear problem directly instead of directly trying to minimize the cost function in Equation \eqref{eq:LSCost}. The new cost function minimized in IV iterations is: \begingroup\makeatletter\def\f@size{8}\check@mathfonts \begin{multline*} J_{iv} = \sum_{i=1}^l |W(j\omega_{i})|^2\ \sum_{\nu=0}^r \left(\left\|\frac{\partial \tilde{J}_i}{\partial N_\nu}\right\|_F^2+ \left\|\frac{\partial \tilde{J}_i}{\partial d_\nu}\right\|_F^2\right) \end{multline*} \endgroup where $\tilde{J}_i(j\omega)=(H_r^{(k)}(j\omega)-H_i)(H_r^{(k)}(j\omega)-H_i)^T$ \begingroup\makeatletter\def\f@size{8}\check@mathfonts \begin{equation*} \begin{aligned} \frac{\partial \tilde{J}_i}{\partial N_\nu}&= -2 \left(H_r^{(k)}(j\omega_i)-H_i\right)\left(\frac{\partial H_r^{(k)}(j\omega_)}{\partial N_\nu}\right)^T \\ \frac{\partial \tilde{J}_i}{\partial d_\nu}&= -2 \left(H_r^{(k)}(j\omega_i)-H_i\right)\left(\frac{\partial H_r^{(k)}(j\omega_i)}{\partial d_\nu}\right)^T \end{aligned} \end{equation*} \endgroup This is a NLS problem due to nonlinear dependence of $H_r^{(k)}$ and its partial derivatives on denominator coefficients. The terms that appear in the denominator of the nonlinear cost function is replaced by the estimates from the previous iteration, similar to SK iterations. This again yields linear least squares problems for each iteration step. The IV iterations can be utilized standalone for solving the nonlinear optimization problem in Eq. \eqref{eq:LSCost}, as in \cite{van_herpen_optimally_2014}. The implementation in \verb"tfest" uses IV iterations in succession to the SK iterations. This is to find a solution near a reasonable local optima first with the SK iterations, which is then utilized for initializing the IV iterations. This is a conservative approach given that IV iterations are frequently successful at good local optima points. This conservative choice was done to increase the chance of finding good solutions given that \verb"tfest" has a wide user base and it is used for a large variety of datasets. \section{Contributions} \label{sec:main} \subsection{Domain Mapping} \label{sec:domainMapping} Numerical issues with continuous-time transfer function estimation are most prevalent when the measurement points $s_i=j\omega_i$ span a wide frequency range. The large magnitude variations of $s_i$ are observed across the rows of the least squared matrices, and lead to numerical issues. It is possible to perform the estimation on a different domain, and transform the final result back. This idea, for instance in the form of scaling the $s$-domain (\cite{pintelon_scaling_2005}), is found in the early continuous-time transfer function estimation literature where monomial bases are commonly used. More recent literature (VF, OVF) typically do not perform such domain mapping, and instead just scale the columns of the least squares matrices at each iteration step. This is because column scaling, combined with the use of barycentric or orthonormal rational polynomials, yields sufficient condition numbers on a wider range of estimation problems. Any invertible mapping can be utilized, and the bilinear mapping $q(s)=\frac{\alpha+s}{\alpha-s}$ is proposed here. This idea is from \cite{balas_2002}. Bilinear mappings provide more flexibility over simple scaling, while maintaining a linear relationship between the transfer function coefficients in the original and transformed domains. The latter property is utilized for efficient handling of parameter constraints discussed in Section \ref{sec:ParameterConstraints}. The mapping $q(s)$ maps the halfplane $s=j\omega$ for $\omega\in[0,\infty)$ to the upper half of the unit circle. The scalar $\alpha\in(0,\infty)$ is a design parameter. The mapped points in $q$ domain have unit magnitude. This eliminates the magnitude variations across the basis function rows observed in $s$ domain. An idea for $\alpha$ is to maximize the distance between the endpoints in $s$-domain through the transformation $q(s)$. Precisely, $\argmax_\alpha |q(j\omega_1)-q(j\omega_l)|$. The unique stationary point of this problem is $\alpha=\sqrt{\omega_1\omega_l}$. Effect of the bilinear domain mapping on the matrix condition numbers is tested with barycentric basis functions (VF) for a single SK iteration step in Equation \eqref{eq:LSCostIter}. The first denominator parameter is fixed to 1. The data $H_i(j\omega_i)$ is from the $20^{th}$ order model in Section \ref{sec:experimentsUnconstrained}. The weight $W(j\omega_i)$ and $d^{(k-1)}(j\omega_i)$ is set to 1. The interpolation points $\lambda_i$ are set as the true poles of the model. For domain mapping, the interpolation points are also mapped as $\bar{\lambda}_i=q(\lambda_i)$. These interpolation points represent a good candidate for the final SK-iteration step, hence the observed condition numbers have a direct impact on the final accuracy. The barycentric bases $1/(s-\lambda_i)$ yield the condition number $4.9{\cdot}10^{8}$. Barycentric bases in transformed domain $1/(q-\bar{\lambda}_i)$ yield $9.1{\cdot}10^{5}$. VF literature typically suggests scaling the columns of the least squares matrices by Euclidean norm. There are also suggestions against this scaling (\cite{drmac_vector_2015}) because it may amplify the impact of noise. \verb"tfest" does not use column scaling, but the impact of the column scaling is also tested. This scaling reduces the condition number with bases in original domain $1/(s-\lambda_i)$ to $1.2{\cdot}10^{3}$, and in the transformed domain $1/(q-\bar{\lambda}_i)$ to $8.1{\cdot}10^{1}$. These show that the basis functions in the transformed domain $q$ are better conditioned with or without column scaling. \subsection{Grey-Box Estimation: Parameter Constraints} \label{sec:ParameterConstraints} Many grey-box transfer function estimation scenarios involve fixed or bounded numerator or denominator coefficients, for instance known integrators, relative degree or bounds on a mass-spring-damper system parameters. It is straightforward to handle affine constraints within the SK iteration framework when the numerators and the denominator are expressed in terms of monomials. Then the estimated parameters are the transfer function coefficients themselves. This corresponds to solving a constrained least-squares problem in each iteration step, instead of an unconstrained one. When barycentric representation (VF) or orthonormal rational basis functions (OVF) are used, the estimated parameters are no longer the transfer function coefficients. The key point is that there is a linear relationship between the estimated parameters and the transfer function coefficients. Here a numerical method is presented to calculate this relationship for barycentric or orthonormal rational basis functions. Domain transformation (Section \ref{sec:domainMapping}) is also accounted for in this treatment. Make the following three assumptions without loss of generality to ease the notational burden: the transfer function is SISO, there are constraints only on the denominator coefficients, and barycentric representation (VF) is utilized. Let $d(s)=\sum_{\nu=0}^rd_{\nu}s^\nu$ be the denominator of the sought after transfer function. Let $d(q)=d_{q0}+\sum_{\nu=1}^{r}d_{q\nu}B_\nu(q)$ be the estimated denominator, in domain $q=\frac{\alpha+s}{\alpha-s}$, with basis functions $B_\nu(q)=\frac{1}{q-\lambda_\nu}$. The aim is to find the linear relationship between transfer function coefficients $d_{\nu}$ and estimated coefficients $d_{q\nu}$ for $\nu=0,1,...,r$. \begingroup\makeatletter\def\f@size{8}\check@mathfonts \begin{align} d(q) & = d_{q0}+\sum_{\nu=1}^r d_{q\nu}\, B_{\nu}(q) \label{eq:paramConstQ} \\ \nonumber & = \frac{\sum_{\nu=0}^r \bar{d}_{q\nu} \, q^\nu}{\prod_{\nu=1}^r (q-\lambda_\nu)} \\ \nonumber & = \frac{\sum_{\nu=0}^r \bar{d}_{q\nu} \, (\alpha+s)^\nu \, (\alpha-s)^{r-\nu}} {(\alpha-s)^r \, \prod_{\nu=1}^r (q-\lambda_\nu)} \\ & = \frac{\sum_{\nu=0}^r d_{\nu} \, s^\nu}{(\alpha-s)^r \, \prod_{\nu=1}^r (q-\lambda_\nu)} \label{eq:paramConstZ} \end{align} \endgroup Here $\bar{d}_{q\nu}$ are some intermediate variables. The right-hand side of Equation \eqref{eq:paramConstQ} contains the estimated parameters $d_{q\nu}$, and the numerator of the right-hand side of Equation \eqref{eq:paramConstZ} contains the transfer function coefficients of interest $d_\nu$. Both quantities appear linearly in the respective equations. This linear mapping can be extracted by evaluating these equations at $r+1$ points: \begingroup\makeatletter\def\f@size{8}\check@mathfonts \begin{align} \nonumber \begin{bmatrix} \vdots & \vdots & \vdots & \vdots \\ 1 & B_1(q) & \cdots & B_r(q) \\ \vdots & \vdots & \vdots & \vdots \end{bmatrix} \begin{bmatrix} d_{q0} \\ \vdots \\ d_{qr} \end{bmatrix} = \\ \begin{bmatrix} \ddots & 0 & 0 \\ 0 & \frac{1}{(\alpha-s)^r \, \prod_{\nu=1}^r (q-\lambda_\nu)} & 0 \\ 0 & 0 & \ddots \end{bmatrix} \begin{bmatrix} \vdots & \vdots & \vdots & \vdots \\ 1 & s & \cdots & s^r \\ \vdots & \vdots & \vdots & \vdots \end{bmatrix} \begin{bmatrix} d_{0} \\ \vdots \\ d_{r} \end{bmatrix} \label{eq:paramConstMatS} \end{align} \endgroup The aim is to solve this equation for $[d_0 \dots d_r]^T$. The first matrix on the right hand side of \eqref{eq:paramConstMatS} is diagonal, hence easy to invert. $[1\; s \cdots s^r]$ is a Vandermonde matrix. Choose the evaluation points $s$ uniformly spaced on the unit disk to make the Vandermonde matrix unitary (after scaling with $1/\sqrt{r+1}$), which is then also easy to invert. The constructed linear constraints can be utilized in SK iterations without modifying the cost function in Equation \eqref{eq:LSCostIter}. Each iteration step is a constrained LLS problem in this case. For IV iterations, the cost function also needs to be modified because stationary points of the NLS problem can be at points where derivative is not zero. The method of Lagrange multipliers can be used to extend the IV iteration cost function to find the stationary points of the constrained nonlinear problem (\cite{bertsekas_nonlinear_1999}). \subsection{Measurement Scaling} \label{sec:DataScaling} The measured frequency response $H_1,\dotsc,H_l\in\mathbb{C}^{p \times m}$ enters the SK and IV iteration linear least-squares matrices. Therefore, the choice of measurement units has an impact on the condition number of the least-squares matrices, and in turn the final fit quality. This is important when using monomial, barycentric (VF) or orthonormal rational polynomial (OVF) basis functions. Let $B^{(k)}=[B_1^{(k)} \dots B_r^{(k)}]$ be the basis function matrix at iteration $k$. For a given input-output channel $(i,j)$, the corresponding rows of the least squares matrices contain $B^{(k)}$ once unscaled, and once row-scaled by measurements $H(i,j)$. A heuristic is to balance the row-scaling induced by $H(i,j)$ around 1 by scaling $c(i,j)=\sqrt{\max_t|H_t(i,j)| \min_t|H_t(i,j)|}$, which yields scaled measurements $\bar{H}(i,j)=c(i,j)H(i,j)$. This magnitude scaling is tested on the example in Section \ref{sec:experimentsUnconstrained}. This scaling reduces the worst condition number observed during the SK and IV iterations from $4.2{\cdot}10^{14}$ to $2.6{\cdot}10^{12}$. This is a modest improvement due to good scaling of the original data. Nevertheless, measurement scaling is straightforward, computationally cheap and useful. \subsection{State-Space Realization for Orthonormal Rational Basis Functions} \label{sec:ovfSSRealization} The orthonormal rational basis functions on the unit disk are due to \cite{ninness_basisfcns_1997}. Let $\xi_\nu\in\mathbb{C}^1$ be the $\nu^{th}$ interpolation point where $\nu=1,2,\ldots,r$. These basis functions have the form in Eq. \eqref{eq:orthBasisReal} if $\xi_\nu$ is real, and the form in Eq. \eqref{eq:orthBasisCplx} if $(\xi_\nu,\xi_{\nu+1})$ are complex-conjugate pairs. \begin{equation} \label{eq:orthBasisReal} B_\nu(q) = \frac{\sqrt{1-|\xi_\nu|^2}}{q-\xi_\nu} \prod_{\nu=1}^{n-1} \frac{1-\xi_\nu^*q}{q-\xi_\nu} \end{equation} \begin{equation} \begin{aligned} \label{eq:orthBasisCplx} B_\nu(q) = \frac{\sqrt{1-|\xi_\nu|^2}(\beta_1q+\mu_1)}{q^2-2Re(\xi_\nu)+|\xi_\nu|^2} \prod_{\nu=1}^{n-2} \frac{1-\xi_\nu^*q}{q-\xi_\nu}\\ B_{\nu+1}(q) = \frac{\sqrt{1-|\xi_\nu|^2}(\beta_2q+\mu_2)}{q^2-2Re(\xi_\nu)+|\xi_\nu|^2} \prod_{\nu=1}^{n-2} \frac{1-\xi_\nu^*q}{q-\xi_\nu} \end{aligned} \end{equation} Here $B_n:\mathbb{C}^1\to\mathbb{C}^1$ is the $n^{th}$ basis function, $Re(\cdot)$ is the real part of a complex number. There are infinite number of choices for $\beta_1$, $\beta_2$, $\mu_1$, $\mu_2$ (\cite{ninness_basisfcns_1997}). Each SK and IV iteration step estimate a denominator polynomial $d^{(k)}(q)=d_0+\sum_{\nu=1}^rd_\nu B_\nu(q)$, then extract its zeros to find the interpolation points $\xi_\nu$ for the next iteration. A minimal state-space realization of $d^{(k)}(q)$ is needed for this operation. A state-space realization for orthonormal rational polynomials on the imaginary axis is in \cite{deschrijver_orthonormal_2007}. Here a similar construction is presented for orthonormal rational polynomials defined on the unit disk. \begin{figure}[h!] \setlength{\unitlength}{1pt} \centering \begin{picture}(210, 31)(0,0) \multiput(0,20)(37,0){6}{\vector(1,0){7}} \put(7,10){\framebox(30,21){$ \frac{1-\xi_1^*q}{q-\xi_1}$}} \put(12,0){$\scriptstyle G_1(q)$} \put(44,10){\framebox(30,21){$ \frac{1-\xi_2^*q}{q-\xi_2}$}} \put(49,0){$\scriptstyle G_2(q)$} \put(91,17){$\cdots$} \put(118,10){\framebox(30,21){$ \frac{1-\xi_\nu^*q}{q-\xi_\nu}$}} \put(123,0){$\scriptstyle G_\nu(q)$} \put(165,17){$\cdots$} \put(192,10){\framebox(30,21){$ \frac{1-\xi_r^*q}{q-\xi_r}$}} \put(198,0){$\scriptstyle G_r(q)$} \end{picture} \caption{Components of the basis functions $B_n(q)$} \label{fig:cascadeIC} \end{figure} \vspace{-.1in} Denote the state-space model matrices to be constructed as $(A,B,C,D)$. The idea is to choose $(A,B)$ such that the states $(qI-A)^{-1}B$ correspond to the basis functions $[B_1(q) B_2(q) \dots B_r(q)]^T$, except the $\sqrt{1-|\xi_\nu|^2}$ scalings. Consider the cascade connection in Fig. \ref{fig:cascadeIC} for this purpose. Realize the $n^{th}$ component is with $(a_n,b_n,c_n,d_n)$ such that its output is $\frac{1-\xi_\nu^*q}{q-\xi_\nu}$, and its state is $\frac{1}{{q-\xi_\nu}}$. Then the series connection of all components per Equation \eqref{eq:ovfSSAB} has its $\nu^{th}$ state as $B_\nu(q)$ (except the scaling $\sqrt{1-|\xi_\nu|^2}$). If $(\xi_\nu,\xi_{n+1})$ are complex-conjugate pairs, $G_n(q)G_{n+1}(q)$ must be realized together in order to have a realization with real coefficients. In this case the output is $\frac{|\xi_\nu|^2q^2-2Re(\xi_\nu)+1}{q^2-2Re(\xi_\nu)+|\xi_\nu|^2}$, and the states are $\frac{1}{{q^2-2Re(\xi_\nu)+|\xi_\nu|^2}}[\beta_1q+\mu_1 \;\; \beta_2q+\mu_2]^T$. A realization that fits these requirements is in Table \ref{table:ssRealization}. This specific realization is obtained in three steps. First, fix $b_v=[\beta_1\;\beta_2]^T$ and $d_v=|\xi_\nu|^2$ to avoid over-parametrization. Second, solve $(qI-a_v)^{-1}b_v=\frac{1}{{q^2-2Re(\xi_\nu)+|\xi_\nu|^2}}[\beta_1q+\mu_1 \; \beta_2q+\mu_2]$ for the elements of $a_v$. Finally, solve $c_v(qI-a_v)^{-1}b_v+d_v=\frac{|\xi_\nu|^2q^2-2Re(\xi_\nu)+1}{q^2-2Re(\xi_\nu)+|\xi_\nu|^2}$ for the elements of $c_v$. The matrices $(C,D)$ capture the scalings $\sqrt{1-|\xi_\nu|^2}$ and the estimated parameters $d_0,\dots,d_r$. Specifically, $D=d_0$ and $\nu^{th}$ element of $C\in\mathbb{R}^{1 \times r}$ is $d_\nu\sqrt{1-|\xi_\nu|^2}$. \begin{table}[ht] \centering \begin{tabular}{ | l | l | } \hline $\xi_\nu$ real & If $\xi_\nu\neq0$, $(a_v,b_v,c_v,d_v)=(\xi_\nu,1,1-\xi_\nu^2,-\xi_\nu)$ \\ \cline{2-2} & If $\xi_\nu=0$, $(a_v,b_v,c_v,d_v)=(0,1,1,0)$ \\ \hline $(\xi_\nu, \xi_{v+1})$ & $a_v\in\mathbb{R}^{2\times2},b_v=[\beta_1\; \beta_2]^T,c_v\in\mathbb{R}^{1\times2},d_v=|\xi_\nu|^2$ \\ complex pair & $a_v(1,1)=\frac{|\xi|^2+(\mu_1\mu_2)/(\beta1\beta2)+(2Re(\xi_\nu)\mu_2)/\beta_2}{\mu_2/\beta_2-\mu_1/\beta_1}$ \\ & $a_v(1,2)=\frac{\mu_1+2\beta_1Re(\xi_\nu)-\beta_1a_v(1,1)}{\beta_2}$ \\ & $a_v(2,1)=\frac{\mu_2+\beta_2a_v(1,1)}{\beta_1}$ \\ & $a_v(2,2)=2Re(\xi_\nu)-a_v(1,1)$ \\ & $c_v(1,1)=\frac{2\mu_2Re(\xi_\nu)(|\xi|^2-1)+(|\xi_\nu|^4-1)\beta_2}{\beta_1\mu_2-\mu_1\beta_2}$ \\ & $c_v(1,2)=\frac{2Re(\xi_\nu)(|\xi_\nu|^2-1)-c_v(1,1)\beta_1}{\beta_2}$\\ \hline \end{tabular} \caption{Realizations of components of interconnection in Fig. \ref{fig:cascadeIC}} \label{table:ssRealization} \end{table} \begin{equation} \begin{aligned} \label{eq:ovfSSAB} A&=\begin{bmatrix} a_1 & 0 & \cdots & 0 \\ b_2c_1 & a_2 & \cdots & 0 \\ b_3d_2c_1 & b_3c_2 & \cdots & 0 \\ b_4d_3d_2c_1 & b_4d_3c_2 & \cdots & \vdots \\ \cdots & \cdots & \cdots & 0 \\ b_kd_{k-1} \cdots d_2c_1 & b_kd_{k-1} \cdots d_3c_2 & \cdots & a_k \\ \end{bmatrix} \\ B&=[b_1 \;\;\; b_2d_1 \;\;\; b_3d_2d_1 \cdots \;\;\; b_kd_{k-1}\dots d_1]^T \end{aligned} \end{equation} \subsection{Algorithm Summary} The algorithm in \verb"tfest" can be summarized as: \begin{enumerate} \item Map $s$ domain to $q$ via $q(s)=\frac{\alpha+s}{\alpha-s}$ \item Scale measurements $H_i$ (Section \ref{sec:DataScaling}). \item Initial fit: Use monomial basis with $d^{(0)}(q)=1$. \item SK iterations: Use orthonormal rational polynomial basis functions on unit disk. Iterate until maximum number of iterations, or convergence. Update basis functions at each step. \item IV iterations: Use the final set basis functions used in SK iterations. Iterate until maximum number of iterations, or convergence. \item Use the best solution found for the NLS problem throughout all steps (initial fit, SK and IV iterations). Calculate the corresponding zero-pole-gain model. \item Revert $s$ to $q$ domain mapping via $s=\frac{\alpha(q-1)}{q+1}$. \item Revert measurement scaling. \item Convert zero-pole-gain to transfer function model. \end{enumerate} \section{Numerical Experiments} \label{sec:numerical} The \verb"tfest" command in System Identification Toolbox (Release 2016b, for use with MATLAB) is used for the experiments. The results compare very favorably with previous releases and other existing algorithms for many frequency domain datasets. Two experiments are presented here due to space constraints. \subsection{Experiment Without Parameter Constraints} \label{sec:experimentsUnconstrained} A randomly generated model was used for the experiment: \begin{equation} G(s) = \sum_{k=1}^{10} \frac{r_k\;\omega_k^2}{s^2 + 2\zeta_k\omega_ks + \omega_k^2} \end{equation} where the parameters $r_k$, $\omega_k$, $\zeta_k$ are randomly generated from a predetermined range. This is a dynamically rich model: it contains ten lightly damped modes that are spread over seven decades. The frequency response of $G(s)$ was extracted at 700 logarithmically spaced points in $[10^{-1},10^{6}]$ $rad/s$. Multiplicative noise was added as $G(j\omega_i)(1+\epsilon(j\omega_i))$ where $\epsilon\in\mathbb{C}^1$ is $\epsilon(j\omega_i)=n_1e^{jn_2}$. $n_1$ and $n_2$ are zero mean Gaussian random numbers with variance 0.01, so the signal-to-noise ratio is approximately $20\,dB$. \begin{figure}[ht] \centering \includegraphics[width=0.46\textwidth]{blackBoxEstimate.eps} \caption{$20^{th}$ order black-box model estimate} \label{fig:blackBoxExample} \end{figure} Figure \ref{fig:blackBoxExample} presents the measurements and the $20^{th}$ order model estimate. The model captures the dynamics of interest, despite the high model order for transfer functions. Only the valley near $2.74\cdot10^4\;rad/s$ was missed. The main reason is the limited contribution of the small magnitude data at and around the valley to the cost function. Using a frequency-based weight $W(s)$ is a straightforward remedy to this problem, which is available through the WeightingFilter option of the \verb"tfestOptions" command. \subsection{Experiment with Parameter Constraints} \label{sec:experimentsConstrained} The model $G(s)$ in Equation \eqref{eq:constrainedModel} is used for the experiment. $G(s)$ has an integrator and a relative degree 3. Three leading numerator coefficients and the last denominator coefficient are fixed to 0 during estimation. The frequency response is extracted at 300 logarithmically spaced points in $[1,10^4]$ $rad/s$. Multiplicative noise is added as explained in Section \ref{sec:experimentsConstrained}, to have a $20dB$ signal-to-noise ratio. \begin{equation} \label{eq:constrainedModel} \small G(s) = \frac{1.2{\cdot}10^8\,s^2 + 4.8{\cdot}10^7\,s + 5{\cdot}10^{10}}{s^5 + 200\,s^4 + 4{\cdot}10^6\,s^3 + 1.7{\cdot}10^6\,s^2 + 1.6{\cdot}10^9\,s} \end{equation} Figure \ref{fig:constrainedFitBode} shows the Bode magnitude plot for the noisy measurements and the fitted model. The estimated model in Equation \eqref{eq:constrainedModelEst} captures system dynamics well. The approach in Section \ref{sec:ParameterConstraints} successfully enforces the poles and zeros in transformed domain $q$ to honor the transfer function coefficient constraints in $s$ domain. \begin{figure}[ht] \centering \includegraphics[width=0.46\textwidth]{grayBoxEstimate.eps} \caption{Measured response and grey-box model estimate} \label{fig:constrainedFitBode} \end{figure} \begin{equation} \label{eq:constrainedModelEst} \small G_{est}(s) = \frac{1.17{\cdot}10^8\,s^2 + 3.32{\cdot}10^7\,s + 4.91{\cdot}10^{10}}{s^5 + 211\,s^4 + 4.0{\cdot}10^6\,s^3 + 1.4{\cdot}10^6\,s^2 + 1.6{\cdot}10^9\,s} \end{equation} \section{Concluding Remarks} \label{sec:conclusion} The SK and IV iterations are commonly implemented with monomial, barycentric or orthonormal rational polynomial basis functions on the imaginary axis. Transforming the problem domain onto the unit disk frequently leads to important numerical improvements. A simple scaling of the measurement data is also helpful. Even though use of barycentric or orthonormal rational polynomials lead to zero-pole-gain models, the numeric benefits of these bases can still be utilized for grey-box transfer function estimation. The linear relationship between these basis functions and transfer function coefficients can be efficiently calculated by utilizing the properties of the Vandermonde matrices. \begin{ack} We would like to acknowledge and thank Professor Peter J. Seiler for sharing his expertise with us, which was instrumental for this work. We would also like to acknowledge and thank Professor Lennart Ljung for his continuous support, feedback and encouraging comments which motivated us and improved the quality of this work. \end{ack}

The views expressed in this article reflect the results of research conducted by the author and do not necessarily reflect the official policy or position of the Department of the Navy, Department of Defense, nor the United States Government. I am a contracted employee of the United States government. This work was prepared as part of my official duties. Title 17 U.S.C. 105 provides that 'copyright protection under this title is not available for any work of the United States Government'. Title 17 U.S.C. 101 defines a U.S. Government work as work prepared by a military service member or employee of the U.S. Government as part of that person's official duties. Introduction {#s1} ============ Concussion, or mild traumatic brain injury, significantly impacts warfighter readiness for events such as deployment and combat. With over 380,000 diagnosed concussions among US service members (SMs) since 2000 ([@B1]), rehabilitation approaches to expedite symptom recovery after concussion are highly relevant to health and readiness in this population. A growing body of evidence has shown the importance of monitoring and regulating activity level in the first hours to days following a concussion ([@B2]--[@B8]); however, little attention has been given to the contribution of post-acute (i.e., ≥1 month\[s\]) activity level on symptom recovery in the weeks and months that follow. Furthermore, most of these studies have focused on sports-related concussion ([@B2]--[@B5], [@B7], [@B8]), while only a limited number provide information about active duty military personnel ([@B9]), who may have a greater risk for concussion, and among whom the consequences of persistent impairments impact military force readiness. Research has also been sparse in the evaluation of symptom recovery against activity by specific categories (e.g., cognitive, physical, vestibular/balance). This type of information could help to improve clinical guidance for primary care managers to educate and provide guidance to their patients on the most appropriate type and intensity of activities at different stages of recovery to optimize return to pre-injury activities and symptom resolution. Activity participation, whether too little or too much, during the acute recovery period from concussion is known to impact recovery rates. Unrestricted physical activity during acute stages of concussion has been shown to negatively impact recovery ([@B2], [@B4]--[@B6]), and increased cognitive activity shortly following injury to lengthen recovery time ([@B3]). In contrast, studies have also demonstrated that too much rest may not provide clear benefits and may even negatively influence recovery ([@B6], [@B10]). In a systematic review of sports-related concussion studies, an initial period of moderate physical and cognitive rest has been shown to provide benefit during the acute post-injury phase ([@B11]). Prolonged physical and cognitive rest beyond the currently recognized recommendation by expert consensus of 2 days has also been demonstrated to be associated with higher levels of total symptoms over 10 days after injury ([@B6], [@B12]). Considering activity levels beyond the acute stage of injury, findings also support the importance of engagement in some activity, but an understanding of the appropriate activity level and activity type remains largely unknown. One study evaluating a cohort of student-athletes suggested that moderate physical and cognitive activity (vs. minimal activity) within 30 days of injury may be necessary for symptom recovery and better neurocognitive performance ([@B5]). In a study of concussed athletes and non-athletes, ages 16--53, with symptoms lasting at least 6 weeks post-injury but stable for 2--3 weeks, a treadmill exercise for 5--6 days per week was found to significantly decrease the level and number of symptoms after 6 weeks from measures prior to the treadmill exercise ([@B7]). Overall, these studies suggest that unrestricted activity at the acute stage of concussion may exacerbate symptoms or delay recovery; however, some level of activity in both acute and post-acute stages of concussion may be beneficial for recovery. Regardless of outcome, existing studies have been limited to cohorts of pediatric and adult athletes ([@B2]--[@B5], [@B7], [@B8]), and findings may not be generalizable to a military population whose occupational environment differs significantly, potentially impacting the optimal types and intensities of activity during recovery. For athletes, the Consensus Statement on Concussion in Sport has recommended immediate physical and cognitive rest after a concussion, followed by a stepwise return-to-play after clearance by treating healthcare providers ([@B13]). A similar set of guidelines, the Progressive Return to Activity (PRA) Clinical Recommendation (CR), which emphasize gradual return to activity after rest at the acute stage of concussion, was developed specifically for SMs by military and civilian subject-matter experts and published by the Defense and Veterans Brain Injury Center (DVBIC) ([@B1]). Recent research among SMs suggests that activity participation within the acute stage of injury impacts symptom resolution over time ([@B9]); however, it is unknown how post-acute levels of activities, particularly the type of activities conducted to include military-specific tasks (e.g., combat training), may influence continued improvement in symptomatology. Optimal patterns of activity during the course of concussion recovery may depend on the severity of symptoms experienced in the acute stage. In a recent study by our group, greater activity level at acute stages of concussion was associated with higher levels of symptoms over time, but only among those with high levels of acute symptoms ([@B9]). In the current study, we build upon our previous findings by evaluating the contribution of activities at later stages of concussion rehabilitation post-concussion. The primary objectives of this study were: 1) to evaluate the relationship between activity level (in total and by categories: cognitive, lifestyle, physical, vestibular/balance, and military-specific) at post-acute stages of concussion (specifically, at 1 and 3 month\[s\] post-injury) and subsequent symptom level among an active duty military population; and 2) to determine whether this association differs by acute symptom severity. Materials and Methods {#s2} ===================== Data for this study were drawn from the broader DVBIC (Defense and Veterans Brain Injury Center, [RRID:SCR_004505](https://scicrunch.org/resolver/RRID:SCR_004505)) PRA CR Study, which investigated the impact of implementing the DVBIC PRA CR focused on gradual return to unrestricted activity post-concussion among SMs ([@B1], [@B14]). The parent study has been previously described in detail ([@B15]). Data from the current study included only concussed SM participants who received "usual treatment" from providers who had not yet received focused training on the PRA CR. "Usual treatment" refers to any treatment deemed appropriate by the treating clinicians; it was not experimentally controlled. This study was carried out in accordance with the recommendations of the Naval Medical Center San Diego Institutional Review Board, with concurrence from Womack Army Medical Center, and Human Research Protections Program administrative review by the Defense Health Agency. These committees approved the protocol in compliance with all applicable federal regulations governing the protection of human subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. Study Participants ------------------ Participants who received "usual treatment" included 64 SMs, 18--48 years of age, recruited from clinics and operational medical units at three U.S. military installations (Army \[southeast U.S.\], Navy \[southwest U.S.\], and Marine Corps \[southwest U.S.\]). Eligible participants had sustained a concussion within the previous 72 h of enrollment into the study. Those who suffered other concussions within 12 months of the concussion in question were excluded. Electronic medical records were reviewed to confirm concussion diagnoses, which must have met the Veterans Administration/Department of Defense definition for concussion ([@B16]). After enrollment, participants completed a face-to-face baseline assessment (i.e., within 72 h of injury), and were followed-up at 1 week, 1 month, 3 months, and 6 months post-injury via a telephone or in-person interview. Of the 64 participants from the parent study, there were 39 with complete data to assess activity level at 1 month and symptom level at 3 months; 38 had complete data to assess activity level at 1 month and symptom level at 6 months; and 33 had complete data to assess activity level at 3 months and symptom level at 6 months. The demographic and military characteristics of the 39 participants evaluated for activity level at 1 month and symptom level at 3 months (the greatest number of participants in any of the proposed study analyses) were comparable to those of the "usual treatment" group from the parent study consisting of 64 participants. Measures and Procedures ----------------------- ### Assessment of Activities This study utilized a 60-item Activity Questionnaire ([@B9], [@B15]), which included activities that are recommended to be avoided or encouraged at various stages of recovery according to the DVBIC PRA CR education materials ([@B17]). For each activity item, participants were asked "Since your concussion, did you \[activity item stated\]" during baseline and at 1 week, or, for other follow-up interviews, "In the last 2 weeks, did you \[activity item stated\]." Although psychometric properties are not yet available, this questionnaire was streamlined for analysis by removal of 11 items due to low variance, low correlation with other activity items, and inconsistent interpretation by SMs as demonstrated by questions asked of the interviewers during follow-up. An interdisciplinary group of investigators representing Epidemiology, Neuropsychology, Primary Care Sports Medicine and Physical Therapy (RR, JB, KS, KM, AC) reviewed the remaining 49 items and determined 6 individual categories: (1) cognitive (12 items); (2) lifestyle (10 items); (3) physical (20 items); (4) vestibular/balance (17 items); and (5) military-specific (6 items) ([@B9]). Eighteen items were included in multiple categories (e.g., walk briskly was categorized as both a physical and vestibular/balance activity). To ensure that scores represented the same direction, seven items (e.g., sleep 6--8 h a night) were reverse coded. Three items did not fit into any category ("wear dark glasses or sunglasses" and "do familiar tasks \[e.g., vehicle maintenance check\]," "rest all day"), but were included in the total activity score. Each item was scored 0--4 with responses such as "never" (0), "every few days" (1), "some days" (2), "most days" (3), and "every day" (4). ### Assessment of Neurobehavioral Symptoms The Neurobehavioral Symptom Inventory (NSI) was used to assess concussion symptoms ([@B18]). Participants were asked to evaluate 22 symptom items since their concussion (during baseline and at 1 week), or, in the last 2 weeks, during the 1-, 3-, and 6-month follow-up interviews. Based on previous exploratory analyses, these items were categorized into 4 factors: (1) cognitive (4 items); (2) vestibular (3 items); (3) somatosensory (7 items); and (4) affective (6 items) ([@B19], [@B20]). Two items regarding hearing difficulty and changes in appetite were only included in the total symptom score as they did not fit into any of the 4 factors. Responses ranged from 0 to 4 and included "none" (0), "mild" (1), "moderate" (2), "severe" (3), and "very severe" (4). Among SMs, NSI has been shown to have high internal consistency (total alpha = 0.95; subscale alpha = 0.88--0.92) ([@B21]). Further, external validity was exemplified with moderate correlation (*r* = 0.41) showing NSI differentiating veterans with TBI status from those without ([@B21]). ### Calculation of Activity and Symptom Scores To maximize the use of available data, a prorated summary score was calculated for activity and symptom level for participants with at least one missing item needed to calculate each variable ([@B22]). All non-missing values per variable were summed and multiplied by the total number of possible items. This, in turn, was divided by the actual number of items with non-missing values. For each category of activity or symptom, prorated scores were based on the items included within each category. A list of items within each category has been reported previously ([@B9]). In the current study, only one participant had 1 missing item for the calculation of activity level. Prorated scores were transformed into z-scores for comparability across categories. Each mean and standard deviation were based on the type of activity or symptom level in question. For example, activity level was assessed as a total score and by categories such as cognitive activity. To calculate the z-score for total score, this took into consideration the mean and standard deviation of all of the activity item responses from all participants. To calculate the z-score for cognitive activity, this took into consideration the mean and standard deviation of activity items that were categorized as cognitive activity, also from all participants. All of these calculations were done at each time point. ### Assessment of Covariates Demographic characteristics (i.e., age, sex, and education) and military information (i.e., branch affiliation, current rank, and number of deployments) were assessed at baseline. Statistical Analyses -------------------- To ascertain the contribution of demographic characteristics and military history on activity level during post-acute stages of symptom recovery, Spearman correlation was used for age, a continuous variable, and one-way analyses of variance for categorical variables. The modifying impact of symptom level at the acute stage of concussion was also analyzed and presented as a dichotomized variable (high vs. low) based on median cut-offs (total and by categories). Although evaluating this variable continuously would provide more statistical power, dichotomization was chosen to enhance clinical relevance. First-order interaction terms were created as a product of this variable with post-acute activity level. In these investigations, linear regression was utilized to assess the relationship between activity level at either 1 or 3 month(s) and subsequent symptoms (i.e., at 3 and 6 months, or at 6 months, respectively) among those with either high vs. low level of symptoms at the acute stage of injury (objective \#2). Partial correlation (r~*p*~) was used to evaluate relationships that did not differ by acute symptoms (objective \#1). All models were adjusted for activity level within 72 h of injury and symptom level at the time of each activity variable being examined. To investigate assumptions for the utility of linear regression and Pearson correlation in our study, we used the following techniques: to identify non-linearities in the data, we used scatter plots and graphs of augmented component-plus-residual plots with LOWESS smoothing; to assess the assumption of normality, we evaluated graphs of standardized normal probability plots (P-P plots) to assess non-normality in the middle range of data, and quantiles of the residual against quantiles of a normal distribution (Q-Q plots) to assess non-normality at the tail-ends of data; to test the assumption of homoscedasticity, a plot of residuals vs. fitted (predicted) values were examined; and to evaluate evidence of auto-correlation, we utilized the Durbin-Watson statistic. All assessments were done between each activity level score (total and by categories) against each symptom level score (total and by categories). To address the presence of outliers that may significantly affect results, both the activity and symptom level data (z-scores) were truncated to +/−3 standard deviation from the mean. Main effects and interaction were considered significant at a *p* \< 0.05. All statistical analyses were completed using Stata statistical software (Stata, [RRID:SCR_012763](https://scicrunch.org/resolver/RRID:SCR_012763)), release 15 (StataCorp, 2017, College Station, TX). Results {#s3} ======= There was an overall decrease in symptom progression (total and by categories) over time (see [Figures 1A](#F1){ref-type="fig"},**B**). Among the 39 participants evaluated for activity level at 1 month and symptom level at 3 months, a significant decrease in total, cognitive and vestibular symptoms were found from 1 month to 3 months post-injury. Although the reduction continues to 6 months, the progression was not statistically significant. See [Figure 1A](#F1){ref-type="fig"}. Among the 33 participants evaluated for activity level at 3 months and symptom level at 6 months, the decrease in symptom progression (total and by categories) was not statistically significant (see [Figure 1B](#F1){ref-type="fig"}). {#F1} [Table 1](#T1){ref-type="table"} illustrates the characteristics of the study sample. Participants had a mean age of 26 years (range = 19--41) and were mostly men (89.7%), married (46.0%), and had 12--15 years of education (82.0%). Most served in the US Army (69.2%), were non-commissioned officers (E4 and E5, 51.3%) and were deployed at least once (56.4%). There were no significant differences in total activity or symptom score at 3 months post-injury by any assessed demographic or military characteristic. Study participants had a median symptom score of 20 (range = 2-61) at the acute stage of concussion. ###### Demographic and military characteristics of service members and their relation to activity and symptom score at 3 months post-injury, overall (*n* = 39). **Variable** **Value** **Total activity score** **Total symptom score** --------------------------------------------------------- ----------------------------- -------------------------- ------------------------- ------------- ------- Age in years, mean (SD) 26 (6.2) (Range = 19--41) NA 0.938 NA 0.393 Sex, *n* (%) 0.804 0.283    Men 35 (89.7) 107.3 (19.6) 15.2 (12.5)    Women 4 (10.3) 109.8 (8.5) 8.3 (5.3) Education, *n* (%) 0.754 0.624    High school diploma or GED 13 (33.3) 103.6 (14.1) 12.3 (9.0)    Some college (1--3 years) 19 (48.7) 111.2 (22.5) 16.1 (13.5)    College graduate (4+ years) 4 (10.3) 103.0 (18.8) 19.5 (16.3)    Some graduate school 1 (2.6) 118.0 (NA) 8.0 (NA)    Graduate/professional program 2 (5.1) 102.0 (2.8) 6.0 (7.1) Marital status, *n* (%)[^a^](#TN1){ref-type="table-fn"} 0.821 0.496    Never married 15 (40.5) 104.3 (19.2) 17.4 (14.4)    Married 17 (46.0) 107.4 (20.3) 13.9 (10.8)    Divorced 4 (10.8) 114.5 (15.2) 10.5 (11.1)    Separated 1 (2.7) 104.0 (NA) 1.0 (NA) Branch, *n* (%) 0.102 0.351    US Navy 3 (7.7) 96.3 (13.3) 5.7 (8.1)    US Marine Corp 9 (23.1) 118.3 (13.7) 17.4 (11.6)    US Army 27 (69.2) 105.2 (19.4) 14.4 (12.4) Rank, *n* (%) 0.870 0.883    Junior enlisted (E1-E3) 8 (20.5) 108.8 (15.9) 15.8 (10.0)    Non-commissioned officer (E4-E5) 20 (51.3) 103.4 (19.7) 13.7 (13.0)    Non-commissioned officer staff (E6-E9) 8 (20.5) 102.6 (22.9) 16.5 (14.6)    Officers (O1-O10) 3 (7.7) 111.3 (9.9) 10.7 (2.5) Number of deployments, *n* (%) 0.895 0.211    0 17 (43.6) 107.1 (20.1) 17.2 (9.8)    1+ 22 (56.4) 108.9 (18.0) 12.3 (13.4) Acute symptom score, median (IQR) 20 (15--37) (Range = 2--61) NA NA NA NA *IQR = Interquartile Range*. *n = 37 for this variable*. Activity Level at 1 Month and Subsequent Symptoms at 3 and 6 Months ------------------------------------------------------------------- The association between activity level at 1 month (total or by categories) and subsequent symptoms at 3 and 6 months did not differ by the level of symptoms within 72 h of injury. [Table 2](#T2){ref-type="table"} provides the correlations between activity level at 1 month and symptom level at 3 and 6 months post-injury. Greater total activity at 1 month post-injury was significantly correlated with lower cognitive (r~*p*~ = −0.41, *p* = 0.012) and somatosensory (r~*p*~ = −0.36, *p* = 0.028) symptoms at 3 months. Subscale analyses suggested that total score correlations were driven primarily by physical and vestibular/balance activities. Greater physical activity at 1 month was significantly correlated with lower cognitive (r~*p*\ =~−0.44, *p* = 0.006) and vestibular (r~*p*\ =~−0.37, *p* = 0.025) symptoms at 3 months post-injury. Greater vestibular/balance activity at 1 month was also significantly correlated with lower total (r~*p*\ =~−0.39, *p* = 0.017) and cognitive (r~*p*\ =~−0.55, *p* \< 0.001) symptoms at 3 months post-injury. No significant correlations were found between total, cognitive, lifestyle and military-specific activities at 1 month and symptoms (total or by category) at either 3 or 6 months. ###### Partial correlation between activity level at 1 month (total and by categories) and symptom level at 3 and 6 months (total and by categories), adjusted for activities at acute stage of concussion and symptoms at 1 month post-injury. **Symptom level months post-injury** **Activity level at 1 month post-injury, correlation coefficient (*****p*****-value)** -------------------------------------- ---------------------------------------------------------------------------------------- --------------- --------------- ---------------------------------------------------- ------------------------------------------------------ --------------- **3 MONTHS (*****n*** **=** **39)** Total score −0.30 (0.070) 0.002 (0.992) −0.06 (0.730) −0.30 (0.068) **−0.39 (0.017)[^\*^](#TN2){ref-type="table-fn"}** −0.07 (0.661) Cognitive **−0.41 (0.012)[^\*^](#TN2){ref-type="table-fn"}** −0.08 (0.622) 0.03 (0.868) **−0.44 (0.006)[^\*^](#TN2){ref-type="table-fn"}** **−0.55 (\<0.001)[^\*^](#TN2){ref-type="table-fn"}** −0.14 (0.423) Vestibular −0.17 (0.310) 0.13 (0.451) 0.18 (0.288) **−0.36 (0.026)[^\*^](#TN2){ref-type="table-fn"}** −0.31 (0.066) 0.09 (0.606) Somatosensory **−0.36 (0.028)[^\*^](#TN2){ref-type="table-fn"}** −0.07 (0.683) −0.17 (0.327) −0.30 (0.076) −0.30 (0.067) −0.21 (0.203) Affective −0.17 (0.312) 0.03 (0.852) 0.03 (0.880) −0.19 (0.270) −0.30 (0.067) −0.03 (0.847) **6 Months (*****n*** **=** **38)** Total score −0.11 (0.506) 0.06 (0.731) −0.06 (0.727) −0.19 (0.275) −0.16 (0.343) 0.09 (0.617) Cognitive −0.21 (0.227) 0.10 (0.558) 0.08 (0.663) −0.31 (0.064) −0.30 (0.080) −0.02 (0.893) Vestibular −0.15 (0.370) 0.10 (0.567) −0.07 (0.695) −0.28 (0.101) −0.17 (0.319) 0.05 (0.761) Somatosensory −0.07 (0.667) 0.11 (0.514) −0.04 (0.822) −0.14 (0.411) −0.13 (0.460) 0.08 (0.661) Affective −0.10 (0.576) −0.10 (0.550) −0.11 (0.535) −0.12 (0.500) −0.10 (0.560) 0.09 (0.585) *Significant p-value at a level \< 0.05 (bolded)*. Activity Level at 3 Months and Subsequent Symptoms at 6 Months -------------------------------------------------------------- The relationships between activity level at 3 months and subsequent symptoms at 6 months post-injury did not vary by the level of symptoms within 72 h of injury. Further, no significant relationships were found between activity level at 3 months (total or by categories) and symptoms at 6 months post-injury (data not shown). Discussion {#s4} ========== This study provided preliminary evidence that activity level at post-acute stages of concussion may play an important role in improving symptoms among SMs; specifically, SMs with higher activity levels at 1-month post-injury reported lower level of post-concussion symptoms 2 months later. These findings held even after controlling for severity of symptoms at 1 month and activity level shortly after injury, suggesting that the effects of post-acute activity levels on later symptom outcomes were not simply due to pre-existing differences in trajectory of recovery. The results of this study suggest that although low levels of activity have been previously shown to be beneficial during the acute stages of injury among SMs ([@B9]), higher levels of activity may provide benefit at later stages in this population. The effects of activity level on later symptoms appeared to level off after approximately 3 months, possibly due to a greater proportion of SMs having achieved successful symptom recovery around this time. As we reported previously in data drawn from the parent study ([@B9]), only 26.8% of SMs demonstrated clinically significant levels of post-concussive symptoms by 3 months post-injury. Of interest, the type of activity was important to symptom resolution; specifically, higher rates of physical and vestibular/balance activities at 1 month were associated with reduced total, cognitive, and vestibular symptoms at 3 months. In contrast, cognitive, lifestyle and military-specific activities at 1 month were not associated with symptoms at 3 months. These findings were consistent with our previous study evaluating the impact of early post-injury activity, in which both physical and vestibular/balance activities were also shown to drive the relationship between activity level and symptoms over time ([@B9]). The current study did not find the association between activities at post-acute stages of concussion and subsequent symptoms to vary by the level of acute symptoms; however, other factors such as lifetime history of concussion could modify this relationship, as symptoms from previous injury may lower cerebral reserve that would otherwise be used for symptom recovery related to the current concussion. Future studies that evaluate the contribution of lifetime concussion history on symptom recovery among concussed SMs are warranted. The findings of this study demonstrate the importance of primary care managers to monitor specific types of activity for symptom recovery at the post-acute stages of concussion, as focusing on activity in total may overlook the specific impact of physical and vestibular/balance activities on symptom resolution. This could, in turn, lead to poor recovery. The preliminary findings from this paper also support exploration of rehabilitation programming/clinical care that includes physical and vestibular/balance activities, notably at chronic stages of concussion. A limited number of studies have evaluated the contribution of activities at the post-acute stages of concussion on improving symptoms ([@B7]), and none among an active duty military population. Aside from our recent work examining acute activities ([@B9]), to our knowledge, no previous studies have considered multiple activity categories, including military-specific items, at a granular level as investigated in the current analyses. The PRA CR developed by DVBIC, from which the activities assessed in this study are obtained, provides an algorithm for activity progression, which can be standardized across military populations as a part of acute and post-acute concussion management ([@B17]). The use of this CR may improve outcomes by providing guidance to primary care managers, particularly in reengaging in physical and/or vestibular/balance activities as supported by our preliminary findings, to optimize the speed of recovery. This may not only guide the prescription of specific patterns of activity, but also, the determination of whether patients should receive medical waivers from engaging in usual duties. Future studies will include the evaluation of activity progression on improving symptom levels, comparing those who receive usual care to those who receive care according to the guidelines published in the DVBIC PRA CR. The strengths of this study included the use of longitudinal activity and symptom data that extend 6 months post-injury within a population of military SMs. This study also evaluated demographic and military service information to assess for potential confounding factors. Additionally, data on activity items and categorization specific to a military population were collected to best address the objectives of this study. There were also limitations to consider. Although activity level was evaluated against symptom level at a later time point, controlling for symptoms at the time of activity assessment to support directionality of the association, we cannot infer causality in the significant relationships found. In addition, the Activities Questionnaire used in this study was developed *de novo* in order to be consistent with the recommendations found in the PRA CR education materials. It should be noted, however, that careful item screening was used to narrow the final list of items used in analyses. Further, this questionnaire not only included military-specific activities, but it also allowed for activity categorization not queried in previously studied and/or published surveys, most of which were aimed to evaluate a different population of individuals with moderate to severe traumatic brain injury. Along with activity level, the level of symptoms was self-reported which may be subject to recall bias. Although self-reported instruments are time- and cost-effective, future studies utilizing more objective assessment of activity (e.g., wearable devices, clinically-monitored activity, and symptom level) might help to validate the results of our findings. Finally, sample size was limited and, in part due to participant attrition over time, conservative corrections for multiple comparisons were not feasible in this study. However, even with a limited sample size, significant results were found. Nonetheless, the findings in this study were preliminary and analyses were exploratory. Future longitudinal studies with a larger sample size may be necessary not only to confirm the findings in this study, but also, to potentially detect significant effects that may have been missed in this study due to limited sample size. Additionally, with a larger sample, the trajectories of the residuals in the relationship between activity and symptom level might be better defined and might perhaps suggest the use of a more complex analytic approach. Our findings provide preliminary evidence that activity level during the post-acute stages of concussion remains an important factor in improving symptoms, thus, advocating for continued monitoring and patient management even up to 1 month post-injury. Taken together with our recent finding that greater activity levels immediately following injury (i.e., within 72 h) result in poorer symptom status among SMs ([@B9]), these findings support recommendations for a gradual increase in activity ([@B1], [@B10], [@B13], [@B14], [@B23]--[@B26])---rather than extended rest---to promote symptom resolution among SMs with recent concussion. Also consistent with our previous work ([@B9]), the results highlight the importance of physical and vestibular/balance activities in managing symptoms over time. Further research is necessary to evaluate the mechanisms of these associations and potential differences in the impact of various types of activities on symptom resolution. Future studies that evaluate how changes in activity level over time influence symptom recovery will also be critical to the continued development of clinical protocols for improving health outcomes of military personnel who have been diagnosed with concussion. Cumulative findings may inform how primary care managers and rehabilitation providers educate and guide their patients regarding progressive return to activity to optimize outcomes and military readiness. Data Availability {#s5} ================= The datasets analyzed for this study are not publicly available but will be submitted to FITBIR beginning in 2019. Requests to access the datasets should be directed to Dr. Rosemay Remigio-Baker (<rosemay.a.remigio-baker.ctr@mail.mil>). Ethics Statement {#s6} ================ This study was carried out in accordance with the recommendations of the Naval Medical Center San Diego Institutional Review Board, with concurrence from Womack Army Medical Center, and Human Research Protections Program administrative review by the Defense Health Agency. These committees approved the protocol in compliance with all applicable federal regulations governing the protection of human subjects. All subjects gave written informed consent in accordance with the Declaration of Helsinki. Author Contributions {#s7} ==================== RR-B, ME, JB, EG, and WC were involved in the development of concept and analytical approach. All analyses were conducted by RR-B. Screening and categorization of activity items used in the study were done by EG, RR-B, JB, KS, KM, and AC. KM and AC served as subject-matter experts regarding activity involvement during concussion recovery among service members. KS served as a topic expert on TBI and education, particularly as it pertains to the Progressive Return to Activity Clinical Recommendation. AC and TA were involved in data collection and provided information regarding study procedures and interaction with participants to interpret data. EG and FQ provided knowledge in the structure and content of the parent study that served as the source of data for this investigation. KS, PS, and LM provided insight into military physical and psychiatric environment. All authors contributed to the development and editing of the manuscript. Conflict of Interest Statement ------------------------------ RR-B was employed by Venesco LLC. JB and KS were employed by the General Dynamics Health Solutions. ME was employed by the American Hospital Services Group LLC. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The authors express gratitude to all of the research coordinators, research assistants, research associates and others who were involved in the progression and completion of the study. Additionally, and most importantly, we would like to thank the service members and providers who provided valuable data and time as participants in this study, which could not have been possible without their contribution. **Funding.** This work was supported by the Defense Health Program 6.7 (award number: D6.7_14_C2_I\_14_J9_1077) and the Defense and Veterans Brain Injury Center. SM : Service Members DVBIC : Defense and Veterans Brain Injury Center PRA : Progressive Return to Activity CR : Clinical Recommendation. [^1]: Edited by: Noah D. Silverberg, University of British Columbia, Canada [^2]: Reviewed by: Nick Reed, Holland Bloorview Kids Rehabilitation Hospital, Canada; Tamara McLeod, A.T. Still University, United States [^3]: This article was submitted to Neurorehabilitation, a section of the journal Frontiers in Neurology

\section{Generalized Metropolis with force bias \label{appendix:fb}} In this appendix, we describe our second approach using the generalized Metropolis procedure to accelerate the sampling of paths in auxiliary field (AF) space. We introduce a dynamic force bias, analogous to what is employed in the branching random walk methods in constrained path or phase-free AFQMC \cite{AFQMC-lecture-notes-2013}, in proposing the updates of the field values, which improves the acceptance ratio and hence the MC efficiency. To facilitate the description of the sampling algorithm we first give a brief sketch of the standard path-integral AFQMC approach, on which more detailed descriptions can be found in, for example, Refs.~\cite{assaad-lec-note} and \cite{AFQMC-lecture-notes-2013}. Ground state AFQMC measures the static properties by \begin{equation} \label{eqn:projection} \langle \hat{O}\rangle =\frac{\langle \psi_T|\,\exp(-\beta \hat{H}/2)\,\hat{O}\,\exp(-\beta\hat{H}/2)\,|\psi_T \rangle}{ \langle \psi_T|\,\exp(-\beta \hat{H})\,|\psi_T \rangle}\,, \end{equation} where the Hamiltonian $\hat{H}\equiv \hat{K}+\hat{V}$ is given by Eq.~(\ref{eq:H}). We apply the usual Trotter-Suzuki breakup \begin{equation} e^{-\Delta \tau \hat{H}} \backsimeq e^{-\Delta \tau \hat{K}/2} e^{-\Delta \tau \hat{V}} e^{-\Delta \tau \hat{K}/2}\, \end{equation} and the Hubbard-Stratonovich (HS) decomposition \cite{PhysRevB.28.4059} \begin{align} e^{\Delta \tau U n_{i\uparrow}n_{i\downarrow}} &=\frac{1}{2} \sum_{x_i=\pm1} e^{(\gamma x_i -\Delta \tau U/2) (n_{i\uparrow} + n_{i\downarrow}-1)} \label{eq:HS-Hirsch-charge} \\ &\equiv \frac{1}{2} \sum_{x_i=\pm1} \hat{b}_i(x_i)\,, \nonumber \end{align} with $ \cosh(\gamma) = \exp( -\Delta \tau U/2 )$, arriving at the form \begin{equation} e^{-\Delta\tau {\hat H}}= \int d{\mathbf x}\, p({\mathbf x}) {\hat B}({\mathbf x})\,, \label{eq:HS} \end{equation} where ${\mathbf x}=\{x_1, x_2, \cdots,x_{{\mathcal N}_s}\}$ The probability density function $p({\mathbf x})$ is uniform for the $2^{{\mathcal N}_s}$ AF configurations under the choice of HS in Eq.~(\ref{eq:HS-Hirsch-charge}), and the one-body propagator is ${\hat B}({\mathbf x})= e^{-\Delta \tau \hat{K}/2} \prod_i \hat{b}_i(x_i)\,e^{-\Delta \tau \hat{K}/2} $. The expression in Eq.~(\ref{eqn:projection}) is then re-written as a path integral of $M\equiv \beta/\Delta\tau$ time slices. Let us consider the $l$-th time slice, and introduce the notation \begin{eqnarray*} \langle\psi_l| & = & \langle \psi_T|\, {\hat B}({\mathbf x}^{(M)}){\hat B}({\mathbf x}^{(M-1)}) \cdots {\hat B}({\mathbf x}^{(l+1)})\,e^{-\Delta \tau \hat{K}/2} \\ |\psi_r\rangle & = & e^{-\Delta \tau \hat{K}/2}\,{\hat B}({\mathbf x}^{(l-1)}){\hat B}({\mathbf x}^{(l-2)}) \cdots {\hat B}({\mathbf x}^{(1)}) \,|\psi_T\rangle\,, \end{eqnarray*} which are both single Slater determinant wave functions if we choose $|\psi_T\rangle$ to be a Slater determinant. The integrand of the path-integral in the denominator of Eq.~(\ref{eqn:projection}) then becomes \begin{equation} {\mathcal W}({\mathbf x})=p({\mathbf x})\, \langle \psi_l | \prod_{i=1}^{{\mathcal N}_s} \hat{b}_i(x_i) |\psi_r\rangle\,, \end{equation} where ${\mathbf x}$ denotes the collection of AF at time slice $l$. In the standard way of sampling ${\mathcal W}$, one proposes to flip each auxiliary-field $x_i$ one by one, and sweeps through ${\mathbf x}$. We will update the entire configuration ${\mathbf x}$ (or a sub-cluster of ${\mathbf x}$ for very large system sizes), simultaneously. We define a force bias \cite{AFQMC-lecture-notes-2013}: \begin{equation} \bar{n}_{i\sigma}= \frac{ \langle \psi_l |n_{i\sigma} |\psi_r\rangle } {\langle \psi_l |\psi_r\rangle}\,, \label{eq:FB-dynamic} \end{equation} and propose updates of the fields with the probability density: \begin{equation} {\mathcal P}({\mathbf x})\propto p({\mathbf x}) \prod_{i=1}^{{\mathcal N}_s} e^{\gamma x_i (\bar{n}_{i\uparrow} + \bar{n}_{i\downarrow}-1)}\, \end{equation} which can be sampled directly. Detailed balance then leads to a Metropolis acceptance probability given by \begin{equation} {\mathcal A}({\mathbf x}\rightarrow {\mathbf x}') = \min\{1,\frac{{\mathcal W}({\mathbf x}^{\prime})\,{\mathcal P}({\mathbf x})}{{\mathcal W}({\mathbf x})\,{\mathcal P}({\mathbf x}^{\prime})}\}\,. \end{equation} Note that the probability function for proposing transitions does not depend on the ``current'' configuration of AF, i.e., ${\mathcal P}({\mathbf x}\rightarrow {\mathbf x}')={\mathcal P}({\mathbf x}')$. If ${\mathcal P = W}$, all updates will be accepted. Because of the force bias, ${\mathcal P}$ approximates ${\mathcal W}$ up to ${\mathcal O}(\sqrt{\Delta \tau})$, leading to typically high acceptance ratio. Although we have used the discrete charge HS decomposition, the algorithm generalizes straightforwardly to continuous HS transformations. We comment that the use of the dynamic force bias in Eq.~(\ref{eq:FB-dynamic}) effectively introduces a background subtraction \cite{PhysRevB.88.125132,AFQMC-lecture-notes-2013} in the decomposition of Eq.~(\ref{eq:HS-Hirsch-charge}). That is, if one were to employ the standard updating algorithms \emph{without the force bias}, one would find Eq.~(\ref{eq:HS-Hirsch-charge}) much less efficient than a continuous charge decomposition which subtracts a constant background. This discrepancy in efficiency grows more as the system density decreases, which is especially relevant since the systems studied here are at the low density limit. (See Ref.~\cite{PhysRevB.88.125132} for an analysis of the efficiency of HS transformations, and Ref.~\cite{PhysRevE.70.056702} for discussion on how the dynamic force bias automatically introduces an optimal constant background shift.) Some other features of our algorithm are: \begin{itemize} \item Since we always work in the dilute limit, the memory is saved by only storing the wave function and calculating the Green function on the fly. We divide the path of $M$ slices into $\sqrt{M}$ blocks, and only track one block each time. The wave function at the beginning of each block is stored. The largest number of wave functions stored in our code is $\sim2\sqrt{M}$. \item The wave function is transformed between real and momentum space by fast Fourier transformation, so that all the one-body operators during projection are diagonal, and Green functions in different space are easily obtained. \item When we only need the energy, we separate it into kinetic and potential energy. They are diagonal either in momentum or real space, where we do not need to calculated the whole Green function. To improve statistics, we measure the energy anywhere along the path and combine them, including the mixed estimator on both side. \item The standard determinantal QMC formalism as sketched above turns out to have a divergence of the Monte Carlo variance. We discuss the variance problem and its solution separately elsewhere \cite{HShi-SZhang-to-be-published}. The solution involves the introduction of a bridge link, which we have implemented in the calculations presented here. The force bias and basic sampling algorithm described above remain unchanged. \end{itemize} \section{Extrapolation to the continuum limit \label{appendix:fss}} We have described the extrapolation procedure of our lattice results to the continuum limit, and the subsequent analysis to reach the thermodynamic limit. Here we illustrate the finite size extrapolation in few-body systems. The extrapolation to the continuum limit, for a fixed number of particles, must be consistent and independent of the type of kinetic energy dispersion. For a two-body problem on the lattice, exact results can be obtained for large system sizes by mapping to a one-body problem in the center of mass system. The results are shown in Fig.~\ref{fig:fss}(a), which fit well a 4th-order polynomial function in $1/L$. We see from the inset that the coefficient on the linear term is zero within numerical precision. We also show the finite size effect in the four-body problem from QMC, in Fig.~\ref{fig:fss}(b), reaching large lattice sizes. The same general behavior is seen as in the two-body problem. We have also studied the finite-size behavior of the BCS solution, finding similar trends but with different slopes. In the many-body system, our QMC data are consistent with these observations as well. They are thus fitted with a 4th-order polynomial function with a vanishing $1/L$ coefficient, as described in the main text. \begin{figure*}[htbp] \includegraphics[scale=0.5]{Size-scaling.eps} \caption{\label{fig:fss} (Color online) Extrapolation of finite-size lattice to the continuum limit in few body problems. Panel (a) shows exact diagonalization results for the two-body problem at $\ln(a k_F)=0.5$, while panel (b) shows QMC solutions for the four-body problem at $\ln(a k_F)=0.0$. In each case, results are obtained for both the Hubbard and the quadratic dispersions. A 4th-order polynomial function in $1/L$ fits well both dispersions, and the extrapolated results in continuum limit agree well with each other. The insets indicate that the coefficients on $1/L$ are negligible in both cases. } \end{figure*} \section{Equation of state data \label{appendix:eos}} We list the data for the equation of state in Fig.~\ref{fig:energy}. The QMC energy data are calculated by our branching random walk approach with BCS trial wave functions. \begin{table} \caption{Data of the equation of state in Fig.~\ref{fig:energy}. The interaction strength, given in the first column, are $\ln(a k_F)=y+\ln(2)/2$, with $y$ from $-0.75$ to $6$ (in increments of $0.25$ up to $y=2$, then increments of 0.5 up to $y=5$). }. \begin{ruledtabular} \begin{tabular}{cccc} $\ln(a k_F)$ & $E_{QMC}/E_{FG}$ & Error bar & $E_{BCS}/E_{FG}$\\ \hline -0.403426 & -5.512634 & 0.000619 & -4.651252 \\ -0.153426 & -3.262997 & 0.000487 & -2.427641 \\ 0.096574 & -1.884889 & 0.000325 & -1.078969 \\ 0.346574 & -1.027841 & 0.000453 & -0.260958 \\ 0.596574 & -0.487667 & 0.000335 & 0.235190 \\ 0.846574 & -0.137058 & 0.000272 & 0.536119 \\ 1.096574 & 0.096228 & 0.000203 & 0.718642 \\ 1.346574 & 0.256943 & 0.000167 & 0.829348 \\ 1.596574 & 0.371799 & 0.000162 & 0.896491 \\ 1.846574 & 0.456471 & 0.000141 & 0.937204 \\ 2.096574 & 0.521804 & 0.000173 & 0.961859 \\ 2.346574 & 0.572904 & 0.000111 & 0.976740 \\ 2.846574 & 0.647340 & 0.000103 & 0.990927 \\ 3.346574 & 0.700067 & 0.000067 & 0.997096 \\ 3.846574 & 0.737144 & 0.000128 & 0.997307 \\ 4.346574 & 0.767283 & 0.000099 & 0.997765 \\ 4.846574 & 0.793547 & 0.000068 & 1.000206 \\ 5.346574 & 0.813073 & 0.000053 & 1.000436 \\ 6.346574 & 0.842689 & 0.000036 & 1.000654 \end{tabular} \end{ruledtabular} \end{table} \section{Full digits for Table \ref{tab:fitparam} \label{appendix:fitparam}} \begin{table} \caption{\label{tab:fitparamfull} Final parameter values (full digits) in the parametrization [Eqs.~(\ref{eqn:fitcross}-\ref{eqn:fitbec})] of the exact EOS from QMC. } \begin{ruledtabular} \begin{tabular}{cccc} & $a^l_0$ & -11.804127317953723 & \\ & $a^l_1$ & 14.675499370762239 &\\ & $a^l_2$ & -4.855080880566919 &\\ & $a_0$ & -0.819842357425408 &\\ & $a_1$ & 0.1273251139440354 &\\ & $a_2$ & 0.0685123559420463 &\\ & $a_3$ & -0.014505432043856327&\\ & $a_4$ & -0.009191602440101383 &\\ & $a_5$ & 0.004190575139056055 &\\ & $a_6$ & -0.0006367374265820822 &\\ & $a_7$ & 3.431232818866204$\times 10^{-5}$ &\\ & $a^r_2$ & -0.060852400057644876 &\\ & $a^r_3$ & 0.36401186693517423 &\\ & $a^r_4$ & -0.61531422724189 \end{tabular} \end{ruledtabular} \end{table} \begin{acknowledgments} We thank J.~Carlson for useful discussions. This research was supported by DOE (grant no.~DE-SC0008627), NSF (grant no.~DMR-1409510), and the Simons Foundation. Computing was carried out at the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725, and at the computational facilities at the College of William and Mary. \end{acknowledgments}

\section*{Introduction} The $q$-de Rham cohomology of a polynomial ring is a $\Z[q]$-linear complex given by replacing the usual derivative with the Jackson $q$-derivative $\nabla_q(x^n)= [n]_qx^{n-1}dx$, where $[n]_q$ is Gauss' $q$-analogue $\frac{q^n-1}{q-1}$ of the integer $n$. In \cite{scholzeqdef}, Scholze discussed the $(q-1)$-adic completion of this theory for smooth rings, explaining relations to $p$-adic Hodge theory and singular cohomology, and conjecturing that it is independent of co-ordinates. We show that $q$-de Rham cohomology naturally arises as a functorial invariant of $\Lambda$-rings (Theorem \ref{mainthm}), and that its $(q-1)$-adic completion depends only on a $\Lambda_P$-ring structure (Theorem \ref{mainthm3}), for $P$ the set of residue characteristics; a $\Lambda_P$ ring has a lift of Frobenius for each $p \in P$. This recovers the known equivalence between de Rham cohomology and complete $q$-de Rham cohomology over the rationals, while giving no really new functoriality statements for smooth schemes over $\Z$. However, in mixed characteristic, it means that complete $q$-de Rham cohomology depends only on a lift $\Psi^p$ of absolute Frobenius locally generated by co-ordinates with $\Psi^p(x_i)=x_i^p$. Given such data, we construct (Proposition \ref{cartierprop}) a quasi-isomorphism between Hodge cohomology and $q$-de Rham cohomology modulo $[p]_q$, extending the local lift of the Cartier isomorphism in \cite[Proposition 3.4]{scholzeqdef}. Taking the Frobenius stabilisation of the complete $q$-de Rham complex of $A$ yields a complex resembling the de Rham--Witt complex. We show (Theorem \ref{mainthm4}) that up to $(q^{1/p^{\infty}}-1)$-torsion, the $p$-adic completion of this complex depends only on the $p$-adic completion of $A[\zeta_{p^{\infty}}]$ (where $\zeta_n$ denotes a primitive $n$th root of unity), with no requirement for a lift of Frobenius or a choice of co-ordinates. The main idea is to show that the stabilised $q$-de Rham complex is in a sense given by applying Fontaine's period ring construction $A_{\inf}$ to the best possible perfectoid approximation to $A[\zeta_{p^{\infty}}]$. As a consequence, this shows (Corollary \ref{finalcor}) that after attaching all $p$-power roots of $q$, $q$-de Rham cohomology in mixed characteristic is independent of choices, which was already known after base change to a period ring, via the comparisons of \cite{BhattMorrowScholze} between $q$-de Rham cohomology their theory $A\Omega^{\bt}$. We expect that the dependence of these cohomology theories either on Adams operations at the residue characteristics (for de Rham) or on $p$-power roots of $q$ (for variants of de Rham--Witt) is unavoidable, and that the conjectures of \cite{scholzeqdef} might thus be slightly optimistic. Some of the strongest evidence for the conjectures is provided by the lifts of the Cartier isomorphisms, which rely on a choice of Frobenius. On the other hand, the comparison theorems of \cite{BhattMorrowScholze} can be seen as a manifestation of $q$-de Rham--Witt complexes; although they do not require a lift of Frobenius, they involve all $p$-power roots of $q$. The essence of our construction of $q$-de Rham cohomology of $A$ over $R$ is to set $q$ to be an element of rank $1$ for the $\Lambda$-ring structure, and to look at flat $\Lambda$-rings $B$ over $R[q]$ equipped with morphisms $A \to B/(q-1)$ of $\Lambda$-rings over $R$. If these seem unfamiliar, reassurance should be provided by the observation that $(q-1)B$ carries $q$-analogues of divided power operations (Remark \ref{DPrmk}). For the variants of de Rham--Witt cohomology in \S \ref{DRWsn}, the key to giving a characterisation independent of lifts of Frobenius is the factorisation of the tilting equivalence for perfectoid algebra via a category of $\Lambda_p$-rings. I would like to thank Peter Scholze for many helpful comments, in particular about the possibility of a $q$-analogue of de Rham--Witt cohomology, and Michel Gros for spotting a missing hypothesis. \tableofcontents \section{Comparisons for $\Lambda$-rings} We will follow standard notational conventions for $\Lambda$-rings. These are commutative rings equipped with operations $\lambda^i$ resembling alternating powers, in particular satisfying $\lambda^k(a+b)= \sum_{i=0}^{k} \lambda^i(a)\lambda^{k-i}(b)$, with $\lambda^0(a)= 1$ and $\lambda^1(a)=a$. For background, see \cite{borgerBasicGeomI} and references therein. The $\Lambda$-rings we encounter are all torsion-free, in which case the $\Lambda$-ring structure is equivalent to giving ring endomorphisms $\Psi^n$ for $n \in \Z_{>0}$ with $\Psi^{mn}=\Psi^m \circ \Psi^n$ and $\Psi^p(x) \equiv x^p \mod p$ for all primes $p$. If we write $\lambda_t(f):=\sum_{i \ge 0} \lambda^i(f)t^i$ and $\Psi_t(f):= \sum_{n \ge 1} \Psi^n(f)t^n$, then the families of operations are related by the formula $\Psi_t= -t\frac{d\log\lambda_{-t} }{d t}$. We refer to elements $x$ with $\lambda^i(x)=0$ for all $i>1$ (or equivalently $\Psi^n(x)=x^n$ for all $n$) as elements of rank $1$. \subsection{The $\Lambda$-ring $\Z[q]$} \begin{definition} Define $\Z[q]$ to be the $\Lambda$-ring with operations determined by setting $q$ to be of rank $1$. \end{definition} We now consider the $q$-analogues $[n]_q:= \frac{q^n-1}{q-1} \in \Z[q]$ of the integers, with $[n]_q!=[n]_q[n-1]_q\ldots [1]_q$, and $\binom{n}{k}_q= \frac{[n]_q!}{[n-k]_q![k]_q!}$. \begin{remark} To see the importance of regarding $\Z[q]$ as a $\Lambda$-ring observe that the binomial expressions \begin{align*} \lambda^k(n)= \tbinom{n}{k},\quad \lambda^k(-n)= (-1)^k\tbinom{n+k-1}{k} \end{align*} have as $q$-analogues the Gaussian binomial theorems \begin{align*} \lambda^k([n]_q)= q^{k(k-1)/2}\tbinom{n}{k}_q,\quad \lambda^k(-[n]_q) = (-1)^k\tbinom{n+k-1}{k}_q, \end{align*} as well as Adams operations \[ \Psi^i([n]_q)= [n]_{q^i}. \] \end{remark} For any torsion-free $\Lambda$-ring, localisation at a set of elements closed under the Adams operations always yields another $\Lambda$-ring, since $\Psi^p(a^{-1})-a^p = (\Psi^p(a)a^p)^{-1}(a^p- \Psi^p(a))$ is divisible by $p$. \begin{lemma}\label{lambdalemma} For the $\Lambda$-ring structure on $\Z[x,y]$ with $x,y$ of rank $1$, the elements \[ \lambda^n(\tfrac{y-x}{q-1}) \in\Z[q, \{(q^n-1)^{-1}\}_{n \ge 1},x,y] \] are given by \begin{align*} \lambda^k(\tfrac{y-x}{q-1}) &=\frac{(y-x)(y-qx)\ldots(y-q^{k-1}x)}{(q-1)^k[k]_q!},\\ &= \sum_{j=0}^k\frac{q^{j(j-1)/2} (-x)^jy^{k-j}}{[j]_q![k-j]_q!}. \end{align*} \end{lemma} \begin{proof} The second expression comes from multiplying out the Gaussian binomial expansions. The easiest way to prove the first is to observe that $\lambda^k(\tfrac{y-x}{q-1})$ must be a homogeneous polynomial of degree $k$ in $x,y$, with coefficients in the integral domain $\Z[q, \{(q^n-1)^{-1}\}_{n \ge 1}]$, and to note that \[ \lambda^k(\tfrac{q^nx-x}{q-1})= \lambda^k([n]_qx)= q^{k(k-1)/2}\tbinom{n}{k}_qx^k. \] Thus $\lambda^k(\frac{y-x}{q-1})$ agrees with the homogeneous polynomial above for infinitely many values of $\frac{y}{x}$, so must be equal to it. \end{proof} \begin{remark}\label{DPrmk} Note that as $q\to 1$, Lemma \ref{lambdalemma} gives $ (q-1)^k\lambda^k(\frac{y-x}{q-1})\to \frac{(x-y)^k}{k!}$. Indeed, for any rank $1$ element $x$ in a $\Lambda$-ring we have $\lambda_{(q-1)t} (\frac{x}{q-1})=e_q(xt)$, the $q$-exponential, with multiplicativity and universality then implying that $\lambda_{(q-1)t} (\frac{a}{q-1})$ is a $q$-deformation of $\exp(at)$ for all $a$. Thus $(q-1)^k\lambda^k(\frac{a}{q-1})$ is a $q$-analogue of the $k$th divided power $(a^k/k!)$. An explicit expression comes recursively from the formula \[ [k]_q(q-1)\lambda^k(\tfrac{a}{q-1})=\sum_{i>0} \lambda^i(a)\lambda^{k-i}(\tfrac{a}{q-1}). \] \end{remark} \begin{lemma}\label{subringbasis} For elements $x,y$ of rank $1$, the $\Lambda$-subring of $\Z[q, \{(q^n-1)^{-1}\}_{n \ge 1},x,y]$ generated by $q,x,y, \frac{y-x}{q-1}$ has basis $\lambda^k(\frac{y-x}{q-1})$ as a $\Z[q,x]$-module. \end{lemma} \begin{proof} The $\Lambda$-subring clearly contains the $\Z[q,x]$-module $M$ generated by the elements $\lambda^k(\frac{y-x}{q-1})$, which are also clearly linearly independent. Since $\Z[x,q]$ is a $\Lambda$-ring, it suffices to show that $M$ is closed under multiplication. By Lemma \ref{lambdalemma}, we know that \[ \lambda^i(\tfrac{y-x}{q-1})\lambda^j(\tfrac{y-q^ix}{q-1})= \tbinom{i+j}{i}_q\ \lambda^{i+j}(\tfrac{y-x}{q-1}). \] We can rewrite $\frac{y-q^ix}{q-1}= \frac{y-x}{q-1}- [i]_qx$, so $\lambda^j(\frac{y-q^ix}{q-1})-\lambda^j(\frac{y-x}{q-1})$ lies in the $\Z[q,x]$-module spanned by $\lambda^m(\frac{y-x}{q-1}) $ for $m<j$. By induction on $j$, it thus follows that \[ \lambda^i(\tfrac{y-x}{q-1})\lambda^j(\tfrac{y-q^ix}{q-1}) - \lambda^i(\tfrac{y-x}{q-1})\lambda^j(\tfrac{y-x}{q-1}) \in M, \] so the binomial expression above implies $ \lambda^i(\frac{y-x}{q-1})\lambda^j(\frac{y-x}{q-1}) \in M $. \end{proof} \subsection{$q$-cohomology of $\Lambda$-rings } \begin{definition} Given a $\Lambda$-ring $R$, say that $A$ is a $\Lambda$-ring over $R$ if it is a $\Lambda$-ring equipped with a morphism $R \to A$ of $\Lambda$-rings. We say that $A$ is a flat $\Lambda$-ring over $R$ if $A$ is flat as a module over the commutative ring underlying $R$. \end{definition} \begin{definition} Given a morphism $R \to A$ of $\Lambda$-rings, we define the category $\Strat^q_{A/R}$ to consist of flat $\Lambda$-rings $B$ over $R[q]$ equipped with a compatible morphism $A \to B/(q-1)$, such that the map $A \to B/(q-1)$ admits a lift to $B$; a choice of lift is not taken to be part of the data, so need not be preserved by morphisms. \end{definition} More concisely, $\Strat^q_{A/R}$ is the Grothendieck construction of the functor \[ (\Spec A)_{\strat}^q\co B \mapsto \im(\Hom_{\Lambda,R}(A,B)\to \Hom_{\Lambda,R}(A,B/(q-1))) \] on the category $f\Lambda(R[q])$ of flat $\Lambda$-rings over $R[q]$. \begin{definition} Given a flat morphism $R \to A$ of $\Lambda$-rings, define $\qDR(A/R)$ to be the cochain complex of $R[q]$-modules given by taking the homotopy limit of the functor \begin{align*} \Strat^q_{A/R} &\to \Ch(R[q])\\ B &\mapsto B. \end{align*} \end{definition} Equivalently, can we follow the approach of \cite{Gr,simpsonHtpy} towards the de Rham stack by regarding $\qDR(A/R)$ as the quasi-coherent cohomology complex of $(\Spec A)_{\strat}^q$. Writing $\O\co f\Lambda(R[q])\to \Mod_{R[q]}$ for the forgetful functor to the category of $R[q]$-modules, and $[f\Lambda(R[q]),\Set]$ for the category of set-valued functors on $f\Lambda(R[q])$, we have \[ \qDR(A/R)= \oR\Hom_{[f\Lambda(R[q]),\Set]}((\Spec A)_{\strat}^q,\O), \] coming from the right-derived functor of the functor $\Hom_{[f\Lambda(R[q]),\Set]}((\Spec A)_{\strat}^q,-)$ of natural transformations with source $(\Spec A)_{\strat}^q$. \begin{definition} Given a polynomial ring $R[x]$, recall from \cite{scholzeqdef} that the $q$-de Rham (or Aomoto--Jackson) cohomology $q\mbox{-}\Omega^{\bt}_{ R[x]/R}$ is given by the complex \[ R[x][q] \xra{\nabla_q} R[x][q]dx,\quad\text{ where }\quad \nabla_q(f) = \frac{f(qx) -f(x)}{x(q-1)}dx, \] so $\nabla_q(x^n)= [n]_qx^{n-1}dx$. Given a polynomial ring $R[x_1, \ldots,x_d]$, the $q$-de Rham complex $q\mbox{-}\Omega^{\bt}_{ R[x_1, \ldots,x_d]/R}$ is then set to be \[ q\mbox{-}\Omega^{\bt}_{ R[x_1]/R}\ten_{R[q]} q\mbox{-}\Omega^{\bt}_{ R[x_2]/R}\ten_{R[q]}\ldots \ten_{R[q]} q\mbox{-}\Omega^{\bt}_{ R[x_d]/R}, \] so takes the form \[ R[x_1, \ldots,x_d][q] \xra{\nabla_q} \Omega^1_{R[x_1, \ldots,x_d]/R}[q] \xra{\nabla_q} \ldots \xra{\nabla_q}\Omega^d_{R[x_1, \ldots,x_d]/R}[q]. \] \end{definition} \begin{proposition}\label{Qcalc} If $R$ is a $\Lambda$-ring and $x$ of rank $1$, then $\qDR(R[x]/R)$ can be calculated by a cosimplicial module $U^{\bt}$ given by setting $U^n$ to be the $\Lambda$-subring \[ U^n \subset R[q, \{(q^m-1)^{-1}\}_{m \ge 1},x_0, \ldots, x_n] \] generated by $q$ and the elements $x_i$ and $\frac{x_i-x_j}{q-1}$. \end{proposition} \begin{proof} For $X=\Spec R[x]$, the set-valued functor $X_{\strat}^q$ is not representable, but it can be resolved by the simplicial functor $\tilde{X}_{\strat}^q$ given by taking the \v Cech nerve of $\Hom_{\Lambda,R}(A,B)\to \Hom_{\Lambda,R}(A,B/(q-1))$, so \begin{align*} (\tilde{X}_{\strat}^q)_n(B)&:= \overbrace{\Hom_{\Lambda,R}(A,B)\by_{\Hom_{\Lambda,R}(A,B/(q-1))} \ldots \by_{\Hom_{\Lambda,R}(A,B/(q-1))} \Hom_{\Lambda,R}(A,B) }^{n+1}\\ &= \Hom_{\Lambda,R}(A, \overbrace{B\by_{B/(q-1)}\ldots \by_{B/(q-1)} B}^{n+1}). \end{align*} Observe that any element of $(X_{\strat}^q)_n(B)$ gives rise to a morphism $f \co R[q,x_0, \ldots, x_n] \to B$ of $\Lambda$-rings over $R[q]$, with the image of $x_i-x_j$ divisible by $(q-1)$. Flatness of $B$ then gives a unique element $f(x_i-x_j)/(q-1) \in B$, so we have a map $f$ to $B$ from the free $\Lambda$-ring $L$ over $R[q,x_0, \ldots, x_n]$ generated by elements $z_{ij}$ with $(q-1)z_{ij} =x_i-x_j$. Since $B$ is flat, it embeds in $B[\{(q^m-1)^{-1}\}_{m \ge 1} ]$ (the only hypothesis we really need) implying that the image of $f$ factors through the image $U^n$ of $L$ in $R[q, \{(q^m-1)^{-1}\}_{m \ge 1},x_0, \ldots, x_n]$. To see that $(X_{\strat}^q)_n$ is represented by $U^n$, we only now need to check that $U^n$ is itself flat over $R[q]$, which follows because the argument of Lemma \ref{subringbasis} gives a basis \[ x_0^{r_0}\lambda^{r_1}(\tfrac{x_1-x_0}{q-1})\cdots \lambda^{r_n}(\tfrac{x_n-x_{n-1}}{q-1}) \] for $U^n$ over $R[q]$. We therefore have $\qDR( R[x]/R)\simeq U^{\bt}$. \end{proof} \begin{theorem}\label{mainthm} If $R$ is a $\Lambda$-ring, and the polynomial ring $R[x_1, \ldots,x_d]$ is given the $\Lambda$-ring structure for which the elements $x_i$ are of rank $1$, then there are $R[q]$-linear zigzags of quasi-isomorphisms \begin{align*} \qDR( R[x_1, \ldots,x_n]/R) &\simeq (\Omega^*_{R[x_1, \ldots,x_n]/R}[q], (q-1)\nabla_q)\\ \oL\eta_{(q-1)} \qDR( R[x_1, \ldots,x_n]/R) &\simeq q\mbox{-}\Omega^{\bt}_{R[x_1, \ldots,x_n]/R}, \end{align*} where $\oL\eta_{(q-1)}$ denotes derived d\'ecalage with respect to the $(q-1)$-adic filtration. \end{theorem} \begin{proof} It suffices to prove the first statement, the second following immediately by d\'ecalage. Since $(\Spec A\ten_RA')_{\strat}^q(B)= (\Spec A)_{\strat}^q(B)\by (\Spec A')_{\strat}^q(B)$, and coproduct of flat $\Lambda$-rings over $R[q]$ is given by $\ten_{R[q]}$, we have $ \qDR((A\ten_RA')/R) \simeq \qDR(A/R)\ten_{R[q]}^{\oL} \qDR(A'/R)$, so we may reduce to the case $A=R[x]$. Proposition \ref{Qcalc} gives $\qDR( R[x]/R)\simeq U^{\bt}$, and in order to compare this with $q$-de Rham cohomology, we now consider the cochain complexes $\tilde{\Omega}^{\bt}(U^n)$ given by \[ U^n \xra{(q-1)\nabla_q} \bigoplus_i U^ndx_i \xra{(q-1)\nabla_q}\bigoplus_{i<j} U^ndx_i\wedge dx_j \xra{(q-1)\nabla_q}\ldots . \] In order to see that this differential is well-defined, observe that \begin{align*} (q-1)\nabla_{q,y}\lambda^k(\tfrac{y-x}{q-1})&= y^{-1}(\lambda^k(\tfrac{qy-x}{q-1}) - \lambda^k(\tfrac{y-x}{q-1}))dy\\ &= y^{-1}(\lambda^k(y+ \tfrac{y-x}{q-1}) - \lambda^k(\tfrac{y-x}{q-1}))dy\\ &= \lambda^{k-1}(\tfrac{y-x}{q-1})dy, \end{align*} and similarly \[ (q-1)\nabla_{q,x}\lambda^k(\tfrac{y-x}{q-1}) = \sum_{i \ge 1} (-1)^ix^{i-1}\lambda^{k-i}(\tfrac{y-x}{q-1})dx. \] The first calculation also shows that the inclusion $\tilde{\Omega}^{\bt}(U^{n-1})\into \tilde{\Omega}^{\bt}(U^n)$ is a quasi-isomorphism, since \[ (q-1)\nabla_{q,x_n}(f(x_0, \ldots, x_{n-1})\lambda^k(\tfrac{x_n-x_{n-1}}{q-1}))= f(x_0, \ldots, x_{n-1})\lambda^{k-1}(\tfrac{x_n-x_{n-1}}{q-1})dx_n. \] By induction on $n$ we deduce that the inclusion $\tilde{\Omega}^{\bt}(U^0)\into \tilde{\Omega}^{\bt}(U^n)$, and hence the retraction of it given by the diagonal, is a quasi-isomorphism. Now the complexes $\tilde{\Omega}^i(U^{\bt})$ are all acyclic for $i>0$, consisting of cosimplicial tensor products of $U^{\bt}$ with cosimplicial symmetric powers of the acyclic complex given by $\Z dx_0 \oplus \cdots \oplus \Z dx_n$ in level $n$. We therefore have quasi-isomorphisms \[ U^{\bt} \la \Tot \tilde{\Omega}^{\bt}(U^{\bt}) \to \tilde{\Omega}^{\bt}(U^0) \] of flat cochain complexes over $R[q]$, so \[ \qDR( R[x]/R) \simeq \tilde{\Omega}^{\bt}(R[x]), \] and we just observe that $\eta_{(q-1)}\tilde{\Omega}^{\bt}(R[x]) =(\Omega^*_{R[x]/R}[q], (q-1)\nabla_q)$. \end{proof} \begin{remark} Note that Theorem \ref{mainthm} implies that $q\mbox{-}\Omega^{\bt}_{R[x_1, \ldots,x_n]/R}$ naturally underlies the d\'ecalage of a cosimplicial $\Lambda$-ring over $R[q]$. Even the underlying cosimplicial commutative ring structure carries more information than an $E_{\infty}$-structure when $\Q \nsubseteq R$. \end{remark} \begin{remark}\label{crisrmk} The complex $ (\Omega^*_{R[x_1, \ldots,x_n]/R}[q], (q-1)\nabla_q)$ is a more fundamental object than its d\'ecalage $q\mbox{-}\Omega^{\bt}_{R[x_1, \ldots,x_n]/R}$, since it has a vestigial memory of the Hodge filtration. There might be a natural formulation of the theorem not involving d\'ecalage, in terms of a $q$-analogue of the crystalline site for a $\Lambda$-ring $A$ over $R$, regarded as an $R[q]$-algebra via $R=R[q]/(q-1)$. Following Remark \ref{DPrmk}, this would involve extensions $B \to A$ of $\Lambda$-rings over $R[q]$ equipped with $q$-analogues of divided power operations on the augmentation ideals $I$, looking like $ x \mapsto (q-1)^k \lambda^k(\tfrac{x}{q-1}). $ \end{remark} \subsection{Completed $q$-cohomology} \begin{definition} Given a morphism $R \to A$ of $\Lambda$-rings, we define the category $\hat{\Strat}^q_{A/R} \subset \Strat^q_{A/R}$ to consist of those objects which are $(q-1)$-adically complete. \end{definition} \begin{definition} Given a flat morphism $R \to A$ of $\Lambda$-rings, define $\hat{\qDR}(A/R)$ to be the cochain complex of $R\llbracket q-1 \rrbracket$-modules given by taking the homotopy limit of the functor \begin{align*} \hat{\Strat}^q_{A/R} &\to \Ch(R\llbracket q \rrbracket)\\ B &\mapsto B. \end{align*} \end{definition} The following is immediate: \begin{lemma} Given a flat morphism $R \to A$ of $\Lambda$-rings, the complex $\hat{\qDR}(A/R)$ is the derived $(q-1)$-adic completion of $\qDR(A/R)$. \end{lemma} \begin{definition}\label{qhatdef2} As in \cite[\S 3]{scholzeqdef}, given a formally \'etale map $\boxempty\co R[x_1, \ldots,x_d] \to A$, define $\widehat{q\mbox{-}\Omega}^{\bt}_{A/R, \boxempty}$ to be the complex \[ A\llbracket q-1\rrbracket \xra{\nabla_q} \Omega^1_{A/R}\llbracket q-1\rrbracket\xra{\nabla_q} \ldots \xra{\nabla_q} \Omega^d_{A/R}\llbracket q-1\rrbracket, \] where $\nabla_q$ is defined as follows. First note that the ring endomorphisms $\gamma_i$ of $R[x_1, \ldots,x_d]\llbracket q-1\rrbracket$ given by $\gamma_i(x_j)= q^{\delta_{ij}}x_j$ extend uniquely to endomorphisms of $A\llbracket q-1\rrbracket$ which are the identity modulo $(q-1)$, then set \[ \nabla_q(f):= \sum_i \frac{\gamma_i(f)-f}{(q-1)x_i}dx_i. \] \end{definition} Note that $\widehat{q\mbox{-}\Omega}^{\bt}_{ R[x_1, \ldots,x_d]/R}$ is just the $(q-1)$-adic completion of $ q\mbox{-}\Omega^{\bt}_{ R[x_1, \ldots,x_d]/R}$. \begin{proposition}\label{mainthm2} If $R$ is a flat $\Lambda$-ring over $\Z$ and $\boxempty\co R[x_1, \ldots,x_d]\to A$ is a formally \'etale map of $\Lambda$-rings, the elements $x_i$ having rank $1$, then there are zigzags of $R\llbracket q\rrbracket$-linear quasi-isomorphisms \begin{align*} \widehat{\qDR}( A/R) &\simeq (\Omega^*_{A/R}\llbracket q -1\rrbracket, (q-1)\nabla_q), & \oL\eta_{(q-1)} \widehat{\qDR}( A/R) &\simeq \widehat{q\mbox{-}\Omega}^{\bt}_{A/R, \boxempty}. \end{align*} The induced quasi-isomorphisms \begin{align*} \widehat{\qDR}( A/R)\ten^{\oL}_{R\llbracket q-1\rrbracket}R &\simeq (\Omega^*_{A/R}, 0), & (\oL\eta_{(q-1)} \widehat{\qDR}( A/R))\ten^{\oL}_{R\llbracket q-1\rrbracket}R &\simeq \Omega^{\bt}_{A/R} \end{align*} are independent of the choice of framing. \end{proposition} \begin{proof} This is much the same as the proof of Theorem \ref{mainthm}. The complex $\widehat{\qDR}( A/R)$ can be realised as a cosimplicial $\Lambda$-ring $U$, with $U^n$ the $(q-1)$-adically complete $\Lambda$-subring of $A^{\ten_R(n+1)}\llbracket q-1 \rrbracket [ \{(q^m-1)^{-1}\}_{m \ge 1}]$ generated by \[ A^{\ten_R(n+1)}\llbracket q-1 \rrbracket \quad \text{ and }\quad (q-1)^{-1}\ker( A^{\ten_R(n+1)}\to A)\llbracket q-1 \rrbracket. \] Uniqueness of lifts with respect to the formally \'etale framing ensures that the endomorphisms $\gamma_i$ commute with the Adams operations, so are $\Lambda$-ring endomorphisms of $R$. Since the formal completion of $A\ten_RA\to A$ is just the $\Lambda$-ring \[ A\llbracket (x_1-y_1), (x_2-y_2), \ldots, (x_d-y_d)\rrbracket, \] the calculations of Theorem \ref{mainthm} now adapt to give quasi-isomorphisms \[ (\Omega^*_{A/R}\llbracket q \rrbracket, (q-1)\nabla_q) \la \Tot \tilde{\Omega}^{\bt}(U^{\bt}) \to U^{\bt}, \] where $\tilde{\Omega}^{\bt}(U^n)$ is the $(q-1)$-adic completion of $ (U^n\ten_{A^{\ten(n+1)}}(\Omega^*_{A/R})^{\ten(n+1)},(q-1)\nabla_q) ). $ Reduction of this or its d\'ecalage modulo $(q-1)$ replaces $\nabla_q$ with $d$, removing the dependence on co-ordinates. \end{proof} \begin{remark}\label{qconnrmk} As in \cite[Definition 7.3]{scholzeqdef}, there is a notion of $q$-connections on projective $A\llbracket q-1 \rrbracket$-modules $M$. Adapting the ideas of Proposition \ref{mainthm2}, these will be equivalent to projective modules over $X_{\strat}^q$, so flat Cartesian $\widehat{\qDR}( A/R) $-modules $N$ with $N\ten_{\widehat{\qDR}( A/R)}A\llbracket q-1 \rrbracket=M$, together with a condition that the $(\widehat{\qDR}( A/R)/(q-1))$-module $N/(q-1)$ is just given by pullback of the $A$-module $M/(q-1)$. Via Lemma \ref{subringbasis}, these data are equivalent to specifying an operator $\pd^1 \co M \to \hat{\bigoplus}_{k_1, \ldots,k_d} M\lambda^{k_1}(\frac{y_1-x_1}{q-1}) \cdots \lambda^{k_d}(\frac{y_d-x_d}{q-1})$ satisfying a cocycle condition and congruent to the identity modulo $(q-1)$. Such operators then arise from $q$-connections $(\nabla_{1,q}, \ldots, \nabla_{d,q})$ as $q$-Taylor series \[ \pd^1(f):= \sum_{k_1, \ldots,k_d} (q-1)^{\sum k_i} (\nabla_{1,q}^{k_1}\ldots \nabla_{d,q}^{k_d})(f)\, \lambda^{k_1}(\tfrac{y_1-x_1}{q-1}) \cdots \lambda^{k_d}(\tfrac{y_d-x_d}{q-1}). \] \end{remark} \section{Comparisons for $\Lambda_P$-rings} Since very few \'etale maps $R[x_1, \ldots,x_d]\to A$ give rise to $\Lambda$-ring structures on $A$, Proposition \ref{mainthm2} is fairly limited in its scope for applications. We now show how the construction of $\hat{\qDR}$ and the comparison quasi-isomorphism survive when we weaken the $\Lambda$-ring structure by discarding Adams operations at invertible primes. \subsection{$q$-cohomology for $\Lambda_P$-rings} Our earlier constructions for $\Lambda$-rings all carry over to $\Lambda_P$-rings, as follows. \begin{definition} Given a set $P$ of primes, we define a $\Lambda_P$-ring $A$ to be a $\Lambda_{\Z,P}$-ring in the sense of \cite{borgerBasicGeomI}. This means that it is a coalgebra in commutative rings for the comonad given by the functor $W^{(P)}$ of $P$-typical Witt vectors. When a commutative ring $A$ is flat over $\Z$, giving a $\Lambda_P$-ring structure on $A$ is equivalent to giving commuting Adams operations $\Psi^p$ for all $p \in P$, with $\Psi^p(a) \equiv a^p \mod p$ for all $a$. \end{definition} Thus when $P$ is the set of all primes, a $\Lambda_P$-ring is just a $\Lambda$-ring; a $\Lambda_{\emptyset}$-ring is just a commutative ring; for a single prime $p$, we write $\Lambda_p:= \Lambda_{\{p\}}$, and note that a $\Lambda_p$-ring is a $\delta$-ring in the sense of \cite{joyaldeltaWitt}. \begin{definition} Given a $\Lambda_P$-ring $R$, say that $A$ is a $\Lambda_P$-ring over $R$ if it is a $\Lambda_P$-ring equipped with a morphism $R \to A$ of $\Lambda_P$-rings. We say that $A$ is a flat $\Lambda_P$-ring over $R$ if $A$ is flat as a module over the commutative ring underlying $R$. \end{definition} \begin{definition} Given a morphism $R \to A$ of $\Lambda_P$-rings, we define the category $\Strat^{q,P}_{A/R}$ to consist of flat $\Lambda_P$-rings $B$ over $R[q]$ equipped with a compatible morphism $A \to B/(q-1)$, such that the map $A \to B/(q-1)$ admits a lift to $B$. We define the category $\hat{\Strat}^{q,P}_{A/R} \subset \Strat^q_{A/R}$ to consist of those objects which are $(q-1)$-adically complete. \end{definition} More concisely, $\Strat^{q,P}_{A/R}$ (resp. $\hat{\Strat}^{q,P}_{A/R}$) is the Grothendieck construction of the functor \[ (\Spec A)_{\strat}^{q,P}\co B \mapsto \im(\Hom_{\Lambda_P,R}(A,B)\to \Hom_{\Lambda_P,R}(A,B/(q-1))) \] of the category of flat $\Lambda_P$-rings (resp. $(q-1)$-adically complete flat $\Lambda_P$-rings) over $R[q]$. \begin{definition} Given a flat morphism $R \to A$ of $\Lambda_P$-rings, define $\qDR_P(A/R)$ to be the cochain complex of $R[q]$-modules given by taking the homotopy limit of the functor \begin{align*} \Strat^{q,P}_{A/R} &\to \Ch(R[q])\\ B &\mapsto B. \end{align*} Define $\hat{\qDR}_P(A/R)$ to be the cochain complex of $R\llbracket q-1 \rrbracket$-modules given by the corresponding homotopy limit over $\hat{\Strat}^{q,P}_{A/R}$. \end{definition} Thus when $P$ is the set of all primes, we have $\qDR_P(A/R)= \qDR(A/R)$. At the other extreme, for $A$ smooth, $\hat{\qDR}_{\emptyset}(A/R)$ is the Rees construction of the Hodge filtration on the infinitesimal cohomology complex of $A$ over $R$ , with formal variable $(q-1)$. In more detail, there is a decreasing filtration $F$ of $\O_{\inf}$ given by powers of the augmentation ideal $\O_{\inf} \to \O_{\Zar}$, and $\hat{\qDR}_{\emptyset}(A/R) \simeq \bigoplus_{\nu\in \Z} (q-1)^{-\nu}\oR\Gamma(\Spec A, F^{\nu}\O_{\inf})(q-1)^{-\nu}$. \begin{lemma} For a set $P$ of primes, the forgetful functor from $\Lambda$-rings to $\Lambda_P$-rings has a right adjoint $W^{(\notin P)}$. There is a canonical ghost component morphism \[ W^{(\notin P)}(B) \to \prod_{\substack{n \in \N:\\ (n,p)=1 ~\forall p \in P}} B, \] which is an isomorphism when $P$ contains all the residue characteristics of $B$. \end{lemma} \begin{proof} Existence of a right adjoint follows from the comonadic definitions of $\Lambda$-rings and $\Lambda_P$-rings. The ghost component morphism is given by taking the Adams operations $\Psi^n$ given by the $\Lambda$-ring structure on $W^{(\notin P)}(B)$, followed by projection to $B$. When $P$ contains all the residue characteristics of $B$, a $\Lambda$-ring structure is the same as a $\Lambda_P$-ring structure with compatible commuting Adams operations for all primes not in $P$, leading to the description above. \end{proof} Note that the big Witt vector functor $W$ on commutative rings thus factorises as $W= W^{(\notin P)} \circ W^{(P)}$, for $W^{(P)}$ the $P$-typical Witt vectors. \begin{proposition}\label{Ponlyprop} Given a morphism $R \to A$ of $\Lambda$-rings, and a set $P$ of primes, there are natural maps \begin{align*} \qDR_P(A/R) &\to \qDR(A/R), & \hat{\qDR}_P(A/R) &\to \hat{\qDR}(A/R), \end{align*} and the latter map is a quasi-isomorphism when $P$ contains all the residue characteristics of $A$. \end{proposition} \begin{proof} We have functors \begin{align*} (\Spec A)_{\strat}^q\circ W^{(\notin P)} \co B &\mapsto \im(\Hom_{\Lambda,R}(A,W^{(\notin P)} B)\to \Hom_{\Lambda,R}(A,(W^{(\notin P)} B)/(q-1)))\\ (\Spec A)_{\strat}^{q,P} \co B &\mapsto \im(\Hom_{\Lambda_P,R}(A,B)\to \Hom_{\Lambda_P,R}(A,B/(q-1))) \end{align*} on the category of flat $\Lambda_P$-rings over $R[q]$. There is an obvious map \[ (W^{(\notin P)} B)/(q-1) \to W^{(\notin P)}(B/(q-1)), \] and hence a natural transformation $(\Spec A)_{\strat}^q\circ W^{(\notin P)} \to (\Spec A)_{\strat}^{q,P}$, which induces the morphism $\qDR_P(A/R) \to \qDR(A/R)$ on cohomology. When $P$ contains all the residue characteristics of $A$, the map $ (W^{(\notin P)} B)/(q-1) \to W^{(\notin P)}(B/(q-1))$ is just \[ \prod_{\substack{n \in \N:\\ (n,p)=1 ~\forall p \in P}} B/(q^n-1) \to \prod_{\substack{n \in \N:\\ (n,p)=1 ~\forall p \in P}} B/(q-1), \] since the morphism $R[q] \to W^{(\notin P)} B$ is given by Adams operations, with $\Psi^n(q-1)= q^n-1$. We have $(q^n-1)=(q-1)[n]_q$, and $[n]_q$ is a unit in $\Z[\frac{1}{n}]\llbracket q-1 \rrbracket$, hence a unit in $B$ when $n$ is coprime to the residue characteristics. Thus the map $ (W^{(\notin P)} B)/(q-1) \to W^{(\notin P)}(B/(q-1))$ gives an isomorphism whenever $B$ is $(q-1)$-adically complete and admits a map from $A$, so the transformation $(\Spec A)_{\strat}^q\circ W^{(\notin P)} \to (\Spec A)_{\strat}^{q,P}$ is a natural isomorphism on the category of flat $(q-1)$-adically complete $\Lambda_P$-rings over $R[q]$, and so $ \hat{\qDR}_P(A/R) \xra{\simeq} \hat{\qDR}(A/R) $. \end{proof} Over $\Z[\frac{1}{P}]$, every $\Lambda_P$-ring can be canonically made into a $\Lambda$-ring, by setting all the additional Adams operations to be the identity. However, this observation is of limited use in establishing functoriality of $q$-de Rham cohomology, because the resulting $\Lambda$-ring structure will not satisfy the conditions of Proposition \ref{mainthm2}. We now give a more general result which does allow for meaningful comparisons. \begin{theorem}\label{mainthm3} If $R$ is a flat $\Lambda_P$-ring over $\Z$ and $\boxempty\co R[x_1, \ldots,x_d]\to A$ is a formally \'etale map of $\Lambda_P$-rings, the elements $x_i$ having rank $1$, then there is a zigzag of $R\llbracket q\rrbracket$-linear quasi-isomorphisms \[ \oL\eta_{(q-1)} \widehat{\qDR}_P( A/R) \simeq \widehat{q\mbox{-}\Omega}^{\bt}_{A/R, \boxempty} \] whenever $P$ contains all the residue characteristics of $A$. \end{theorem} \begin{proof} The key observation to make is that formally \'etale maps have a unique lifting property with respect to nilpotent extensions of flat $\Lambda_P$-rings, because the Adams operations must also lift uniquely. In particular, this means that the operations $\gamma_i$ featuring in the definition of $q$-de Rham cohomology are necessarily endomorphisms of $A$ as a $\Lambda_P$-ring. Similarly to Proposition \ref{mainthm2}, $\widehat{\qDR}_P( A/R)$ is calculated using a cosimplicial $\Lambda_P$-ring given in level $n$ by the $(q-1)$-adic completion $\hat{U}_{P,A}^{\bt}$ of the $\Lambda_P$-ring over $R[q]$ generated by $A^{\ten_R(n+1)}[q]$ and $(q-1)^{-1}\ker( A^{\ten_R(n+1)}\to A)[q]$. The observation above shows that $\hat{U}_{P,A}^n \cong \hat{U}_{P,R[x_1, \ldots,x_d]}^n\hten_{ R[x_1, \ldots,x_d]}A$, changing base along $\boxempty$ applied to the first factor. As in Proposition \ref{Ponlyprop}, $\hat{U}_{P,R[x_1, \ldots,x_d]}^{\bt}$ is just the $(q-1)$-adic completion of the complex $U^{\bt}$ from Proposition \ref{Qcalc}. Further application of the key observation above then allows us to adapt the constructions of Theorem \ref{mainthm}, giving the desired quasi-isomorphisms. \end{proof} \subsection{Cartier isomorphisms in mixed characteristic}\label{cartiersn} The only setting in which Theorem \ref{mainthm3} leads to results close to the conjectures of \cite{scholzeqdef} is when $R=W^{(p)}(k)$, the $p$-typical Witt vectors of a field of characteristic $p$, and $A=\Lim_n A_n$ is a formal deformation of a smooth $k$-algebra $A_0$. Then any formally \'etale morphism $W^{(p)}(k)[x_1, \ldots,x_d]\to A$ gives rise to a unique compatible lift $\Psi$ of absolute Frobenius on $A$ with $\Psi(x_i)=x_i^p$, so gives $A$ the structure of a topological $\Lambda_p$-ring. The framing still affects the choice of $\Lambda_p$-ring structure, but at least such a structure is guaranteed to exist, giving rise to a complex $\qDR_P(A/R)^{\wedge_p}:= \Lim_n \qDR_p(A/R)\ten_RR_n$ depending only on the choice of $\Psi$, where $R_n=W_n^{(p)}(k)$. Our constructions now allow us to globalise the quasi-isomorphism \[ (\widehat{q\mbox{-}\Omega}^{\bt}_{A/R, \boxempty})^{\wedge_p}/[p]_q \simeq (\Omega^*_{A/R})^{\wedge_p}\llbracket q-1 \rrbracket/[p]_q \] of \cite[Proposition 3.4]{scholzeqdef}, where $\Omega^*_{A/R}$ denotes the complex $A \xra{0} \Omega^1_{A/R} \xra{0} \Omega^2_{A/R} \xra{0}\ldots $. \begin{proposition}\label{cartierprop} Take a smooth formal scheme $\fX$ over $R=W^{(p)}(k)$ equipped with a lift $\Psi$ of Frobenius which \'etale locally admits co-ordinates $\{x_i\}_i$ as above with $\Psi(x_i)=x_i^p$. Then there is a global quasi-isomorphism \[ C_q^{-1}\co (\Omega^*_{\fX/R})^{\wedge_p}\llbracket q-1 \rrbracket/[p]_q\to (\oL\eta_{(q-1)}\hat{\qDR}_p(\sO_{\fX}/R))^{\wedge_p}/[p]_q \] in the derived category of \'etale sheaves on $\fX$. \end{proposition} \begin{proof} Functoriality of the construction $\qDR_p$ for rings with Frobenius lifts gives us a sheaf $\hat{\qDR}_p(\sO_{\fX}/R)^{\wedge_p}$ on $\fX$. We then have maps \begin{align*} \Psi^p \co \qDR_p(\sO_{\fX}/R)^{\wedge_p}&\to \qDR_p(\sO_{\fX}/R)^{\wedge_p}\\ \qDR_p(\sO_{\fX}/R)^{\wedge_p}/(q-1)&\to \qDR_p(\sO_{\fX}/R)^{\wedge_p}/(q^p-1), \end{align*} and thus, denoting good truncation by $\tau$, \[ (q-1)^i\Psi^p \co \tau^{\le i}(\qDR_p(\sO_{\fX}/R)^{\wedge_p}/(q-1))\to (\oL\eta_{(q-1)}\hat{\qDR}_p(\sO_{\fX}/R)^{\wedge_p})/[p]_q \] the left-hand side is quasi-isomorphic to $\bigoplus_{j \le i} (\Omega^j_{\sO_{\fX}/R})^{\wedge_p}[-j]$ by Proposition \ref{mainthm2}. Extending the construction $R[q]$-linearly and restricting to top summands therefore gives us the global map $C_q^{-1}$. For a local choice of framing, the map $\Psi^p$ necessarily corresponds via Theorem \ref{mainthm3} to the chain map $a dx^I \mapsto \Psi^p(a) x^{I(p-1)}dx^I$ on the complex $(\Omega^*_{A/R}\llbracket q-1 \rrbracket, (q-1)\nabla_q)$. This gives equivalences \[ (q-1)^i\Psi^p \simeq \sum_{j \le i} (q-1)^{i-j} (\tilde{C}^{-1})^j \] for Scholze's locally defined lifts $(\tilde{C}^{-1})^j\co (\Omega^j_{A/R})^{\wedge_p}[-j] \to (\widehat{q\mbox{-}\Omega}^{\bt}_{A/R, \boxempty})^{\wedge_p}/[p]_q$ of the Cartier quasi-isomorphism. The local calculation of \cite[Proposition 3.4]{scholzeqdef} then ensures that $C_q^{-1}$ is a quasi-isomorphism. \end{proof} \section{Functoriality via analogues of de Rham--Witt cohomology}\label{DRWsn} In order to obtain a cohomology theory for smooth commutative rings rather than for $\Lambda_P$-rings, we now consider $q$-analogues of de Rham--Witt cohomology. Our starting point is to observe that if we allow roots of $q$, we can extend the Jackson differential to fractional powers of $x$ by the formula \[ \nabla_q (x^{m/n}) = \frac{q^{m/n} -1}{q-1} x^{m/n}d\log x, \] so terms such as $[n]_{q^{1/n}}x^{m/n}$ have integral derivative, where $[n]_{q^{1/n}}= \frac{q-1}{q^{1/n}-1}$. \subsection{Motivation}\label{motivnsn} \begin{definition} Given a $\Lambda_P$-ring $B$, define $\Psi^{1/P^{\infty}}B$ to be the smallest $\Lambda_P$-ring containing $B$ on which the Adams operations are automorphisms. \end{definition} In the case $P= \{p\}$, the $\Lambda_p$-ring $\Psi^{1/p^{\infty}}B$ is thus the colimit of the diagram \[ B \xra{\Psi^{p}} B \xra{\Psi^{p}} B \xra{\Psi^{p}} \ldots. \] The proof of Theorem \ref{mainthm3} allows us to replace $\widehat{\qDR}_p( A/R)$ with the complex $(\Omega^*_{A/R}\llbracket q-1 \rrbracket, (q-1)\nabla_q)$; under this quasi-isomorphism, the Adams operations on $A$ extend to $\Omega^*_{A/R}\llbracket q-1 \rrbracket$ by setting $\Psi^n(dx_i):= x_i^{n-1}dx_i$. As an immediate consequence we have: \begin{lemma}\label{qDRWlemma1} If $R$ is a flat $\Lambda_p$-ring over $\Z$ with $\Psi^{1/p^{\infty}}R=R$ and residue characteristic $p$, then $\Psi^{1/p^{\infty}}\qDR_p( R[x]/R) \simeq (\Omega^*_{R[x^{1/p^{\infty}}]/R}[q^{1/p^{\infty}}], (q-1)\nabla_q)$, so the d\'ecalage $\oL\eta_{(q-1)}\Psi^{1/p^{\infty}}\widehat{\qDR}_p( R[x]/R)$ is quasi-isomorphic to the $(q-1)$-adic completion of the complex \begin{align*} & \{ a \in R[x^{1/p^{\infty}}, q^{1/p^{\infty}}] ~:~ \nabla_qa \in R[x^{1/p^{\infty}}, q^{1/p^{\infty}}]d\log x\} \xra{\nabla_q}\\ &\{ b\, d\log x \in R[x^{1/p^{\infty}}, q^{1/p^{\infty}}]d\log x ~:~ b(0,q)=0\}. \end{align*} \end{lemma} Thus in level $0$ (resp.\ level $1$), $\oL\eta_{(q-1)}\Psi^{1/p^{\infty}}\widehat{\qDR}( R[x]/R)$ is spanned by elements of the form $[p^n]_{q^{1/p^n}}x^{m/p^n}$ (resp.\ $x^{m/p^n} d\log x$), so setting $q^{1/p^{\infty}}=1$ gives a complex whose $p$-adic completion is the $p$-typical de Rham--Witt complex. \begin{lemma}\label{perfectioncoholemma} Let $R$ and $A$ be flat $p$-adically complete $\Lambda_p$-algebras over $\Z_p$, with $\Psi^{1/p^{\infty}}R=R$ and, for elements $x_i$ of rank $1$, a map $\boxempty\co R[x_1, \ldots,x_d]^{\wedge_p}\to A$ of $\Lambda_p$-rings which is a flat $p$-adic deformation of an \'etale map. Then the map \[ (R[q^{1/p^{\infty}}]\ten_{R[q]} \oL\eta_{(q-1)} \widehat{\qDR}_p( A/R))^{\wedge_p} \to \oL\eta_{(q-1)} (\Psi^{1/p^{\infty}}\widehat{\qDR}_p (A/R))^{\wedge_p} \] is a quasi-isomorphism. \end{lemma} \begin{proof} The map $\Psi^p \co A\ten_{R[x_1, \ldots,x_d]}R[x_1^{1/p}, \ldots,x_d^{1/p}] \to A$ becomes an isomorphism on $p$-adic completion, because $\boxempty$ is flat and we have an isomorphism modulo $p$. Thus \[ \Psi^{1/p^{\infty}}A \cong A[x_1^{1/p^{\infty}}, \ldots,x_d^{1/p^{\infty}}]^{\wedge_p}:= (A\ten_{R[x_1, \ldots,x_d]}R[x_1^{1/p^{\infty}}, \ldots,x_d^{1/p^{\infty}}])^{\wedge_p} \] Combined with the calculation of Lemma \ref{qDRWlemma1}, this gives us a quasi-isomorphism between $(\Psi^{1/p^{\infty}}\widehat{\qDR}_p(A/R))^{\wedge_p}$ and the $(p,q-1)$-adic completion of \[ \bigoplus_{\alpha} A\llbracket q-1\rrbracket x_1^{\alpha_1}\ldots x_d^{\alpha_d} dx^I, \] where $\alpha \in p^{-\infty}\Z^d$ such that $0 \le \alpha_i <1$ if $i \notin I$ and $-1< \alpha_i \le 0$ if $i \in I$. We then observe that the contributions to the d\'ecalage $\eta_{(q-1)}$ from terms with $\alpha \ne 0$ must be acyclic, via a contracting homotopy defined by the restriction to $\eta_{(q-1)}$ of the $q$-integration map \[ f x_1^{\alpha_1}\ldots x_d^{\alpha_d} dx^I \mapsto f x_1^{\alpha_1}\ldots x_d^{\alpha_d}\sum_{i \in I} \pm x_i [\alpha_i]_q^{-1} dx^{(I\setminus i)}, \] where $[\frac{m}{p^n}]_q^{-1}= [m]_{q^{1/p^n}}^{-1}[p^n]_{q^{1/p^n}}$ for $m$ coprime to $p$, noting that $[m]_{q^{1/p^n}}$ is a unit in $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$. \end{proof} \begin{remark}\label{Wittrmk} The endomorphism given on $\Psi^{1/P^{\infty}}\hat{\qDR}_P(A/R)$ by \[ a \mapsto \Psi^{1/n}([n]_qa)= [n]_{q^{1/n}}\Psi^{1/n}a \] descends to an endomorphism of $\H^0(\Psi^{1/P^{\infty}}\hat{\qDR}_P(A/R)/(q-1))$, which we may denote by $V_n$ because it mimics Verschiebung in the sense that $\Psi^{n}V_n= n\cdot\id$. For $A$ smooth over $\Z$, we then have \begin{align*} \H^0(\Psi^{1/P^{\infty}}\hat{\qDR}_P(A/\Z)/(q-1))/(V_p~:~p \in P) &\cong A[q^{1/P^{\infty}}]/([p]_{q^{1/p}}~:~p \in P)\\ &\cong A[\zeta_{P^{\infty}}], \end{align*} for $\zeta_n$ a primitive $n$th root of unity. By adjunction, this gives an injective map \[ \H^0(\Psi^{1/P^{\infty}}\hat{\qDR}_P(A/\Z)/(q-1)) \into W^{(P)}A[\zeta_{P^{\infty}}] \] of $\Lambda_P$-rings, which becomes an isomorphism on completing $\Psi^{1/P^{\infty}}\hat{\qDR}(A/\Z)$ with respect to the system $\{([n]_{q^{1/n}})\}_{n\in P^{\infty}}$ of ideals, where we write $P^{\infty}$ for the set of integers whose prime factors are all in $P$. This implies that the cokernel is annihilated by all elements of $(q^{1/P^{\infty}}-1)$, so leads us to consider almost mathematics as in \cite{gabberalmost}. \end{remark} \subsection{Almost isomorphisms} Combined with Lemma \ref{perfectioncoholemma}, Remark \ref{Wittrmk} allows us to regard $\oL\eta_{(q-1)}\Psi^{1/p^{\infty}}\hat{\qDR}_p(A/\Z_p)^{\wedge_p}$ as being almost a $q^{1/p^{\infty}}$-analogue of $p$-typical de Rham--Witt cohomology. (From now on, we consider only the case $P=\{p\}$.) The ideal $(q^{1/p^{\infty}}-1)^{\wedge_{(p,q-1)}}= \ker(\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}} \to \Z_p)$ is equal to the $p$-adic completion of its square, since we may write it as the kernel $W^{(p)}(\m)$ of $W^{(p)}(\bF_p[q^{1/p^{\infty}}]^{\wedge_{(q-1)}})\to W^{(p)}(\bF_p)$, for the idempotent maximal ideal $\m = ( (q-1)^{1/p^{\infty}})^{\wedge_{(q-1)}}$ in $\bF_p[q^{1/p^{\infty}}]^{\wedge_{(q-1)}}$. If we set $h^{1/p^n}$ to be the Teichm\"uller element \[ [q^{1/p^n}-1] = \lim_{r \to \infty} (q^{1/p^{nr}}-1)^{p^r} \in \Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}, \] then $W^{(p)}(\m)=(h^{1/p^{\infty}})^{\wedge_{(p,h)}}$. Although $W^{(p)}(\m)/p^n$ is not maximal in $\Z[h^{1/p^{\infty}}]^{\wedge_{(h)}}/p^n$, it is idempotent and flat, so gives a basic setup in the sense of \cite[2.1.1]{gabberalmost}. We thus regard the pair $(\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}},W{(p)}(\m) )$ as an inverse system of basic setups for almost ring theory. We then follow the terminology and notation of \cite{gabberalmost}, studying $p$-adically complete $(\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}})^a$-modules (almost $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$-modules) given by localising at almost isomorphisms, the maps whose kernel and cokernel are $W^{(p)}(\m)$-torsion. The obvious functor $(-)^a$ from modules to almost modules has a right adjoint $(-)_*$, given by $N_*:=\Hom_{\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}}(W^{(p)}(\m) ,N)$, the module of almost elements. Since the counit $(M_*)^a\to M$ of the adjunction is an (almost) isomorphism, we may also regard almost modules as a full subcategory of the category of modules, consisting of those $M$ for which the natural map $M \to (M^a)_*$ is an isomorphism. We can define $p$-adically complete $(\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}})^a$-algebras similarly, forming a full subcategory of $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$-algebras. \begin{lemma}\label{recoveralmost} For any $\Z\llbracket q-1\rrbracket$-module $M$, we may recover the $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$-module $(M\ten_{\Z[q]}\Z[q^{1/p^{\infty}}])^{\wedge_p}$ as the module of almost elements of the associated almost $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$-module. \end{lemma} \begin{proof} Since $M\ten_{\Z[q]}\Z[q^{1/p^{\infty}}]= \bigoplus_{\alpha} M\ten q^{\alpha}$ for $\alpha \in p^{-\infty}\Z$ with $0 \le \alpha <1$, calculation shows that $(M\ten_{\Z[q]}\Z[q^{1/p^{\infty}}])^{\wedge_p}\to (M\ten_{\Z[q]}\Z[q^{1/p^{\infty}}])^{\wedge_p}_*$ must be an isomorphism. \end{proof} \subsection{Perfectoid algebras} We now relate Scholze's perfectoid algebras to a class of $\Lambda_p$-rings. \begin{definition}\label{perfectoidLambdapdef} Define a perfectoid $\Lambda_p$-ring to be a flat $p$-adically complete $\Lambda_p$-algebra over $\Z_p$, on which the Adams operation $\Psi^p$ is an isomorphism. \end{definition} For a perfectoid field $K$ in the sense of \cite{scholzePerfectoidSpaces}, there is a tilt $K^{\flat}$ (a complete perfect field of characteristic $p$). The subring of power-bounded elements is denoted $K^o \subset K$. \begin{lemma}\label{cfperfectoid} Given a perfectoid field $K$, we have equivalences \[ \xymatrix{ \text{perfectoid almost }K^o\text{-algebras} \ar@<1ex>[d]^{\sA_{\inf}} \\ \text{perfectoid almost }\Lambda_p\text{-rings over }\sA_{\inf}(K^o) \ar@<1ex>[u]^{K^{oa}\ten_{(\sA_{\inf}(K^o)^a)}-} \ar@<-1ex>[d]_{\bF_p\ten_{\bZ_p}-}\\ \text{perfectoid almost }K^{\flat o}\text{-algebras}\ar@<-1ex>[u]_{W^{(p)}} } \] of categories, where $\sA_{\inf}(C):= \Lim_{\Psi^p}W^{(p)} (C)$. \end{lemma} \begin{proof} A perfectoid $\Lambda_p$-ring $B$ is a deformation of the $\bF_p$-algebra $B/p$. As in \cite[Proposition 5.13]{scholzePerfectoidSpaces}, a perfect $\bF_p$-algebra $C$ has a unique deformation $W^{(p)}C$ over $\Z_p$, to which Frobenius must lift uniquely; this gives the bottom pair of equivalences. We then observe that since $B:=\sA_{\inf} (C)$ is a perfectoid $\Lambda_p$-ring for any flat $p$-adically complete $\Z_p$-algebra $C$, we must have $B \cong W^{(p)}(B/p)$. Comparing rank $1$ elements then gives a monoid isomorphism $(B/p)\cong \Lim_{x \mapsto x^p} C$, from which it follows that \[ \bF_p\ten_{\bZ_p}\sA_{\inf}(C)\cong \Lim_{\Phi} (C/p)=C^{\flat} \] whenever $C$ is perfectoid. Since tilting gives an equivalence of almost algebras by \cite[Theorem 5.2]{scholzePerfectoidSpaces}, this completes the proof. \end{proof} We will only apply Lemma \ref{cfperfectoid} to perfectoid almost $\Lambda_p$-rings over $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$, in which case it shows that reduction modulo $[p]_{q^{1/p}}$ (resp. $p$) gives an equivalence with perfectoid $(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a$-algebras (resp. perfectoid $(\bF_p[q^{1/p^{\infty}}]^{\wedge_{(q-1)}})^a$-algebras), \subsection{Functoriality of $q$-de Rham cohomology} Since $(\Psi^{1/p^{\infty}}\widehat{\qDR}_p(A/\Z_p))^{\wedge_p}$ is represented by a cosimplicial perfectoid $\Lambda_p$ ring over $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$ for any flat $\Lambda_p$-ring $A$ over $\Z_p$, it corresponds under Lemma \ref{cfperfectoid} to a cosimplicial perfectoid $(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a$-algebra, representing the following functor: \begin{lemma}\label{perfectoidreplemma} For a perfectoid $(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a$-algebra $C$, and a $\Lambda_p$-ring $A$ over $\Z_p$ with $X=\Spec A$, there is a canonical isomorphism \[ X_{\strat}^q(\Lim_{\Psi^p}W^{(p)} (C)_*) \cong \im( \Lim_{\Psi^p} X(C_*) \to X(C_*)), \] for the ring $C_*$ of almost elements. \end{lemma} \begin{proof} By definition, $X_{\strat}^q(\Lim_{\Psi^p}W^{(p)} (C)_*)$ is the image of \[ \Hom_{\Lambda_p}(A, \Lim_{\Psi^p}W^{(p)} (C)_*) \to \Hom_{\Lambda_p}(A, (\Lim_{\Psi^p}W^{(p)} (C)_*)/(q-1)). \] Since right adjoints commute with limits, we may rewrite the first term as $\Lim_{\Psi^p} \Hom_{\Lambda_p}(A, W^{(p)} (C_*))= \Lim_{\Psi^p} X(C_*)$. Setting $B:=\Lim_{\Psi^p}W^{(p)} (C)_*$, observe that because $[p^n]_{q^{1/p^n}}(q^{1/p^n}-1)=(q-1)$, we have $\bigcap_n [p^n]_{q^{1/p^n}}B=(q-1)B$, any element on the left defining an almost element of $(q-1)B$, hence a genuine element since $B=B_*$ is flat. Then note that since the projection map $\theta \co B \to C_*$ has kernel $([p]_{q^{1/p}})$, the map $\theta \circ \Psi^{p^{n-1}}$ has kernel $([p]_{q^{1/p^n}})$, and so $B \to W^{(p)}C_*$ has kernel $\bigcap_n [p^n]_{q^{1/p^n}}B$. Thus \[ \Hom_{\Lambda_p}(A, (\Lim_{\Psi^p}W^{(p)} (C)_*)/(q-1)\into \Hom_{\Lambda_p}(A,W^{(p)}C_*)= X(C_*). \] \end{proof} In fact, the tilting equivalence gives $\Lim_{\Psi^p} X(C_*) \cong X(C^{\flat}_*)$, so the only dependence of $((\Psi^{1/p^{\infty}}\widehat{\qDR}_p(A/\Z_p))^{\wedge_p})^a$ on the Frobenius lift $\Psi^p$ is in determining the image of $ X(C^{\flat}_*) \to X(C_*)$. Although this map is not surjective, it is almost so in a precise sense, which we now use to establish independence of $\Psi^p$, showing that, up to faithfully flat descent, $\hat{\qDR}_p(A/\Z_p)^{\wedge_p}/[p]_{q^{1/p}}$ is the best possible perfectoid approximation to $A[\zeta_{p^{\infty}}]^{\wedge_p}$. \begin{definition} Given a functor $X$ from $(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a $-algebras to sets and a functor $\sA$ from perfectoid $(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a$-algebras to abelian groups, we write \[ \oR\Gamma_{\Pfd}(X,\sA):=\oR\Hom_{[\Pfd((\Z_p[\zeta_{p^{\infty}}]^{\wedge_p})^a),\Set]}( X , \sA), \] where $\Pfd(S^a) $ denotes the category of perfectoid almost $S$-algebras, and $[\C,\Set]$ denotes set-valued functors on $\C$. When $X$ is representable by a $(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a $-algebra $C$, we simply denote this by $ \oR\Gamma_{\Pfd}(C,\sA)$ --- when $C$ is perfectoid, this will just be $\sA(C)$. \end{definition} The following gives a refinement of \cite[Theorem 1.17]{BhattMorrowScholze}, addressing some of the questions in \cite[Remark 1.11]{BhattMorrowScholze} \begin{theorem}\label{mainthm4} If $R$ is a $p$-adically complete $\Lambda_p$-ring over $\Z_p$, and $A$ a formal $R$-deformation of a smooth ring over $(R/p)$, then the complex \[ \oR\Gamma_{\Pfd}((A[\zeta_{p^{\infty}}]\ten_R \Psi^{1/p^{\infty}}R)^{\wedge_p},\sA_{\inf}) \] of $\Z_p[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$-modules is almost quasi-isomorphic to $(\Psi^{1/p^{\infty}}\hat{\qDR}_p(A/R))^{\wedge_p} $ for any $\Lambda_p$-ring structure on $A$ coming from a framing over $R$ as in Theorem \ref{mainthm3}. \end{theorem} \begin{proof} First observe that $((\Psi^{1/p^{\infty}}\hat{\qDR}_p(A/R))^{\wedge_p} )^a $ is the completion of $\hat{\qDR}_p(A/R)$ with respect to the category of cosimplicial perfectoid almost $\Lambda_p$-rings over $\Z[q^{1/p^{\infty}}]^{\wedge_{(p,q-1)}}$. Combining the definition of $\widehat{\qDR}_p$ with Lemma \ref{cfperfectoid}, it then follows that for $X=\Spec A$ and $Y=\Spec R$, the complex $(\widehat{\qDR}_p(A/R)^{\wedge_p})_* $ is given by the homotopy limit \[ \oR\Gamma_{\Pfd}((X_{\strat}^q \by_{Y_{\strat}^q}Y) \circ (\sA_{\inf})_*, (\sA_{\inf})_*). \] Writing $X^{\infty}(C):= \im( \Lim_{\Psi^p} X(C_*) \to X(C_*))$, Lemma \ref{perfectoidreplemma} then combines with the description above to give \begin{align*} (\widehat{\qDR}_p(A/R)^{\wedge_p})_* &\simeq \oR\Gamma_{\Pfd}(X^{\infty}\by_{Y^{\infty}}\Lim_{\Psi^p}Y, (\sA_{\inf})_*),\\ &\simeq \oR\Gamma_{\Pfd}(X^{\infty}\by_Y\Lim_{\Psi^p}Y, (\sA_{\inf})_*). \end{align*} We now introduce a Grothendieck topology on the category $[ \Pfd_{(\Z[\zeta_{p^{\infty}}]^{\wedge_p})^a},\Set]$ by taking covering morphisms to be those maps $C \to C'$ of perfectoid algebras which are almost faithfully flat modulo $p$. Since $C^{\flat}=\Lim_{\Phi} (C/p)$, the functor $\sA_{\inf}$ satisfies descent with respect to these coverings, so the map \[ \oR\Gamma_{\Pfd}((X^{\infty}\by_Y\Lim_{\Psi^p}Y)^{\sharp}, (\sA_{\inf})_*)\to \oR\Gamma_{\Pfd}(X^{\infty}\by_Y\Lim_{\Psi^p}Y, (\sA_{\inf})_*) \] is a quasi-isomorphism, where $(-)^{\sharp}$ denotes sheafification. In other words, the calculation of $(\widehat{\qDR}_p(A/R)^{\wedge_p})^a$ is not affected if we tweak the definition of $X^{\infty}$ by taking the image sheaf instead of the image presheaf. We then have \[ (X^{\infty})^{\sharp}(C)=\bigcup_{C\to C'} \im( X(C_*)\by_{X(C'_*)} \Lim_{\Psi^p} X(C_*')\to X(C_*)), \] where $C \to C'$ runs over all covering morphisms. Now, $\Lim_{\Psi^p} X$ is represented by the perfectoid algebra $(\Psi^{1/p^{\infty}}A)^{\wedge_p}$, which is isomorphic to $A[x_1^{1/p^{\infty}}, \ldots,x_d^{1/p^{\infty}}]^{\wedge_p}$ as in the proof of Lemma \ref{perfectioncoholemma}. This allows us to appeal to Andr\'e's results \cite[\S 2.5]{andreFacteur} as generalised in \cite[Theorem 2.3]{bhattDirectSummand}. For any morphism $f \co A \to C$, there exists a covering morphism $C \to C_i$ such that $f(x_i)$ has arbitrary $p$-power roots in $C_i$. Setting $C':= C_1\ten_C \ldots \ten_C C_d$, this means that the composite $A \xra{f} C \to C'$ extends to a map $(\Psi^{1/p^{\infty}}A)^{\wedge_p} \to C'$, so $f \in (X^{\infty})^{\sharp}(C)$. We have thus shown that $(X^{\infty})^{\sharp}=X$, giving the required equivalence \[ ((\Psi^{1/p^{\infty}}\hat{\qDR}_p(A/R))^{\wedge_p} )_* \simeq \oR\Gamma_{\Pfd}(X\by_Y\Lim_{\Psi^p}Y, (\sA_{\inf})_*). \] \end{proof} \begin{corollary}\label{finalcor} If $R$ is a $p$-adically complete $\Lambda_p$-ring over $\Z_p$, and $A$ a formal $R$-deformation of a smooth ring over $(R/p)$, then the $q$-de Rham cohomology complex $(q\mbox{-}\Omega^{\bt}_{A/R, \boxempty}\ten_{R[q]}(\Psi^{1/p^{\infty}}R)[q^{1/p^{\infty}}])^{\wedge_p}$ is, up to quasi-isomorphism, independent of a choice of co-ordinates $\boxempty$. It is naturally an invariant of the commutative $p$-adically complete $(\Psi^{1/p^{\infty}}R)[\zeta_{p^{\infty}}]^{\wedge_p}$-algebra $(A[\zeta_{p^{\infty}}]\ten_R \Psi^{1/p^{\infty}}R)^{\wedge_p}$. \end{corollary} \begin{proof} By Theorem \ref{mainthm4}, we know that the complex $((\Psi^{1/p^{\infty}}\hat{\qDR}_p(A/R))^{\wedge_p})_*$ depends only on $(A[\zeta_{p^{\infty}}]\ten_R \Psi^{1/p^{\infty}}R)^{\wedge_p}$. Since \[ \Psi^{1/p^{\infty}}\qDR_p(A/R) = \Psi^{1/p^{\infty}}\qDR_p((A\ten_R \Psi^{1/p^{\infty}}R)/\Psi^{1/p^{\infty}}R), \] Theorem \ref{mainthm3} combines with Lemmas \ref{perfectioncoholemma} and \ref{recoveralmost} to give \[ ( q\mbox{-}\Omega^{\bt}_{A/R, \boxempty}\ten_{R[q]}(\Psi^{1/p^{\infty}}R)[q^{1/p^{\infty}}])^{\wedge_p} \simeq \oL\eta_{(q-1)}((\Psi^{1/p^{\infty}}\hat{\qDR}_p(A/R))^{\wedge_p})_*, \] which completes the proof. \end{proof} \begin{remark}[Scholze's conjectures] If we weaken the conjectures of \cite{scholzeqdef} by attaching roots of $q$, then most of them follow from the results above. Taking $R=\Z_p$ in Corollary \ref{finalcor} establishes the analogue of \cite[Conjecture 3.1]{scholzeqdef} (co-ordinate independence of $q$-de Rham cohomology over $\Z$) via an arithmetic fracture square. Taking more general base rings $R$ in Corollary \ref{finalcor} gives an analogue of \cite[Conjecture 7.1]{scholzeqdef}, further weakened by having to invert all Adams operations on $R$. The description of Theorem \ref{mainthm4} is very closely related to the definition of $A\Omega^{\bt}$ in \cite[Definition 1.12]{BhattMorrowScholze}, giving an analogue of \cite[Conjecture 4.3]{scholzeqdef}, and hence the comparison with singular cohomology in \cite[Conjecture 3.3]{scholzeqdef}. The operations described in \cite[Conjectures 6.1 and 6.2]{scholzeqdef} correspond to the Adams operations on $\qDR$ respectively at and away from the residue characteristics. Remark \ref{qconnrmk} provides a category of $q$-connections as described in \cite[Conjecture 7.5]{scholzeqdef}; these will correspond via Theorem \ref{mainthm} to projective $\sA_{\inf}$-modules on the site of integral perfectoid algebras over $A[\zeta_{p^{\infty}}]^{\wedge_p}$, so are again independent of co-ordinates after base change to $\Z[q^{1/P^{\infty}}]$. \end{remark} \bibliographystyle{alphanum}

\section{Introduction and presentation of the results} Our work falls within the framework of graph theory with a special focus on decomposition problems. A {\it graph} $G=(V(G), E(G))$ (or $G=(V, E)$) consists of a finite set $V$ of {\it vertices} together with a set $E$ of pairs of distinct vertices, called {\it edges}. On the one hand, let $G=(V, E)$ be a graph. The \emph{neighborhood} of $x$ in $G$, denoted by $N_{G}(x)$ or simply $N(x)$, is the set $N_{G}(x) = \{y \in V \setminus \{x\}: \{x,y\} \in E\}$. The \emph{degree} of $x$ in $G$, denoted by $d_G(x)$ ( or $d(x)$), is the cardinal of $N_{G}(x)$. A vertex $x$ with degree one is called a \emph{leaf}, its adjacent vertex is called a \emph{support} vertex and it is denoted by $x^+$. If $x$ is a support vertex in $G$ admitting a unique leaf neighbor, this leaf is denoted by $x^-$. The set of leaves and support vertices in a graph $G$ is denoted by $\mathcal{L}(G)$ and $\mathcal{S}(G)$ respectively. An \emph{internal} vertex is a vertex with a degree greater than or equal to 2. The \emph{distance} between two vertices $u$ and $v$ in $G$ is the length (number of edge) of the shortest path connecting them and is denoted by $dist_G(u,v)$ or simply $dist(u,v)$. The notation $x \sim Y$ for each $Y\subseteq V\backslash \{x\}$ means $x$ is adjacent to all or none vertex of $Y$. The negation is denoted by $x \not\sim Y$. On the other hand, given a graph $G=(V, E)$, with each subset $X$ of $V$, the graph $G[X]= (X, \displaystyle E\cap (^X_2))$ is an {\it induced subgraph} of $G$. For $X \subseteq V$ (resp. $x \in V $), the induced subgraph $G[V \setminus X]$ (resp. $G[ V\setminus \{x\}]$) is denoted by $G-X$ (resp. $G-x$). The notions of isomorphism and embedding are defined in the following way. Two graphs $G = (V, E)$ and $G' = (V', E')$ are {\it isomorphic}, which is denoted by $G \simeq G'$, if there is an {\it isomorphism} from $G$ onto $G'$, i.e., a bijection from $V$ onto $V'$ such that for all $x, y \in V$, $\{x, y\} \in E$ if and only if $\{f(x), f(y)\} \in E'$. We say that a graph $G'$ {\it embeds} into a graph $G$ if $G'$ is isomorphic to an induced subgraph of $G$. Otherwise, we say that $G$ {\it omits} $G'$. Given a graph $G$ and a one-to-one function $f$ defined on a set containing $V(G)$, we denote by $f(G)$ the graph $(f(V(G)); f(E(G)))=(f(V(G)); \{\{f(x), f(y)\}: \{x, y\}\in E(G)\})$. A nonempty subset $C$ of $V$ is a \emph {connected component} of $G$ if for $x \in C$ and $y \in V \setminus C, \, \{x,y\} \notin E$ and if for $x \neq y \in C,$ there is a sequence $x=x_{0}, \, \ldots, \, x_{n}=y$ of $C$ elements such that for each $0\leq i \leq n-1$, $\{x_{i},x_{i+1}\} \in E$. A vertex $x$ of $G$ is \emph{isolated} if $\{x\}$ constitutes a connected component of $G$. The set of the connected components of $G$ is a partition of $V$, denoted $\mathcal{C}(G)$. The graph $G$ is \emph{connected} if it has at most one connected component of $G$. Otherwise, it is called \emph{non-connected}. For example, a \emph{tree} is a connected graph in which any two vertices are connected by exactly one path. In addition, let's consider a graph $G=(V, E)$, a subset $M$ of $V$ is a \emph{module} of $G$ if every vertex outside $M$ is adjacent to all or none of $M$. This concept was introduced in \cite{Spinrad1992P4TreesAS} and independently under the name \emph{interval} in \cite{Cournier1998MinimalIG, Frass1984LIntervalleET, Schmerl1993CriticallyIP} and an \emph{autonomous} set in \cite{Ehrenfeucht1990PrimitivityIH}. The empty set, the singleton sets, and the full set of vertices are \emph{trivial modules}. A graph is \emph{indecomposable} (or {\it primitive}) if all its modules are trivial; otherwise, it is decomposable. Therefore, indecomposable graphs with at least four vertices are prime graphs. This concept was developed in several papers e.g (\cite{Ehrenfeucht1990PrimitivityIH, Frass1984LIntervalleET, Kelly1986InvariantsOF, Schmerl1993CriticallyIP, Sumner1973GraphsIW}), and is now elaborated in a book by Ehrenfeucht, Harju and Rozenberg \cite{Ehrenfeucht1997TheTO}. Properties of the prime substructures of a given prime structure were determined by Schmerl and Trotter \cite{Schmerl1993CriticallyIP} in their fundamental paper. Indeed, several papers within the same sphere of reference have then appeared (\cite{belkhechine2010indecomposable, belkhechine2010les, Bouchaala2013FiniteTW, DBLP:journals/arscom/Boudabbous16, Boussari2007LesG2, Ehrenfeucht1997TheTO, Ille1997IndecomposableG, Elayech2015TheD, Pouzet2009OnMP, Sayar2011PartiallyCT}). For instance, the {\it path} defined on $\mathbb{N}_n=\{1,...,n\}$, denoted by $P_n$, is prime for $n \geq 4$. A path with extremities $x$ and $y$ is referred to as $(x,y)$-path. For example, it is easy to verify that each prime graph is connected. The study of the hereditary aspect of the primality in the graphs revolve around the following general question. Given a prime graph $G$, is there always a proper prime subgraph in $G$ ? Addressing this problematic led to the publication of numerous papers. A first result in this direction dates back to D. P. Sumner \cite{Sumner1973GraphsIW} who asserted that for every prime graph $G$, there exists $X \subseteq V (G)$ such that $|X| = 4$ and $G[X]$ is prime. In 1990, A. Ehrenfeucht and G. Rozenberg \cite{Ehrenfeucht1990PrimitivityIH} reported also that the prime graphs have the following ascendant hereditary property. Let $X$ be a subset of a prime graph $G$ such that $G[X]$ is prime. If $|V (G)\backslash X|\geq 2$, then there exist $x \neq y \in V (G) \backslash X$ such that $G[X \cup \{x, y\}]$ is prime. The latter result was improved in 1993 by J. H. Schmerl and W. T. Trotter \cite{Schmerl1993CriticallyIP} as follows: Each prime graph of order $n$, ($n \geq 7$), embeds a prime graph of order $n - 2$. It is then natural to raise the next question. Given a prime graph $G$ of order $n$, is there always a prime subgraph of $G$ of order $n-1$? The answer to this question is negative and the prime graph $G$ such that $G-x$ is decomposable for each $x \in V(G)$, referred to as \emph{critical graph}, is the counterexample. In 1993, J.H. Schmerl and W.T. Trotter \cite{Schmerl1993CriticallyIP} characterized the critical graphs. Consider now a prime graph $G=(V,E)$. A vertex $x$ of $G$ is said to be {\it critical} if $G-x$ is decomposable. Otherwise, $x$ is a {\it non-critical} vertex. The set of the non-critical vertices of $G$ is denoted by $\sigma(G)$. Moreover, if $G$ admits $k$ non-critical vertices, it is then called a \emph{$(-k)$-critical graph}. Recently, Y.Boudabbous and Ille \cite{BOUDABBOUS20092839} asked about the description of the $(-1)$-critical graphs. Their question was answered by H. Belkhechine, I. Boudabbous and M. Baka Elayech \cite{belkhechine2010les} in the case of graph. More recently, I. Boudabbous and W. Marweni described the triangle-free prime graphs having at most two non critical vertices \cite{DBLP:journals/mvl/BoudabbousM20}. Another intrinsic tool in this work is the notion of \emph{minimal} graphs defined as follows. A prime graph $G$ is {\it minimal} for a vertex subset $X$, or {\it $X$-minimal}, if no proper induced subgraph of $G$ containing $X$ is prime. A graph $G$ is {\it $k$-minimal} if it is minimal for some $k$-element set of $k$ elements. A. Cournier and P. Ille \cite{Cournier1998MinimalIG} in 1998 characterized the $1$-minimal and $2$-minimal graphs. Recently, M. Alzohairi and Y. Boudabbous characterized the 3-minimal triangle-free graphs \cite{ALZOHAIRI20143}. In 2015, M. Alzohairi characterized the triangle-free graphs which are minimal for some nonstable 4-vertex subset \cite{ALZOHAIRI2015159}.\\ Motivated by these two fundamental notions, I. Boudabbous proposes to find the $(-k)$-critical graphs and $k$-minimal graphs for some integer $k$ even in a particular case of graphs. This work resolves what is requested by I. Boudabbous. For this reason, we shall describe the prime tree having exactly $k$ non-critical vertices. Recall that $\lfloor x\rfloor$ denotes the greatest integer $\leq x$. Therefore, we obtain: \begin{theorem} \label{TH10} Let $T=(V,E)$ be a tree with at least $5$ vertices and $\{x_1,...,x_k\}$ be a vertex subset of $G$ where $k$ is an integer. \\ $T$ is $(-k)$-critical and $\sigma(T)= \{x_1,...,x_k\}$ (see Figure \ref{FK} (a)), if and only if $T$ satisfies the four assertions. \begin{enumerate} \item For each $x\neq y \in \mathcal{L}(T)$, $dist(x,y)\geq 3$, \item $\{x_1,...,x_k\} \subseteq \mathcal{L}(T)$ and $1 \leq k\leq \lfloor\frac{n}{2}\rfloor$, \item For each $x \in \mathcal{L}(T)\backslash\{x_1,...,x_k\}$, there is a unique $i\in\{1,...,k\}$ such that $dist(x,x_i)=3$ and $d(x^+)=2$, \item If $d(x_{i}^+)=2$ where $i \in \{1,...,k\}$, then for all $x \in \mathcal{L}(T)\backslash\{x_i\}$, $dist(x_i, x)\geq 4$. \end{enumerate} \end{theorem} Moreover, we shall describe the $k$-minimal trees. As a matter of fact, we obtain: \begin{theorem} \label{TH20} Let $T=(V,E)$ be a tree with at least $5$ vertices and let $\{x_1,...,x_k\}$ be a vertex subset of $G$ where $k$ is a strictly positive integer.\\ $T$ is minimal for $\{x_1,...,x_k\}$ (see Figure \ref{FK} (b)) if and only if $T$ satisfies the three assertions. \begin{enumerate} \item For each $x\neq y \in \mathcal{L}(T)$, $dist(x,y)\geq 3$. \item For each $x \in \mathcal{L}(T)$, $\{x, x^+\}\cap\{x_1,...,x_k\}\neq\emptyset$. \item If $x_i\in \mathcal{S}(T)$ and $x_i^-\notin \{x_1,...,x_k\}$ where $i \in \{1,...,k\}$, then $d(x_i)=2$ and there is $j\neq i \in \{1,...,k\}$ such that $x_j \in \mathcal{L}(T)$ and $d(x_i, x_j)=2$. \end{enumerate} \end{theorem} \begin{figure}[!h] \centering \includegraphics[width=11cm]{TT} \caption{~~$(a)$ $T$ is $(-6)$-critical and $\sigma(T)= \{x_1,...,x_6\}$.~~~~ $(b)$ $T'$ is minimal for $\{x_1,...,x_9\}$.} \label{FK} \end{figure} \section{Proof of Theorem 1.1:} We recall the characterization of the prime tree set forward by to M. Alzohairi and Y. Boudabbous. \begin{lemma}{\rm (\cite{ALZOHAIRI20143})}\label{tree} \begin{enumerate} \item If $M$ is a nontrivial module in a decomposable tree $T$, then $M$ is a stable set of $T$. Moreover, the elements of $M$ are leaves of $T$. \item A tree with at least four vertices is prime if and only if any two distinct leaves do not have the same neighbor. \end{enumerate} \end{lemma} As an immediate consequence of Lemma \ref{tree}, we have the following result. \begin{corollary}\label{wtree} Let $T=(V,E)$ be a tree. $T$ is prime if and only if $d(x,y)\geq 3$, for each $x\neq y \in \mathcal{L}(T)$. \end{corollary} The following observation follows immediately from Lemma \ref{tree}. \begin{observation}\label{O1} Let $T=(V,E)$ be a prime tree with $n$ vertices. Then, $|\mathcal{S}(T)|=|\mathcal{L}(T)|\leq \lfloor\frac{n}{2}\rfloor$. \end{observation} Now, we establish the next lemma that will be needed in the sequel. \begin{lemma} \label{LL1} Let $T=(V,E)$ be a prime tree and $x\in \mathcal{L}(T)$. If $T-x$ is decomposable, then there is $y \in \mathcal{L}(T)\backslash\{x\}$ such that $\{y, x^+\}$ is the unique module of $T-x$. \end{lemma} \noindent{\textbf{Proof:}} Consider a prime tree $T=(V,E)$ and $x\in \mathcal{L}(T)$. Assume that $T-x$ is a decomposable tree. Resting upon Lemma \ref{tree}, there exist two distinct leaves of $T-x$, said $y$ and $z$, which have the same neighbor. Hence, $\{y, z\}$ is a module of $T-x$. Since $T$ is a prime tree, $x \not\sim \{y,z\}$. Thus, $x^+\in\{y,z\}$. Without loss of generality, we may assume that $x^+ =z$ and we have $y\in \mathcal{L}(T)$. Since $T$ is prime and $y\in \mathcal{L}(T)$, then $dist(y,u)\geq 3$ for each $u \neq y\in \mathcal{L}(T)$. Therefore, $\{y, x^+\}$ is the unique module of $T-x$.{\hspace*{\fill}$\Box$\medskip}\\ \noindent{\textbf{Proof of Theorem \ref{TH10}.}} Consider a tree $T=(V,E)$ with $n$ vertices where $n\geq 5$ and $\{x_1,...,x_k\}$ is a subset of $V$ where $k$ is a strictly positive integer.\\ Assume that $T$ is $(-k)$-critical and $\sigma(T)= \{x_1,...,x_k\}$. Since $T$ is prime, by Corollary \ref{wtree}, we have for each $x \neq y \in \mathcal{L}(T),~dist(x,y)\geq 3$. Hence, $T$ satisfies the condition (1) of Theorem \ref{TH10}.\\ Moreover, let $x\in V \backslash \mathcal{L}(T)$, $x$ is an internal vertex of $T$ and $T-x$ is a non-connected graph. Then, $T-x$ is decomposable and $x\notin \sigma(T)$. Thus, $\{x_1,...,x_k\} \subseteq \mathcal{L}(T)$. As $T$ is prime, based on Observation \ref{O1}, we have $1\leq k\leq \lfloor\frac{n}{2}\rfloor$. Hence, $T$ satisfies condition (2) of Theorem \ref{TH10}.\\ Now, consider $x \in \mathcal{L}(T)\backslash\{x_1,...,x_k\}$. Then, $T-x$ is a decomposable tree. By Lemma \ref{LL1}, there is $y \in \mathcal{L}(T)\backslash\{x\}$ such that $\{y, x^+\}$ is the only module of $T-x$. Clearly, $dist(x,y)=3$ and $d(x^+)=2$. Now, prove that $y\in \sigma(T)$. To the contrary, suppose that $y\notin \sigma(T)$, implying that $T-y$ is a decomposable tree. Using again Lemma \ref{LL1}, there is $z \in \mathcal{L}(T)\backslash\{y\}$ such that $\{z, y^+\}$ is the unique module of $T-y$. Thus, $d(y^+)=2$ and $dist(y,z)=3$. This implies that $z=x$ and we obtain that $T$ is with 4 vertices; which contradicts the fact that $T$ is a tree having at least 5 vertices. Hence, $y\in \sigma(T)$. Therefore, $T$ satisfies the condition (3) of Theorem \ref{TH10}.\\ Besides, assume that there is $i \in \{1,...,k\}$ such that $d(x_i^+)=2$. Then, $T-x_i$ is a prime tree and $x_i^+ \in \mathcal{L}(T-x_i)$. Referring to Corollary \ref{wtree}, for all $y\in\mathcal{L}(T-x_i)$, $dist(x_i^{+}, y)\geq3$. Since $\mathcal{L}(T-x_i)\backslash \{x_i^+\}=\mathcal{L}(T) \backslash \{x_i\}$, then for each $y\neq x_i\in \mathcal{L}(T)$, $dist(x_i,y)\geq 4$. Hence, $T$ satisfies condition (4) of Theorem \ref{TH10}. Conversely, assume that $T$ satisfies the conditions (1)-(4) of Theorem \ref{TH10}. Proving that, $T$ is $(-k)$-critical and $\sigma(T)= \{x_1,...,x_k\}$. Since for each $x\neq y \in \mathcal{L}(T)$, $dist(x,y)\geq 3$ and by Corollary \ref{wtree}, $T$ is prime. Clearly, if $x\in V \backslash \mathcal{L}(T)$, $T-x$ is a non-connected graph. Thus, $T-x$ is decomposable and hence $x$ is a critical vertex.\\ Furthermore, if $x \in \mathcal{L}(T)\backslash\{x_1,...,x_k\}$, by assertion (3), there is a unique $i\in\{1,...,k\}$ such that $dist(x,x_i)=3$ and $d(x^+)=2$. Then, $\{x^{+}, x_i\}$ is a module of $T-x$. Hence for each $x\in\mathcal{L}(T)\backslash\{x_1,...,x_k\}$, $T-x$ is decomposable and then $x$ is a critical vertex.\\ Now, given $i \in \{1,...,k\}$; if $d(x^+_i)\geq 3$, then $\mathcal{L}(T)\backslash\{x_i\}=\mathcal{L}(T-x_{i})$. According to first hypothesis of Theorem \ref{TH10}, for each $x\neq y \in\mathcal{L}(T-x_{i})$, $dist(x, y)\geq3$. By Corollary \ref{wtree}, $T-x_i$ is a prime tree.\\ Assume now that $d(x_i^+)=2$. Suppose that $T-x_i$ is a decomposable tree. Then, by Lemma \ref{LL1} there is a unique $y \neq x_i \in \mathcal{L}(T)$ such that $\{y, x_i^+\}$ is the unique module of $T-x_i$. As a matter of fact, $dist(y,x_i)=3$; which contradicts the hypothesis 4 of Theorem \ref{TH10}. Hence, $T-x_i$ is prime. Consequently, $T$ is $(-k)$-critical and $\sigma(T)= \{x_1,...,x_k\}$. {\hspace*{\fill}$\Box$\medskip} Our second objective in this section lies in determining the number of nonisomorphic $(-k)$-critical trees with $n\geq5$ vertices where $k \in \{1,2, \lfloor\frac{n}{2}\rfloor\}$. According to the characterization of critical graphs \cite{Schmerl1993CriticallyIP}, $P_4$ is the a unique critical trees. To specify the the number of nonisomorphic $(-k)$-critical trees where $k \in \{1,2,\lfloor\frac{n}{2}\rfloor\}$, we introduce for all $n\in \mathbb{N}$, the one-to-one function: $$\begin{array}{ccccc} T_n & : & \mathbb{N} & \to & \mathbb{N} \\ & & p & \mapsto & p+n \\ \end{array}$$ Now, we introduce also the following trees. \vspace{-0.2cm} \begin{description} \item[$\bullet$] For integers $m\geq 2$, let $A_{2m+1}$ be the tree defined on $\{0,...,2m\}$ and $E(A_{2m+1})=\{\{0,i\}, \{i,i+m\} : 1\leq i \leq m\}$ (see Figure \ref{AA}). \item[$\bullet$] For integers $k\geq 4$, $t\geq 1$, let $P_{k, t}$ be the tree defined on $\{1,...,2t+k\}$ and $E(P_{k,t})=E(T_{2t}(P_k))\cup\{\{2i-1, 2i\}: 1\leq i \leq t\}\cup \{\{2t+2, 2i\}: 1\leq i \leq t\}$ (see Figure \ref{2CR1}). \item[$\bullet$] For integers $m\geq 4$, $n_1\geq 1$, $n_2\geq 1$, for each $p\in \{1,2\}$, $s_p=n_1+...+n_p$. Let $P_{m, n_1,n_2}$ be the tree defined on $\{1,...,2s_2+m\}$ and $E(P_{m,n_1,n_2})=E(T_{2s_2}(P_m))\cup\{\{2i-1, 2i\}: 1\leq i \leq s_2\}\cup \{\{2s_2+2, 2i\}, \{2s_2+m-1, 2j\}: 1\leq i \leq n_1 ~~and ~~n_1< j \leq s_2\}$ (see Figure \ref{2CR}). \end{description} \begin{figure}[h!] \centering \includegraphics[width=5cm]{AA \caption{The tree $A_{2m+1}$} \label{AA} \end{figure} \begin{proposition} \begin{enumerate} \item Up to isomorphisms, the $(-1)$-critical trees with $n$ vertices are the tree $\displaystyle P_{4,\frac{n-4}{2}}$ where $n$ is an even integer $\geq 6$. \item Up to isomorphisms, the $\displaystyle (-\lfloor\frac{n}{2}\rfloor)$-critical trees with $n$ vertices are the tree $A_{n}$ where $n$ is an odd integer $\geq 5$. \end{enumerate} \end{proposition} \begin{figure}[h!] \centering \includegraphics[width=12cm]{2CR1 \caption{The tree $P_{k, t}$} \label{2CR1} \end{figure} \begin{figure}[h!] \centering \includegraphics[width=12cm]{2CR \caption{The tree $P_{m, n_1,n_2}$} \label{2CR} \end{figure} \noindent{\textbf{Proof:}}\begin{enumerate} \item Clearly, departing from Theorem \ref{TH10}, $\displaystyle P_{4,\frac{n-4}{2}}$ is a $(-1)$-critical tree where $n\geq 6$ and $\displaystyle\sigma(P_{4,\frac{n-4}{2}})=\{n-3\}$. Now, we consider a $(-1)$-critical tree $T$ with $n\geq 5$ vertices such that $\sigma(T)=\{x_1\}$. By Theorem \ref{TH10}, $x_1 \in \mathcal{L}(T)$. If $x\neq x_1\in \mathcal{L}(T)$, then by assertion (3) of Theorem \ref{TH10}, $dist(x, x_1)=3$ and $d(x^+)=2$. On the contrary, suppose that $|\mathcal{L}(T)|=2$. Since $T$ is prime, then $T$ is isomorphic to $P_4$; which contradicts the fact that $T$ has at least 5 vertices. Hence, $|\mathcal{L}(T)|\geq 3$. By assertion (3), for each $y\in\mathcal{L}(T)$, $dist(y, x_1)=3$ and $d(y^+)=2$. Thus, $T$ is isomorphic to $\displaystyle P_{4,\frac{n-4}{2}}$ where $n\geq 6$ is an even integer. \item By Theorem \ref{TH10}, $A_{2m+1}$ is $\displaystyle(-\lfloor\frac{2m+1}{2}\rfloor)$-critical and $\sigma(A_{2m+1})=\mathcal{L}(A_{2m+1})$ where $m\geq 2$. Now, we consider a $\displaystyle(-\lfloor\frac{n}{2}\rfloor)$-critical tree $T$ with $n\geq 5$ vertices. Using Theorem \ref{TH10}, $\sigma(T)\subseteq \mathcal{L}(T)$ implies that $\displaystyle|\sigma(T)|=\lfloor\frac{n}{2}\rfloor \leq |\mathcal{L}(T)|$. By Observation \ref{O1}, $\displaystyle|\mathcal{L}(T)| =|\mathcal{S}(T)|\leq \lfloor\frac{n}{2}\rfloor$ and therefore $\displaystyle|\mathcal{L}(T)| =|\mathcal{S}(T)|=\lfloor\frac{n}{2}\rfloor$. Now, we shall prove that $n$ is odd. To the contrary, suppose that $n$ is even. Then, $\displaystyle|\mathcal{L}(T)| =|\mathcal{S}(T)|=\frac{n}{2}$. Hence, $V(T)=\mathcal{L}(T)\cup\mathcal{S}(T)$. Since $T[\mathcal{S}(T)]$ is a tree, there exists a vertex $y\in \mathcal{S}(T)$ with $d_{T[\mathcal{S}(T)]}(y)=1$. Hence, $d_{T}(y)=2$. We may assume that $N(y)=\{y^-,x\}$ where $x\in \mathcal{S}(T)$. Thus, $\{y, x^-\}$ is a module of $T-y^-$; which contradicts the fact that $y^-$ is not a critical vertex. Accordingly, $n$ is odd, $\displaystyle |\mathcal{L}(T)|=|\mathcal{S}(T)|=\frac{n-1}{2}$, and $V(T)=\mathcal{L}(T)\cup\mathcal{S}(T)\cup\{z\}$. We can assume that $\displaystyle\mathcal{L}(T)=\{x_1,...,x_{(\frac{n-1}{2})}\}$ and $\displaystyle\mathcal{S}(T)=\{x^+_1,...,x^+_{(\frac{n-1}{2})}\}$. Since $T[\mathcal{S}(T)\cup\{z\}]$ is a tree, there exists a vertex $x_i^+\in \mathcal{S}(T)$ where $\displaystyle 1\leq i \leq \frac{n-1}{2}$ with $d_{T[\mathcal{S}(T)]}(x_i^+)=1$ and hence $d_{T}(x_i^+)=2$. By Theorem \ref{TH10}, $dist(x_i, x_j)\geq 4$ for each $\displaystyle j\neq i \in \{1,...,\frac{n-1}{2}\}$. Hence, $T$ is isomorphic to $A_{n}$ where $n\geq 5$.{\hspace*{\fill}$\Box$\medskip} \end{enumerate} As a consequence of Theorem \ref{TH10}, we get the following result. \begin{proposition}\label{p1} Up to isomorphisms, the $(-2)$-critical trees with $n\geq 5$ vertices are the trees $P_n$, $P_{k,t}$ where $k\geq 4$, $t\geq 1$ and $n=k+2t$, and $P_{m, n_1, n_2}$ where $m\geq4$, $n_1, n_2\geq 1$ and $n=m+2(n_1+n_2)$. \end{proposition}\label{2CT} \noindent{\textbf{Proof:}} By Theorem \ref{TH10}, $P_n$, $P_{k,t}$ where $k\geq 4$, $t\geq 1$, and $P_{m, n_1, n_2}$ where $m\geq4$, $n_1, n_2\geq 1$ are $(-2)$-critical trees and $\sigma(P_n)=\{1, n\}$, $\sigma(P_{k,t})=\{2t+1,2t+k\}$, and $\sigma(P_{m, n_1, n_2})=\{2s_2 +1, 2s_2+ m\}$. Now, assume that $T$ is a $(-2)$-critical tree with $n\geq 5$ vertices such that $\sigma (T)=\{x_1,x_2\}$. By Theorem \ref{TH10}, $x_1, x_2 \in \mathcal{L}(T)$. As $T$ is a prime tree, then the $(x_1,x_2)$-path is isomorphic to $P_k$ where $k \geq4$. If $|\mathcal{L}(T)|=2$, then $T$ is isomorphic to $P_n$ and $n=k$. Assume that $|\mathcal{L}(T)|\geq 3$, then by Theorem \ref{TH10}, for each $x\in \mathcal{L}(T)\backslash\{x_1, x_2\}$; there is a unique $i\in \{1,2\}$ such that $dist(x, x_i)=3$ and $d(x^+)=2$. Hence, $T$ is isomorphic to $P_{k,t}$ where $k\geq 4$, $t\geq 1$ and $n=k+2t$ or $T$ is isomorphic to $P_{k, n_1, n_2}$ where $k\geq4$, $n_1, n_2\geq 1$ and $n=k+2(n_1+n_2)$.{\hspace*{\fill}$\Box$\medskip} \begin{theorem} \label{TH2} The number of nonisomorphic $(-2)$-critical trees with $n$ vertices equals:\\ $\bullet$ $\displaystyle\left\lfloor\frac{n}{4}\right\rfloor^{2}-1$ if $n\equiv 0~(mod~4)$.\\ $\bullet$ $\displaystyle\left\lfloor\frac{n}{4}\right\rfloor^{2}$ if $n\equiv 1~(mod~4)$.\\ $\bullet$ $\displaystyle\left\lfloor\frac{n}{4}\right\rfloor\left(\left\lfloor\frac{n}{4}\right\rfloor+1\right)-1$ if $n\equiv 2~(mod~4)$.\\ $\bullet$ $\displaystyle\left\lfloor\frac{n}{4}\right\rfloor\left(\left\lfloor\frac{n}{4}\right\rfloor+1\right)$ otherwise. \end{theorem} \noindent{\textbf{Proof:}} At the beginning, it is not difficult to verify that there are not two isomorphic different trees in the union $\{P_m: m\geq 5\}\cup\{P_{k,t}: k\geq5, t\geq 1\}\cup\{P_{m,n_1,n_2}: m\geq4, n_1\geq1 ~~and~~ n_2\geq 1\}$. \\ By Proposition \ref{p1}, $P_5$ is the unique $(-2)$-critical tree with five vertex and $P_6$ is the unique $(-2)$-critical tree with six vertices. As a matter of fact, the result holds.\\ Now, assume that $n\geq 7$. By Proposition \ref{2CT}, the nonisomorphic $(-2)$-critical trees with $n$ vertices are $P_n$, the family of $P_{k,t}$ where $t \geq1$, $k\geq5$, and $n=2t+k$, or the family of $P_{m, n_1, n_2}$ where $1\leq n_1 \leq n_2$, $m\geq4$, and $n=2(n_1+n_2)+m$. Therefore, it is sufficient to determine the number of the family of $P_{k,t}$ and the number of the family of $P_{m, n_1, n_2}$.\\ Let $S_{m}=\{(n_1, n_2)\in \mathbb{N}\times \mathbb{N}: 1\leq n_1\leq n_2, ~~n_1+n_2=\frac{n-m}{2}\}$, where $4\leq m\leq n-4$ and let $C_{t}=\{k\in \mathbb{N}: 5\leq k~~and ~~k=n-2t\}$, where $1\leq t\leq \frac{n-5}{2}$. Since $n-m= 2(n_1+n_2)$, it is obvious that $n$ and $m$ are of the same parity. Hence, we distinguish two cases.\\ $\underline{\textbf{Case 1:}}$ If $n=2p$ where $4 \leq p$ and $m=2q$ where $2 \leq q \leq p-2$. \\ Consider $\displaystyle S=\bigcup_{q=2}^{p-2}S_{2q}$ and $\displaystyle C=\bigcup_{t=1}^{p-3}C_t$. First, it is clear that the number of the family of $P_{k,t}$ is the cardinality of the set $C$. Moreover, it is clear that $|C_t|=1$ where $1\leq t\leq p-3$. Hence, $\displaystyle |C|=\sum_{t=1}^{p-3}|C_t|=p-3$. Second, obviously the number of the family of $P_{m, n_1, n_2}$ is the cardinality of the set $S$. Furthermore, we have $\displaystyle |S|=\sum_{q=2}^{p-2}|S_{2q}|$. It is noticeable that for each $2\leq q \leq p-2$, $|S_{2q}|= P_2(\frac{n-2q}{2})$, where $P_i(j)$ is the number of partitions of $j$ to $i$ parts. Recall that for an integer $k \geq 3$, $P_2(k) =\lfloor\frac{k}{2}\rfloor$ where $\lfloor x \rfloor$ is the greatest integer $\leq x$ \cite{anderson_2006}. We get then $$\begin{aligned} \displaystyle |S| &= \sum_{q=2}^{p-2} P_{2}\left(\frac{n-2q}{2}\right)~~\\ & = \sum_{q=2}^{p-2} \left\lfloor\frac{n-2q}{4}\right\rfloor \\ & =\sum_{q=2}^{p-2} \left\lfloor\frac{\frac{n}{2}-q}{2}\right\rfloor\\ & =\sum_{i=0}^{p-2} \left\lfloor\frac{i}{2}\right\rfloor.\\ & =\displaystyle \left\{ \begin{array}{ll} ~\displaystyle (k-1)^2 & if ~p=2k, \hbox{} \\ ~~\\ ~\displaystyle (k-1)k & if ~p=2k+1. \hbox{} \end{array} \right. \end{aligned}$$ $\underline{\textbf{Case 2:}}$ Let $n=2p+1$ where $4 \leq p$ and $m=2q+1$ where $2 \leq q \leq p-2$. Let $\displaystyle S=\bigcup_{q=2}^{p-2}S_{2q+1}$ and $\displaystyle C=\bigcup_{t=1}^{p-2}C_t$. Clearly, the number of the family of $P_{k,t}$ is the cardinality of the set $C$. Hence, $\displaystyle |C|=\sum_{t=1}^{p-2}|C_t|=p-2$. In addition, the number of the family of $P_{m, n_1, n_2}$ is the cardinality of the set $S$. Therefore, we have $\displaystyle |S|=\sum_{q=2}^{p-2}|S_{2q+1}|$.\\ Since for each $2\leq q \leq p-2$, $|S_{2q+1}|= P_2\left(\frac{(n-1)-2q}{2}\right)$. Proceeding in the same manner as case 1, if $p=2k$ where $k\geq 2$, then $|S|=(k-1)^2$. Otherwise, $|S|=(k-1)k$.\\ Consequently, the number of nonisomorphic $(-2)$-critical trees with $n$ vertices equals: $$\displaystyle \left\{ \begin{array}{ll} ~\displaystyle \left\lfloor\frac{n}{4}\right\rfloor^{2}-1 & if ~n\equiv 0~(mod~4), \hbox{} \\ ~~\\ ~\displaystyle\left\lfloor\frac{n}{4}\right\rfloor^{2} &if ~n\equiv 1~(mod~4), \hbox{} \\ ~~\\ ~\displaystyle \left\lfloor\frac{n}{4}\right\rfloor\left(\left\lfloor\frac{n}{4}\right\rfloor+1\right)-1& if ~n\equiv 2~(mod~4), \hbox{}\\ ~~\\ ~\displaystyle\left\lfloor\frac{n}{4}\right\rfloor\left(\left\lfloor\frac{n}{4}\right\rfloor+1\right) &if ~n\equiv 3~(mod~4). \hbox{} \end{array} \right. $$ {\hspace*{\fill}$\Box$\medskip} \section{Proof of Theorem 1.2:} We set two major objectives throughout this section. First, to characterize the $k$-minimal trees. Second, to determine the number of nonisomorphic $k$-minimal trees with $n$ vertices where $k \in \{1,2,3\}$. \noindent{\textbf{Proof of Theorem 1.2.} Let $T=(V,E)$ be a tree with $n\geq 5$ vertices and let $\{x_1,...,x_k\}$ be a vertex subset of $G$ where $k$ is a strictly positive integer. Assume that $T$ is minimal for $\{x_1,...,x_k\}$. Since $T$ is prime, it satisfies the first condition of Theorem \ref{TH20}. Suppose, on the contrary, that there is $y \in \mathcal{L}(T)$ such that $\{y, y^+\}\cap \{x_1,...,x_k\}=\emptyset$. Since $T-y$ is a decomposable tree, by Lemma \ref{LL1}, there is $x\in \mathcal{L}(T)$ such that $\{x, y^+\}$ is a module of $T-y$. Thus, $d(y^+)=2$ and $d(x, y)=3$. If $x\notin \{x_1,...,x_k\}$, then $T-x$ is a decomposable tree. By using again Lemma \ref{LL1}, $d(x^+)=2$ and so $T$ is isomorphic to $P_4$; which contradicts the fact that $n\geq 5$.\\ Moreover, assume that $x\in \{x_1,...,x_k\}$ and $d(x^+) \geq 3$, then $\mathcal{L}(T)\backslash\{y\}=\mathcal{L}(T-\{y, y^+\})$. By Lemma \ref{tree}, $T-\{y, y^+\}$ is a prime tree containing $\{x_1,...,x_k\}$; which contradicts the fact that $T$ is minimal for $\{x_1,...,x_k\}$. Hence, for each $x\in \mathcal{L}(T)$, $\{x, x^+\}\cap\{x_1,...,x_k\}\neq\emptyset$ and $T$ satisfies the second condition of Theorem \ref{TH20}.\\ Now, assume that there is $x_i\in \mathcal{S}(T)$ and $x_i^-\notin \{x_1,...,x_k\}$ where $1\leq i \leq k$. On the contrary, suppose that $d(x_i) \geq 3$, then $\mathcal{L}(T)\backslash\{x_i^-\}=\mathcal{L}(T-x_i^-)$. By Lemma \ref{tree}, $T-x_i^-$ is a prime tree containing $\{x_1,...,x_k\}$; which is impossible. Hence, $d(x_i)=2$. On the contrary, suppose that for each $j\neq i \in \{1,...,k\}$ such that $x_j \in \mathcal{L}(T)$, $d(x_i, x_j)\geq 3$. Since $T-x_i^-$ is a decomposable tree, then by Lemma \ref{LL1} there is $y\in \mathcal{L}(T)$ such that $d(x_i, y) =2$ and hence $y\notin \{x_1,...,x_k\}$. By Condition 2, $y^+ \in \{x_1,...,x_k\}$ and so $d(y^+)=2$. Thus, $T$ is isomorphic to $P_4$; which is impossible. Therefore, $T$ satisfies the third condition. Conversely, let $T=(V,E)$ be a tree with $n\geq 5$ vertices. Since for each $x\neq y \in \mathcal{L}(T)$, $d(x, y)\geq 3$, $T$ is a prime tree. Let $X$ be a subset of $V$ such that $\{x_1,...,x_k\} \subseteq X$ and $T[X]$ is prime. Consider $x\in \mathcal{L}(T)$. If $x\in \{x_1,...,x_k\}$, then $x\in X$.\\ Now, assume that $x\notin \{x_1,...,x_k\}$. On the contrary, suppose that $x\notin X$. By assertion (2) of Theorem \ref{TH20}, $x^+\in \{x_1,...,x_k\}$. Since $T$ satisfies assertion (3) of Theorem \ref{TH20}, then $d(x^+)=2$ and there is $i \in \{1,...,k\}$ such that $x_i\in \mathcal{L}(T)$ and $d(x_i, x^+)=2$. Thus, $\{x^+, x_i\}$ is a module of $T[X]$; which is impossible. Hence, $x\in X$. We conclude that $\mathcal{L}(T)\subset X$. Since $T[X]$ is a prime, it is connected. Therefore, $T[X]$ is a tree containing $\mathcal{L}(T)$. Hence, $X=V$. Thus, $T$ is minimal for $\{x_1,...,x_k\}$.{\hspace*{\fill}$\Box$\medskip} The following corollary is an immediate consequence of Theorem \ref{TH20}. \begin{corollary} For any distinct vertices $x_1, x_2$,..., and $x_k$ in a prime tree $H$, there is an induced subtree $T$ of $H$ that contains $\{x_1, x_2,...,x_k\}$, and satisfies the assertions of Theorem \ref{TH20}. \end{corollary} Our second objective is to determine the number of nonisomorphic $k$-minimal trees with $n$ vertices where $k \in \{1,2,3\}$. According to the characterization of 1-minimal and 2-minimal graphs, $P_4$ is the unique 1-minimal tree and $P_k$, where $k\geq 4$, is the unique 2-minimal tree \cite{Cournier1998MinimalIG}. To specify the number of nonisomorphic $3$-minimal trees with $n$ vertices, we introduce the following tree. \begin{description} \item[$\bullet$] For positive integers $k$, $m$, $n$ with $k \leq m \leq n$, let $S_{k,m,n}$ be the $(k+m+n+1)$-vertex tree with the union of the paths of lengths $k$, $m$, and $n$ having common endpoint $r$. Let $a_1,...,a_k$, $b_1,...,b_m$, and $c_1,...,c_n$ denote the other vertices on these paths, indexed by their distance from $r$ (see Figure \ref{3MINM}). \end{description} \begin{figure}[h!] \centering \includegraphics[width=10cm]{3min \caption{$S_{k,m,n}$} \label{3MINM} \end{figure} As an immediate consequence of Theorem \ref{TH20}, we get the following result which was already obtained by M. Alzohairi and Y. Boudabbous in \cite{ALZOHAIRI20143}. \begin{corollary}\label{3mw} Let $x$, $y$, and $z$ be distinct vertices in a tree $T$. The tree $T$ is minimal for $\{x,y,z\}$ if and only if $T$ and $\{x,y,z\}$ have one of the following forms: \begin{enumerate} \item $T \simeq P_4$. \item $T \simeq P_k$ with $k\geq5$ such that $\{x,y,z\}$ contains the leaves. \item $T \simeq S_{k,m,n}$ with $m\geq 2$ such that $x$, $y$, and $z$ are the leaves. \item $T \simeq S_{1,2,n}$ such that $\{x,y,z\}=\{a_1, b_1, c_n\}$. \item $T \simeq S_{1,2,2}$ such that $\{x,y,z\} = \{a_1, b_1, c_1\}$. \end{enumerate} \end{corollary} \begin{proposition} The number of nonisomorphic $3$-minimal trees with $n$ vertices equals:\\ $\bullet$ 1 if $n\in \{4,5\}$.\\ $\bullet$ 2 if $n=6$.\\ $\bullet$ $\left[\frac{(n-1)^2}{12}\right]-\left\lfloor\frac{n-4}{2}\right\rfloor + \left\lfloor\frac{n-2}{2}\right\rfloor-1$ if $n\geq 7$, where $[x]$ is the nearest integer from $x$. \end{proposition} \noindent{\textbf{Proof:}} At the beginning, it is not difficult to verify that there are not two isomorphic different graphs in the union $\{P_k: k\geq 4\}\cup \{S_{k,m,n} : m=2\}$. \\ By Corollary \ref{3mw}, $P_4$ is the unique 3-minimal tree with four vertices and $P_5$ is the unique 3-minimal tree with five vertices. In addition, the only $3$-minimal tree with six vertices are $P_6$ and $S_{1,2,2}$. Therefore, the result holds for $n \in \{4,5,6\}$.\\ Now, assume that $n\geq 7$. By Corollary \ref{3mw}, the non isomorphic $3$-minimal $n$-vertex trees are $P_n$ and the family of $S_{k,m,t}$, where $k \leq m\leq t$, $m\geq 2$, and $k+m+t+1=n$. Therefore, it is sufficient to prove that the cardinality of the set $S=\{(k,m,t)\in \mathbb{N}\times\mathbb{N}\times\mathbb{N}:1\leq k \leq m\leq t, m\geq 2, k+m+t=n-1\}$ equals $$\displaystyle \left[\frac{(n-1)^2}{12}\right]-\left\lfloor\frac{n-4}{2}\right\rfloor+\left\lfloor\frac{n-2}{2}\right\rfloor-2.$$ Let $S_2=\{(k,m,t)\in S: k\geq 2\}$. It is easy to infer that $|S-S_2|=P_2(n-2)-1$. Notice that $|S_2|=|\{(p,q,r)\in \mathbb{N}\times\mathbb{N}\times\mathbb{N}: 1\leq p \leq q\leq r, p+q+r=n-4\}|=P_3(n-4)$. Moreover, grounded on \cite{anderson_2006}, the number of partitions of $k$ with at most $3$ parts is equal to $\left[\frac{(k+3)^2}{12}\right]$. It follows that $P_3(k)=\left[\frac{(k+3)^2}{12}\right]-\left\lfloor\frac{k}{2}\right\rfloor-1$ \cite{anderson_2006}. Therefore, $$|S_2|=\left[\frac{(n-1)^2}{12}\right]-\left\lfloor\frac{n-4}{2}\right\rfloor-1$$ and $$|S-S_2|=\left\lfloor\frac{n-2}{2}\right\rfloor-1.$$ Thus, $$|S|=\displaystyle \left[\frac{(n-1)^2}{12}\right]-\left\lfloor\frac{n-4}{2}\right\rfloor+\left\lfloor\frac{n-2}{2}\right\rfloor-2.$$ {\hspace*{\fill}$\Box$\medskip}\\ \section{Conclusion} The problems of finding the $(-k)$-critical graphs and the $k$-minimal graphs seem to be challenging where $k$ is an integer ($k\geq 2$). At least, we solve these problems in the particular case of trees. In addition, we determine the number of nonisomorphic $(-k)$-critical trees with $n\geq5$ vertices where $k \in \{1,2, \lfloor\frac{n}{2}\rfloor\}$. Besides, we specify the number of nonisomorphic $3$-minimal trees with $n$ ($n\geq 4$) vertices. \nocite{*} \bibliographystyle{abbrvnat}

Introduction ============ *Pseudomonas* spp. and their genomic features (e.g., multifaceted oxidative enzymatic repertoire) have been extensively explored as cell-factories and functional blocks for biocatalysis ([@evu254-B85]; [@evu254-B83]; [@evu254-B18]; [@evu254-B69]; [@evu254-B59]), particularly in the production of fine chemicals for several industries (e.g., terpenoids) ([@evu254-B13]; [@evu254-B55]; [@evu254-B79]; [@evu254-B21]; [@evu254-B23]; [@evu254-B9]). The instability of some monoterpenes directed biotechnological research to their derivatives which are known to retain or enhance some properties of the original compound ([@evu254-B79]) and are usually easier to incorporate into hydrophilic/aqueous formulations. The utilization of monoterpenes and their derivatives is widespread in industry because of their unique properties ([@evu254-B37]; [@evu254-B17]; [@evu254-B61]). One of the most promising monoterpenes, displaying a plethora of industrial applications, is β-myrcene (7-Methyl-3-methylene-1,6-octadiene). β-myrcene is mainly used for the synthesis of top selling flavors and fragrances (e.g., linalool, nerol, geraniol, menthol) ([@evu254-B8]) and compounds with pharmacological potential (e.g., antimutagenics, analgesics, tyrosinase inhibitors), as well as a starting material in production of polymers, biodegradable surfactants, pheromones, and insect repellents ([@evu254-B43]; [@evu254-B53]; [@evu254-B8]). In 1999, a *Pseudomonas* sp. (strain M1) was isolated from soil sediments of the Rhine River ([@evu254-B40]), which exhibited the ability to use β-myrcene as sole carbon and energy sources. Transposon random mutagenesis led to the identification of four genes potentially coding for β-myrcene-biotransforming enzymes: An aldehyde dehydrogenase (*myrA*), an alcohol dehydrogenase (*myrB*), an acyl-CoA synthetase (*myrC*), and an enoyl-CoA hydratase (*myrD*) ([@evu254-B40]). Later, a proteomic analysis ([@evu254-B68]) provided further insights into the β-myrcene catabolic pathway and associated adaptive mechanisms of *Pseudomonas* sp. M1. However, during that proteomic analysis, the authors outlined that a significant number of β-myrcene-responsive proteins were not identified due to the lack of information in the public databases, suggesting that the β-myrcene catabolic apparatus of *Pseudomonas* sp. M1 would be based, at least partially, on enzymes that have evolved divergently with respect to what have been reported in the literature so far. Curiously, in another work using *Pseudomonas* sp. M1, it was reported that its ability to use phenol and benzene as sole carbon sources relied on a unique σ^54^-dependent gene cluster organization ([@evu254-B67]), which, apparently, has evolved differently, when compared with other reported phenol catabolic pathways. Altogether, these hints suggest that *Pseudomonas* sp. M1 harbors, in its genomic repertoire, a unique code for biotechnologically relevant enzymes and functional modules, probably due to particular evolutionary constrains. Nowadays, *Pseudomonas* sp. M1 is one of the most promising gram-negative bacteria able to metabolize β-myrcene, for which a catabolic pathway draft has been proposed ([@evu254-B40]; [@evu254-B68]). *Pseudomonas* sp. M1 is also capable of mineralizing other terpenes (e.g., citronellol, citral), as well as several toxic and/or recalcitrant phenolic compounds, such as phenol, 4-chlorophenol, benzene, and toluene ([@evu254-B66]; [@evu254-B68]). The genome reconstruction conducted in this work updated the current available draft reported by [@evu254-B72] for exploitation of *Pseudomonas* sp. M1 full potential. A genomic analysis led to the identification of genomic loci that suggest a versatile metabolism of terpene-backbone compounds. Transcriptome sequencing of β-myrcene-challenged M1 cells revealed that the core-code of β-myrcene metabolism is located in a unique genetic locus and allowed the identification of the full set of M1 genes involved in β-myrcene-dependent molecular mechanisms of catabolism, adaption, and regulation. Materials and Methods ===================== Bacterial Strains, Culture Conditions, and Sampling --------------------------------------------------- *Pseudomonas* sp. M1 ([@evu254-B40]) was grown in Lysogenic Broth Lennox for DNA extraction. For gene expression analysis, cells were cultivated in 50 ml of mineral medium (MM; 8.9 mM K~2~HPO~4~, 6.2 mM NaH~2~PO~4~, 34.2 μM EDTA, 7 μM ZnSO~4~, 6.8 μM CaCl~s~, 18 μM FeSO~4~, 0.8 μM Na~s~MoO~4~, 0.7 μM CuSO~4~, 1.7 μM CoCl~2~, 1.9 μM MnCl~2~, 15.1 mM (NH~4~)~2~SO~4~, and 0.5 mM MgCl~2~) supplemented with 0.4% lactate or 100 μl of β-myrcene (Merck) (CAS number: 123-35-3) as sole carbon source, in 250-ml Erlenmeyer-sealed flasks, at 30 °C and with orbital shaking (200 rpm). As the water solubility of β-myrcene is very low (density: 0.796 g/ml; log K~ow~: 4.17; water solubility at 25 °C approximately 6 mg/l ([@evu254-B31]), the amount (100 μl) of β-myrcene used in the experiments assured a constant water phase saturation with β-myrcene to avoid eventual carbon starvation effects being reflected on the overall transcriptome analysis. For total RNA sample preparation, after overnight growth, M1 cells were inoculated to an initial OD~600nm~ of 0.08--0.1 in MM supplemented with respective carbon sources (0.4% lactate or 100 μl of β-myrcene) and cultures diluted again to OD~600nm~ of 0.08--0.1 when OD~600nm~ of 0.5 was reached to synchronize cell growth. Samples for RNA extraction were collected in triplicate (lactate: L1, L2, L3; β-myrcene: M1, M2, M3) when fresh cultures reached an OD~600nm~ of about 0.3--0.4 (2 h of exponential growth) and were immediately flash-frozen in liquid nitrogen. Next-Generation Genome Sequencing, Assembly, and Annotation ----------------------------------------------------------- The genomic DNA from *Pseudomonas* sp. M1 was extracted and purified using the Wizard Genomic purification kit (Promega). The DNA samples were processed according to Illumina instructions generating Nextera XT paired-end libraries (2 × 250 bp) which were sequenced using the Illumina MiSeq system at Yale Center for Genomic Analysis. Raw paired-reads obtained from Illumina MiSeq platform (2 × 250 bp with insert size around 470 bp) and the previously obtained data set using Illumina GAIIx (2 × 50 bp paired-end reads with insert size around 320 bp) ([@evu254-B72]) were both corrected with QUAKE ([@evu254-B44]). These QUAKE-corrected read-data sets and the previously obtained data sets: Illumina GAIIx 2 × 50 bp mate-pair reads with insert size around 5,200 bp, and 454 FLX (single reads with average length of 523 bp) technologies ([@evu254-B72]) were then all processed using fastq-mcf tool ([@evu254-B5]), with which a quality filter was applied on all reads based on Phred quality scores ([@evu254-B20]) and eventual adapter contamination, low quality, and ambiguous nucleotides were trimmed off from the remaining reads. A comparison of the overall quality of the resulting corrected and filtered data sets with the original data sets was carried out using FastQC ([@evu254-B2]). Paired-end reads from MiSeq were de novo assembled using Edena version 3.131028 ([@evu254-B34]) and Spades version 2.5.1 ([@evu254-B6]) using different parameterization. To analyze the assembly quality, internal SNVs and indels were estimated from reference assembly mapping of the filtered Illumina GAIIx paired-end reads against the respective de novo genome assemblies, using bowtie2 ([@evu254-B45]). Samtools version 0.1.19 ([@evu254-B46]) and SNVer version 0.5.1 ([@evu254-B82]) were used to convert .SAM files to sorted .BAM files and detection of variant nucleotides, respectively. The tuned assemblies from Edena and Spades assemblers were merged using GAM-NGS assembly merger ([@evu254-B80]), followed by automatically improvement of the resulting de novo genome assembly, using Pilon version 1.6 (<http://www.broadinstitute.org/software/pilon/>, last accessed October 1, 2014). Afterwards, filtered data sets from Illumina MiSeq and GAIIx (both) were used for scaffolding with SSPACE version 2.3 ([@evu254-B10]) and gap closure by using Gapfiller version 1.5 ([@evu254-B11]) and GapCloser version 1.12 ([@evu254-B50]). The final genome sequence draft of *Pseudomonas* sp. M1 was obtained following automatic improvement using Pilon, mapping of Sanger sequencing reads of regions evidencing abnormal coverage and/or gaps, mapping of 454 Roche FLX data set and manual curation. Genome annotation was carried out using National Center for Biotechnology Information (NCBI)'s Prokaryotic Genomes Annotation Pipeline 2.0 ([@evu254-B3]) and metabolic pathways were predicted in silico by KAAS ([@evu254-B54]). *Total RNA Extraction from Pseudomonas*sp. *M1* ----------------------------------------------- Total RNA samples were obtained using the Aurum Total RNA Mini kit (Bio-Rad), as described by the manufacturer's instructions. Prior to RNA sequencing (RNA-seq) and reverse transcription-quantitative polymerase chain reaction (RT-qPCR) experiments, all genomic DNA remaining in RNA samples was digested with RNA-free DNase I. RNA integrity was verified in the Agilent 2100 Bioanalyzer (Agilent Technologies, CA) and only samples with RNA Integrity Number over 8.0 were used in RNA-seq and RT-qPCR experiments. Differential Expression Analysis of RNA-Seq Samples from M1 Cells Grown in Lactate or β-Myrcene ----------------------------------------------------------------------------------------------- Total RNA samples extracted from M1 cells were processed according to Illumina instructions generating libraries of 50-bp single reads, which were sequenced by Illumina GAIIx platform at BaseClear (the Netherlands), resulting in three data sets from M1 cells grown in lactate medium and three data sets from M1 cells grown in β-myrcene medium, which ranged from 20 million to 56 million reads. Raw reads were trimmed and clipped for adaptor sequences and ambiguous nucleotides (N's) using fastq-mcf tool ([@evu254-B5]). Expression levels were estimated by mapping and counting filtered RNA-seq data sets against the M1 draft genome using EDGE-pro pipeline ([@evu254-B51]). Differential expression levels were determined as fold change (FC) by the ratio of normalized β-myrcene-derived read count values per lactate values using the EdgeR Exact Test ([@evu254-B63]) with false discovery rate (FDR) *P* value correction. Genes were defined as significantly differentially expressed if they evidenced an altered FC ratio of 1.5 (or higher) and FDR smaller than 0.05. Genes having less than five reads mapped were considered to not be expressed. *Characterization of the 28-kb Genomic Island Harboring β-Myrcene Core-Code* In silico prediction of the transcriptional units (TUs), promoter and terminator regions was carried out by FGENESB tool from Softberry server (<http://www.softberry.com/>, last accessed October 1, 2014). RNA-seq based prediction of operons and transcription start sites (TSS) within this genomic island (GI) was carried out using ReadXplorer ([@evu254-B36]). Quantitative Expression of β-Myrcene-Biotransforming Key Genes by RT-qPCR ------------------------------------------------------------------------- For RT-qPCR experiments, reverse transcription of RNA to cDNA was performed with the iScript cDNA Synthesis Kit (Bio-Rad). A total of 500 ng of RNA (from the same batch used for RNA-seq) was used for each sample and cDNA synthesized according to manufacturer's instructions (5 min at 25 °C, 30 min at 40 °C, and 5 min at 85 °C). Expression of *PM1_0216305*, *PM1_0216325*, *PM1_0216335*, *PM1_0216350*, *myrA* (*PM1_0216395*), *myrD* (*PM1_0216400*), and *recA* (*PM1_0217885*) was analyzed in a CFX96 Touch Real-Time PCR Detection System (Bio-Rad) using SensiFAST SYBR No-ROX Kit (Bioline) as the detecting agent. Primers used were synthesized by Stab Vida with following sequences: *PM1_0216305* (fwd: 5′-TATCGCAGCTTGGTCTTGGC-3′; rev: 5′-CTAACCAGTCACCAGACGCC-3′), *PM1_0216325* (fwd: 5′-AAAGGACTGGGCAAACTGGG-3′; rev: 5′-GCTGACGGCTTTGTCTTTGG-3′), *PM1_0216335* (fwd: 5′-CCGGATTACCGAGCTTCTGG-3′; rev: 5′-GGTGTTGGCTGTGTCAGGC-3′), *PM1_0216350* (fwd: 5′-TGTTCTCAAGGTCGCTGACG-3′; rev: 5′-TCCTGCAAATGCTTCAGGGG-3′), *myrA* (fwd: 5′-CGATACCGCTCTGCAACTCC-3′; rev: 5′-CCGCTGCGATGATTTTGTCG-3′), *myrD* (fwd: 5′-GGCGACTTACAGCGTTTTGC-3′; rev: 5′-GCTGGAGTGGTATGCAGAGC-3′), and *recA* (fwd: 5′-GGCCGAGTTCCAGATCCTCTAC-3′; rev: 5′-GATTTCCGCGTTCTCTTCCAGG-3′). Efficiency of each pair of primers (E) was determined with cDNA dilution curves (from 1 μg to 10 pg) from M1 cultures grown in lactate or β-myrcene, using the equation E = 10^(−1/slope)^ − 1 ([@evu254-B65]). Primers used showed efficiencies between 90% and 110%. Each reaction was set up in a total volume of 20 μl, containing 100 ng of cDNA, SensiFAST SYBR No-ROX mix (Bio-Rad), 2 pmol of each gene-specific primer, and nuclease-free water. Amplification was performed at 95 °C for 2 min followed by 40 cycles at 95 °C for 5 s, 55 °C for 10 s, and 72 °C for 20 s. Specificity of primer annealing and presence of amplification artifacts were assessed by the dissociation temperatures of amplification products in a melting curve analysis of temperatures between 65 °C and 95 °C (stepwise increase of 0.5 °C/5 s), with a continuous measurement of fluorescence. Negative control without cDNA template and control of genomic DNA contamination in RNA samples were performed for each pair of primers and each batch of RNA extracted. Eight independent RNA samples were used and all reactions were performed in triplicated. Quantification cycle values (Cq) were determined with Bio-Rad CFX Manager software as the threshold cycle. Relative quantification of gene expression as FC was calculated by the 2^−ΔΔCq^ method taking into account primer efficiencies, as described by [@evu254-B58] and using *recA* (*PM1_02298*) as reference gene for normalization ([@evu254-B52]; [@evu254-B29]). Error propagation was calculated according to [@evu254-B57]. Phylogeny of Genes from the 28-kb GI ------------------------------------ Phylogenetic analysis was carried for all 22 genes comprised in the 28-kb GI and for the *GMP synthase* (*guaA*, *PM1_216145*). Each gene was compared with sequences deposited in GenBank by BLASTn ([@evu254-B1]) and the 20 most homologous entries were chosen for multialignment with MAFFT v7 ([@evu254-B42]). Using these alignments, maximum likelihood (ML) trees were inferred with PHYML v3.0 ([@evu254-B30]), using aproximate likelihood-ratio test nonparametric branch support based on a Shimodaira--Hasegawa-like procedure. For all the genes of the GI, the preferred model of nucleotide substitution was GTR (General Time Reversible) with a proportion of invariable sites. Similarities/dissimilarities between phylogenetic tree topologies may provide important clues on the relatedness of the evolutionary track of each gene of the GI under analysis in this work. Therefore, the phylogenetic tree-based parameters Log-likelihood score and patristic distance between multialigned sequences for each gene were deduced from the phylogenetic tree reconstruction and used to build a multivariate matrix. Hierarchical clustering analysis of the resulting matrix was carried by estimating the squared Euclidean distance between the different objects (gene alignment tree parameters), followed by agglomerative clustering with UPGMA (Unweighted Pair Group Method with Arithmetic Mean) method, using R statistical software with the gplots package ([@evu254-B81]). The number of clusters was estimated by partitioning around mediods clustering algorithm, using the function pamk() in the R package fpc ([@evu254-B33]). Results and Discussion ====================== Whole-Genome Sequencing and Assembly ------------------------------------ Previous work describing key players of β-myrcene catabolism in *Pseudomonas* sp. M1 ([@evu254-B40]; [@evu254-B68]) was limited by the absence of reference genomic data. To set the proper background for exploring the biotransformation potential of *Pseudomonas* sp. M1, its genome was resequenced with Illumina MiSeq technology and assembled together with sequencing data sets obtained previously with Illumina GAIIx and 454 Roche FLX technologies ([@evu254-B72]). The assembly pipeline is summarized in [figure 1](#evu254-F1){ref-type="fig"}. F[ig]{.smallcaps}. 1.---Schematic representation summarizing the custom assembly pipeline used to reconstruct the genome of *Pseudomonas* sp. M1. Gray arrows indicate the pipeline stepwise flow. The differently colored lines associate each sequencing dataset with the phase(s) of the pipeline in which they were used. MiSeq-sequencing yielded 5,431,166 paired-end reads of 250 bp. Raw reads were then corrected for sequencing errors, trimmed, quality-filtered, and approximately 70% of the original data set (3,847,594 reads) was de novo assembled using Edena ([@evu254-B34]) and Spades ([@evu254-B6]). Assembly-quality and number of SNVs and indels were assessed by mapping-filtered Illumina GAIIx paired-end data set (4,052,990 reads) against de novo draft genome assemblies using SNver ([@evu254-B82]). Merging Spades and Edena draft genomes with GAM-NGS assembly merger ([@evu254-B80]) and improvement with Pilon (<http://www.broadinstitute.org/software/pilon/>, last accessed October 1, 2014) yielded 52 contigs (longest contig size of 506,988 bp and N50 of 252,542 bp) with 9 SNVs and 12 indels. Scaffolding and gap closure were performed using the filtered data sets from Illumina MiSeq (3,847,594 paired-end reads) and GAIIx (4,052,990 paired-end reads and 4,698,623 mate-pair reads), which resulted in 20 contigs organized in two scaffolds with 9,303 unknown nucleotides (N's). After improvement using Pilon, mapping Sanger sequencing reads of regions evidencing abnormal coverage and/or gaps, mapping of the filtered 454 Roche FLX data set (253,532 single reads) and manual curation, the final improved genome draft of *Pseudomonas* sp. M1 was obtained in 4 contigs organized in 1 scaffold, with 21 unknown nucleotides (N's), 7 SNVs, 5 indels, and a total size of 6,982,449 bp. The resulting sequence was annotated with PGAP (Prokaryotic Genome Annotation Pipeline, NCBI) and deposited at DDBJ/EMBL/GenBank (accession number: ANIR00000000.2). The improved draft of *Pseudomonas* sp. M1 genome includes 6,163 genes (6,023 coding sequences, 58 pseudogenes, 5 copies of each 5 S, 16 S, 23 S ribosomal RNA \[rRNA\] genes, 66 transfer RNA \[tRNA\] genes, and 1 noncoding RNA), with an average GC content of 67.2%. The major improvements of the current draft, when compared with the previous version, are reflected in the number of annotated genes and, importantly, in the reliability of base call along the genomic sequence, as deduced from the very low number of internal SNVs. Such improvement may have a significant impact on downstream functional analyses such as reference genome read mapping and protein identification. The genome of *Pseudomonas* sp. M1 comprises a large set of oxygenases/hydroxylases-coding genes (99). Several enzymes belonging to this family have been shown to play an essential role on the initial oxidation steps of several natural and anthropogenic recalcitrant compounds. In silico prediction of M1 versatile metabolism by KAAS ([@evu254-B54]) suggested that the genomic repertoire of *Pseudomonas* sp. M1 may include the ability to engage in biotransformation/mineralization of recalcitrant compounds such as cresol and xylene, styrene, chloroalkanes and chloroalkenes, polycyclic aromatic compounds, naphtalene, benzoates, bisphenol, and atrazine ([supplementary table S1](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online). Additionally, M1 nutritional versatility is probably prompted by respiratory flexibility in using N-oxides as alternative electron donors, as suggested by the presence of denitrifying clusters in the genome. Such genomic features enhance the potential application of M1 for bioremediation at contaminated oxygen-depleted foci and as a biosensor for in situ monitorization of aliphatic and aromatic hydrocarbons. Genome sequencing also allowed identification of several genetic features putatively involved in the metabolism of plant-derived volatiles other than β-myrcene ([supplementary table S1](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online), such as the acyclic terpene utilization cluster "atuRABCDEFGH" (*PM1_0225865--PM1_0225910*) that is described to be involved in citronellol and related terpenes catabolism in *Pseudomonas citronellolis* ([@evu254-B23]) and *Pseudomonas aeruginosa* strain PAO1 ([@evu254-B21]), and some oxygenases and dehydrogenases putatively involved in the metabolism of limonene- and alpha-pinene-related terpenes. Thus, *Pseudomonas* sp. M1 seems to harbor an attractive set of enzymes for biocatalysis of some high-value terpenoids. RNA-Seq of M1 Cells Challenged with β-Myrcene --------------------------------------------- Identification of genes involved in β-myrcene-stimulon was performed by comparative RNA-seq analysis of the transcriptomes of M1 cells grown in lactate (L) (control) and β-myrcene (M). Raw reads were trimmed and clipped for adaptor sequences and ambiguous nucleotides (N's) and mapped to a pool of rRNA and tRNA gene sequences from M1 draft genome to filter RNA-seq data sets exclusively for mRNA reads. To assess expression levels, mRNA reads were 99% uniquely mapped against M1 draft genome using EDGE-pro pipeline ([@evu254-B51]) and the Bioconductor package EdgeR ([@evu254-B63]). Among the large number of currently available tools for RNAS-seq analysis, this strategy was chosen based on: 1) EDGE-pro has been designed toward analysis of bacterial RNA-seq samples (in contrast with most of other read count applications which are directed toward the analysis of eukaryotic samples), and 2) EdgeR has been reported as one of the most used and reliable methods for data set normalization and statistical inference ([@evu254-B19]). As shown in [table 1](#evu254-T1){ref-type="table"} (and more detailed in [supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online), the comparative transcriptome analysis revealed a number of different functions being altered when cells are cultivated with β-myrcene. In M1 challenged cells, 627 genes were found to be upregulated at least 1.5-fold (all evidencing an FDR value lower than 0.05) and 1,246 genes were found to be downregulated at least 1.5-fold (all evidencing an FDR value of 0.05 or lower), suggesting an extensive molecular reprogramming of cell physiology. Globally, the cultivation of M1 cells in β-myrcene induced a change in the expression levels of over 30% of the genes in its genomic repertoire. Table 1Summary of Differential Expressed Genes in Myrcene-Grown *Pseudomonas* sp. M1 Cells, Organized in COG CategoriesFunctional CategoriesGenomeRNA-SeqNumber of GenesUp ≥ 1.5-FoldDown ≥ 1.5-Fold\[A\] RNA processing and modification110\[B\] Chromatin structure and dynamics502\[C\] Energy production and conversion35962108\[D\] Cell cycle control, cell division, chromosome partitioning3940\[E\] Amino acid transport and metabolism54333202\[F\] Nucleotide transport and metabolism100615\[G\] Carbohydrate transport and metabolism2111076\[H\] Coenzyme transport and metabolism1892635\[I\] Lipid transport and metabolism2435167\[J\] Translation, ribosomal structure and biogenesis1841215\[K\] Transcription4624463\[L\] Replication, recombination and repair14487\[M\] Cell wall/membrane/envelope biogenesis2252540\[N\] Cell motility178650\[O\] Posttranslational modification, protein turnover, chaperones1722119\[P\] Inorganic ion transport and metabolism2923590\[Q\] Secondary metabolites biosynthesis, transport and catabolism1601851\[R\] General function prediction only49040101\[S\] Function unknown4204645\[T\] Signal transduction mechanisms3012152\[U\] Intracellular trafficking, secretion, and vesicular transport1642241\[V\] Defense mechanisms621317\[W\] Extracellular structures101Total genes with COGs[^a^](#evu254-TF1){ref-type="table-fn"}4,463469974No COG assigned1,700158272Total6,1636271,246[^3] Having in mind the assigned Clusters of Orthologous Groups (COG) categories, the most extensive alterations were registered in genes assigned to energy production and conversion; amino acid transport and metabolism; carbohydrate transport and metabolism; lipid transport and metabolism; transcription; inorganic ion transport and metabolism; and secondary metabolites biosynthesis, transport and catabolism. Importantly, β-myrcene metabolism seems to be dependent on expression of specific genetic loci, as deduced by a registered FC ratio greater than 10 for 52 of the upregulated genes. *A Unique 28-kb Genomic Island Harbors the* myr+ *Genetic Trait* ---------------------------------------------------------------- β-myrcene supplementation induced the expression of 22 genes of a 28-kb genomic locus (*PM1_0216305--PM1_0216410*), from which 15 registered the highest expression levels from all β-myrcene-dependent transcriptome ([fig. 2](#evu254-F2){ref-type="fig"} and [supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online). Interestingly, the *myrDABC* cluster (*PM1_0216385--PM1_0216400*), previously identified by [@evu254-B40] as essential for the initial steps of β-myrcene catabolism, is located in this 28-kb locus. Elements involved in conjugative transfer were found to be flanking this locus, which has an average GC content of 61.7% (genome average is 67.2%), indicating a hypothetical foreign nature for this genomic region. F[ig]{.smallcaps}. 2.---Global organization and functionality of the newly identified 28-kb GI harboring β-myrcene core-code. (*A*) Genomic locus organization and functional elements detected using in silico prediction. Genes transcribed in forward orientation are represented by rightward arrows, whereas genes located in antisense strand are represented by leftward arrows. Genes are organized in TU, promoter regions (P) are illustrated by blue arrows, and terminator sites indicated by red circles. (*B*) Detailed analysis of the RNA-seq data mapping with the GI. In the top panel, read mapping coverage per base was plotted for samples derived from M1 growth in lactate (gray line) and in β-myrcene (blue line). Prediction of the operon-like transcripts (horizontal orange bars) and transcription starting sites (TSS) (red arrows) was carried out with ReadXplorer ([@evu254-B36]). Expression levels for each gene are shown, in the bottom panel, as the ratio (FC) of normalized transcripts between cells grown in lactate (L) and cells grown in β-myrcene (M). To characterize this GI, TU were defined in silico by predicting operon-like organization of genes from the GI, the respective promoter regions and terminator sequences. The 22 genes seem to be organized in eight TUs: TU1, TU2, TU3, TU4, TU5, TU6, TU7, and TU8 ([fig. 2](#evu254-F2){ref-type="fig"}*A*). Transcriptional organization was validated by mapping RNA-seq reads and de novo assembled contigs against the 28-kb GI, suggesting the polycistronic transcription of five of the eight TUs ([fig. 2](#evu254-F2){ref-type="fig"}*B*). Moreover, RNA-seq based analysis of this GI also allowed the prediction of seven TSS, six of which overlapping putative promoter regions deduced from in silico analysis ([fig. 2](#evu254-F2){ref-type="fig"}*A*). The detection of two TSS within TU3 may suggest that this TU may actually lead to different transcripts, depending on the growth conditions. The first predicted TU comprises two acyl-CoA dehydrogenases (*PM1_0216305* and *PM1_0216310*) and an acetyl-CoA acetyltransferase*,* (*PM1_0216315*) putatively involved in β-myrcene catabolism, whose expression was increased approximately 365-, 370-, and 160-fold in β-myrcene-grown cells. A second cluster comprises two genes coding for a putative chemotaxis sensory transducer protein and a putative membrane protein (*PM1_0216320* and *PM1_0216325*). In the presence of β-myrcene, expression of this TU increased over 15-fold. From all 172 genes annotated on M1 genome as being involved in chemosensory systems and in signal transduction, 42 genes registered altered expression of at least 2-fold, from which only *PM1_0216320* and other signal protein-coding gene (*PM1_0226790*) were found to be upregulated in β-myrcene grown-cells, suggesting that TU2, together with *PM1_0226790*, might play essential role in recognition, transduction of the initial signal triggered by β-myrcene stimulus and possible uptake of β-myrcene (or derivative) molecules. OmpW family porins have been described to transport hydrophobic molecules across the membrane in some *Pseudomonas* spp. and *Escherichia coli* ([@evu254-B38]; [@evu254-B56]; [@evu254-B76]). Moreover, [@evu254-B56] showed that mutation in OmpW of *Pseudomonas fluorescens* resulted in loss of ability to catabolize polycyclic aromatic molecules, such as naphtalene. TU3 codes for six products that might also be involved in β-myrcene oxidation: Another acetyl-CoA acetyltransferase (*PM1_0216330*) with increased expression of approximately 320-fold, a monooxygenase (*PM1_0216335*) increased by 431-fold, an enoyl-CoA hydratase (*PM1_0216340*) increased by 412-fold, an acyl-CoA dehydrogenase family protein (*PM1_0216345*) upregulated by 263-fold, an epoxide hydrolase (*PM1_0216350*) overexpressed by 102-fold, and a hypothetical protein (*PM1_0216355*) with increased expression by 51-fold. Moreover, TU4 is composed by a single gene, coding for a putative LuxR-family transcriptional regulator (*PM1_0216360*). This putative regulator registered increased expression of 190-fold in β-myrcene-grown cells. The fifth transcriptional unit is composed by two genes which code a rubredoxin (*PM1_0216365*), whose expression was increased by 260-fold, and a fatty acid desaturase (*PM1_0216370*), whose expression was increased by 202-fold. The sixth TU, comprises five genes which code for another LuxR-family transcriptional regulator (*PM1_0216375*), an oxidoreductase (*PM1_0216380*), the acyl-CoA synthase *myrC* (*PM1_0216385*), the alcohol dehydrogenase *myrB* (*PM1_0216390*), and the aldehyde dehydrogenase *myrA* (*PM1_0216395*). Expression levels were upregulated by 5-fold, 71-fold, 129-fold, 86-fold, and 44-fold, respectively. Gene coding for the enoyl-CoA hydratase *myrD* (*PM1_0216400*) comprises a seventh transcriptional unit and registered an increased expression of 220-fold. The eighth transcriptional unit predicted is composed by two genes which code for a putative LysR-family transcriptional regulator (*PM1_0216405*) and a hypothetical protein (*PM1_0216410*), whose expression was increased by 95- and 70-fold, respectively. The high expression of such set of catabolic enzymes strongly suggests that this GI is the main genetic trait that provides *myr +*phenotype to *Pseudomonas* sp. M1. Particularly, the monooxygenase-coding gene *PM1_0216335* showed the highest expression registered in all transcriptome and might be involved in initial oxidation of β-myrcene into a more soluble substrate. The understanding of its substrate range would be of utmost interest for terpene-based industry and future fine-tuning of its activity for biotechnological applications. Moreover, the differential expressed genes encoding for putative regulators (*PM1_0216360, PM1_0216405*, and *PM1_0216375*) suggest that they may be key players in the modulation of expression of the catabolic enzymes encoded within the GI. Expression of predicted TUs was validated by RT-qPCR according to Pfaffl ([@evu254-B58]) and using *recA* (*PM1_02298*) as reference gene for normalization ([@evu254-B52]; [@evu254-B29]) ([fig. 3](#evu254-F3){ref-type="fig"}). Similar to expression levels measured by RNA-seq, key genes *PM1_0216305, PM1_0216325*, *PM1_0216335*, *PM1_0216350*, *myrA,* and *myrD* were highly expressed in M1 cells challenged with β-myrcene, as showed by the FC ratio in [figure 3](#evu254-F3){ref-type="fig"}. In general, the FC ratios obtained by RT-qPCR were higher when compared with RNA-seq-based estimated FC ratios. Although both methods have been proved to provide significant information regarding the evaluation of mRNA levels among different conditions, they rely on different methodological approaches which can differently affect the estimation of FC ratios (e.g., library preparation, data normalization, amount of starting template, and template amplification procedures). Nonetheless, both approaches clearly show that the expression of the selected genes is strongly stimulated by β-myrcene. F[ig]{.smallcaps}. 3.---Expression levels of key genes from the 28-kb GI detected by RNA-seq and validated by RT-qPCR, shown as the ratio of normalized transcripts between cells grown in lactate and cells grown in β-myrcene (FC). The 28-kb GI Evolved from the Assembly of Functional Blocks ----------------------------------------------------------- As the GC content percentage within this GI has an average of 61.7% which is significantly different of the whole-genome average (67.2%), it was hypothesized that this GI may have been acquired as a whole DNA sequence segment through horizontal gene transfer. However, a thorough inspection of the public databases for DNA sequence deposition indicated that this GI was not present in any records besides the ones related with *Pseudomonas* sp. M1. In fact, the closest homologs of each predicted gene of the GI indicated a wide range of percentage of identity variation ranging from 56% to 87% (for the top hits), *PM1_261405* evidencing a closest similarity with the matched homologs (87%) and *PM1_216375* evidencing a lowest similarity (56%). To obtain clues on the origin of this GI, a phylogenetic analysis of all the genes of the 28-kb GI identified in this work was carried out. In this context, and due to the registered fairly low identity percentage between the genes of the GI and the respective homologs, only the 20 closest related sequences for all the genes (except *PM1_216355* for which only ten sequences resembling this gene were found) were selected for this analysis. Additionally, the analysis of the *guaA* gene of M1 strain and the respective 20 closest sequences was also performed. This gene was selected as a representative gene of M1 core genome as it evidences a GC content % similar to the average of the whole genome of *Pseudomonas* sp. M1 and it often used in multilocus sequence typing of *Pseudomonas* strains. In [figure 4](#evu254-F4){ref-type="fig"}, and in more detail in [supplementary table S3](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online, the results of the comparison of each gene of GI with their homologs are presented. The comparison of each gene of the GI with the respective 20 homologs indicates 1) a significantly lower average percentage of identity (57--77%) with respect to the 93% of identity for the reference gene *guaA* and 2) a significant difference in the GC% content of M1 genes with respect with the closest homologs (ranging 1--12% difference). These differences are supported, at least partially, by the diversified taxonomic origin of the closest homolog sequences, as depicted in [figure 4](#evu254-F4){ref-type="fig"}. As expected, the results for the representative *guaA* gene are in agreement with previously published phylogenetic analysis of *Pseudomonas* sp. M1 based on 16 S rRNA sequencing ([@evu254-B67]) and groups M1 within the Gammaproteobacteria class (and *Pseudomonas* genus). In contrast, most of the genes of the GI, with the exception of *PM1_216320*, share significant identity with homolog sequences from organisms belonging to a taxonomic class not related with *Pseudomonas* sp. M1. In the case of *PM1_216305*, *PM1_216310*, *PM1_216315*, *PM1_216320*, *PM1_216325*, *PM1_216330*, *PM1_216335*, *PM1_216340*, *PM1_216355*, *PM1_216360*, *PM1_216365*, *PM1_216375*, *PM1_216405*, and *PM1_216410* genes, 75% (or more) of their homolog sequences belong to organisms of a specific taxonomic class. Distinctively, in the case of *PM1_216345*, *PM1_216350*, *PM1_216370*, *PM1_216380*, *PM1_216385*, *PM1_216390*, *PM1_216395,* and *PM1_216400* genes, the taxonomic distribution of their homolog sequences is spread among different classes. This asymmetric taxonomic distribution of the homolog sequences of the GI may suggest that the genetic reservoirs of some of the genes within the GI are confined to specific taxonomic groups whereas homologs of the other genes are widely distributed across different taxonomic classes. Moreover, this asymmetry also suggests that, probably, the GI resulted from acquisition of DNA segments from different environmental reservoirs which were assembled together evolving toward gene specialization for β-myrcene utilization. This hypothesis is in agreement with the multiple parallel evolutionary origins of GIs outlined before ([@evu254-B41]). F[ig]{.smallcaps}. 4.---Graphical representation of the analysis of the results of multialignments of each gene of the GI implicated in β-myrcene catabolism in *Pseudomonas* sp. M1 and PM1_216145 (*guaA*), and the respective most homolog genes, retrieved using BLASTn. The dotted line indicates the average percentage of identity with respect to M1 genes. The average of GC percentage of each M1 gene or of the homologs sequences is represented by blue or red lines, respectively. The taxonomic origin of each homolog sequence was assessed and grouped according the respective taxonomic class. The bars in graph represent relative distribution (in percentage of the total number homolog sequences) of Alphaproteobacteria (green), Actinobacteria (light blue), Betaproteobacteria (red), Gammaproteobacteria (dark blue), and other classes (light orange). Each gene of the GI, the *guaA,* and the most the selected most homolog sequences (analyzed above) were aligned using MAFFT. The resulting multialignments were used to construct a phylogenetic tree for each gene (GI and *guaA*) using an ML method. The respective −Log likelihood (−ln *L*) scores and divergence distance with respect to each *Pseudomonas* sp. M1 gene (estimated as the average of the patristic distances between tree tips and the respective M1 gene) were extracted. The −ln *L* score reflects the overall phylogenetic tree topology and branch lengths as a consequence of the relatedness of the sequences under analysis, being higher values an indication of higher phylogenetic distances. In fact, in this study it was verified that, in general, the −ln *L* increased when the number of taxonomic genus represented in the multialignment sequences increased (data not shown). Presumably, M1 genes evidencing closer −ln *L* scores and divergence distances, based on the comparison with the selected homolog sequences, may share more similar evolutionary histories. Therefore, a multivariable matrix including the −ln *L* scores and divergence distances for all the genes of the GI and *guaA* was constructed. In [figure 5](#evu254-F5){ref-type="fig"}, the agglomerative clustering analysis of the resulting matrix is shown. As expected from the results presented in [figure 4](#evu254-F4){ref-type="fig"}, the reference gene *guaA* clustered distantly from the genes within the GI as the homolog sequences share a higher degree of identity. Remarkably, this analysis suggests that most of the genes within TU3 and TU6 ([fig. 2](#evu254-F2){ref-type="fig"}) may share a close evolutionary history as suggested by their inclusion in the same cluster partition, C9 and C6 of [figure 5](#evu254-F5){ref-type="fig"}, respectively. Additionally, *PM1_216320/PM1_216325* and *PM1_216405/PM1_216410* pairs of genes cluster close to each other and correspond to the predicted TU2 and TU8, respectively. These observations are in agreement with the correlation between coexpression and coevolution of genes encoding enzymes with coupled metabolic fluxes previously reported (Seshasayee et al. 2009). Moreover, *PM1_216320, PM1_216360,* and *PM1_216375* genes, that putatively code for three of the four transcriptional regulators present in the GI, cluster closer (clusters C1, C2, and C3 of [fig. 5](#evu254-F5){ref-type="fig"}, respectively) and apart from the catabolic genes. Significantly, the highest values for the −ln *L* and/or average patristic distance parameters were registered for these genes. Our results may indicate that the control of the catabolic genes of the GI requires an intricate tuning of the regulatory sequences towards an effective responsiveness to β-myrcene (and possibly other terpenes), probably due to the unusual nature of the substrates. Thus, it suggests that the regulatory and the catabolic components of the GI may have evolved separately. This hypothesis is in agreement with a modular assembly concept of emerging catabolic pathways that includes evolution from constitutive to regulated expression of the catabolic enzymes, as previously described ([@evu254-B14]). F[ig]{.smallcaps}. 5.---Multivariate clustering analysis of the inferred phylogenetic tree-based parameters −log likelihood score (−ln *L*) and average patristic distance (Divergence) for each gene of the GI. For a comparison purpose, the parameters for the representative gene *guaA* were also included in the analysis. Each M1 gene, represented in the figure, was multialigned together with the most homolog sequences and clustering analysis of the inferred parameters was carried by estimating the squared Euclidean distance between the different sequences in each multialignment, followed by agglomerative clustering with UPGMA method. The number of clusters was estimated by partitioning around mediods clustering algorithm. Each cluster is outlined by a different color. β-Myrcene-Induced Genetic Repertoire Might Code for Enzymatic Redundancy toward a Broader Range of Terpene-Substrates --------------------------------------------------------------------------------------------------------------------- Transcriptome data not only support the pathway described by [@evu254-B40] but also strongly suggest the involvement of several other gene products in β-myrcene oxidation and raise new hypotheses about β-myrcene metabolic pathway that will be the focus of following studies. At this moment, it is unclear whether all 22 genes from the 28-kb GI are actually necessary for the metabolism of the aliphatic C8-chain of β-myrcene or whether the high expression levels of the island are induced by a chemosensory signal common to a broader metabolic potential toward several terpene-backbone compounds. The expression of the putative epoxide hydrolase (*PM1_0216350*) when *Pseudomonas* sp. M1 is grown in β-myrcene medium may indicate a possible epoxidation of β-myrcene for subsequent hydrolysis by PM1_0216350 into a β-myrcene alcohol. However, terpene-epoxide microbial biotransformation has been reported mainly as result of monooxygenase activity toward cyclic molecules (e.g., alpha-pinene and limonene) ([@evu254-B79]). Therefore, it is also possible that the putative epoxide hydrolase encoded by the GI may be actually involved in epoxidation of substrates other than β-myrcene. β-myrcene biotransformation has been reported in other gram-negative bacteria. *Rhodococcus erythropolis* MLT1 was described to biotransform aerobically β-myrcene into geraniol ([@evu254-B75]). In *Castellaniella defragans* 65Phen β-myrcene is hydratated to linalool, following isomerization to geraniol and oxidation to geranic acid, under anaerobic conditions ([@evu254-B12]; [@evu254-B48]; [@evu254-B47]; [@evu254-B49]), which indicates a cross-talk between catabolic pathways of different terpenes. In M1 cells challenged with β-myrcene, the majority of genes from the acyclic terpene utilization cluster *atuRABCDEFGH* and the leucine/isovalerate degradation cluster *liuRABCDE* (*PM1_0223355*--*PM1_0223380*) were found to be upregulated (average cluster expression induced by 3.8-and 4.1-fold, respectively). The overexpression of these two clusters suggests that, similarly to *R. erythropolis* and *C. defragans* ([@evu254-B21]; [@evu254-B23]), a citronellol/geraniol-like compound might be an intermediate/byproduct of β-myrcene metabolism and channel β-myrcene-derivatives through propionyl-CoA into central metabolism ([fig. 6](#evu254-F6){ref-type="fig"}). Another hypothesis is that the sensory system present in the GI may be involved in the recognition of a spectrum of terpene-backbone molecules, including β-myrcene, and trigger expression of several genetic loci with complementary/redundant degradative functions ([fig. 6](#evu254-F6){ref-type="fig"}). F[ig]{.smallcaps}. 6.---Schematic prediction of β-myrcene metabolic pathway and redistribution of central metabolic fluxes in M1 cells challenged with β-myrcene. Upregulated genes are shown by upward black (\>2-fold) or gray (\<2-fold) arrows and downregulated genes are shown by downward black (\>2-fold) or gray (\<2-fold) arrows. Genes with similar expression levels in *Pseudomonas* sp. M1 cells grown in lactate medium and β-myrcene medium are indicated by white squares. Several other genes located outside of the 28-kb locus and coding for beta-oxidation-like enzymes had their expression altered in M1 cells grown in β-myrcene medium. As shown in [supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online, increased expression by at least 2-fold was detected for 8 acyl-CoA-like dehydrogenases (*PM1_0206520, PM1_0208500, PM1_0211935, PM1_0220850, PM1_0221270, PM1_0224730, PM1_0224740,* and *PM1_0224920*), a putative acetyl-CoA acetyltransferase (*PM1_0224735*), a putative acetyl-CoA synthetase (*PM1_0208785*), 2 putative acyl-CoA thiolases (*PM1_0221265* and *PM1_0212440*), 2 putative enoyl-CoA hydratases (*PM1_0221070* and *PM1_0221075*), a putative enoyl-CoA reductase (*PM1_0225900*), and 2 putative aldehyde dehydrogenases (*PM1_0226785* and *PM1_0230170*). In β-myrcene-grown *C. defragans* and *R. erythropolis*, overexpression of additional acyl-CoA dehydrogenases and oxidoreductases was detected ([@evu254-B75]; [@evu254-B49]). Furthermore, kinetic assays showed that geraniol dehydrogenase GeoA of *C. defragans* was able to catalyze either cyclic (e.g., perillyl-alcohol) and acyclic monoterpenes (e.g., geraniol) ([@evu254-B47]), and its deletion did not impair completely growth on β-myrcene medium, suggesting the existence of secondary enzyme system acting on acyclic terpene-backbone molecules ([@evu254-B47]). Similarly, mutagenesis trials in *Pseudomonas* sp. M1 (e.g., *ΔmyrB* mutant) did not abolish completely the ability of M1 to grow in β-myrcene medium (data not shown), supporting the redundancy of beta-oxidation-like enzymes overexpressed during growth in β-myrcene medium and suggesting a cross-talk between metabolic fluxes of terpene catabolism and a complementary substrate range for these set of enzymes ([fig. 6](#evu254-F6){ref-type="fig"}). Central Metabolic Flux Is Changed in β-Myrcene-Grown Cells ---------------------------------------------------------- Concomitant with the upregulation of the 28-kb, expression of almost all genes coding enzymes of the Entner--Doudoroff pathway was downregulated ([supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online). On the other hand, genes coding for enzymes of the tricarboxylic acid cycle (TCA) show similar expression levels in M1 cells grown in lactate or β-myrcene ([fig. 6](#evu254-F6){ref-type="fig"} and [supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online), with exception of the citrate conversion to isocitrate as the *acnA* gene (*PM1_0223060*) coding for aconitate hydratase registered 2.4-fold increase in expression. Subsequently, isocitrate might be channeled to the glyoxylate shunt, mediated by a 2.6-fold increase in expression for *PM1_0221765* (coding for a isocitrate lyase AceA-like). The glyoxylate bypass might increase levels of succinate for energy production, surpassing the metabolic branching of TCA intermediaries (e.g., 2-oxoglutarate) for amino acid biosynthesis. Also, glyoxylate shunt might increase levels of malate, which is further converted to oxaloacetate, a key branching step in the central metabolism. *Pseudomonas* sp. M1 genome harbors *mqoB* gene (*PM1_0214580*) coding for a malate:quinone oxidoreductase and a gene coding for a putative malate dehydrogenase (*PM1_0229630*), being the former upregulated 2.0-fold in M1 cells grown in β-myrcene. In *P. citronellolis*, *mqoB* expression has been described to be essential for the metabolism of the acyclic terpenes citronellol and citronellic acid ([@evu254-B22]). A possible bypass of isocitrate toward the glyoxylate shunt in β-myrcene-grown cells supports previous proteomic data ([@evu254-B68]), which suggested that TCA flux might be directed toward oxaloacetate synthesis as this intermediate might be used in condensation reaction with a β-myrcene derivative, such as propionyl-CoA ([@evu254-B74]). This would channel the final intermediates of β-myrcene catabolism to the central metabolism through 2-methycitrate cycle (e.g., in form of succinate and pyruvate) ([fig. 6](#evu254-F6){ref-type="fig"}). In fact, the expression of all five genes coding for 2-methycitrate cycle enzymes (*PM1_0215080--PM1_0215100*) was also increased in β-myrcene-grown cells (average upregulation of 10.6-fold). β-Myrcene Catabolism Induces an Anaerobiose-Like Physiology ----------------------------------------------------------- RNA-seq data also suggested that M1 cells may use alternative pathways for energy production during β-myrcene catabolism, resembling an oxygen-depletion-like physiology ([supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online). Expression levels of four clustered genes coding for Dnr-like regulators (*PM1_0227665, PM1_0227670, PM1_0227660,* and *PM1_0227675*) were upregulated 7.7-, 2.5-, 2.1-, and 1.9-fold, respectively. Furthermore, Dnr-like regulators were probably responsible for inducing the transcription of several denitrification operons: Average 2.3-fold induction of the *narK1K2GHJI* (*PM1_0215670--PM1_0215695*), an average 38.5-fold induction of the *norCBD* cluster (*PM1_0227680--PM1_0227690*), an average induction of 6.2-fold for the *nirQOP* cluster (*PM1_0227695--PM1_0227705*), 7.1-fold increase in expression of *nirM* (*PM1_0227710*), and an average induction of 3.3-fold for *nosRZDFYL* cluster (*PM1_0227715--PM1_0227740*), coding the set of reductases and cytochromes necessary for N-oxides respiration ([@evu254-B4]). Moreover, expression of genes from the Anr-dependent operon *arcDABC* (*PM1_0202140--PM1_0202155*), coding for enzymes of the arginine deiminase pathway, was also induced at least 4-fold in β-myrcene-grown M1 cells. This operon is responsible for arginine fermentation with the concomitant formation of ATP ([@evu254-B27]). Arginine fermentation might have been induced at earlier stages as a source of energy production that would be gradually replaced by nitrate respiration as Anr--Dnr transcriptional cascade activated their target genes and as β-myrcene metabolism demanded higher yields of energy production. Catabolism of some terpenes in nitrate-respiring conditions has already been reported for *C. defragrans* (e.g., alpha-pinene, β-myrcene, and limonene) ([@evu254-B25]; [@evu254-B35]; [@evu254-B12]) and *Thauera* spp. (e.g., linalool, menthol, and eucalyptol) ([@evu254-B24]), mainly based on metabolite identification. In this study, a molecular snapshot shed some light in what might be a common mechanism for terpene-backbone metabolism under denitrifying conditions. M1 Cells Reprogram Their Physiology during β-Myrcene Catabolism --------------------------------------------------------------- As described above, β-myrcene stimulus activates specific pathways on *Pseudomonas* sp. M1, which in turn appears to switch off genes that might become a metabolic burden for growing cells. M1 cells grown in β-myrcene show several secondary metabolic pathways and respective transport systems downregulated when compared with cells grown in lactate ([supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online), mainly those related to amino acids biosynthesis (e.g., glutamate, glutamine, serine, histidine, lysine, arginine, proline), utilization of alternative carbon sources (e.g., aromatic compounds, sugars), degradation of urea, amines, and polyamines. The altered expression of genes involved in lipopolysaccharide (LPS) and exopolysaccharide synthesis suggested a reorganization of cell envelope in β-myrcene grown cells. Expression of a cluster harboring the *wzm-wzt-wbpXYZ* genes (*PM1_0228115--PM1_0228170*) ([@evu254-B64]; [@evu254-B28]) involved in LPS biogenesis was downregulated over 1.5-fold, as well as a gene coding for the LptA protein (*PM1_0228970*), an element of the LPS transport system across the outer membrane ([@evu254-B77]), whose expression was downregulated by 1.4-fold. The expression of the main cluster associated with alginate synthesis, *algD-8-44-KEGXLIJFA* (*PM1_0215545--PM1_0215600*) ([@evu254-B15]) was downregulated about 2-fold, suggesting a reduction in the synthesis of this polymer. Changes in outer cell envelope were, apparently, followed by changes in membrane lipid composition, membrane stability and permeability as many genes coding for the synthesis of lipid polar head group, transporters/facilitators, porins and membrane proteins had their expression changed when comparing with lactate grown-cells. In β-myrcene-grown cells, expression of two genes coding for OprD homologous (*PM1_0203585* and *PM1_0218170*) was increased by 70.8-fold and decreased by 10.0-fold, respectively. Moreover, expression of a set of efflux pumps was increased more than 2-fold, including six genes coding for Resistance-Nodulation-Cell Division family efflux systems (*PM1_0222160, PM1_0210745, PM1_0222165, PM1_0217125, PM1_0217120,* and *PM1_0202020*), a gene homologous to fusaric acid-inducible resistance pumps (*PM1_0210750*) and two genes coding for a multidrug resistance transporter system (*PM1_0202030* and *PM1_0202025*) with homology to *emrAB* cluster of *E. coli*. On the other hand, OmpQ-family protein-coding genes *PM1_0230840* and *PM1_0213495* were downregulated to 1.8- and 2.0-fold, respectively; *opdE* (*PM1_0220290*) was downregulated to 1.8-fold; and *opdH* (*PM1_0216660*) expression was reduced to 2.0-fold. Abundance of porins changes in response to nutritional requirements of the cell and control membrane permeability toward stressors ([@evu254-B84]; [@evu254-B73]; [@evu254-B62]). Additionally, efflux systems are usually described to be upregulated in stressful conditions, being responsible for cell homeostasis by exporting reactive and toxic compounds ([@evu254-B26]; [@evu254-B60]; [@evu254-B39]). Furthermore, such cell envelope dynamics might account for the adaptive mechanisms toward β-myrcene hydrophobicity, which might affect overall membrane integrity and induce cellular stress ([@evu254-B71]; [@evu254-B78]; [@evu254-B16]). Increased membrane hydrophobicity was previously described in *Pseudomonas* strains challenged by other hydrophobic compounds, such as n-alkanols and phenanthrene, as a mechanism to enhance biotransformation ([@evu254-B32]; [@evu254-B7]). Overall, M1 cells grown in β-myrcene, apparently, leads to a reorganization of membrane structure and functions, most probably to cope with the highly hydrophobic character of β-myrcene and eventual generation of β-myrcene-derived toxic intermediates, as previously evidenced ([@evu254-B68]). Additionally, changes in M1 cell envelope structure and functions may account for a strategy to increase substrate availability/contact surface at the oil--water interface by changing cell surface polarity and control levels of stress by changing membrane permeability and stiffness. Detailed characterization of expression kinetics and regulation are required to fully understand the β-myrcene-stimulon. The sensory system of β-myrcene shows high sensitivity in detecting β-myrcene highly hydrophobic molecules, being an interesting candidate for biosensory applications. Currently, three enzymes, the monooxygenase PM1_0216335, the epoxide hydrolase PM1_0216350 and the 2-methylcitrate synthase PrpC (PM1_0215095), are in protein crystallization trials (data not shown) for future fine-tuning and source of novel biomolecules. Nevertheless, key players involved in the initial steps of β-myrcene metabolism by *Pseudomonas* sp. M1 and the range of plant-derived volatiles that may be biotransformed by the 28-kb GI, as a functional block and by M1 as a cell-factory, still remain to be elucidated and will be focus of following work. Supplementary Material ====================== [Supplementary tables S1--S3](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) are available at *Genome Biology and Evolution* online (<http://www.gbe.oxfordjournals.org/>). ###### Supplementary Data This work was supported by FEDER through POFC---COMPETE and by national funds from Foundation for Science and Technology (Portugal) through the projects PEst-C/BIA/UI4050/2011, PTDC/EBB-BIO/104980/2008 and PTDC/BIA-MIC/113733/09, and through a PhD grant (grant number SFRH/BD/76894/2011) to P.S.-C. [^1]: **Data deposition:** The Whole Genome Shotgun project of *Pseudomonas* sp. M1 has been deposited at DNA Data Bank of Japan/EMBL/GenBank under the accession ANIR00000000. The version described in this paper is ANIR00000000.2. [^2]: **Associate editor:** José Pereira-Leal [^3]: ^a^In total, 482 genes of the whole-genome annotation have multi-COG assignment. Further details are provided in [supplementary table S2](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1), [Supplementary Material](http://gbe.oxfordjournals.org/lookup/suppl/doi:10.1093/gbe/evu254/-/DC1) online.

INTRODUCTION {#sec1-1} ============ World Health Organization (WHO) elaborated scientifically grounded recommendations on normal birth management in 1996. The recommendations provided the basis for clinical protocols of diagnostics and treatment in Kazakhstan implemented in the country since 2007. Until then, labor management in Kazakhstan included a great number of routine procedures (cleansing enema, amniotomy, and others) and did not differ from such management as applied in the times of the USSR. Having found its independence, the Kazakhstan public health service settled down to a course of reforming and improvement. Over the last 10 years, the evidential practice has been developing in the country including midwifery. A questionnaire survey was carried out in all maternity departments of Almaty. Almaty was chosen as it is a Megapolis of Kazakhstan with the maximum number of obstetricians and midwives. Thus, according to the Statistics Digest,\[[@ref1]\] more than 763 obstetricians and 856 midwives are currently being employed in Almaty that is higher than these indices in the other cities of the country. The questionnaire survey was voluntary and anonymous, and only those respondents were involved in the survey, who had signed an informed consent. Inclusion criteria {#sec2-1} ------------------ Direct physiological labor management: Physicians and midwives of the reception ward and obstetric departments. Exclusion criteria {#sec2-2} ------------------ Physicians and midwives of the Maternal Child Unit and the Functional Diagnostics Department, those from the administration and others, who do not participate in the labor management. When drawing up the questionnaire, we relied on the similar surveys of other authors (evaluation of the medical-social and economic efficiency and the introduction of protocols on diagnostics and treatment in Kazakhstan-thesis). The representativeness of the sample was confirmed subject to the extent of the general totality --- in 2013, the total of obstetricians in Almaty in maternity departments was 113, of whom those, who were subject to the questionnaire survey −50 (44%), and the total of midwives −316, of whom those, who were subject the questionnaire survey-72 (22,7%).\[[@ref1]\] However, as the experience of some foreign countries shows, it is practically not always that one can fully succeed in the introduction of the evidential practice and the relevant WHO\'s recommendations. For example, in a number of both developed and developing countries, there is a low level of adherence to WHO\'s recommendations even in case of normal birth.\[[@ref2][@ref3][@ref4][@ref5][@ref6][@ref7][@ref8][@ref9][@ref10][@ref11][@ref12][@ref13][@ref14]\] The compliance with the protocols of diagnostics and treatment in Kazakhstan in the normal (physiological) birth management was analyzed by us. The aim of the research is to identify the degree of the usage of WHO\'s recommendations on intranatal care in Kazakhstan. MATERIALS AND METHODS {#sec1-2} ===================== To conduct research, we made a questionary particularly for obstetrician-gynecologists and midwives. Design {#sec2-3} ------ Qualitative descriptive research based on the questionnaire survey. Research methods {#sec2-4} ---------------- Sociological method based on the questionnaire survey, statistical analysis, analytical method. To carry on the questionnaire survey, an approval by the local ethical commission of the S.D. Asfendiyarov Kaz NMU was received (minutes of the meeting No. 4 dated 7.04.2014). Obstetricians and midwives were subject to the questionnaire survey conducted by us in 4 large Maternity Hospitals of Almaty. The survey was carried out by the mass observation where in all the personnel participated, who was involved in the normal birth management in the clinic (doctors, midwives, interns, trainee doctors, and hospital physicians). The evidentiary medicine principles have been introduced in the Kazakhstan obstetrics during the last more than 10 years, and this process is being carried on a planned and consistent basis supported by the Government of Kazakhstan\[[@ref15]\] and international organizations (EuroWHO, ZdravPlus, and others).\[[@ref16][@ref17]\] Some foreign and domestic specialists provided trainings of medical professionals (obstetricians and midwives) in modern scientifically proved principles of easy delivery management. However by law, the rules of managing easy delivery are enshrined in such protocols on diagnostics and treatment as accepted in 2007, update-2010, 2013. The findings were given in the absolute numbers and percent. The nominal data was compared by the Pearson Chi-square criterion. The critical significance level *P* = 0.05 when testing statistical hypotheses in the survey. Statistical analysis was made in the SPSS Program, Version 20.0 for Windows. RESULTS {#sec1-3} ======= At the first stage, the following practice of medical personnel is applied in Kazakhstan as per the questionnaire survey as shown in Figures [1](#F1){ref-type="fig"} and [2](#F2){ref-type="fig"}. {#F1} {#F2} At the first stage of labor, nonpharmacological methods of pain relief (point massage, self-massage and others) are used often-48.0% (χ^2^ = 25.750 df = 2 *P* \< 0.001) according to the obstetricians and 51.4% (χ^2^ = 43, 702 df = 2 *P* \< 0.001) that corresponds to the said WHO\'s recommendations. Second stage of labor \[Figures [3](#F3){ref-type="fig"} and [4](#F4){ref-type="fig"}\]. Pharmacological pain relief in labor is applied seldom in the Maternity Hospitals concerned 48% (χ^2^ = 4.0 df = 1, *P* = 0,046) based on the results of the questionnaire survey of the obstetricians, 50% (χ^2^ = 4.0 df = 1, *P* \< 0.001) of the midwives. {#F3} {#F4} Up to now, the amniotomy is used in case of the cervical dilatation at 8 cm or more-often - about 12.0% of the obstetricians so surveyed (χ^2^ = 9.6 df = 2 *P* \< 0.001), and 41.7% of the midwives (χ^2^ = 19.702 df = 2 *P* \< 0.001). The case follow-up for a maternity patient at the first stage of labor is carried out by measuring of temperature, pulse and blood pressure. According to the questionnaire survey, 54.0% (χ^2^ = 60.25 df = 3 *P* \< 0.001) of the obstetricians monitor the state of the woman --- every 4 h and 41.7% of the midwives (χ^2^ = 69.118 df = 5 *P* \< 0.00), fetal heart beats --- every 30 min (the obstetricians-48.0% (χ^2^ = 48.4 df = 3 *P* \< 0.001), and 58.3% of the midwives (χ^2^ = 100.0 df = 3 *P* \< 0.001). The vaginal examination is made by 72% of the surveyed obstetricians every 4 h that corresponds to the WHO\'s recommendations as per the partograph. At the second stage of labor, the Kristeller\'s method is used according to 8% of the surveyed obstetricians (χ^2^ = 10.4 df = 2 *P* = 0.006), 11.1% --- of the midwives (χ^2^ = 19.659 df = 2 *P* \< 0.001). At the third stage of labor, the WHO\'s recommendations are complied with as follows: Controlled cord traction noted by 50% of the surveyed midwives (χ^2^ = 57.8 df = 2 *P* \< 0.001), and 60% of the obstetricians (χ^2^ = 24.5 df = 1 *P* \< 0.00), preventive medication of Oxytocin-63.9% of the midwives (χ^2^ = 35.280 df = 1 *P* \< 0.001). DISCUSSION {#sec1-4} ========== According to such protocols of diagnostics and treatment as accepted in Kazakhstan, (Order of the Kazakhstan Ministry of Health, 2010), at the time of the survey, at the first stage of labor, the following was provided for in the RK: Case follow-up (measuring of temperature, pulse and blood pressure, uterine contractions, monitoring the fetal heart rate, partograph, and vaginal examination --- every 4 h), and point massage, at the second stage of labor, in addition to the case follow-up, obstetrical aide; and at the third stage-controlled cord traction and preventive prescription of oxytocin. According to the results of the questionnaire survey, despite the fact that the routine procedures such as routine use of enema fall in the Category B WHO (1996)\[[@ref18]\] the WHO\'s classification, that is, they are practices which are clearly harmful or ineffective and should be eliminated, the same has been still used in the Almaty Maternity Hospitals. At the second stage of labor, unfortunately, the Christeller\'s method is used that can be noted in the answers of both obstetrician-gynecologists and midwives. It should be particularly noted that some part of respondents both among the physicians and the midwives noted the usage of CTG at the first and second stages to monitor the state of a fetus even in the normal (physiological) birth management. CONCLUSION {#sec1-5} ========== Thus, in analyzing the questionnaire survey of the said obstetricians and midwives, who directly manage normal (physiological) births, it was found that the Almaty Maternity Hospitals complied with the WHO\'s recommendations but not in all cases of normal births. Up to now, there are inadvisable procedures and measures such as cleansing enema, sanitization, Kristeller\'s method that requires the more detailed study of the causes. Financial support and sponsorship {#sec2-5} --------------------------------- Nil. Conflicts of interest {#sec2-6} --------------------- The authors have no conflicts of interest. AUTHORS' CONTRIBUTION {#sec1-6} ===================== ADK and ABK contributed in the conception of the work, conducting the study, revising the draft, approval of the final version of the manuscript, and agreed for all aspects of the work. Aizhan D. Aizhan D. Kyzayeva contributed in the conception of the work, drafting and revising the draft, approval of the final version of the manuscript, and agreed for all aspects of the work. This study was the part of dissertation work (Kazakh National Medical University named after S.D. Asfendiyarov). Project 618.1.Protocol no 4 (7/04/2014).

INTRODUCTION ============ Choline is a water-soluble organic cation that is crucial for the normal function of all cells. Its cellular absorption occurs via choline transporters. After uptake, choline is metabolized and used for various functions (e.g., regulation of osmotic pressure, a precursor for methyl donors in epigenetics, phospholipid synthesis) ([@b13-bt-28-195]; [@b7-bt-28-195]). Choline is an essential component of the membrane phospholipids phosphatidylcholine (PC) and sphingomyelin and plays a critical role in the structure and function of biological membranes in all cells. Thus, large amounts of choline uptake and PC biosynthesis are needed for new membrane synthesis. Choline is also a precursor for the synthesis of the neurotransmitter acetylcholine at the cholinergic nerve terminal. Finally, choline is converted to betaine through oxidative metabolism, which provides a source of methyl groups for the production of S-adenosylmethionine (SAM). SAM participates in DNA and histone methylation and is therefore required for the establishment and maintenance of the epigenome ([@b1-bt-28-195]). ^11^C- or ^18^F-labelled choline derivatives are used as diagnostic tools for positron emission tomography (PET) and PET/computed tomography (PET/CT) in clinical oncology ([@b24-bt-28-195]; [@b28-bt-28-195]). Choline PET is highly effective for imaging several tumor types, including prostate carcinoma ([@b12-bt-28-195]), brain tumors ([@b20-bt-28-195]), lung metastasis of esophageal cancer ([@b11-bt-28-195]), and mediastinal lymph node metastases ([@b5-bt-28-195]). Choline PET can also be used to differentiate between malignant and benign lesions in various regions of the body, including the brain, head, bone, and soft tissue ([@b26-bt-28-195]). Choline PET or PET/CT is increasingly used to visualize primary and recurrent prostate cancer ([@b6-bt-28-195]; [@b28-bt-28-195]). Because tumor cells grow very quickly, the biosynthesis of their cell membranes is also rapid. Consequently, the uptake of choline by tumors is linked to the rate of tumor cell proliferation. Intracellular uptake of choline via the choline transporter is the rate-limiting step in the synthesis of phospholipids in the plasma membrane, and a prerequisite for cancer cell proliferation. However, the uptake system for choline and the functional expression of choline transporters in human prostate cancer are not completely understood. To date, the choline transport system has been categorized into three major transporter families: (I) high-affinity choline transporter 1 (CHT1/SLC5A7), (II) choline transporter-like proteins (CTL1-5/SLC44A1-5), and (III) polyspecific organic cation transporters (OCT1-2/SLC22A1-2) with low affinity for choline ([@b14-bt-28-195]). The families are expressed in different organisms and cell types and are involved in a variety of physiological functions. Our studies have shown that many cancer cells express high levels of CTL1, and inhibition of CTL1 function induces apoptotic cell death ([@b7-bt-28-195]; [@b25-bt-28-195]; [@b18-bt-28-195]; [@b17-bt-28-195]). Thus, CTL1 may be a target for cancer therapy. In this study, we investigated which transporters could mediate choline uptake in LNCaP prostate cancer cells, which are androgen-dependent. We also examined the correlations between choline uptake and the effects of anticancer drugs on cell viability. MATERIALS AND METHODS ===================== Cell culture ------------ The human prostate adenocarcinoma cell line, LNCaP (RCB2144), PC-3 (RCB2145) and DU-145 (RCB2143) were provided by the RIKEN BRC through the National Bio-Resource Project of the MEXT (Ibaraki, Japan). The cells were grown in RPMI 1640 medium (Wako, Osaka, Japan) supplemented with 10% fetal bovine serum (Gibco BRL, Gaithersburg, MD, USA) and 20 mg/L kanamycin (Gibco BRL) in non-coated flasks. Cultures were maintained in a humidified atmosphere of 5% CO~2~ and 95% air at 37°C, and the medium was changed every 3 to 4 days. RNA extraction and real-time polymerase chain reaction (PCR) ------------------------------------------------------------ Total RNA was extracted from cells using ISOGEN (Wako) according to the manufacturer's instructions. The primer pairs and TaqMan probes for the target (CHT1, OCT1-3, CTL1-5) and housekeeping (glyceraldehyde-3-phosphate dehydrogenase \[GAPDH\]) mRNAs were designed based on human mRNA sequences using TaqMan^®^ Gene Expression Assays (Applied Biosystems, Foster City, CA, USA) ([Table 1](#t1-bt-28-195){ref-type="table"}). One-step realtime PCR was performed on total RNA (50 ng) using the TaqMan^®^ RNA-to-CT^TM^ 1-Step Kit (Applied Biosystems) according to the manufacturer's instructions. Real-time PCR data were analyzed using the LightCycler^®^ 96 system (Roche Diagnostics, Mannheim, Germany). Target gene expression levels were calculated relative to GAPDH using the comparative cycle time (C~t~) method (relative mRNA expression=2^−(Ct\ target--Ct\ GAPDH)^×100%). Immunocytochemistry ------------------- LNCaP cells were grown in 35 mm glass base dishes (IWA-KI, Chiba, Japan), washed twice with D-PBS, and fixed with 100% MeOH for 20 min at room temperature. Fixed cells were incubated with detector blocking solution (Kirkegaard and Perry Laboratories, Gaithersburg, MD, USA) overnight at 4°C. Cells were incubated with 2 μg/mL anti-CTL1 rabbit polyclonal antibody (ab110767, Abcam plc, Cambridge, UK), 2 μg/mL anti-CTL2 mouse monoclonal antibody (clone 2D11, Abnova, Taipei City, Taiwan), or anti-COX IV rabbit polyclonal antibody (ab16056, Abcam plc) in fresh blocking solution overnight at 4°C. After washing with wash solution (Kirkegaard and Perry Laboratories), the cells were incubated with 1 μg/mL Alexa Fluor 488 goat anti-mouse IgG (Molecular Probes, Eugene, OR, USA) or Alexa Fluor 568 goat anti-rabbit IgG (Molecular Probes) for 1 h at room temperature. After washing, the specimens were mounted using VECTASHIELD mounting medium containing DAPI (Vector Laboratories, Burlingame, CA, USA). Fluorescence images were obtained using a Confocal Laser Scanning Biological Microscope (FV10i-DOC, Olympus, Tokyo, Japan). Immunoblotting -------------- Cells were washed twice with D-PBS (Wako) and extracted on ice in RIPA buffer (Santa Cruz, CA, USA) containing 1 mM EDTA and protease inhibitors. The extracts were centrifuged for 15 min at 4°C. The supernatant was incubated for 5 min at 100°C in a 1:1 (v:v) ratio with Tris--SDS β-ME sample solution (Cosmo Bio, Tokyo, Japan). The samples electrophoresed on 10% SDS-PAGE gels and then transferred to PVDF membranes. Following protein transfer, the membranes were blocked with Detector Block Solution (Kirkegaard and Perry Laboratories) overnight at 4°C. Membranes were then incubated with 1 μg/mL rabbit anti-CTL1 polyclonal antibody (ab110767, Abcam plc) in new Detector Block Solution overnight at 4°C. Membranes were washed three times in wash solution and then incubated with 1 μg/mL horseradish peroxidase-conjugated anti-rabbit and anti-goat IgG (Kirkegaard and Perry Laboratories) for 1 h at room temperature. Following washing, protein bands were visualized using the ECL Prime Western Blotting Detection System (GE Healthcare Life Sciences, Marlborough, MA, USA) on the ChemiDoc XRS Plus System (Bio-Rad Laboratories, Hercules, CA, USA). \[^3^H\]Choline uptake into LNCaP cells --------------------------------------- \[^3^H\]Choline uptake was measured as previously described ([@b17-bt-28-195]). Following removal of the growth medium, the cells were washed twice with uptake buffer (125 mM NaCl, 4.8 mM KCl, 1.2 mM CaCl~2~, 1.2 mM KH~2~PO~4~, 5.6 mM glucose, 1.2 mM MgSO~4~, and 25 mM HEPES adjusted to pH7.4 with Tris) and then incubated with \[^3^H\]choline (PerkinElmer, Boston, MA, USA). When Na^+^-free buffer was used, NaCl was replaced with an equimolar concentration of N-methyl-D-glucamine chloride (NMDG-Cl). Choline uptake was terminated by removal of the uptake buffer and three rapid washes with ice-cold uptake buffer. The cultures were dissolved in 0.1 M NaOH containing 0.1% Triton X-100. The radioactivity was measured using a liquid scintillation counter (Tri-Carb^®^2100 TR, Packard, Palo Alto, CA, USA). Uptake buffers of varying pH (pH 6.0, 6.5, 7.0, 7.5, 8.0, and 8.5) were prepared by mixing 25 mM MES (pH 5.5) and 25 mM Tris (pH 8.5). Both buffers contained 125 mM NaCl, 4.8 mM KCl, 1.2 mM CaCl~2~, 1.2 mMKH~2~PO~4~, 5.6 mM glucose, and 1.2 mM MgSO~4~. In saturation kinetics experiments, the concentration of \[^3^H\]choline was kept constant at 10 nM, and unlabeled choline was added to give the desired choline concentration. The specific uptake of \[^3^H\]choline was calculated as the difference between total \[^3^H\]choline uptake in the presence and absence of 30 mM unlabeled choline. Protein concentrations were determined using the DC Protein Assay Kit (Bio-Rad Laboratories). Cell viability assay -------------------- Cells were plated, and then inhibitors were added 24 h later. Cell numbers were measured using ATPLite^TM^ (PerkinElmer) and the luminescence ATP detection assay system according to the manufacturer's instructions. Luminescence was measured on a FilterMax F5 Multi-Mode Microplate Reader (Molecular Devices, San Jose, CA, USA). Caspase-3/7 assay ----------------- LNCaP cells were seeded at a density of 1×10^4^ cells/well in 24-well plates. Twenty-four hours after plating, cells were incubated with test drugs for 24 h. Caspase-3/7 activity was determined using the Caspase-Glo^®^ 3/7 Assay kit (Promega Corporation, Madison, WI, USA) according to the manufacturer's instructions. Luminescence was measured on a FilterMax F5 Multi-Mode Microplate Reader (Molecular Devices). Data analysis ------------- Analysis of variance (ANOVA) followed by the Dunnett's multiple comparison test for data from multiple groups was performed using InStat 3 (GraphPad Software, San Diego, CA, USA). The time courses of choline uptake were compared by fitting into single-exponential decay equations. The kinetic parameters were calculated by non-linear regression methods. Eadie-Hofstee plots were fitted using linear regression (Prism 6, GraphPad Software). The *K~i~* values were derived from the IC~50~ values as described by [@b2-bt-28-195]: *K~i~*=IC~50~/(1+\[L\]/*K~m~*), where \[L\] is the concentration of radiolabeled ligand. RESULTS ======= Expression of choline transporters in LNCaP, PC-3 and DU-145 cells ------------------------------------------------------------------ The expression of CHT1, CTL1-5, and OCT1-3 in prostate cancer cell lines (LNCaP, PC-3, and DU-145) was investigated using real-time PCR ([Fig. 1A](#f1-bt-28-195){ref-type="fig"}). In the prostate cancer cell lines, CTL1, CTL2, and CTL4 were highly expressed, and their levels did not differ from each other. In contrast, CTL3 and OCT1 were expressed at very low levels, and CHT1, CTL5, OCT2, and OCT3 were not detectable. The expression patterns of the transporters closely resembled it in all prostate cancer cell lines. Therefore, we conducted a molecular and functional analysis using LNCaP cells. It has been reported that human CTL4 protein functions as thiamine pyrophosphate (TPP) transporter, and CTL4 transport TPP but not thiamine and choline ([@b16-bt-28-195]). Therefore, CTL4 may not be involved in choline uptake in LNCaP cells. Expression of both CTL1 and CTL2 in LNCaP cells on the protein level was established by immunocytochemical staining ([Fig. 1B](#f1-bt-28-195){ref-type="fig"}). CTL1 was located in the plasma membrane. In contrast, CTL2 co-localized with COX IV, indicating that it was located in mitochondria. Properties of \[^3^H\]choline uptake into LNCaP cells ----------------------------------------------------- We examined the time course of \[^3^H\]choline uptake in LNCaP cells in the presence and absence of extracellular Na^+^ for 60 min ([Fig. 2A](#f2-bt-28-195){ref-type="fig"}). \[^3^H\]Choline uptake increased in a time-dependent manner and was linear for up to 20 min. When the NaCl in the uptake buffer was replaced by NMDG-Cl, the uptake of \[^3^H\]choline under Na^+^-free conditions was slightly decreased compared to uptake under normal conditions. The kinetics of \[^3^H\]choline uptake were determined by nonlinear regression. This analysis yielded a Michaelis--Menten constant (*K~m~*) of 4.7 ± 0.7 μM and maximal velocity (*V~max~*) of 93.4 ± 3.7 pmol/mg protein/min ([Fig. 2B](#f2-bt-28-195){ref-type="fig"}). The Eadie--Hofstee plot shows single straight lines. These data suggest that \[^3^H\] choline uptake in LNCaP cells is mediated by a single Na^+^-independent, intermediate-affinity transport system. It has been reported that choline uptake through CTL1 is pH-dependent ([@b8-bt-28-195]). Therefore, the influence of extracellular pH on the uptake of \[^3^H\]choline in LNCaP cells was examined by varying the pH of the pre-incubation and incubation media from pH 6.0 to 8.5 ([Fig. 3A](#f3-bt-28-195){ref-type="fig"}). \[^3^H\]Choline uptake significantly decreased when the extracellular pH was lowered from 7.5 to 6.0. Conversely, \[^3^H\]Choline uptake was enhanced when the pH of the extracellular medium was raised from 7.5 to 8.5. HC-3 is a choline uptake inhibitor ([@b7-bt-28-195]). We investigated the inhibitory effects of this compound on the uptake of \[^3^H\]choline into LNCaP cells ([Fig. 3B](#f3-bt-28-195){ref-type="fig"}). HC-3 inhibited \[^3^H\] choline uptake in a concentration-dependent manner with a *K~i~* of 67.3 μM. Effects of flutamide and bicalutamide on LNCaP cells ---------------------------------------------------- The anticancer drugs, flutamide and bicalutamide, inhibited cell viability and \[^3^H\]choline uptake in a concentration-dependent manner ([Fig. 4](#f4-bt-28-195){ref-type="fig"}, [5](#f5-bt-28-195){ref-type="fig"}). The correlations between the effects of both anticancer drugs on cell viability and \[^3^H\]choline uptake were significant (flutamide *R* square=0.9697, *p*=0.0023; bicalutamide *R* square=0.9329, *p*=0.0075). Caspase-3/7 activity was significantly increased following 24 h treatment with 80 and 100 μM flutamide or 100 μM bicalutamide. Because both flutamide and bicalutamide reduced choline uptake in LNCaP cells, we determined whether these two drugs had an effect on CTL1 protein levels in this cell line and found that both drugs (100 μM for 24 h treatment) significantly decreased the levels of this transporter ([Fig. 6](#f6-bt-28-195){ref-type="fig"}). DISCUSSION ========== The levels of choline and phospholipid metabolites in cancer tissues have been reported to correlate with the Gleason score and tumor growth in prostate cancer patients ([@b27-bt-28-195]). Choline PET is used clinically as a diagnostic imaging technique for prostate cancer and for monitoring prostate cancer recurrence ([@b22-bt-28-195]; [@b23-bt-28-195]). However, which choline transporters are expressed in prostate cancer is currently unknown. Therefore, we analyzed the functional expression of choline transporters in the human prostate cancer cell line LNCaP. We found that CTL1, CTL2, and CTL4 were highly expressed. Human CTL4 has been shown to transport TPP but not thiamine or choline ([@b16-bt-28-195]). Therefore, CTL4 is not likely involved in choline transport in LNCaP cells. We previously reported that CTL1 and CTL2 are highly expressed in various cancer cells types ([@b9-bt-28-195]; [@b25-bt-28-195]; [@b18-bt-28-195]; [@b17-bt-28-195]). Therefore, we examined the intracellular localization of CTL1 and CTL2 in LNCaP cells. CTL1 was clearly distributed in the plasma membrane, while CTL2 was localized to the intracellular compartment, specifically in mitochondria. These results suggest that it is CTL1 that is involved in the uptake of extracellular choline into cells, whereas CTL2 is involved in mitochondrial choline transport. These findings are similar to previous reports in which CTL1 is a target molecule for cancer therapy in esophageal cancer cells ([@b17-bt-28-195]). We believe that CTL1 is a therapeutic target molecule not only for esophageal cancer but also for prostate cancer. Choline oxidase is located in the inner membrane of mitochondria in the liver and kidney ([@b21-bt-28-195]; [@b10-bt-28-195]; [@b19-bt-28-195]). Therefore, intracellular choline must cross the mitochondrial inner membrane before oxidation can occur. CTL2 may play an important role as the rate-limiting step of choline uptake into mitochondria. We examined whether CTL1 was involved in choline transport in LNCaP cells. LNCaP cells have a Na^+^-independent and intermediate-choline transport systems. The *K~m~* value was 4.7 μM, which is similar to those in the human cell line, and these cell types show properties similar to those of CTL1. [@b15-bt-28-195] also evaluated the sodium dependence of choline uptake in LNCaP cells. They reported that choline uptake in LNCaP cells consists of two components, 50% NaCl-dependent, and 50% NaCl-independent ([@b15-bt-28-195]). However, our results showed only a Na^+^-independent choline uptake mechanism. These differences may be due to the problem in the replacement of NaCl with NMDG. In our experiment, the Cl ion concentration was maintained by replacing NaCl with NMDG chloride (NMDG-Cl). It is considered that choline uptake was inhibited by 50% as a result of replacing NaCl with NMDG alone and using a buffer solution containing 140 mM of less Cl ions. CTL1 is known to be pH-dependent and can be completely inhibited by HC-3 in the μM range ([@b9-bt-28-195]; [@b18-bt-28-195]; [@b17-bt-28-195]). In this study, \[^3^H\]choline uptake in LNCaP cells was decreased by acidification and increased by alkalinization of the extracellular medium, indicating that choline may be transported by a choline/H^+^ antiport system in these cells. Moreover, the *K~i~* value of HC-3 is 67.3 μM, which is very close to the *K~i~* value for CTL1. Thus, \[^3^H\]choline uptake in LNCaP cells may be mediated by CTL1. Flutamide and bicalutamide are nonsteroidal androgen receptor antagonists that are used in the treatment of prostate cancer ([@b4-bt-28-195]). LNCaP cells are androgen-dependent prostate cancers, and these antiandrogenic drugs appear to have antiproliferative effects. Therefore, we investigated the relationship between the antitumor effects of these antiandrogens and choline uptake. Both flutamide and bicalutamide inhibited cell viability and choline uptake in LNCaP cells, and these inhibitory effects were significantly correlated. In addition, both drugs increased caspase-3/7 activity at high concentrations. Thus, both flutamide and bicalutamide were found to have not only antiandrogenic activity but also apoptosis-inducing activity by inhibiting choline uptake. Flutamide and bicalutamide are organic cationic drugs that may competitively inhibit the uptake of choline, which is an organic cation. Many studies have reported that various organic cations inhibited choline uptake and cell viability ([@b29-bt-28-195]; [@b18-bt-28-195]; [@b17-bt-28-195]). These results demonstrate that the functional inhibition of choline uptake through CTL1 could promote apoptotic cell death. After the androgen testosterone is enzymatically converted to dihydrotestosterone, it binds to the androgen receptor, which is then translocated to the nucleus, and binds to DNA, leading to the growth of prostate cancer ([@b3-bt-28-195]). Thus, the androgen receptor complex may regulate the expression of the CTL1 gene in LNCaP cells. Indeed, both flutamide and bicalutamide caused a decrease in CTL1 protein levels in LNCaP cells. Thus, the antitumor activity of flutamide and bicalutamide is thought to be involved not only in direct inhibition of CTL1-mediated choline uptake but also in suppression of CTL1 expression. The current study suggests that LNCaP cells express the intermediate-affinity choline transporter CTL1, which uses a directed H^+^ gradient as a driving force. Choline uptake through CTL1 is correlated with cell viability, and the functional inhibition of CTL1 could promote apoptotic cell death. The antiandrogens flutamide and bicalutamide may express antitumor activity through the suppression of CTL1 function and expression. Thus, the CTL1-mediated choline transport system may be a potential new target for prostate cancer therapy. This work was supported by JSPS KAKENHI Grant Number 17K08315. **CONFLICT OF INTEREST** The authors have declared that there are no conflicts of interest. {#f1-bt-28-195} ![Time course and Na^+^-dependence of \[^3^H\]choline uptake in LNCaP cells. (A) Time course of \[^3^H\]choline uptake (10 μM) in the presence (●) and absence (○) of extracellular Na^+^ in LNCaP cells over 60 min. Na^+^-free buffer was modified by replacing NaCl with an equimolar concentration of NMDG-Cl. The time course of \[^3^H\]choline uptake was fitted to the experimental data using linear regression. (B) The kinetics of \[^3^H\]choline uptake in LNCaP cells. Cells were incubated for 20 min with 2.31 to 50 μM \[^3^H\]choline. Inset: Eadie-Hofstee plots of \[^3^H\]choline uptake show single straight lines. All data are presented as the mean ± SD (n=4).](bt-28-195f2){#f2-bt-28-195} ![Effects of pH and HC-3 treatment on \[^3^H\]choline uptake in LNCaP cells. (A) Influence of extracellular pH on \[^3^H\]choline uptake in LNCaP cells. \[^3^H\]choline uptake (10 μM) was measured for 20 min under different pH conditions. Data are presented as a percentage of uptake relative to that at pH 7.5. ^\*^*p*\<0.05, ^\*\*^*p*\<0.01, and ^\*\*\*^*p*\<0.001 compared to pH 7.5. (B) Effect of HC-3 on \[^3^H\]choline uptake in LNCaP cells. Cells were pre-incubated for 20 min with increasing concentrations of HC-3 followed by the measurement of \[^3^H\]choline uptake for 20 min. Results are shown as a percentage of the uptake for the vehicle control. All data are presented as the mean ± SD (n=4).](bt-28-195f3){#f3-bt-28-195} ![Effects of flutamide on cell viability, \[^3^H\]choline uptake, and caspase-3/7 activity in LNCaP cells. (A) Cell viability was measured 24 h after treatment with increasing concentrations of flutamide. The results are presented as the percentage of the vehicle control. (B) LNCaP cells were incubated for 20 min with increasing concentrations of flutamide followed by the measurement of \[^3^H\] choline uptake. \[^3^H\]Choline uptake was calculated as uptake per protein. The results are presented as the percentage of uptake in the vehicle control sample. (C) Cells were incubated for 24 h with increasing concentrations of flutamide followed by the measurement of caspase-3/7 activity and cell number. Caspase activity was calculated as the activity per cell number. The results are shown as a percentage of the vehicle control. (D) Correlation analysis between inhibition of \[^3^H\]choline uptake and cell viability for LNCaP cells treated with increasing concentrations of flutamide. The correlation was highly statistically significant (*p*=0.0023) with a correlation coefficient of *R* square=0.9697. Data for all assays are presented as the mean ± SD (n=4). ^\*\*^*p*\<0.01 compared with vehicle.](bt-28-195f4){#f4-bt-28-195} ![Effect of bicalutamide on cell viability, \[^3^H\]choline uptake, and caspase-3/7 activity in LNCaP cells. (A) Cell viability was measured 24 h after treatment with increasing concentrations of flutamide. The results are presented as the percentage of the vehicle control. (B) LNCaP cells were incubated for 20 min with increasing concentrations of bicalutamide followed by the measurement of \[^3^H\]choline uptake. \[^3^H\]Choline uptake was calculated as uptake per protein. The results are presented as the percentage of uptake in the vehicle control sample. (C) Cells were incubated for 24 h with increasing concentrations of bicalutamide followed by the measurement of caspase-3/7 activity and cell number. Caspase activity was calculated as the activity per cell number. The results are shown as a percentage of the vehicle control. (D) Correlation analysis between inhibition of \[^3^H\]choline uptake and cell viability for LNCaP cells treated with increasing concentrations of bicalutamide. The correlation was highly statistically significant (*p*=0.0075) with a correlation coefficient of *R* square=0.9329. Data for all assays are presented as the means ± SD (n=4). ^\*\*^*p*\<0.01 compared with vehicle.](bt-28-195f5){#f5-bt-28-195} {#f6-bt-28-195} ###### TaqMan^®^ gene expression assay Gene Accession number Assay ID Exon boundary Assay location Amplicon length ---------------- ------------------ --------------- --------------- ---------------- ----------------- CHT1 (SLC5A7) NM_021815 Hs00222367_m1 4--5 749 66 CTL1 (SLC44A1) NM_080546 Hs00223114_m1 5--6 759 66 CTL2 (SLC44A2) NM_020428 Hs01105936_m1 20--21 2060 67 CTL3 (SLC44A3) NM_001114106 Hs00537043_m1 5--6 625 64 CTL4 (SLC44A4) NM_001178044 Hs00228901_m1 12--13 1177 58 CTL5 (SLC44A5) NM_152697 Hs01120485_m1 9--10 673 73 OCT1 (SLC22A1) NM_003057 Hs00427554_m1 8--9 1493 66 OCT2 (SLC22A2) NM_003058 Hs01010723_m1 10--11 1771 120 OCT3 (SLC22A3) NM_021977 Hs01009568_m1 10--11 1657 73 GAPDH NM_002046 Hs99999905_m1 3--3 243 122

\section{Introduction and Motivation} Let $(G,+)$ be an abelian group and $W\subseteq G$ be a nonempty subset. A nonempty set $W'\subseteq G$ is said to be an \textit{additive complement} to $W$ if $W + W' = G.$ Additive complements have been studied since a long time in the context of representations of the integers e.g., they appear in the works of Erd\H{o}s, Hanani, Lorentz and others. See \cite{Lorentz54, Erdos54, ErdosSomeUnsolved57} etc. The notion of minimal additive complements for subsets of groups was introduced by Nathanson in \cite{NathansonAddNT4}. An additive complement $W'$ to $W$ is said to be minimal if no proper subset of $W'$ is an additive complement to $W$, i.e., $$W + W' = G \,\text{ and }\, W + (W'\setminus \lbrace w'\rbrace)\subsetneq G \,\,\, \forall w'\in W'.$$ Minimal complements are intimately connected with the existence of minimal nets in groups. See \cite[Section 2]{NathansonAddNT4} and \cite[Section 2.1]{MinComp1}. Further, in case of the additive group $\mathbb{Z}$, they are related to the study of minimal representations. See \cite[Section 3]{NathansonAddNT4}. Given two nonempty subsets $A, B$ of a group $G$, they are said to form a co-minimal pair if $A \cdot B = G$, and $A' \cdot B \subsetneq G$ for any $\emptyset \neq A' \subsetneq A$ and $A\cdot B' \subsetneq G$ for any $\emptyset \neq B' \subsetneq B$. Thus, they are pairs $(A\in G,B\in G)$ such that each element in a pair is a minimal complement to the other. Co-minimal pairs in essence capture the tightness property of a set and its complement. Further, they are a strict strengthening of the notion of minimal complements in the sense that a non-empty set $A$ might admit a minimal complement, but might not be a part of a co-minimal pair. See \cite[Lemma 2.1]{CoMin1}. Which sort of sets can be a part of a co-minimal pair is an interesting question. It was shown in \cite[Theorem B]{CoMin1} that if $G$ is a free abelian group (of any rank $\geq 1$), then, given any non-empty finite set $A$, there exists another set $B$ such that $(A,B)$ forms a co-minimal pair. Moreover, in a very recent work of Alon, Kravitz and Larson, they establish that any nonempty finite subset of an infinite abelian group is a minimal complement to some subset \cite[Theorem 2]{AlonKravitzLarson}, and this implies that the statement of \cite[Theorem B]{CoMin1} holds for nonempty finite subsets of any infinite abelian group. However, the existence of non-trivial co-minimal pairs involving infinite subsets $A$ and $B$ were unknown, until recently. It was shown in \cite{CoMin2} that such pairs exist and explicit constructions of two such pairs in the integers $\mathbb{Z}$ were provided. The aim of this article is to establish that infinite co-minimal pairs are abundant along majority of lacunary sequences in the integers and from there, draw conclusions on which sort of infinite subsets can be a part of a co-minimal pair. We are considering the underlying set associated with the sequence and this implies that there is some sparseness between successive elements of the underlying set. These subsets also satisfy a number of algebraic and combinatorial properties. Moreover, they can be generalised to any free abelian group of finite rank. \subsection{Statement of results} First, let us recall the notion of lacunary sequences. \begin{definition}[Lacunary sequence] A Lacunary sequence is a sequence of numbers $\lbrace x_{n}\rbrace_{n\in \mathbb{N}}$ such that $\frac{x_{n+1}}{x_{n}}\geqslant \lambda >1 \forall n\in \mathbb{N}$. \end{definition} Additive complements of lacunary sequences also has a long history. For a brief history see \cite[Section 1]{RuzsaLacunary}. However, the study of minimal additive complements and co-minimal pairs involving lacunary sequences is new. Our first result concerns co-minimal pairs of lacunary sequences. To avoid introducing cumbersome notation from the beginning, we state a simplified version of Theorem \ref{Thm} below \begin{theoremIntro} \label{ThmA} In the additive group of the integers, a ``majority'' of lacunary sequences have the property that they belong to a co-minimal pair. \end{theoremIntro} By a ``majority'', we mean the following: we know that a lacunary sequence $\lbrace x_{n}\rbrace_{n\in \mathbb{N}}$ is defined by the growth condition $\frac{x_{n+1}}{x_{n}}\geqslant \lambda$, with $\lambda \in (1,+\infty)$. The lacunary sequences which satisfy Theorem \ref{ThmA} have $\lambda \in[6,+\infty)$ and in some cases even $\lambda \in (3,+\infty)$. Further, they can be generalised to $\mathbb{Z}^{d}$. See Theorem \ref{Thm} for the complete statement. It is worth mentioning that in \cite{CoMin2}, it has been established that the set consisting of the terms of the lacunary sequence $1, 2, 2^2, 2^3, \cdots$ is a part of a co-minimal pair. Next, we consider the subsets of $\ensuremath{\mathbb{Z}}$ of the following types. \begin{enumerate}[(Type 1)] \item Symmetric subsets of $\ensuremath{\mathbb{Z}}$ containing the origin, \item Symmetric subsets of $\ensuremath{\mathbb{Z}}$ not containing the origin, \item Subsets of $\ensuremath{\mathbb{Z}}$ which are bounded from the above or from the below. \end{enumerate} We investigate whether a subset of $\ensuremath{\mathbb{Z}}$ of the above types can be a part of a co-minimal pair with a subset of $\ensuremath{\mathbb{Z}}$ of the above types. We establish certain sufficient conditions for this to be true. In fact, we have the following result. \begin{theoremIntro} For (I, II) equal to any one of $$(1, 1), (1, 2), (1, 3), (2, 1), (2, 2), (2, 3), (3, 3),$$ there are uncountably many subsets of $\ensuremath{\mathbb{Z}}$ of Type I which form a co-minimal pair with a subset of $\ensuremath{\mathbb{Z}}$ of Type II. \end{theoremIntro} Again, the above is a specialised version of a more general theorem which holds for $\mathbb{Z}^{d}$. See Theorem \ref{Thm:Uncountable}. \section{Co-minimal pairs involving lacunary sequences and generalisations} Let $d\geq 1$ be an integer. Denote the lexicographic order on $\ensuremath{\mathbb{Z}}^d$ by the symbol $<$. The element $(1, \cdots, 1)$ of $\ensuremath{\mathbb{Z}}^d$ is denoted by $\ensuremath{\mathbf{1}}$. For $?\in \{<, \leq, > , \geq\}$ and $x\in \ensuremath{\mathbb{Z}}^d$, the set $\{n\in \ensuremath{\mathbb{Z}}^d\,|\, n?x\}$ is denoted by $\ensuremath{\mathbb{Z}}^d_{?x}$. Let $t_0 < t_1 < t_2 < \cdots$ be elements of $\ensuremath{\mathbb{Z}}^d$. Assume that $t_0 > 0$, $$t_n \geq 2 t_{n-1} \quad \text{ for all } n \geq 1,$$ and for some $n\geq 0$, all the coordinates of $t_n$ are positive. Let $\ensuremath{\mathcal{T}}, \ensuremath{\mathcal{V}}, \ensuremath{\mathcal {W}}$ be subsets of $\ensuremath{\mathbb{Z}}^d$ defined by \begin{align*} \ensuremath{\mathcal{T}} & = \{t_0 , t_1, t_2, \cdots \},\\ \ensuremath{\mathcal{V}} & = \ensuremath{\mathcal{T}} \cup (-\ensuremath{\mathcal{T}}),\\ \ensuremath{\mathcal {W}} & = \ensuremath{\mathcal{T}} \cup \{0\} \cup (-\ensuremath{\mathcal{T}}). \end{align*} Then, the following holds, \begin{theorem} \label{Thm} \quad \begin{enumerate} \item If $$t_n > 3 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal {W}}$ and a symmetric subset $\ensuremath{\mathcal{E}}$ of $\ensuremath{\mathbb{Z}}^d$ containing the origin form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \item If $$t_n > 3 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal {W}}$ and a symmetric subset $\ensuremath{\mathcal{F}}$ of $\ensuremath{\mathbb{Z}}^d$ not containing the origin form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \item If $$t_n \geq 6 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal {W}}$ and a subset $\ensuremath{\mathcal{G}}$ of $\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0}$ form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \item If $$t_n > 3 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal{V}}$ and a symmetric subset $\ensuremath{\mathcal{P}}$ of $\ensuremath{\mathbb{Z}}^d$ containing the origin form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \item If $$t_n > 3 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal{V}}$ and a symmetric subset $\ensuremath{\mathcal{Q}}$ of $\ensuremath{\mathbb{Z}}^d$ not containing the origin form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \item If $$t_n \geq 6 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal{V}}$ and a subset $\ensuremath{\mathcal{R}}$ of $\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0}$ form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \item If $$t_n \geq 2 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $\ensuremath{\mathcal{T}}$ is a minimal complement to some subset of $\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0}$. In addition, if at least one coordinate of the sequence of points $\{t_n - 2t_{n-1}\}_{n\geq 1}$ goes to $\infty$, then the set $\ensuremath{\mathcal{T}}$ and a subset $\ensuremath{\mathcal{S}}$ of $\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0}$ form a co-minimal pair in $\ensuremath{\mathbb{Z}}^d$. \end{enumerate} \end{theorem} \begin{theorem} \label{Thm:UncountableSubsets} Any infinite subset of any finitely generated abelian group has uncountably many subsets which is a part of a co-minimal pair. In particular, any infinite subset of $\ensuremath{\mathbb{Z}}^d$ has uncountably many subsets which admit minimal complements. \end{theorem} It turns out that there are plenty of subsets of $\ensuremath{\mathbb{Z}}^d$ each of which is a part of a co-minimal pair. \begin{theorem} \label{Thm:Uncountable} The set of integers $\ensuremath{\mathbb{Z}}^d$ has uncountably many \begin{enumerate} \item symmetric subsets containing the origin, each of which forms a co-minimal pair together with a symmetric subset of $\ensuremath{\mathbb{Z}}^d$ containing the origin, \item symmetric subsets containing the origin, each of which forms a co-minimal pair together with a symmetric subset of $\ensuremath{\mathbb{Z}}^d$ not containing the origin, \item symmetric subsets containing the origin, each of which forms a co-minimal pair together with a subset of $\ensuremath{\mathbb{Z}}^d$ which avoids $\ensuremath{\mathbb{Z}}^d_{\geq 0}$, \item symmetric subsets not containing the origin, each of which forms a co-minimal pair together with a symmetric subset of $\ensuremath{\mathbb{Z}}^d$ containing the origin, \item symmetric subsets not containing the origin, each of which forms a co-minimal pair together with a symmetric subset of $\ensuremath{\mathbb{Z}}^d$ not containing the origin, \item symmetric subsets not containing the origin, each of which forms a co-minimal pair together with a subset of $\ensuremath{\mathbb{Z}}^d$ which avoids $\ensuremath{\mathbb{Z}}^d_{\geq 0}$, \item subsets contained in $\ensuremath{\mathbb{Z}}^d_{\geq 0}$, each of which forms a co-minimal pair together with a subset of $\ensuremath{\mathbb{Z}}^d$ which avoids $\ensuremath{\mathbb{Z}}^d_{\geq 0}$. \end{enumerate} \end{theorem} \section{Proofs} For two points $P, Q\in \ensuremath{\mathbb{Z}}^d$ satisfying $P\leq Q$, let $\ensuremath{\mathcal {X}}_{P, Q}$ denote the subset of $\ensuremath{\mathbb{Z}}^d$ defined by $$ \ensuremath{\mathcal {X}}_{P, Q} = \ensuremath{\mathbb{Z}}^d_{\geq P} \setminus \ensuremath{\mathbb{Z}}^d_{\geq Q} . $$ \begin{lemma} \label{Lemma:SumBound} Let $P, Q$ be points in $\ensuremath{\mathbb{Z}}^d$ satisfying $P \leq Q$, and $A$ be a nonempty subset of $\ensuremath{\mathbb{Z}}^d$. For any $v\in \ensuremath{\mathbb{Z}}^d$, the inclusion $$\ensuremath{\mathcal {X}}_{P, Q} + A_{\leq v} \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq Q+v} $$ holds. \end{lemma} \begin{proof} Note that \begin{align*} \ensuremath{\mathcal {X}}_{P, Q} + A_{\leq v} & \subseteq \cup _{a\in A , a \leq v} (\ensuremath{\mathcal {X}}_{P, Q} +a)\\ & \subseteq \cup _{a\in A , a \leq v} \ensuremath{\mathcal {X}}_{P+a, Q+a}\\ & \subseteq (\cup _{a\in A , a \leq v} \ensuremath{\mathbb{Z}}^d_{\geq P+a} ) \setminus (\cup _{a\in A , a \leq v} \ensuremath{\mathbb{Z}}^d_{\geq Q+a} )\\ & \subseteq (\cup _{a\in A , a \leq v} \ensuremath{\mathbb{Z}}^d_{\geq P+a} ) \setminus \ensuremath{\mathbb{Z}}^d_{\geq Q+v} \\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq Q+v} . \end{align*} \end{proof} Set $$t_{-2} = t_{-1} = 0.$$ Define the subsets $\{\ensuremath{\mathcal I}_n\}_{n\geq 0}$ of $\ensuremath{\mathbb{Z}}^d$ as follows. $$\ensuremath{\mathcal I}_n = \ensuremath{\mathcal {X}}_{-t_n, -t_{n-1} }. $$ Consider the subsets $\{\ensuremath{\mathcal J}_n\}_{n\geq 0}$ of $\ensuremath{\mathbb{Z}}^d$ defined by $$\ensuremath{\mathcal J}_n = \ensuremath{\mathcal {X}}_{-t_n, -t_{n-1} -t_{n-2}}. $$ These subsets make sense since $t_n \geq 2 t_{n-1}$ holds for $n\geq 0$. Note that for $n\geq 0$, $$ -t_{n-1} - t_{n-2} \leq -t_{n-1} $$ holds, this implies $$ \ensuremath{\mathbb{Z}}^d_{\geq -t_{n-1} - t_{n-2} } \supseteq \ensuremath{\mathbb{Z}}^d_{\geq -t_{n-1}} , $$ which yields $$\ensuremath{\mathcal J}_n \subseteq \ensuremath{\mathcal I}_n.$$ Since all the coordinates of $t_n$ are positive for some $n\geq 0$, we obtain $-t_n \to (-\infty, \cdots, -\infty)$. Thus it follows that $$\cup _{n\geq 0} \ensuremath{\mathcal I}_n = \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq 0}.$$ \begin{proposition} \label{Prop:M} \quad \begin{enumerate} \item Each of the sets $$ \cup _{m\geq n+2} \ensuremath{\mathcal J}_m + \ensuremath{\mathcal {W}} , \ensuremath{\mathcal J}_{n+1} + (\ensuremath{\mathcal {W}}\setminus \{t_n\})$$ contains no point of $\ensuremath{\mathcal I}_n$ for all $n\geq 0$. \item For any $n\geq 1$, the inclusion $$ (\ensuremath{\mathcal J}_0 \cup \cdots \cup (\ensuremath{\mathcal J}_n\cap \ensuremath{\mathbb{Z}}^d_{\geq -2t_{n-1}})) + \ensuremath{\mathcal {W}} \subseteq (\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq -t_n}) \cup \ensuremath{\mathbb{Z}}^d_{\geq -3t_{n-1}} $$ holds. \item If $$t_n > 3 t_{n-1} \quad \text{ for all } n \geq 1,$$ then the set $$ \cup _{m\geq n+1}(- (\ensuremath{\mathcal J}_m\cap \ensuremath{\mathbb{Z}}^d_{\geq -2t_{m-1} })) + \ensuremath{\mathcal {W}} $$ contains no point of $\ensuremath{\mathcal I}_n$ for all $n\geq 0$. \item For any $n\geq 1$, the inclusion $$ (-(\ensuremath{\mathcal J}_1 \cup \cdots \cup \ensuremath{\mathcal J}_n)) + (\ensuremath{\mathcal {W}} \setminus \{-t_n\}) \subseteq \ensuremath{\mathbb{Z}}^d_{\leq -2t_n} \cup (\ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq -t_{n-1}}) $$ holds if $$t_n \geq 3 t_{n-1} \quad \text{ for all } n \geq 1.$$ \end{enumerate} \end{proposition} \begin{proof} For $n\geq 0$, the inclusions \begin{align*} \ensuremath{\mathcal J}_{m} + \ensuremath{\mathcal {W}}_{\leq t_{m-2}} & \subseteq \ensuremath{\mathcal {X}}_{-t_m, - t_{m-1} - t_{m-2} } + \ensuremath{\mathcal {W}}_{\leq t_{m-2}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq - t_{m-1} - t_{m-2} + t_{m-2} }\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq - t_{m-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq - t_n}\\ \end{align*} hold for $m\geq n+1$ (the second inclusion follows from Lemma \ref{Lemma:SumBound}), the inclusions \begin{align*} \ensuremath{\mathcal J}_{m} + t_{m-1} & \subseteq \ensuremath{\mathcal {X}}_{-t_m, - t_{m-1} - t_{m-2} } + t_{m-1}\\ & = \ensuremath{\mathcal {X}}_{-t_m+t_{m-1}, - t_{m-1} - t_{m-2} + t_{m-1}}\\ & = \ensuremath{\mathcal {X}}_{-t_m+t_{m-1}, - t_{m-2} }\\ & \subseteq \ensuremath{\mathcal {X}}_{-t_m+t_{m-1}, - t_n } \end{align*} hold for $m\geq n+2$, the inclusions \begin{align*} \ensuremath{\mathcal J}_{m} + \ensuremath{\mathcal {W}}_{\geq t_m} & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - t_{m}} + \ensuremath{\mathcal {W}}_{\geq t_m}\\ & = \cup _{r\geq m} \ensuremath{\mathbb{Z}}^d_{\geq - t_{m} + t_r}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - t_{m} + t_{m}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq 0} \end{align*} hold for $m\geq n+1$. This proves part (1). For $n\geq 1$, the inclusions \begin{align*} (\ensuremath{\mathcal J}_0 \cup \cdots \cup \ensuremath{\mathcal J}_n) + \ensuremath{\mathcal {W}}_{\leq -t_n} & \subseteq \ensuremath{\mathcal {X}}_{-t_n, 0} +\ensuremath{\mathcal {W}}_{\leq -t_n}\\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - t_n} \end{align*} hold (the second inclusion follows from Lemma \ref{Lemma:SumBound}), the inclusions \begin{align*} (\ensuremath{\mathcal J}_0 \cup \cdots \cup (\ensuremath{\mathcal J}_n\cap \ensuremath{\mathbb{Z}}^d_{\geq -2t_{n-1}})) + \ensuremath{\mathcal {W}}_{\geq -t_{n-1}} & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq -2t_{n-1} } + \ensuremath{\mathcal {W}}_{\geq -t_{n-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq -3t_{n-1}} \end{align*} hold for $r\leq n-1$. This proves part (2). For $n\geq 0$, $m\geq n+1$, the inclusions \begin{align*} (- (\ensuremath{\mathcal J}_m\cap \ensuremath{\mathbb{Z}}^d_{\geq -2t_{m-1} })) - t_r & \subseteq \ensuremath{\mathbb{Z}}^d_{\leq 2t_{m-1} } - t_r \\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\leq 2t_{m-1} -t_m }\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{< - t_{m-1}} \\ & \subseteq \ensuremath{\mathbb{Z}}^d_{< - t_n } \end{align*} hold for $r\geq m$, the inclusions \begin{align*} (-\ensuremath{\mathcal J}_m) - t_r & \subseteq ( \ensuremath{\mathbb{Z}}^d_{\leq t_m} \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_{m-1}} )- t_r \\ & \subseteq ( \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_{m-1}} )- t_r \\ & = \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_{m-1}-t_r}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_{m-1}-t_{m-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq 0} \end{align*} hold for $r\leq m-1$. This proves part (3). For $n\geq 1$, the inclusions \begin{align*} -(\ensuremath{\mathcal J}_1 \cup \cdots \cup \ensuremath{\mathcal J}_n) - t_r & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_0} -t_r\\ & = \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_0 -t_r}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq t_0 -t_{n-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\leq -t_{n-1}} \end{align*} hold for $r\leq n-1$, the inclusions \begin{align*} -(\ensuremath{\mathcal J}_1 \cup \cdots \cup \ensuremath{\mathcal J}_n) - t_r & \subseteq \ensuremath{\mathbb{Z}}^d_{\leq t_n } - t_r\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\leq t_n - t_{n+1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\leq -2t_n }\\ \end{align*} hold for $r\geq n+1$. This proves part (4). \end{proof} \begin{lemma} \label{Lemma:Finiteness} Let $S$ and $T$ be nonempty subsets of an abelian group $G$ such that $S + T = G$. If the set $S$ is countable, and each element of $G$ can be expressed as a sum of an element of $S$ and an element of $T$ only in finitely many ways, then some nonempty subset of $S$ is a minimal complement to $T$. \end{lemma} \begin{proof} The lemma follows when $S$ is finite. Let us consider the case when $S$ is infinite. Let $s_1, s_2, \cdots$ be elements of $G$ such that $S = \{s_1, s_2, \cdots\}$. Define $S_1 = S$ and for each positive integer $i\geq 1$, define $$S_{i+1} : = \begin{cases} S_i \setminus \{-s_i\} & \text{ if $S_i \setminus \{-s_i\}$ is a complement to $T$,}\\ S_i & \text{ otherwise.} \end{cases} $$ Let $\ensuremath{\mathcal{S}}$ denote the subset $\cap _{i \geq 1} S_i$ of $G$. We claim that the set $\ensuremath{\mathcal{S}}$ is a minimal complement to $T$. For each $y\in G$ and for each $i\geq 1$, there exist elements $s_{y, i}\in S_i , t_{y, i}\in T$ such that $$y = s_{y, i} + t_{y, i}.$$ Consequently, for some element $s_y \in S$, we have the equality $s_y = s_{y, i}$ for infinitely many $i$. Further, for such integers $i$, we have $t_y = t_{y, i}$ where $t_y: = y - s_y\in T$. This implies that $y - t_y = s_{y, i}$ holds for infinitely many $i$. Hence, for each integer $i\geq 1$, there exists an integer $\ell_i \geq i$ such that $y - t_y = s_{y, \ell_i}$, which yields $$y\in t_y + S_{\ell_i} \subseteq t_y + S_i.$$ As a consequence, $y$ lies in $t_y + S$, i.e., $y\in S + T$. Hence $\ensuremath{\mathcal{S}}$ is an additive complement of $T$. It follows that $\ensuremath{\mathcal{S}}$ is a minimal complement to $T$. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(1)] Define the subsets $\ensuremath{\mathcal{E}}_n$ of $\ensuremath{\mathbb{Z}}^d$ for $n\geq 0$ as follows. $$\ensuremath{\mathcal{E}}_n = \begin{cases} \{0\} & \text{ if } n = -1,\\ \emptyset & \text{ if } n = 0,\\ \{ - t_{n-1} \} + (\ensuremath{\mathcal I}_{n-1} \setminus ( (\cup_{-1\leq m \leq n-1} (\ensuremath{\mathcal{E}}_m \cup (-\ensuremath{\mathcal{E}}_m))) + \ensuremath{\mathcal {W}})) & \text{ if } n \geq 1. \end{cases} $$ Define the subset $\ensuremath{\mathcal{E}}$ of $\ensuremath{\mathbb{Z}}^d$ by $$\ensuremath{\mathcal{E}} : = \cup _{n\geq -1} (\ensuremath{\mathcal{E}}_n \cup (-\ensuremath{\mathcal{E}}_n)).$$ For $n\geq 0$, the inclusions \begin{align*} \ensuremath{\mathcal{E}} + \ensuremath{\mathcal {W}} & \supseteq (\ensuremath{\mathcal{E}}_{n+1} + t_n) \cup ( (\cup_{-1\leq m \leq n} (\ensuremath{\mathcal{E}}_m \cup (-\ensuremath{\mathcal{E}}_m))) + \ensuremath{\mathcal {W}})\\ & \supseteq (\ensuremath{\mathcal I}_{n} \setminus ( (\cup_{-1\leq m \leq n} (\ensuremath{\mathcal{E}}_m \cup (-\ensuremath{\mathcal{E}}_m))) + \ensuremath{\mathcal {W}}))\cup ( (\cup_{-1\leq m \leq n} (\ensuremath{\mathcal{E}}_m \cup (-\ensuremath{\mathcal{E}}_m))) + \ensuremath{\mathcal {W}})\\ & \supseteq \ensuremath{\mathcal I}_n \end{align*} hold and hence $-\ensuremath{\mathcal I}_n \subseteq (-\ensuremath{\mathcal{E}}) + (-\ensuremath{\mathcal {W}}) = \ensuremath{\mathcal{E}} + \ensuremath{\mathcal {W}}$. Moreover, the inclusions $$ \ensuremath{\mathcal{E}} + \ensuremath{\mathcal {W}} \supseteq \ensuremath{\mathcal{E}}_{-1} + \ensuremath{\mathcal {W}} \supseteq \{0\}$$ hold. It follows that $\ensuremath{\mathcal {W}}$ is an additive complement to $\ensuremath{\mathcal{E}}$. We claim that $\ensuremath{\mathcal {W}}$ is a minimal complement of $\ensuremath{\mathcal{E}}$. Since $t_n > 3t_{n-1}$ for $n\geq 1$, from Proposition \ref{Prop:M}, it follows that for $n\geq 0$, no point of $\ensuremath{\mathcal{E}} \times \ensuremath{\mathcal {W}}$ other than $(0, -t_n)$ goes to $-t_n$ under the addition map $\ensuremath{\mathcal{E}} \times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$, and hence no point of $\ensuremath{\mathcal{E}} \times \ensuremath{\mathcal {W}}$ other than $(0, t_n)$ goes to $t_n$ under the addition map $\ensuremath{\mathcal{E}} \times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. Thus $\ensuremath{\mathcal {W}}$ is a minimal complement of $\ensuremath{\mathcal{E}}$. We claim that $\ensuremath{\mathcal{E}}$ is a minimal complement to $\ensuremath{\mathcal {W}}$. On the contrary, let us assume that $\ensuremath{\mathcal{E}}$ is not a minimal complement to $\ensuremath{\mathcal {W}}$. Hence $\ensuremath{\mathcal{E}}\setminus \{e\}$ is an additive complement to $\ensuremath{\mathcal {W}}$ for some $e\in \ensuremath{\mathcal{E}}$. Note that $e \neq 0$. Since $\ensuremath{\mathcal{E}}$ is symmetric, we may assume that $e$ lies in $\ensuremath{\mathcal{E}}_{n+1}$ for some $n\geq 0$. Thus $t_n +e$ lies in $\ensuremath{\mathcal I}_n$. It follows from Proposition \ref{Prop:M} that no element of $\ensuremath{\mathcal I}_n$ lies in $$ ((\cup _{m\geq n+2} \ensuremath{\mathcal{E}}_m) + \ensuremath{\mathcal {W}}) \cup ((\cup _{m\geq n+1} (-\ensuremath{\mathcal{E}}_m)) + \ensuremath{\mathcal {W}}) \cup (\ensuremath{\mathcal{E}}_{n+1} + (\ensuremath{\mathcal {W}} \setminus \{t_n\})).$$ So $t_n+e$ belongs to $( (\cup_{0 \leq m \leq n} (\ensuremath{\mathcal{E}}_m \cup (-\ensuremath{\mathcal{E}}_m))) + \ensuremath{\mathcal {W}} ) \cup ((\ensuremath{\mathcal{E}}_{n+1} \setminus\{e\}) + \{t_n\})$. Since $e\in \ensuremath{\mathcal I}_{n+1}$, it follows that $t_n + e$ lies in $(\ensuremath{\mathcal{E}}_{n+1} \setminus\{e\}) + \{t_n\}$, which yields $e\in \ensuremath{\mathcal{E}}_{n+1} \setminus\{e\}$. This contradicts the hypothesis that $\ensuremath{\mathcal{E}}$ is not a minimal complement of $\ensuremath{\mathcal {W}}$. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(2)] Define the subsets $\ensuremath{\mathcal{F}}_n$ of $\ensuremath{\mathbb{Z}}^d$ for $n\geq 0$ as follows. $$\ensuremath{\mathcal{F}}_n = \begin{cases} \{-t_0\} & \text{ if } n = 0,\\ \{ - t_{n-1} \} + (\ensuremath{\mathcal I}_{n-1} \setminus ( (\cup_{0\leq m \leq n-1} (\ensuremath{\mathcal{F}}_m \cup (-\ensuremath{\mathcal{F}}_m))) + \ensuremath{\mathcal {W}})) & \text{ if } n \geq 1. \end{cases} $$ Define the subset $\ensuremath{\mathcal{F}}$ of $\ensuremath{\mathbb{Z}}^d$ by $$\ensuremath{\mathcal{F}} : = \cup _{n\geq 0} (\ensuremath{\mathcal{F}}_n \cup (-\ensuremath{\mathcal{F}}_n)).$$ For $n\geq 0$, the inclusions \begin{align*} \ensuremath{\mathcal{F}} + \ensuremath{\mathcal {W}} & \supseteq (\ensuremath{\mathcal{F}}_{n+1} + t_n) \cup ( (\cup_{0\leq m \leq n} (\ensuremath{\mathcal{F}}_m \cup (-\ensuremath{\mathcal{F}}_m))) + \ensuremath{\mathcal {W}})\\ & \supseteq (\ensuremath{\mathcal I}_{n} \setminus ( (\cup_{0\leq m \leq n} (\ensuremath{\mathcal{F}}_m \cup (-\ensuremath{\mathcal{F}}_m))) + \ensuremath{\mathcal {W}}))\cup ( (\cup_{0\leq m \leq n} (\ensuremath{\mathcal{F}}_m \cup (-\ensuremath{\mathcal{F}}_m))) + \ensuremath{\mathcal {W}})\\ & \supseteq \ensuremath{\mathcal I}_n \end{align*} hold and hence $-\ensuremath{\mathcal I}_n \subseteq (-\ensuremath{\mathcal{F}}) + (-\ensuremath{\mathcal {W}}) = \ensuremath{\mathcal{F}} + \ensuremath{\mathcal {W}}$. Moreover, the inclusions $$ \ensuremath{\mathcal{F}} + \ensuremath{\mathcal {W}} \supseteq \ensuremath{\mathcal{F}}_{0} + \ensuremath{\mathcal {W}} \supseteq \{0\}$$ hold. It follows that $\ensuremath{\mathcal {W}}$ is an additive complement to $\ensuremath{\mathcal{F}}$. We claim that $\ensuremath{\mathcal {W}}$ is a minimal complement of $\ensuremath{\mathcal{F}}$. Since $t_n > 3t_{n-1}$ for $n\geq 1$, from Proposition \ref{Prop:M}, it follows that for $n\geq 0$, no point of $\ensuremath{\mathcal{F}} \times \ensuremath{\mathcal {W}}$ other than $(-2t_n, t_n)$ goes to $-t_n$ under the addition map $\ensuremath{\mathcal{F}} \times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$, and hence no point of $\ensuremath{\mathcal{F}} \times \ensuremath{\mathcal {W}}$ other than $(2t_n, -t_n)$ goes to $t_n$ under the addition map $\ensuremath{\mathcal{F}} \times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. Thus $\ensuremath{\mathcal {W}}$ is a minimal complement of $\ensuremath{\mathcal{F}}$. We claim that $\ensuremath{\mathcal{F}}$ is a minimal complement to $\ensuremath{\mathcal {W}}$. On the contrary, let us assume that $\ensuremath{\mathcal{F}}$ is not a minimal complement to $\ensuremath{\mathcal {W}}$. Hence $\ensuremath{\mathcal{F}}\setminus \{f\}$ is an additive complement to $\ensuremath{\mathcal {W}}$ for some $f\in \ensuremath{\mathcal{F}}$. Since $\ensuremath{\mathcal{F}}$ is symmetric, we may assume that $f$ lies in $\ensuremath{\mathcal{F}}_{n+1}$ for some $n\geq -1$. Since no point of $\ensuremath{\mathcal{F}} \times \ensuremath{\mathcal {W}}$ other than $(-t_0, -t_0)$ goes to $-2t_0$ under the addition map $\ensuremath{\mathcal{F}} \times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$, it follows that $f \neq -t_0$, i.e., $f\notin \ensuremath{\mathcal{F}}_0$. So $f$ lies in $\ensuremath{\mathcal{F}}_{n+1}$ for some $n\geq 0$. Thus $t_n +f$ lies in $\ensuremath{\mathcal I}_n$. It follows from Proposition \ref{Prop:M} that no element of $\ensuremath{\mathcal I}_n$ lies in $$ ((\cup _{m\geq n+2} \ensuremath{\mathcal{F}}_m) + \ensuremath{\mathcal {W}}) \cup ((\cup _{m\geq n+1} (-\ensuremath{\mathcal{F}}_m)) + \ensuremath{\mathcal {W}}) \cup (\ensuremath{\mathcal{F}}_{n+1} + (\ensuremath{\mathcal {W}} \setminus \{t_n\})).$$ So $t_n+f$ belongs to $( (\cup_{0 \leq m \leq n} (\ensuremath{\mathcal{F}}_m \cup (-\ensuremath{\mathcal{F}}_m))) + \ensuremath{\mathcal {W}} ) \cup ((\ensuremath{\mathcal{F}}_{n+1} \setminus\{f\}) + \{t_n\})$. Since $f\in \ensuremath{\mathcal I}_{n+1}$, it follows that $t_n + f$ lies in $(\ensuremath{\mathcal{F}}_{n+1} \setminus\{f\}) + \{t_n\}$, which yields $f\in \ensuremath{\mathcal{F}}_{n+1} \setminus\{f\}$. This contradicts the hypothesis that $\ensuremath{\mathcal{F}}$ is not a minimal complement of $\ensuremath{\mathcal {W}}$. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(3)] Let $\{m_k\}_{k\geq 0}$ be an increasing sequence such that $m_0 \geq 2$ and $t_{m_k -2}\geq (k+1)\ensuremath{\mathbf{1}}$ for all $k\geq 0$. We define the subsets $G_n$ of $\ensuremath{\mathbb{Z}}^d$ for $n\geq 0$ as follows. $$G_n = \begin{cases} \ensuremath{\mathcal I}_0 & \text{ if } n = 0, \\ \ensuremath{\mathcal {X}}_{-2t_{n-1}, -t_{n-1} - t_{n-2}} & \text{ if } n \geq 1 \text{ and } n \neq m_k \text{ for all } k \geq 0,\\ \ensuremath{\mathcal {X}}_{- t_n + k\ensuremath{\mathbf{1}}, - t_n + (k+1)\ensuremath{\mathbf{1}}} \cup \ensuremath{\mathcal {X}}_{-2t_{n-1}, -t_{n-1} - t_{n-2}} & \text{ if } n \geq 1 \text{ and } n = m_k \text{ for some } k \geq 0.\\ \end{cases} $$ Let $G$ denote the union $\cup_{n\geq 0 } G_n$. Note that $\ensuremath{\mathcal{T}}$ is an additive complement of $G$. Indeed, the inclusions \begin{align*} G + \ensuremath{\mathcal{T}} & \supseteq \left(\cup_{n\geq 1} (\ensuremath{\mathcal {X}}_{-2t_{n-1}, -t_{n-1} - t_{n-2}}+ \ensuremath{\mathcal{T}}) \right) \bigcup \left(\cup_{k\geq 0} (\ensuremath{\mathcal {X}}_{- t_{m_k} + k\ensuremath{\mathbf{1}}, - t_{m_k} + (k+1)\ensuremath{\mathbf{1}}}+ \ensuremath{\mathcal{T}})\right) \\ & \supseteq (\cup_{n\geq 1} \ensuremath{\mathcal I}_{n-1} ) \cup \ensuremath{\mathbb{Z}}^d_{\geq 0} \\ & = \ensuremath{\mathbb{Z}}^d \end{align*} hold, which shows that $G + \ensuremath{\mathcal{T}} = \ensuremath{\mathbb{Z}}^d$. So $\ensuremath{\mathcal {W}}$ is an additive complement to $G$. We claim that $\ensuremath{\mathcal {W}}$ is a minimal complement of $G \setminus (\{-2t_0\}\cup \{- t_n - 3t_{n-1} \,|\,n \geq 1\})$. For $k\geq 0$, the inequalities \begin{align*} t_{m_k} - t_{m_k-1} - t_{m_k-2} & = (t_{m_k} - 2t_{m_k-1}) + (t_{m_k-1} - 2t_{m_k-2}) + t_{m_k-2} \\ & \geq t_{m_k-2} \\ & \geq (k+1)\ensuremath{\mathbf{1}} \end{align*} hold, which implies $$-t_{m_k} + (k+1)\ensuremath{\mathbf{1}} \leq - t_{m_k-1} - t_{m_k-2} ,$$ this yields $$ \ensuremath{\mathbb{Z}}^d_{\geq -t_{m_k} + (k+1)\ensuremath{\mathbf{1}}} \supseteq \ensuremath{\mathbb{Z}}^d_{\geq - t_{m_k-1} - t_{m_k-2} },$$ and hence $$ \ensuremath{\mathcal {X}}_{-t_{m_k}, -t_{m_k} + (k+1)\ensuremath{\mathbf{1}}} \subseteq \ensuremath{\mathcal {X}}_{-t_{m_k}, - t_{m_k-1} - t_{m_k-2} }.$$ Thus $G_n \subseteq \ensuremath{\mathcal J}_n$ for all $n\geq 0$. Let $n$ be a positive integer. Note that $-3t_{n-1}, -4t_{n-1}$ lie in $\ensuremath{\mathcal I}_n$. The inclusions \begin{align*} \cup_{0 \leq m < n} G_m + \ensuremath{\mathcal {W}}_{\leq -t_n} & \subseteq (\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0}) + \ensuremath{\mathbb{Z}}^d_{\leq - t_n} \\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq -t_n}\\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - 4t_{n-1} } \end{align*} hold, and the inclusions \begin{align*} \cup_{0 \leq m < n} G_m + \ensuremath{\mathcal {W}}_{\geq -t_{n-1}} & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - t_{n-1} } + \ensuremath{\mathbb{Z}}^d_{\geq - t_{n-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - 2t_{n-1} } \\ \end{align*} hold. Using Proposition \ref{Prop:M}, it follows that the set $$ \cup _{m \neq n, n+1} G_m + \ensuremath{\mathcal {W}}$$ contains none of $-3t_{n-1}, -4t_{n-1}$. Since $-3t_{n-1} , -4 t_{n-1}\in \ensuremath{\mathcal I}_n$, and $G_{n+1} + t_n$ contains $\ensuremath{\mathcal I}_n$, it follows that $- t_n - 3t_{n-1}, -t_n -4t_{n-1} \in G_{n+1}$. By Proposition \ref{Prop:M}, $G_{n+1} + (\ensuremath{\mathcal {W}} \setminus \{t_n\})$ contains no element of $\ensuremath{\mathcal I}_n$. So no point of $\cup _{m \neq n} G_m + \ensuremath{\mathcal {W}}$ other than $(- t_n - 3t_{n-1}, t_n)$ goes to $-3t_{n-1}$, and no point of $\cup _{m \neq n} G_m + \ensuremath{\mathcal {W}}$ other than $(- t_n - 4t_{n-1}, t_n)$ goes to $-4t_{n-1}$. Note that the inclusions \begin{align*} (\{-t_{n-1}\} + \ensuremath{\mathcal I}_{n-1} )+ \ensuremath{\mathcal {W}}_{> -t_{n-1}} & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq -2t_{n-1} } + \ensuremath{\mathbb{Z}}^d_{> -t_{n-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{> -2t_{n-1} -t_{n-1}}\\ & = \ensuremath{\mathbb{Z}}^d_{> -3t_{n-1} } \end{align*} hold, the inclusions \begin{align*} (\{-t_{n-1}\} + \ensuremath{\mathcal I}_{n-1} )+ \ensuremath{\mathcal {W}}_{\leq -t_{n}} & \subseteq \ensuremath{\mathcal {X}}_{-t_{n-1}, -t_{n-2}} + \ensuremath{\mathcal {W}}_{\leq -t_{n}}\\ & \subseteq \ensuremath{\mathcal {X}}_{-t_{n-1}, 0} + \ensuremath{\mathcal {W}}_{\leq -t_{n}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq -t_n} \end{align*} hold. Moreover, the inclusion $$ \ensuremath{\mathcal {X}}_{- t_n + k\ensuremath{\mathbf{1}}, - t_n + (k+1)\ensuremath{\mathbf{1}}} + \ensuremath{\mathcal {W}}_{\geq t_n} \subseteq \ensuremath{\mathbb{Z}}^d_{\geq 0} $$ holds when $n = m_k$ for some $k\geq 0$, and the inclusions \begin{align*} \ensuremath{\mathcal {X}}_{- t_n + k\ensuremath{\mathbf{1}}, - t_n + (k+1)\ensuremath{\mathbf{1}}} + \ensuremath{\mathcal {W}}_{\leq t_{n-1}} & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - t_n + (k+1)\ensuremath{\mathbf{1}} + t_{n-1}} \\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - 6t_{n-1} + (k+1)\ensuremath{\mathbf{1}} + t_{n-1}} \\ & = \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - 5t_{n-1} + (k+1)\ensuremath{\mathbf{1}}} \\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - 4t_{n-1} - t_{n-2} + (k+1)\ensuremath{\mathbf{1}}} \\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq - 4t_{n-1} } \end{align*} hold when $n = m_k$ for some $k\geq 0$. Also note that $G_n$ contains $- 2t_{n-1}$. It follows that no element of $G_n \times \ensuremath{\mathcal {W}}$ other than $(-2t_{n-1}, -t_{n-1})$ goes to $-3t_{n-1}$ under the addition map $G_n\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. Hence no element of $G \times \ensuremath{\mathcal {W}}$ other than $(-2t_{n-1}, -t_{n-1}), (- t_n - 3t_{n-1}, t_n)$ goes to $-3t_{n-1}$ under the addition map $G\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. Moreover, $G_n + \ensuremath{\mathcal {W}}$ does not contain $-4t_{n-1}$. Hence no element of $G \times \ensuremath{\mathcal {W}}$ other than $(- t_n - 4t_{n-1}, t_n)$ goes to $-4t_{n-1}$ under the addition map $G\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. Proposition \ref{Prop:M} implies that $$ - t_0 \notin (\cup_{m \geq 2} G_m + \ensuremath{\mathcal {W}}) \cup (G_{1} + (\ensuremath{\mathcal {W}} \setminus \{t_{0}\})). $$ Also note that the inclusions \begin{align*} G_0 + \ensuremath{\mathcal {W}}_{\geq t_0} & \subseteq \ensuremath{\mathcal {X}}_{-t_0, 0} + \ensuremath{\mathcal {W}}_{\geq t_0}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq -t_0} + \ensuremath{\mathcal {W}}_{\geq t_0}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq 0} \end{align*} hold and the inclusions \begin{align*} G_0 + \ensuremath{\mathcal {W}}_{\leq -t_0} & \subseteq \ensuremath{\mathcal {X}}_{-t_0, 0} + \ensuremath{\mathcal {W}}_{\leq -t_0}\\ & \subseteq \ensuremath{\mathbb{Z}}^d \setminus \ensuremath{\mathbb{Z}}^d_{\geq -t_0} \end{align*} hold, and hence no element of $G\times \ensuremath{\mathcal {W}}$ other than $(-2t_0, t_0), (-t_0, 0)$ goes to $-t_0$ under the addition map $G\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. It follows that $\ensuremath{\mathcal {W}}$ is a minimal complement to $G \setminus (\{-2t_0\}\cup \{- t_n - 3t_{n-1} \,|\,n \geq 1\})$. We claim that $\ensuremath{\mathcal {W}}$ and some subset of $G \setminus (\{-2t_0\}\cup \{- t_n - 3t_{n-1} \,|\,n \geq 1\})$ form a co-minimal pair. By Proposition \ref{Prop:M}, each element of $\ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0} = \cup_{n\geq 0} \ensuremath{\mathcal I}_n$ can be expressed as a sum of an element of $G$ and an element of $\ensuremath{\mathcal {W}}$ only in finitely many ways. Note that the inclusions \begin{align*} G_m + \ensuremath{\mathcal {W}}_{\leq t_{m-1}} & \subseteq \ensuremath{\mathcal {X}}_{-t_m, - t_{m-1} } + \ensuremath{\mathcal {W}}_{\leq t_{m-1}}\\ & \subseteq \ensuremath{\mathbb{Z}}^d\setminus \ensuremath{\mathbb{Z}}^d_{\geq 0} \end{align*} hold for $m\geq 1$ and $r\leq m-1$, the inclusions \begin{align*} (\{-t_{m-1}\} + \ensuremath{\mathcal I}_{m-1} ) + \ensuremath{\mathcal {W}}_{\geq t_m} & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - 2t_{m-1} } + \ensuremath{\mathbb{Z}}^d_{\geq t_m}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq t_m- 2t_{m-1} } \end{align*} hold for $m\geq 1$, the inequality $$ \ensuremath{\mathcal {X}}_{-t_{m_k} + k\ensuremath{\mathbf{1}}, -t_{m_k} + (k+1)\ensuremath{\mathbf{1}} } + t_r \geq \ensuremath{\mathbb{Z}}^d_{\geq - t_{m_k} + k\ensuremath{\mathbf{1}} + t_r} = \ensuremath{\mathbb{Z}}^d_{\geq k\ensuremath{\mathbf{1}}} $$ holds for $k\geq 0$ and for $r\geq m_k$, and hence $$\cup_{m\geq M} G_m + \ensuremath{\mathcal {W}}$$ does not contain a given point of $\ensuremath{\mathbb{Z}}^d_{\geq 0}$ for some large enough $M$. Hence any given element element of $\ensuremath{\mathbb{Z}}^d_{\geq 0}$ can be expressed as a sum of an element of $G$ and an element of $\ensuremath{\mathcal {W}}$ only in finitely many ways. So each element of $\ensuremath{\mathbb{Z}}^d$ can be expressed as a sum of an element of $G$ and an element of $\ensuremath{\mathcal {W}}$ only in finitely many ways. In particular, each element of $\ensuremath{\mathbb{Z}}^d$ can be expressed as a sum of an element of $G \setminus (\{-2t_0\}\cup \{- t_n - 3t_{n-1} \,|\,n \geq 1\})$ and an element of $\ensuremath{\mathcal {W}}$ only in finitely many ways. By Lemma \ref{Lemma:Finiteness}, it follows that some nonempty subset $\ensuremath{\mathcal{G}}$ of $G \setminus (\{-2t_0\}\cup \{- t_n - 3t_{n-1} \,|\,n \geq 1\})$ is a minimal complement to $\ensuremath{\mathcal {W}}$. Since $\ensuremath{\mathcal {W}}$ is a minimal complement to $G \setminus (\{-2t_0\}\cup \{- t_n - 3t_{n-1} \,|\,n \geq 1\})$, it follows that $\ensuremath{\mathcal {W}}$ is a minimal complement to $\ensuremath{\mathcal{G}}$. Hence $(\ensuremath{\mathcal{G}}, \ensuremath{\mathcal {W}})$ is a co-minimal pair. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(4)] Define the subsets $\ensuremath{\mathcal{P}}_n$ of $\ensuremath{\mathbb{Z}}^d$ for $n\geq 0$ as follows. $$\ensuremath{\mathcal{P}}_n = \begin{cases} \{0\} & \text{ if } n = -1,\\ \{-t_0\} & \text{ if } n = 0,\\ \{ - t_{n-1} \} + (\ensuremath{\mathcal I}_{n-1} \setminus ( (\cup_{-1\leq m \leq n-1} (\ensuremath{\mathcal{P}}_m \cup (-\ensuremath{\mathcal{P}}_m))) + \ensuremath{\mathcal{V}})) & \text{ if } n \geq 1. \end{cases} $$ Define the subset $\ensuremath{\mathcal{P}}$ of $\ensuremath{\mathbb{Z}}^d$ by $$\ensuremath{\mathcal{P}} : = \cup _{n\geq -1} (\ensuremath{\mathcal{P}}_n \cup (-\ensuremath{\mathcal{P}}_n)).$$ For $n\geq 1$, the inclusions \begin{align*} \ensuremath{\mathcal{P}} + \ensuremath{\mathcal{V}} & \supseteq (\ensuremath{\mathcal{P}}_{n+1} + t_n) \cup ( (\cup_{-1\leq m \leq n} (\ensuremath{\mathcal{P}}_m \cup (-\ensuremath{\mathcal{P}}_m))) + \ensuremath{\mathcal{V}})\\ & \supseteq (\ensuremath{\mathcal I}_{n} \setminus ( (\cup_{-1\leq m \leq n} (\ensuremath{\mathcal{P}}_m \cup (-\ensuremath{\mathcal{P}}_m))) + \ensuremath{\mathcal{V}}))\cup ( (\cup_{-1\leq m \leq n} (\ensuremath{\mathcal{P}}_m \cup (-\ensuremath{\mathcal{P}}_m))) + \ensuremath{\mathcal{V}})\\ & \supseteq \ensuremath{\mathcal I}_n \end{align*} hold and hence $-\ensuremath{\mathcal I}_n \subseteq (-\ensuremath{\mathcal{P}}) + (-\ensuremath{\mathcal{V}}) = \ensuremath{\mathcal{P}} + \ensuremath{\mathcal{V}}$. Moreover, the inclusions $$\ensuremath{\mathcal I}_0\subseteq \ensuremath{\mathcal{P}}_1 + t_0, $$ $$-\ensuremath{\mathcal I}_0\subseteq (-\ensuremath{\mathcal{P}}_1) + (- t_0), $$ $$ \ensuremath{\mathcal{P}} + \ensuremath{\mathcal{V}} \supseteq \ensuremath{\mathcal{P}}_{0} + \ensuremath{\mathcal{V}} \supseteq \{0\}$$ hold. It follows that $\ensuremath{\mathcal{V}}$ is an additive complement to $\ensuremath{\mathcal{P}}$. We claim that $\ensuremath{\mathcal{V}}$ is a minimal complement of $\ensuremath{\mathcal{P}}$. Since $t_n > 3t_{n-1}$ for $n\geq 1$, from Proposition \ref{Prop:M}, it follows that for $n\geq 0$, no point of $\ensuremath{\mathcal{P}} \times \ensuremath{\mathcal{V}}$ other than $(0, -t_n)$ goes to $-t_n$ under the addition map $\ensuremath{\mathcal{P}} \times \ensuremath{\mathcal{V}} \to \ensuremath{\mathbb{Z}}^d$, and hence no point of $\ensuremath{\mathcal{P}} \times \ensuremath{\mathcal{V}}$ other than $(0, t_n)$ goes to $t_n$ under the addition map $\ensuremath{\mathcal{P}} \times \ensuremath{\mathcal{V}} \to \ensuremath{\mathbb{Z}}^d$. Thus $\ensuremath{\mathcal{V}}$ is a minimal complement of $\ensuremath{\mathcal{P}}$. We claim that $\ensuremath{\mathcal{P}}$ is a minimal complement to $\ensuremath{\mathcal{V}}$. On the contrary, let us assume that $\ensuremath{\mathcal{P}}$ is not a minimal complement to $\ensuremath{\mathcal{V}}$. Hence $\ensuremath{\mathcal{P}}\setminus \{p\}$ is an additive complement to $\ensuremath{\mathcal{V}}$ for some $p\in \ensuremath{\mathcal{P}}$. Note that $p \neq 0$. Since $\ensuremath{\mathcal{P}}$ is symmetric, we may assume that $p\in \ensuremath{\mathcal{P}}_{n+1}$ for some $n\geq -1$. Since no point of $\ensuremath{\mathcal{P}} \times \ensuremath{\mathcal{V}}$ other than $(-t_0, -t_0)$ goes to $-2t_0$ under the addition map $\ensuremath{\mathcal{P}} \times \ensuremath{\mathcal{V}} \to \ensuremath{\mathbb{Z}}^d$, it follows that $p \neq -t_0$, i.e., $p\notin \ensuremath{\mathcal{P}}_0$. So $p$ lies in $\ensuremath{\mathcal{P}}_{n+1}$ for some $n\geq 0$. Thus $t_n +p$ lies in $\ensuremath{\mathcal I}_n$. It follows from Proposition \ref{Prop:M} that no element of $\ensuremath{\mathcal I}_n$ lies in $$ ((\cup _{m\geq n+2} \ensuremath{\mathcal{P}}_m) + \ensuremath{\mathcal{V}}) \cup ((\cup _{m\geq n+1} (-\ensuremath{\mathcal{P}}_m)) + \ensuremath{\mathcal{V}}) \cup (\ensuremath{\mathcal{P}}_{n+1} + (\ensuremath{\mathcal{V}} \setminus \{t_n\})).$$ So $t_n+p$ belongs to $( (\cup_{0 \leq m \leq n} (\ensuremath{\mathcal{P}}_m \cup (-\ensuremath{\mathcal{P}}_m))) + \ensuremath{\mathcal{V}} ) \cup ((\ensuremath{\mathcal{P}}_{n+1} \setminus\{p\}) + \{t_n\})$. Since $p\in \ensuremath{\mathcal I}_{n+1}$, it follows that $t_n + p$ lies in $(\ensuremath{\mathcal{P}}_{n+1} \setminus\{p\}) + \{t_n\}$, which yields $p\in \ensuremath{\mathcal{P}}_{n+1} \setminus\{p\}$. This contradicts the hypothesis that $\ensuremath{\mathcal{P}}$ is not a minimal complement of $\ensuremath{\mathcal{V}}$. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(5)] Define the subsets $\ensuremath{\mathcal{Q}}_n$ of $\ensuremath{\mathbb{Z}}^d$ for $n\geq 0$ as follows. $$\ensuremath{\mathcal{Q}}_n = \begin{cases} \{-t_0\} & \text{ if } n = 0,\\ \{ - t_{n-1} \} + (\ensuremath{\mathcal I}_{n-1} \setminus ( (\cup_{0\leq m \leq n-1} (\ensuremath{\mathcal{Q}}_m \cup (-\ensuremath{\mathcal{Q}}_m))) + \ensuremath{\mathcal{V}})) & \text{ if } n \geq 1. \end{cases} $$ Define the subset $\ensuremath{\mathcal{Q}}$ of $\ensuremath{\mathbb{Z}}^d$ by $$\ensuremath{\mathcal{Q}} : = \cup _{n\geq 0} (\ensuremath{\mathcal{Q}}_n \cup (-\ensuremath{\mathcal{Q}}_n)).$$ For $n\geq 0$, the inclusions \begin{align*} \ensuremath{\mathcal{Q}} + \ensuremath{\mathcal{V}} & \supseteq (\ensuremath{\mathcal{Q}}_{n+1} + t_n) \cup ( (\cup_{0\leq m \leq n} (\ensuremath{\mathcal{Q}}_m \cup (-\ensuremath{\mathcal{Q}}_m))) + \ensuremath{\mathcal{V}})\\ & \supseteq (\ensuremath{\mathcal I}_{n} \setminus ( (\cup_{0\leq m \leq n} (\ensuremath{\mathcal{Q}}_m \cup (-\ensuremath{\mathcal{Q}}_m))) + \ensuremath{\mathcal{V}}))\cup ( (\cup_{0\leq m \leq n} (\ensuremath{\mathcal{Q}}_m \cup (-\ensuremath{\mathcal{Q}}_m))) + \ensuremath{\mathcal{V}})\\ & \supseteq \ensuremath{\mathcal I}_n \end{align*} hold and hence $-\ensuremath{\mathcal I}_n \subseteq (-\ensuremath{\mathcal{Q}}) + (-\ensuremath{\mathcal{V}}) = \ensuremath{\mathcal{Q}} + \ensuremath{\mathcal{V}}$. Moreover, the inclusions $$ \ensuremath{\mathcal{Q}} + \ensuremath{\mathcal{V}} \supseteq \ensuremath{\mathcal{Q}}_{0} + \ensuremath{\mathcal{V}} \supseteq \{0\}$$ hold. It follows that $\ensuremath{\mathcal{V}}$ is an additive complement to $\ensuremath{\mathcal{Q}}$. We claim that $\ensuremath{\mathcal{V}}$ is a minimal complement of $\ensuremath{\mathcal{Q}}$. Since $t_n > 3t_{n-1}$ for $n\geq 1$, from Proposition \ref{Prop:M}, it follows that for $n\geq 0$, no point of $\ensuremath{\mathcal{Q}} \times \ensuremath{\mathcal{V}}$ other than $(-2t_n, t_n)$ goes to $-t_n$ under the addition map $\ensuremath{\mathcal{Q}} \times \ensuremath{\mathcal{V}} \to \ensuremath{\mathbb{Z}}^d$, and hence no point of $\ensuremath{\mathcal{Q}} \times \ensuremath{\mathcal{V}}$ other than $(2t_n, -t_n)$ goes to $t_n$ under the addition map $\ensuremath{\mathcal{Q}} \times \ensuremath{\mathcal{V}} \to \ensuremath{\mathbb{Z}}^d$. Thus $\ensuremath{\mathcal{V}}$ is a minimal complement of $\ensuremath{\mathcal{Q}}$. We claim that $\ensuremath{\mathcal{Q}}$ is a minimal complement to $\ensuremath{\mathcal{V}}$. On the contrary, let us assume that $\ensuremath{\mathcal{Q}}$ is not a minimal complement to $\ensuremath{\mathcal{V}}$. Hence $\ensuremath{\mathcal{Q}}\setminus \{q\}$ is an additive complement to $\ensuremath{\mathcal{V}}$ for some $q\in \ensuremath{\mathcal{Q}}$. Since $\ensuremath{\mathcal{Q}}$ is symmetric, we may assume that $q\in \ensuremath{\mathcal{Q}}_{n+1}$ for some $n\geq -1$. Since no point of $\ensuremath{\mathcal{Q}} \times \ensuremath{\mathcal{V}}$ other than $(-t_0, -t_0)$ goes to $-2t_0$ under the addition map $\ensuremath{\mathcal{Q}} \times \ensuremath{\mathcal{V}} \to \ensuremath{\mathbb{Z}}^d$, it follows that $q \neq -t_0$, i.e., $q\notin \ensuremath{\mathcal{Q}}_0$. So $q$ lies in $\ensuremath{\mathcal{Q}}_{n+1}$ for some $n\geq 0$. Thus $t_n +q$ lies in $\ensuremath{\mathcal I}_n$. It follows from Proposition \ref{Prop:M} that no element of $\ensuremath{\mathcal I}_n$ lies in $$ ((\cup _{m\geq n+2} \ensuremath{\mathcal{Q}}_m) + \ensuremath{\mathcal{V}}) \cup ((\cup _{m\geq n+1} (-\ensuremath{\mathcal{Q}}_m)) + \ensuremath{\mathcal{V}}) \cup (\ensuremath{\mathcal{Q}}_{n+1} + (\ensuremath{\mathcal{V}} \setminus \{t_n\})).$$ So $t_n+q$ belongs to $( (\cup_{0 \leq m \leq n} (\ensuremath{\mathcal{Q}}_m \cup (-\ensuremath{\mathcal{Q}}_m))) + \ensuremath{\mathcal{V}} ) \cup ((\ensuremath{\mathcal{Q}}_{n+1} \setminus\{q\}) + \{t_n\})$. Since $q\in \ensuremath{\mathcal I}_{n+1}$, it follows that $t_n + q$ lies in $(\ensuremath{\mathcal{Q}}_{n+1} \setminus\{q\}) + \{t_n\}$, which yields $q\in \ensuremath{\mathcal{Q}}_{n+1} \setminus\{q\}$. This contradicts the hypothesis that $\ensuremath{\mathcal{Q}}$ is not a minimal complement of $\ensuremath{\mathcal{V}}$. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(6)] Let $G$ be as in the proof of Theorem \ref{Thm}(3). Since $G + \ensuremath{\mathcal{T}} = \ensuremath{\mathbb{Z}}^d$, it follows that $\ensuremath{\mathcal{V}}$ is an additive complement to $G$. We claim that $\ensuremath{\mathcal{V}}$ is a minimal complement of $G \setminus \{- t_n - 3t_{n-1} \,|\,n \geq 1\}$. For $n\geq 1$, no element of $G \times \ensuremath{\mathcal {W}}$ other than $(-2t_{n-1}, -t_{n-1}), (- t_n - 3t_{n-1}, t_n)$ goes to $-3t_{n-1}$ under the addition map $G\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$, and no element of $G \times \ensuremath{\mathcal {W}}$ other than $(- t_n - 4t_{n-1}, t_n)$ goes to $-4t_{n-1}$ under the addition map $G\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. Moreover, no element of $G\times \ensuremath{\mathcal {W}}$ other than $(-2t_0, t_0), (-t_0, 0)$ goes to $-t_0$ under the addition map $G\times \ensuremath{\mathcal {W}} \to \ensuremath{\mathbb{Z}}^d$. It follows that $\ensuremath{\mathcal{V}}$ is a minimal complement to $G \setminus \{- t_n - 3t_{n-1} \,|\,n \geq 1\}$. Each element of $\ensuremath{\mathbb{Z}}^d$ can be expressed as a sum of an element of $G$ and an element of $\ensuremath{\mathcal {W}}$ only in finitely many ways. In particular, each element of $\ensuremath{\mathbb{Z}}^d$ can be expressed as a sum of an element of $G \setminus \{- t_n - 3t_{n-1} \,|\,n \geq 1\}$ and an element of $\ensuremath{\mathcal{V}}$ only in finitely many ways. By Lemma \ref{Lemma:Finiteness}, it follows that some nonempty subset $\ensuremath{\mathcal{R}}$ of $G \setminus \{- t_n - 3t_{n-1} \,|\,n \geq 1\}$ is a minimal complement to $\ensuremath{\mathcal{V}}$. Since $\ensuremath{\mathcal{V}}$ is a minimal complement to $G \setminus \{- t_n - 3t_{n-1} \,|\,n \geq 1\}$, it follows that $\ensuremath{\mathcal{V}}$ is a minimal complement to $\ensuremath{\mathcal{R}}$. Hence $(\ensuremath{\mathcal{R}}, \ensuremath{\mathcal{V}})$ is a co-minimal pair. \end{proof} \begin{proof}[Proof of Theorem \ref{Thm}(7)] Let $G$ be the set as in the proof of Theorem \ref{Thm}(3). Note that $\ensuremath{\mathcal{T}}$ is an additive complement of $G$. Let $n\geq 0$ be an integer. Proposition \ref{Prop:M} implies that $$ - t_n \notin (\cup_{m \geq n+2} G_m + \ensuremath{\mathcal{T}}) \cup (G_{n+1} + (\ensuremath{\mathcal{T}} \setminus \{t_{n}\})). $$ Note that the inclusions \begin{align*} G_m + \ensuremath{\mathcal{T}} & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - t_m } + \ensuremath{\mathbb{Z}}^d_{\geq t_0} \\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq - t_m + t_0}\\ & \subseteq \ensuremath{\mathbb{Z}}^d_{\geq -t_{n} + t_0} \end{align*} hold for any $m\leq n$. So $$ - t_n \notin (\cup_{m \neq n, n+1} G_m + \ensuremath{\mathcal{T}}) \cup (G_{n+1} + (\ensuremath{\mathcal{T}} \setminus \{t_{n}\})). $$ Hence $\ensuremath{\mathcal{T}}$ is a minimal complement of $G$. Since each element of $\ensuremath{\mathbb{Z}}^d$ can be expressed as a sum of an element of $G$ and an element of $\ensuremath{\mathcal{V}}$ only in finitely many ways, it follows that each element of $\ensuremath{\mathbb{Z}}^d$ can be expressed as a sum of an element of $G$ and an element of $\ensuremath{\mathcal{T}}$ only in finitely many ways. By Lemma \ref{Lemma:Finiteness}, some nonempty subset $\ensuremath{\mathcal{S}}$ of $G$ is a minimal complement to $\ensuremath{\mathcal{T}}$. Since $\ensuremath{\mathcal{T}}$ is a minimal complement to $G$, it follows that $(\ensuremath{\mathcal{S}}, \ensuremath{\mathcal{T}})$ is a co-minimal pair. \end{proof} \begin{proof} [Proof of Theorem \ref{Thm:UncountableSubsets}] Let $G$ be a finitely generated abelian group. Then $G$ is isomorphic to the direct product of a finite group $G_\ensuremath{\mathrm{tors}}$ and a free abelian group $\ensuremath{\mathbb{Z}}^d$. Given any infinite subset $X$ of $G$, it contains an infinite subset $Y$ such that the projections of all the elements of $Y$ to $G_\ensuremath{\mathrm{tors}}$ are equal. It suffices to show that any infinite subset of $\ensuremath{\mathbb{Z}}^d$ has uncountably many subsets which admit minimal complements. Let $X$ be an infinite subset of $\ensuremath{\mathbb{Z}}^d$. Let $S$ be the subset consisting of the integers $1\leq i\leq d$ such that the absolute values of the $i$-th coordinate of the elements of $X$ form an unbounded set. Note that $S$ is nonempty. Thus $X$ has an infinite subset $Y$ such that the $i$-th coordinate of all the elements of $Y$ are equal for any $i\in \{1, 2, \cdots, d\}\setminus S$, and the absolute values of the $i$-th coordinate of the elements of $Y$ form an unbounded set for any $i\in S$. Replacing $X$ by one of its translate, we may assume that the $i$-th coordinates of the elements of $Y$ are equal to $0$ for any $i\in \{1, 2, \cdots, d\}\setminus S$. Replacing $X$ by one of its image under an automorphism of $\ensuremath{\mathbb{Z}}^d$, we may assume that $Y$ has a subset $Z$ such that the $i$-th coordinates of the elements of $Z$ form an infinite subset of $\ensuremath{\mathbb{Z}}_{\geq 1}$ for any $i\in \{1, 2, \cdots, d\}\setminus S$. Let $d'$ denote the cardinality of $S$. It suffices to show that if $A$ an infinite subset of $\ensuremath{\mathbb{Z}}^{d'}$ such that for any $1\leq i \leq d'$, the $i$-th coordinates of the points of $A$ form an infinite subset of $\ensuremath{\mathbb{Z}}_{\geq 1}$, then $A$ has uncountably many subsets which admit minimal complements. Note that there is a sequence $\{x_n\}_{n\geq 0}$ contained in $A$ such that $x_n\geq 6x_{n-1}$ for all $n\geq 1$. For any subsequence $\{x_{n_k}\}$ of this sequence, the inequality $x_{n_k} \geq 6 x_{n_{k-1}}$ holds for any $k\geq 1$ and moreover, $x_{n_0} > 0$ holds and all the coordinates of any term of this subsequence are positive. By Theorem \ref{Thm}, the subset $\{x_{n_k} \,|\, k \geq 0\}$ of $A$ admits a minimal complement in $\ensuremath{\mathbb{Z}}^{d'}$. Since the finite subsets of $\ensuremath{\mathbb{N}}$ are precisely the bounded subsets of $\ensuremath{\mathbb{N}}$, it follows that the number of finite subsets of $\ensuremath{\mathbb{N}}$ is countable. Thus the number of infinite subsets of $\ensuremath{\mathbb{N}}$ is uncountable. So $\{x_n\}_{n\geq 0}$ has uncountably many subsequences. It follows that $A$ has uncountably many subsets which admit minimal complements. \end{proof} \begin{proof} [Proof of Theorem \ref{Thm:Uncountable}] Since the finite subsets of $\ensuremath{\mathbb{N}}$ are precisely the bounded subsets of $\ensuremath{\mathbb{N}}$, it follows that the number of finite subsets of $\ensuremath{\mathbb{N}}$ is countable. Thus the number of infinite subsets of $\ensuremath{\mathbb{N}}$ is uncountable. So any sequence has uncountably many subsequences. Note that if $\{x_n\}_{n\geq 0}$ is a sequence in $\ensuremath{\mathbb{Z}}^d$ such that $x_0 = (1, \cdots, 1)$ and $$x_n \geq 6 x_{n-1} \quad \text{ for all } n \geq 1,$$ then any subsequence of $\{x_n\}_{n\geq 0}$ satisfies the hypothesis of each of the seven parts of Theorem \ref{Thm}. Thus Theorem \ref{Thm:Uncountable} follows from the existence of a sequence $\{x_n\}_{n\geq 0}$ in $\ensuremath{\mathbb{Z}}^d$ such that $x_0 = (1, \cdots, 1)$ and $$x_n \geq 6 x_{n-1} \quad \text{ for all } n \geq 1,$$ which exists, for instance, consider the sequence $x_n = (6^n, \cdots, 6^n)$. \end{proof} As an application of Theorem \ref{Thm}, we provide several examples of subsets of $\ensuremath{\mathbb{Z}}$ each of which is a part of a co-minimal pair. \begin{corollary} For each of the following subsets $S_{1},S_{2}$, there exists subsets $S'_{1},S'_{2}$ such that $(S_{1},S'_{1})$ and $(S_{2},S'_{2})$ form co-minimal pairs. \begin{enumerate} \item $S_{1} := \{n^k \,|\, k \geq 0\}$ for any $n\geq 3$. \item $S_{2}:= \{2^k + k\,|\, k \geq 0\}$. \end{enumerate} \end{corollary} \begin{proof} The growth condition of the elements of the subsets (considered as sequences) in this corollary clearly satisfies the condition of Theorem \ref{Thm}. \end{proof} \section{Acknowledgements} The first author would like to thank the Department of Mathematics at the Technion where a part of the work was carried out. The second author would like to acknowledge the Initiation Grant from the Indian Institute of Science Education and Research Bhopal, and the INSPIRE Faculty Award from the Department of Science and Technology, Government of India. \def\cprime{$'$} \def\Dbar{\leavevmode\lower.6ex\hbox to 0pt{\hskip-.23ex \accent"16\hss}D} \def\cfac#1{\ifmmode\setbox7\hbox{$\accent"5E#1$}\else \setbox7\hbox{\accent"5E#1}\penalty 10000\relax\fi\raise 1\ht7 \hbox{\lower1.15ex\hbox to 1\wd7{\hss\accent"13\hss}}\penalty 10000 \hskip-1\wd7\penalty 10000\box7} \def\cftil#1{\ifmmode\setbox7\hbox{$\accent"5E#1$}\else \setbox7\hbox{\accent"5E#1}\penalty 10000\relax\fi\raise 1\ht7 \hbox{\lower1.15ex\hbox to 1\wd7{\hss\accent"7E\hss}}\penalty 10000 \hskip-1\wd7\penalty 10000\box7} \def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth \lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace} \providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR } \providecommand{\MRhref}[2]{% \href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2} } \providecommand{\href}[2]{#2}

All relevant data are within the paper and its Supporting Information files. Introduction {#sec001} ============ Social organization is an important determinant of a species' survival \[[@pone.0202407.ref001]\], foraging efficiency \[[@pone.0202407.ref002]\] and ability to reproduce successfully \[[@pone.0202407.ref003]\], factors which ultimately affect their potential to recover from disturbances. Social structures may be as simple as monogamous pairing or as complex as a eusocial colony with division of labour and non-reproductive castes. Social organization may be influenced by broad ecological \[[@pone.0202407.ref004]\] or life-history factors \[[@pone.0202407.ref005]\], within-group social interactions \[[@pone.0202407.ref006]\] or genetic relatedness \[[@pone.0202407.ref007]\] and even individual variation in physiology \[[@pone.0202407.ref008]\], neurophysiology and genetics \[[@pone.0202407.ref009]\]. Each social structure provides benefits to its constituents, but often at a cost to their reproduction or access to some other resource \[[@pone.0202407.ref010]--[@pone.0202407.ref012]\]. That is to say, there are trade-offs associated with different social structures that individuals must consider. Group living is thought to have evolved in many lineages as a response to genetic (kinship) and environmental factors \[[@pone.0202407.ref010]--[@pone.0202407.ref012]\]. With respect to environmental factors, many hypotheses point toward variability in ecological factors as influencing the evolution of sociality \[[@pone.0202407.ref004], [@pone.0202407.ref013]\]. Hypotheses such as the benefits of philopatry \[[@pone.0202407.ref014]--[@pone.0202407.ref016]\] and ecological constraints models \[[@pone.0202407.ref004], [@pone.0202407.ref016]--[@pone.0202407.ref019]\] examine the idea that ecological factors, such as habitat quality (e.g. habitat size, resource availability, defence) or the availability of suitable breeding territory (respectively), influence the decision of subordinates to either disperse from their habitat or remain within a group. These two hypotheses are often viewed as two sides of the same coin as they both look at aspects of ecology to explain social evolution and maintenance \[[@pone.0202407.ref020]\]. The benefits of philopatry hypothesis focuses on the benefits conferred from residing in a high-quality habitat (e.g. inheritance of breeding status \[[@pone.0202407.ref021]\], increased fitness \[[@pone.0202407.ref022]\]). High-quality habitat is typically colonized rapidly \[[@pone.0202407.ref013]\]. An individual living on low-quality habitat may therefore increase its fitness by moving to a high-quality habitat as a subordinate \[[@pone.0202407.ref013]\]. However, this benefit must be traded off against the associated costs (e.g. delayed reproduction, risk of movement). In contrast, the ecological constraints hypothesis concentrates on factors of ecology that may restrict subordinate individuals already residing in a group from dispersing (e.g. habitat saturation \[[@pone.0202407.ref023]\], predation risk \[[@pone.0202407.ref024]\]). These two hypotheses are not mutually exclusive and often operate alongside other effects (e.g. kinship, life-history). However, the question of which combination of effects best describes social group formation and maintenance is still of interest as each one emphasizes different costs and benefits \[[@pone.0202407.ref025]\]. While these hypotheses have been well studied in terrestrial organisms, they have only recently been tested in marine environments \[[@pone.0202407.ref020], [@pone.0202407.ref026]--[@pone.0202407.ref028]\]. Of the marine taxa tested so far, habitat-specialist coral-reef fishes are emerging as a useful model species to study theories of social evolution and maintenance \[[@pone.0202407.ref020], [@pone.0202407.ref027], [@pone.0202407.ref028]\] and have shown similar responses to habitat manipulation as terrestrial species (e.g. \[[@pone.0202407.ref023]\]). Many social fishes have a pelagic larval phase which suggests low levels of kinship within groups, reducing the potential confounding factor of relatedness (e.g. \[[@pone.0202407.ref020], [@pone.0202407.ref029], [@pone.0202407.ref030]\] but see \[[@pone.0202407.ref031]\]). Given the apparent influence of ecological factors on the formation and maintenance of social groups, we would expect that disturbances capable of altering a species' habitat, such as severe weather events, would have a strong impact on social organization \[[@pone.0202407.ref023], [@pone.0202407.ref032]\]. Many species of coral-reef fishes, especially habitat-specialists, can be found in social groups \[[@pone.0202407.ref033], [@pone.0202407.ref034]--[@pone.0202407.ref036]\]. The size of these social groups is often related directly or indirectly to the size of the habitat in which they reside \[[@pone.0202407.ref036]--[@pone.0202407.ref039]\]. Complex social structures such as size-based dominance hierarchies, in which the largest dominants breed and smaller subordinates are reproductively suppressed, have been documented in these groups \[[@pone.0202407.ref023], [@pone.0202407.ref038]\]. Further, they are known to exhibit sequential hermaphroditism or bi-directional sex-change \[[@pone.0202407.ref038], [@pone.0202407.ref040], [@pone.0202407.ref041]\]. In such systems, the loss of a breeding individual results in the next subordinate in the queue taking its place \[[@pone.0202407.ref006]\]. This social organization may provide a level of redundancy which could help a social species recover quickly following a major disturbance. For example, Rubenstein \[[@pone.0202407.ref042]\] argued that cooperative breeding could be a bet hedging strategy in variable environments as it may buffer variance in fecundity between years. Duffy and Macdonald \[[@pone.0202407.ref026]\] also found that eusociality conferred advantages to sponge-dwelling shrimps allowing them to occupy a greater number of host sponge species and more sponges overall than less social sister species. This finding, combined with research on host specialization and extinction by Munday \[[@pone.0202407.ref043]\], could imply that more social species face lower extinction risk following a disturbance because their sociality allows them to monopolize a greater host range. However, Courchamp et al. \[[@pone.0202407.ref044]\] found that obligate cooperative breeders were more at risk of group extinction because of their reliance on subordinates to reproduce and survive. These studies show that while complex social structures may provide advantages allowing species to survive a severe disturbance and to re-colonize afterwards, they may also result in localized extinctions. Further research into the effects of ecological disturbance on social organization and how varying social systems, such as pair- or group-forming, are able to cope with disturbance are clearly required. Extreme climatic events such as tropical cyclones are known to have devastating impacts on the physical structure of coral reefs \[[@pone.0202407.ref045]--[@pone.0202407.ref048]\]. The effects on fish and invertebrate communities which depend on the coral structure for food and shelter are likewise devastating \[[@pone.0202407.ref045], [@pone.0202407.ref049]--[@pone.0202407.ref051]\]. The destructive forces of cyclones can have a strong influence on the re-distribution of species and their relative abundances following the event \[[@pone.0202407.ref052]\]. However, relatively little is known about the impacts that cyclones may have on the social organization of coral-reef inhabitants and whether social organization may mediate disturbance-induced population trends in species with different social structures. Given the importance of social organization for factors such as reproduction \[[@pone.0202407.ref003]\], foraging efficiency \[[@pone.0202407.ref002]\] and ultimately the ability to recover from a major disturbance, it is plausible that destructive events such as tropical cyclones may have a detectable effect on a species' social organization. We evaluated the effects of cyclones on the social organization of coral gobies of the genus *Gobiodon*. These species are small (3--4 cm) microbenthic \[[@pone.0202407.ref036]\] habitat-specialist fishes that live within the structures of branching and plate-forming acroporid and pocilloporid corals \[[@pone.0202407.ref053], [@pone.0202407.ref054]\]. These fishes are highly site attached once settled, but have been shown to move between corals \[[@pone.0202407.ref041]\]. *Gobiodon* spp. display a wide variety of social phenotypes from pair-forming (PF) species to group-forming (GF) species that typically live in groups ranging from 3 to 12 individuals \[[@pone.0202407.ref055]\]. Social groups usually consist of two breeding individuals and one or more non-breeding subordinates which form a size-based hierarchy and queue for a breeding position. However there is some evidence to support multiple breeding individuals in larger group sizes for some species \[[@pone.0202407.ref055]\]. In this study, we investigated how extreme climatic events influence the social organization of colonies of *Gobiodon* fishes and discuss how these effects may impact their survival. Opportunistic investigations of such disturbances (extreme climatic events) are important for theory development as they can test well developed theory under extreme conditions \[[@pone.0202407.ref056]\]. Specifically, we examined the effects of two successive category 4 cyclones that impacted the Great Barrier Reef, on the group size (social structure) and coral size (ecological factor) of GF and PF species of *Gobiodon*. As habitat patch size is known to be related to mean group size in some species, smaller corals should be less capable of supporting larger groups \[[@pone.0202407.ref055]\]. Therefore, we expected that physical damage caused by the cyclones would result in smaller corals, and that as coral size decreased, so too would mean group size of both GF and PF coral gobies. We also expected that advantages conferred from sociality would help GF species to recover from these disturbances \[[@pone.0202407.ref026], [@pone.0202407.ref043]\]. Additionally, we used the occurrence of these cyclones as a 'natural experiment' to examine the related effects of ecological constraints and benefits of philopatry on the formation of social groups in the GF species. Munday \[[@pone.0202407.ref057]\] demonstrated that coexistence between two species of *Gobiodon* occurred through a competitive lottery, meaning that whichever species colonized a particular coral was able to hold that territory. Our own observations show that while coral gobies do show distinct preferences for certain species of coral, they can and will colonize a wide range of species. It is therefore likely that *Gobiodon* species will colonize any available habitat following a severe disturbance. If ecological constraints (lack of available habitat) were responsible for the formation of social groups, we would expect coral vacancy to be very low as gobies would preferentially colonize vacant habitat over residing as a subordinate in a group. That is, subordinates should disperse to seek independent breeding opportunities if there is suitable vacant habitat. In contrast, if benefits of philopatry were driving group living, we would expect greater coral vacancy as the GF species would vacate lower-quality corals in favor of taking up residence as a subordinate in higher-quality corals. While we do not fully understand what constitutes high- or low quality habitat in these species, we consider coral size to be a reasonable proxy of habitat quality as Kuwamura et al. \[[@pone.0202407.ref058]\] and Hobbs and Munday \[[@pone.0202407.ref059]\] demonstrated that growth, survival and reproductive success increased in larger habitats for other species of coral associated fishes. Materials and methods {#sec002} ===================== Ethics statement {#sec003} ---------------- This research was conducted under research permits issued by the Great Barrier Reef Marine Park Authority (G13/36197.1 and G15/37533.1) and with the approval of the University of Wollongong Animal Ethics Committee (AE14-04). Study area and survey sites {#sec004} --------------------------- The study took place at Lizard Island, Great Barrier Reef, Queensland, Australia (14° 40.729' S, 145° 26.907' E) ([Fig 1](#pone.0202407.g001){ref-type="fig"}) between 2014 and 2016. Twenty three sites were surveyed in total over four survey times, eleven of which were located within the sheltered lagoon. The remaining twelve sites were located on the fringing reefs around Lizard Island. As this study was designed to examine how sociality of *Gobiodon* spp. varied over successive impacts at Lizard Island as a whole, we did not assess variation in sociality at smaller spatial scales (e.g. sites). As such, survey sites were chosen to give reasonable coverage of the reefs at Lizard Island. Not all sites were assessed during each survey time as several sites were scoured down to bare rock after each cyclone. These sites were not surveyed as our interest was in the surviving goby colonies (see [S1 Data](#pone.0202407.s001){ref-type="supplementary-material"} for the range of sites covered at each survey time). The number of sites visited during each survey time was 15, 14, 11, 17 respectively. All measurements were made on scuba at depths ranging from less than one meter to five meters. {#pone.0202407.g001} Cyclone activity and sampling periods {#sec005} ------------------------------------- Two cyclones impacted the study site in consecutive years. Cyclone Ita impacted Lizard Island in April 2014 as a category 4 system and cyclone Nathan in March 2015, also as a category 4 system. Both cyclones caused substantial damage to the fringing and lagoonal reefs including greatly reduced coral cover and associated changes in reef fish diversity and abundance \[[@pone.0202407.ref060]--[@pone.0202407.ref063]\]. We conducted surveys on coral sizes and group sizes of 13 *Gobiodon* spp. during February and March 2014 (1 month prior to cyclone Ita), August and September 2014 (4 months after cyclone Ita), January and February 2015 (1 month prior to cyclone Nathan and 9 months after cyclone Ita) and January and February 2016 (10 months after cyclone Nathan) ([Fig 2](#pone.0202407.g002){ref-type="fig"}). These repeated surveys provided us with a broad overview of the effects that multiple disturbances had on the social organization of coral gobies. {#pone.0202407.g002} Survey methods {#sec006} -------------- Two types of transects were deployed over the four surveys. For this study however, we did not attempt to assess any spatial patterns between sites. Transects were only used as a guide to locate corals. Haphazardly placed 30 m line transects were used to locate corals one meter either side during the first and fourth survey times. Cross transects (two 4 m x 1 m belt transects laid in a cross, designed to measure the community around a focal colony) were used during the first (Palfrey reef only; [Fig 1](#pone.0202407.g001){ref-type="fig"}, sites 4 and 4a), second, third and fourth survey times. Line transects were placed roughly parallel to each other and separated by at least 10 m and cross transects were placed at least 8 m (twice the length of the transect on either axis) from each other to ensure that any given coral was not measured twice during the survey period. As coral gobies show strong preferences for certain species of branching and plate forming (mostly) Acroporid corals \[[@pone.0202407.ref035], [@pone.0202407.ref053], [@pone.0202407.ref064]\], only these species of corals were counted on the transects. In total, 23 species of coral were surveyed for goby occupancy ([S1 Data](#pone.0202407.s001){ref-type="supplementary-material"}). Each coral's living part was measured along three axes (length (L) width (W) and height (H)) and the simple average diameter calculated as (L + W + H)/3. Simple average diameter was used in this study (as opposed to geometric mean diameter (L x W x H)^1/3^) as it provides a better representation of the major axis of the coral \[[@pone.0202407.ref058]\]. All goby supporting corals occurring on the transects were measured and searched for gobies. The number of adult gobies living within each coral head was counted by visual inspection using a torch. Adults could be easily distinguished from juveniles by their distinct coloration and markings. While the number of juveniles (if present) was recorded for each coral, they were not included in the group size observations as juveniles had been observed moving between multiple corals during each survey (Hing pers. obs.). Additionally, juvenile abundance was extremely low during all surveys and there was no difference in abundance for either PF or GF species during any survey time ([S2 Fig](#pone.0202407.s003){ref-type="supplementary-material"}). In contrast, adults displayed remarkable coral-host fidelity, even tolerating extreme hypoxia and severe coral bleaching \[[@pone.0202407.ref065], [@pone.0202407.ref066]\]. The number of transects at each site varied depending on the size of the reef. The number of transects conducted at each site also varied from year to year depending on the perceived abundance of suitable corals for habitation, and ranged from 1 to 44 transects. In total, the number of transects placed around Lizard Island during each survey time was 56, 141, 109 and 140 for the February 2014, August 2014, January 2015 and January 2016 surveys respectively. The methods of measuring goby group sizes and coral sizes (described in detail below) remained exactly the same regardless of the different number and size of transects that were used throughout the study. There was a significant difference in coral size measured between the two transect types, however this was most likely a site effect as line transects were used extensively on the fringing reefs in January 2016, after both cyclones. Corals at these sites sustained heavy damage and were therefore smaller on average. We therefore pooled the data from both types of transect and included site as a random effect in the statistical models. Sociality in *Gobiodon* {#sec007} ----------------------- We documented 15 species of *Gobiodon* at Lizard Island during the present study ([Table 1](#pone.0202407.t001){ref-type="table"}). However, two species (sp. A and sp. D) were excluded from later analyses as they were uncommon at the study site. The remaining species displayed a range of sociality ranging from solitary individuals to pairs and groups reaching up to 21 individuals. From here on we will use the term "group" to refer to any colony with a group size of three or more. We used a sociality index formulated by Avilés and Harwood \[[@pone.0202407.ref067]\] to categorize each species as either GF or PF: $$Sociality = \frac{\left( \frac{A_{d}}{A_{a}} \right) + \left( \frac{N_{g}}{N_{g} + N_{p} + N_{i}} \right) + \left( \frac{I_{n}}{I_{r} + I_{n}} \right)}{3}$$ Where A~d~ = age at dispersal, A~a~ = age when adulthood is reached, N~g~ = number of groups, N~p~ = number of pairs, N~i~ = number of solitary adults, I~r~ = number of reproducing adults and I~n~ = number of non-reproducing (subordinate) adults. The three components in the numerator of [Eq 1](#pone.0202407.e001){ref-type="disp-formula"} represent the proportion of a species' life-cycle spent in a group, the proportion of groups in the population and the proportion of subordinates in the population (respectively). 10.1371/journal.pone.0202407.t001 ###### List of *Gobiodon* spp. and sociality categorization. {#pone.0202407.t001g} Species Individuals Groups Sociality index Categorization --------------------------------------------------------- ------------- -------- ----------------- ---------------- *G*. *axillaris* 15 9 0.33 Pair *G*. *brochus* 70 35 0.36 Pair *G*. *ceramensis* 36 20 0.36 Pair *G*. *erythrospilus* 138 69 0.41 Pair *G*. *histrio* 79 43 0.43 Pair *G*. *oculolineatus* 59 30 0.39 Pair *G*. *okinawae* 33 19 0.46 Pair *G*. *quinquestrigatus* 114 59 0.38 Pair *G*. *acicularis* 48 17 0.56 Group *G*. *citrinus* 37 9 0.63 Group *G*. *fuscoruber*[^a^](#t001fn002){ref-type="table-fn"} 142 51 0.57 Group *G*. *rivulatus* 145 45 0.65 Group Unknown species 28 8 0.63 Group *Gobiodon* spp. observed at Lizard Island with their social index. The number of individuals and groups of each species recorded during the February 2014 survey are provided. Species were categorized as group-forming (below dotted line) if their social index was greater than 0.5. Otherwise they were categorized as pair-forming (above dotted line). ^a^ *G*. *fuscoruber* is synonymous with *G*. *unicolor* \[[@pone.0202407.ref076]\] Using this equation, we calculated a sociality index for each *Gobiodon* spp., making some necessary but biologically relevant assumptions. Once coral gobies settle onto a coral as juveniles, they are not known to move frequently unless forcefully evicted from the coral \[[@pone.0202407.ref006], [@pone.0202407.ref068]\]. Although we do not have a precise estimate of the age at settlement for each species, Brothers et al. \[[@pone.0202407.ref069]\] estimated the larval life of three species of *Gobiodon* ranging from 22 to 41 days. Given that *Gobiodon* spp. live in the order of years \[[@pone.0202407.ref029], [@pone.0202407.ref070]\], we assume that each species spends the majority of its life-cycle in a single coral. We therefore set the maximum proportion of the life-cycle spent in the group ($\frac{A_{d}}{A_{a}}$) as 1 for each species. While there may be natural variation in this parameter, this assumption is biologically realistic and enables us to make relative comparisons between species primarily based on the remaining two factors in [Eq 1](#pone.0202407.e001){ref-type="disp-formula"}. The last two components of the index were calculated as per [Eq 1](#pone.0202407.e001){ref-type="disp-formula"}. Having calculated the sociality indices, species were categorized as GF if their sociality index was greater than 0.5 and remaining species with sociality indices less than 0.5 were categorized as PF ([Table 1](#pone.0202407.t001){ref-type="table"}). The index value of 0.5 was defined as the cut-off value between PF and GF species because it lies directly in the middle of the observed index range where there was a natural split in the data ([S1 Fig](#pone.0202407.s002){ref-type="supplementary-material"}). It should be noted however that "PF" species were sometimes observed in groups (i.e. 3 or more individuals) and "GF" species were sometimes observed in pairs or as singles. The terminology used here therefore indicates the tendency of particular species to form either groups or pairs. Importantly, calculations of sociality indices and subsequent categorization was based on data from surveys obtained before any recent cyclone activity (February 2014). We acknowledge that these reefs have been subjected to Crown of Thorns Starfish (COTS) outbreaks in the past. Our measure of sociality may therefore vary from sociality recorded at other locations. Unfortunately, COTS outbreaks are a relatively frequent occurrence on the Great Barrier Reef and we therefore consider our measure of sociality to be representative of the 'normal' social organization of the species in question. Group size {#sec008} ---------- To assess the effect of cyclone activity on social organization, we used a generalized linear mixed model with a zero-truncated negative binomial distribution to analyze the effects of sociality and survey time and their interaction on the group size of coral gobies. The zero truncated distribution was used as it does not allow predictions of group size less than one. A negative binomial distribution was used to account for over-dispersion which rendered an initially employed zero-truncated Poisson model unsuitable. The model contained survey time (Feb-14, Aug-14, Jan-15 and Jan-16), social organization (PF or GF) and the interaction between these factors as fixed effects. Site, coral species and goby species were included as crossed random effects. Root mean square error (RMSE) was used to assess model performance. RMSE is a measure of the overall agreement between model predictions and the observed data and is measured in the same unit as the response variable. Generalized linear mixed models were conducted in R using the glmmADMB package \[[@pone.0202407.ref071], [@pone.0202407.ref072]\] and pairwise comparisons conducted with the emmeans package \[[@pone.0202407.ref073]\]. Figures were produced using the ggplot2 package \[[@pone.0202407.ref074]\]. Coral size and abundance of empty corals {#sec009} ---------------------------------------- To investigate changes in coral sizes for PF and GF species over the four survey times, we tested the relationship between social organization, survey time and coral size. We used a generalized linear mixed model with survey time, social organization and their interaction as fixed effects and site, goby species and coral species as crossed random effects. A gamma distribution was used to account for positive skew and heteroscedasticy in the data and because it gave a better fit than models conducted with log-normal distributions. RMSE was used to assess model performance. To test the hypothesis that subordinates (i.e. non-reproducing individuals) in colonies of GF species might be constrained by a lack of available habitat, we also assessed whether the mean number of empty corals on a transect was different for transects with or without groups of GF species. We used a generalized linear model for this analysis with the number of empty corals as the dependent variable and survey time and transect type (with or without groups of GF species) as independent variables. The model was run with a zero-inflated negative binomial distribution to handle the large number of zero counts and this produced a better model than a zero-inflated Poisson model when compared with Akaike Information Criterion (AIC). The model was conducted using the R package glmmADMB \[[@pone.0202407.ref071], [@pone.0202407.ref072]\]. Proportion of inhabited corals and probability of coral occupancy {#sec010} ----------------------------------------------------------------- We qualitatively reviewed the mean proportion of corals occupied on each transect to determine whether cyclone activity would change the relative proportion of either social organization's occupancy. Since we expected coral size to change with cyclone activity, we assessed whether coral size (a potential aspect of habitat quality) was related to the type of goby species (PF or GF) that occupied it during each survey. This was examined by assessing the multinomial probability that corals would be inhabited by either GF species, PF species or neither. These data were modeled as a multinomial response with coral size and survey time as predictors. Prior to cyclone Ita (survey 1), these data were only collected at one site (Palfrey; [Fig 1](#pone.0202407.g001){ref-type="fig"}). For each of the remaining time points (surveys 2--4), data were collected from various sites around Lizard Island ([Fig 1](#pone.0202407.g001){ref-type="fig"}; [S1 Data](#pone.0202407.s001){ref-type="supplementary-material"}). Misclassification error is the proportion of false classifications predicted by the model and was used to assess model performance. The multinomial model was conducted in R using the nnet package \[[@pone.0202407.ref075]\]. Results {#sec011} ======= Categorization of social organization {#sec012} ------------------------------------- Of the 13 *Gobiodon* spp. surveyed at Lizard Island, five species were categorized as "GF" species and eight species were classified as "PF" (see above for definitions) ([Table 1](#pone.0202407.t001){ref-type="table"}). Group size {#sec013} ---------- Prior to cyclone Ita (Feb 2014), GF species were observed with mean group sizes of 2.71 (± 0.17 SE) individuals per coral. The mean group size of GF species decreased to 2.13 (± 0.11 SE) following cyclone Ita (Aug 2014). Five months later (Jan 2015, 9 months after cyclone Ita) the mean group size of GF species appeared to show some sign of recovery, increasing to 2.58 ± 0.11 (SE). This trend of recovering group sizes was not evident 10 months after cyclone Nathan (Jan 2016), when mean group sizes for GF species was 2.27 ± 0.15 (SE), similar to those just four months after cyclone Ita. Meanwhile PF species had a mean group size of 1.88 (± 0.05 SE) individuals per coral at the beginning of the study (Feb 2014) and maintained their group sizes at a similar level through both cyclones. The mean group sizes of PF species was 1.81 (± 0.03 SE), 1.74 (± 0.04 SE) and 1.74 (± 0.04 SE) for the Aug 2014, Jan 2015 and Jan 2016 surveys respectively. Group-forming species had larger mean group sizes than PF species at every survey time ([Fig 3](#pone.0202407.g003){ref-type="fig"}), although the difference in group size between GF and PF species reduced substantially following cyclone Ita (Aug 2014; [Fig 3](#pone.0202407.g003){ref-type="fig"}). This was due to the reduction in the mean group size of the GF species after cyclone Ita. These patterns were supported by the statistical model which had a RMSE of 1.36 ([S1 Table](#pone.0202407.s004){ref-type="supplementary-material"}). The model predicted an initial decrease in the mean group size of GF species following cyclone Ita (pairwise comparison ratio 1.58 (Feb-14/group:Aug-14/group, 95% CI (0.76, 1.83)). However, the predicted mean group sizes remained at these lower sizes for the subsequent surveys ([S2 Table](#pone.0202407.s005){ref-type="supplementary-material"}; [Fig 3](#pone.0202407.g003){ref-type="fig"}). The model did show a slight increase in mean group size of GF species in the Jan-15 survey (confidence interval was relatively large; estimated marginal mean 1.78, 95% CI (1.10, 2.81); [Fig 3](#pone.0202407.g003){ref-type="fig"}). {#pone.0202407.g003} Mean group sizes of PF species did not change significantly throughout the study. The statistical model showed very little variation in group size during any survey time ([Fig 3](#pone.0202407.g003){ref-type="fig"}), but predicted lower mean group sizes than observed, ranging from 1.28 ± 0.16 (SE) before the cyclones to 1.09 ± 0.19 (SE) after both cyclones. Coral size {#sec014} ---------- Over each successive survey, the mean size of corals inhabited by GF species, PF species and the mean size of uninhabited corals all decreased (pairwise comparison ratio 1.23 (Feb-14:Jan16, 95% CI (1.16, 1.32); [Fig 4](#pone.0202407.g004){ref-type="fig"}). The number of very large corals (greater than 50 cm mean diameter) also decreased substantially following the first cyclone (cyclone Ita; [Fig 4](#pone.0202407.g004){ref-type="fig"}) and were detected in low numbers in all subsequent surveys. The interaction between sociality and survey time was not significant (analysis of deviance χ^2^ = 3.36, df = 3, P = 0.34), indicating that the coral size decreased at a similar rate across the four survey times for each category of social organization. As this interaction was non-significant, pairwise comparisons were conducted on the main effects only. On average, GF species inhabited larger corals (26.93 ± 0.56 (SE)) than the PF species (19.76 ± 0.19 (SE)) during each survey and the mean size of uninhabited corals (12.86 ± 0.25 (SE)) was always less than that of inhabited corals ([Fig 4](#pone.0202407.g004){ref-type="fig"}). The pattern of decreasing coral size was supported by the statistical model (RMSE = 7.42; [Fig 4](#pone.0202407.g004){ref-type="fig"}). The model also supported the pattern of GF species inhabiting larger corals than PF species on average (pairwise comparison ratio 0.80 (PF:GF), 95% CI (0.61, 1.06); [Fig 5](#pone.0202407.g005){ref-type="fig"}). Vacant corals were smaller than corals inhabited by either PF or GF species (pairwise comparison ratio 0.63 (vacant:PF), 95% CI (0.37, 0.85); pairwise comparison ratio 0.51 (vacant:GF), 95% CI (0.37, 0.69) respectively). {#pone.0202407.g004} {#pone.0202407.g005} To assess whether habitat saturation (an ecological constraint) was acting as a constraint on subordinate dispersal, we looked at whether the number of vacant corals differed between transects with or without groups (colonies with 3 or more individuals) of GF species. Corals that were uninhabited were present on transects where at least one group of GF species was present ([S1 Data](#pone.0202407.s001){ref-type="supplementary-material"}). This means that there was vacant habitat available for subordinates to disperse to. However, there was no difference in the mean number of empty corals on transects with or without a group of GF species during any survey time detected by the model (pairwise comparison ratio 1.19 (no groups:groups present), 95% CI (0.89, 1.44). This could indicate that some coral vacancy was due to reduced abundance of coral gobies overall, but the fact that groups of GF species were present on transects where there were corals available to disperse to demonstrates that either; some constraint was restricting dispersal from the group or subordinate gobies were receiving a benefit from remaining within the group. Proportion of inhabited corals and probability of occupation {#sec015} ------------------------------------------------------------ PF species occupied proportionally more corals on average during each survey than GF species ([Fig 5](#pone.0202407.g005){ref-type="fig"}). There was a similar proportion of corals occupied by GF species as there were vacant corals during each survey. The proportion of corals inhabited by PF species decreased from 0.61 ± 0.04 (SE) at the beginning of the study (Feb 2014) to 0.54 ± 0.02 (SE) after cyclone Ita (Aug 2014). This downward trend continued into the next survey (Jan 2015) where the proportion of corals inhabited by PF species was 0.46 ± 0.02 (SE). However, the proportion of corals inhabited by PF species increased after cyclone Nathan (Jan 2016) to 0.56 ± 0.03 (SE). The GF species on the other hand showed relative stability in the proportion of corals they occupied during the study. There was an initial increase in the proportion of corals inhabited by GF species from 0.14 ± 0.03 (SE) at the beginning of the study to 0.22 ± 0.02 (SE) after cyclone Ita (Aug 2014). The mean proportion of corals occupied by GF species then remained at similar levels for the remaining two surveys ([Fig 5](#pone.0202407.g005){ref-type="fig"}). The proportion of vacant corals was also relatively unchanged throughout the study except for a small increase nine months after cyclone Ita (Jan 2015; 0.31 ± 0.02 (SE); [Fig 5](#pone.0202407.g005){ref-type="fig"}). The multinomial model of coral occupancy had a misclassification rate of 0.404 indicating that the predictions may not be reliable. Nevertheless, the trends agree reasonably well with our observations and we give a qualitative account of these, recognizing that probability estimates may have large error. Odds ratios and associated confidence intervals for the model coefficients are available in [S3 Table](#pone.0202407.s006){ref-type="supplementary-material"}, however, we urge the same caution in their interpretation. Prior to cyclone activity (February 2014), there was a low probability that the smallest corals would remain vacant and this probability decreased rapidly for corals of increasing mean diameter ([Fig 6A](#pone.0202407.g006){ref-type="fig"}). This was consistent with our observations as larger corals were rarely vacant ([Fig 4](#pone.0202407.g004){ref-type="fig"}, pink and blue points). After cyclone Ita, there was a similar pattern of decreasing probability of corals remaining vacant with increasing coral size ([Fig 6A](#pone.0202407.g006){ref-type="fig"}), but there was a higher probability of the smallest corals remaining vacant. Again, this pattern was consistent with our observations of coral size ([Fig 4](#pone.0202407.g004){ref-type="fig"}). The probability that a PF species would occupy a coral increased initially with increasing coral size, but then decreased after reaching an apparent optimal coral size around 15 cm ([Fig 6B](#pone.0202407.g006){ref-type="fig"}, solid orange line). This pattern of increasing to an optimum size is certainly plausible if we consider that GF species typically inhabited the larger corals ([Fig 4](#pone.0202407.g004){ref-type="fig"}) posing an upper restraint on occupancy by PF species. Corals in the smaller range may have been less desirable as they may not support successful feeding, reproduction or protection from predators. Furthermore, the coral size model had predicted the mean coral size for PF species within this coral size range ([Fig 4](#pone.0202407.g004){ref-type="fig"}). The 'optimal' coral size for PF species appeared to increase to 20 cm-- 30 cm in the survey times after cyclone Ita ([Fig 6B](#pone.0202407.g006){ref-type="fig"}). Consistent with the concept of the PF species having lower probability of occupancy at higher coral sizes, the probability that a GF species would occupy a coral increased as coral size increased ([Fig 6C](#pone.0202407.g006){ref-type="fig"}). This relationship between coral size and probability of inhabitance by a GF species did not change with respect to survey time ([Fig 6C](#pone.0202407.g006){ref-type="fig"}). {#pone.0202407.g006} Discussion {#sec016} ========== The effects of cyclones on the social organization of coral-reef fish are poorly understood despite clear links between social organization and factors that could affect species persistence and recovery following environmental disturbances \[[@pone.0202407.ref001]--[@pone.0202407.ref003]\]. Here, we investigated the impacts of two successive cyclones (Ita 2014 and Nathan 2015) on the social organization of coral-gobies over three years, and at the same time shed light on the possible factors influencing the formation of social groups. Effects of cyclones on social organization and coral size {#sec017} --------------------------------------------------------- Both cyclones had a small, but detectable effect on the social organization of GF species. Similar impacts on social organization were not evident in the PF species. The group size of GF species declined, while the group sizes of PF species showed little variation over time. Despite the general decline in their group sizes, GF species exhibited some recovery eight months after cyclone Ita. However, there was no such recovery exhibited after cyclone Nathan. The lack of apparent recovery after cyclone Nathan indicates that multiple impacts of this nature can have longer lasting negative impacts on the social structure of GF species. The relative stability of group sizes in the PF species on the other hand, suggests a level of resilience in social structure in the face of natural disturbance. Overall, mean coral size and the presence of very large corals (greater than 50 cm mean diameter) decreased with each cyclone. This was consistent with damage reported in studies on these cyclones \[[@pone.0202407.ref060]--[@pone.0202407.ref063]\] and others \[[@pone.0202407.ref045]\]. Implications for pair- and group-forming species {#sec018} ------------------------------------------------ The overall reduction in coral sizes meant that both GF and PF species were more frequently observed in corals of smaller sizes including some of a size that were unoccupied before the cyclones (Feb 2014). Therefore, the recovery in group size of GF species following cyclone Ita (Jan 2015) occurred despite the fact that the corals they inhabited were smaller on average compared to pre-cyclone (Feb 2014). This result was unexpected, given the positive relationship between coral size and group size regularly reported for social habitat-specialist reef fishes \[[@pone.0202407.ref055], [@pone.0202407.ref077], [@pone.0202407.ref078]\]. This may indicate that GF species of gobies will tolerate greater coral saturation (i.e. more subordinates in smaller corals) following a disturbance, especially if they benefit in future reproduction or survival from doing so \[[@pone.0202407.ref079]\]. Despite this small recovery following cyclone Ita, group sizes of GF species remained relatively lower following cyclone Nathan. This may be due to social conflict \[[@pone.0202407.ref006]\] and recruitment prevention \[[@pone.0202407.ref037]\], demonstrated in other social fishes at high rates of habitat saturation. Smaller group sizes suggest lower numbers of subordinates which may have a negative impact on future reproductive efforts \[[@pone.0202407.ref079]\]. Smaller group sizes could also be problematic under a regime of repeated disturbance as larger group size may provide a level of redundancy and buffer effects of future disturbance \[[@pone.0202407.ref026], [@pone.0202407.ref042], [@pone.0202407.ref043]\]. However, when group sizes are reduced, so too is this redundancy. The proportion of corals inhabited by PF species did decrease following cyclone Ita, but had returned to pre-cyclone levels in the period following cyclone Nathan. At all survey times, PF species inhabited a substantially higher proportion of corals than GF species. This suggests that PF species might be better able to colonize vacant corals than GF species, for example by out-competing GF species for habitat \[[@pone.0202407.ref057], [@pone.0202407.ref080]\]. However, most of the GF species in our study tended to prefer different species of coral to the PF species and we therefore consider competitive effects unlikely. Instead, the greater proportion of corals inhabited by PF species could be due to their tendency to live in intermediate sized (20--30 cm) corals as shown by our analysis of the probability of occupation by a PF species. Corals in this size range were relatively common in the surveys following cyclone Ita (compared to the larger corals that GF species tend to inhabit). Group-forming species on the other hand showed a relatively lower probability of occupying corals in this intermediate size range. This ability or preference of PF species to occupy corals in the range of sizes most commonly found after the cyclones could be advantageous at the population scale, as long as these habitats were of sufficient quality to enable foraging, protection from predators and successful breeding \[[@pone.0202407.ref058], [@pone.0202407.ref059]\]. Ecological constraints and benefits of philoparty {#sec019} ------------------------------------------------- In theory, subordinates living in a group could maximize their lifetime reproductive success if they dispersed to pursue independent breeding rather than remaining in a group as a subordinate. In practice however, various ecological constraints and benefits of remaining philopatric amongst other factors (e.g. life-history and phylogeny), alter the advantages of dispersing from or remaining in their current group \[[@pone.0202407.ref011], [@pone.0202407.ref012], [@pone.0202407.ref020]\]. For example, a lack of vacant habitat to disperse to in order to pursue independent breeding (an ecological constraint) would increase the benefit of remaining in the group, even as a non-breeding subordinate, especially if the subordinate stands to inherit the breeding position in the future (a benefit of philopatry). Habitat saturation (i.e. lack of available suitable habitat) is often invoked as a key ecological constraint leading to group formation and maintenance in a variety of taxa (e.g. birds \[[@pone.0202407.ref013]\]; mammals \[[@pone.0202407.ref081]\]; fish \[[@pone.0202407.ref020]\]). Other studies on a closely related coral goby \[[@pone.0202407.ref023]\] and on social freshwater fishes \[[@pone.0202407.ref022]\] have found the combination of habitat saturation and benefits of philopatry promote group-living. However, we found little evidence to support habitat saturation acting as a constraint on dispersal in coral gobies following these disturbances. Our analysis of vacant corals on transects with and without groups of GF species indicate that GF groups were present even when alternative corals were available for subordinate dispersal. As we only included corals of a size that pairs of gobies had been observed in, these alternative corals are assumed to be of a size capable of supporting at least a breeding pair. Our study therefore indicates that habitat saturation alone was unlikely to explain group formation. Instead, and consistent with the benefits of philopatry hypothesis, subordinates of GF species stayed within the group, presumably obtaining benefits that group living provides (e.g. inheritance of a breeding position in a good quality habitat). Additionally, our analysis of the probability of occupation showed that GF species were increasingly more likely to inhabit a coral as coral size increased. Coral size has been shown to be related to individual growth, survival and reproductive success in some coral-associated fishes and may therefore be considered a reasonable proxy for habitat quality \[[@pone.0202407.ref058], [@pone.0202407.ref059]\]. This strong association between coral size (quality) and probability of occupancy by a GF species is consistent with the benefits of philopatry model as we would expect larger group sizes (characteristic of more social species) in higher-quality habitat. Conversely, under a habitat saturation model we would expect a much weaker association between coral size and the probability of occupancy by a GF species as subordinates would be expected to disperse to vacant habitat of any size that could support independent breeding. Furthermore, if habitat saturation (availability of corals) was acting as a constraint on dispersal following the cyclone, we would expect the proportion of inhabited corals to approach 100% as subordinates would quickly fill any vacant habitat to pursue independent breeding \[[@pone.0202407.ref013]\]. Alternatively, if there were sufficient benefits of residing in a high-quality habitat, we would expect the proportion of inhabited corals to be substantially lower than 100% after the cyclones as individuals living in low-quality habitat would vacate and take up residence in a higher-quality habitat as a subordinate. We found the proportion of corals inhabited by social species was very low and relatively constant (\< 25%) throughout the study, even though there were vacant corals present (approximately 20% per transect), suggesting that benefits of philopatry and not habitat saturation was responsible for group formation. Conclusion {#sec020} ========== Few studies thus far have examined the effects that extreme climatic events such as tropical cyclones could have on social organization of social species. While two cyclones in consecutive years may be rare, the frequency of the most intense cyclones is projected to increase as sea surface temperatures continue to rise in the future and repeated disturbances may become more prevalent \[[@pone.0202407.ref082]\]. The destructive nature of these events on coral-reef communities has been well documented \[[@pone.0202407.ref046], [@pone.0202407.ref048], [@pone.0202407.ref083]\]. However, changes to social organization from such events have been less studied. Here we demonstrated that repeated cyclones are likely to negatively impact social organization in a genus of coral-reef fishes through flow-on effects of the destruction of habitat, but only in GF species. Pair-forming species appear to be able to monopolize smaller corals and maintain their social organization in response to extreme climatic events. Additionally, we suggest that the most likely mechanism for the maintenance of group sizes in GF species are benefits of philopatry, but these benefits only promote group living when the habitat is of sufficient size. Cyclones are capable of reducing whole areas of coral to well below what appears to be the minimum size threshold for GF coral gobies to form their usual group structures, which may be linked to their ability to recover from such disasters. In fact, we observed several sites that were completely devoid of corals (and hence coral gobies) following each cyclone. With the frequency of more intense cyclones and other stressors on coral reefs (e.g. coral bleaching) set to increase in the near future, population declines and localized extinctions of GF species of coral gobies through habitat loss and lowered recovery ability due to impacts on their social organization are a real possibility. While PF species appear to buffer these effects somewhat, they are still vulnerable to habitat destruction caused by these catastrophic events. Supporting information {#sec021} ====================== ###### Raw Data. Data collected from Lizard Island over four survey times between February 2014 and February 2016. Includes some data from collected by MW and SK in 2013. This 2013 data was only used in the categorization of *Gobiodon* species, not in the statistical models. (CSV) ###### Click here for additional data file. ###### Sociality index for each species of *Gobiodon* observed at Lizard Island. Sociality indices (red dot) calculated for each species. Jittered point clouds indicate the relative number of colonies that were available to calculate the index from. There is a natural split in the species' indices around 0.5. (DOCX) ###### Click here for additional data file. ###### *Gobiodon* juvenile abundance at Lizard Island. Predicted mean juvenile abundance and 95% CI for pair- and group-forming species (pink and blue respectively) across each survey time. Raw data is shown as jittered point clouds. (DOCX) ###### Click here for additional data file. ###### Summary of statistical models. Results from statistical models. Abbreviations are: Akaike Information Criterion (AIC); standard error (SE); degrees of freedom (df); standard deviation (SD); root mean squared error (RMSE). (DOCX) ###### Click here for additional data file. ###### Pairwise comparisons for the fixed effects terms of each of the group size, coral size, empty corals and predicted probabilities of inhabitance. Pairwise comparisons were conducted in R using the emmeans package \[[@pone.0202407.ref001]\]. For a given contrast A/B, ratios greater than 1.00 indicate that A is greater than B and ratios less than 1.00 indicate that B is greater than A. (DOCX) ###### Click here for additional data file. ###### Model coefficients for the multinomial probability of occupancy. Odds and associated confidence interval (CI) for each model coefficient. Vacant corals were the reference group. Odds = 1 indicate an equal chance that the coral would remain vacant or be inhabited by either a pair- or group-forming species. Odds \> 1 indicate a greater chance of the coral being inhabited by either pair- or group-forming species rather than remaining vacant. Odds \< 1 indicates a greater chance of the coral remaining vacant. (DOCX) ###### Click here for additional data file. The authors express our thanks to Grant Cameron and Karen Hing for their help in the field without which this research could not have taken place. We also wish to acknowledge the staff and volunteers of the Lizard Island Research Station for the smooth operation of this vital research hub. This manuscript was prepared with the help of the Australian Coral Reef Society Writing Workshop (2016). Thanks also to Simon Brandl, two anonymous reviewers and the editor, Heather Patterson, for extremely helpful and detailed advice in the preparation of the manuscript. [^1]: **Competing Interests:**The authors have declared that no competing interests exist.

1.. Introduction ================ High Voltage Direct Current (HVDC) transmission systems are more favorable than the conventional High Voltage Alternating Current (HVAC) systems because of their economical, technical, and environmental advantages for long distance and bulk power transmission. To date, a ±800 kV HVDC transmission system has been in commercial operation and a ±1000 kV transmission system is being studied on a research basis in China \[[@b1-sensors-15-03540]\]. The electrical environment under HVDC transmission lines is characterized by several different electrical parameters. These parameters are the electric field, the ion current density, the space charge density, the radio interference, the audible noise and the magnetic field \[[@b2-sensors-15-03540]--[@b4-sensors-15-03540]\]. The electric field under transmission lines is one of the important parameters for evaluation of the electromagnetic environment of concern to people. Electric field measurement systems based on Wireless Sensor Networks (WSNs) have been applied extensively in the electrical power system on account of their convenience. During the process of electric field measurement, it is necessary to obtain the electric field distribution at different locations under the transmission lines, therefore, the location information of each sensor node is crucial to the analysis of the measurement results. The main focus of the research presented in this paper will be how to obtain the node location information. To measure the electric field strength at different positions under transmission lines, electric field sensors are arrayed along the direction perpendicular to the transmission lines. The wireless node is connected with the each electric field sensor through twisted-pair braid shielded cable. The wireless nodes\' primary function is to acquire and process the output analog signals of the sensor and send the digital electric field data to the remote computer. The key requirement to interpret the received electric field data is to determine the locations of the electric field sensors. The construction of the wireless nodes can be reduced to a one-dimensional linear topology where nodes have a spatial ordering. This spatial characteristic indicates the relative locations for nodes and can facilitate the establishment of the corresponding relationship between the wireless node and electric field sensor. The relative location relationship among the wireless nodes can be described as a node location permutation. A relative localization method based on Received Signal Strength Indicator (RSSI) is elaborated to obtain the correct node location permutation to get the ordering of the wireless nodes. Compared with the existing approach which gathers the location information by manually labelling sensors during deployment, automatic localization can benefit the deployment in two aspects. Firstly, in the case of a large number of nodes, the relative localization method would be more efficient and less error prone. Secondly, the relative localization method can easily re-establish the ordering in situations where a failed node is replaced or a new node is added to the current network. The developed localization techniques can be divided into two categories: range-based and range-free \[[@b5-sensors-15-03540]\]. Range-based localization estimates distance or angle between the unknown nodes and the anchor node, while range-free localization exploits radio connectivity to confirm proximity or exploits the sensing capabilities of each sensor. Range-based localization should estimate distance using different methods such as Time of Arrival (TOA) \[[@b6-sensors-15-03540]\], Time Difference of Arrival (TDOA) \[[@b7-sensors-15-03540]\], Angle of Arrival (AOA) \[[@b8-sensors-15-03540]\], and RSSI-based ones \[[@b9-sensors-15-03540]--[@b13-sensors-15-03540]\], and then use the trilateration location algorithm, triangulation location algorithm or maximum likelihood estimation method to calculate the location of the node \[[@b14-sensors-15-03540],[@b15-sensors-15-03540]\]. All four distance measurement methods, except for RSSI-based methods, have superior measuring precision, hence range-based localization can accomplish accurate location. The RSSI-based localization methods can be divided into two kinds: range-based methods and range-free methods. Range-based methods assume that nodes can estimate the distance, which needs extra hardware and thus is not appropriate for large-scale outdoor sensor networks. Range-free methods localize nodes based on simple sensing features, such as wireless connectivity, anchor proximity, or localization events detection \[[@b16-sensors-15-03540],[@b17-sensors-15-03540]\]. Among these, connectivity-based solutions feature a low cost, but at the expense of localization accuracy. Radio Frequency (RF) signals are vulnerable to environmental impact which will lead to the instability of RSSI. The environment under high voltage transmission lines is extremely complex. Electromagnetic interference, topography, vegetation, and many other factors may all have an impact on the RSSI. This will cause a large error when estimating the distance through the direct use of RSSI. Although RSSI is not considered as the best choice for distance estimation in the situation, it does provide some useful distance related information which can be used. The experimental result in the paper confirms that the radio signal strength weakens approximately monotonically with the physical distance and the RSSI could provide potential information about which neighbor node is closer and which is farther. Thus, this paper focuses on the research of range-free methods to obtain the nodes relative location information and accordingly realize the nodes sorting. In \[[@b10-sensors-15-03540]\], a range-free localization beyond connectivity is presented to capture the relative distance between 1-hop neighboring nodes from their neighborhood orderings that serve as unique high-dimensional location signatures for nodes in the network. The signatures are subsequently utilized to estimate the distance between nodes. In \[[@b11-sensors-15-03540]\], one-dimensional localization algorithm based on RSSI is proposed. The algorithm utilizes the ratios of signal strengths, instead of signal strength, to determine the locations of the unknown nodes. The algorithm only needs to establish the data set for the ratio of the signal strengths once. In \[[@b12-sensors-15-03540],[@b13-sensors-15-03540]\], the proposed algorithms assume that closer nodes receive higher RSSI. In \[[@b12-sensors-15-03540]\], the presented algorithm can only be used in the case of the known coordinates of the database, and the wireless sensor network topology is a linear structure or a mesh structure. In \[[@b13-sensors-15-03540]\], the algorithm that does not require the anchor node is proposed to accomplish the relative localization. Based on the aforementioned points for network deployment, a novel and practical RSSI-based localization algorithm for WSN which assumes that closer nodes receive higher RSSI is proposed for the implementation of relative localization. The algorithm is able to correct the exceptional situation of nodes in the output location permutation to improve the accuracy of the localization. As our localization algorithm can correct RSSI-based relative localization, it is named as "Correct RSSI-Based Relative for Localization", or "CRBR-Localization" for short. This paper is organized as follows: in Section 2, the motivation for the CRBR-Localization algorithm is described with empirical data and comparisons with existing methods. In Section 3, the CRBR-Localization algorithm is proposed and elaborated. In Section 4, based on the experimental results, the effectiveness of the algorithm is evaluated in four different deployments. Finally, a conclusion section ends this paper. 2.. Motivation for the Correct RSSI-Based Relative-Localization (CRBR-Localization) Algorithm ============================================================================================= 2.1.. Empirical Data as Motivation ---------------------------------- 1. Received Signal Strength Indicator (RSSI) and physical distance Theoretically, RSSI monotonically varies with the physical distance between the sender node and the receiver node. To evaluate the relationship of RSSI and physical distance, outdoor experiments were conducted in three types of environment: a park lawn, a school playground and the China National Grid HVDC test base. In the experiments, 10 sender nodes were arranged at different distances from the receiver node. All the nodes are located in a straight line and the interval of the sender node is 2 m. Each sender node broadcasted 19-bytes test data packets with 18 dB·m sending power, and the receiver node records the RSSI upon receiving the packet. The test is conducted 100 times and the average value of RSSI of each node from different distance is calculated. The experimental results are shown in [Figure 1](#f1-sensors-15-03540){ref-type="fig"}. From [Figure 1](#f1-sensors-15-03540){ref-type="fig"}, the use of RSSI for distance estimation is not reasonable since the same RSSI values could correspond to different distances. However, for each individual curve the RSSI values mostly decreased monotonically with increasing distance. There are some abnormal and exceptional situations in the curve which can be attributed to many factors, such as the RF transmission parameters, antenna issues, random noise, propagation path loss and node-level sensing discrepancy. Therefore, because of these factors, the direct use of RSSI values to calculate the physical distance of nodes will result in inaccurate results. 2. Received Signal Strength Indicator (RSSI) ordering and node location permutation In a linear topology WSN, an ordered node location permutation can be formed by the nodes\' physical distance. For any sender node, say *Si*, the sort order of the node location permutation, say *S*, is obtained by the sender node *Si* and the receiver node communication distance from near and far. Another permutation, say *R*, can be obtained by ordering the RSSI value from sender node *Si* to the receiver node in descending order and we define this sorted permutation as RSSI ordering. Ideally, if the RSSI decreases monotonically with the increasing distance, *S* and *R* should be identical. But due to some unexpected factors existing in the process of signal propagation, two permutations *S* and *R* could not be consistent. As can be seen from [Figure 1](#f1-sensors-15-03540){ref-type="fig"}, the direct use of RSSI to determine the distance between the sender and receiver is not reliable. When sender nodes are far away from the receiver node, it is less accurate to infer their distance by RSSI values. Actually, this is reasonable when the process of signal propagation is considered. Signals spread during propagation and become weaker and weaker. The same noise level can influence weaker signals more easily. As a result, the RSSI becomes less reliable. It is obvious that at different distances from the receiver, the nearest node observes the best signal. In other words, the influence caused by a disturbance will be reduced as the distance between the sender and receiver decreases. This assertion has been verified in the following experiment. In the experiment, four sender nodes are arranged at different distances from the receiver node. All the nodes are in a straight line. The interval between each sender node is set to 1 m, 2 m, and 3 m, respectively. The distance between receiver and the first sender node (Sender 1) is correspondingly set to 1 m, 2 m, and 3 m. The layout of the network is shown in [Figure 2a](#f2-sensors-15-03540){ref-type="fig"}. Each sender node broadcasts 19-bytes test data packets with 18 dB m sending power, and the receiver node records the RSSI upon receiving the packet. The experiment is conducted 100 times at different intervals and the average value of RSSI of each node from different distance is calculated. The experimental result is shown in [Figure 2b](#f2-sensors-15-03540){ref-type="fig"}. Through the experiment results, the RSSI ordering can be checked as to whether it reveals the true distance relationship among the senders. In [Figure 2](#f2-sensors-15-03540){ref-type="fig"}, the vertical ordinate indicates the accuracy of rank in the RSSI ordering compared with the corresponding rank in node location permutation. The accuracy of the nearest sender node (rank 1 in the node list) is the highest and over 94% in the case of different intervals. This percentage decreases with the increase of rank. The accuracy of the 4th nearest node (rank 4 in the node list) is lower than 32%. Although the percentage decreases with the increase of the interval, the nearest node (rank 1) still obtains a better signal than other nodes. Therefore, the assertion that the nearest node observes the best signal among nodes in different distance from the receiver is verified in the experiment and can be employed to realize node sorting. 2.2.. Existing Methods as Motivation ------------------------------------ The experimental results show that the nearest node usually observes the largest RSSI. Some relative localization methods employing this principle have been proposed \[[@b12-sensors-15-03540],[@b13-sensors-15-03540]\]. The Kcd location algorithm \[[@b12-sensors-15-03540]\] obtains the positon of unknown nodes through the relative localization between the unknown node and the anchor node. The output node permutation can be obtained through the RSSI ordering of unknown nodes. Once the location of an unknown node is known, the node can be upgraded to an anchor node and help to locate the neighboring unknown nodes. This progress is repeated until the localization of as many unknown nodes as possible has been accomplished. However, the Kcd location algorithm has two limitations: (1) it is only suitable for the case where the deployment of the sensor network has a grid or linear topology and where the interval between any two neighboring nodes is regular rather than arbitrary; (2) the Kcd location algorithm upgrades the unknown nodes which have localized their position to an anchor node status, but once an unknown node has failed to localize its positon, the subsequent unknown nodes will all fail to localize their positions. In \[[@b13-sensors-15-03540]\], a spatial relative localization method without the anchors has been proposed. The method consists of two parts: the first part uses a specific method to find candidate orderings, while the second part verifies the consistency of each candidate ordering and outputs the final results. For each node, it broadcasts the localization command to obtain RSSI and ID from its neighbors. All the neighbors are sorted in descending order according to their observed RSSI. The sorted list is the node neighbor list. The first part of the method is to find the location permutation of nodes with the neighbor list. It iteratively derives the node ordering and starts with each node because there is no information about the initial node. During the iteration, the method deletes the completed one from the node\'s neighbor list, and prepares for the next iteration until the output of node permutation. If the process cannot proceed, then the method outputs a null. The second part of the method is to verify the consistency of all the reversion of candidate permutations coming from the first part. Compared with the Kcd location algorithm, the spatial ordering method in \[[@b13-sensors-15-03540]\] can judge whether the output is correct. However, it cannot handle the exception of output node permutation. In practice, three kinds of results will appear when using this method to locate nodes. The first one is all the candidate permutations are wrong, then the method outputs a null; the second result is multiple candidate permutations are simultaneously consistent with the conditions, then the method outputs a null; the third result is only one permutation is consistent with the conditions and the method outputs the right permutation. The spatial ordering method also has several limitations: (1) it does not know the initial node, so it has to iteratively start with every node and derive all the nodes; (2) in the three kinds of results, only the third one can output the right permutation and the others all output a null; (3) it can only judge the correctness of the output permutation, but it cannot handle the exception of the output permutation. 3.. Correct RSSI-Based Relative-Localization (CRBR-Localization) Algorithm ========================================================================== With respect to the limitations of the localization methods described in \[[@b10-sensors-15-03540]--[@b13-sensors-15-03540]\], certain improvements have been carried out and the CRBR-Localization algorithm based on the localization method in \[[@b12-sensors-15-03540],[@b13-sensors-15-03540]\] is achieved. Compared with the localization methods in the literature \[[@b12-sensors-15-03540],[@b13-sensors-15-03540]\], the CRBR-Localization algorithm described in the paper is more applicable in the relative localization of wireless nodes in the electric field measurement environment. In the localization method of the paper, the nodes that can communicate with node *i* are arranged in descending order according to the return RSSI values, thus obtaining the neighbor list of the node *i*. We select the first three nodes in the neighbor list during the localization, as shown in [Figure 3](#f3-sensors-15-03540){ref-type="fig"}. In [Figure 3](#f3-sensors-15-03540){ref-type="fig"}, the nodes are arranged in 1--2--3--4--5--6--7--8, which is the actual nodes location permutation. *N~i~* is the neighbor list consisting of three nodes with largest return RSSI values in descending order among nodes communicating with node *i*, while *N* is the set of neighbor list *N~i~* of all nodes. Before locating the nodes, it is necessary to mark the initial node *A* of the unknown nodes to be sorted. Taking [Figure 3](#f3-sensors-15-03540){ref-type="fig"} for example, *S* is the output location permutation after the algorithm, and the initial node *S* (1) is node 1.We select the first node (node 2) of the neighbor list of *S* (1) as *S* (2) and delete *S* (1) in the neighbor list of all nodes; select the first node (node 3) of the neighbor list of *S* (2) as *S* (3) and delete *S* (2) in the neighbor list of all nodes. The above process is repeated until the neighbor list of the last node of the ordering is empty, and the output nodes location permutation is *S*. Now we have the first solution and process summarized in Algorithm 1. **Algorithm 1** Deriving the permutation with initial node**Input**: all nodes\' neighborhood *N*; the initial node *A*.**Output**: node location permutation *S*.1.*S* (1) = *A*; *i* = *2;*2.**while** *N~s~* ~(~*~i~*~−1)~ is not empty **do**3. *S* (*i*) = the first element of *N~s~* ~(~*~i~*~−1)~4. delete *S* (*i* − 1) in the *N*5. *i* = *i* + 16.**end while**7.output the location permutation *S* 3.1.. Exception Handling ------------------------ From [Figure 2](#f2-sensors-15-03540){ref-type="fig"}, when the space between nodes is 1 m, over 94% of the RSSI orderings can identify the nearest node correctly, but there are still 6% of RSSI orderings failing to do so. Supposing that the neighbor node with largest RSSI value in the neighbor list of node *i* is not the nearest one to the receiver node during the location, then an inconsistent location permutation will be output. Since there is a small possibility that the above-mentioned anomalous condition may occur, we only analyze the situation where there is only one exceptional node, and the method is not applicable to the situation where several nodes are in an anomalous condition. When there is only one exceptional node among the unknown nodes to be located, two possible results arise: one is that the number of nodes in the output location permutation is less than the actual node number and another one is that the number of nodes in the output location permutation is equal to the actual node number: (1)The number of nodes in the output location permutation is less than the actual node number.When there are many nodes to be sorted, there are often two things occurring when the neighbor node with the largest RSSI value is not the one nearest to the receiver node, as shown in [Figures 4](#f4-sensors-15-03540){ref-type="fig"} and [5](#f5-sensors-15-03540){ref-type="fig"}. In these figures, the red nodes refer to faulty neighbor nodes while the blue ones are correct neighbor nodes, but the different location of the blue nodes in the neighbor list will cause different output location permutations. In [Figure 4](#f4-sensors-15-03540){ref-type="fig"}, the output location permutation according to Algorithm 1 is 1--2--4--5--6--7--8, where the node number is one less than the actual number, while in [Figure 5](#f5-sensors-15-03540){ref-type="fig"}, the output location permutation according to Algorithm 1 is 1--2--4--3, where the node number is four less than the actual number. In these two cases, the correct location permutation fails to be output due to a certain exceptional node, so a method is desired to correct the abovementioned conditions, thus obtaining the correct output location permutation: (a)For the case where the number of nodes in the output location permutation is different from the actual node number by one, the following method can be employed to handle it: choose the node failing to appear in the location permutation, and check whether its first two neighbor nodes adjoin in the ordering; if so, insert the missing node between its two neighbor nodes to get a new ordering, or the sorting fails. Verify the new ordering repeatedly to determine whether it is correct, or the sorting fails. As is shown in [Figure 4](#f4-sensors-15-03540){ref-type="fig"}, Node 3 is missing in the output location permutation, and then check whether the first two nodes in the neighbor list of Node 3 adjoins in the output ordering. The first two nodes in the neighbor list of Node 3 are Node 4 and Node 2, which adjoin in the ordering 1--2--4--5--6--7--8, and then insert Node 3 between Node 4 and Node 2 to get a new ordering 1--2--3--4--5--6--7--8. Select the last node (Node 8) of the new ordering as the initial node and make reverse deduction according to Algorithm 1. If the output ordering is 8--7--6--5--4--3--2--1, it is verified to be successful and output the correct location permutation 1--2--3--4--5--6--7--8. The method to handle such an exception is as shown in Algorithm 2: **Algorithm 2** Exception handling algorithm**Input**: all nodes\' neighborhood *N*; error location permutation *S*\'.**Output**: node location permutation *S*.1.Find the missing node *M*;2.**If** the first two elements in *N~M~* are neighborhood in the error ordering *S*\' **then**3.put the missing node *M* in the *S*\' between the two elements in *N~M~*;4.**else**5.**break**;6.*S* (1) = *S*\' (*n*), use Alogorithm1 to derive ordering *S*7.**If** *S* is the reversion of *S*\' **then**8.*S* = *S*\';9.**else**10.*S* = null;11.**end if** 7. output the location permutation *S*(b)For the case where the difference between the number of nodes in the output location permutation and the actual node number is more than one, the following method can be employed to handle it: select the neighbor list of the last but one node in the permutation, and delete the last node of the output location permutation in its neighbor list; if its neighbor list is empty, the output is an empty set and the sorting fails; if there still is a node in its neighbor list, select this node as the next one to continue sorting. When the difference between the node number of the location permutation and the actual node number is more than one, it means more than one node are exceptional, and the sorting fails. When the difference between the node number of the location permutation and the actual node number is one, we handle it according to algorithm 2. Taking [Figure 5](#f5-sensors-15-03540){ref-type="fig"} for example, the difference between the node number of location permutation and the actual node number is more than one and the location permutation is 1--2--4--3. Select the neighbor list *N*~4~ of the last but one node (Node 4) in the ordering and delete Node 3 from *N*~4~. Then the first neighbor node in *N*~4~ is Node 5 and make deduction according to algorithm 1 to get the order 1--2--4--5--6--7--8. Since the difference between the number of nodes in the location permutation and the actual node number is one, Algorithm 2 can be employed to handle. It can be seen that the location permutation is the same as the one in [Figure 4](#f4-sensors-15-03540){ref-type="fig"}, so Algorithm 2 can be employed to get the correct node location ordering 1--2--3--4--5--6--7--8. The detailed process is shown in Algorithm 3 to handle such an abnormity: **Algorithm 3** Exception handling algorithm**Input**: all nodes\' neighborhood N; error location permutation *S*\'.**Output**: node location permutation *S*.1.*k* = the number of *S*\'; *n* = the actual node number2.delete the last element of *S*\' in the *N~s~*~\'\ (~*~k~*~−1)~3.**If** *N~s~*~\'\ (~*~k~*~−1)~ is null **then**4. break;5.**else**6. *S*\' (*k*) = the first element of *N~s~*~\'\ (~*~k~*~−1)~; delete *S*\' (*k* − 1) in *N*7.**end if**8.**while** *N~s~*~\'\ (~*~k~*~)~ is not empty **do**9. *S*\'(*k* + 1) = the first element of *N~s~*~\'\ (~*~k~*~−1)~10. delete *S\'* (*k*) in *N*11. *k* = *k +* 112.**end while**13.*k*\' = the number of the new ordering *S*\'14.**If** *k*\' = = *n* **then**15. *S* (1) = *S*\' (*n*), use Algorithm 1 to derive the ordering *S*16.   **If** *S* is the reversion of *S*\' **then**17.     *S* = *S*\'18.   **else**19.     *S* = null20.   **end if**21.**else if** *n* − *k*\' = = 1 **then**22. turn to Algorithm *2*23.**else**24. *S* = null25.**end if**(2)The number of nodes in the output location permutation is equal to the actual node number. There are often the following three things occurring when the neighbor node with the largest RSSI value measured is not the one nearest to the receiving node, as is shown in [Figure 6](#f6-sensors-15-03540){ref-type="fig"}, [7](#f7-sensors-15-03540){ref-type="fig"} and [8](#f8-sensors-15-03540){ref-type="fig"}. In [Figure 6](#f6-sensors-15-03540){ref-type="fig"}, the actual order of nodes is 1--2--3--4--5--6--7--8, but the exceptional neighbor node of Node 5 causes the ordering output according to algorithm 1 to be 1--2--3--4--5--7--8--6. When in such a case, the reverse ordering of the output location permutation, *S* (*n*) *S* (*n* − 1) *...*, *S* (1), can be verified to determine whether the output location permutation is correct, that is to take *S* (*n*) as the initial node and make ordering deduction according to algorithm 1 to get the location permutation, which is called the reverse verification permutation S1. If the output reverse verification permutation is identical to the reverse ordering of permutation S, the sorting is correct, and if not, the sorting is faulty. In [Figure 6](#f6-sensors-15-03540){ref-type="fig"}, permutation *S* is 1--2--3--4--5--7--8--6, and the reverse verification ordering deducted with Node 6 as the initial node is 6--7--8--5--4--3--2--1, which is not consistent with the expected reverse verification permutation, 6--8--7--5--4--3--2--1, so the node sorting is faulty. In [Figure 7](#f7-sensors-15-03540){ref-type="fig"}, the actual order of nodes is 1--2--3--4--5--6--7--8, and although the neighbor node of Node 7 is exceptional, the ordering output according to Algorithm 1 is 1--2--3--4--5--6--7--8. However, when taking Node 8 as the start node, the reverse verification ordering output according to Algorithm 1 is 8--7--5--4--3--2--1, which is not consistent with the expected reverse verification ordering, 8--7--6--5--4--3--2--1. In [Figure 8](#f8-sensors-15-03540){ref-type="fig"}, the actual order of nodes is consistent with that in [Figure 7](#f7-sensors-15-03540){ref-type="fig"}, and its ordering permutation according to Algorithm 1 is still 1--2--3--4--5--6--7--8, but its reverse verification permutation according to Algorithm 1 is 8--7--5--6, which is not consistent with the expected reverse verification ordering, 8--7--6--5--4--3--2--1. The three cases mentioned above can be handled according to the following algorithm. When permutation *S* is deducted according to Algorithm 1,whether the output location ordering is correct can be determined by comparing whether the reverse verification permutation *S*1 is consistent with the reverse ordering of permutation *S*. If the node number of permutation *S*1 is identical to that of permutation *S*, and it is the reverse ordering of permutation *S*, the sorting is correct; if the difference between the node numbers of permutation *S*1 and permutation *S* is one, it can be corrected according to algorithm 2 using *S*1 as input; if the difference between the node numbers of permutation *S*1 and permutation *S* is more than one, it can be corrected according to Algorithm 3 using *S*1 as input; if it cannot be corrected, the sorting fails. When in the three above-mentioned cases, only [Figure 6](#f6-sensors-15-03540){ref-type="fig"} fails to output correct ordering while in the other two cases, relevant algorithms can be used to get correct node location ordering. The detailed process is shown in Algorithm 4. **Algorithm 4** Exception handling algorithm**Input**: all nodes\' neighborhood *N*; verified permutation *S***Output**: node location permutation *S*.1.use Alogorithm1 to derive the reverse verification permutation *S*12.*k* = the number of ordering *S*1; *n* = the actual node number3.**If** *k* = = *n* **then**4. **If** *S*1 is reversion of *S* **then**5. *S* = *S*6. **else**7. *S* = null8. **end if**9.**else if** *n* -- *k* = = 1 **then**10. turn to Algorithm 2 *using S*1 *as input*11.**else**12. turn to Algorithm 3 *using S*1 *as input*13.**end if** 3.2.. Realization of Correct RSSI-Based Relative-Localization *(*CRBR-Localization*)* Algorithm ----------------------------------------------------------------------------------------------- The CRBR-Localization algorithm combines the two methods in \[[@b12-sensors-15-03540],[@b13-sensors-15-03540]\], and applies a correction algorithm with respect to their respective shortcomings, which make it possible to handle the exception where one node fails. However, CRBR-Localization algorithm is only capable of correcting the cases in [Figures 4](#f4-sensors-15-03540){ref-type="fig"}, [5](#f5-sensors-15-03540){ref-type="fig"}, [6](#f6-sensors-15-03540){ref-type="fig"}, [7](#f7-sensors-15-03540){ref-type="fig"} and [8](#f8-sensors-15-03540){ref-type="fig"}, and for the case where several nodes fail, the algorithm output ordering is empty. The realization of the CRBR-Localization algorithm is shown in Algorithm 5: **Algorithm 5** CRBR-Localization Algorithm**Input**: all nodes\' neighborhood *N*; the initial node *A*; the actual node number *n*.**Output**: node location permutation *S*.1.use Algorithm 1 to derive the permutation *S*2.*k* = the number of ordering *S*3.**If** *k* = = *n* **then**4. turn to Algorithm 45.**else if** *n* − *k* = = *1*6. turn to Algorithm 27.**else**8. turn to Algorithm 39.**end if** 4.. System Implementation and Evaluation ======================================== To evaluate the CRBR-Localization algorithm, three indexes are chosen: (1) percentage of output, which is the ratio of the occurrences that the output permutation is a non-empty set after sorting to the experiment occurrences; (2) output accuracy, which is the ratio of the occurrences that the output permutation is a right sequence to the occurrences the output permutation is a nonempty set after sorting; (3) accuracy, the ratio of the occurrences that the output permutation is a right one after sorting to the experiment occurrences. The proposed three indexes can be defined as [Equations (1)](#FD1){ref-type="disp-formula"}--[(4)](#FD4){ref-type="disp-formula"}: $$R_{po} = \frac{N_{\textit{non}}}{N_{\textit{all}}}$$ $$R_{oa} = \frac{N_{\textit{acu}}}{N_{\textit{non}}}$$ $$R_{\textit{acu}} = \frac{N_{\textit{acu}}}{N_{\textit{all}}}$$ $$R_{\textit{acu}} = R_{po} \times R_{oa}$$where *R~po~* is percentage of output, *R~oa~* is output accuracy, *R~acu~* is accuracy, *N~non~* is the occurrences that the output permutation is a nonempty set after sorting, *R~acu~* is the occurrences that the output permutation is a right sequence after sorting, *R~all~* is the total experiment occurrences. From the definitions of the three indexes, we can know that accuracy (*R~acu~*) equals to the arithmetic product of the percentage of output (*R~po~*) and the output accuracy (*R~oa~*). The ideal algorithm has not only a high percentage of output but also high output accuracy. Meanwhile, the index of accuracy could be a comprehensive evaluation of the algorithm\'s percentage of output and output accuracy. 4.1.. Experimental System ------------------------- To validate and verify the proposed localization algorithm, a double circuit experimental line on the same tower was erected at the HVDC test base of China Electric Power Research Institute in Beijing. [Figure 9](#f9-sensors-15-03540){ref-type="fig"} shows our hardware system for the localization. There are two kinds of nodes which are wireless nodes (unknown nodes) and a sink node in the system. The sink node is connected to the computer through the USB bus. The wireless nodes which are arranged in a line perpendicular to the transmission lines are the subject of the localization. The sink node broadcasts the localization command message to the wireless nodes. Upon receiving the message, the wireless nodes will communicate with their neighbor nodes and obtain the neighbor list. In the same way, all the wireless nodes in turn get their neighbor lists and then transmit them to the sink node. As a result, the computer connected with the sink node executes the CRBR-Localization algorithm. The detailed workflow of the CRBR-Localization algorithm is shown in [Figure 10](#f10-sensors-15-03540){ref-type="fig"}. The sink node has two main functions. The first one is to broadcast the localization commands and another one is to receive all the unknown nodes\' neighbor lists used for the CRBR Localization algorithm. There is a high-performance CPU and enough memory in each unknown node, so the neighbor list of each unknown node is created by the unknown node itself and then sent to the sink node. Therefore in the experimental system, only one node which is connected with the computer through the USB interface is the sink node. The sink node could be outside or at one end of the network of unknown nodes. It is convenient for the operator to always take as the sink node the one with the connected computer. A block diagram of the proposed wireless node is presented in [Figure 11](#f11-sensors-15-03540){ref-type="fig"}. The wireless nodes\' primary function is to acquire and process the output analog signals from the electric field sensors and send the digital data to the sink node. The core of the wireless node is an STM32F embedded microcontroller. The wireless communication system selected for the sensor node is the 2.4 GHz Xbee Pro Radio Frequency (RF) module based on the IEEE 802.15.4 protocol from Digi^TM^ (Minnetonka, MN, USA) 4.2.. Node Number Experiments ----------------------------- In the electric field measurement environment, since the transmission line spacing is different, different numbers of sensor nodes are needed. In order to verify the validity of the algorithm, we did sorting experiments using different node numbers. In these experiments, the numbers of unknown nodes to be positioned are 5, 10, 15, and 20, respectively. The nodes are arranged in a line perpendicular to the transmission lines, and the interval between nodes is 2.5 m. The RSSI values that are returned by the unknown nodes in each sequence are sampled five times and averaged in order to use the average for sorting. Three hundred sequences are obtained with each sorting, and the sorting accuracy is counted and analyzed. The result of the experiment is shown in [Figure 12](#f12-sensors-15-03540){ref-type="fig"}. Since the output accuracies of the experiments are all 100%, the accuracy equals the percentage of output, thus only accuracy is used as the evaluation criterion. From [Figure 12](#f12-sensors-15-03540){ref-type="fig"} we can see that when sorting five nodes, the sorting accuracies of nodes before and after exception handling are both 98.4%, thus when the number of nodes is small, exception handling has an insignificant influence on the result. Node sorting accuracy before exception handling falls dramatically with the increase of the number of sorting nodes. When the number of sorting nodes is 10, the accuracy is 94.4%, and when the number of sorting nodes is 15, the accuracy is 91.25%. When the number of sorting nodes is 20, the accuracy is only 86.6%. Node sorting accuracy after exception handling falls slightly with the increase of the number of sorting nodes. When the number of sorting nodes is 10 the accuracy is 98.1%, and when the number is 15, the accuracy is 97.5%. When the number of sorting nodes is 20, the accuracy is still above 96.3%, so when the number of nodes in the experiment is large, the sorting accuracy can be improved markedly through exception handling by the CRBR-Localization algorithm. 4.3.. Experiment in Four Deployments ------------------------------------ Theoretically, when placing sensor nodes, every node should be arranged in a line perpendicular to the transmission lines. However, since there exist personal deployment errors, not all the nodes are in a straight line, so the adaptability of the CRBR-Localization algorithm to different node deployment structures needs to be verified. In the experiment, we sequence 10 unknown wireless nodes in four different deployment structures. The RSSI values that are returned by the unknown nodes in each sorting are sampled five times and averaged for further sorting. Three hundred sequences are obtained with each structure. The results are obtained to compare with the Spatial Ordering method presented in \[[@b13-sensors-15-03540]\] and the Kcd location algorithm presented in \[[@b12-sensors-15-03540]\]. ### 4.3.1.. Linear Homogeneous Deployment In this experiment, the interval between neighboring nodes is 3 m. The nodes are arranged in a line perpendicular to the transmission line, as shown in [Figure 13](#f13-sensors-15-03540){ref-type="fig"}. The distance between the first node and the last node is 27 m, and they can communicate normally. The result is shown in [Figure 14](#f14-sensors-15-03540){ref-type="fig"}. From [Figure 14](#f14-sensors-15-03540){ref-type="fig"} we can see that the percentage of output and the output accuracy of the CRBR-Localization algorithm are both 100%, higher than the values of the other two algorithms. Since the Spatial Ordering method doesn\'t process exceptional conditions and may have several candidate orderings, its percentage of output is lower than that of the CRBR-Localization algorithm and Kcd location algorithm. Because the Kcd location algorithm can\'t judge the correctness of its output sequences, although the percentage of output is 100%, the output accuracy is still lower than those of the CRBR-Localization algorithm and Spatial Ordering method. Comprehensively comparing the three algorithms, we can see that the three indexes of the CRBR-Localization algorithm are all higher than those of the other two algorithms, and it has higher localization accuracy. ### 4.3.2.. Linear Inhomogeneous Deployment In this experiment, the intervals between neighboring nodes are not the same. The maximum interval is 4 m, while the minimum is 1.5 m. The nodes are arranged in a line perpendicular to the transmission lines, as shown in [Figure 15](#f15-sensors-15-03540){ref-type="fig"}. The first node and the last node can communicate normally, as shown in [Figure 16](#f16-sensors-15-03540){ref-type="fig"}. From [Figure 16](#f16-sensors-15-03540){ref-type="fig"}, Kcd location algorithm has the highest percentage of output, and the CRBR-Localization algorithm is in the middle, while the Spatial Ordering method has the lowest, which is less than 90%. Although the Kcd location algorithm has the highest percentage of output, its accuracy and output accuracy are both lower than 90%, apparently lower than those of the CRBR-Localization algorithm. The percentage of output of the CRBR-Localization algorithm is higher than 99%, and its accuracy is higher than 96%, both of which are higher than Kcd location algorithm and Spatial Ordering method. Under linear inhomogeneous deployment conditions, the percentage of output of CRBR-Localization algorithm and Spatial Ordering method are both lower than under a linear homogeneous alignment. This is because when the nodes are placed homogeneously, there may be a situation that some nodes are placed closely and others are placed loosely. When the nodes are close, the distance between nodes is small, so the RSSI values measured are close. Since the RSSI values of each node are close, elements like electromagnetic components, topographic relief and surface vegetation will interfere with the radio signal, which will have a greater influence on obtaining the RSSI accurately, whereas the CRBR-Localization algorithm and the Spatial Ordering method can judge the correctness of output location permutation, and when the output permutation is judged as a wrong one, then output will not be allowed. In addition to that, the CRBR-Localization algorithm can not only judge the correctness of output permutations, but also further correct the wrong permutations, so it has a higher percentage of output and accuracy than the Spatial Ordering method. ### 4.3.3.. Cross Deployment In this experiment, each node is first arranged in a line with a 3 m interval. Then we make each node move 0.5 m to the left and right, respectively, as is shown in [Figure 17](#f17-sensors-15-03540){ref-type="fig"}. From [Figure 18](#f18-sensors-15-03540){ref-type="fig"}, the CRBR-Localization algorithm has an output accuracy of 100%, and has a higher accuracy than the Kcd location algorithm and Spatial Ordering method. Though the Kcd location algorithm has a high percentage of output, it can\'t judge the correctness of the output location permutations, so its percentage of output and accuracy are both lower than those of the CRBR-Localization algorithm. Comprehensively comparing the three algorithms, we can see that in cross deployment structure scenarios, the overall performance of the CRBR-Localization algorithm is better than that of the other two algorithms. ### 4.3.4.. Random Deployment In actual measurement, since there exist personal deployment errors, so not all the nodes are arranged according to the ideal structure (linear and homogeneous). Therefore, an approximate random arrangement is used to simulate the situations in an actual measurement environment. Under this deployment, the nodes are arranged randomly in a rectangle 30 m long and 1 m wide, and the minimum interval of adjacent nodes is 1.8 m, while the maximum is 3.6 m, as shown in [Figure 19](#f19-sensors-15-03540){ref-type="fig"}. Under random deployment, the intervals between nodes are not the same, and not all the nodes are in the same line, but they are randomly arranged in a rectangle area. Under this circumstance, the RSSI is sensitive to the interference factors in the environment. From [Figure 20](#f20-sensors-15-03540){ref-type="fig"}, we can see that the CRBR-Localization algorithm has accuracy and percentage of output higher than 97%, and its output accuracy is 100%. Both its output accuracy and its percentage of output are higher than those of the Kcd location algorithm and Spatial Ordering method. Though the Spatial Ordering method has a higher percentage of output, its accuracy and output accuracy are both lower than 94%, which is lower than the CRBR-Localization algorithm. Comprehensively comparing the three algorithms, we can see that under random arrangement conditions, the overall performance of the CRBR-Localization algorithm is better than that of the other two algorithms. In the four different localization experiments, the accuracy and percentage of output of the CRBR-Localization algorithm are apparently higher than those of the Kcd Location algorithm and the Spatial Ordering method. Four different deployment structures, the accuracies of the CRBR-Localization algorithm are all above 96%. Taking the three performance indexes into consideration, we can see that compared with the other two algorithms, the CRBR-Localization algorithm has a better overall performance, and has a better adaptability of different applied environments of electric field measurement. 5.. Conclusions =============== In this paper, a RSSI-based algorithm is proposed for achieving the relative localization of one-dimensional linear topology wireless networks under HVDC transmission lines. Starting from the observation that the closest node always has the highest RSSI, the algorithm utilizes unknown nodes\' neighbor list based on the RSSI values to determine the relative locations of nodes. A detailed interpretation has been made around the solutions of the exceptional situations for the output permutation which can effectively improve the localization accuracy. The performance of this algorithm in real circumstances has been evaluated through several experiments with different numbers of nodes and different deployments at the China State Grid HVDC test base. Results show that in the experiments with different number of nodes, the accuracy of the CRBR-Localization algorithm is above 96.2% and in the experiments with different deployments the accuracy of the CRBR-Localization algorithm, which is over 96%, is higher than that of the other two algorithms. The authors would like to thank all the reviewers for their helpful comments. The research was supported by National Natural Science Foundation of China (61273165, 51207005) and Basic Research Foundation of Beihang University (YWF-14-ZDHXY-13). All authors have made significantly contributions to the paper. Cui Yong carried out the literature survey and wrote the paper section about the relative localization methods for the measurement of the electric field under HVDC transmission lines. Yuan Haiwen proposed the concept of relative localization in the WSN. The experimental system was developed and operated by Hu Xuemin under the supervision of Wang Qiusheng. Song Xiao also contributed to the data collection and modification of the manuscript. Zhao Luxing provided experimental conditions and gave valuable suggestions for improvement in the China State Grid HVDC test base. The authors declare no conflict of interest. {#f1-sensors-15-03540} ###### The correspondence between the Received Signal Strength Indicator (RSSI) ordering and node location permutation; (**a**) Layout of the network; (**b**) Experimental results.   {#f3-sensors-15-03540} {#f4-sensors-15-03540} {#f5-sensors-15-03540} {#f6-sensors-15-03540} {#f7-sensors-15-03540} {#f8-sensors-15-03540} {#f9-sensors-15-03540} {#f10-sensors-15-03540} {#f11-sensors-15-03540} {#f12-sensors-15-03540} {#f13-sensors-15-03540} {#f14-sensors-15-03540} {#f15-sensors-15-03540} {#f16-sensors-15-03540} {#f17-sensors-15-03540} {#f18-sensors-15-03540} {#f19-sensors-15-03540} {#f20-sensors-15-03540} [^1]: Academic Editors: Luciano Lavagno and Mihai T. Lazarescu

The electronic version of this article is the complete one and can be found at: <http://f1000.com/reports/b/4/9> Introduction {#s01} ============ The adaptive immune response of vertebrates occurs through clonal selection, an elegant process that involves large numbers of lymphocytes. Each lymphocyte bears a randomly generated receptor that, if bound by its ligand in an appropriate context (indicative of potential infection), stimulates the cell to proliferate, thereby generating many more cells with this receptor. Because the specific cells are pre-selected to remove those that might recognize "self" components, the ligands that might engage the receptors are "foreign" and, thus, may identify an invading organism and orchestrate a response. The responding cells proliferate and functionally differentiate, and ultimately clear the invader. The numbers of responding cells then decline, ultimately leaving memory cells capable of a recall response should the foreign ligands reappear. The evolution of the adaptive immune response to many pathogens is sometimes viewed as a sort of "arms race," since the invading organism may have the potential to replicate much more rapidly than the lymphocytes that would control it. It is perhaps not surprising, then, that activated lymphocytes (B and T cells) have one of the shortest cell cycles seen in mature vertebrate animals, replicating every four to six hours \[[@bib-001]\]. However, this proliferative burst is preceded by a lag phase of about 24 hours, during which the cell grows (increases mass) prior to the first cell cycle entry. As we will see, this lag may be essential for the proliferative burst and for subsequent function of the clones, and the events surrounding these changes represent an emerging area of intense interest. Preparing T cell metabolism for proliferation {#s02} ============================================= In shifting from a small, resting cell to a rapidly cycling cell, a naïve lymphocyte effectively "reprograms" its metabolism, changing from a cell that relies on fatty acid oxidation (and some glycolysis), to one that engages robust aerobic glycolysis and glutaminolysis \[[@bib-002]\] (see [Figure 1](#fig-001){ref-type="fig"}; here we focus on T lymphocytes, but similar events may well occur in activated B lymphocytes). This change relies on signaling events early in the activation process. For example, a hierarchical signaling cascade downstream of the cell surface receptors involved in MAPK (mitogen activated protein kinase)/ERK (extra-cellular signal-regulated kinase), PI3K (phosphoinositide 3-kinase)/Akt, mTOR (mammalian target of rapamycin) kinase and NFκB (nuclear factor-κB) pathways rapidly engages the expression of the transcription factor Myc, which, in turn, induces the expression of transporters for glucose and glutamine, and many of the enzymes involved in glycolysis, the pentose phosphate pathway, and glutaminolysis \[[@bib-002],[@bib-003]\]. This reprogramming, thereby, directs nutrients to the production of nucleotides, lipids, amino acids, and other biosynthetic products needed for proliferation. Meanwhile, the activation of AKT and ERK facilitates post-translational surface expression of glucose transporter Glut-1 and glutamine transporter SNAT2 (system A neutral amino acid transporter 2), respectively \[[@bib-004],[@bib-005]\]. However, AKT is dispensable for increased glycolysis and proliferation in cytotoxic CD8 T cells; although these effects depend on PDK1 (phosphoinositide-dependent protein kinase-1), a well-known upstream regulator of AKT \[[@bib-006]\]. TORC (TOR complex) 1, a downstream signaling component of Akt, also plays a role in maintaining the metabolic gene transcriptome in naïve quiescent T cells \[[@bib-007]\]. Meanwhile, the suppression of the nuclear receptor LXR (liver X receptor) enhances cholesterol synthesis, an essential component of membranes \[[@bib-008]\], and the activation of an orphan steroid receptor, ERRα (estrogen-related receptor α), promotes lipid production, and perhaps mitochondrial biogenesis \[[@bib-009]\]. Coordinately, these events most likely prepare the activated T cells for their entry into their rapid cell cycle. {#fig-001} T cell metabolism and differentiation {#s03} ===================================== As T lymphocytes begin to proliferate they also undergo differentiation into functional subsets in response to extracellular signals, and these subsets determine the nature of the immune response. CD4 T cells differentiate into Th (T helper) 1 cells that mediate responses to intracellular pathogens, Th2 cells that control responses to extracellular bacteria and helminthes, Th17 cells that are important in anti-fungal defense and inflammation, and induced Treg (regulatory T) cells that dampen immune responses. Of these, only the Treg cells rely on lipid oxidation as an energy source, and forcing proliferating T cells to utilize free fatty acids for energy tends to drive enhanced Treg differentiation \[[@bib-010]\]. In contrast, increased glycolysis is seen in differentiating Th17 cells, and this is a function of the transcription factor HIF1 (hypoxia-inducible factor 1) \[[@bib-010],[@bib-011]\]. Thus, damping of glycolysis with low dose 2-deoxyglucose inhibits Th17 and promotes Treg generation \[[@bib-011]\]. HIF1 also directly enhances RORγt activity and represses Foxp3 activity, producing a reciprocal increase in Th17 and decrease in iTreg differentiation \[[@bib-012]\]. Meanwhile, the choice between Th1, Th2 and Th17 differentiation is mediated in part by TORC1/2, the activation of which requires a surplus of intracellular amino acids pool; TORC1 promotes Th1 and Th17 differentiation, while TORC2 promotes Th2 differentiation \[[@bib-013]\]. In addition, the depletion of extracellular amino acids, either by amino acid catabolic enzymes such as IDO (indoleamine 2,3-dioxygenase), Arg I (arginase I) and asparaginase or by the small molecule halofuginone, results in the activation of the protein kinase GCN2 (general control nonrepressed 2) in T cells \[[@bib-014]-[@bib-017]\]. Consequently, Th17 differentiation is suppressed, whereas Treg development and T cell anergy are enhanced \[[@bib-014],[@bib-018]\]. Each of these regulatory "nodes" (HIF1, TORC, GCN2) are responsive to metabolic status (e.g. oxygen availability, intracellular ATP and amino acid pool), highlighting the connections between metabolism and T cell signaling. Following the peak of proliferation and differentiation, there is a contraction phase as cells undergo apoptosis dependent on the pro-apoptotic proteins BIM and PUMA \[[@bib-019]\]. Cells destined to be memory T cells, responsible for enhanced immunity upon rechallenge, persist by employing fatty acid oxidation, the activity of which is associated with increased mitochondrial respiratory capacity and is negatively regulated by TORC1 activity \[[@bib-020]-[@bib-022]\]. The inhibition of TORC1 may require the function of the TSC (tuberous sclerosis protein) 1/2 complex, as cells lacking TSC1 do not generate immune memory \[[@bib-007]\]. Summary and future directions {#s04} ============================= We suspect that these studies merely "scratch the surface" of metabolic control of lymphocyte function. And of course, the immune response is not restricted to lymphocytes --- the metabolic functions in cells of the innate immune system, including macrophages, dendritic cells, and others, are topics of ongoing investigation as well. Further, while much current work is directed at understanding how signaling pathways regulate metabolic changes, there are intriguing observations in these and other systems that support the converse; that is, that changes in metabolic substrates and products directly affect signaling. For example, the intracellular ATP and amino acid pools directly impact the activities of AMPK (AMP-activated protein kinase), TORC1 and GCN2 \[[@bib-014], [@bib-023], [@bib-024]\]. Further, enzymes in the metabolic pathways can also function in signaling. For example, glyceraldehyde 3-phosphate dehydrogenase (GAPDH) has been implicated in the control of gene expression \[[@bib-025]\], and the glycolytic enzyme pyruvate kinase M2 isoform (PKM2) (expressed in cancers \[[@bib-026]\] and activated lymphocytes \[[@bib-002]\]) engages β-catenin signaling independently of Wnt \[[@bib-027]\]. We may intuit that these, and other critical enzymes, are likely to link metabolic status to signaling. A great deal of excitement has followed the metabolic control of cancer in pointing the way to new therapies. The same may well apply to metabolic control of dysregulated autoimmune disease. These studies hold the promise for a renaissance in nutritional research, far beyond the "immune diet" fads, based on fundamental principles of the emerging connections between signaling pathways and metabolism. We thank the members of the Green laboratory for valuable discussions. This work was supported by the George J. Mitchell fellowship from St Jude Children Hospital (R.W.), NIH grants AI40646 and GM52735 (D.R.G.) and the American Lebanese and Syrian Associated Charities. Competing Interests =================== The authors declare that they have no competing financial interests. ERK : extra-cellular signal-regulated kinase GCN2 : general control nonrepressed 2 HIF1 : hypoxia-inducible factor 1 Th : T helper cell TORC : TOR complex Treg : regulatory T TSC : tuberous sclerosis protein

Introduction ============ Hypertension is one of the severe public health problems worldwide and one of the leading causes of cardiovascular disease and death.[@b1-ARYA-12-109],[@b2-ARYA-12-109] Annually, about 7.5 million deaths (13% of all deaths) are caused by hypertension.[@b3-ARYA-12-109],[@b4-ARYA-12-109] Report given by the global burden of hypertension, expressed that, almost one billion adults suffer of hypertension in 2000, and it is expected to enlarge to 1.56 billion until 2025.[@b5-ARYA-12-109] In spite of wide efforts to handle hypertension, just half of treated individuals are controlled and the rest is cases with resistant hypertension.[@b6-ARYA-12-109],[@b7-ARYA-12-109] Resistant hypertension is a common clinical problem of blood pressure that is not controlled despite the simultaneous application of multiple antihypertensive agents.[@b8-ARYA-12-109],[@b9-ARYA-12-109] Physician inertia, medication side effects, non-conformity to lifetime pharmacological remedy by patients, and drug incompetence have been expressed as reasons of failure in the pharmacological strategy.[@b10-ARYA-12-109],[@b11-ARYA-12-109] Several studies have indicated that kidney has an important impact on blood pressure regulation; it has been proved that augmented activity in the sympathetic nervous system contributes to the hypertension pathogenesis.[@b12-ARYA-12-109],[@b13-ARYA-12-109] Hence, ablation of renal afferent nerves has been applied and proved to decrease hypertension and injuries produced by severe sympathetic hyperactivity.[@b14-ARYA-12-109],[@b15-ARYA-12-109] Traditionally, surgical sympathectomy had been successfully applied in decreasing blood pressure in individuals with chronic hypertension.[@b16-ARYA-12-109] Given that, this method was founded to have abundant large scale perioperative mortality and long-lasting complications, was forsaken until the advent of catheter-based method.[@b17-ARYA-12-109] Catheter-based renal sympathetic denervation has been expressed as a safe, helpful, and cost-effective intervention in patients with resistant hypertension.[@b18-ARYA-12-109],[@b19-ARYA-12-109] Reduction in blood pressure and diminishing the possibility of stroke, left ventricular hypertrophy, chronic renal, and heart failure by catheter-based renal sympathetic denervation in combination with pharmacologic remedy have been reported.[@b20-ARYA-12-109] However, the long-term effect of renal sympathetic denervation and its complications in patients with treatment-resistant hypertension is not clearly understood. Toward this end, the present survey tries to investigate the long-term effect of renal artery sympathetic ablation and its complications in patients with treatment-resistant hypertension. This study designed to assess the feasibility, safety, and effectiveness of renal sympathetic denervation in patients with resistant arterial hypertension. Materials and Methods ===================== This single-center, prospective study was approved by the Ethical Review Committees of Isfahan University of Medical Sciences, Isfahan, Iran. After a full explanation of the study, written informed consent was obtained from all the patients. In total 30 patients with resistant arterial hypertension despite treatment with ≥ 3 antihypertensive drugs (at least one of the antihypertensive medications was required to be a diuretic) were enrolled in this study. Inclusion criteria were age more than 15 years in both sex, with systolic blood pressure (SBP) ≥ 160 mm Hg or diastolic blood pressure (DBP) ≥ 90 mm Hg. Exclusion criteria were pregnancy; have any known secondary cause of hypertension; severe renal artery stenosis, previous renal stenting or angioplasty, or known dual renal arteries type 1 diabetes, hemodynamically major renal artery stenosis, previous renal artery intervention, renal artery anatomy that precluded treatment (\< 4 mm diameter, or \< 20 mm length), an estimated glomerular filtration rate of \< 45 ml/minutes, heart disease, planned pregnancy during the study, and a history of myocardial infarction (MI) and unstable angina in the previous 6 months. Resistant hypertension is defined as blood pressure that remains above goal despite concurrent use of three antihypertensive agents of different classes, one of which should be a diuretic.[@b1-ARYA-12-109] Patients whose blood pressure is controlled with four or more medications are considered to have resistant hypertension. Selected patients were treated with the renal denervation procedure, the femoral artery was accessed with the standard endovascular technique, and the Symplicity catheter was advanced into the renal artery and connected to a radiofrequency generator. Blood pressure was measured twice in sitting position after 5 minutes resting by mercury sphygmomanometer. The mean of the two recordings was reported as patient's blood pressure. Subjects who had three or more of the criteria defined by National Cholesterol Education Program (NCEP) were diagnosed with metabolic syndrome. The criteria of NCEP include: (1) Central obesity as the waist circumference \> 102 cm in men and \> 88 cm in women; (2) fasting plasma triglycerides ≥ 150 mg/dl; (3) low high-density lipoprotein cholesterol (HDL-C) \< 40 mg/dl in men and \< 50 mg/dl in women; (4) SBP ≥ 130 mm Hg and/or DBP 85 mm Hg and/or antihypertensive agents (5) hyperglycemia with fasting plasma glucose (FPG) ≥ 100 mg/dl and/or hypoglycemic medications.[@b10-ARYA-12-109] Before and 12 months after renal denervation procedure waist, body mass index (BMI), SBP, DBP, metabolic syndrome, fasting blood sugar (FBS), HDL, and triglyceride were measured in all patients. Change in the mean of measurements of SBP and DBP from baseline to 12 months after renal denervation were the main effectiveness endpoint of this study, also, chronic procedural safety such as death, MI, stroke, congestive heart failure (CHF), and renal arterial stenosis were assessed in all patients after 12 months. To measure renal arterial stenosis 12 months after renal denervation, angiography performed in any of patients who were willing and ready. With a sample of 30 patients, we calculated that the study would have at least 80% power to show benefit of renal denervation, assuming at least a 10 mm Hg difference with a 21 mm Hg standard deviation (SD) of the change in SBP from baseline to 12 months. The pattern of medication use was defined according to the type of drugs used to control and/or treat hypertension. Medicines were classified according to pharmacological category. The collected data were analyzed statistically with SPSS software (version 20, SPSS Inc., Chicago, IL, USA). Kolmogorov-Smirnov normality test was used before analysis, and in case of abnormality in data transformation was used. Continuous variables present as mean ± SD and categorical variables as number (%). Studied variables after renal denervation in compare to baseline were assessed using pair t-test. The frequency of metabolic syndrome was tested by McNamara's test. P \< 0.050 was considered significant. Results ======= During the study period (March 2012 to November 2013), a total of 37 patients were consented for enrollment. During the screening process, seven patients were excluded (three patients did not met inclusion criteria and four refused informed consent). In total, 30 patients completed baseline evaluation and underwent the renal denervation procedure. Within the 12 months follow-up period one patient died, and finally, 29 patients completed the study and analyzed. The mean age of studied patients was 56.3 ± 10.8 years old, 14 of patients (47%) were male and 16 patients (53%) were female. Clinical characteristics of studied patients at baseline in compare to 12 months follow-up after renal denervation are shown in [table 1](#t1-ARYA-12-109){ref-type="table"}. BMI and waist after renal denervation significantly decreased in compare to baseline (P \< 0.050). SBP at baseline was 169.8 mm Hg and after renal denervation meaningfully reduced to 147.5 mm Hg (P \< 0.001). DBP was significantly decreased of 95.7 mm Hg at baseline to 83.8 mm Hg after renal denervation (P \< 0.001). The distribution of metabolic syndrome was not significantly different after renal denervation in compare to baseline (P = 0.170). Moreover, a significant decreased in FBS and triglyceride after renal denervation was observed in compare to baseline (P = 0.001). HDL significantly increased in compare to baseline (P = 0.002). Distribution of SBP levels at baseline in compare to 12 months post-renal denervation. At baseline, all studied patients had SBP higher than 140 mm Hg and only 7% had SBP between 140 and 160 mm Hg and 93% had SBP higher than 160 mm Hg. 12 months after renal denervation 3.3% of subjects had SBP lower than 140 mm Hg and 20.0% had SBP between 140 and 160 mm Hg. Decline in SBP in 7% of patients was lower than 10 mm Hg, which were defined as no decrease in SBP. 93% of patients had the SBP reduction of ≥10 mm Hg. Moreover, also 12 months after renal denervation 23.3% of patients achieved the SBP of \< 140 mm Hg. [Table 2](#t2-ARYA-12-109){ref-type="table"} shows there are no significant differences in the pattern of ntihypertensive medication at baseline and 12 months follow-up after renal denervation. The mean of decreases in both SBP and DBP was 22.3 and 11.83 mm Hg, respectively. The mean of decreases in BMI, waist, FBS and triglyceride after follow-up was 0.9, 0.34, 4.5 and 19.4, respectively; and the mean of increased in the level of HDL was 3.27. Post-procedure within the follow-up period we detected the following side effects: MI was occurred in two patients. One of patient's died from a MI. One of the patients had stroke and one patient had CHF. Angiography was done in eight patients who were willing and renal arterial stenosis (50%) was observed in two of them. Discussion ========== In present single-center, prospective study showed renal sympathetic denervation as a modern and secure catheter-based method for sustained reduction hypertension in treatment-resistant cases is useful method so renal denervation provides harmless and continued drop of blood pressure to 2 years. Similar to our results Esler et al. demonstrated control hypertensive subjects who crossed over to renal denervation with the Symplicity system demonstrated a meaningful decline in blood pressure alike to that observed in subjects obtaining urgent denervation.[@b21-ARYA-12-109] As with the former available reports concerning this method, our findings, reaffirm the effectiveness of catheter-based therapy for hypertension resistant subjects and clarified that renal sympathetic denervation creates a secure and sustained blood pressure reduction during 12 months after treatment follow-up. A highly significant reduction of SBP and DBP were observed. SBP at baseline was 169.8 mm Hg and after renal denervation significantly decreased to 147.5 mm Hg (P \< 0.001). Furthermore, DBP was significantly decreased from 95.7 mm Hg at baseline to 83.8 mm Hg after renal denervation (P \< 0.001). Our findings are keeping with Katholi ýand Rocha-Singh,[@b19-ARYA-12-109] they have expressed that catheter-based renal denervation offers sustained and considerable reduction of blood pressure in resistant hypertension patients. They also reported that catheter-based renal denervation causes no severe unpleasant complication, and mentioned the necessities of prospective randomized clinical trials for proving their findings.[@b19-ARYA-12-109] Hypertension reduction using catheter-based renal sympathetic denervation along with pharmaceutical remedy have been shown to be effective in decreasing the stroke risk, heart failure, left ventricular hypertrophy, and severe renal failure.[@b21-ARYA-12-109] Moreover, it has been suggested that after 6 months, analogs significant reduction in blood pressure like as individuals receiving immediate renal sympathetic denervation were observed and the secure and sustained reduction of blood pressure exist in 1 year follow-up. Our results confirm this safety and sustainability of renal sympathetic denervation method in 21 months follow-up period.[@b21-ARYA-12-109] It's worth noting that we measured the amount of triglyceride, HDL and FBS after renal sympathetic denervation according to the results triglyceride, HDL, and FBS meaningfully decreased from baseline to 21 months after renal sympathetic denervation. We have to acknowledge the investigation limitation due to the fact that inasmuch as, there are no other therapeutic strategies to be compared with renal sympathetic denervation, we are not capable to judge against the renal sympathetic denervation effectiveness with other remedy options. Another study limitation possibly will be the relatively little sample size. Conclusion ========== Renal sympathetic denervation presents a modern and secure catheter-based method for sustained reduction hypertension in treatment-resistant cases. We demonstrate for the first time to our knowledge that triglyceride, HDL, and FBS are also controlled by renal sympathetic denervation. I owe the accomplishment of this work to the only living almighty, merciful, and compassionate God, in whom is hidden all the treasures of knowledge and wisdom. This study could not have been completed without the support, assistance, and generous cooperation of several people to whom I wish to express my greatest appreciation. My greatest thanks and deepest respect goes to my teachers Dr. Massoud Pourmoghadas and Dr. Alireza Khosravi for all their love and support to me, my husband, and especially our children. **Conflicts of Interest** Authors have no conflict of interests. ###### Clinical characteristics of studied patients at baseline and 12 months follow-up after renal denervation Variable Baseline 12 months after renal denervation P between two phase -------------------------- -------------- ----------------------------------- ---------- ------------------- BMI (kg/m2) 28.3 ± 2.8 27.9 ± 2.5 0.008 -0.90 ± 1.20 Waist circumference (cm) 92.3 ± 10.4 91.1 ± 10.2 0.003 -0.34 ± 0.02 SBP (mm Hg) 169.8 ± 10.5 147.5 ± 14.9 \< 0.001 -22.30 ± 10.04 DBP (mm Hg) 95.7 ± 9.7 83.8 ± 7.5 \< 0.001 -11.83 ± 5.63 FBS (mg/dl) 111.7 ± 15.7 107.2 ± 12.9 0.001 -4.50 ± 2.31 HDL (mg/dl) 45.1 ± 5.5 48.3 ± 5.4 0.002 3.27 ± 1.03 Triglyceride (mg/dl) 165.9 ± 53.8 146.5 ± 38.1 0.001 -19.43 ± 10.02 P values calculated by pair t-test. SD: Standard deviation; BMI: Body mass index; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; FBS: Fasting blood sugar; HDL: High-density lipoprotein ###### Prevalence of antihypertensive medication at baseline and 12 months follow-up after renal denervation Drug Value (%) Value (%) ------------------------------ ----------- ----------- ACE inhibitor 51.0 49.9 Angiotensin receptor blocker 67.3 67.5 Calcium channel blocker 27.6 28.1 Diuretic 58.8 59.2 Aldosterone antagonist 18.4 17.8 β blocker 81.6 82.2 α androgenic blocker 8.2 8.3 Direct renin inhibitor 16.3 16.7 Thiazides 4.0 4.3 ACE: Angiotensin-converting-enzyme

\section{Introduction} In this work we introduce the prioritising exclusion process (PEP): a stochastic scheduling mechanism for a priority queue, where high priority customers overtake low priority customers in order to receive service sooner. The queue of customers is represented by a one dimensional lattice, which grows and shrinks as customers arrive and are served. Lattice sites are either empty or occupied by a single particle, representing low and high priority customers respectively; particles hop forwards stochastically into empty sites, corresponding to a high priority customer overtaking the low priority customer immediately ahead of them, but the \textit{exclusion rule} prevents particles hopping into or over occupied sites. The PEP is closely related to a priority queuing model first introduced by Kleinrock \cite{Kleinrock64}, and the subject of more recent work \cite{StanfordTZ14}. Priority queueing systems are relevant in healthcare applications, both as a way to efficiently manage hospital queues with patients of differing urgency \cite{StanfordTZ14}, and to describe actual practise in emergency rooms\cite{HayVB06}. The hopping and exclusion in the PEP is analogous to that of the totally asymmetric simple exclusion process (TASEP) \cite{Spitzer70,ASEP2}, which is one of the most thoroughly studied and central models of non-equilibrium statistical mechanics \cite{Derrida98,Schuetz00,GolinelliM06,BlytheE07}. The TASEP is a microscopic model of a driven system \cite{SchmittmannZ95}, and has been the focus of much mathematical interest due to the fact that it is integrable. Many tools have been applied to or developed for the TASEP. Its exact stationary distribution is known \cite{DerridaDM92,SchutzD93} and can be written in matrix product form \cite{DerridaEHP93,BlytheE07}. Its dynamic properties are studied by means of the Bethe ansatz \cite{GwaS92Burgers,GolinelliM06,deGierE05} and for the infinite lattice powerful techniques from random matrix theory are available \cite{PraehoferS02}. Domain wall theory \cite{KolomeiskySKS98} provides a phenomenological explanation of the stationary behaviour of the TASEP. Domain wall theory is amenable to generalisation to more complicated models that may not be integrable, and we will describe it in more detail later. Recently, several generalisations of the TASEP have been proposed, which allow the lattice length to vary dynamically as is the case in the PEP. Such models are of theoretical interest in statistical physics as they are grand-canonical analogues of the TASEP. The review \cite{ChouMZ11} highlights many biological applications, such as modelling filament growth \cite{SugdenEPR07, SchmittS11, DoroszMP10} and length regulation \cite{JohannEK12,MelbingerRF12}. There have been applications to queueing theory as well. The exclusive queueing process (EQP) \cite{Arita09, YanagisawaTJN10} uses an exclusion process to model the motion of customers waiting in a queue. In the EQP, customers (all of a single priority class) are represented by particles, with empty lattice sites for the space between them. The hopping of particles represents customers shuffling forwards as space becomes available. Beyond applications, these models are also of interest because of the rich phase structure the varying lattice length introduces \cite{SugdenE07, NowakFC07, AritaS12,AritaS13}. As is typical of a queueing system, the PEP has a phase transition from a phase with finite expected queue length, to one with an \textit{unbounded} queue length increasing with time. This transition occurs when the rate of arrival of customers exceeds the service rate. In the latter case, we treat the lattice as infinite in length in order to study the late time limit. The PEP has a natural domain wall structure, and the domain wall dynamics can be derived directly from the microscopic transition rules, similarly to \cite{CividiniHA14}. In the unbounded phase, in the infinite lattice limit, the solution of the domain wall equations gives exact local density profiles. The domain wall solution reveals a second phase transition where the `jam' of high priority customers waiting at the service end becomes infinite. When the expected lattice length remains finite (the \textit{bounded} phase), the domain wall theory leads to approximate solutions only, but the structure from the unbounded phase carries over. We see a remnant of the jamming transition from the unbounded phase as a crossover where the jam of high priority customers at the service end delocalises, and the expected jam length becomes comparable to the queue length. Then by defining `aggregated correlation functions', we find that the form of the unbounded solution can be applied in the bounded phase as an alternative to mean field theory, giving a very accurate calculation of customer waiting times. \subsection{The model} In the lattice bulk, the PEP behaves as a TASEP: sites are either occupied by a single particle or empty, and particles hop forwards into empty sites with rate $p$. At the boundaries the PEP differs from the TASEP. The PEP lattice can be extended on the left by the addition of a filled or empty lattice site, with rates $\la_1$ and $\la_2$ respectively. At the other boundary, the rightmost site is removed with rate $\mu$, irrespective of its occupation. These rules, summarised in Figure~\ref{fig:queuerules}, allow both the lattice length and particle number to vary. \begin{figure}[h] \centering \includegraphics[width=0.6\textwidth]{queuerules} \caption{PEP transition rates, filled circles are occupied sites.} \label{fig:queuerules} \end{figure} We specify a PEP configuration by binary variables $\tau_i$ with $\tau_i = 1$ for a filled site and $\tau_i = 0$ for an empty site. Usually we will number sites from right to left and write a length $n$ configuration as \begin{equation*} \bm{\tau} = \tau_n\tau_{n-1}\ldots\tau_1. \end{equation*} \subsection{A queueing system} \label{sec:QueueingSystem} The PEP can be interpreted as a priority queueing system with two classes of customers. The lattice, itself, is the queue of customers, with filled sites representing high priority customers (class $1$) and empty sites representing low priority customers (class $2$). The rates $\la_1$ and $\la_2$ are the arrival rates of high and low priority customers respectively, and the rate $\mu$, is the rate at which customers are served and leave the queue. In this interpretation, a particle hopping forward one site corresponds to a high priority customer stepping ahead of the low priority customer immediately in front of them. The stochastic overtaking is the scheduling mechanism in this priority queue, giving high priority customers preferential treatment over low priority. The larger the overtake rate $p$, the greater the advantage. The PEP is modelled on a well studied priority queueing system introduced by Kleinrock \cite{Kleinrock64,KleinrockVolII} and now known as the accumulating priority queue (APQ) \cite{StanfordTZ14}. In the APQ, customers have a priority value which accumulates linearly with time. Class $1$ customers accumulate priority faster than class $2$, thus overtaking them in the service queue. The key difference between the APQ and the PEP is that, for a given sequence of arrivals, overtaking in the APQ is deterministic, but in the PEP the overtakes occur stochastically. The PEP is also related to a simpler queueing system, the $M/M/1$ queue (see, for example, \cite{KleinrockVolI}). The total arrival rate of customers to the PEP is $\la = \la_1 + \la_2$, and the service rate is $\mu$. Both these rates are independent of the internal arrangement of the queue, and the prioritising parameter $p$. So, if we are interested only in the total length of the queue, we can treat the system as a $M/M/1$ queue with arrival rate $\la$ and service rate $\mu$. The state of a $M/M/1$ queue is characterised simply by the length, $n$, with probability distribution $P_n$ obeying the master equation \begin{eqnarray} \frac{\dd P_0}{\dd t} = \mu P_1 - \la P_0 \\ \frac{\dd P_n}{\dd t} = \lambda P_{n-1} + \mu P_{n+1} - (\mu+\lambda)P_n, \qquad n > 0. \end{eqnarray} The stationary length distribution of the $M/M/1$ queue (and hence for the PEP) is the solution of $\dd P_n / \dd t = 0$, which is \ \begin{equation} P_n = \left(1 - \frac{\la}{\mu}\right)\left(\frac{\la}{\mu}\right)^n, \label{eq:lengthDist} \end{equation} when $\la < \mu$, i.e. when the total arrival rate is less than the service rate. In this case the system is described as \textit{stable}, because the queue length does not grow without bound. The expected queue length is finite, given by \begin{equation} \langle n \rangle = \frac{\la}{\mu - \la}. \label{eq:avLength} \end{equation} We will call this the \textit{bounded} phase of the PEP. When $\la > \mu$, the system is \textit{unstable} and the expected queue length grows as \begin{equation} \langle n \rangle \sim (\la - \mu) t. \end{equation} In the late time limit, we can treat the queue as infinite in length. We call this the \textit{unbounded} phase of the PEP. At the special value $p = 0$, the PEP really does reduce to a $M/M/1$ queue. Customers arriving at rate $\la$ are high priority with probability $\la_1 / \la$ or low priority with probability $\la_2 / \la$. But as there is no overtaking, there is no reordering of customers, and no advantage in being a class $1$ customer. The probability of high or low at any site is the same as at arrival, i.e. $\la_1/\la$ or $\la_2 / \la$, respectively. In the bounded phase, the probability of the length $n$ configuration $\tau_n\ldots\tau_1$ is then \begin{equation} P(\tau_n\ldots\tau_1) = P_n \left(\frac{\la_1}{\la}\right)^h \left(\frac{\la_2}{\la}\right)^l, \label{eq:pZeroExact} \end{equation} where $h$ is the number of high priority customers, and $l$ the number of low priority customers, i.e. \begin{equation} h = \sum_{i=1}^n \tau_i, \qquad l = n - h. \end{equation} We can contrast the phase behaviour of the PEP with that of the EQP, where the length of the lattice is defined by the position of the last customer, and so the lattice length depends on how fast customers step into the space ahead of them (i.e. the particle hopping rate). The EQP has bounded and unbounded length phases\footnote{convergent and divergent in their terminology}, but with phase boundaries dependent on the hopping rate \cite{Arita09}. \subsection{Density profiles and waiting times} To define a density profile for the PEP we must specify both the site, $i$, and the lattice length, $n$, so that \begin{equation*} \corr{\tau_i}_n = P(\text{queue length is } n \text{, and site } i \text{ is occupied}). \end{equation*} These are one-point functions. We can similarly define higher order correlations \begin{equation} \corr{\tau_{i_1} \tau_{i_2} \ldots \tau_{i_m}}_n, \qquad n \ge i_1 > i_2 > \ldots > i_m \ge 1. \end{equation} The rate equations for the one-point functions are \begin{eqnarray} \fl \frac{\dd}{\dd t} \corr{\tau_1}_1 = \la_1 P_0 + \mu \corr{\tau_2}_2 - (\la + \mu) \corr{\tau_1}_1, \label{eq:rate11} \\ \fl \frac{\dd}{\dd t} \corr{\tau_1}_n = \la \corr{\tau_1}_{n-1} + \mu \corr{\tau_2}_{n+1} + p \corr{\tau_2(1 - \tau_1)}_n - (\la + \mu)\corr{\tau_1}_n, \qquad n > 1, \label{eq:rate1n} \\ \fl \frac{\dd}{\dd t} \corr{\tau_i}_i = \la_1 P_{i-1} + \mu \corr{\tau_{i+1}}_{i+1} -p \corr{\tau_i(1 - \tau_{i-1})}_i - (\la + \mu) \corr{\tau_i}_i, \qquad i > 1, \label{eq:rateii} \\ \fl \frac{\dd}{\dd t} \corr{\tau_i}_n = \la \corr{\tau_i}_{n-1} + \mu \corr{\tau_{i+1}}_{n+1} + p \corr{\tau_{i+1}(1 - \tau_i)}_n \nonumber \\ {} - p \corr{\tau_i(1 - \tau_{i-1})}_n - (\la + \mu) \corr{\tau_i}_n, \qquad i > 1, n > i. \label{eq:ratein} \end{eqnarray} These couple the one-point functions to the two-point correlations, and length $n$ to length $n \pm 1$. The rate equations imply a conserved current of particles across the lattice, but because of the coupling between lengths, some care is required in how this current is defined. We will return to this for the bounded and unbounded phases separately. Viewing the PEP as a queueing system, we are interested in performance measures, and how these differ for high and low priority customers. The current tells us the rate at which customers pass through the system, and from the density profile we can calculate the average waiting time for customers of each class. To calculate waiting times, we use Little's result (see Chapter 2.1 of \cite{KleinrockVolI}), which states that the average waiting time, $\overline{W}_i$, is related to the average number of waiting customers, $\overline{N}_i$, for each class $i = 1, 2$, by \begin{equation} \overline{N}_i = \la_i \overline{W}_i. \label{eq:littles} \end{equation} The average number of high priority customers can be calculated from the density profile as \begin{equation} \overline{N}_1 = \sum_{n=1}^\infty \sum_{i=1}^n \corr{\tau_i}_n, \label{eq:pepNhigh} \end{equation} and the average number of low priority customers is \begin{equation} \overline{N}_2 = \corr{n} - \overline{N}_1 = \frac{\la}{\mu - \la} - \overline{N}_1. \label{eq:pepNlow} \end{equation} Here we take the total time from arrival to removal from the system as the waiting time for a customer. Our aim, now, is to compute the density profile for the PEP. \subsection{Domain wall theory} \label{sec:DWT} Domain wall theory \cite{KolomeiskySKS98, SantenA02} reduces the multi-particle dynamics of the TASEP\footnote{Domain wall theory applies more generally to the \textit{partially} asymmetric simple exclusion process.} to the motion of a single random walker on the lattice. The TASEP boundary conditions (the particle entry and exit rates) create domains of low or high density at the boundaries. These domains extend through the lattice, and where they meet a shock, or domain wall, forms. Domain wall theory models the motion of this shock as a random walk, with the simplifying assumptions that density is constant throughout each domain, and that there is a sharp transition between domains so that the shock can be localised to a single site. Though the stationary solution of the TASEP is known exactly, domain wall theory provides a simple physical explanation of the stationary behaviour \cite{KolomeiskySKS98}, and beyond this it allows accurate approximation of some dynamic properties \cite{SantenA02}. Domain wall theory can also be applied to more complex models \cite{PopkovSSS01,CookZ09} where the exact solution is not known. For the EQP, a domain wall approach was used to describe the global density profile in the divergent length (i.e. unbounded) phase \cite{AritaS12}. In \cite{CividiniHA14} domain wall theory provided an exact solution and we will show that this also occurs in the unbounded phase of the PEP. The PEP has a natural domain wall structure. As high priority customers overtake and reach the service end, they form a jam (Figure \ref{fig:jam}): a jam is a section of high priority customers (filled sites) at the service end ahead of any low priority customer (empty site). The jam is characterised by $k$, the number of consecutive high priority customers. As there are no gaps, there is no overtaking in the jammed region, and the length of the jam reduces only as customers are served. This is similar to the situation in \cite{CividiniHA14}, where a TASEP with parallel update and deterministic bulk motion is considered. In \cite{deGierFS11}, a jam of particles was suggested as the cause of a reduced effective lattice length in the reverse bias regime of the partially asymmetric simple exclusion process. \begin{figure} \centering \subfigure[A queue configuration with jam length $k = 3$] { \includegraphics[width=0.4\textwidth]{lattice_jam} \label{fig:jam} } \subfigure[Conditional probability that site $i$ is filled given queue length $n$ and jam length $k$] { \includegraphics[width=0.4\textwidth]{domain_wall_prob} \label{fig:dwprob} } \caption{} \end{figure} Let us assume that the region beyond the jam has uniform density, and that the conditional probability that site $i$ is filled, given the queue length, $n$, and jam length, $k$, is (Figure~\ref{fig:dwprob}) \begin{equation} P(\tau_i = 1 | n, k) = \cases{ 1 & $1 \le i \le k$ \\ 0 & $i = k + 1$ \\ \alpha & $k + 2 \le i \le n$. } \label{eq:dwprob} \end{equation} Then the bulk equation ($n > k+1, k > 0$) for $P(n,k)$, the probability of a length $k$ jam in a length $n$ queue, is \begin{equation} \eqalign{ \fl \frac{\dd}{\dd t} P(n, k) = & \la P(n - 1, k) + \mu P(n + 1, k + 1) + \mu (1 - \alpha) \alpha^k P(n + 1, 0) \\ & {} + p \alpha P(n, k - 1) - (\la + \mu + p \alpha) P(n, k). } \label{eq:effDW} \end{equation} Let us explain the meaning of each of the terms in equation \eref{eq:effDW}. The term \begin{equation*} \la P(n - 1, k) \end{equation*} is the entry into the $(n, k)$ configuration from a length $n$ queue due to the arrival of a customer, and the terms \begin{equation*} \mu P(n + 1, k + 1) + \mu (1 - \alpha) \alpha^k P(n + 1, 0), \end{equation*} represent the service of a customer. The second term is the transition into the $k$-jam state from the $(k = 0)$-jam state by the service of a low priority customer who was followed by $k$ consecutive high priority customers. The term \begin{equation*} p \alpha P(n, k - 1), \end{equation*} is a $(k-1)$-jam extending to length $k$ with rate $p \alpha$: there is a high priority customer at site $k + 1$ with probability $\alpha$, which overtakes with rate $p$ the low priority customer at site $k$ (the low priority customer marking the end of the $(k-1)$-jam). The low priority customer thus moves to position $k+1$ defining the new end of the jam. The loss terms \begin{equation*} -(\la + \mu + p \alpha) P(n, k), \end{equation*} are the rate at which the $(n, k)$ configuration is left due to a customer arrival or service, or growth of the jam. In the next section, we will show that the $n \to \infty$ limit of \eref{eq:effDW}, and the corresponding equation for $k = 0$, follows from the unbounded phase master equation. We find the exact stationary solutions of these equations, describing the behaviour of the jam on an infinite lattice. In the bounded phase (Section~\ref{sec:bounded_phase}), domain wall theory leads to two complementary approximations. One reveals information about the length dependence and the other about waiting times. \section{The unbounded phase} We consider first the unbounded phase of the PEP, where the total arrival rate exceeds the service rate ($\la > \mu$). Recall that the expected lattice length grows as $\corr{n} \sim (\la - \mu)t$. In our domain wall picture, the jam increases with rate $p \alpha$ and decreases with rate $\mu$ (ignoring for the moment the $\mu(1 - \alpha)\alpha^k P(n + 1, 0)$ term in \eref{eq:effDW}). If $p \alpha > \mu$, the jam will grow as $(p \alpha - \mu) t$ unless it reaches the arrival end of the queue, but if $p \alpha < \mu$, the jam length will fluctuate near $0$. \begin{figure}[ht] \centering \subfigure[ With $p = 3$, there is a growing jam. ] { \includegraphics[width=0.45\textwidth]{unbounded_unbounded_time} \label{fig:unbounded_unbounded_time} } \qquad \subfigure[ With $p = 1.8$ the jam remains finite. ] { \includegraphics[width=0.45\textwidth]{unbounded_bounded_time} \label{fig:unbounded_bounded_time} } \caption{ Time evolution of an unbounded queue with arrival rates $\la_1 = 1.1$, $\la_2 = 0.1$, and service rate $\mu = 1$. Density $\corr{\tau_i}(t)$ is averaged over a small time period, and the inset shows mapping of density to colour. } \label{fig:unbounded_evolution} \end{figure} Figure~\ref{fig:unbounded_evolution} shows simulation results for a growing jam (Figure~\ref{fig:unbounded_unbounded_time}), and a jam fluctuating near $0$ (Figure~\ref{fig:unbounded_bounded_time}). The figures show the time evolution of the density profile, starting from an empty queue, with the density at each site calculated by averaging over a short time period. The queue length grows with rate $\la - \mu$, and the low density region beyond the jam of high priority customers is fairly regular. We will focus on the situation in Figure~\ref{fig:unbounded_bounded_time}, where the jam length remains finite. In Figure~\ref{fig:unbounded_unbounded_time} where the jam grows, we see that it nevertheless grows more slowly than the queue length, and we will comment on this later. We would like to understand the late time behaviour in the unbounded phase, but as for any fixed configuration $\tau_n\ldots\tau_1$, \begin{equation} \lim_{t \to \infty} P(\tau_n \ldots \tau_1) = 0, \end{equation} the \textit{global} description is uninformative. Instead, we follow the approach of \cite{SugdenE07} and consider \textit{local} behaviour relative to a specified reference frame. The \textit{service frame} is fixed at the right hand end of the lattice where customers are served and depart. For a general but finite section of length $m$, \begin{equation*} \tau_m\ldots\tau_1, \end{equation*} the service frame probability in a length $n$ lattice is defined as \begin{equation} \Pserv(\tau_m\ldots\tau_1; n) = \sum_{\tau_n,\ldots,\tau_{m+1}=0,1} P(\tau_n\ldots\tau_{m+1}\tau_m\ldots\tau_1). \end{equation} We will also consider the \textit{arrival frame}, fixed at the left hand end of the lattice. In this case we number sites left to right. A general length $m$ section is written \begin{equation*} \tau_1\ldots\tau_m \end{equation*} and the arrival frame probability in a length $n$ lattice is defined by \begin{equation} \Parr(\tau_1\ldots\tau_m; n) = \sum_{\tau_{m+1},\ldots,\tau_n=0,1} P(\tau_1\ldots\tau_{m}\tau_{m+1}\ldots\tau_n). \end{equation} In the $t \to \infty$ limit, the expected lattice length is infinite, and thus we are interested in the $n \to \infty$ limits \begin{equation} \eqalign{ \Pserv(\tau_m\ldots\tau_1) = \lim_{n \to \infty} \Pserv(\tau_m\ldots\tau_1; n), \\ \Parr(\tau_1\ldots\tau_m) = \lim_{n \to \infty} \Parr(\tau_1\ldots\tau_m; n), \\ } \label{eq:ub:localProbs} \end{equation} with the assumption that this limit exists. We will write the rate equations for the arrival and service frame probabilities \eref{eq:ub:localProbs} in this limit, then seek the stationary solution. To do this, we use a domain wall ansatz for \eref{eq:ub:localProbs}, and show that this leads to the exact stationary solution for these quantities. \subsection{Domain wall ansatz} We use domain wall theory to form an ansatz for the service frame probabilities. We consider a general but finite section of length $m$, with a length $k$ jam, \begin{equation} \tau_m\ldots\tau_{k+2}01^k = \bm{\tau}01^k, \label{eq:ub:kjamconfig} \end{equation} where $1^k$ indicates a string of $k$ $1$'s. The configuration of any finite segment can be written this way, as long as we can take $m \ge k + 1$. We will assume that the conditional probability for a high at site $i$ given a jam of length $k$ is \begin{equation} P(\tau_i = 1 | k) = \cases{ 1 & $1 \le i \le k$ \\ 0 & $i = k + 1$ \\ \alpha & $k + 2 \le i$, } \end{equation} which is \eref{eq:dwprob} in the $n \to \infty$ limit. Then the probability of the finite segment \eref{eq:ub:kjamconfig} is \begin{equation} \eqalign{ \Pserv(\bm{\tau}01^k) &= \sum_{\tau_\infty,\ldots,\tau_{m+1}=0,1}P(\ldots\tau_{m+1}\tau_m\ldots\tau_{k+2}01^k) \\ &= \alpha^h (1 - \alpha)^l \Pjam(k), } \label{eq:ub:jamAnsatz} \end{equation} where $h$ is the number of highs in the configuration beyond the jam up to position $m$, and $l$ is the number of lows, that is \begin{equation} h = \sum_{i = k+2}^m \tau_i, \qquad l = m - h - k - 1, \end{equation} and $\Pjam(k)$ is the probability of a length $k$ jam. The jam probabilities are normalised such that \begin{equation} \sum_{k = 0}^\infty \Pjam(k) = 1. \label{eq:ub:jamNorm} \end{equation} Equation \eref{eq:ub:jamAnsatz} is the domain wall ansatz for the stationary service frame probabilities. \subsection{The service frame} \label{sec:ub:service_frame} In this section we write the general service frame rate equations, then apply the domain wall ansatz \eref{eq:ub:jamAnsatz}. We use the notation $\bm{\tau}|_{(i,i-1)}$ to indicate the exchange of customers in places $i$ and $i - 1$. That is, for $\bm{\tau} = \tau_r\ldots\tau_1$ \begin{equation*} \bm{\tau}|_{(i,i-1)} = \tau_r\ldots\tau_{i+1}\tau_{i-1}\tau_i\tau_{i-2}\ldots\tau_1, \end{equation*} and \begin{equation*} 0\bm{\tau}|_{(r+1,r)} = \tau_r 0 \tau_{r-1}\ldots\tau_1. \end{equation*} The stationary rate equation for the $k$-jam configuration \eref{eq:ub:kjamconfig} with $k \geq 1$ is \begin{eqnarray} \fl 0 &= \frac{\dd}{\dd t} \Pserv(\bm{\tau}01^k) \nonumber \\ \fl &= \mu \Pserv\left(\bm{\tau}01^{k+1}\right) + \mu \Pserv\left(\bm{\tau}01^k0\right) \nonumber \\ \fl & {} + p \tau_m \Pserv\left(0\bm{\tau}01^k|_{(m+1,m)}\right) + \sum_{i = k + 2}^m p (1 - \tau_i)\tau_{i-1} \Pserv\left(\bm{\tau}01^k|_{(i,i-1)}\right) + p \Pserv\left(\bm{\tau}101^{k-1}\right) \nonumber \\ \fl & {} -\mu \Pserv(\bm{\tau}01^k) - \sum_{i = k + 2}^m p \tau_i (1 - \tau_{i-1}) \Pserv(\bm{\tau}01^k) - p(1 - \tau_m) \Pserv(1\bm{\tau}01^k). \label{eq:ub:statrate} \end{eqnarray} Let us again explain the various terms. The terms \begin{equation*} \mu \Pserv\left(\bm{\tau}01^{k+1}\right) + \mu \Pserv\left(\bm{\tau}01^k0\right), \end{equation*} give the rate of arrival to the $k$-jam configuration after, respectively, a high or low priority customer is served. Then there are the hopping terms. A high in $m$th place in $\bm{\tau}01^k$ can arrive from place $m + 1$: \begin{equation*} p \tau_m \Pserv\left(0\bm{\tau}01^k|_{(m+1,m)}\right). \end{equation*} Overtaking within the low density region behind the jam is given by \begin{equation*} \sum_{i = k + 2}^m p (1 - \tau_i)\tau_{i-1} \Pserv\left(\bm{\tau}01^k|_{(i,i-1)}\right), \end{equation*} and a $(k-1)$-jam extends to a $k$-jam when a high hops onto the end: \begin{equation*} p \Pserv\left(\bm{\tau}101^{k-1}\right). \end{equation*} The loss term \begin{equation*} -\mu \Pserv(\bm{\tau}01^k), \end{equation*} is the reduction of the jam as a customer is served, and \begin{equation*} -\sum_{i = k + 2}^m p \tau_i (1 - \tau_{i-1}) \Pserv(\bm{\tau}01^k) \end{equation*} are overtakings within the $m$ places of $\bm{\tau}01^k$. The final loss term \begin{equation*} - p(1 - \tau_m) \Pserv(1\bm{\tau}01^k) \end{equation*} arises if the configuration has a low in $m$th place, which can be overtaken by a high from place $m+1$. Finally we note that the terms involving the arrival rates $\lambda_1$ and $\lambda_2$ do not appear in \eref{eq:ub:statrate} as they cancel from the stationary rate equations in the $n\rightarrow\infty$ limit. Substituting the ansatz \eref{eq:ub:jamAnsatz}, the terms representing overtaking within the $m$ sites of $\bm{\tau}01^k$ combine and telescope to \begin{eqnarray*} p \left(\sum_{i=k+2}^m (1 - \tau_i)\tau_{i-1} - \sum_{i=k+2}^m \tau_i(1 - \tau_{i-1})\right) \alpha^h (1 - \alpha)^l \Pjam(k) \\ = p \left(\tau_{k+1} - \tau_m\right)\alpha^h (1 - \alpha)^l \Pjam(k) \\ = - p \tau_m\alpha^h (1 - \alpha)^l \Pjam(k); \end{eqnarray*} recall that $\tau_{k+1} = 0$. The factor $\alpha^h (1 - \alpha)^l$ is common to all terms in the rate equation. Cancelling, and simplifying leaves \begin{equation} \fl 0 = p \alpha \Pjam(k-1) + \mu \Pjam(k+1) + \mu (1 - \alpha)\alpha^k \Pjam(0) - (\mu + p \alpha) \Pjam(k). \label{eq:ub:exactDW} \end{equation} This agrees with the $n \to \infty$ limit of \eref{eq:effDW}, but we have derived it from the full PEP rate equations. Were it not for the $\Pjam(0)$ term, this equation for the position of the jam would have the same form as the domain wall theory for the TASEP \cite{SantenA02}. The $k = 0$ case differs only slightly. In this case the rate equation is \begin{eqnarray} \fl 0 &= \frac{\dd}{\dd t} \Pserv(\bm{\tau}0) \nonumber \\ \fl &= \mu \Pserv\left(\bm{\tau}01\right) + \mu \Pserv\left(\bm{\tau}00\right) \nonumber \\ \fl & {} + p \tau_m \Pserv\left(0\bm{\tau}0|_{(m+1,m)}\right) + \sum_{i = 2}^m p (1 - \tau_i)\tau_{i-1} \Pserv\left(\bm{\tau}0|_{(i,i-1)}\right) \\ \fl & {} - \mu P\left(\bm{\tau}0\right) - \sum_{i = 2}^m p \tau_i (1 - \tau_{i-1}) \Pserv\left(\bm{\tau}0\right) - p(1 - \tau_m) \Pserv\left(1\bm{\tau}0\right) \nonumber \end{eqnarray} Substituting the ansatz \eref{eq:ub:jamAnsatz}, this reduces to \begin{equation} 0 = \mu \Pjam(1) - (\mu \alpha + p \alpha) \Pjam(0), \label{eq:ub:zeroDW} \end{equation} and rearranging gives \begin{equation} \Pjam(1) = \frac{p \alpha}{\mu} \Pjam(0) + \alpha \Pjam(0). \end{equation} With this as the base case, we use \eref{eq:ub:exactDW} to show by induction that \begin{equation} \Pjam(k) = \frac{p \alpha}{\mu} \Pjam(k-1) + \alpha^k \Pjam(0), \qquad k \ge 1. \end{equation} This recurrence for $\Pjam(k)$ has solution \begin{eqnarray} \Pjam(k) &= \sum_{i=0}^k \left(\frac{p \alpha}{\mu}\right)^{k-i} \alpha^i \Pjam(0) \nonumber \\ &= \frac{p \left(\frac{p \alpha}{\mu}\right)^k - \mu \alpha^k}{p - \mu} \Pjam(0). \label{eq:ub:PjamSol} \end{eqnarray} The normalisation condition \eref{eq:ub:jamNorm} fixes \begin{equation} \Pjam(0) = (1 - \alpha)(1 - \frac{p \alpha}{\mu}), \label{eq:ub:PjamNorm} \end{equation} subject to the constraint \begin{equation} p \alpha < \mu. \label{eq:ub:jamConstraint} \end{equation} The domain wall picture makes the meaning of this constraint clear. The jam of high priority customers grows with rate $p \alpha$ and is reduced with rate $\mu$. If $p \alpha > \mu$ the jam grows with rate $p \alpha - \mu > 0$, that is \begin{equation} \langle k \rangle \sim (p \alpha - \mu) t, \end{equation} and as $t \to \infty$ the expected length of the jam becomes infinite. In contrast, when \eref{eq:ub:jamConstraint} is satisfied, the service rate is fast enough to prevent a backlog of high priority customers, and the expected jam length is finite and given by \begin{equation} \langle k \rangle = \sum_{k=1}^\infty k \Pjam(k) = \frac{\alpha}{1 - \alpha} + \frac{p \alpha}{\mu - p \alpha}. \end{equation} \subsection{The arrival frame} To determine $\alpha$, we examine the PEP in the arrival frame. Sites are now numbered left to right, and the interchange operation is defined as \begin{equation*} \bm{\tau}|_{(i,i+1)} = \tau_1\ldots\tau_{i-1}\tau_{i+1}\tau_i\tau_{i+2}\ldots\tau_r, \end{equation*} for $\bm{\tau} = \tau_1\ldots\tau_r$. We will assume that the jam is always far from the arrival end. In the late time limit this is guaranteed if \begin{equation} p \alpha < \la. \label{eq:ub:arrivalConstraint} \end{equation} This condition is clearly met when \eref{eq:ub:jamConstraint} is satisfied, but we will show that $\alpha$ can be determined consistently with this requirement. Then in the arrival frame, the ansatz \eref{eq:ub:jamAnsatz} implies that for a configuration on the first $m$ sites, \begin{equation*} \bm{\tau} = \tau_1\ldots\tau_m, \end{equation*} the arrival frame probability has the form \begin{equation} \Parr(\bm{\tau}) = \alpha^h(1 - \alpha)^l, \label{eq:ub:arrAnsatz} \end{equation} where \begin{equation*} h = \sum_{i = 1}^m \tau_i, \qquad l = m - h. \end{equation*} The stationary rate equation for this configuration is \begin{eqnarray} 0 &= \frac{\dd}{\dd t}\Parr(\bm{\tau}) \nonumber \\ &= \la_1 \tau_1 \Parr(\tau_2\ldots\tau_m) + \la_2 (1 - \tau_1) \Parr(\tau_2\ldots\tau_m) \nonumber \\ & {} + \sum_{i=1}^{m-1} p (1 - \tau_i)\tau_{i+1} \Parr(\bm{\tau}|_{(i,i+1)}) + p (1 - \tau_m) \Parr(\bm{\tau}1|_{(m,m+1)}) \\ & {} -\la \Parr(\bm{\tau}) - \sum_{i = 1}^{m-1} p \tau_i(1 - \tau_{i+1}) \Parr(\bm{\tau}) - p\tau_m \Parr(\bm{\tau}0). \nonumber \end{eqnarray} Substituting \eref{eq:ub:arrAnsatz}, the summed hopping terms again combine and telescope, and the factors of $\alpha$ and $1 - \alpha$ common to all terms can be cancelled. This leaves \begin{equation*} 0 = -\la \alpha(1 - \alpha) + p \alpha^2(1 - \alpha) - p \tau_1 \alpha (1 - \alpha) + \tau_1 \la_1(1 - \alpha) + (1 - \tau_1) \la_2 \alpha, \end{equation*} which for both $\tau_1 = 0$ and $\tau_1 = 1$ reduces to \begin{equation*} p \alpha^2 - (p + \la) \alpha + \la_1 = 0. \end{equation*} The two solutions are \begin{equation} \alpha_\pm = \frac{p + \la \pm \sqrt{(p - \la)^2 + 4 p \la_2}}{2p}, \label{eq:ub:alphapm} \end{equation} with $0 < \alpha_- < 1$ and $\alpha_+ > 1$ for $p, \la_1, \la_2 > 0$. As $\alpha$ is a density value we must take $\alpha = \alpha_-$. Substituting this value for $\alpha$, it is seen that $p \alpha < \la$ (the constraint \eref{eq:ub:arrivalConstraint}) is satisfied for all physical parameter values. The jam always grows more slowly than the queue length, as illustrated by the simulation results in Figure~\ref{fig:unbounded_unbounded_time}. In the limit $p \to 0$, $\alpha \to \la_1 / \la$, the expected occupancy of any site when there is no overtaking. With $\alpha$ given by \eref{eq:ub:alphapm}, and the jam probabilities, $\Pjam(k)$ by \eref{eq:ub:PjamSol} and \eref{eq:ub:PjamNorm}, we thus have the stationary domain wall solution describing the motion of the high priority jam on an infinite lattice. Through \eref{eq:ub:jamAnsatz} and \eref{eq:ub:arrAnsatz}, this then gives the exact stationary service and arrival frame probabilities $\Pserv(\tau_m\ldots\tau_1)$ and $\Parr(\tau_1\ldots\tau_m)$. In the next section, we compare this solution to simulation results, where the lattice length, $n$, is large, but necessarily finite. In doing so we make the assumption that the large $n$ behaviour converges to the $n \to \infty$ limit. We note also that the result \eref{eq:ub:PjamNorm}, the probability to not have a jam, can be understood by the following heuristic argument\footnote{We thank an anonymous referee for pointing this out.}. The number of low priority customers in the system at time $t$ can be expressed as $(1-\alpha)(\lambda- \mu)t$, i.e. the ratio of low priority customers $(1-\alpha)$ times the expected length of the queue. It can also be expressed as $(\lambda_2-P_{\rm jam}(0)\mu)t$, i.e. (arrival rate $-$ service rate) times time. Comparing the two expressions and using \eref{eq:ub:alphapm} we obtain $\eref{eq:ub:PjamNorm}$. \subsection{Density profile, conserved currents, and service rates} The density at site $i$ in the service frame, $\corrinf{\tau_i}$, is computed from the domain wall solution as \begin{eqnarray} \corrinf{\tau_i} &= \alpha \sum_{k = 0}^{i - 2}\Pjam(k) + \sum_{k = i}^\infty\Pjam(k) \nonumber \\ &= \alpha + (1 - \alpha)\left(\frac{p \alpha}{\mu}\right)^i. \label{eq:ub:onePoint} \end{eqnarray} Figure~\ref{fig:unbounded_density} shows $\corrinf{\tau_i}$ computed from \eref{eq:ub:onePoint} plotted against simulation results for a range of parameters chosen, necessarily, with $p \alpha < \mu$. In all cases we see excellent agreement. The only deviation occurs for $p = 1.8$ in Figure~\ref{fig:unbounded_density_a}, which with $p \alpha \simeq 0.98$ is very close to the critical value $p \alpha = \mu$. We believe that this difference occurs because of the slowing convergence of simulation results as we near the critical point. A study of the critical behaviour, as for example the numerical study\cite{AritaS13} for the EQP, would be required to confirm this. \begin{figure}[t] \centering \subfigure[Density profile at the service end of the unbounded queue; $\la_1 = 1.1, \la_2 = 0.1, \mu = 1$ with $p = 1$ (black), $p = 1.4$ (mid-gray), and $p = 1.8$ (light gray). ] { \includegraphics[width=0.4\textwidth]{unbounded_density_a} \label{fig:unbounded_density_a} } \subfigure[Density profile at the service end of the unbounded queue; $\la_1 = 0.9, \la_2 = 0.3, \mu = 1$ with $p = 1$ (black), $p = 1.8$ (mid-gray), and $p = 3$ (light gray). ] { \includegraphics[width=0.4\textwidth]{unbounded_density_b} \label{fig:unbounded_density_b} } \caption{ Density profiles of the unbounded PEP in the service frame. Triangle markers for simulation results plotted against calculated profile $\corr{\tau_i}$. } \label{fig:unbounded_density} \end{figure} With $p \alpha > \mu$, the jam grows to fill any finite section at the service end. Checking Figure~\ref{fig:unbounded_unbounded_time}, we see that the rate of growth is consistent with $\corr{k} \sim (p \alpha - \mu) t$. We can check that the domain wall solution satisfies the rate equations for the one point functions \eref{eq:rate11} -- \eref{eq:ratein}. To obtain the service frame rate equations, we take the $n \to \infty$ limit, allowing us to neglect the boundary cases. The bulk equations can be written \begin{equation} \frac{\dd}{\dd t} \corrinf{\tau_i} = J_\infty^{(i+1)} - J_\infty^{(i)}, \qquad i \ge 1, \label{eq:ub:dtJ} \end{equation} where \begin{eqnarray} J_\infty^{(1)} = \mu \corrinf{\tau_1} \nonumber \\ J_\infty^{(i)} = \mu \corrinf{\tau_{i}} + p \corrinf{\tau_i(1 - \tau_{i-1})}, \qquad i \ge 2. \label{eq:ub:J_eqns} \end{eqnarray} In the stationary state the time derivatives are zero so \eref{eq:ub:dtJ} defines a conserved current \begin{equation} J_\infty = J_\infty^{(1)} = J_\infty^{(2)} = \ldots \end{equation} The bulk current, $J_\infty^{(i)}$, has the usual TASEP hopping term, $p\corrinf{\tau_i(1 - \tau_{i-1})}$. The additional term $\mu \corrinf{\tau_i}$ arises due to the choice of reference frame. The two-point correlation $\corrinf{\tau_{i+1}(1 - \tau_i)}$ is computed as \begin{eqnarray} \corrinf{\tau_{i+1}(1-\tau_i)} &= \alpha (1 - \alpha) \sum_{k = 0}^{i-2}\Pjam(k) + \alpha \Pjam(i-1) \nonumber \\ &= \alpha(1 - \alpha)\left(1 - \left(\frac{p \alpha}{\mu}\right)^i\right). \label{eq:ub:twoPoint} \end{eqnarray} The resulting service frame current is \begin{equation} J_\infty = p \alpha(1 - \alpha) + \mu \alpha, \end{equation} and is the rate at which particles exit the system. In terms of the queueing model, the current is the average rate at which high priority customers leave the queue. As the total rate at which customers leave the system is $\mu$,\footnote{If the queue was ever empty, the rate at which customers leave would be less than the service rate, but in the unbounded phase this is not a concern.} low priority customers leave the queue at rate \begin{equation} \mu - J_\infty = (\mu - p \alpha)(1 - \alpha). \end{equation} The constraint $p \alpha < \mu$ (equation \eref{eq:ub:jamConstraint}) ensures that this rate is greater than zero and low priority customers always receive a share of the service. For $p \alpha > \mu$, the jam of high priority customers at the service end becomes unbounded, and low priority customers can no longer reach the front of the queue to be served. Thus the low priority current has a second order phase transition at $p \alpha = \mu$ . The phase transition subdivides the unbounded phase of the PEP. By fixing values for $\la = \la_1 + \la_2$ and $\mu$, we can plot illustrative two dimensional phase diagrams with $\la_1$ and $p$ as the axes. Using \eref{eq:ub:alphapm} for $\alpha = \alpha_-$, the curve where $p \alpha = \mu$ is given by \begin{equation} \la_1^{(\infty)}(p) = \cases{ \la & $p < \mu$ \\ \mu\left(1 + \frac{\la - \mu}{p}\right) & $p \ge \mu$, } \end{equation} and $p \alpha < \mu$ for $\la_1 < \la_1^{(\infty)}(p)$. Figure~\ref{fig:phase_diagram_unbounded} shows the phase diagram for $\la = 1.2, \mu = 1$. \begin{figure}[ht] \centering \includegraphics{phase_diagram_unbounded} \caption{ Subdivisions of the unbounded phase with $\la = 1.2, \mu = 1$. } \label{fig:phase_diagram_unbounded} \end{figure} The function $\la_1^{(\infty)}(p)$ is decreasing in $p$ and $\lim_{p \to \infty} \la_1(p) = \mu$. Therefore, the transition into the `infinite jam' phase occurs only if $\la_1 > \mu$. This is marked by the lower dashed line. The upper dashed line marks the $\la_1 = \la, \la_2 = 0$ boundary. \section{The bounded phase} \label{sec:bounded_phase} In the queueing theory interpretation, it is the bounded phase of the PEP (with $\la < \mu$) that is of greatest interest. In this phase the queue lengths and waiting times remain finite, and we can compare how the waiting time varies with the overtake rate, $p$, for high and low priority class customers. The fluctuating lattice length proves a challenge in applying domain wall theory directly to the bounded phase. We will take two approaches, each leading to an approximate solution revealing different aspects of the system. The first method tells us about the shape and length dependence of the density profiles, while the second method allows us to calculate customer waiting times. \subsection{Domain wall ansatz} \label{sec:boundeddw} To apply the domain wall ansatz directly in the bounded phase, we consider a general length $n$ configuration with a length $k$ jam, \begin{equation} \tau_n\ldots\tau_{k+2}01^k = \bm{\tau}01^k. \end{equation} The stationary rate equation for $k > 0$, $n > k +1$ is \begin{eqnarray} \fl 0 &= \frac{\dd}{\dd t} P(\bm{\tau}01^k) \nonumber \\ \fl &= \la_1 \tau_n P(\tau_{n-1}\ldots\tau_{k+2}01^k) + \la_2 (1 - \tau_n) P(\tau_{n-1}\ldots\tau_{k+2}01^k) \nonumber \\ \fl & {} + \sum_{i=k+2}^n p (1 - \tau_i) \tau_{i-1} P(\bm{\tau}01^k|_{(i,i-1)}) + p P(\bm{\tau}101^{k-1}) + \mu P(\bm{\tau}01^{k+1}) + \mu P(\bm{\tau}01^k0) \nonumber \\ \fl & {} - (\la + \mu)P(\bm{\tau}01^k) - \sum_{i=k+2}^n p \tau_i(1 - \tau_{i-1}) P(\bm{\tau}01^k), \label{eq:bd:kJamFull} \end{eqnarray} and for $k = 0$, $n > 1$ \begin{eqnarray} \fl 0 &= \frac{\dd}{\dd t} P(\bm{\tau}0) \nonumber \\ \fl &= \la_1 \tau_n P(\tau_{n-1}\ldots\tau_{2}0) + \la_2 (1 - \tau_n) P(\tau_{n-1}\ldots\tau_{2}0) \nonumber \\ \fl & {} + \sum_{i=2}^n p (1 - \tau_i) \tau_{i-1} P(\bm{\tau}0|_{(i,i-1)}) + \mu P(\bm{\tau}01) + \mu P(\bm{\tau}00) \nonumber \\ \fl & {} -(\la + \mu)P(\bm{\tau}0) - \sum_{i=2}^n p \tau_i(1 - \tau_{i-1}) P(\bm{\tau}0). \label{eq:bd:0JamFull} \end{eqnarray} These are ``bulk" equations, valid when the jam is away from the arrival end of the queue and are of the form discussed in Section~\ref{sec:DWT}. To see this, define \begin{equation} P(n, k) = \sum_{\tau_n, \ldots, \tau_{k+2} = 0,1} P(\tau_n\ldots\tau_{k+2}01^k), \end{equation} which is the probability of a length $k$ jam in a length $n$ queue. For $k > 0, n > k + 1$, summing \eref{eq:bd:kJamFull} and applying the domain wall ansatz \eref{eq:dwprob} gives \begin{equation} \eqalign{ \fl 0 = \frac{\dd}{\dd t} P(n, k) = & \la P(n - 1, k) + \mu P(n + 1, k + 1) + \mu (1 - \alpha) \alpha^k P(n + 1, 0) \\ & {} + p \alpha P(n, k - 1) - (\la + \mu + p \alpha) P(n, k). } \label{eq:bd:Pnk} \end{equation} which is exactly \eref{eq:effDW}. And for $k = 0, n > 1$, summing \eref{eq:bd:0JamFull} gives \begin{equation} \eqalign{ \fl 0 &= \frac{\dd}{\dd t} P(n, 0) \\ \fl &= \la P(n - 1, 0) + \mu P(n + 1, 1) + \mu (1 - \alpha) P(n+1, 0) - (\la + \mu + p \alpha) P(n, 0). } \label{eq:bd:Pn0} \end{equation} The simple domain wall picture breaks down when the jam reaches the arrival end, i.e. for configurations $01^{n-1}$ or $1^n$. Our strategy is to find a solution of the bulk equations, without requiring it to satisfy these boundary equations. We can hope that this will give an approximation to the true solution. What we will show is that, within the range of validity, the approximation is very good. \subsubsection{Length assumption.} To solve the bulk equations we assume the length dependence factorises as \begin{equation} P(n, k) = P_n \Pjam^*(k), \label{eq:bd:lengthAssumption} \end{equation} where $P_n$ is the length distribution \eref{eq:lengthDist}. Then equation \eref{eq:bd:Pnk}, for $k > 0$, becomes \begin{equation} 0 = p \alpha \Pjam^*(k - 1) + \la \Pjam^*(k + 1) + \la (1 - \alpha) \alpha^k \Pjam^*(0) - (\la + p \alpha) \Pjam^*(k), \quad \label{eq:bd:kjam} \end{equation} and equation \eref{eq:bd:Pn0}, for $k = 0$, gives \begin{equation} 0 = \la \Pjam^*(1) - (\la \alpha + p \alpha) \Pjam^*(0). \end{equation} These have the same form as the unbounded queue domain wall equations, \eref{eq:ub:exactDW}, \eref{eq:ub:zeroDW}, but with $\la$ in place of $\mu$. Therefore they are solved by \begin{eqnarray} \Pjam^*(k) &= \sum_{i=0}^k \left(\frac{p \alpha}{\la}\right)^{k-i} \alpha^i \Pjam^*(0) \nonumber \\ &= \frac{p \left(\frac{p \alpha}{\la}\right)^k - \la \alpha^k}{p - \la} \Pjam^*(0). \label{eq:bd:Pjam} \end{eqnarray} The normalisation of $\Pjam^*(k)$ must be independent of $n$. For the solution to be valid for $n \to \infty$ (as there is no cap on queue length) we must require \begin{equation} \sum_{k = 0}^\infty \Pjam^*(k) = 1, \label{eq:bd:jamNorm} \end{equation} fixing \begin{equation} \Pjam^*(0) = (1 - \alpha)\left(1 - \frac{p \alpha}{\la}\right), \end{equation} subject to the constraint \begin{equation} p \alpha < \la. \label{eq:bd:jamConstraint} \end{equation} The total probability at each length, $n$, must sum to the length distribution \eref{eq:lengthDist}, that is \begin{equation} \sum_{k = 0}^{n - 1} P(n, k) + P(1^n) = P_n. \end{equation} As only $P(1^n)$ is undetermined, we must have that \begin{eqnarray} P(1^n) & = P_n - \sum_{k=0}^{n-1} P(n, k) \nonumber \\ & = P_n \sum_{k = n}^\infty \Pjam^*(k) \nonumber \\ & = P_n \frac{p(1 - \alpha)\left(\frac{p \alpha}{\la}\right)^n - \la\left(1 - \frac{p \alpha}{\la}\right) \alpha^n}{p - \la}. \label{eq:bd:PAll1s} \end{eqnarray} This is analogous to the unbounded queue. There the probability that the first $n$ sites from the service end are filled is $\sum_{k=n}^\infty \Pjam(k)$. To determine $\alpha$ we return to the general $k$-jam equation \eref{eq:bd:kJamFull}\footnote{This is for $k>0$, but \eref{eq:bd:0JamFull} for $k=0$ gives the same result.}, and apply the domain wall ansatz \eref{eq:dwprob} and the length assumption \eref{eq:bd:lengthAssumption}, leaving \begin{equation} \eqalign{ \fl 0 = & \frac{\la_1 \mu}{\la} \tau_n (1 - \alpha) \Pjam^*(k) + \frac{\la_2 \mu}{\la} (1 - \tau_n) \alpha \Pjam^*(k) + p \alpha^2(1 - \alpha) \Pjam^*(k-1) \\ \fl & {} + \la \alpha(1 - \alpha) \Pjam^*(k+1) + \la (1 - \alpha)^2 \alpha^{k+1} \Pjam^*(0) \\ \fl & {} - (\la + \mu + p \tau_n)\alpha(1 - \alpha) \Pjam^*(k). } \label{eq:bd:kJamSubbed} \end{equation} Multiplying \eref{eq:bd:kjam} by $\alpha(1 - \alpha)$ and subtracting from \eref{eq:bd:kJamSubbed}, we then consider $\tau_n = 0$ and $\tau_n = 1$ separately. Both cases reduce to \begin{equation} 0 = p \alpha^2 - (p + \mu) \alpha + \frac{\la_1 \mu}{\la}, \label{eq:bd:alphaQuadratic} \end{equation} with solutions \begin{equation} \alpha_\pm = \frac{p + \mu \pm \sqrt{(p+\mu)^2 - 4 p \frac{\la_1 \mu}{\la}}}{2p}. \label{eq:bd:alphapm} \end{equation} Using the inequalities \begin{equation} \la_1 < \frac{\la_1 \mu}{\la} < \mu, \end{equation} (the first inequality holds as the queue is bounded) we see that $0 < \alpha_- < 1$, and $\alpha_+ > 1$ when $p, \la_1, \la_2, \mu > 0$. Again, we must take $\alpha = \alpha_-$ to have a proper density value. As for the unbounded queue, in the limit $p \to 0$, $\alpha \to \la_1 / \la$, the expected occupancy of any site when there is no overtaking. \subsubsection{Density profile.} To summarise, we have solved the bulk equations in the domain wall approximation, giving the solution in the form \eref{eq:bd:lengthAssumption}, which with \eref{eq:lengthDist} and \eref{eq:bd:Pjam} results in \begin{equation} P(n,k) = \left(1-\frac{\lambda}{\mu}\right) \left(1-\frac{p\alpha}{\lambda}\right)\frac{1-\alpha}{p-\lambda} \left(\frac{\lambda}{\mu}\right)^n \left(p \left(\frac{p \alpha}{\la}\right)^k - \la \alpha^k\right). \label{eq:bd:explicitPnk} \end{equation} In general this solution does not satisfy the boundary equations for $k=n, n-1$, i.e. the cases we neglected were where the jam extends to the length of the queue. The jam grows with rate $p \alpha$, so the constraint $p \alpha < \la$ (equation \eref{eq:bd:jamConstraint} arising from the normalisation condition) requires that the queue length grows faster on average then the jam. Since we have totally neglected the boundary equations, we expect our approximation to be best when $p\alpha \ll \la$. In fact when $p \to 0$ \eref{eq:bd:explicitPnk} reduces to \begin{equation} \lim_{p \to 0} P(n, k) = \left(1 - \frac{\la}{\mu}\right)\left(\frac{\la}{\mu}\right)^n \frac{\la_2 \la_1^k}{\la^{k+1}}, \end{equation} as expected from the exact solution in this case, \eref{eq:pZeroExact}. The density at site $i$ in a length $n$ queue, computed from \eref{eq:bd:explicitPnk}, \eref{eq:bd:PAll1s} is \begin{eqnarray} \corr{\tau_i}_n &= \alpha \sum_{k=0}^{i-2}P(n,k) + \sum_{k=i}^{n-1}P(n,k) + P(1^n) \nonumber \\ &= \alpha \sum_{k=0}^{i-2}P(n,k) + \sum_{k=i}^{\infty}P(n,k) \nonumber \\ &= P_n \left( \alpha + (1 - \alpha) \left( \frac{p \alpha}{\la} \right)^i \right). \label{eq:bd:dwdensity} \end{eqnarray} Figure~\ref{fig:lengthdep_density} shows length dependent density profiles for $p \alpha < \la$ (Figure~\ref{fig:lengthdep_dw}) and $p \alpha > \la$ (Figure~\ref{fig:lengthdep_nondw}), scaled by dividing out the length distribution $P_n$. \begin{figure}[ht] \centering \subfigure[ $\la_1 = 0.7, \la_2 = 0.1, \mu = 1, p = 0.5$, giving $p \alpha < \la$. ] { \includegraphics[width=0.4\textwidth]{boundeddw_bylength_high_density} \label{fig:lengthdep_dw} } \qquad \subfigure[ $\la_1 = 0.7, \la_2 = 0.1, \mu = 1, p = 3$, giving $p \alpha > \la$. ] { \includegraphics[width=0.4\textwidth]{bounded_nondw_bylength_high_density} \label{fig:lengthdep_nondw} } \caption{ Scaled density profiles $\corr{\tau_i}_n / P_n$ for $n = 2, 4, 6, 10, 14$. Triangle markers for simulation results with points for each length $n$ connected by dashed lines. The analytical expression, calculated from \eref{eq:bd:dwdensity}, is plotted as the solid curve in (a). } \label{fig:lengthdep_density} \end{figure} Triangle markers show simulation results, with points for each length connected by dashed lines. In Figure~\ref{fig:lengthdep_dw}, the solid curve shows $\corr{\tau_i}_n/P_n$ calculated from \eref{eq:bd:dwdensity} and \eref{eq:lengthDist}. We see that as $n$ increases, the simulation results converge to the calculated profile, and even for $n = 10$ the match is very good. In Figure~\ref{fig:lengthdep_nondw}, where the solution \eref{eq:bd:dwdensity} no longer applies, the profiles become almost linear and are reminiscent of the behaviour of the TASEP on the coexistence line. In that phase, the domain wall occurs at all positions with equal probability, resulting, on average, in a linear profile \cite{SchutzD93}. In the bounded PEP the evidence suggests that the constraint $p \alpha < \la$ marks a crossover between the localised jam, for $p \alpha < \la$ and the delocalised jam for $p \alpha > \la$. \subsection{Aggregate density profile and current} In this section we take a different approach, working with the rate equations for the one-point functions. We start by defining a conserved current for the bounded phase. To do so, we sum the density at each position over all lengths, thus aggregating the effect of the length fluctuations. Define the summed one-point functions \begin{equation} \corrbar{\tau_i} = \sum_{n=i}^\infty \corr{\tau_i}_n. \label{eq:bd:densitySummed} \end{equation} Note that as $\corr{\tau_i}_n \le P_n$, \begin{equation} \corrbar{\tau_i} \le \sum_{n=i}^\infty P_n = \left(\frac{\la}{\mu}\right)^i, \label{eq:bd:sum_converges} \end{equation} so the sum is bounded, and it converges as $\sum_{n=i}^M \corr{\tau_i}_n$ is monotone increasing in $M$. Thus $\corrbar{\tau_i}$ is well defined. Higher order summed correlations are defined similarly, e.g. \begin{equation} \corrbar{\tau_{i+1}(1 - \tau_i)} = \sum_{n=i}^\infty \corr{\tau_{i+1}(1 - \tau_i)}_n. \end{equation} Summing the rate equations \eref{eq:rate11} -- \eref{eq:ratein} gives \begin{eqnarray} \fl \frac{\dd}{\dd t} \corrbar{\tau_1} = \la_1 P_0 + \mu \corrbar{\tau_2} + p \corrbar{\tau_2(1 - \tau_1)} - \mu \corrbar{\tau_1}, \label{eq:bd:summedRate1} \\ \fl \frac{\dd}{\dd t} \corrbar{\tau_i} = \la_1 P_{i-1} + \mu \corrbar{\tau_{i+1}} + p \corrbar{\tau_{i+1}(1 - \tau_i)} -p \corrbar{\tau_i(1 - \tau_{i-1}} - \mu \corrbar{\tau_i}, \qquad i > 1. \label{eq:bd:summedRatei} \end{eqnarray} These can be written as a conservation equation for three currents \begin{equation} \frac{\dd}{\dd t} \corrbar{\tau_i} = J_{\text{ext}}^{(i)} + \overline{J}^{(i+1,i)} - \overline{J}^{(i,i-1)}, \qquad i \ge 1, \label{eq:bd:3Current} \end{equation} where \begin{eqnarray} J_{\text{ext}}^{(i)} = \la_1 P_{i - 1} \label{eq:bd:Jext}\\ \overline{J}^{(1,0)} = \mu \corrbar{\tau_1} \label{eq:bd:Jbar1} \\ \overline{J}^{(i,i-1)} = \mu \corrbar{\tau_i} + p \corrbar{\tau_i(1-\tau_{i-1})}, \qquad i \ge 2. \label{eq:bd:Jbari} \end{eqnarray} $\overline{J}^{(i,i-1)}$ is the site-to-site current with a hopping term and frame current term. But customers can also step directly into place at the end of the queue, which gives the external current $J_{\text{ext}}^{(i)}$. In the stationary distribution the time derivatives are zero, and so \eref{eq:bd:3Current} gives a recurrence for the site-to-site current \begin{equation} \overline{J}^{(i + 1,i)} = \overline{J}^{(i,i-1)} - J_{\text{ext}}^{(i)}, \end{equation} which reduces to \begin{equation} \overline{J}^{(i + 1,i)} = \overline{J}^{(1,0)} - \la_1 \sum_{n=0}^{i-1} P_n. \label{eq:bd:JBarRecReduced} \end{equation} We have $\lim_{i \to \infty} \overline{J}^{(i+1,i)} = 0$, as both terms on the right hand side of \eref{eq:bd:Jbari} go to zero. Therefore \eref{eq:bd:JBarRecReduced} implies that \begin{equation} \overline{J}^{(1,0)} = \la_1 \sum_{n=0}^\infty P_n = \la_1, \end{equation} and \begin{equation} \overline{J}^{(i,i-1)} = \la_1 \left(1 - \sum_{n=0}^{i-2} P_n \right) = \la_1 \left(\frac{\la}{\mu}\right)^{i-1}, \qquad i \ge 1. \end{equation} Now \eref{eq:bd:Jbar1} tells us that \begin{equation} \corrbar{\tau_1} = \frac{\lambda_1}{\mu}. \label{eq:bd:aggcorr1exact} \end{equation} This result is exact -- it is the probability that the lattice is at least length one with a particle in site $1$. We could have derived this result directly from the notion of the PEP as a queue: high priority customers leave the system at an average rate $\mu \corrbar{\tau_1}$. The rate at which high priority customers leave the system cannot be higher than $\la_1$, the rate they arrive. But as the service capacity exceeds the total arrival rate ($\la = \la_1 + \la_2$) there is no bottleneck at the server, so high priority customers\footnote{A corresponding argument applies to low priority customers.} leave at the rate they arrive, that is $\mu \corrbar{\tau_1} = \la_1$. The presence of two-point correlations prevent us from calculating exact densities for $i = 2, 3$, etc. The mean field method \cite{DerridaDM92,SugdenE07} is the standard way to deal with this, assuming the correlation between neighbouring sites is small so that the two-point correlations can be approximated as products of the one-point functions. But for the PEP, there is a similarity to the unbounded system, which we can exploit to solve the one-point rate equations. By the definition \eref{eq:bd:densitySummed}, $\corrbar{\tau_i}$ is the probability \begin{equation} \corrbar{\tau_i} = P(\tau_i = 1, \text{length } n \ge i). \end{equation} We can instead work with the conditional probability \begin{equation} \corr{\tau_i | n \ge i} = P(\tau_i = 1 | \text{length } n \ge i); \end{equation} the two are related by \begin{equation} \corrbar{\tau_i} = P(\text{length } n \ge i)\corr{\tau_i | n \ge i} = \left( \frac{\la}{\mu} \right)^i \corr{\tau_i | n \ge i}. \label{eq:bd:aggregated_conditional} \end{equation} Similarly, define $\corr{\tau_i(1 - \tau_{i-1}) | n \ge i}$ through \begin{equation} \corrbar{\tau_i (1 - \tau_{i-1})} = \left( \frac{\la}{\mu} \right)^i \corr{\tau_i (1 - \tau_i)| n \ge i}. \label{eq:bd:aggregated_conditional_twopoint} \end{equation} Substituting into \eref{eq:bd:Jbar1}, \eref{eq:bd:Jbari} gives \begin{eqnarray} \frac{\la_1 \mu}{\la} & = \mu \corr{\tau_1 | n \ge 1} \nonumber \\ & = \mu \corr{\tau_i | n \ge i} + p \corr{\tau_i(1 - \tau_{i-1}) | n \ge i}, \qquad i \ge 2. \label{eq:bd:effJ_eqns} \end{eqnarray} These have the same form as the current equations for the unbounded queue, \eref{eq:ub:J_eqns}, with an effective current \begin{equation} \widetilde{J} = \frac{\la_1 \mu}{\la}, \end{equation} so are solved by the unbounded queue one- and two-point functions \eref{eq:ub:onePoint}, \eref{eq:ub:twoPoint}. That is \begin{equation} \corr{\tau_i | n \ge i} = \alpha + (1 - \alpha)\left(\frac{p \alpha}{\mu}\right)^i, \qquad i \ge 1, \label{eq:bd:corr_conditional} \end{equation} and \begin{equation} \corr{\tau_i(1-\tau_{i-1}) | n \ge i} = \alpha(1 - \alpha)\left(1 - \left(\frac{p \alpha}{\mu}\right)^{i-1}\right), \qquad i \ge 2. \label{eq:bd:two_corr_conditional} \end{equation} To determine $\alpha$, we substitute \eref{eq:bd:corr_conditional}, \eref{eq:bd:two_corr_conditional} into \eref{eq:bd:effJ_eqns}, and take $i \to \infty$ (assuming $p \alpha < \mu$). This gives back the quadratic for $\alpha$ \eref{eq:bd:alphaQuadratic}, so we again must take $\alpha = \alpha_-$, given by \eref{eq:bd:alphapm}. Note that $p \alpha < \mu$ if $p, \mu, \la_2 > 0$, so the density profile is always exponentially decaying. This solution satisfies the current equation \eref{eq:bd:effJ_eqns}, and therefore the resulting aggregated one- and two-point functions \eref{eq:bd:aggregated_conditional}, \eref{eq:bd:aggregated_conditional_twopoint} satisfy the one-point rate equations \eref{eq:bd:summedRate1}, \eref{eq:bd:summedRatei}. But as it does not in general satisfy the higher order rate equations it is approximate only. Figure~\ref{fig:density_aggregated} compares simulated and calculated aggregated density profiles, $\corrbar{\tau_i}$. We have divided out the length dependence factor $(\la/\mu)^i$ so in fact are plotting $\corr{\tau_i | n \ge i}$ computed from \eref{eq:bd:corr_conditional}. \begin{figure}[ht] \centering \subfigure[$\la_1 = 0.1, \la_2 = 0.7, \mu = 1$.] { \includegraphics[width=0.4\textwidth]{bounded_summed_low_density} \label{fig:bounded_summed_low_density} } \qquad \subfigure[$\la_1 = 0.1, \la_2 = 0.3, \mu = 1$.] { \includegraphics[width=0.4\textwidth]{bounded_summed_short} \label{fig:bounded_summed_short} } \subfigure[$\la_1 = 0.7, \la_2 = 0.1, \mu = 1$.] { \includegraphics{bounded_summed_high_density} \label{fig:bounded_summed_high_density} } \qquad \subfigure[$\la_1 = 0.4, \la_2 = 0.1, \mu = 1$.] { \includegraphics{bounded_summed_nodw} \label{fig:bounded_summed_nodw} } \caption{ Aggregated density profiles scaled by dividing out $(\la/\mu)^i$. Triangle markers for simulation results, plotted against calculated profile $\corr{\tau_i | n \ge i}$. For $p = 0.1$ (black), $p = 1$ (mid-gray), $p = 5$ (light gray). } \label{fig:density_aggregated} \end{figure} We see that at $i = 1$, where the exact value $\corr{\tau_1 | n \ge 1}$ is known, and asymptotically for large $i$, the simulated and calculated density profiles agree. At intermediate values of $i$ we see the greatest discrepancy, indicating that the aggregated domain wall solution is approximate only, although the agreement is still very good. Note we could also compare the aggregated profiles with $\corr{\tau_i}_n$ \eref{eq:bd:dwdensity} summed over $n$. However, even at position $1$ the summed $\corr{\tau_1}_n$ does not agree with $\corrbar{\tau_1}$ \eref{eq:bd:aggcorr1exact} for which the exact result is known. The direct application of the domain wall ansatz in Section~\ref{sec:boundeddw} gave an indication of the length dependence in the system, but the approach in this section, following from the current conservation equation, is in much better agreement numerically with simulation results. \subsection{Waiting times} We can use the aggregated one-point functions to compute the average number of customers in the queue, and in turn the average waiting times for both classes of customers. Switching the order of the sums in \eref{eq:pepNhigh}, we can write $\overline{N}_1$, the average number of high priority customers, as \begin{equation} \overline{N}_1 = \sum_{i=1}^\infty \sum_{n=i}^\infty \corr{\tau_i}_n = \sum_{i=1}^\infty \corrbar{\tau_i}. \label{eq:bd:Nhi} \end{equation} Substituting \eref{eq:bd:aggregated_conditional}, \eref{eq:bd:corr_conditional} into \eref{eq:bd:Nhi}, Little's result \eref{eq:littles} gives the average high priority waiting time, \begin{equation} \overline{W}_1 = \frac{1}{\la_1} \overline{N}_1 = \frac{1}{\la_1} \left( \alpha \frac{\la}{\mu - \la} + (1 - \alpha) \frac{p \alpha \la}{\mu^2 - p \alpha \la} \right). \label{eq:bd:W1} \end{equation} With \eref{eq:pepNlow}, the average low priority waiting time is \begin{equation} \overline{W}_2 = \frac{1}{\la_2} \left(\corr{n} - \overline{N}_1 \right) \nonumber = \frac{1}{\la_2}(1 - \alpha) \left( \frac{\la}{\mu - \la} - \frac{p \alpha \la}{\mu^2 - p \alpha \la} \right). \label{eq:bd:W2} \end{equation} Though \eref{eq:bd:W1}, \eref{eq:bd:W2} come from an approximate solution, in the $p \to 0$ and $p \to \infty$ limits they give the correct waiting times. Taking first the limit $p \to 0$, we find \begin{equation} \lim_{p \to 0} \overline{W}_1 = \lim_{p \to 0} \overline{W}_2 = \frac{1}{\mu - \la}. \label{eq:bd:waitPZero} \end{equation} With $p = 0$, high and low priority customers are treated identically. The PEP reduces to a $M/M/1$ queue with arrival rate $\la$ and service rate $\mu$, for which the average waiting time is as given by \eref{eq:bd:waitPZero}. Conversely, if we make the overtake rate infinite, then high priority customers arriving at the queue will immediately overtake any waiting low priority customers. In this limit, high priority customers see an $M/M/1$ queue with arrival rate $\la_1$ and service rate $\mu$, and indeed we find \begin{equation} \lim_{p \to \infty} \overline{W}_1 = \frac{1}{\mu - \la_1}. \end{equation} The average waiting time for low priority customers is \begin{equation} \lim_{p \to \infty} \overline{W}_2 = \frac{1}{(1 - \la / \mu)(\mu - \la_1)}. \end{equation} This can be found by directly taking the limit, or via the requirement that $\overline{N}_1 + \overline{N}_2 = \corr{n}$. Figure~\ref{fig:waiting_times} shows $\overline{W}_1$, $\overline{W}_2$ plotted as a function $p$. Again we see good agreement between simulation results and the calculated values. \begin{figure}[ht] \centering \subfigure[$\la_1 = 0.1, \la_2 = 0.7, \mu = 1$] { \includegraphics{waiting_times_low_density} \label{fig:waiting_times_low_density} } \qquad \subfigure[$\la_1 = 0.1, \la_2 = 0.3, \mu = 1$] { \includegraphics{waiting_times_short} \label{fig:waiting_times_short} } \subfigure[$\la_1 = 0.7, \la_2 = 0.1, \mu = 1$] { \includegraphics{waiting_times_high_density} \label{fig:waiting_times_high_density} } \qquad \subfigure[$\la_1 = 0.4, \la_2 = 0.1, \mu = 1$] { \includegraphics{waiting_times_nodw} \label{fig:waiting_times_nodw} } \caption{ Average waiting time for high priority customers (black) and low priority customers (gray) plotted against overtake rate $p$. Dashed lines show the asymptotic values. } \label{fig:waiting_times} \end{figure} Increasing $p$ interpolates between a first come first served queue ($p = 0$) and strict prioritisation according to customer class ($p \to \infty$). In designing a queueing system, one would choose $p$ to give the desired high priority waiting time, within the constraints imposed by the asymptotic limits. The greatest response in $\overline{W}_1$, defined as the maximum value of $|\dd \overline{W}_1 / \dd p|$, occurs at $p = 0$ in Figure~\ref{fig:waiting_times_low_density}, and so there is a strong relative benefit to high priority customers using even small values of $p$. In Figure~\ref{fig:waiting_times_high_density} the value of the parameters give rise to an inflection point at $p > 0$, and hence the largest response in $\overline{W}_1$ occurs at some positive value of $p$. A sufficient condition for such an inflection point to occur\footnote{This is simpler than trying to solve $\dd \overline{W}_1^2 / \dd p^2 = 0$.} is $\dd^2 \overline{W}_1 / \dd p^2 |_{p=0} < 0$, which happens if and only if \begin{equation} \frac{\la_2}{\la_1} < \frac{\la_1 + \la_2}{\mu}. \end{equation} Then $\dd^2 \overline{W}_1 / \dd p^2$ must change sign as $\lim_{p \to \infty} \dd^2 \overline{W}_1 / \dd p^2 > 0$. For these values of the parameters the benefit to high priority customers of switching on $p$ is relatively small compared to the penalty for low priority customers. \section{Conclusion} In this paper we introduce the prioritising exclusion process: a priority queueing model in which high priority customers are allowed to push ahead in the queue, and thus gain their advantage. The PEP is the exclusion process analog of a well studied priority queueing model, the APQ, a connection which is interesting in itself. But the PEP also has a natural domain wall structure, which allows domain wall dynamics to be derived from the microscopic transition rules. This has recently been achieved for a TASEP with deterministic bulk motion and stochastic boundary conditions \cite{CividiniHA14}. In contrast, the PEP is fully stochastic, but the unique boundary conditions result in the regular behaviour of the jam of high priority customers, the key descriptor in our domain wall model. The PEP exhibits a phase transition from a phase with finite expected lattice length, to one with an unbounded lattice length. In the unbounded phase, we find the exact solution of the domain wall equations in the $n \to \infty$ limit, which reveals a further subdivision of this phase into phases with finite or infinite jam length. We find the condition for an infinite jam, in which case low priority customers will, with probability one, never get served. When the jam remains finite we calculate exact stationary density profiles in appropriately defined local reference frames. In the bounded phase, domain wall theory does not give exact results but leads to two complementary approximate solutions. From a direct application of the domain wall ansatz we find that the shape of the density profile can again be understood in terms of a jam, in this case either localised at the service end or able to grow and fill the lattice. In a second approach, a current conservation equation implies that domain wall theory can be naturally applied to aggregate densities. We thus give very good approximations for these observables and consequently accurately estimate average customer waiting times. We also give the condition for which the stochastic overtaking of low priority customers is most effective as a scheduling mechanism. The behaviour of the jam of high priority customers plays a key role in understanding all phases of the PEP. There is an interesting analog in the APQ in the notion of \textit{accredited} customers: class $1$ customers with accumulated priority greater than the maximum possible priority of any class $2$ customer \cite{StanfordTZ14}. This differs from the definition of a jam, since the last accredited customer may be followed by an unaccredited class $1$ customer, whereas in the PEP, a jam is always terminated by a class 2 (low priority) customer. Understanding this connection may allow us to compute complete waiting time distributions for the PEP, as has already been done for the APQ \cite{StanfordTZ14}. \section*{Acknowledgement} We thank Peter Taylor for suggesting the PEP to us as well as for discussions, and Alexandre Lazarescu, Chikashi Arita, Guy Latouce, and Kirone Mallick for discussions and encouragement. We are grateful to the Australian Research Council and the ARC Centre of Excellence for Mathematical Frontiers (ACEMS) for financial support. \section*{References}

1. Introduction {#sec1-high-throughput-09-00008} =============== In the era of precision medicine, the identification of germline variants related to the inter-individual variability observed in response to the same drug represents a great opportunity for evidence-based drug prescription. Individual drug response is influenced by multiple and highly variable factors which impact on the pharmacokinetic (PK) and/or pharmacodynamic (PD). In fact, both physiological and pathological conditions (aging, kidney and liver function, comorbidities, environmental conditions) together with genetic background may interfere with drug efficacy and/or toxicity. Although many studies have identified germline polymorphic variants in genes involved in Adsorption, Distribution, Metabolism and Excretion (ADME) of drugs in relation to inter-individual variability, their identification as part of common clinical practice remains an active challenge. In the post genomic era the terms pharmacogenetics and pharmacogenomics are interchangeable and are here referred to as PGx. The objectives of PGx include safer prescriptions, powerful and appropriate drugs, reduction of healthcare costs and treatment toxicities. In the field of oncology, the role of biomarkers for the development of precision medicine provides a strategic opportunity for technological developments to improve human health and reduce health-care costs. While oncogenic somatic alterations are well known to influence drug response, germline mutations, such as those in *BRCA*, *PALB2*, *ATM*, and *CHEK2* genes, were thought to simply determine individual predisposition to develop breast, ovarian and other cancers \[[@B1-high-throughput-09-00008],[@B2-high-throughput-09-00008],[@B3-high-throughput-09-00008]\]. However, among germline alterations, the polymorphic variants in genes encoding for drug-metabolizing enzymes can be used as predictive biomarkers for drug efficacy and drug-induced toxicity \[[@B4-high-throughput-09-00008]\]. Therefore, germline and somatic variants are both potential PGx predictive biomarkers for drug targets, drug related toxicity and inter-individual variability in drug response, in the perspective of precision medicine. In PGx studies, the genetic variants evaluated are Single Nucleotide Polymorphisms (SNPs), nucleotide insertions, deletions, short tandem repeats, copy number variations (CNV) and chromosomal translocations \[[@B5-high-throughput-09-00008]\]. SNPs are common inherited variations (90%) that exert a biological role only when they occur within a regulatory region or in a gene coding sequence. SNPs are inherited within haplotype blocks and exist in strong linkage disequilibrium (LD) with a specific genetic variant. In other words, the identification of a specific SNP allows one to assume the presence of the co-inherited genetic variant. For this reason, SNPs can be used as tag markers for unseen causative alleles. The closely linked SNP alleles are in blocks, separated by regions of high recombination (hotspots), in which there are few or many polymorphic variants which also could be associated with a disease or drug-response phenotype \[[@B6-high-throughput-09-00008]\]. Keeping this in mind, the genotype-phenotype correlation assumes high importance in PGx studies and in determining the individual response to pharmacological therapy. The identification of polymorphic variants with impact on phenotype has evolved through three approaches: (1) the candidate gene approach, where a small number of well-known PK or PD-related markers are tested in a small sample size; (2) the genome wide association study (GWAS) approach, a hypothesis-free method performed on large populations where high numbers of markers are simultaneously tested, identifying only common variants but with the need of stringent statistical correction \[[@B7-high-throughput-09-00008]\]; (3) the pre-defined SNPs panel approach which includes only thousands of candidate SNPs relevant pharmacogenes with putative importance. Advancement in technologies, such as next generation sequencing (NGS), contemplate a diagnostic tool where whole genome or exome sequences are interrogated as comprehensive PGx genotyping tool in a rapid and large-scale DNA sequencing technology \[[@B8-high-throughput-09-00008]\]. Therefore, there has been a revolution in tumor treatment prescription and in drug's labeling PGx recommendation towards a more tailored therapy. The candidate gene approach has the limitation of high rate of false positives, replication of results and overestimation of effect size, while the multiple comparisons correction used in GWAS leads to false negatives. In addition, GWAS arrays have a good coverage of PK genes but include a limited number of known PD genes. The SNP panel represents an ideal compromise between the other two approaches through the simultaneous genotyping of SNPs with known relevance in PK-PD and a limited need for statistical stringency for multiple comparisons \[[@B9-high-throughput-09-00008]\]. Users have the option to customize a SNP-panel choosing the candidate genes and SNPs that suite best their needs. The DMET^TM^ platform represents an example of the SNPs panel widely used to discover biomarkers correlated to efficacy or toxicity in common and rare diseases \[[@B10-high-throughput-09-00008],[@B11-high-throughput-09-00008],[@B12-high-throughput-09-00008],[@B13-high-throughput-09-00008],[@B14-high-throughput-09-00008],[@B15-high-throughput-09-00008]\]. The difficulty in analyzing the mole of information generated by DMET^TM^ platform led to the development and implementation of algorithms for statistical as well as data mining analysis. Moreover, in order to integrate and validate big-data derived from omics technologies, high-performance software tools, that are able to deal with these vast amounts of data, are necessary. Software tools allow efficient handling of omics data to validate the explorative biomarkers identified by DMET assay and to correlate them with drug efficacy, toxicity and/or cancer susceptibility. These analyses should lead to identifying potentially predictive biomarkers to be confirmed in well-designed clinical trials for translating purposes. In this review, we aimed to describe software tools developed specifically for DMET genotyping analysis. In addition, we describe how to employ the available software tools in order to integrate information derived from analysis of matched data (disease, treatment, stadiation, race). The latter might represent a new approach for biomarkers validation. 2. DMET^TM^ Genotyping Platform {#sec2-high-throughput-09-00008} =============================== The DMET™ Plus array (Thermo Fisher Scientific, Inc., Waltham, MA, USA) is a PGx high-throughput genotyping assay, based on microarray technology. It allows to analyze comprehensively, by a single multiplexed assay, 1931 SNPs and five copy number variations (CNVs) located in 231 genes encoding phase I and II metabolism enzymes, drug transporters, drug targets and modifiers, approved by Food and Drug Administration (FDA) as involved in drugs and carcinogenic metabolism ([Supplementary Table S1](#app1-high-throughput-09-00008){ref-type="app"}) \[[@B16-high-throughput-09-00008]\]. The DMET^TM^ platform is recognized for research application only. The assay starts from 1 μg of genomic DNA extracted from peripheral blood, formalin fixed-paraffin-embedded tissue, or buccal swab. The DMET^TM^ Plus assay uses Molecular Inversion Probe (MIP) technology \[[@B17-high-throughput-09-00008]\] to amplify the sequence-specific targets for each 1936 marker. MIPs are single-stranded oligonucleotides consisting of a common linker containing universal PCR primer binding sites flanked by target-specific probes. The assay probe pool contains one or more MIPs for each of the genotyping markers. The target-specific probes simultaneously hybridize to the same DNA fragment, forming a circular structure with the intended target captured between the probes (annealing); addition of polymerase and ligase results in gap filling and completion of the circular form by incorporation of the intervening target sequence (gap fill and ligation). Library formation is performed by cleaving the circular form (exonuclease digestion and inversion), followed by standard multiplex PCR. PCR products were than fragmented, labeled with biotin and hybridized overnight (16--18 h) to the array. Finally, arrays were scanned to generate the file containing the intensity signal for all probes. After normalization of raw signal values, genotypes are generated using the Affymetrix® DMET Console software (Version 1.3, Thermo Fisher Scientific, Inc., Waltham, MA, USA) as single-sample genotyping and converted to standardized name (star nomenclature) to track known polymorphic variants. Genotypes are determined for each SNP site and reported as homozygous wild-type, heterozygous, homozygous variant, 'no call' or Possible Rare Allele (PRA) in the case of lack of genotype call. The reproducibility of genotyping results with Sanger sequencing or TaqMan SNP genotyping assays is 100% \[[@B18-high-throughput-09-00008]\]. The DMET^TM^ Console generates CEL files containing the raw data produced by chip-array scans. The CEL files are translated in CHP files and after are exported as xlsx (excel) or plain text files, formats suitable for statistical or data mining analysis. The analysis of microarray data is conducted through four steps: (a) preprocessing, which allows background correction, summarization and normalization; (b) annotation (process which associates to each gene a set of functional information) and translation (process which converts the genotype calls (reported in CHP files) to tracked functional allele calls); (c) statistical/data mining analysis; and (d) biological interpretation. The conversion of intensity value in actionable knowledge can be conducted by DMET^TM^ Console, apt-DMET-genotype and DMET Analyzer. Affymetrix® DMET Console and apt-DMET-genotype allow the summarization/normalization of CEL files, the annotation and translation of genotype calls, then the CHP file and ARR sample files are merged and annotated using standardized nomenclature. DMET-Analyzer allows the automatic analysis of data provided by Affymetrix® DMET Console \[[@B19-high-throughput-09-00008]\]. A scheme of the DMET analysis workflow is represented in [Figure 1](#high-throughput-09-00008-f001){ref-type="fig"}. 3. Software Tools to Analyze Genotyping Data {#sec3-high-throughput-09-00008} ============================================ Software tools have become needful to support researchers in omics science and to efficiently process the vast amounts of data produced by high-throughput technologies. The continuing improvements in microarrays and NGS mainly due to the continuous reduction of costs, along with the improvement in the number of variants and/or genes that can be investigated in a single experiment, spur for the development of software frameworks able to transform the available amount data in actionable knowledge. The development of appropriate software frameworks and algorithms for the efficient and scalable data analysis will allow researchers to analyze considerable amounts of genotyping data in a short time as well as to provide enhancement in the accuracy of the results due to the capability to investigate the problems from a boarder perspective. Thus the most suitable software tools can contribute to move the first steps towards the so called Predictive, Preventive, Personalized and Participatory Medicine (P4 medicine) \[[@B20-high-throughput-09-00008]\]. Although vendors of genomics platforms provide its proprietary software frameworks with which it is possible only to handle raw data, i.e., convert the intensity data file in gene expressions or SNPs value. Vendors do not offer any software tools capable to extract actionable knowledge from the data. Algorithms and software frameworks able to manage the converted, annotated and summarized genotyping raw data, they can support researcher in the phase of data analysis to identify knowledge buried into the data and not easily available. For example, in a case-control study, to identify the SNPs related to an adverse drug reaction into the population under investigation, without the support of a specific analysis framework, the analysis should be done manually, a long and error prone task considering the huge amount of data that must be analyzed. Instead, the use of specific software frameworks can speed up and simplify the knowledge extraction, since can be accomplished automatically. In this section we present suite of bioinformatic frameworks for the preprocessing and analysis of DMET-SNPs data. The Affymetrix® DMET^TM^ platform allows us to identify genes involved in ADME functions obtained in case-control studies. Below we review the open-source bioinformatics frameworks developed at the Bioinformatics Laboratory of the University of Catanzaro, along with five other tools providing an overview of their main technical and functional features \[[@B21-high-throughput-09-00008]\]. A description of the investigated software tools is presented below. Furthermore, a methodology to integrate genomic data by using GMQL-Web and a case study are provided. - *DMET-Analyzer* \[[@B22-high-throughput-09-00008]\] is a software framework for the automatic association analysis between the variations present in the patients' genome and their clinical conditions, i.e., the different response to drugs. DMET-Analyzer simplifies the identification of relevant variations into the population under investigation through the automation of the whole statistical analysis workflow. Moreover, DMET-Analyzer can annotate SNP data by automatically retrieving the information available in existing SNP databases, e.g., dbSNP, as well as to interpreter the biological process in which variants are involved through the association of SNP within a pathway, by automatically retrieving the information stored in specialized databases, e.g., PharmaGKB. DMET-Analyzer is written in Java making it compatible with Unix/Linux, MacOS and Windows operating systems, it presents a simple graphical user interface that allows non-programming users (e.g., doctors and biologists) to analyze DMET files interactively produced using the Affymetrix DMET-Console. Moreover, DMET-Analyzer implements the FDR and Bonferroni statistical correctors, the Odds-Ratio and HardyWeinberg equilibrium calculator. DMET-Analyzer can analyze only DMET files in xlsx (i.e., excels files) or plain tab delimited format (i.e., plain text files). Results can be easily saved by clicking on it and saved in textual format (txt), markup language (html) and so on. DMET-Analyzer is freely available under the GNU General Public License version 2.0 \[[@B23-high-throughput-09-00008]\]. - *DMET-Miner* \[[@B24-high-throughput-09-00008]\] is a software tool for mining association rules from DMET SNP datasets. DMET-Miner through the association rules can correlate the presence of multiple allelic variants with the clinical condition of the patients. Allowing the users to overcome the limitation of the univariate statistics implemented in DMET-Analyzer that can extract associations among a single allelic variant and the clinical conditions of samples. For example, the most frequent association among alleles responsible for the different response to a treatment. DMET-Miner enables users to automatically mine association rules from a whole DMET datasets, conversely from other available tool e.g., Weka requiring to the user to preprocess the input file in order to handle missing value, trivial data and so on. DMET-Miner is written in Java, making it available for all the operating systems compatible with Java. It presents a simple graphical user interface, allowing the users to analyze a dataset through some mouse's clicks. DMET-Miner is distributed under Creative Commons license and is freely downloadable for academic and not-for-profit institutions \[[@B25-high-throughput-09-00008]\]. - *OS-Analyzer* \[[@B26-high-throughput-09-00008]\] is a software framework implemented in Java for the analysis of SNP microarray datasets enriched with survival events. OS-Analyzer comes with a simple, effective and intuitive graphical user interface for the automatic computation and visualization of Overall Survival (OS) and Progression Free Survival (PFS) curves of case patients, evaluating their association with ADME gene variants. Moreover, to simplify the researchers work, results according to statistical significance obtained by comparing the area under the ROC (Receiver Characteristic Curve) curves are ranked. Statistical Relevance is computed by using the log-rank test, allowing a quick and easy analysis and visualization of high-throughput data. OS-Analyzer is distributed under Creative Commons License, is freely available for academic and not-for-profit institutions \[[@B27-high-throughput-09-00008]\]. - *Affymetrix® DMET Console* allows to preprocess the raw data file generated by the Affymetrix DMET for building a comprehensive table containing, for each probe and for each sample, the detected SNP or a No call value (i.e., ambiguous nucleotide in the SNP). DMET Console support probe-set summarization of a complete dataset of binary ".CEL" files (containing the probe-level intensities), the management of resulting preprocessed files ".CHP" (containing the gene-level information) and the building of tabular dataset containing the genotype call for all the probesets and all the samples of an experiment. Once the preprocessing phase is completed, the relationship between the detected SNPs and the response to drugs must be tested. Because of DMET Console does not allow this test, in order to discover statistically significant associations, researches have to export and manually process SNPs tables produced by the DMET Console thought the use of external tools (e.g., statistical software). - *Affymetrix Power tools suite* (APT) is a set of command line programs that implement different algorithms for preprocessing Affymetrix microarray data. Two of the most popular programs in APT are: apt-dmet-genotype, for making genotype calls from Affymetrix genotyping arrays, and apt-probeset-summarize for analyzing gene expression arrays. Such as DMET Console, both programs are generally focused on .CEL file analysis. - *The GenoMetric Query Language* (GMQL) \[[@B28-high-throughput-09-00008]\] is a high-level query language, inspired by classic traditions of data-base management, that extends conventional algebraic operations with bioinformatics domain-specific operations specifically designed for genomics; thus, it supports knowledge discovery across thousands or even millions of samples comparing genomic regions on the basis of metric properties but also arbitrary attributes and metadata that concern regions and samples, respectively. In particular, datasets are described by the Genomic Data Model (GDM) \[[@B29-high-throughput-09-00008]\], based on the notion of genomic region, which provides interoperability between several data formats. In addition, GDM combines abstractions for genomic region data with the associated experimental, biological and clinical metadata. GMQL system can be used online through a specific Web interface which provides a user-friendly intuitive environment for bioinformaticians and biologists who need to query genomic processed data (e.g., sourcing from big consortia such as ENCODE, TCGA, GENCODE or RefSeq) as well as to combine them with their private datasets (i.e., datasets created by a specific user as result of their own experiments and studies, e.g., output from Affymetrix® DMET^TM^ Platform). Such environment provides portable and scalable genomic data management on powerful servers and clusters (based on Apache Spark). - *PLINK* \[[@B30-high-throughput-09-00008]\] is a free, open-source tool for GWAS and research in population genetics. PLINK works on five core functional domains: data management, summary statistics, population stratification, association analysis and identity-by-descent estimation. Association tests can be run by PLINK to evaluate case-control data to determine if an SNP has an effect on disease status. PLINK can run either as a stand-alone tool (from the command line or via shell scripting) or in conjunction with gPLINK, a Java-based graphical user interface (GUI) that offers a simple project management framework to track PLINK analyses and facilitates the integration with Haploview. - *Haploview* \[[@B31-high-throughput-09-00008]\] is a comprehensive suite of tools, written in Java language, for analysis and visualization of LD and haplotype maps. Haploview accepts input in a variety of formats and generates marker quality statistics, LD information, haplotype blocks, population haplotype frequencies and single marker association statistics. Haploview is fully compatible with data dumps from the HapMap project. HapMap genome browser allows researchers to explore a particular region of the genome and extract HapMap genotype data for all genotyped markers in the selected region in a format accepted by Haploview. Haploview currently supports visualization and plotting of PLINK whole genome association results \[[@B32-high-throughput-09-00008]\]. The significant improvements because of the use of the microarrays have led to an increase in the number of variants that it is possible to investigate in a single experiment as well as a reduction of the analysis time, allowing to produce considerable amounts of data in ever shorter times. Therefore, it is necessary to develop software tools and algorithms capable of managing ever-increasing volumes of data. The following is a summarization of the parallel and web applications developed to efficiently deal with huge genotyping datasets. - *coreSNP* \[[@B33-high-throughput-09-00008]\] is a tool implemented in Java language, for the parallel pre-processing and statistical analysis of DMET SNP datasets. The scalable implementation is obtained exploiting the multi-threading capabilities of modern CPUs, allowing core SNP to manage huge amount of DMET data. The automatic association analysis between possible genome variations of the patient and the clinical conditions through the well-known Fisher's Test is obtained. Moreover, multiple-statistical correctors i.e., Bonferroni and False Discovery Rate, with which to improve the statistical significance of results, are available. The visualization of the detected SNPs as heatmap plot provides a visual feedback that simplifies the interpretation of the results. - Parallel Association Rules Extractor from SNPs (*PARES*) \[[@B34-high-throughput-09-00008]\] is a multi-thread software tool developed in Java for the parallel extraction of association rules by which to correlate the presence of a multiple allelic variants with the patients' clinical condition, i.e., the most likely set of alleles responsible for the onset of adverse drug reactions. PARES is a multi-thread version based on the optimized version of the Frequent Pattern Growth (FP-Growth) algorithm. PARES encompasses a customized SNP dataset preprocessing approach based on a Fisher's Test Filter to prune trivial transactions, allowing to shrink the search space as well as to reduce the FP-Tree size enabling a better management of the main memory. PARES comes with a simple and intuitive graphic user interface, where specific skills are not necessary to extract multiple relations between genomic factors buried into the datasets. PARES is distributed by BioinfoLabUnicz, under Creative Commons license, is freely downloadable for academic and not-for-profit institutions \[[@B35-high-throughput-09-00008]\]. - *GenotypeAnalytics* \[[@B36-high-throughput-09-00008]\] is a representational state transfer (RESTFul) service by which to mine association rules from SNP-dataset through the use of a common web-browser. GenotypeAnalytics can speed up and simplify the analysis of massive amount of SNPs data, highlighting in a remarkable way only the SNP involved in the development of the disease or in adverse drug reaction. GenotypeAnalytics can support researcher and medical doctors to discern new molecular markers that can be (after further opportune validations) used into the clinical contexts. Possible new molecular markers can be used to perform pathway enrichment to understand which SNPs are responsible for the pathway's anomalies or the SNPs that influence drug responses in subjects with the same pathology following the same therapy. - *Cloud4SNP* \[[@B37-high-throughput-09-00008]\] is a Cloud software tool for the parallel pre-processing and statistical analysis of DMET SNPs data. Cloud4SNP is the Cloud version of DMET-Analyzer \[[@B22-high-throughput-09-00008]\] that has been implemented on the Cloud using the Data Mining Cloud Framework, a software environment for the design and the execution of knowledge discovery workflows on the Cloud. Cloud4SNP is developed in Java and presents a simple graphical user interface that can be accessed by means of a common web-browser. Providing the analysis workflow as a service, it allows the users to upload and analyze the data without to buy expensive hardware or to setup the analysis environment. Cloud4SNP allows one to identify the relevance of the presence of SNPs in one of the two classes of samples using the well-known Fisher test, along with the use of multiple-statistical correctors such as Bonferroni and FDR. The main features of the presented software frameworks are summarized in [Table 1](#high-throughput-09-00008-t001){ref-type="table"}. 4. A New Methodology to Integrate Genomic Public Data with Genotyping Private Data {#sec4-high-throughput-09-00008} ================================================================================== The developments of high-throughput technologies have accelerated the accumulation of massive amounts of omics data from multiple sources. In contrast to the traditional approaches in which data from each source have been analyzed in isolation, integrative analysis of multi-omics and clinical data has become a key for new biomedical discoveries and advancements in precision medicine \[[@B20-high-throughput-09-00008]\]. Indeed, isolated omics studies frequently fall short when identifying the cause of complex diseases such as cancer. Although there has been much technological progress, the integration across data sets and data types remains limited. In this section we illustrate a methodology \[[@B38-high-throughput-09-00008]\], based on the use of GMQL System, for combining private datasets such as DMET Platform \[[@B19-high-throughput-09-00008]\] output with public datasets already available in the GMQL Repository as well as external (i.e., external public dataset can be uploaded following the steps of workflow in [Figure 1](#high-throughput-09-00008-f001){ref-type="fig"} such as for private dataset). A private dataset is an alternative to public dataset and it is created by a specific user through some uploading operations or it is a result of some GMQL queries. The methodology consists of the following main three phases: data preprocessing, data ingestion and query-download results ([Figure 2](#high-throughput-09-00008-f002){ref-type="fig"}). 4.1. Data Preprocessing Stage {#sec4dot1-high-throughput-09-00008} ----------------------------- This stage prepares raw data for further processing and it is necessary since most of the datasets obtained as outputs from various systems are noisy, incomplete, inconsistent and contain outliers. Data preprocessing is a method of resolving such issues. Once preprocessed, the dataset needs to be transformed into a GDM compliant format in order to employ it within GMQL-web. Each GDM dataset is a collection of samples with the same region data schema and with the same type features; each sample, in turn, corresponds to a pair of files which contain: \(i\) region data: describing the physical coordinates of the genomic areas (e.g., the genomic position of the Affymetrix probes) and other optional genomic features (e.g., p-value, peak, q-value); \(ii\) metadata: unstructured attribute-value pairs describing general properties (i.e., biological, clinical, and experimental) of the sample; It must be considered that conversion of genome coordinates between different assemblies is often required for many integrative and comparative studies (e.g., Affymetrix DMET results are aligned to GRCh37/hg19 genome build thus they could need to be remapped to current genome build GRCh38. 4.2. Data Ingestion Stage {#sec4dot2-high-throughput-09-00008} ------------------------- GMQL-Web interface allows one to combine private datasets with public datasets through the "Add/Upload new dataset" feature that allows the user to choose among two options for uploading private datasets files: the *standard file-format mode* allows one to use a number of file standard formats directly supported by the system \[[@B39-high-throughput-09-00008]\] (e.g., BED, NarrowPeak, BroadPeak, VCF, etc);the *custom file-format mode* can be chosen by selecting the "Custom (GTF or tab/delimited)" option and it allows one to use a user-defined format following the guidelines of Gene Transfer Format (GTF) \[[@B40-high-throughput-09-00008]\] or TAB-delimited formats. In this case it is required the definition of an additional Extended Markup Language (XML) format Schema file describing the structure of the dataset to upload. Once uploaded, the private datasets are shown in the "interface datasets viewer" under the "Private" folder and it can be managed independently or in combination with other public datasets. 4.3. Query/Download Results Stage {#sec4dot3-high-throughput-09-00008} --------------------------------- A GMQL query is a sequence of operations, applied on one or more datasets, resulting in the creation of new datasets. A query is expressed as a sequence of GMQL operations, each with the following structure: \<OV\> = OPERATOR (\<params\>) \<IV1\>\<IV2\> Where OV stands for output variable, an operator can be tuned using optional parameters, and IV1 and IV2 are input variables (Data-Driven Genomic Computing \[[@B41-high-throughput-09-00008]\]). Among all GMQL unary and binary operators, we here focus only on three of them, since they are used in the reported following example: SELECT creates a new dataset from an existing one by extracting a subset of samples from the input dataset. Several conditions can be combined using Boolean operators; MATERIALIZE writes the content of a dataset to a file and it registers the saved dataset in the repository to make it usable in other queries; MAP applies to two datasets, a reference and an experiment. For each sample in the experiment dataset, it computes aggregates over the values of the regions that intersect with at least one region, in at least one reference sample. The count aggregate counts the number of experiment regions that intersect a specific reference region. The right side of [Figure 1](#high-throughput-09-00008-f001){ref-type="fig"} shows GMQL queries can be applied to public datasets optionally combined with private ones (derived from the previous preprocessing and ingestion phases). The GMQL Web "Query editor" allows one to write the query that will be later compiled and executed. Once results are produced (i.e., materialized) they become available in the "Datasets viewer" for browsing, further processing or downloading. 5. Case Study {#sec5-high-throughput-09-00008} ============= 5.1. Integration of Private DMET Platform Data with GENCODE Annotation Database {#sec5dot1-high-throughput-09-00008} ------------------------------------------------------------------------------- In this Section we illustrate a representative example of how to applicate our methodology for integrating DMET platform genotype results, with dataset sourcing from GENCODE annotation database \[[@B42-high-throughput-09-00008]\]. The methodology used the workflow shown in [Figure 3](#high-throughput-09-00008-f003){ref-type="fig"} involving DMET platform, GMQL system and, in general, public genomic databases. [Figure 4](#high-throughput-09-00008-f004){ref-type="fig"} focuses on the case study workflow that specifically involved GENCODE as a public genomic database. 5.2. Data Preprocessing Stage {#sec5dot2-high-throughput-09-00008} ----------------------------- DMET-Console output is a tab-delimited file structured as a matrix with 1936 rows (probes) and a number of columns related to the number of subjects enrolled in the analysis. The value contained in the (*i-th,j-th*) cell is the SNP detected in the *i-th* probe and belonging to the *j-th* subject. The tab-delimited file obtained from DMET Console needs be transformed into a GDM compliant format. We located the position of the DMET SNPs within the genomic region by using the DMET \[[@B43-high-throughput-09-00008]\] annotation file which provided the genomic coordinates for each Affymetrix probe. In order to integrate the DMET SNPs output with the GENCODE v27 dataset, already available in the GMQL Repository, the DMET probes genomic coordinates were mapped to current genome build GRCh38 by using the UCSC Lift Genome Annotations tool \[[@B44-high-throughput-09-00008]\]. 5.3. Data Ingestion {#sec5dot3-high-throughput-09-00008} ------------------- As depicted in [Figure 2](#high-throughput-09-00008-f002){ref-type="fig"} (central part) the user can choose among two options for uploading private datasets files. We have chosen the *custom file-format mode*. We created a tab-delimited file containing the DMET dataset where the columns are defined as follows: the first is the Affymetrix Probe identifier (i.e., a unique Affymetrix identifier for the probe set or SNP); the next four contain the values describing the genomic regions (i.e., chromosome, start, stop, strand); the fifth and sixth columns contain respectively the dbSNP RS ID (i.e., RS ID from dbSNP database, e.g., rs11584174, rs2501870, etc.) and the associated identifiers in the PharmGKB database. The remaining columns contain the detected SNPs for each subject. The XML schema file describing the structure of the dataset must be created before loading the DMET dataset. Once the dataset file and its schema are obtained, we uploaded them in GMQL-Web using the feature "Add/Upload". 5.4. Query/Download Results {#sec5dot4-high-throughput-09-00008} --------------------------- Once the DMET output file (containing the SNPs detected by 33 probes on 28 subjects) was uploaded into GMQL Web system, the query in Listing 1 ([Figure 5](#high-throughput-09-00008-f005){ref-type="fig"}) was executed in order to:select SNPs from DMET Dataset;select exon regions from GENCODE annotation dataset;map exon regions on SNPs regions;select only SNPs regions that overlap at least one exon region. The result of the MAP operation (EXON_Map) was a dataset which contained SNPs equipped with the count (SNP_count) of exons overlapping with that nucleotide. Many SNPs did not overlap with exon, thus the SNP_count is zero. The select statement filters out samples containing a region with null count. Results of this query ([Figure 6](#high-throughput-09-00008-f006){ref-type="fig"}) contain SNPs region coordinates (chromosome, start and end positions, strand) and additional characterizing attributes (gene name, dbSNP ID, all SNPs detected by DMET in that region and the number of overlapped exon regions). 6. Conclusions {#sec6-high-throughput-09-00008} ============== Comprehension of the role of common variations in ADME genes has the potential to significantly improve clinical research by predicting the contribution of an individual genetic make up to PK and PD activity. This knowledge may be translated into reduction of ADRs, increase of efficacy as well as healthcare outcomes improvement and economic benefits. Many high-throughput tools are available on the market for the genotyping, at different scale, of SNPs known to be related to drugs and xenobiotics metabolism. Among these the DMET technology has been widely used in clinical research. Understanding how genes are affected by SNPs associated with a pharmacological phenotype is crucial for the prescription of the right drug for the right disease in the era of precision medicine. The advancement in high-throughput technologies has urged to consult and integrate massive amounts of public omics data from different sources as well as to validate private genotyping results. For instance, DMET plus SNPs platform applied to a limited and private sample of patients, allows one to identify genotypes that may support the discovery of biomarkers in the optic of precision medicine. However, genetic associations identified in the initial study need to be confirmed in an additional non overlapping study samples in order to replicate the statistically correlation between the same genetic variants and the trait of interest. Moreover, the biomarker validation process includes two important concepts i) analytical validation and ii) clinical qualification. The first refers to the reanalysis of all, or a subset of genetic variants investigated in the initial study, and includes the process to assess the accuracy, the robustness and reproducibility of the assay. The clinical qualification regards the correlation of the identified biomarker with biologic processes and clinical endpoints. The clinical translation of a biomarker is regulated in the European Community (CE) and United States (US) by different steps to achieve the inclusion in clinical practice guidelines \[[@B45-high-throughput-09-00008]\] as FDA-cleared or CE-IVD marked clinical diagnostic tests. In this scenario, annotated SNP effects need to be interpreted in the broader context of genes, LD and molecular pathways, and so functional information could be integrated from different repositories. The integration of genotype data with biological and clinical information, could be possible using public databases including gene expression, methylation and sequencing data. The functional validation of the impact of a genetic variant, hypothesized through the integration of GWAS/exome, RNA-seq, and protein data, will contribute to reveal further translational and protein structural effects of potential biomarkers. In this work we described all the algorithms developed up to now to analyze genotyping data, generated through DMET Console software and GMQL System, with the aim to integrate private DMET Platform results, across data sets and data types output, with public datasets already available in the GMQL Repository and external public dataset. In our model of analysis, the opportunity to insert specific queries to interrogate a public dataset with the aim to identify predictive biomarkers of drug efficacy/toxicity or disease susceptibility, allows us to characterize also the role of tag SNPs providing useful confirmation of hypothesized models for gene and genome dynamics. The integration of genotype data with biological and clinical information through the approaches described is important for the understanding, expansion and validation of the knowledge on the SNP-association and its clinical translation. The identification of predictive biomarkers is the major goal of biomedical research. However, the ability to consult and integrate large-scale molecular data have not been largely applied, precluding further advances in precision medicine. Additional efforts are needed to translate into clinical practice the promise of genomic and other molecular data. The following are available online at <https://www.mdpi.com/2571-5135/9/2/8/s1>. ###### Click here for additional data file. This research received no external funding. The authors declare no conflict of interest. {#high-throughput-09-00008-f001} {#high-throughput-09-00008-f002} {#high-throughput-09-00008-f003} {#high-throughput-09-00008-f004} {#high-throughput-09-00008-f005} {#high-throughput-09-00008-f006} high-throughput-09-00008-t001_Table 1 ###### Main features of the presented software frameworks. ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Software Availability Main Features IA Interface OS DM Statistical Prep CS Weaknesses Strengths ------------------------------------ -------------- --------------------------------------------------------------------------------- --------------------------------------------------- ----------------------- ---------------------- ----- ------------- ------ ---- ------------------------------------------------------------------------- ------------------------------------------------------------------------------------- DMET-Analyzer Free Analysis of Variants\ Annotations and direct link to external databases Command line, GUI Platform-independent NO YES YES L Single allelic\ It automatically analyses data in\ based\ variants discovery;\ case-control association studies;\ on Statistical test Unable to dig with\ Affymetrix raw files; DMET-Miner Free Analysis of Variants\ Annotations and direct link to external databases GUI Platform-independent YES NO YES L Manage small Dataset Multiple allelic variants discovery using Association Rules mining OS-Analyzer Free computation and visualization of OS and PFS curves Integration of genomic and clinical data GUI Platform-independent NO YES YES L It cannot analyse gene expression\ High performance with respect to other statistical tools;\ data in order to plot OS curves (3) Automatic analysis of whole DMET SNPs Dataset Affymetrix DMET-Console Free\* Preprocessing of raw data generated by DMET NO GUI Windows NO NO YES L Lacks in the possibility of doing statistical and data mining analysis; It allows only the preprocessing of binary data Affymetrix Power tools suite (APT) Free\* They implement algorithms for analyzing and working with Affymetrix Microarrays YES command line Platform-independent NO NO YES M Lacks in the possibility of doing statistical and data mining analysis Analysis of intensity microarray data to produce final tabular\ dataset The GenoMetric Query Language\ Free Query, Download and Integrate\ Integrate Public with Private Genomic Datasets Web interface Platform-independent NO NO NO M Allows to perform only genometric queries It combines private dataset with publicly available datasets (GMQL) Public with Private Genomic Datasets PLINK Free Analysis of genotype/phenotype data YES command line, GUI (5) Platform-independent NO YES YES H Allows to perform only statistical analysis Data management, statistical analysis, association analysis of whole-genome studies Haploview Free Haplotype analysis YES GUI Platform-independent NO YES NO M Allows to perform only statistical analysis Graphical computation of LD statistics and population haplotype patterns coreSNP Free Parallel analysis of Variants\ NO GUI Platform-independent NO YES YES L single allelic\ Massive parallel analysis of SNPs dataset based\ variants discovery;\ on Statistical test Unable to dig with\ Affymetrix raw files PARES Free Parallel association rules extraction from SNP Datasets NO GUI Platform-independent YES NO YES L Manage small Dataset Multiple allelic variants discovery Genotype\ Free Web Services for bioinformatics YES Web interface Platform-independent YES YES YES L Data mining and statistical Web Services to analyse SNP Datasets Analytics Cloud4SNP Free Cloud Serices to analyse SNP Datasets YES Web interface Platform-independent NO YES YES L Allows to perform only statistical analysis Easy to use through Web Browser ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Notes: This table summarizes the main features of the listed genomics framework tools. In the table, the column "Availability" provides information about the license of use of each tool. Free\* in Availability means that the tool is free to use after registration on the vendor web site. "Main features" column shows the main functionalities of the listed tools. "IA" stands for Integrative Analysis specifying if IA is provided by the tool. "Interface" column provides information about how to access the tool functions. "OS" stands for Operating System, indicating the compatibility between OS and each tool. "DM" is short for Data Mining, meaning if DM analysis is provided by the tool. "Statistical" indicates if Statistical Analysis is provided by the tool. "Prep" stands for Preprocessing and it shows if Preprocessing is provided by the tool. "CS" stands for Computer Skills and it describes the levels of competency in using tools (i.e., L indicates basic computer knowledge, M indicates limited computer knowledge, H indicates expert to manage software). [^1]: These authors have contributed equally to this work.

The retained ability to cycle despite freezing of gait (FOG) has been reported to be typical of patients with PD,[@b1], [@b2] and the "bicycle sign" (i.e., the loss of the ability to cycle) has been suggested as a red flag indicative of atypical parkinsonism.[@b3] However, we present here a patient with Fahr\'s syndrome and severe FOG, but a remarkably preserved ability to cycle 7 years after disease onset. Fahr\'s syndrome encompasses a group of neurodegenerative disorders associated with calcification of the basal ganglia, cerebellum, and other brain regions.[@b4] Parkinsonism with FOG and early falls are part of the clinical phenotype of this disorder.[@b5] This 57-year-old patient developed difficulty walking with falls at the age of 53, followed swiftly by stuttering, erectile dysfunction, and urinary urgency. There was no relevant family history. On examination, he had hypomimia, festinant speech, echolalia, and scanning dysarthria. There was bilateral, but asymmetrical, bradykinesia and rigidity, intermittent rest tremor and bilateral postural arm tremor, dysmetria, and dysdiadochokinesia. He had marked FOG and tended to festinate backward on the pull test (see Video, Segment 1). The Mini--Mental State Examination was 25/30. A CT brain scan showed widespread calcification within the basal ganglia and dentate nuclei consistent with Fahr\'s syndrome. Extensive investigation showed no other abnormalities, and a dopamine transporter (DAT) scan was normal. He was treated with levodopa ([l]{.smallcaps}-dopa) with a mild improvement of his symptoms. The patient reported that despite the marked FOG, he could get around his local village by cycling with no difficulty (see Video, Segment 2). Here, we report a patient with atypical parkinsonism and severe FOG who, nonetheless, has a negative bicycle sign, which has been suggested to distinguish patients with PD from atypical parkinsonism. This case suggests that this sign should be used with some caution in this regard. It remains unclear why, in some patients, there is a dissociation between severe difficulty with generating leg movements while walking, manifesting as FOG, and preserved ability to generate leg movements while cycling, although some possible suggestions have been made.[@b2] For example, the action of cycling might represent a type of an external pacing cue that helps to overcome freezing.[@b7] Another notable aspect is that the speed of leg movements seems to improve when cycling. This is not simply an improvement related to the "on" state in PD patients, as this was also observed in our patient who had a normal DAT scan and little response to [l]{.smallcaps}-dopa. The improvement may be related to "paradoxical kinesia," a brief, sudden period of mobility in response to stress or life-threatening events.[@b6] The bicycle sign has been suggested as a new red flag for distinguishing PD from atypical parkinsonism.[@b3] One issue with the previous report is that the atypical parkinsonian patients, taken together as a group, were significantly older and more impaired in terms of UPDRS, postural instability, and ataxia than PD patients. It is unclear whether patients in the earlier stages of atypical parkinsonian conditions also lose the ability to cycle. In conclusion, here, we demonstrate a patient with atypical parkinsonism and marked FOG resulting from Fahr\'s syndrome with a perfectly preserved ability to cycle 7 years after disease onset. We suggest that the bicycle sign should be used with caution as a red flag to distinguish PD from atypical parkinsonism. Legend to the Video =================== The first part of the video demonstrates the patient with Fahr\'s syndrome and severe freezing of gait, which improves when using a visual cue stick. The second part of the video shows the patient cycling without any difficulties. M.S. is supported by an EFNS scientific fellowship. This work was undertaken at UCLH/UCL, which received a proportion of funding from the Department of Health\'s NIHR Biomedical Research Centers funding scheme. We thank the patient and his caregivers for their consent to publish the video. Supplementary material ====================== Additional Supporting Information may be found in the online version of this article. [^1]: **Relevant conflicts of interest/financial disclosures:** M.J.E. receives royalties from the publication of the *Oxford Specialist Handbook of Parkinson\'s Disease and Other Movement Disorders* (Oxford University Press, 2008) and receives research support from a National Institute for Health Research (NIHR) grant, where he is the PI. He has received support for conference travel and accommodation from Teva Pharmaceuticals and UCB Pharmaceuticals. K.B. received funding for travel from GlaxoSmithKline, Orion Corporation, Ipsen, and Merz Pharmaceuticals, LLC, serves on the editorial boards of *Movement Disorders* and *Therapeutic Advances in Neurological Disorders*, receives royalties from the publication of the *Oxford Specialist Handbook of Parkinson\'s Disease and Other Movement Disorders* (Oxford University Press, 2008), received speaker honoraria from GlaxoSmithKline, Ipsen, Merz Pharmaceuticals, LLC, and Sun Pharmaceutical Industries Ltd., and received research support from Ipsen and the Halley Stewart Trust through Dystonia Society UK and the Wellcome Trust MRC strategic neurodegenerative disease initiative award (Ref. no.: WT089698). M.S. and M.K. have no disclosures.Full financial disclosures and author roles may be found in the online version of this article.

\section{Introduction} Connecting microscopic kinetics to emergent rates has a long history for equilibrium and nonequilibrium systems. Complex systems often settle into metastable basins with rare transitions regulating the rates of switching between the emergent states. In equilibrium, transition state theory~\cite{eyring1935activated, laidler1983development}, variational transition state theory~\cite{truhlar1980variational}, Grote-Hynes theory~\cite{grote1980stable,hynes1985chemical}, and transition path sampling (TPS)~\cite{dellago1998transition,bolhuis2002transition} have given strategies to estimate the transition rate. Away from equilibrium, theories built upon a free energy landscape lose their applicability, yet trajectory sampling approaches can still extract rates from microscopic dynamics, either through a noise-guided TPS approach~\cite{crooks2001efficient, gingrich2015preserving} or using forward flux sampling (FFS)~\cite{allen2005sampling,allen2006simulating,allen2009forward}. For all their merits, sampling approaches bring intrinsic challenges. How many samples are required to estimate the rate? Is a good approximate reaction coordinate needed, and if so, how is it found \cite{ma2005automatic}? How does the expense grow with larger or more complicated microscopic systems? Motivated by those ubiquitous concerns, we set out to consider a different approach\textemdash one that effectively samples all possible trajectories as an ensemble evolution instead of propagating individual realizations. Such a proposition may appear absurd given the exploding state space presented by the many-body problem, but we show that for the class of discrete state, continuous-time, reaction-diffusion master equations (RDME), we can evolve the full distribution of a classical many-body systems in a numerically controllable manner by combining two well-developed methods, the Doi-Peliti (DP) formalism with tensor network (TN) algorithms. While TPS and FFS have been developed for a variety of dynamical equations of motion (Hamiltonian, Langevin, Markov jump, etc.), we focus our attention solely on the chemical master equation (CME) dynamics, the setting where FFS was first developed \cite{allen2005sampling}. More specifically, we study discrete reaction-diffusion dynamics of \(L\) sites or voxels, \cite{isaacson2006incorporating,erban2007practical}. The rationale for this focus is twofold. From a practical perspective, the RDME is important as a popular method to model chemical and biological processes \cite{fange2010stochastic,smith2019spatial}. From a theoretical perspective, the RDME is interesting because it provides a setting where one can analyze the influence of stochasticity on the emergence of patterns \cite{craciun2006understanding,krause2021modern,kondo2010reaction}. In many pattern-formation problems, it is possible to understand the underlying physics in terms of a mean-field partial differential equation that tracks spatiotemporal evolution of deterministically evolving fields~\cite{turing1952chemical, cross1993pattern}. It is, however, well-appreciated that stochastic effects can quantitatively and even qualitatively impact kinetics~\cite{mcquarrie1967stochastic,lee2006non,mcadams1997stochastic,thattai2001intrinsic,vilar2002mechanisms,paulsson2004summing,lee2006non}. It is therefore important that the methodology we discuss captures more than just the typical dynamics. Accurately computing transition rates between metastable states requires quantitatively resolving even rare tails of the state space distribution since these rare fluctuations can be instrumental in triggering transitions~\cite{tian2006stochastic,allen2006simulating}. Historically, similar schemes for the CME ensemble evolution have been hampered by the curse of dimensionality. As a result, exact solutions are rare \cite{heuett2006grand} and numerical solutions often resort to sampling methods built upon the kinetic Monte-Carlo stochastic simulation algorithm (SSA) \cite{gillespie2013perspective,cao2006efficient}, also known as the Gillespie algorithm. Another approach is to evolve a subset of the state space through the finite state projection (FSP) method~\cite{munsky2006finite}. FSP has previously been combined with TNs to approximately describe the CME \cite{kazeev2014direct,vo2017adaptive}, resembling our use of TNs to tame the many-body problem. Unlike those projective methods, however, the approach we describe uses the time-dependent variational principle (TDVP) to evolve the dynamics with probability conservation at all times, akin to prior work with diffusion and no reactions~\cite{strand2022using,strand2022computing}. Since we estimate rare events by measuring the distribution's small probability fluxes, conservation of probability and high-fidelity dynamical evolution are materially important. While the formalism we describe applies quite generally to different well-mixed and reaction-diffusion models, we center our paper around a single model problem: rate calculations for transitions between basins of the Schl\"ogl model~\cite{schlogl1972chemical} in the bistable regime, both with and without diffusion. In Sec.~\ref{sec:wellmixed}, we first illustrate the rare event problem and extract rate constants from a propagator in the well-mixed regime. For more complicated systems (either systems with more chemical species or more voxels), the curse of dimensionality precludes the exact calculation of the propagator, but the remainder of the paper shows how to employ TNs to follow the same conceptual path. In Sec.~\ref{sec:dp}, we review the Doi-Peliti framework for writing reaction-diffusion dynamics in a second-quantized field theoretic form. Sec.~\ref{sec:tn} then reviews the TN matrix product state (MPS) ansatz that enables ensemble evolution via TDVP. Those methods are applied to the many-body rate calculation in Sec.~\ref{sec:mbrates}. Finally, in Sec.~\ref{sec:Results}, we show results for the RDME with diffusion between \(L\) well-mixed Schl\"ogl voxels, demonstrating that the methodology can extend numerical transition rate calculations to regimes where the SSA becomes impractical. The approach effectively evolves the probability of occupying each microstate in an unfathomably large many-body state space, yet the calculations scale sub-exponentially with system size and do not assume detail-balanced equilibrium. \section{Ensemble rate calculations}\label{sect:EstRate} \label{sec:wellmixed} Consider an archetypal two-state rate process between reactants and products $\ce{R <=> P}$, initiated entirely in the reactant state. After some reaction timescale, the concentration of products grows before eventually leveling off at its equilibrium value. The reaction rate can simply be defined and measured in terms of how quickly the product concentration rises. An alternative formulation is to study the fate of an individual reactant on its path to transitioning into a product. The TPS and FFS path sampling approaches extract the reaction rate by focusing on the statistical properties of these individual trajectories. Algorithms exist to propagate those single trajectories, algorithms that can be practically implemented even when the system grows large and complex. By contrast, propagating the entire ensemble of trajectories has typically been limited to simple low-dimensional problems. In this section, we illustrate an ensemble-evolution rate calculation for a simple case in order to develop notation and lay the groundwork for the many-body problem. The remainder of the paper will show that the methodology can be extended beyond such simple problems by employing TNs. \subsection{Well-mixed Schl\"ogl Model} \label{sec:WMSchlogl} Consider the well-mixed Schl\"ogl model, a quintessential example of bistability in chemical reaction networks (CRNs)~\cite{vellela2009stochastic}. The model tends to be studied under chemostatted conditions that fix the concentrations of species A and B while allowing species X's concentration to evolve according to the four reactions: \begin{align} \ce{$2$X + A &<=>[\tilde{c}_1][\tilde{c}_2] $3$X}\label{rxn:S1} \\ \ce{B &<=>[\tilde{c}_3][\tilde{c}_4] X} \label{rxn:S2} \end{align} where $\tilde{c}_1,\tilde{c}_2,\tilde{c}_3$ and $\tilde{c}_4$ are the stochastic rate coefficients. Our starting point is microscopic such that the rate \(\tilde{c}_1\) should be interpreted as the reciprocal of the waiting time for a particular set of $2$ X molecules to react and form 3 X molecules when mediated by a fixed concentration of A molecules. In the large system, well-mixed limit of chemical kinetics, it is common to also define macroscopic, or kinetic rate constants \(k_r\) for each reaction \(r\). The relation between $k_r$ and the microscopic rate coefficients \(\tilde{c}_r\) are only equivalent for linear reactions and in the thermodynamic limit~\cite{higham2008modeling}. By varying the number of A and B molecules along with the rate coefficients, the Schl\"{o}gl system can occupy either a unimodal or a bimodal regime. In the bimodal regime, the number $n$ of species X fluctuates, switching between two metastable states. A representative trajectory, shown in Fig.~\ref{fig:ER}, oscillates between a state with \(n \approx 5\) and another with \(n \approx 50\) molecules. We first set out to compute the rate of bistable switching between those two metastable states, making use of the fact that the dynamics is one-dimensional (1D)\textemdash \(n\) executes a random walk between 0 and some large maximum number \(M\) that we impose for the sake of calculation. \begin{figure}[htb] \centering \includegraphics[width=.45\textwidth]{Figures/Fig1.pdf} \caption{\label{fig:ER}(Top) A typical trajectory in the bistable regime ($c_1 = 2.676, c_2 = 0.040, c_3 = 108.102,$ and $c_4 = 37.881$) of the well-mixed Schl\"{o}gl model of Sec.~\ref{sec:WMSchlogl}. The system stochastically switches between the two metastable basins, where basin $\mathcal{A}$ is the top colored region and $\mathcal{B}$ is the bottom colored region. (Bottom) Starting with a distribution localized in basin $\mathcal{A}$, the transition rate can be extracted from the probability growth in the linear regime of the evolution, $P(\mathcal{B}_t|\mathcal{A}_0) \propto k_{\mathcal{B}\mathcal{A}}t$. The inset shows that growth on a much longer timescale, revealing the eventual approach toward equilibrium.} \end{figure} Rather than track \(n(t)\) for a single trajectory as in Fig.~\ref{fig:ER}, the ensemble approach tracks evolution of the distribution \(p_t(n)\), measuring the probability that an individual trajectory would have \(n\) molecules at time \(t\). Foreshadowing the analogies with quantum mechanical methodologies, we choose to write that distribution as a ket: \(\ket{p_t} = \left[p_t(0), p_t(1) \hdots, p_t(M)\right]^T\), where \(T\) indicates the transpose to give a column vector. The distribution evolves according to the master equation \begin{equation} \frac{d\ket{p_t}}{dt} = H\ket{p_t}. \label{eq:ClSch} \end{equation} Here, \(H\) is a rate matrix whose off-diagonal elements \(H_{n'n}\) are the rates of transitioning from \(n\) X molecules to \(n'\) and whose diagonal elements \(H_{nn}\) are minus the rate of escaping from the state with \(n\) X molecules. The rate matrix conserves probability, a fact compactly expressed as $\bra{\mathbf{1}}H = 0\bra{\mathbf{1}}$, with $\bra{\mathbf{1}}$ denoting a vector of ones. The distribution \(\ket{p_t}\) can always be formally evolved in terms of a propagator \(e^{Ht}\) as \begin{equation} \ket{p_t} = e^{Ht}\ket{p_0}. \label{eq:Propagator} \end{equation} Due to the analogy with quantum-mechanical time evolution, we equivalently call \(H\) an effective Hamiltonian. Assuming $H$ is irreducible, the long-time limit for this distribution tends to the unique steady-state $\ket{\boldsymbol{\pi}} = \lim_{t\to \infty} e^{Ht}\ket{p_0}$. \(H\) can be decomposed into contributions from the two different reversible mechanisms of Eqs.~\eqref{rxn:S1} and~\eqref{rxn:S2}, labeled \(H^{A}\) and \(H^{B}\), respectively, based on whether the transitions are mediated by the chemostatted species A or B. Then, the Schl\"ogl mechanism at the level of the master equations corresponds to tridiagonal rate matrices with elements \begin{align} \label{eq:tridiagonal} H^A_{n'n} = &\frac{\tilde{c}_1 n (n-1)}{2}\delta_{n'-1,n} \frac{\tilde{c}_2 n (n-1)(n-2)}{6}\delta_{n'+1,n} - \nonumber \\ &\frac{n (n-1)}{2}\left(\tilde{c}_1 + \frac{\tilde{c}_2(n-2)}{3}\right)\delta_{n',n} \end{align} and \begin{equation} H^B_{n',n} = \tilde{c}_3\,\delta_{n'-1,n} + \tilde{c}_4n \delta_{n'+1,n} - \left(\tilde{c}_3 + \tilde{c}_4n \right)\delta_{n',n}, \end{equation} where $H = H^A + H^B$. Given the specific tridiagonal \(H\) and the restriction \(0 \leq n \leq M\), both \(H\) and the propagator \(e^{Ht}\) are \((M+1) \times (M+1)\) matrices, amenable to evolving the entire distribution of realizations, effectively averaging over all the realizations. To extract a rate from the evolution of distribution, we must initialize the distribution in one of the two metastable states and measure the rate of relaxation. In analogy with free energy basins, we will call the two metastable regions basins \(\mathcal{A}\) and \(\mathcal{B}\) (red and blue regions of Fig.~\ref{fig:ER} respectively) with \(\mathcal{A} \cap \mathcal{B} = \emptyset\). We write \(\hat{\mathcal{P}}_{\omega}\) for the non-negative idempotent ($\hat{\mathcal{P}}_{\omega}^2 = \hat{\mathcal{P}}_\omega$) operator that projects onto the density lying in a region \(\omega\), where \(\omega \in \{\mathcal{A},\mathcal{B}\}\). If \(\ket{p_t}\) has support over a state space \(\Omega\), then the projection \(\hat{\mathcal{P}}_\omega \ket{p_t}\) is an unnormalized distribution with support restricted to \(\omega\). The total probability in \(\omega\) at time \(t\) can then be expressed as \(\bra{\mathbf{1}}\hat{\mathcal{P}}_\omega\ket{p_t}\). \subsection{Rate Constants} \label{sec:rateconstants} In general, basins can be arbitrary subsets of $\Omega$. For the current example, $\mathcal{A}$ and $\mathcal{B}$ are basins of attraction for the Schl\"{o}gl model (see Fig.~\ref{fig:ER}a), and we would like to measure the rate of going from $\mathcal{A} \rightarrow \mathcal{B}$. Rate constants can be defined either for the unidirectional flux from $\mathcal{A} \rightarrow \mathcal{B}$, the quantity we focus on in this work, or for the timescale of the relaxation to equilibrium, which also involves the $\mathcal{B} \rightarrow \mathcal{A}$ transitions~\cite{roux2022transition}. Additionally, the precise definition of the rate depends on whether one seeks trajectories that move from the boundary of one basin to another or from the interior of one basin to another. For instance, studying the time to transition from the boundary of \(\mathcal{A}\) to the boundary of \(\mathcal{B}\) yields the Transition Path Theory (TPT) rate related to the committor function~\cite{e2010transition}. Here, we will have access to the distribution over all states in the interior of $\mathcal{A}$ and $\mathcal{B}$, allowing for the calculation of the unidirectional rate \(k_{\mathcal{B} \mathcal{A}}\). That $k_{\mathcal{B}\mathcal{A}}$ is the rate to start in $\mathcal{A}$ at time zero with probability $P(\mathcal{A}_0)$, and transition to any state in $\mathcal{B}$ at time $t$ with conditional probability $P(\mathcal{B}_t|\mathcal{A}_0)$. We initialize with a distribution over microstates $\hat{\mathcal{P}}_\mathcal{A}\ket{\boldsymbol{\pi}}/\bra{\mathbf{1}}\hat{\mathcal{P}}_{\mathcal{A}}\ket{\boldsymbol{\pi}}$, formed by projecting the equilibrium distribution onto $\mathcal{A}$ and normalizing, making $P(\mathcal{A}_0)=1$ by construction. With this initialization, the rate for transitioning from $\mathcal{A}$ to $\mathcal{B}$ can then be written in terms of the time derivative of the conditional probability of occupying $\mathcal{B}$ a time $t$ later \cite{allen2005sampling}: \begin{align} k_{\mathcal{B} \mathcal{A}} &\equiv \frac{d}{dt} P(\mathcal{B}_t | \mathcal{A}_0)P(\mathcal{A}_0)\biggr\rvert_{t >\tau_{\rm mol}},\nonumber \\ &= \frac{d}{dt} \frac{\bra{\mathbf{1}}\hat{\mathcal{P}}_{\mathcal{B}}\,e^{Ht}\, \hat{\mathcal{P}}_{\mathcal{A}} \ket{\boldsymbol{\pi}}}{\bra{\mathbf{1}}\hat{\mathcal{P}}_\mathcal{A}\ket{\boldsymbol{\pi}}}\biggr\rvert_{t >\tau_{\rm mol}}. \label{eq:rate} \end{align} That derivative is evaluated at some timescale exceeding $\tau_{\rm mol}$, a molecular time scale of the system that is long enough for the first trajectories to reach \(\mathcal{B}\) from \(\mathcal{A}\) but much shorter than the typical transit time $k_{\mathcal{B}\mathcal{A}}^{-1}$. To the extent that there is a separation of timescales between the time to explore the basin and the time to transit between basins, distinctions between this rate and a TPT rate are insignificant. For rare events, no transitions can occur faster than the molecular time scale \(\tau_{\rm mol}\), beyond that point, the number of transitions accumulates over time as \(P(\mathcal{B}_t|\mathcal{A}_{0}) \propto k_{\mathcal{B}\mathcal{A}} t\). At times long compared to \(k_{\mathcal{B} \mathcal{A}}^{-1}\), \(P(\mathcal{B}_t|\mathcal{A}_{0})\) must eventually stop growing linearly to level off at the stationary value \(\bra{\mathbf{1}}\hat{\mathcal{P}}_\mathcal{B} \ket{\boldsymbol{\pi}}\), but our interest is in extracting \(k_{\mathcal{B} \mathcal{A}}\) from the growth at much shorter times. A typical realization of the process will not require observation much beyond \(\tau_{\rm mol}\) units of time, to ascertain the rate, meaning our ensemble evolutions must be propagated for roughly $\tau_{\rm mol}$. Our ability to compute \(k_{\mathcal{B} \mathcal{A}}\) from the ensemble evolution traces back to the fact that it is practical to numerically calculate \(e^{Ht}\) for the well-mixed model with a single species and a modest \(M\). While Eq.~\ref{eq:rate} generalizes naturally for harder problems (those with more than one species or with diffusion between multiple well-mixed voxels), the state space for \(\ket{p_t}\) grows exponentially with additional stochastic degrees of freedom. We next show that the Doi-Peliti framework re-expresses the rate matrix in a second-quantized form that allows us to generalize beyond the case of a single well-mixed species. \begin{figure*}[!ht] \centering \includegraphics[width=0.95\textwidth]{Figures/Fig2.pdf} \caption{\label{fig:one} The reaction-diffusion master equation (RDME) consists of voxels in which well-mixed chemical reactions occur (a) and between which species can diffuse (b). This work centers around a one-dimensional chain of $L$ such voxels, each with a fluctuating number of molecules of species X in a Schl\"{o}gl chemical reaction. The Doi-Peliti (DP) framework turns the reaction-diffusion evolution into a second-quantized effective Hamiltonian that governs the evolution of the distribution over classical reaction-diffusion microstates. (c) That effective Hamiltonian is represented by a high-dimensional tensor, with each vertical leg taking one of $(M+1)$ possible values corresponding to the number of molecules of species X which could occupy each voxel. The DP form of the effective Hamiltonian lends itself to a compact representation of the effective Hamiltonian tensor of rates for transitioning from one microstate to another.} \end{figure*} \section{Doi-Peliti} \label{sec:dp} The Doi-Peliti (DP) formalism is a classical version of the second quantization methods from quantum field theory \cite{ohkubo2013algebraic}. Originally introduced by Doi \cite{doi1976stochastic} as a critique to the assumption of the Smoluchowski equation, it was later developed by Peliti \cite{peliti1985path} and others \cite{rey1999asymptotic} as a path integral formalism for solving chemical reaction networks. The DP framework has also found success in other areas such as studying kinetically constrained lattice models in the context of glassy dynamics~\cite{garrahan2007dynamical,garrahan2009first}. The DP representation builds operators which encode the dynamics onto the high-dimensional space that accompanies many-body problems. Crucially, writing the chemical mechanism in an operator form guarantees it can be easily converted into a matrix product operator (MPO)~\cite{schollwock2011density,causer2022finite}, as we will do in Sec.~\ref{sec:tn}.\\ To start, we go beyond Sec.~\ref{sec:wellmixed} by considering well-mixed reactions with multiple dynamic species as well as heterogeneous systems where molecules can diffuse between neighboring voxels. Let $X_1, X_2,\ldots X_L$ represent each chemical species in each voxel. We could be talking about a single species \(X\) with different values in \(L\) different voxels, a single well-mixed voxel with \(L\) distinct dynamic species, or a mixture with multiple species and multiple voxels. The number of molecules in a given voxel can change due to a chemical reaction or a diffusion event, with both events being cast as, \begin{equation} \ce{ $\sum_j \eta_j^r$ X$_j$ ->[\tilde{c}_r] $\sum_j \mu_j^r$ X$_j$} \label{eq:rxn} \end{equation} where $\tilde{c}_r$ are again the rate coefficients for reaction $r$. $\eta_j^r$ and $\mu_j^r$ reflect the stochiometry of the reactions and give the stochiometric vector $\nu^r = \mu^r - \eta^r$ which evolves the system from state $\mathbf{n} \in \Omega$ in state space $\Omega$, to $\mathbf{n}' = \mathbf{n} + \nu^r$. In the previous section, these reactions induced changes between the \(M+1\) different microstates, specified by the scalar \(n\). Now, the microstates $\mathbf{n}$ are specified by \(\mathbf{n} = [n_1, n_2, \hdots, n_L]\), counting the occupation of each species within each voxel. For a dilute well-mixed solution, where well-mixed means molecules are evenly distributed around a given sub-volume, the change in probability of microstate $\mathbf{n}$ follows the gain-loss equation \cite{isaacson2006incorporating}, \begin{equation} \frac{dp_t(\mathbf{n})}{dt} = \sum_{r}^N\left[ \alpha_{\mathbf{n} - \nu^r}^r p_t({\mathbf{n} - \nu^r}) - \alpha_{\mathbf{n}}^r p_t({\mathbf{n}})\right], \label{eq:chgPr} \end{equation} where the sum is taken over \(N\) different reactions of the form Eq.~\eqref{eq:rxn}, each with their own stochiometric vector \(\nu^r\). The transition rate between any two states is the sum of contributions from each reaction \(H_{\mathbf{n}'\mathbf{n}} = \sum_r \alpha^r_{\mathbf{n}}\). Each propensity $\alpha^r$ is a product of a stochastic rate coefficient $\tilde{c}_r$ and the combinatorial number of ways to select either $\eta_j$ or $\mu^r$ molecules from reaction $r$, depending on the direction of the reaction. For a ``forward" reaction the propensity is, $\alpha_r(\mathbf{n}) = \tilde{c}_r\prod_{n_j\in r}{n_j\choose\eta_j}$. Encoding the reactions into a matrix form allows us to evolve the elements of the joint distribution $p_t(\mathbf{n})$ analogously to the previous section. While it is appealing to try and directly solve these more complicated dynamics as we did before, it is impractical. The curse of dimensionality demands a less onerous way to work with both \(H\) and \(p_t(\mathbf{n})\). Since particles of the same species in the same voxel are indistinguishable, it is beneficial to work in an occupation number representation with \begin{equation} \ket{p_t} = \sum_{\mathbf{n}} p_t(\mathbf{n}) \ket{\mathbf{n}}. \end{equation} The basis vectors \begin{align} \nonumber \ket{\mathbf{n}} &= \ket{n_1, n_2, \hdots n_L}\\ &= \ket{n_1} \otimes \ket{n_2} \otimes \hdots \otimes \ket{n_L} \end{align} corresponds to a possible microstate, and the set of basis vectors forms a tensor product space called a Fock space $\mathcal{F}^L$ with $\ket{\mathbf{n}} \in \mathcal{F}^L$ and \(\bra{\mathbf{m}}\ket{\mathbf{n}} = \delta_{\mathbf{m}\mathbf{n}}\)~\cite{del2022probabilistic}. Without loss of generality, we will discuss the case of a single chemical species and \(L\) physical sites, in which case \(n_l\) has the interpretation of the number of molecules at site \(l\). The second-quantization procedure rewrites \(H\) in terms of creation and annihilation operators, \(x_l^\dagger\) and \(x_l\), that respectively raise and lower the number of X molecules at site \(l\). That is to say, \begin{align} x_l^\dagger\ket{n_l} &= \ket{n_l+1},\nonumber \\ x_l\ket{n_l} &= n_l\ket{n_l-1},\nonumber \\ x_l\ket{0} &= 0. \label{eq:raiselower} \end{align} where $\ket{0}$ is the vacuum state, i.e., no molecules. By $x_l$ or $x_l^\dagger$ we are implying that the operator only acts on the $l^{\rm th}$ site: \begin{equation} x_l = \mathbb{I}_1 \otimes \mathbb{I}_2\otimes\ldots\otimes \mathbb{I}_{l-1}\otimes x \otimes \mathbb{I}_{l+1} \otimes \ldots \otimes \mathbb{I}_L \end{equation} Because they act on the single-site space, the general one-site operators $x$ and $x^\dagger$ can be written as the \(M+1 \times M+1\) matrices \begin{equation} x^\dagger = \begin{bmatrix} 0 && 0 && 0 && \ldots \\ 1 && 0 && 0 && \ldots \\ 0 && 1 && 0 && \\ \vdots && && \vdots && \ddots \end{bmatrix} \text{ and } \qquad x = \begin{bmatrix} 0 && 1 && 0 && \ldots \\ 0 && 0 && 2 && \ldots \\ 0 && 0 && 0 && \\ \vdots && && \vdots && \ddots \end{bmatrix}. \end{equation} That these matrices yield Eq.~\eqref{eq:raiselower} can be readily confirmed. More abstractly, the existence of such raising and lowering operators in this classical setting originates from the classical commutation relation \([x_i x_j^\dagger] = \delta_{ij}\) and the existence of a positive semi-definite number operator $\hat{n}_l = x_l^\dagger x_l$, so named because \(\hat{n}_l\ket{n_l} = n_l \ket{n_l}\). Together, the set $\{\Omega,\mathcal{F}^L,x,x^\dagger\}$ are also sometimes called an interacting Fock space \cite{ohkubo2013algebraic}. Our use of the raising and lowering operators is less formal and more pragmatic. As in the construction of the number operator, we will be able to decompose contributions to \(\hat{H}\) in terms of the \(x^\dagger\) and \(x\)'s. \begin{figure*}[!ht] \centering \includegraphics[width=.9\textwidth]{Figures/Fig3.pdf} \caption{\label{fig:mps} The ensemble evolution of the many-body state requires that one solve for the probability of every microstate at all times. That high-dimensional joint probability distribution can be well approximated by a matrix product state (MPS). An MPS approximation for the joint distribution can be advantageous whether there are $N_s$ species in a well-mixed reaction [(a) top], one dynamical (unchemostatted) species with diffusion [(a) middle], or multiple species and diffusion [(a) bottom]. In all three cases, each circle represents a tensor, either a rank-$2$ at the first and last site or rank-$3$ in the body. Horizontal lines represent the bond-indices connecting tensors and vertical lines are physical indices. For an RDME, the physical indices represent the possible number of molecules at a given tensor. If two lines are connected, then the index is summed over. (b) An effective Hamiltonian $H$ can be likewise decomposed into lower-rank tensors to form a matrix product operator (MPO). Now there are two physical indices per tensor, representing transitions from state $n_l^X$ to $m_l^X$. Vertical connected lines are the bond indexes which are contracted over.} \end{figure*} One benefit of defining operators in terms of creation and annihilation operators is that expectation values for observables can be expressed easily in terms of inner products with operators. Consider, for example, the observable that counts all particles in all voxels, \(\hat n = \sum_l \hat n_l\). The operator \(\hat{n}\) acts on a many-body state \(\ket{\mathbf{n}}\), yet each \(\hat{n_l}\) is a single-body operator acting on site $l$ and leaving all other sites unchanged. It is therefore understood that when local operators such as $\hat n_l$ act on $\ket{\mathbf{n}}$, it is written as a shorthand for \begin{equation} \hat n_l = \mathbb{I} \otimes \hdots \otimes x^\dagger_lx_l \otimes \hdots \otimes \mathbb{I}. \end{equation} In terms of that shorthand, $\hat n_l \ket{\mathbf{n}} = n_l \ket{\mathbf{n}}$ counts the number of particles in site $l$ for the many-body state $\ket{\mathbf{n}}$. The time-dependent expected number of particles in the whole system is thus \(\langle n(t) \rangle = \sum_l\bra{\mathbf{1}}\hat n_l \ket{p_t}\), where the vector of ones \(\bra{\mathbf{1}}\) serves to sum over all microstates. The time-dependence of a distribution can also be extracted from the action of a second-quantized \(\hat{H}\) on \(\ket{p_t}\), and this operation can be practically computed when \(\hat{H}\) is expressed in terms of the single-site operators. Mirroring the discussion in Sec.~\ref{sec:wellmixed}, the generator $\hat H_l$ for the Schl\"ogl model reactions within the $l^{\rm th}$ voxel is a sum of contributions from the elementary reactions $\hat H_l = \sum_r \hat H^r$. Fig.~\ref{fig:one}a shows each reaction, expressed as a second-quantized contribution to the effective Hamiltonian, as derived in App.~\ref{app:DPEnc}. Combining together the forward and reverse reactions mediated by species A, we see that the reversible reaction Eq.~\eqref{rxn:S1} contributes \begin{equation} \hat{H}^A_l = x_l^{\dagger 2}\left[\left(c_1x^\dagger_l + c_2\left(\mathbb{I} - x_l^\dagger\right)x_l\right)x_l^2 - c_1x_l^2\hat{\mathcal{P}}^c_{M_l}\right]. \label{eq:HA} \end{equation} to the effective Hamiltonian. For convenience we have absorbed the combinatorial terms into that rate coefficient, defining $c_r \equiv \tilde{c}_r / \gamma!$, where $\gamma$ is either $\eta^r$ or $\mu^r$ for forward and reversed reactions, respectively. Likewise, the contribution \begin{equation} \hat{H}_l^B = x_l^\dagger(c_3-c_4)x_l + c_4x_l - c_3\hat{\mathcal{P}}^c_{M_l}. \label{eq:HB} \end{equation} comes from Eq.~\eqref{rxn:S2}. We show in App.~\ref{app:DPEnc} that the truncation of the state space at finite \(M\) requires the addition of a projector, $\hat{\mathcal{P}}^c_{M_l}$ to prevent \(\hat{H}_l\) from promoting any probability from the \(M^{\rm th}\) state into an \((M+1)^{\rm th}\) that has been truncated away. In the $M \rightarrow \infty$ limit, the projector can be discarded, as reflected by the expressions in Fig.~\ref{fig:one}a. The Hamiltonian operator for \(L\) independent, well-mixed Schl\"ogl voxels is thus \(\hat{H}^{\rm rxn} = \sum_{l}^L \hat{H}^A_l + \hat H^B_l\). To add diffusion, consider the well-mixed voxels arranged in a 1D lattice as in Fig.~\ref{fig:one}b. Each chemical species is able to move between neighboring voxels as regulated by a diffusion operator \begin{equation} \hat{H}^{\rm diff} = \sum_{l = 1}^{L-1} d \left[\left(x^\dagger_{l+1} - x^\dagger_l\right)\left(x_l - x_{l+1}\right)\right], \label{eq:Hdiff} \end{equation} with hopping rate \(d = D/h^2\), consistent with macroscopic diffusion constant $D$ for voxel width $h$ \cite{smith2019spatial}. Combining reaction with diffusion: \begin{equation} \hat{H} = \hat{H}^{\rm rxn} + \hat{H}^{\rm diff}, \end{equation} is straightforward at the level of the effective Hamiltonian, but the diffusive coupling between neighbors has the consequence of introducing nontrivial interactions between voxels. \section{Tensor Networks}\label{sec:tn} We have now seen how to transform the matrix-form rate operator $H$ of Sec.~\ref{sect:EstRate} into a Doi-Peliti form $\hat{H}$, expressed in terms of local creation and annihilation operators. That change of representation is not merely cosmetic. As the number of chemical species or number of lattice sites grows, the size of $H$ grows exponentially. For $N$ dynamical (unchemostatted) species and $L$ lattice sites, the state space has $(M+1)^{NL}$ states. To put this scaling in perspective, the largest numerical result that follows uses $N=1$, $M=85$ and $L=8$, meaning the system contains $2.99\times 10^{15}$ microstates. Calculations on a matrix $H$, even a very sparse matrix $H$, are untenable. By contrast, it is practical to construct a TN decomposition of the second quantized form of $\hat H$ in terms of an MPO, as shown in Fig.~\ref{fig:mps}b, and demonstrated explicitly for the Schl\"ogl mechanism in Appendix \ref{app:MPO}. Given the second quantized form, the MPO can be readily constructed in a practical form through SVD decomposition using automated packages such as autoMPO in ITensor~\cite{itensor}. Much like the operator \(\hat{H}\) can be compressed as an MPO, the states that operators act upon can also be expressed more compactly in terms of TNs. Consider \(\ket{p_t} = p_t(\mathbf{n}) = p_t(n_1, n_2, \hdots, n_L)\), the probability of microstate \(\mathbf{n}\) at time \(t\). This \(\ket{p_t}\) can be thought of as a rank \(L\) tensor. By specifying the occupancy of each of the \(L\) sites, \(\ket{p_t}\) outputs a number\textemdash a probability\textemdash associated to that microstate. By leveraging a matrix product ansatz, $\ket{p_t}$ can be approximated as a partial contraction over $L$ different low-rank tensors (see Fig.~\ref{fig:mps}a). The low rank tensors of an MPS~\cite{orus2014practical} are organized in a 1D chain of tensors \(Q_{s_1}^{n_1}, Q_{s_1 s_2}^{n_2}, \hdots, Q_{s_{L-2} s_{L-1}}^{n_{L-1}},\) and \(Q_{s_{L-1}}^{n_L}\) whose upper index \(n_{l}\) specifies the occupancy of site \(l\) and whose lower indices are so-called bond indices, nonphysical indices which are contracted over to generate correlations between nearby sites. For suitably chosen \(Q\) tensors, we therefore approximate \begin{align} \ket{p_t} &= \sum_{n_1, n_2, \hdots n_L} p_t(n_1,n_2,\ldots,n_L) \ket{n_1, n_2, \hdots n_L} \nonumber\\ &\approx \hspace{-3mm}\sum_{s_1,\ldots,s_{L-1}}\hspace{-3mm} Q^{n_1}_{s_1}Q^{n_2}_{s_1s_2}\ldots Q^{n_L}_{s_{L-1}}\ket{n_1, n_2, \hdots n_L} \nonumber\\ &\equiv \ket{q_t}, \label{eq:Pr_MPS} \end{align} where $\ket{q_t}$ is the approximate distribution over microstates generated from the MPS. Though not explicitly written, the \(Q\) tensors are time-dependent. There is a diagrammatic representation of the tensors, shown in Fig.~\ref{fig:mps}, which aids in visualizing tensor operations. The tensors \(Q\) are given by circles, one per site, and vertical lines feeding into these circles are the physical indices that pick out a particular slice of the tensor based on the occupation number of that site. Circles also have horizontal lines emanating from them, representing the non-physical bond indices. Connection of two circles by a line represents a shared index being contracted over. Fig.~\ref{fig:mps} emphasizes that the TN decomposition is flexible enough to address varied CME's. The MPS of Fig.~\ref{fig:mps}a middle, has been the focus. Each site from 1 to \(L\) reflects the number of X molecules in that voxel, but Fig.~\ref{fig:mps}a (top) and (bottom) show that the MPS decomposition can equally well be constructed when multiple species X$_i$, $i = 1,2,\ldots,N_{\rm s}$ can occupy a single voxel or when $N_{\rm s}$ different species diffuse between $L/N_{\rm s}$ voxels to yield an MPS with $L$ sites. What is practically important is that \(Q\)'s be arranged such that the physical dynamics correlates the neighbors. For example, for the RDME Schl\"{o}gl model, neighboring \(Q\)'s correspond to neighboring sites in the 1D Schl\"{o}gl lattice. An attractive feature of the MPS ansatz is that the size and accuracy of the MPS are controllable through the bond dimension of the $s$ auxiliary indices. If the dimension of those indices is allowed to grow exponentially from $s_1$ through $s_{L/2}$, the ansatz in Eq.~\eqref{eq:Pr_MPS} can be exact, but even when the bond dimension of the indices $s$ are capped at a maximum dimension of $\kappa$, the ansatz can be a very good approximation. Crucially, capping that maximum bond dimension makes it practical to work with $\ket{q_t}$ because the rank-$L$ tensor $\ket{p_t}$ has been replaced by the $L$ different $Q$ tensors, each with no more than $(M+1)\kappa^2$ elements. Having approximated both $\ket{q_t}$ and $\hat{H}$ by an MPS and MPO, respectively, we can now revisit the otherwise intractable dynamics problem of Eq.~\eqref{eq:ClSch}, only now we wish to evolve $\ket{q_t}$ in lieu of $\ket{p_t}$. As $\ket{q_t}$ is an MPS state, we wish for the dynamics to be constrained for all times to the manifold $\mathcal{M}$ of possible MPS states. Merely replacing $\ket{p_t}$ by $\ket{q_t}$ in Eq.~\eqref{eq:ClSch}: \begin{equation} \frac{d \ket{q_t}}{dt} = \hat{H} \ket{q_t} \label{eq:qeom} \end{equation} would not impose such a constraint because $\hat{H}$ in general evolves $\ket{q_t}$ off $\mathcal{M}$. An appealing resolution is given by the Time-Dependent Variational Principle (TDVP)~\cite{dirac1930note,haegeman2011time}, which proposes to first project the right-hand side onto the tangent space of the MPS manifold: \begin{equation} \frac{d \ket{q_t}}{dt} = \mathcal{P}_{T_{\mathcal{M}}} \hat{H} \ket{q_t}. \label{eq:tdvpeom} \end{equation} In this way, dynamics is restricted to the desired sub-space of MPS states~\cite{vanderstraeten2019tangent}. It turns out that the time-discretized numerical integration of this $\ket{q_t}$ is amenable to efficient TN operations. The technical details of the time-evolution algorithm have been reported elsewhere \cite{paeckel2019time,haegeman2011time,haegeman2016unifying,lubich2015time}. We give only a brief, high-level overview of that procedure (see also App.~\ref{app:tdvp} for more details). The central step in deriving the TDVP time-evolution algorithm is to demonstrate that the projector $\mathcal{P}_{T_{\mathcal{M}}}$ can be decomposed in terms of a sum over $L$ terms (associated to each site of the MPS) and $L-1$ terms (associated to each bond connecting sites). In the limit of a small time step $\delta t$, the propagator $e^{\mathcal{P}_{T_{\mathcal{M}}} \hat{H} \delta t}$ can thus be evaluated as a composition of time-evolutions involving the $2L-1$ individual terms. That composition of time-evolutions is executed by a one-site TDVP algorithm that ``sweeps'' from lattice site 1 through $L$, advancing each term for the small time step. The efficiency of the TDVP time-evolution algorithm relies on the flexibility to represent the same MPS state in redundant ways due to a gauge freedom. Consider, for example, neighboring tensors $Q^{n_l}_{s_{l-1}s_l}$ and $Q^{n_{l+1}}_{s_ls_{l+1}}$ in an MPS. For an invertible $R$, replacing those tensors by $Q^{n_l}_{s_{l-1}s_l} R$ and $R^{-1} Q^{n_{l+1}}_{s_ls_{l+1}}$, respectively, does not change the MPS state; upon contracting over the neighboring sites, the identity $R R^{-1}$ would fall out. By leveraging the gauge freedoms, it is always possible to transform Eq.~\eqref{eq:Pr_MPS} into a so-called mixed canonical form~\cite{vidal2003efficient}: \begin{equation} \ket{q_t} = A^{n_1} \hdots A^{n_{l-1}} Q^{n_l} B^{n_{l+1}} \hdots B^{n_l} \ket{n_1, \hdots, n_L}, \label{eq:mixedcanonical} \end{equation} where each $A$ is a left-orthogonal tensor satisfying $A^T A = \mathbb{I}$ and each $B$ is a right-orthogonal tensor satisfying $B B^T = \mathbb{I}$, where here, $T$ represents the transpose. For notational compactness, we have suppressed implied summations over the bond indices between tensors in Eq.~\eqref{eq:mixedcanonical}. The mixed canonical form is said to be centered about the tensor $Q^{n_l}$. From that tensor's perspective, the chain of $A$'s and chain of $B$'s constitute the environment, and it is beneficial to contract over the links between $A$'s and links between $B$'s to obtain so-called environment tensors. These environment tensors inherit an orthogonality from the orthogonality of the $A$'s and $B$'s, significantly reducing the computational complexity to advance $Q^{n_l}$ in time by $\delta t$ provided the MPS was already centered about the $Q^{n_l}$ we aim to evolve. \begin{figure*}[hbt] \centering \includegraphics[width=1\textwidth]{Figures/Fig4.pdf} \caption{\label{fig:Snaps} Snapshots of the joint density in two neighboring voxels from an $L=3$ site lattice of Schl\"{o}gl voxels. Using $c_1 = 2.676, c_2 = 0.040, c_3 = 108.102, c_4 = 37.881$ and $d = 8.2207$, the steady-state distribution was computed with $\kappa = 30$ by evolving an initial uniform distribution according to $\hat{H}$ via TDVP. That steady-state distribution was projected into $\mathcal{A}$ at time 0, renormalized, then propagated according to the modified no-recrossing dynamics $\tilde{H}$ using TDVP. For clarity of visualization, the density within $\mathcal{A}$ is not plotted as it would overwhelm the small probability that leaks into $\mathcal{B}$. By $t = 0.01$ (left image), probability has begun to leak from $\mathcal{A}$ but none has reached $\mathcal{B}$. Probability has just begun to reach $\mathcal{B}$ at at $t = 0.14$, roughly the $\tau_{\rm mol}$ timescale for this problem. By $t = 0.44$, probability is collecting in $\mathcal{B}$, and by $t = 0.75$, the flux passing from $\mathcal{A}$ to $\mathcal{B}$ has attained an approximately constant steady-state value corresponding to the linear growth phase where $P(\mathcal{B}_t|\mathcal{A}) \propto k_{\mathcal{B}\mathcal{A}}t$.} \end{figure*} As made more explicit in App.~\ref{app:tdvp}, the one-site TDVP algorithm initially centers the MPS around the first tensor, propagates that tensor forward in time by $\delta t$, performs a QR decomposition to change the gauge, propagates backward in time by $\delta t$ the R term of that QR decomposition then passes that term to the neighboring site to re-center the MPS around the neighboring site. The neighboring site is then propagated forward in time as the algorithm iterates. After the complete sweep through the MPS, the probability distribution $\ket{q_t}$ has been propagated by a time step $\delta t$. Whether that propagation well approximates the evolution of the full distribution $\ket{p_t}$ depends on the time-step error but also on the severity of the MPS approximation, which can be systematically controlled through the maximum bond dimension $\kappa$. Notably, the time-evolved MPS state $\ket{q_t}$ allows us to estimate the probability of \emph{any} microstate of interest, even though it would be impossible to enumerate probabilities of all microstates. Alternatively, sets of microstates can be summed over to yield tractable marginal distributions. As an example, Fig.~\ref{fig:Snaps} shows the results of calculations on a 1D chain of $L = 3$ Schl\"{o}gl voxels. Though there are three voxels, site one is traced over to leave the two-dimensional joint probability of finding $n_2^X$ X molecules in site 2 and $n_3^X$ X molecules in site 3. At time zero, the distribution is initialized within region $\mathcal{A}$ according to $\hat{\mathcal{P}}_\mathcal{A}\ket{\boldsymbol{\pi}}/\bra{\mathbf{1}}\hat{\mathcal{P}}_{\mathcal{A}}\ket{\boldsymbol{\pi}}$. As time progresses, a small flux of probability leaks from $\mathcal{A}$ to $\mathcal{B}$. Because the overwhelming preference is to stay in $\mathcal{A}$, Fig.~\ref{fig:Snaps} highlights the rare transitions by only plotting the joint distribution outside of $\mathcal{A}$. In the next section, we illustrate the fact that the MPS state's ability to capture the rare time-dependent flux is key to calculating the rate of traversing between basins. \section{Many-body rate calculation}\label{sec:mbrates} Having established the TDVP dynamics of an MPS state, we can now repeat the strategy of Eq.~\eqref{eq:rate} to compute a transition rate between states $\mathcal{A}$ and $\mathcal{B}$: \begin{equation} \label{eq:tdvpdynamicsrate} k_{\mathcal{B} \mathcal{A}} = \frac{d}{dt} \frac{\bra{\mathbf{1}}\hat{\mathcal{P}}_{\mathcal{B}}\,e^{\mathcal{P}_{T_{\mathcal{M}}} \hat{H}t}\, \hat{\mathcal{P}}_{\mathcal{A}} \ket{\boldsymbol{\pi}}}{\bra{\mathbf{1}}\hat{\mathcal{P}}_\mathcal{A}\ket{\boldsymbol{\pi}}}\biggr\rvert_{t >\tau_{\rm mol}}. \end{equation} In doing so, there are a few important differences from Eq.~\eqref{eq:tdvpdynamicsrate}. Most obviously, the TDVP dynamics includes the projection onto the MPS tangent space and $\ket{\boldsymbol{\pi}}$ must now be an MPS approximation to the steady-state distribution. Additionally, the meaning of $\hat{\mathcal{P}}_{\mathcal{A}}$ and $\hat{\mathcal{P}}_{\mathcal{B}}$ must be adapted to now define many-body analogs of the projectors onto the two metastable states. The single-voxel Schl\"{o}gl model had an $\mathcal{A}$ state defined as all microstates with occupancy $n$ above a threshold $q_A^*$. Similarly, $\mathcal{B}$ included all microstates with occupancy less than some other threshold $q_B^*$. The $L$-voxel Schl\"{o}gl model generalizes to involve hypercubic metastable states; $\mathcal{A}$ requires that each voxel satisfies $n \geq q_A^*$ and $\mathcal{B}$ requires that each voxel satisfies $n \leq q_B^*$. With these generalizations, the transition rate of Eq.~\eqref{eq:tdvpdynamicsrate} would measure the flux crossing into $\mathcal{B}$ through its boundary, but if we are to compare with a transition path theory rate, we must exclude re-crossing events. In other words, we want to include trajectories which exit $\mathcal{A}$ then some time later enter $\mathcal{B}$ without having re-entered $\mathcal{A}$, but we want to exclude the flux that leaves $\mathcal{B}$ and re-enters $\mathcal{B}$ some time later without having returned to $\mathcal{A}$. In Sec.~\ref{sec:rateconstants}, $k_{\mathcal{BA}}$ was sufficiently large that those re-crossings could be ignored, but as we generalize to handle smaller $k_{\mathcal{BA}}$, those recrossings can become significant. We note that the TDVP dynamics can just as well be executed with a modified absorbing-boundary-condition dynamics with effective Hamiltonian \begin{equation} \label{eq:modified} \tilde H = \tilde H^{\rm rxn} +\tilde H^{\rm diff}, \end{equation} where the reaction and diffusion Hamiltonians exclude events that exit $\mathcal{B}$. An explicit construction of that modified dynamics is provided in App.~\ref{app:modifieddynamics}, allowing us to directly remove the re-crossing events. \begin{figure*}[!ht] \centering \includegraphics[width=.9\textwidth]{Figures/Fig5.pdf} \caption{\label{fig:RateL} (a) The rate $k_{\mathcal{B}\mathcal{A}}$ for switching between metastable states of the $L$-voxel Schl\"{o}gl model is calculated using the SSA (Red line) and using TDVP (dashed lines) with varying bond-dimensions $\kappa$. For sufficiently large bond-dimensions, TDVP agrees with a SSA rate estimate built using $400$ transitions from $\mathcal{A}$ to $\mathcal{B}$ at each $L$. (b) As $L$ grows a larger bond-dimension is required to capture the dynamics. For $L=3$ we see that for $\kappa=5$ and $\kappa=6$ the flow of probability from $\mathcal{A}$ to $\mathcal{B}$ is unstable. As $\kappa$ grows, the solutions converge. For this system, $\kappa=20$ is stable up to $L=6$. While $\kappa=30$ gives stable solutions up to $L=8$. (c) The total computation time in hours to estimate $k_{\mathcal{B}\mathcal{A}}$ from both the SSA and TDVP. Despite using different bond-dimensions to estimate rates on larger lattices, we see that the total time to evolve the TDVP ensemble evolution scales sub-exponentially while the SSA scales exponentially.} \end{figure*} \section{Numerical Evaluation of the $L$-Voxel Schl\"{o}gl Model}\label{sec:Results} The formalism of the preceding sections has laid out a controllable approximation to extract rates between metastable states of RDMEs. Numerical experiments are necessary to evaluate whether that formalism can be practically useful. We set out to demonstrate that utility by computing the transition rates between high-occupancy and low-occupancy states of a $L$-voxel Schl\"{o}gl model as a function of $L$. The model is parameterized by the four stochastic rate coefficients discussed in Sec.~\ref{sect:EstRate}: $c_1 = 2.676, c_2 = 0.040, c_3 = 108.102,$ and $c_4 = 37.881$, chosen so the $L=1$ Schl\"{o}gl model will exhibit the bistablity reflected in Fig.~\ref{fig:ER}. With those chosen rates, $\mathcal{A}$ and $\mathcal{B}$ are identified by choosing $n_a = 15$ and $n_b = 25$, and the dynamics lends itself to a finite maximum-occupancy truncation of $M = 85$. By linking multiple voxels together in a 1D chain with a diffusive hopping rate $d = 8.2207$, it becomes rarer to switch between the two metastable states than it would be with a single isolated voxel. With multiple voxels, a switching event must flip the state of each voxel, and diffusion from a neighbor can stabilize the original state of each voxel. Increasing $L$ thus gives a way to systematically probe rarer transition rates. The most straightforward way to numerically detect the $L$-dependence of the transition rate is via sampling of trajectories. Realizations of the continuous-time RDME, subject to the $M = 85$ maximum voxel occupancy, were generated using the SSA. First, $5\times 10^5$ independent SSA realizations were initialized from a uniform distribution and propagated up to $t=5$ to ensure the steady state distribution was reached. Sampled configurations that fell within $\mathcal{A}$ were then uniformly drawn as an initial configuration for 400 independent trajectories which were evolved via the SSA until the trajectory first reached $\mathcal{B}$. From each of those $400$ trajectories, the time between the final exit of $\mathcal{A}$ and the first entry of $\mathcal{B}$, $\tau_{\mathcal{BA}}$ was recorded and the transition rate was computed from the mean of those waiting times: $k_{\mathcal{BA}} = \langle \tau_{\mathcal{BA}}\rangle^{-1}$. Fig.~\ref{fig:RateL}a (red line) shows that the rate exponentially decreases as $L$ is increased, a result consistent with the exponential dependence of rate on volume in the well-mixed Schl\"{o}gl model~\cite{vlysidis2018differences}. While the rate can be computed by trajectory sampling, the difficulty to do so (via brute force sampling) grows in proportion to the typical $\tau_{\mathcal{BA}}$, meaning the computational cost grows exponentially with $L$ (see Fig.~\ref{fig:RateL}c). The poor scaling of brute-force SSA sampled rate calculations has motivated so-called advanced sampling methods like FFS, which focus computational effort on the progress along a reaction coordinate. Fig.~\ref{fig:RateL} shows that the DP/TDVP approach provides a radically different strategy for circumventing the scaling problem, reproducing the SSA rates at lower computational expense. The SSA rate calculations can be made more accurate by collecting more trajectories, and the DP/TDVP calculations likewise offer a systematic way to improve accuracy by increasing the computational expense. The two most prominent ways to adjust that accuracy are through the time step $\delta t$ for the TDVP evolution and through the maximum bond dimension $\kappa$ for the MPS state. In this work, a time step of $\delta t = 10^{-4}$ was used throughout, and the trade-off between expense and accuracy was entirely tuned through the choice of $\kappa$. The single-site TDVP algorithm we used to propagate an MPS state does not alter the bond dimension of the MPS, so $\kappa$ was entirely set by the bond dimension of the initial MPS. We therefore needed to approximate the steady state $\ket{\boldsymbol{\pi}}$ with different values of $\kappa$. Approximations were generated by starting with an initially uniform state using a bond dimension $\kappa$ set to several different values between 5 and 30. Those uniform distributions were evolved with single-site TDVP for $10,000$ time-steps allowing for the natural dynamics of $\hat{H}$ to approach the limit $\ket{\boldsymbol{\pi}}$. We found that evolving that $10,000$ time steps was sufficient to approximate the steady-state distribution, as reflected by the average rate of change of the distribution. That is to say, $\bra{p_t}\ket{\dot{p}_t} = \bra{p_t}\hat{H}\ket{p_t} \approx 0$ by time $t = 1$, indicating proximity to the actual steady state, which satisfies $\ket{\boldsymbol{\dot \pi}} = \ket{\mathbf{0}}$. To get rates, the $\kappa$-dependent MPS approximation to the steady state was projected onto $\mathcal{A}$ and renormalized. That renormalized distribution was then propagated under single-site TDVP with the modified dynamics for another $10,000$ time steps, and projected onto $\mathcal{B}$ to add up the total probability to have reached $\mathcal{B}$. Note that these 10,000 steps are not nearly enough to reach the steady state $\ket{\boldsymbol{\pi}}$ because the slow rate of transitioning from $\mathcal{A}$ to $\mathcal{B}$ is a bottleneck, one that did not need to be surmounted to relax from the uniform distribution to $\ket{\boldsymbol{\pi}}$ in $10,000$ steps. As additional time steps of dynamics were taken, the growth of the transition flux from $\mathcal{A}$ to $\mathcal{B}$ was measured and plotted in Fig.~\ref{fig:RateL}b. That figure shows that at very short times ($t < \tau_{\rm mol}$), the flux from $\mathcal{A}$ to $\mathcal{B}$ has not reached a steady state value because there hasn't been enough time for any transitions. After that initial time $t_{\rm mol}$, the plot plateaus at a value corresponding to the rate $k_{\mathcal{BA}}$, a plateau that is reached orders of magnitude faster than the time for a typical SSA trajectory to make a transition. We find that if the bond dimension is insufficient (i.e., $\kappa = 5$ for $L = 3$), then the dynamics becomes unstable, reflected in erratic estimates for the conditional probability $P(\mathcal{B}_t|\mathcal{A}_0)$. As the bond dimension is grown to $\kappa = 9$, the estimates converge, allowing a stable rate to be extracted. Larger values of $L$ demand an even greater value of bond dimension. Fig.~\ref{fig:RateL}a shows that SSA rates agree with $\kappa = 20$ calculations up to $L = 6$, but $\kappa = 30$ was required to push up to $L = 8$. Provided that larger bond dimension is used, the DP/TDVP rates reproduce the SSA rates across five orders of magnitude, up to and beyond the point that the brute-force SSA rates are practical. Though the required bond dimension grows with $L$, Fig.~\ref{fig:RateL}c shows that the total time to estimate the rate grows sub-exponentially. The blue and black dashed lines in that plot are the total computational time to estimate the rate using TDVP with $\kappa=30$ and $\kappa=20$ respectively. Both SSA and TDVP were run using codes written in the Julia programming language \cite{bezanson2017julia} on equivalent CPUs. While the exact wall time, of course, depends on details of the hardware and the implementation, the plot clearly demonstrates the difference in scaling between brute-force SSA and DP/TDVP. An increase in bond dimension increases the computation time, but it does not fundamentally alter the favorable scaling in $L$. In addition to the favorable scaling at large $L$, we highlight that the DP/TDVP formalism allows us to extract many observables with similar ease. For example, Fig.~\ref{fig:ExpMol} shows $\langle n_l(t)\rangle = \bra{\mathbf{1}}\hat n_l\ket{p_t}$, the average number of molecules in site $l$ with $l$ ranging from 1 through 4 for an $L = 8$ site lattice. The state $\ket{p_t}$ was initialized in the normalized steady-state confined to $\mathcal{A}$, the high-occupation-number state. For $t < \tau_{\rm mol}$, a rapid loss of particles near the boundary of $\mathcal{A}$ leads to a sharp drop. After that $\tau_{\rm mol}$ timescale, the rate of decrease in $\langle n_l(t)\rangle$ becomes linear, reflecting the steady-state flux from $\mathcal{A}$ to $\mathcal{B}$. \begin{figure}[th] \centering \includegraphics[width=.4\textwidth]{Figures/Fig6.pdf} \caption{\label{fig:ExpMol}Average number of molecules for different lattice sites when $L=8$. Colors correspond to a different lattice sites. The expected number of molecules $\langle n_l(t)\rangle = \bra{\mathbf{1}}\hat n_l\ket{p_t}$ for sites $1$ to $\lfloor L/2\rfloor$ (shown) are mirrored by sites $\lfloor L/2+1\rfloor$ to $L$. The initial non-linear flow of probability out of $\mathcal{A}$ is shown by the steep drop in average number of molecules. The system passing $\tau_{mol}$ is shown by the transition to a linear change in particle number $\langle n_l(t)\rangle$.} \end{figure} \section{Discussion} In this work we have shown how macroscopic rates can be efficiently calculated from the microscopic dynamics of a RDME. We have illustrated that capability in a minimal toy model, the $L$-voxel Schl\"{o}gl model, with each voxel chemostatted such that it contains a single dynamical species (X). In actuality, the formalism and numerical methodology extend far beyond this initial demonstration. Any reaction-diffusion mechanism which can be written as a set of elementary reactions will have a DP representation, one which can involve raising and lowering operators for different particle types, not only for X. For example, Table~\ref{tbl:Rxn} shows the effective Hamiltonians for several common (unchemostatted) reactions: unimolecular, bimolecular, and auto-catalytic. Given the set of reactions, it is comparatively straightforward to derive the DP representation, and computational tools already exist to generate an MPO~\cite{schollwock2011density} and propagate via TDVP from that DP form of the effective Hamiltonian. \begin{table}[ht] \centering \begin{tabular}{c|c} Reaction & Hamiltonian \\ \hline \ce{X ->[$\tilde{c}$] Y} & $c[(y^\dagger - x^\dagger) x]$ \\[1pt] \ce{X +Y ->[$\tilde{c}$] Z} & $c[(z^\dagger - x^\dagger y^\dagger)x y]$ \\[1pt]\ce{X + Y ->[$\tilde{c}$] 2X} & $c[x^{\dagger }(x^\dagger - y^\dagger)x y]$ \\[1pt] \hline \end{tabular} \caption{\label{tbl:Rxn} Examples of the mapping between a common chemical reaction and the Doi-Peliti form of the contribution to an effective Hamiltonian. The procedure to derive such correspondences is systematic~\cite{schulz2005exact} and typically discussed for a Fock space with occupation number ranging from 0 to infinity. To truncate to a finite Fock space, one must additionally add a projector to each effective Hamiltonian, as in this work, to ensure conservation of probability. } \end{table} As those computational tools become increasingly well developed, one can imagine that the TN tools will offer generic ``turn the crank'' analysis of microscopic models of reaction-diffusion at the ensemble level. That vision motivates more future work at the interface of reaction-diffusion kinetics and tensor network methods. Here, we have approximated the microscopic distribution with an MPS, allowing us to capitalize on the efficient TDVP algorithm for MPS states. The downside is that an RDME with many dynamic species will not always be easily approximated by a 1D chain of tensors. If two chemical species are strongly correlated but far apart in the MPS structure, the MPS calculation will demand a very high bond dimension. It will be important for future work either to avoid the need for high bond dimension, for example by leveraging TN topologies which mirror that of an underlying CRN or to adapt the bond dimension through subspace expansion \cite{hubig2015strictly}. Those technical advances may eventually lay the groundwork to move the methodology from this Schl\"{o}gl study to the complex, biologically-motivated models where SSA realizations presently reign supreme. \section{Acknowledgments} We gratefully acknowledge Hadrien Vroylandt, Nils Strand, and Luis Pedro Garcia-Pintos for many insightful discussions. We are also grateful to Miles Stoudenmire, Matthew Fishman, Steven White, and other developers of ITensor, a library for implementing tensor network calculations, upon which this work was built. The material presented in this manuscript is based upon work supported by the National Science Foundation under Grant No.\ 2141385. TRG is thankful that Phill Geissler could see this work in a preliminary form, but is deeply saddened that Phill's much-too-soon passing prevented him from sharing more of his wisdom.

\section{Introduction}\label{sec:introduction} \subsection{Secure communication}\label{sec1:securecomm} For thousands of years, code-makers and code-breakers have been fighting for supremacy. With the recent rise of Internet of Things, cyber security has become a hot topic. Cyber warfare that can undermine the security of critical infrastructures, such as smart power grids and financial systems, threatens the well-being of individual countries and the global economy. In conventional cryptography, two distant parties, traditionally called Alice and Bob, share a communication channel and they would like to communicate privately in the presence of an eavesdropper, Eve. The Holy Grail of secure communication is information-theoretical security. It is known that one could achieve information-theoretically secure communication via the one-time-pad (OTP) method~\cite{vernam1926cipher}, if the two users, Alice and Bob, share a long random string that is kept secret from Eve. Note that, for the OTP scheme to be information-theoretically secure, it is important not to re-use the key~\cite{shannon1949communication}. That is to say that the key has to be as long as the message itself and can only be used once. How to distribute such a long key in the presence of Eve is called the key distribution problem. In fact, the key distribution problem is a central challenge in all kinds of encryption methods. In principle, \emph{all} conventional key distribution schemes that rely on classical physics and mathmatics can only provide computational security, because in classical physics, there is nothing to prevent an eavesdropper from copying the key during the key distribution process. Now, if Eve and Bob have the same key, whatever Bob can decrypt, Eve can decrypt too. Currently, the key distribution problem is often solved by public key cryptography. In public key cryptography, there are a pair of keys: a public key and a private key. An intended recipient Bob will publish the public key so that anyone, such as an intended sender Alice, can encrypt a message, called a plain text, with the public key and send the encrypted message, a cipher text, to Bob. On the other hand, only Bob with the private key can decrypt the cipher text to recover the plain text efficiently. The security of public key cryptography is based on computational assumptions. Given the public key, there is no efficient known algorithm for Eve to work out the private key or to recover the plain text, from the cipher text. For instance, the security of the best-known public key crypto-system, RSA~\cite{rivest1978method}, is based on the presumed hardness of factoring large integers. Unfortunately, public key cryptography is vulnerable to unanticipated advances in hardware and software. Moreover, in 1994, Peter Shor then at AT\&T invented an efficient quantum algorithm for factorization~\cite{shor1999polynomial}. For this reason, if a large scale quantum computer is ever constructed, much of conventional cryptography will fall apart! After more than two decades of intense theoretical and experimental efforts, primitive small scale quantum computers have already been built. Several big companies and a number of labs and start-ups are racing to build the world's first practical quantum computer. For instance, Google AI Quantum Laboratory\footnote{Google Q: research.google/teams/applied-science/quantum} has realized the quantum advantage (or supremacy) over state-of-the-art classical supercomputer for a specific computational task~\cite{arute2019quantum}, and plans to commercialize quantum computers within a few years~\cite{mohseni2017commercialize}; IBM Q has already put its sixteen-qubit quantum processor online for client use\footnote{IBM Q: www.research.ibm.com/ibm-q}; Rigetti has also provided the quantum cloud service\footnote{Rigetti: www.rigetti.com}; Chinese Academy of Sciences (CAS) and Alibaba have established the Quantum Computing Laboratory to advance the research of quantum computing\footnote{CAS-Alibaba: quantumcomputer.ac.cn/index.html}; Other companies, such as Intel, Microsoft, Baidu, Tencent, IonQ, Xanadu, Zapata and so forth, have also joined the international race to build a quantum computer. Moreover, China is building the National Laboratory for Quantum Information Science to support the revolutionary research in quantum information; The European Commission is planning to launch the flagship initiative on quantum technologies\footnote{ec.europa.eu/digital-single-market/en/news/quantum-europe-2017-towards-quantum-technology-flagship}; USA has already launched the National Quantum Initiative Act in 2018\footnote{www.congress.gov/bill/115th-congress/house-bill/6227}. All in all, the risk of successful construction of a quantum computer in the next decade can no longer be ignored. Note that some data such as our DNA data and health data need to kept secret for decades. This is called \emph{long-term security}. However, cryptographic standards could take many years to change. An eavesdropper intercepting encrypted data sent in 2019 may save them for decades as they wait for the future successful construction of a quantum computer. The eavesdropper could then retro-actively successfully crack an encryption scheme, therefore cryptographic standards need to consider potential future technological advances of the next few decades. For instance, Canadian Census Data is required to be kept confidential for 92 years\footnote{www12.statcan.ca/English/census01/Info/chief.cfm} or until 2111. To ensure such security, we need to predict the future technology in the next century. As a comparison, the first general-purpose electronic computer, ENIAC, was formally dedicated in 1946, which was less than 92 years ago. This meant that general-purpose electronic computers did not even exist 92 years ago. Therefore, if history is any guide, we think that it is not realistic for one to predict with any confidence what types of technology would exist 92 years from now. In 2015, the US National Security Agency (NSA) announced a plan for transition to quantum-safe crypto-systems. For instance, the US National Institute of Standards and Technology (NIST) has made a call for quantum-safe candidate algorithm nominations, which was due November 30, 2017\footnote{csrc.nist.gov/Projects/Post-Quantum-Cryptography}. Over the next few years, those candidate algorithms will be evaluated. Broadly speaking, there are two approaches to a quantum-safe encryption scheme. The first approach is to use conventional cryptography and to develop alternative public-key encryption schemes, such as hash-based or code-based encryption schemes, in which known quantum attacks such as Shor's algorithm~\cite{shor1999polynomial} do not apply. This approach is called \emph{post-quantum cryptography} and it has the advantages of being compatible with existing crypto infrastructure while having high key rates that are available over long distances. Recently, Google has performed a test deployment of a post-quantum crypto algorithm in Transport Layer Security (TLS)\footnote{security.googleblog.com/2016/07/experimenting-with-post-quantum}. One drawback of post-quantum algorithms is that those conventional algorithms are only shown to be secure against \emph{known} quantum attacks. There is always a possibility that some smart conventional or quantum physicist or computer scientist might one day come up with clever algorithms for breaking them efficiently. As said, this would lead to a retroactive security breach in future for data transmitted today with potentially disastrous consequences. The second approach is to use quantum cryptography~\cite{bennett1984quantum,ekert1991quantum}, particularly quantum key distribution (QKD). It has the advantage of promising information-theoretical security based on the fundamental laws of quantum physics, i.e., the security is independent of all future advances of algorithm or computational power. Note however that quantum cryptography cannot replicate all the functionalities of public key cryptography. In future, quantum cryptography is likely to be combined with the post-quantum cryptography to form the infrastructure of quantum-safe encryption scheme. For instance, the post-quantum cryptography can be used to perform the initial authentication. This authentication is only required in a short time, and once it is done, the generated QKD key will be secure. Therefore, we believe that the two approaches---post-quantum cryptography and quantum cryptography---are \emph{complementary} to each other (rather than mutually exclusive). \subsection{Quantum key distribution (QKD)}\label{sec1:QKD} The main goal of QKD is to achieve information-theoretical security by harnessing the laws of physics~\cite{bennett1984quantum,ekert1991quantum}. The quantum no-cloning theorem dictates that an unknown quantum state cannot be cloned reliably~\cite{Wootters1982,dieks1982communication}. If Alice distributes a key via quantum (e.g., single-photon) signals, because there is only a single copy of the key to begin with, there is \emph{no} way for Eve to clone the quantum state reliably to produce two copies of the same quantum state. Therefore, if Eve tries to eavesdrop in QKD, she will unavoidably introduce disturbance to the quantum signals, which will then be detected by the users, Alice and Bob. Alice and Bob can then simply discard such a key\footnote{Note that a key is simply a random string of numbers and if a key is aborted, it will not be used. So, there is no loss in security in aborting.} and try the key distribution process again. Note that an important advantage of QKD is that, since the communication is quantum, once a QKD session is over, there is \emph{no} classical transcript for Eve to keep. Therefore, an eavesdropper has to break a QKD session real-time or it will be secure forever. This is very different from conventional key distribution schemes. \subsubsection{BB84 protocol}\label{sec1:BB84} The best-known QKD scheme is the Bennett-Brassard-1984 (BB84) protocol~\cite{bennett1984quantum}. The BB84 protocol allows two users, Alice and Bob, who share a quantum channel (e.g., an optical fiber or free-space) and an authenticated conventional classical channel, to generate a secure key in the presence of an eavesdropper with unlimited quantum computing powers. In the BB84 protocol, a sequence of single photons carrying qubit states are sent by Alice to Bob through a quantum channel. A schematic diagram of the BB84 protocol is illustrated in Fig.~\ref{Fig:bb84}, and the steps of the protocol are listed in Box~\ref{tab:bb84}. \begin{tcolorbox}[title = {Box I.B.1: BB84 protocol.}] (1) For each signal, Alice randomly encodes a single photon with one of the four polarization states, namely, vertical, horizontal, 45-degree and 135-degree, and sends the photon through a quantum channel to Bob. \\ (2) For each signal, Bob chooses one of the two bases, rectilinear and diagonal, to perform a measurement on the polarization of a received photon. After detection, Alice and Bob publicly announces their basis choices through an authenticated conventional channel. \\ (3) Alice and Bob discard the polarization data that have been encoded and detected in different bases. They keep only those polarization data in the same basis. This remaining data forms the sifted key. Alice and Bob can choose a random sample of the sifted key bits and compare them to compute the quantum bit error rate (QBER).\\ (4) If the computed QBER is too high, they abort. Otherwise, they proceed with classical postprocessing such as error correction and privacy amplification to generate a secret key. \end{tcolorbox}\label{tab:bb84} \begin{figure} [htbp] \begin{center} \includegraphics[scale=0.6]{bb84.pdf} \end{center} \caption{Schematic diagram of the BB84 protocol. Alice encodes random bits on the polarization states of single photons. Bob randomly selects measurement bases, rectilinear ($+$) or diagonal ($\times$), to perform measurements using two detectors. They keep only those polarization data that have been encoded and detected in the same basis as the Sifted key, and perform additional classical postprocessing on the Sifted key to produce the final Secret key. \label{Fig:bb84}} \end{figure} \subsubsection{Intuition of security}\label{sec1:intuition} The quantum no-cloning theorem guarantees that Eve cannot copy the unknown quantum state sent by Alice reliably~\cite{Wootters1982,dieks1982communication}. Furthermore, a key feature in quantum mechanics is the complementarity between the two conjugate bases, rectilinear and diagonal. Since the two measurements corresponding to the two bases do {\it not} commute with each other, there is no way to measure the two observables simultaneously without disturbing the state. Therefore, Eve who tries to eavesdrop and extract information on the polarization data will inevitably introduce disturbance to the state. Bob on the other hand, with the authenticated classical channel, has a fundamental advantage over Eve because he can compare his basis choice with Alice and determine the QBER for data that is encoded and detected in the same basis. What happens if Eve attacks the quantum channel? A simple example of an eavesdropping strategy is the \emph{intercept-resend attack}~\cite{bennett1984quantum}. In this attack, for each photon sent from Alice, Eve performs a measurement in a randomly chosen basis and re-sends a new photon to Bob according to her measurement result. Let us focus on those cases when Alice and Bob happen to use the same basis since they will throw away the rest. If Eve happens to use the correct basis (50\%), then both she and Bob will decode Alice's bit value correctly. No error is introduced by Eve. On the other hand, if Eve uses the wrong basis (50\%), then both she and Bob will have random measurement results. This suggests that if Alice and Bob compare a subset of the sifted key, they will see a significant amount of errors. Here, for these bits, the photons will be passed on to Bob in the wrong basis, so regardless of Eve's measurement result, Bob will have a 50\% probability of measuring the opposite of Alice's bit value. In other words, Eve's attack will introduce 50\% QBER for half of the total bits, and thus a total of 25\% QBER. This example illustrates the basic principle behind QKD: \emph{Eve can only gain information at the cost of introducing disturbance}, which will expose her interference. \subsubsection{Overview of recent developments} \label{sec1:overview} \emph{Theoretical developments.} On the theoretical side, the first security proof of QKD was based on the uncertainty principle by Mayers~\cite{mayers2001unconditional}. Mayers's proof was put into a conceptually simple framework based on entanglement distillation by Lo and Chau~\cite{lo1999unconditional}, building on the earlier work of quantum privacy amplification \cite{deutsch1996quantum} and entanglement distillation \cite{Bennett1996Mixed}. Later on, Shor and Preskill employed the idea of the Calderbank-Shor-Steane (CSS) quantum error correcting code \cite{PhysRevA.54.1098,PhysRevA.54.4741} to simplify the entanglement-based proof to a prepare-and-measure protocol~\cite{shor2000simple}. See also~\cite{biham2000proof,Devetak2005Distillation,koashi2009simple} for security proofs of QKD. Rather interestingly, the rigorous definition of secure keys was presented afterwards in 2000s \cite{ben2005universal,Renner2005Security}, where the composable security definition in conventional cryptography \cite{Canetti2001} was introduced to quantum cryptography \cite{ben2005universal}. A further development was the security proof for the consideration of finite-key effects in a more rigorous manner~\cite{renner2008security,scarani2008quantum,tomamichel2012tight}. Device imperfections in practical systems were investigated in security analyses \cite{lutkenhaus2000security,inamori2007unconditional}, and the remarkable framework of the security analysis for realistic devices was established by Gottesman-Lo-L\"{u}tkenhaus-Preskill (GLLP) \cite{gottesman2004security}. Moreover, new protocols, such as the decoy-state~\cite{hwang2003quantum,lo2005decoy,wang2005beating}, differential-phase shift (DPS)~\cite{DPS:2002}, SARG-04~\cite{Scarani2004}, coherent-one-way (COW)~\cite{stucki2005fast}, measurement-device-independent (MDI)~\cite{lo2012measurement} [see also~\cite{braunstein2012side}] and round-robin DPS~\cite{sasaki2014practical}, were proposed to address the issues of device imperfections. In particular, the decoy-state protocol enables secure QKD with weak coherent pulses and the MDI protocol removes all side channels in the detection. Furthermore, device-independent QKD was proposed~\cite{mayers1998quantum,barrett2005no,acin2007device} to allow QKD with uncharaterized devices. Its security was proven effective against collective attacks~\cite{pironio2009device,masanes2011secure} [see also~\cite{hanggi2010efficient}] and later against general attacks~\cite{vazirani2014fully,Arnon2018Practical}. \emph{Experimental developments.} After more than two decades of efforts~\cite{gisin2002quantum,Lo2014NP}, QKD developments include the first laboratory demonstration performed in 1992 over 32.5-cm free space~\cite{bennett1992experimental}, to the recent landmark accomplishment of quantum satellite QKD experiment in 2017 over 1200 km by China~\cite{Liaosate}, and 7600 km in 2018 between China and Austria~\cite{liao2018satellite}. Note that this is a seven order of magnitude of improvement in terms of the distance of QKD. There are also on-going efforts on satellite-based quantum communications by Europe, USA, Canada, Japan, and Singapore~\cite{joshi2018space}. In fiber, the distance has been pushed to 500-km ultra-low loss fiber~\cite{Fang2019surpassing,chen2019sending}. In addition to long distances, high secret key rate is important for practical applications. Researchers have recently pushed the secret key rate of QKD from 1 Mbits/s over 50-km fiber~\cite{Luca2013} to more than 10 Mbits/s~\cite{Yuan:18,islam2017provably}. Commercial QKD systems are currently available on the market by several companies such as ID Quantique, Quantum CTek, Qasky and Toshiba Europe. Several institutes, e.g., European Telecommunications Standards Institute (ETSI), International Organization for Standardization (ISO), and International Telecommunication Union (ITU), have made great efforts to address the standardization issues in QKD. Besides point-to-point link, a number of field-test QKD networks have been conducted in USA~\cite{elliott2005current}, Europe~\cite{peev2009secoqc,stucki2011long}, Japan~\cite{sasaki2011field}, China~\cite{chen2009field,chen2010metropolitan,wang2010field}, UK~\cite{dynes2019cambridge} and so forth. Based on trustful relays\footnote{In trusted-relay scenario, Alice and Bob respectively share a secret key with a relay in the middle, and then the relay announces the XOR results of both keys publicly. With the announced result, Alice and Bob can get each other's key via the XOR with her/his own key. The negative side for this method is that the relay must be trustful. However, the positive side is reducing the cost and complexity as compared to the all-connected point-to-point links, and extending the transmission distance.}, remote users can be connected beyond point-to-point links. Recently, China has successfully completed the 2000-km-long fiber-optic backbone link between Beijing to Shanghai~\cite{Yuao2019}. UK has launched the Quantum Communications Hub project that aims to build quantum networks in England\footnote{www.quantumcommshub.net/about-us/}. US is deploying their first dark fiber quantum network connecting Washington DC to Boston over 800-km\footnote{techcrunch.com/2018/10/25/new-plans-aim-to-deploy-the-first-u-s-quantum-network-from-boston-to-washington-dc/}. Overall, QKD is already mature for several real-life applications~\cite{qiu2014quantum}. For instance, QKD was used to encrypt security communications in the 2007 Swiss election and the 2010 World Cup. In China, QKD is being widely used to ensure long-term security for numerous users in government, financial and energy industry~\cite{Yuao2019}, including the People's Bank of China, the China Banking Regulatory Commission, and the Industrial and Commercial Bank of China. Figure~\ref{Fig:backbone} shows a schematic diagram of the space-ground integrated quantum network~\cite{Yuao2019}, constructed already in China, which spans more than 2000~km coverage area and has more than 600 QKD links. \begin{figure*} \begin{center} \includegraphics[scale=0.5,angle=0]{backbone.pdf} \end{center} \caption{(Color online) Schematic diagram of the space-ground integrated quantum network in China~\cite{Yuao2019}, consisting of four quantum metropolitan area networks in the cities of Beijing, Jinan, Shanghai, and Heifei, a backbone network over 2000 km, and ground-satellite links. There are three types of nodes in the network: user nodes, all-pass optical switches, and trusted relays. The backbone network is connected by trusted intermediate relays. The satellite is connected to a ground satellite station near Beijing, which can provide ultralong distance communications~\cite{liao2018satellite}.}\label{Fig:backbone} \end{figure*} \subsection{Focus of this review}\label{sec1:motivation} In the Code Book by Simon Singh~\cite{singh2000code}, the author boldly proclaimed that quantum cryptography achieves the Holy Grail of cryptography by offering unconditional security. Therefore, quantum cryptography presents the final stage of evolution of cryptography. After quantum cryptography, cryptography will no longer continue to evolve. Is this really true? In principle, QKD promises unconditional security based on the laws of physics. In practice, however, the realistic devices display imperfections, which might seldom conform to idealized theoretical models used in the security analysis by theorists. The deviations might also be vulnerable by some special attacks, i.e., quantum hacking. For this reason, an arms race has been going on in quantum cryptography among quantum code-makers and quantum code-breakers. The main goal is to assess the deviations between the system and the ideal, thus establishing the \emph{practical security} for real QKD systems. Table~\ref{Tab1} summarizes the quantum hacking strategies developed in the last two decades. See also~\cite{jain2016attacks} for an earlier review on the subject. Right after the QKD security proofs, in which ideal devices were presented, a well-known hacking strategy was proposed --- photon number splitting (PNS) attack~\cite{lutkenhaus2000security,brassard2000limitations} that targets practical QKD source. The source device imperfection severely undermines the performance of a QKD system, typically below 30-km fiber~\cite{lutkenhaus2000security,gottesman2004security,Ma2006low}. In order to close this side channel for a QKD source, the decoy state method has been proposed by quantum code-makers to make QKD practical with standard weak coherent pulses (WCPs) that are generated by attenuated lasers~\cite{hwang2003quantum,lo2005decoy,wang2005beating}. Decoy-state QKD presents dramatic performance improvement over the conventional security proofs~\cite{gottesman2004security}, and it has become a standard technique in current QKD experiments. Table~\ref{Tab2} provides a list of decoy-state QKD experiments. \begin{table*}[ht!] \scriptsize \caption{List of quantum hacking strategies.} \label{Tab1} \begin{tabular}{ l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l } \hline \hline \textbf{Attack} & \textbf{Source/Detection} & \textbf{Target component} & \textbf{Manner} & \textbf{Year} \\ \hline Photon-number-splitting~\cite{lutkenhaus2000security,brassard2000limitations} & Source & WCP (multi-photons) & Theory & 2000 \\ Detector fluorescence~\cite{kurtsiefer2001breakdown} & Detection & Detector & Theory & 2001 \\ Faked-state~\cite{makarov2005faked,Makarov2006} & Detection & Detector & Theory & 2005 \\ Trojan horse~\cite{vakhitov2001large,gisin2006trojan} & Source\&Detection & Backflection light & Theory & 2006 \\ Time shift~\cite{Qi2007,Zhao2008} & Detection & Detector & Experiment$^{*}$ & 2007 \\ Time side-channel~\cite{lamas2007breaking} & Detection & Timing information & Experiment & 2007 \\ Phase remapping~\cite{Fred2007,xu2010} & Source & Phase modulator & Experiment$^{*}$ & 2010 \\ Detector blinding~\cite{makarov2009controlling,Lars2010} & Detection & Detector & Experiment$^{*}$ & 2010 \\ Detector blinding~\cite{gerhardt2011full,gerhardt2011experimentally} & Detection & Detector & Experiment & 2011 \\ Detector control~\cite{lydersen2011controlling,wiechers2011after} & Detection & Detector & Experiments & 2011 \\ Faraday mirror~\cite{Sun2011}& Source & Faraday mirror & Theory & 2011\\ Wavelength~\cite{Li2011,huang2013quantum}& Detection & Beam-splitter & Experiment & 2011\\ Dead-time~\cite{Wei2011} & Detection & Detector & Experiment & 2011\\ Channel calibration~\cite{Jain2011}& Detection & Detector & Experiment$^{*}$ & 2011\\ Intensity~\cite{Jiang2012,Shihan2015} & Source & Intensity modulator & Experiment & 2012 \\ Phase information~\cite{Sun2012,Tang2013,Sun2015} & Source & Phase randomization & Experiment & 2012 \\ Memory attacks~\cite{barrett2013memory} & Detection & Classical memory & Theory & 2013 \\ Local oscillator~\cite{jouguet2013preventing,ma2013local}$^{**}$ & Detection & Local oscillator & Experiment & 2013 \\ Trojan horse~\cite{jain2014trojan,jain2015risk} & Source\&Detection & Backflection light & Experiment & 2014 \\ Laser damage~\cite{Audun2014,Vadim2016} & Detection & Detector & Experiment & 2014\\ Laser seeding~\cite{Sun2015} & Source & Laser phase/intensity & Experiment & 2015 \\ Spatial mismatch~~\cite{sajeed2015security,chaiwongkhot2019eavesdropper} & Detection & Detector & Experiment & 2015 \\ Detector saturation~\cite{qin2016quantum}$^{**}$ & Detection & Homodyne detector & Experiment & 2016 \\ Covert channels~\cite{curty2017quantum} & Detection & Classical memory & Theory & 2017 \\ Pattern effect~\cite{yoshino2018quantum} & Source & Intensity modulator & Experiment & 2018 \\ Detector control~\cite{qian2018hacking} & Detection & Detector & Experiment & 2018 \\ Laser seeding~\cite{Sun2015,huang2019laser,pang2019hacking} & Source & Laser & Experiment & 2019 \\ Polarization shift~\cite{wei2019implementation} & Detection & SNSPD & Experiment & 2019 \\ \hline \hline $*$Demonstration on commercial QKD system \\ $**$Continuous-variable QKD \\ \end{tabular} \end{table*} \begin{table*}[htb!] \scriptsize \caption{List of decoy-state QKD experiments and their performance.} \label{Tab2} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l } \hline \hline \textbf{Reference} & \textbf{Clock rate} & \textbf{Encoding} & \textbf{Channel} & \textbf{Maximal distance} & \textbf{Key rate (bps)} & \textbf{Year} \\ \hline \cite{Yi2006,zhao2006simulation} & 5MHz & Phase & Fiber & 60km & 422.5 & 2006 \\ \cite{Peng2007} & 2.5MHz & Polarisation & Fiber & 102km & 8.1 & 2007\\ \cite{rosenberg2007long} & 2.5MHz & Phase & Fiber & 107km & 14.5 & 2007\\ \cite{Tobias2007} & 10MHz & Polarisation & Free-space & 144km & 12.8$^{\ast}$ & 2007\\ \cite{yuan2007unconditionally} & 7.1MHz & Phase & Fiber & 25.3km & 5.5K & 2007\\ \cite{Yin2008} & 1MHz & Phase & Fiber & 123.6km & 1.0 & 2008\\ \cite{Qin2008}$^{\ast\ast}$ & 0.65MHz & Phase & Fiber & 25km & 0.9 & 2008\\ \cite{Dixon2008} & 1GHz & Phase & Fiber & 100.8km & 10.1K & 2008\\ \cite{peev2009secoqc} & 7MHz & Phase & Fiber network & 33km & 3.1K & 2009\\ \cite{Rose2009} & 10MHz & Phase & Fiber & 135km & 0.2 &2009\\ \cite{Yuan2009} &1.036GHz & Phase & Fiber & 100Km & 10.1K & 2009\\ \cite{chen2009field} & 4MHz & Phase& Fiber network & 20km & 1.5K & 2009\\ \cite{liu2010decoy} & 320MHz & Polarisation &Fiber & 200km & 15.0 & 2010 \\ \cite{chen2010metropolitan} & 320MHz & Polarisation & Fiber network & 130km & 0.2K &2010 \\ \cite{sasaki2011field} & 1GHz & Phase & Fiber network & 45km & 304.0K & 2011\\ \cite{Wang2013} & 100MHz & Polarisation & Free space & 96km & 48.0 &2013\\ \cite{Bernd2013} & 125MHz & Phase & Fiber network & 19.9km &43.1K & 2013\\ \cite{Luca2013} & 1GHz & Phase & Fiber & 80km &120.0K &2013\\ \cite{frohlich2017long} & 1GHz& Phase & Fiber & 240km$^{\ddagger}$ & 8.4 & 2017\\ \cite{Liaosate} & 100MHz & Polarisation & Free space& 1200km & 1.1K & 2017\\ \cite{Yuan:18} & 1GHz & Phase & Fiber & 2dB & 13.7M & 2018\\ \cite{boaron2018secure} & 2.5GHz & Time-bin & Fiber & 421km$^{\ddagger}$ & 6.5 & 2018\\ \hline \hline $^{\ast}$Asymptotic key rate\\ $^{\ast\ast}$Heralded single-photon source\\ $^{\ddagger}$Ultralow loss fiber \\ \end{tabular} \end{table*} After the decoy-state method, however, various quantum hacking attacks have been performed by quantum code-breakers against other components in practical QKD systems (see Table~\ref{Tab1}). To counter those attacks, a few important concepts have been proposed by quantum code-makers. One practical counter-measure against quantum hacking is the measurement-device-independent quantum key distribution (MDI-QKD)~\cite{lo2012measurement} [see also~\cite{braunstein2012side}]. MDI-QKD completely removes all security loopholes in the detection system and ensures a QKD network security with \emph{untrusted} relays. It is practical with current technology. Table~\ref{Tab3} summarizes the MDI-QKD experiments after its invention. In addition, an efficient version of MDI-QKD --- twin-field (TF) QKD --- has the potential to greatly extend the secure distance. Table~\ref{Tab:tfqkd} summarizes the recent TF-QKD experiments. Also, we note some of the recent developments of continuous-variable (CV) QKD (see Table~\ref{Tab:cvqkd}), chip-based QKD (see Table~\ref{Tab:chipqkd}), and other QKD protocols and implementations (see Table~\ref{Tab:otherqkd}). We also summarize a list of some developments of recent quantum cryptographic protocols, besides QKD (see Table~\ref{Tab5}). \begin{table*}[ht!] \scriptsize \caption{List of MDI-QKD experiments and their performance.} \label{Tab3} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l } \hline \hline \textbf{Reference} & \textbf{Clock rate} & \textbf{Encoding} & \textbf{Distance/loss} & \textbf{Key rate (bps)} & \textbf{Year} & \textbf{Notes} \\ \hline \cite{Rub2013}$^{\ddagger}$ & 2MHz & Time-bin & 81.6km & $0.24^{\ast}$ & 2013 & Field-installed fiber \\ \cite{Liu2013} & 1MHz & Time-bin & 50km & $0.12$ & 2013 & First complete demonstration \\ \cite{Silva2013}$^{\ddagger}$ & 1MHz & Polarisation & 17km & $1.04^{\ast}$ & 2013 & Multiplexed synchronization\\ \cite{Tang2014} & 0.5MHz & Polarisation & 10km & $4.7\times10^{-3}$ & 2014 & Active phase randomization\\ \cite{yanlin2014} & 75MHz & Time-bin & 200km & $0.02$ & 2014 & Fully automatic system\\ \cite{Yanlin2015} & 75MHz & Time-bin& 30km & $16.9$ & 2015 & Field-installed fiber\\ \cite{Wang2015} & 1MHz & Time-bin & 20km & $8.3^{\ast}$ & 2015 & Phase reference free\\ \cite{valivarthi2015measurement} & 250MHz & Time-bin & 60dB & $5\times10^{-2}$ & 2015 & Test in various configurations\\ \cite{pirandola2015high}$^{\ddagger}$ & 10.5MHz & Phase & 4dB & 0.1 & 2015 & Continuous variable\\ \cite{Yanlin2016} & 75MHz & Time-bin & 55km & $16.5$ & 2016 & First fiber network \\ \cite{Yin2016} & 75MHz & Time-bin & 404km & $3.2\times10^{-4}$ & 2016 & Longest distance \\ \cite{tang2016experimental} & 10MHz & Polarisation & 40km & 10 & 2016 & Include modulation errors \\ \cite{Comandar}$^{\ddagger}$ & 1GHz & Polarisation & 102km & 4.6K & 2016 & High repetition rate\\ \cite{kaneda2017quantum}$^{\ddagger}$ & 1MHz & Time-bin & 14dB & 0.85 & 2017 & Heralded single-photon source\\ \cite{wang2017measurement} & 1MHz & Time-bin & 20km & $6.3\times10^{-3}$ & 2017 & Stable against polarization change \\ \cite{valivarthi2017cost} & 20MHz & Time-bin & 80km & $100$ & 2017 & Cost-effective implementation \\ \cite{liu2018polarization} & 50MHz & Time-bin & 160km & $2.6^{\ast}$ & 2018 & Phase reference free \\ \cite{liu2018experimental} & 75MHz & Time-bin & 100km & $14.5$ & 2019 & Asymmetric channels \\ \cite{wei2019high} & 1.25GHz & Polarisation & 20.4 dB & 6.2K & 2019 & Highest repetition/key rate\\ \hline \hline $^{\ast}$Asymptotic key rate \\ $^{\ddagger}$No random modulations \\ \end{tabular} \end{table*} \begin{table}[ht!] \scriptsize \caption{List of TF-QKD experiments.} \label{Tab:tfqkd} \begin{tabular}{l @{\hspace{0.2cm}} l @{\hspace{0.3cm}} l @{\hspace{0.2cm}} l @{\hspace{0.2cm}} l } \hline \hline \textbf{Reference} & \textbf{Distance/loss} & \textbf{Key rate (bps)} & \textbf{Year} \\ \hline \cite{minder2019experimental} & 90.8dB & 0.045$^\ast$ & 2019 \\ \hline \cite{wang2019beating} & 300km & 2.01k$^\ast$ & 2019 \\ \hline \cite{LiuTF2019} & 300km & 39.2 & 2019 \\ \hline \cite{ZhongTF2019} & 55.1dB & 25.6$^\ast$ & 2019 \\ \hline \cite{Fang2019surpassing} & 502km$^{\ddagger}$ & 0.118 & 2019 \\ \hline \cite{chen2019sending} & 509km$^{\ddagger}$ & 0.269 & 2019 \\ \hline \hline $^{\ast}$Asymptotic key rate \\ $^{\ddagger}$Ultra-low loss fiber \\ \end{tabular} \end{table} \begin{table*}[ht!] \scriptsize \caption{List of some recent CV-QKD experiments and their performance.} \label{Tab:cvqkd} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l } \hline \hline \textbf{Reference} & \textbf{Clock rate} & \textbf{Distance/loss} & \textbf{Key rate (bps)} & \textbf{Year} & \textbf{Notes} \\ \hline \cite{jouguet2013experimental} & 1MHz & 80.5km & $\sim$250 & 2013 & Full implementation \\ \hline \cite{qi2015generating} &25MHz & --- & --- & 2015 & Local LO \\ \hline \cite{soh2015self} & 250KHz & --- & --- & 2015 & Local LO \\ \hline \cite{huang2015high} & 100MHz & 25km & 100K & 2015 & Local LO \\ \hline \cite{pirandola2015high} & 10.5MHz & 4dB & 0.1 & 2015 & CV MDI-QKD \\ \hline \cite{Huang:15} & 50MHz & 25km & $\sim$1M & 2015 & High key rate \\ \hline \cite{Kumar_2015} & 1MHz & 75km &490 & 2015 & Coexistence with classical \\ \hline \cite{Zhang_2019} & 5MHz & 50km& 5.8K & 2019 & Field test \\ \hline \cite{zhang2020long} & 5MHz & 202.8km$^{\ddagger}$ & 6.2 & 2020 & Long distance \\ \hline \hline $^{\ddagger}$Ultra-low loss fiber \\ \end{tabular} \end{table*} \begin{table*}[ht!] \scriptsize \caption{List of chip-based QKD experiments.} \label{Tab:chipqkd} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l } \hline \hline \textbf{Reference} & \textbf{Clock rate} & \textbf{Distance/loss} & \textbf{Key rate (bps)} & \textbf{Year} & \textbf{Notes} \\ \hline \cite{ma2016silicon} &10MHz & 5km & 0.95K & 2016 & Silicon, decoy-BB84 \\ \hline \cite{sibson2017chip} & 1.72GHz & 4dB & 565K & 2017 & InP, DPS \\ \hline \cite{sibson2017integrated} & 1.72GHz & 20km & 916K & 2017 & Silicon, COW \\ \hline \cite{bunandar2018metropolitan} & 625MHz & 43km & 157K & 2018 & Silicon, decoy-BB84 \\ \hline \cite{ding2017high} & 5KHz & 4dB & $\sim$7.5 & 2018 & Silicon, high-dimension \\ \hline \cite{zhang2019integrated} & 1MHz & 16dB & 0.14K & 2019 & Silicon, CV-QKD \\ \hline \cite{paraiso2019modulator} & 1GHz & 20dB & 270K & 2019 & InP, modulator-free \\ \hline \cite{wei2019high} & 1.25GHz & 140km & 497 & 2019 & Silicon, MDI-QKD \\ \hline \hline \end{tabular} \end{table*} \begin{table*}[ht!] \scriptsize \caption{List of recent experiments of other QKD protocols.} \label{Tab:otherqkd} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l } \hline \hline \textbf{Reference} & \textbf{Clock rate} & \textbf{Distance/loss} & \textbf{Key rate (bps)} & \textbf{Year} \\ \hline Quantum access network~\cite{Bernd2013} & 125MHz & 19.9km & 259 & 2013 \\ \hline Centric network~\cite{hughes2013network} & 10MHz & 50km & --- & 2013 \\ \hline RR-DPS~\cite{guan2015experimental} & 500MHz & 53km & $\sim$118.0 & 2015 \\ \hline RR-DPS~\cite{takesue2015experimental} & 2GHz & 20km & 2.0K & 2015 \\ \hline RR-DPS~\cite{wang2015experimental} & 1GHz & 90km & $\sim$800 & 2015 \\ \hline RR-DPS~\cite{li2016experimental} & 10KHz & 18dB & 15.5 & 2016 \\ \hline High-dimension~\cite{lee2014entanglement} & 8.3MHz & --- & 456 & 2014 \\ \hline High-dimension~\cite{zhong2015photon} & CW & 20km & 2.7M & 2015 \\ \hline High-dimension~\cite{mirhosseini2015high} & 4KHz & --- & 6.5 & 2015 \\ \hline High-dimension~\cite{sit2017high} & --- & 0.3km & $\sim$30K & 2017 \\ \hline High-dimension~\cite{islam2017provably} & 2.5GHz & 16.6dB & 1.07M & 2017 \\ \hline Coherent-one-way~\cite{korzh2015provably} & 625MHz & 307km & 3.2 & 2015 \\ \hline Modulator-free~\cite{yuan2016directly} & 1GHz & 40dB & $\sim$10 & 2018 \\ \hline \hline \end{tabular} \end{table*} \begin{table*}[ht!] \scriptsize \caption{List of recent developments of other quantum cryptographic protocols beyond QKD.} \label{Tab5} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l} \hline \hline \textbf{Protocol} & \textbf{Theory/Experiment} & \textbf{Notes} \\ \hline Noisy quantum storage~\cite{konig2012unconditional,wehner2008cryptography,damgaard2008cryptography} & Theory & Unconditional security \\ Oblivious transfer~\cite{erven2014experimental} & Experiment & Noisy-storage model \\ Bit commitment~\cite{ng2012experimental} & Experiment & Noisy-storage model \\ Bit commitment~\cite{kent2012unconditionally} & Theory & Relativistic assumption \\ Bit commitment~\cite{lunghi2013experimental,liu2014experimental} & Experiment & Relativistic assumption \\ Bit commitment~\cite{lunghi2015practical,chakraborty2015arbitrarily,verbanis2016hour} & Experiment & Long commitment time \\ Digital signature~\cite{clarke2012experimental} & Experiment & First demonstration \\ Digital signature~\cite{collins2014realization,dunjko2014quantum} & Experiment & No quantum memory \\ Digital signature~\cite{donaldson2016experimental,yin2017experimental} & Experiment & Insecure channel \\ Coin flipping~\cite{berlin2011experimental,pappa2014experimental} & Experiment & Loss tolerance \\ Data locking~\cite{Fawzi2013low,seth2013quantum,lupo2014robust} & Theory & Loss tolerance \\ Data locking~\cite{liu2016experimental,lum2016quantum} & Experiment & Loss tolerance \\ Blind quantum computing~\cite{broadbent2009universal,barz2012demonstration} & Theory,Experiment & No quantum memory \\ Blind quantum computing~\cite{reichardt2013classical,huang2017experimental} & Theory,Experiment & Classical clients \\ \hline \hline \end{tabular} \end{table*} Note that the side channels are common problems to \emph{any} cryptosystems, i.e., not only to quantum cryptography but also to conventional cryptographic systems. For instance, the power consumption of the CPU performing encryption and decryption and the timing of the signals are common side channels, which can threaten implementations of both quantum and conventional cryptographic systems~\cite{kocher1999differential,brumley2005remote}. Therefore, closing the side channels are essentially required in all cryptographic technologies. It is only through painstaking battle-testing that the security of a practical crypto-system could be established with confidence. The arms race between code-makers and code-breakers will continue in cryptographic systems. Nonetheless, QKD is a physics-based crypto system and its security is working on the physical layer. Compared to the conventional mathematical-based cryptography, QKD can provide an accurate description of the physical realization of a cryptographic system, and the security can be proved based on this description. More importantly, QKD has the fundamental advantage of promising information-theoretical security, which is independent of all future advances of computational power. Furthermore, the recent advances, such as MDI-QKD, have closed the critical side channels in the detection of physical implementations, paving the way for secure QKD with realistic devices. Therefore, we believe that QKD does represent an important chapter in the history of code-making. We hope that QKD will play an important role in the quantum-safe encryption infrastructure for real applications, and it will bring us one step closer to the dream of information-theoretical security. \subsection{Outline of this review}\label{sec1:scope} This review will focus mainly on the practical security of realistic QKD systems. We begin with a discussion of security analysis in Section~\ref{Sc:Security} and the basic implementation of QKD in Section~\ref{Sc:Implement}. In Section~\ref{sec:3}, we review various quantum hacking attacks against QKD implementations. In Section~\ref{sec:4}, we review the security of a practical QKD source. In particular, we focus on the decoy-state protocol which is a standard method for secure QKD with attenuated lasers. In Section~\ref{sec:5}, we turn to detector security. We primarily review the MDI-QKD protocol and how it automatically foils all attacks on the detection system. In Section \ref{sec:CVQKD}, we review the developments of CV-QKD schemes and their practical security aspects. Section~\ref{sec:6} contains a review of other quantum cryptographic protocols. In Section~\ref{sec:9}, we present some concluding remarks. For those readers who want to learn further basics of QKD, we refer to the two earlier reviews published in Review of Modern Physics, one by Gisin \emph{et al.} that introduces the basic experimental elements and systems~\cite{gisin2002quantum} and the other one by Scarani \emph{et al.} that discusses the basic security analysis tools of various QKD protocols~\cite{scarani2009security}. An early review on the first stage of development of QKD can be found in the book by Sergienko~\cite{sergienko2018quantum}. An earlier review on quantum attacks can be seen in~\cite{jain2016attacks}. A brief overview of the implementation security of QKD can be found in a survey article in~\cite{Lo2014NP} and an ETSI white paper\footnote{https://www.etsi.org/images/files/ETSIWhitePapers/} by Lucamarini \emph{et al.} A short overview of the practical challenges associated with QKD can be found in~\cite{Diamanti2016}. Moreover, the entropy uncertainty relation, an important tool to analyze the security of QKD, can be seen in~\cite{coles2017entropic}, and the quantum random number generator, a basic element in a practical QKD system, can be found in~\cite{Ma2016,herrero2017quantum}. A review on various techniques of single-photon detectors can be seen in~\cite{Hadfield:2009,zhang2015advances}. Furthermore, we may not cover too much on some important topics, but we refer the readers to other review articles on the topics of CV-QKD~\cite{weedbrook2012gaussian,diamanti2015distributing,laudenbach2018continuous}, high-dimensional QKD~\cite{xavier2020quantum}, quantum repeaters~\cite{sangouard2011quantum,pan2012multiphoton,munro2015inside}, quantum Internet~\cite{kimble2008quantum,wehner2018quantum}, Bell nonlocality and device-independent protocols~\cite{brunner2014bell}, and blind quantum computing~\cite{fitzsimons2017private}. These related review articles are summarized in Table~\ref{Tab6}. \begin{table*}[ht!] \scriptsize \caption{List of related reviews to QKD.} \label{Tab6} \begin{tabular}{l @{\hspace{1.2cm}} l } \hline \hline \textbf{Reference} & \textbf{Subject} \\ \hline \cite{gisin2002quantum} & Experimental basics of QKD \\ \cite{scarani2009security} & Theoretical basics of QKD \\ \cite{Lo2014NP,Diamanti2016,Zhang:18} & Practical challenges of QKD \\ \cite{jain2016attacks} & Quantum hacking attacks \\ \cite{Xu:IEEE} & Measurement-device-independent QKD \\ \cite{Hadfield:2009,zhang2015advances} & Single-photon detector \\ \cite{Ma2016,herrero2017quantum} & Quantum random number generator \\ \cite{coles2017entropic} & Entropy uncertainty relation \\ \cite{weedbrook2012gaussian,diamanti2015distributing,laudenbach2018continuous} & Continuous-variable QKD \\ \cite{sangouard2011quantum,pan2012multiphoton,munro2015inside} & Quantum repeaters \\ \cite{kimble2008quantum,wehner2018quantum} & Quantum internet \\ \cite{brunner2014bell} & Bell nonlocality/device-independent QKD \\ \cite{fitzsimons2017private} & Blind quantum computing \\ \cite{xavier2020quantum} & High-dimensional QKD \\ \hline \hline \end{tabular} \end{table*} \section{Security analysis} \label{Sc:Security} We review the security aspects of QKD, including the security definition, various security proofs and implementation assumptions. We present a general framework to address device imperfections in security analysis. While we mainly focus on the widely implemented BB84 protocol, most of the results can be extended to other QKD protocols. We shall leave the MDI-QKD case in Section~\ref{sec:5:mdi}, and the DI-QKD case in Section~\ref{sec:diQKD}. \subsection{Security definition} \label{Sc:SecurityDef} To prove the security of QKD, one needs to define the security criteria first. Ideally, a secure key satisfies two requirements. First, the key bit strings possessed by Alice and Bob need to be identical, i.e., be \emph{correct}. Second, the key bit string should be uniformly distributed to anyone (say Eve) other than Alice and Bob, i.e., be \emph{secret}. Due to practical issues, such as the finite data size and non-ideal error correction, Alice and Bob cannot generate an ideal key. In reality, it is reasonable to allow the key to have a small failure probability $\epsilon$. For some $\epsilon_{\text{cor}}$ and $\epsilon_{\text{sec}}$, we say that the QKD protocol is $\epsilon$-secure with $\epsilon=\epsilon_{\text{cor}}+\epsilon_{\text{sec}}$, if it is $\epsilon_{\text{cor}}$-correct and $\epsilon_{\text{sec}}$-secret~\cite{ben2005universal,Renner2005Security}. We define $K_A$ and $K_B$ (with the same length $m$) to be the key bit strings obtained by Alice and Bob, respectively. The secret key can be correlated to a quantum state $\rho_E$ held by Eve. The joint state $\rho_{ABE}$ is a \emph{classical-classical-quantum (c-c-q)} state, \begin{equation} \label{eq:rho_ABE} \rho_{ABE} = \sum_{k_A, k_B} Pr(k_A,k_B) \ket{k_A}\bra{k_A} \otimes \ket{k_B}\bra{k_B} \otimes \rho_E^{(k_A, k_B)}, \end{equation} where $k_A, k_B \in \{0,1\}^{m}$ are the bit values. In particular, an ideal key state held by Alice and Bob is described by the private state, \begin{equation} \label{eq:rho_ideal} \rho_{ABE}^{ideal} = 2^{-m} \sum_{k} \ket{k}_A\bra{k} \otimes \ket{k}_B\bra{k} \otimes \rho_E, \end{equation} where $k_A=k_B=k$ implies that Alice and Bob hold the same string, and $\rho_E$ is independent of $k$, i.e., Eve has no information on the key string variable $K$. A QKD protocol is defined to be $\epsilon_{\text{cor}}$-correct, if the probability distribution $Pr(k_A, k_B)$ of the final state $\rho_{ABE}$ in Eq.~\eqref{eq:rho_ABE} satisfies, \begin{equation} Pr(k_A \neq k_B) \leq \epsilon_{cor}. \end{equation} A QKD protocol is defined to be $\epsilon_{sec}$-secret~\cite{Renner2005Security}, if the state $\rho_{AE}$ is close in trace distance to the single-party private state $\rho_{AE}^{ideal}$ \begin{equation} \min_{\rho_E} \dfrac{1}{2}(1-p_{\text{abort}})|| \rho_{AE} - \rho_{AE}^{ideal} ||_1 \leq \epsilon_{sec}, \end{equation} where $p_{\text{abort}}$ is the probability that the protocol aborts, $\rho_{AE}^{ideal}\equiv 2^{-m} \sum_{s} \ket{s}_A\bra{s} \otimes \rho_E$, and $||A||_1 \equiv Tr[\sqrt{A^\dag A}]$ is the trace norm. It turns out that the security definition from the trace-distance metric owns a composable security property \cite{ben2005universal,Renner2005Security}. In general, following the definition from Ben-Or \emph{et al.}'s work \cite{ben2005universal}, a QKD protocol can be defined to be $\epsilon$-secure, if the final distilled \emph{c-c-q} state $\rho_{ABE}$ is \emph{$\epsilon$-close} to the ideal key state $\rho_{ABE}^{ideal}$ given in Eq.~\eqref{eq:rho_ideal} with a proper chosen $\rho_E$ \begin{equation} \label{eq:tracedis} \min_{\rho_E} \dfrac{1}{2}(1-p_{\text{abort}})|| \rho_{ABE} - \rho_{ABE}^{ideal} ||_1 \leq \epsilon. \end{equation} Note that if a distilled state is $\epsilon$-close to the ideal key state, then the guessing probability for Eve on the final key is also bounded by $\epsilon$. Here, we want to emphasize that one should not interpret the security parameter $\epsilon$ used in the above definition as the guessing probability. In fact, the statement, \emph{a key is $\epsilon$-close to the ideal key}, is much stronger than the statement, \emph{Eve's guessing probability on a key is bounded by $\epsilon$}. Let us show a simple example. Denote $l=-\log\epsilon$ and $l<m$. We consider an $m$-bit key $K_{bad}$, which concatenates a uniformly distributed $l$-bit string with $m-l$ bit of 0's. Obviously, this key $K_{bad}$ does not satisfy the trance-distance (statistical distance in this case since everything is classical here) $\epsilon$-security definition used in Eq.~\eqref{eq:tracedis}, because the statistical distance between $K_{bad}$ and $K_{ideal}$ is close to 1 when $m\gg l$. However, the guessing probability of Eve on the key $K_{bad}$ is bounded by $\epsilon$. Clearly, the guessing probability alone is \emph{not} a proper security parameter definition. This is a common mistake for those who are confused about the security foundation of (quantum) cryptography, see for example \cite{Yuen7403842}. This common mistake has also been pointed out and clearly explained by Renner in \cite{Renner2012Reply}. \subsection{Security proofs} \label{sub:SecurityProof} \subsubsection{Lo-Chau security proof} In the Lo-Chau security proof \cite{lo1999unconditional}, the joint quantum state shared by Alice and Bob before the final key measurement is one of the Bell states, \begin{equation} \label{eq:4Bell} \begin{aligned} \ket{\Phi^{\pm}} &= \frac1{\sqrt2}(\ket{00}\pm\ket{11}) \\ \ket{\Psi^{\pm}} &= \frac1{\sqrt2}(\ket{01}\pm\ket{10}). \end{aligned} \end{equation} To see how security of QKD is related to entanglement, consider the case where Alice and Bob share $m$-pairs of perfect EPR pairs $\ket{\Phi^+}^{\otimes m}$. It is not hard to verify that if both of them perform the local $Z$ measurement, $M_{zz}$, on their halves of $m$ pairs, they will share the ideal key state $\rho_{ABE}^{ideal}$ in Eq.~\eqref{eq:rho_ideal}. In other words, the amount of distillable entanglement from quantum transmission would give a lower bound on the key generation rate. The main job for a security analysis is to make sure that Alice and Bob eventually share (almost perfect) EPR pairs before they make the final $ZZ$ measurement to obtain secure key bits. The procedure to extract perfect EPR pairs from imperfect ones is called entanglement distillation \cite{Bennett1996Mixed}. The main idea of the Lo-Chau security proof lies on \emph{quantum error correction} \cite{lo1999unconditional}, which proved the security of an entanglement-based QKD protocol. Let us recap the Bennett-Brassard-Mermin-1992 (BBM92) \cite{bennett1992quantum} protocol, an entanglement version of BB84 in Box~\ref{tab:procedure}. For the simplicity of description, we assume Alice and Bob own quantum memories, which will be removed shortly in the Shor-Preskill security proof~\cite{shor2000simple}. \begin{tcolorbox}[title = {Box II.B.1: BBM92 protocol with quantum memories, an entanglement version of BB84.}] (1) Alice prepares an EPR pair, $\ket{\Phi^+}$, stores one half of it locally, and sends the other half to Bob. \\ (2) Upon receiving a qubit, Bob stores the half of the EPR pair in a quantum memory. If the qubit is lost in the channel or the quantum storage fails, they discard the pair. \\ (3) Repeat the above two steps many times until Alice and Bob store $N$ pairs of qubits. \\ (4) With the help of pre-shared perfect EPR pairs, Alice and Bob apply a quantum error correcting code to correct all the errors in the $N$ pairs. \\ (5) After a random hashing test, Alice and Bob share almost perfect EPR pairs. They return the cost of pairs in the previous step and measure the rest in the local $Z$ basis to obtain the final key. \end{tcolorbox}\label{tab:procedure} The (quantum) random hashing test happens in the two conjugate bases separately. In each basis, Alice and Bob can compare the parities of the qubits. Comparison of each parity will cost Alice and Bob an EPR pair. Once they agree on an enough number of parities, the states are stabilized by the operations, $X\otimes X$ and $Z\otimes Z$, with a small failure probability. This step comes from the error verification in classical error correction~\cite{Fung2010Finite}. There are a few notes on this scheme. \begin{enumerate} \item This scheme is source-device-independent, which means that the source can be fully untrusted~\cite{Koashi2003BasisInd}. In the first step, the state preparation can be done by Eve. Then, Eve prepares qubits pairs (designed to be EPR pairs) and sends to Alice and Bob who store the quantum states in memories. The rest steps, 4 and 5, are the same. \item After quantum transmission, Alice and Bob share $N$ EPR pairs. Due to channel disturbance or Eve's interference, these $N$ EPR pairs are generally imperfect and might be entangled with each other and Eve's system. Here, we consider the most general coherent attacks. \item In a security proof, it is crucial to evaluate the number of EPR pairs cost in Step 4. \end{enumerate} When Alice and Bob both measure in the local $Z$ basis, an error occurs when the outcomes are different. We call it a \emph{bit error}. Similarly, when they both measure in the $X$ basis, a \emph{phase error} occurs when the outcomes are different. Denote the bit and phase error rates to be $e_b$ and $e_p$, respectively, \begin{equation} \label{eq:biterror} \begin{aligned} e_b &= \frac{\text{\# of bit errors}}{N}, \\ e_p &= \frac{\text{\# of phase errors}}{N}. \\ \end{aligned} \end{equation} Since we are considering the most general coherent attacks, the errors are in general not independent but correlated. Note that bit and phase errors can be defined in any two complementary bases in the qubit case. For quantum signals measured in a particular basis, where the bit error is defined, the phase error denotes the hypothetical error if these signals were measured in its complementary basis. For higher dimension cases, such definitions would be slightly trickier with more than one types of phase errors. In order to distill perfect EPR pairs from imperfect ones with errors defined in Eq.~\eqref{eq:biterror}, Alice and Bob can employ quantum error correction. Entanglement distillation can be done in two steps via bit and phase error correction. In bit error correction, Alice hashes her qubits in the $Z$ basis by applying \emph{Control-NOT} (C-NOT) to ancillary perfect EPR pairs, as shown in Fig.~\ref{fig:BitPhaseEC}. Alice sends the measurement results of ancillary qubits to Bob, which serves as error syndrome in error correction. In the infinite data size limit, the number of perfect EPR pairs cost in this procedure is given by the Shannon entropy, $N H(e_b)$. By applying Hadamard gates, one can switch between bit and phase spaces. Then, similarly, the phase error correction will cost additional $N H(e_p)$ EPR pairs. \begin{figure}[hbt!] \centering \includegraphics[width=8.5cm]{hashingec.pdf} \caption{Illustration of bit error correction. By adding Hadamard gates, the circuit can also be used for phase error correction.} \label{fig:BitPhaseEC} \end{figure} Finally, the net rate of EPR pairs generated is given by, \begin{equation} \label{eq:ShorPreskillR} \begin{aligned} r \ge 1-H(e_b)-H(e_p), \\ \end{aligned} \end{equation} where $H(e)=-e\log e-(1-e)\log(1-e)$ is the binary Shannon entropy function. Note that this formula is not tight in general. If two-way classical communication is allowed in quantum error correction, more key can be distilled \cite{Gottesman2003twoway,Chau2002twoway}. In reality, when the data size is finite, the entanglement distillation might fail with a small failure probability, which can be understood as the failure probability of quantum error correction. In original security proofs \cite{lo1999unconditional,shor2000simple,koashi2009simple}, the fidelity between the key state $\rho_{ABE}$ to the ideal state $\rho_{ABE}^{ideal}$ is often used as an intermediate measure to finally bound the mutual information between the final key and Eve's system. In fact, this definition is not composable. In order to make the security parameter composable~\cite{ben2005universal,Renner2005Security}, one can apply the connections between fidelity and trace distance using a general inequality relating them [see Section III in~\cite{Fung2010Finite}]. \subsubsection{Shor-Preskill: reduction to prepare-and-measure schemes} In general, this quantum error correction based entanglement distillation procedure, which is the essence of the Lo-Chau security proof, requires quantum memories and quantum computers. However, these quantum memories and quantum computers are not available with the current technology~\cite{lo1999unconditional}. In order to remove this quantum memory or quantum computer requirement, one can move the final measurement ahead of the two error correction steps. The bit error correction becomes classical error correction, and the phase error correction becomes privacy amplification \cite{shor2000simple}. There are a few steps for this permutation of operations to work. \begin{enumerate} \item Quantum bit and phase error correction operations commute, as shown in Fig.~\ref{fig:BitPhaseEC}. This is due the fact that Alice and Bob use EPR pairs as ancillary qubits. \item The $Z$-basis measurement on the ancillary EPR qubits commutes with all the operations for error correction. This is straightforward to see since there are only two possible operations on ancillary qubits, $I$ and $X$ (from C-NOT), both of which commute with the $Z$ measurement. \item The $Z$-basis measurement on the Alice and Bob qubits commutes with the bit error correction. This is true since the $Z\otimes Z$ measurement commutes with C-NOT. After moving the $Z$ measurement ahead, the C-NOT operation becomes regular \emph{exclusive-OR} ($XOR$) on the two outcome bits. \item The $Z$-basis measurement on the Alice and Bob qubits commutes with the phase error correction. This relies on the usage of EPR pair ancillary states. \item In phase error correction, after locating the errors, phase error correction does not affect the values of final key measurement in the $Z$ basis. Thus, no ``correction'' operation is needed. Of course, the EPR pairs are still cost in this step. \item Then, all the quantum operations become classical bit operations, essentially, hashing. \item In order to perform privacy amplification, one still needs to estimate the phase error rate $e_p$. Now, let us focus on the case that the key bits are measured in the $Z$ basis. The phase error rate can be estimated by measuring the key bits in the $X$ basis. Of course, in order for this estimation to work, one needs to make sure the sampling is fair, which raises the critical assumptions in security proof discussed in Section \ref{sec:assumptions}. \end{enumerate} After considering the permutation of quantum error correction and measurement, Alice and Bob can directly measure the EPR pairs once they receive them. Suppose Alice prepares the original EPR pairs, measures halves of the pairs, and sends the rest halves to Bob. Conditioned on Alice's measurement outcomes, the states sent from Alice to Bob are pure. It is equivalent for Alice to prepare these states directly and send to Bob. Now, the entanglement-based protocol is reduced to a prepare-and-measure one. Reduction from quantum bit error correction to classical error correction is easy to understand. Let us take Fig.~\ref{fig:BitPhaseEC} for example. Alice and Bob need to compare the ancillary qubit measurement results. Since the final $Z\otimes Z$ measurement commute with C-NOT operation, one can measure all the qubits in the $Z$ basis first and $XOR$ the bit values of all the measurement outcomes of the control qubits to the target qubits. The CNOT links shown in Fig.~\ref{fig:BitPhaseEC} can be understood as a hashing matrix, meaning it is equivalent to construct a matrix and multiply with the raw bit string. Of course, such error correction is linear. In general, any error correcting code can be applied, once bit and phase error correction can be decoupled. Reduction from quantum phase error correction to privacy amplification is trickier. In general, after Hadamard gates, C-NOT operation does not commute $Z$ measurement any more. In fact, those two operations become anti-commute. In this case, Alice and Bob can design phase error correcting code such that it commutes with the $Z$ measurement. Again, let us take the linear code as an example. Certain number of parity bits need to be exchanged for error correction. Assuming universal hashing, Alice sends $N H(e_p)$ bits to Bob and Bob corrects the phase errors. Note that final key measurement must commute with this hashing. Then, they can use the null space of the hashing matrix as for the final key space. The equivalence between the phase error correction and privacy amplification is illustrated in Fig.~\ref{Fig:PEPA}. This can also be understood as a random number extraction. Alice and Bob use phase error rate to estimate the randomness in the key and apply universal hashing to extract out true randomness. \begin{figure}[hbt] \begin{center} \includegraphics[scale=0.55,angle=0]{circitphpa.pdf} \end{center} \caption{Equivalence between the phase error correction and privacy amplification. The outcome of a simple 3-qubit repetition phase error correcting code is same as the hashing of three bit values.}\label{Fig:PEPA} \end{figure} In quantum error correction, we assume Alice and Bob to use ancillary EPR pairs. As shown in Fig.~\ref{fig:BitPhaseEC}, with EPR pairs, bit and phase error correction operations commute with each other. That is, one can decouple these two error correction steps \cite{Lo2003Method}. In the Shor-Preskill security proof \cite{shor2000simple}, no ancillary EPR pairs are employed. Instead, the CSS quantum error correcting code \cite{PhysRevA.54.1098,PhysRevA.54.4741} is used to decouple these two steps. After the reduction to prepare-and-measure schemes, the data postprocessing can be divided into two steps: error correction and privacy amplification. Error correction is a step to reconciliate Alice's and Bob's sifted key. If we allow one-way key reconciliation, the cost in this step is $H(A|B)$, where $A$ and $B$ represent the random variables of Alice's and Bob's sifted key, respectively. In a symmetric channel, where the detected numbers of bits 0's and 1's are the same, the cost per bit is given by $H(e_b)$ as shown in Eq.~\eqref{eq:ShorPreskillR}. It turns out that the cost can be reduced if we allow two-way key reconciliation. The optimal key rate is an open question even in the classical key agreement case \cite{748999}. Privacy amplification is a procedure for Alice and Bob to distill a common private key from a raw key about which Eve might have partial information \cite{bennett1995generalized}. The concept of privacy amplification is closely related to the randomness extraction problem \cite{doi:10.1137/100813683,Ma2013Extractor}. The difference is that in privacy amplification local randomness is a free resource, which in randomness extraction any randomness is valuable. Note that initially, the classical treatment on privacy amplification~\cite{bennett1995generalized} is only applicable to QKD under restrictive assumptions, i.e., the adversary has no quantum memory. Later on, however, this treatment has been generalized to the case where the adversary has quantum memory~\cite{renner2008security}. Details of data postprocessing, which distills a secure key from the raw data measured in quantum transmission (see Fig.~\ref{fig:flowchart}), are presented in Section \ref{sc:postprocessing}. In the end, Eq.~\eqref{eq:ShorPreskillR} gives the key rate of the BB84 protocol. Considering the symmetric case where the bit and phase error rates are the same, it is not hard to see that the tolerable error rate of Eq.~\eqref{eq:ShorPreskillR} is 11\%, comparing to 7\% in the Mayers's proof \cite{mayers2001unconditional}. Similar to the entanglement distillation case, this formula is not tight. With two-way classical communication, one can achieve advantage distillation \cite{Gottesman2003twoway} using bit-flip error {\it detection}. Nonetheless, phase error detection remains forbidden in the absence of quantum computers. From the QKD postprocessing point of view, the bit and phase errors might be correlated. Alice and Bob can perform some preprocessing to reduce the total amount of key cost in error correction and privacy amplification. For example, they can group bits into pairs and compare parities and discard the ones with different parities. In this way, one can reduce the errors in remaining bits. This is called B step \cite{Gottesman2003twoway}. It terms out that, such prepossessing is useful in practical QKD processing \cite{Ma2006Twoway}. With two-way classical communication, one can also increase the tolerable error rates \cite{Gottesman2003twoway,Chau2002twoway}. Also, with the six-state protocol \cite{PhysRevLett.81.3018,lo2001proof}, it has been shown that the tolerable error rate is higher. We list all the tolerable error rates in Table \ref{Tab:tolerable}. Apparently, there are gaps between the lower (tolerable) and upper error rate bounds. This is an open question in QKD as well as in entanglement distillation for many years, which is also related to key agreement problem in classical communication case \cite{748999}. \begin{table}[ht!] \caption{List of tolerable error rate bounds for different schemes and proofs. The upper bounds are evaluated by simple individual (intercept-and-resend) attacks \cite{Gottesman2003twoway}.} \label{Tab:tolerable} \begin{tabular}{l @{\hspace{0.6cm}} l @{\hspace{0.6cm}} l @{\hspace{0.6cm}} l} \hline \hline \textbf{Scheme} & \textbf{One-way} & \textbf{Two-way} & \textbf{Upper bound} \\ \hline BB84 & 11.0\% & 20.0\% & 1/4 \\ Six-state & 12.7\% & 27.6\% & 1/3 \\ \hline \hline \end{tabular} \end{table} \subsubsection{Koashi's complementarity approach} The aforementioned security analyses by Lo-Chau and Shor-Presill based on quantum error correction complication certainly enjoy the strong intuition from entanglement to privacy. In fact, it turns out that entanglement (or a quantum channel that is capable of transmitting entangled state) is a precondition for secure QKD \cite{curty2004}. The main drawback of this approach is its complication of introducing a virtual entanglement-based protocol. Although the bit and phase error correction can be decoupled in postprocessing by employing the CSS quantum error correcting code~\cite{shor2000simple} or ancillary EPR pairs \cite{Lo2003Method}, these two steps always mix together in security proofs. Sometimes, constructing a virtual entanglement protocol could be highly nontrivial \cite{PhysRevLett.90.167904,PhysRevA.69.032316,fung2009security}. Error correction and privacy amplification are very different procedures in conventional cryptography. The former is to guarantee that Alice and Bob share an identical key, while the latter is to make sure they share a private key. One key observation is that error correction step is not directly related to quantum laws in the security analysis. That is, if Alice and Bob only want to share an identical key, they can just transmit classical states to do the job. From this observation, Koashi has developed a simplified security proof framework based on \emph{complementarity} \cite{koashi2009simple}. In Koashi's approach, error correction and privacy amplification are decoupled from the beginning. Alice and Bob perform error correction first to make sure that the two bit strings are the same. Now, the problem becomes how many private key bits can be distilled from Alice's (same as Bob's) error corrected key. In this case we only need to deal with two parties, Alice and Eve. Denote the length of Alice's key string to be $N$. Alice's $N$-bit string can be regarded as the $Z$-basis measurement outcome of $N$ qubits $\rho_A\in \mathcal{H}_{2^N}$. Note that, under the most general coherent attacks, these $N$ qubits are correlated (or even entangled) with each other. The key idea is that in a virtual protocol, if each qubit of $\rho_A$ is measured in the complementary $X$-basis measurement and only $+1$ results are obtained, then $\rho_A=\ket{+}^N$, where $\ket{+}$ is the eigenstate of $X$ with the eigenvalue $+1$. In this ideal scenario, no one (including Eve) can predict Alice's key bits without accessing the measurement results directly. Like EPR pairs discussed in the Lo-Chau security proof, this ideal case will render perfect privacy. Interestingly, this unpredictability on the (computational) $Z$-basis is quantified by the coherence measure in the resource theory \cite{Yuan2015Coherence}, which is recently connected to the security of QKD \cite{Ma2019Operational}. In general, $\rho_A$ is not a product of $\ket{+}$ states. In this framework, the phase error rate $e_p$ is defined as the ratio of getting $-1$ eigenstates of the complementary $X$-basis measurement on $\rho_A$. The parameter $e_p$ can be estimated differently in different QKD protocols. For instance, in BB84, essentially Alice randomly chooses some qubits to be measured in the $X$-basis and use random sampling to estimate $e_p$. Details of random sampling for parameter estimation will be discussed in Section \ref{sc:postprocessing}. Now, Alice can perform a virtual phase error correction on her $N$ qubits by the similar means discussed in the Lo-Chau phase error correction. Alice can hash the $X$-basis measurement outcomes and find the error syndrome. After phase error correction, Alice's state becomes close to $\ket{+}^N$, again, measured by fidelity or trace distance. The key difference between Lo-Chau and Koashi security proofs lies on the definition of the phase error rate $e_p$. In Lo-Chau security proof, $e_p$ is defined on Bob's system relative to Alice's, while in Koashi's, it is defined on Alice's (or Bob's) side locally depending on protocols and Bob (Alice) can have an arbitrary system (irrelevant for security). In Koashi's approach, the complementary basis can be chosen arbitrary as long as $e_p$ can be estimated accurately. Meanwhile, along the line of complementarity approach, security proofs based on entropic uncertainty relations~\cite{coles2017entropic} have been developed \cite{Koashi_2006,berta2010uncertainty}. In summary, the Lo-Chau, Shor-Preskill and Koashi security proofs are all based on the phase error correction. Note that, in this line of approach, the estimation of $e_p$ is the core of the security analysis. Sometimes, more sophisticated tools like semi-definite programming is employed to upper bound the phase error rate \cite{wang2019characterising}. Recently, there is an effort to make a connection between Shor-Preskill's type security proof \cite{shor2000simple,koashi2009simple} and Entropic approach~\cite{renner2008security} by Tsurumaru~\cite{tsurumaru2018leftover}. So far, the security proof we reviewed here focuses on the BB84 protocol. Obviously, the security proof based on phase error correction can be extended to other protocols, like Bennett-1992 (B92) \cite{PhysRevLett.68.3121,PhysRevLett.90.167904} and six-state protocols \cite{PhysRevLett.81.3018,lo2001proof}. Meanwhile, this technique can also be employed to general qudit systems \cite{1412037}. Note that there is also security proof based on the idea of twisted states~\cite{PhysRevLett.100.110502,4529275}. Intuitively twisted states include shields. It allows the phase error correction syndrome to be hidden in the shield and thus become unaccessible to Eve. In principle, a virtual conceptual measurement on the joint state of Alice and Bob's shield will allow them to extract the missing phase error correction syndrome to complete the quantum error correction process. In practice, Alice and Bob do not need to perform such a virtual measurement. \subsubsection{Entropic approach} There is another line of security analysis~\cite{renner2008security,renner2005information,scarani2008quantum,tomamichel2012tight,coles2017entropic} that originates from the communication complexity and quantum memory approach \cite{BenOr2002Security,renner2008security}. Based on the entanglement distillation idea, a framework has been established for a general $\rho_{AB}$ by Devetak and Winter~\cite{Devetak2005Distillation}. In this quantum-entropy based framework, Alice and Bob share many i.i.d.~copies of $\rho_{AB}$, on which they perform measurements to obtain key bits. The Devetak-Winter key rate formula is given by \begin{equation} \label{eq:DevetakWinter} \begin{aligned} r &= S(A|E)-H(A|B), \\ S(A|E)&=S(\rho_{AE})-S(\rho_{E}) \\ &=S(\rho_{B})-S(\rho_{AB}), \\ \end{aligned} \end{equation} where $S(A|E)$ is conditional quantum entropy, $S(A)=-\Tr(\rho_A\log\rho_A)$ is the von Neumann entropy. In the derivation, we assume the worst case scenario where $\rho_{ABE}$ is pure. In fact, the privacy amplification term can also be written in a relative entropy form \cite{coles2016numerical}, \begin{equation} \label{eq:DevetakWinterRel} \begin{aligned} S(A|E)=D(\rho_{AB}||\Delta_z(\rho_{AB})), \end{aligned} \end{equation} where $\Delta_z(\rho_{AB})=\sum_{i}\dyad{i}_A\rho_{AB}\dyad{i}_A$ is the partial dephasing operation on system A in the $Z$ basis and the relative entropy function $D(\rho||\sigma)=\Tr\rho\log\rho-\Tr\rho\log\sigma$. This allows us to give an operational interpretation of coherence in QKD \cite{Ma2019Operational}. The density matrix information $\rho_{AB}$ is unknown to Alice and Bob due to Eve's interference. They have to monitor $\rho_{AB}$ in real time, say via tomography. Thus, the Devetak-Winter analysis is normally applied to the i.i.d.~case, where Eve interferes all rounds of QKD identically and independently, i.e., collective attack~\cite{renner2008security,scarani2008quantum}. Nonetheless, the security analysis can be extended to the case of coherent attack by further analysis~\cite{tomamichel2012tight,coles2017entropic}, such as the de Finetti theorem \cite{renner2007symmetry}, the post-selection technique \cite{PhysRevLett.102.020504} and uncertainty relation for smooth entropies \cite{tomamichel2011uncertainty}. Another advantage of this approach is that the security of a complicated QKD scheme can be analyzed numerically \cite{coles2016numerical,Winick2018reliablenumerical}. \subsection{Security assumptions} \label{sec:assumptions} We will discuss the security assumptions made in general security proofs. We focus on the BB84 protocol, but most of the discussions can be applied to other protocols, such as B92, BBM92, and the six-state protocols. In security proofs \cite{lo1999unconditional,shor2000simple,koashi2009simple}, as shown in Section \ref{sub:SecurityProof}, we assume Alice sends ideal qubit states in $\{\ket{0},\ket{1},\ket{+},\ket{-}\}$ and Bob performs ideal qubit $Z$-basis and $X$-basis measurement. The channel, on the other hand, is assumed to be under a full control of Eve. Nevertheless, in actual experiments these assumptions can be problematic. Table~\ref{Tab:secassumptions} summarizes the main differences between the security assumptions of security proofs and typical experimental setups. These differences, if unnoticed, might essentially open the security issue of \emph{basis dependency} between $Z$-basis and $X$-basis, thus causing the problem of quantum hacking attacks (see Table~\ref{Tab1}). \begin{table}[ht!] \scriptsize \caption{Security assumptions and actual setup for BB84.} \label{Tab:secassumptions} \begin{tabular}{l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l @{\hspace{0.3cm}} l} \hline \hline \textbf{Component} & \textbf{Security assumption} & \textbf{Practical setup} \\ \hline Photon source & Ideal single photon & Coherent laser \\ Encoding state & Two-dimension & Arbitrary-dimension \\ Encoding state & Basis-independent & Source flaws \\ Measurement & Two-dimension & Arbitrary-dimension \\ Measurement & Basis-independent & Measurement flaws \\ Photon detection & Ideal SPD & Threshold detector \\ \hline \hline \end{tabular} \end{table} \subsubsection{Source} First, let us relax the requirement on source by considering a more general source. In a prepare-and-measure QKD protocol, Alice randomly prepares system $B$ on one of the four states, $\{\rho_{x0},\rho_{x1},\rho_{z0}, \rho_{z1}\}$, and sends it to Bob. These four states can be denoted as $\rho_{\beta \kappa}$, where $\beta\in\{X, Z\}$ represents the encoding basis, and $\kappa\in\{0,1\}$ represents the encoding key bit. Here, we consider four states with two bases, but such scenario can be easily extended to more general cases with an arbitrary number of states and bases. The prepare-and-measure protocol can be linked to the entanglement-based one as follows. Define the purification of state $\rho_{\beta \kappa}$ as $\ket{\psi_{\beta\kappa}}_{A_0 B}$, where system $A_0$ is an ancillary system. From an entanglement-based view of protocol, Alice sends out state $\rho_{\beta \kappa}$ is equivalent to her preparing \begin{equation} \label{eq:StatePrepare} \begin{aligned} \ket{\Psi_\beta}_{A A_0 B} = \dfrac{1}{\sqrt{2}}\sum_{\kappa} \ket{\beta_\kappa}_A \ket{\psi_{\beta\kappa}}_{A_0 B}, \end{aligned} \end{equation} then measuring system $A$ in the $\beta$-basis, and sending out system $B$ according to measurement result $\kappa$. Here, system $A$ is a qubit system, $\ket{\beta_\kappa}_A$ is the $\beta$-basis eigenstate whose eigenvalue is $\kappa$. For the ideal BB84 protocol, there is no ancillary system $A_0$ (or $A_0$ is just a detached trivial system), since all encoding states $\rho_{\beta\kappa}$ are pure. Then, the states sent by Alice are, \begin{equation} \label{eq:StateB} \begin{aligned} \rho_{\beta\kappa} &= Tr_{A_0}(\ket{\psi_{\beta\kappa}}\bra{\psi_{\beta\kappa}}_{A_0 B}), \end{aligned} \end{equation} the four BB84 states. To send out $\rho_{x\kappa}$, in the entanglement-based equivalent protocol, Alice prepares $\ket{\Phi^+}_{AB} = \dfrac{1}{\sqrt{2}} (\ket{++} + \ket{--})_{AB}$, measures system $A$ on the $X$ basis, and obtains the measurement result $\kappa$. Similarly, to send out $\rho_{z\kappa}$, Alice first prepares $\ket{\Phi^+}_{AB} = \dfrac{1}{\sqrt{2}} (\ket{00} + \ket{11})_{AB}$, measures $A$ on the $Z$ basis, and obtains the measurement result $\kappa$. No matter which basis Alice wants to send, the initial entangled states prepared are the same. Let us denote the $X$-basis state and $Z$-basis state are \begin{equation} \begin{aligned} \rho_x &= \dfrac{1}{2}(\rho_{x0} + \rho_{x1}), \\ \rho_z &= \dfrac{1}{2}(\rho_{z0} + \rho_{z1}), \\ \end{aligned} \end{equation} which are the quantum state transmitted given Alice and Bob choose the $X$ and $Z$ bases, respectively. Thus, in the ideal BB84 source case, the state sent out by Alice is independent of the basis choice, \begin{equation} \label{eq:basisindep} \begin{aligned} \rho_x=\rho_z. \\ \end{aligned} \end{equation} We call this kind of source \emph{basis-independent} \cite{Koashi2003BasisInd,Ma2007Entangled}. In the original proposal of the BB84 protocol, the basis choice is assumed to be unknown to Eve. This is also a crucial assumption in security proofs \cite{lo1999unconditional,shor2000simple}, as shown in Section \ref{sub:SecurityProof}. This is important for phase error estimation. If the source is basis-dependent, one cannot simply use one basis information to estimate the other. This is guaranteed by Eq.~\eqref{eq:basisindep}. In fact, as long as the source is basis-independent, it can be in arbitrary dimension or state. It can even be assumed to be under the control of Eve. In practice, it is hard to construct single-qubit sources. Instead, entangled photon sources are widely used as a basis-independent source. Notice that, for entangled photon sources to work as a basis-independent source, the measurement for heralding has to be basis-independent. Later in Section~\ref{sub:GLLP}, we will show that the security can be guaranteed once the source contains a certain amount of basis-independent components. In some QKD schemes, such as BBM92~\cite{bennett1992quantum}, Alice and Bob choose basis after the quantum signals transmitted through the channel. In BBM92 protocol, Alice prepares an entangled source, holds one part by herself and sends another to Bob. Alice measures her own part in some basis to realize the basis choice and encoding. In these schemes, the quantum source can even be assumed to be in the possession of Eve. Then for these schemes, Eq.~\eqref{eq:basisindep} can be guaranteed by the experimental setting. \subsubsection{Measurement} The requirement on measurement is very similar. Again, take the BB84 protocol as an example. There are four measurement outcomes labeled by two bits, $\beta, \kappa$. The corresponding four POVM elements are $M_{\beta\kappa}$, \begin{equation} \label{eq:MxMz} \begin{aligned} M_x &= M_{x0} + M_{x1}, \\ M_z &= M_{z0} + M_{z1}. \\ \end{aligned} \end{equation} Here, $\{M_{x0}$, $M_{x1}\}$ form the $X$-basis measurement, while $\{M_{z0}, M_{z1}\}$ form the $Z$-basis measurement. We also require the measurement to be basis-independent, \begin{equation} \label{eq:Mbasisindep} \begin{aligned} M_x=M_z. \\ \end{aligned} \end{equation} On the measurement side, the requirement is more strict. For the security proof presented in Section \ref{sub:SecurityProof}, it must be qubit measurements in the $X$ and $Z$ bases. Such requirement can be extended to more general projection measurements. In practice, a \emph{squashing model} is widely employed \cite{gottesman2004security,Beaudry2008squashing,Fung2011universal}. In a squashing model, an arbitrary quantum state (from the channel) is projected to a qubit or vacuum. Then the $X$ or $Z$ measurement is performed. It has been proved that a typical threshold detector model adapts to the squashing model \cite{Beaudry2008squashing,Tsurumaru2008Security}. Now, one can see that the assumptions on the source and measurement are quite different. For source, one only needs to guarantee its basis-independent property in Eq.~\eqref{eq:basisindep}. On measurement, it must be specific projection measurements. In practice, the source requirement is easier to meet comparing to the measurement requirement. Hence, there are more practical security issues on measurement than on source. A full security analysis needs to take account of these measurement deviations. We present it in Section~\ref{sub:GLLP}. This problem is finally resolved by MDI-QKD (see Section~\ref{sec:5:mdi}). \subsubsection{Channel} In security proofs, the channel is assumed to be under the full control of Eve. Thus, in principle, we do not put any requirement on channel. In fact, if any implementation deviation from the ideal QKD protocol can be put into the channel, it will not cause any security problems. For example, detectors normally have a finite efficiency. The loss caused by detectors can be moved to channel. Then, a detector can be replaced by a 100\% efficiency one in security analysis. Now, the question is what kind of implementation deviations can be moved to channel. The implementation deviation can be regarded as some deviation operation acting on an ideal implementation. The key requirement is that the deviation operation must commute with basis switch operation. Alice and Bob each uses a basis switching device (say, a phase modulator in phase-encoding schemes). The channel is defined as the operation on the quantum signals in between the two basis switching devices. \subsection{Practical security analysis} In practice, there are two issues that need to be addressed: device imperfection and statistical fluctuation. In Section \ref{sec:assumptions}, we review the assumptions in the security proofs. In reality, these assumptions might be fully satisfied. Implementation devices might be (slightly) deviated from the ideal case used in the security proofs. When the deviation is small enough, we expect that a secure key can still be generated. In Section \ref{sub:GLLP}, we review the quantification of device imperfections and its effects on the security analysis. In principle, the error rates defined in Eq.~\eqref{eq:biterror} cannot be obtained accurately since they are measured in complementary bases. In the security proofs reviewed in Section \ref{sub:SecurityProof}, we employ random sampling to estimate the error rates. When the data size goes to infinite, the error rates approach error probabilities, which can be estimated accurately. In a finite data size, such parameter estimation would render a finite confidence interval. In Section \ref{sec:randomsampling}, we review the parameter estimation with random sampling by calculating the failure probabilities and parameter bounds. In Section \ref{sc:postprocessing}, we review the classical postprocessing of QKD and explain how can Alice and Bob distill secure keys in the raw bit strings from quantum measurement to final secure keys, with the help of public discussions. Note that some of the discussions need to be encrypted and/or authenticated. \subsubsection{GLLP} \label{sub:GLLP} There exist deviations between realistic QKD systems and the ideal QKD protocol. In order to achieve practical security of a QKD system, Alice and Bob need to characterize these device deviations or imperfections carefully and take them into account in the security analysis. Based on previous works on the topics \cite{lutkenhaus2000security,inamori2007unconditional}, Gottesman, Lo, L\"utkenhaus, and Preskill (GLLP) established a general framework for security analysis with realistic devices \cite{gottesman2004security}. First, Alice and Bob need to characterize their devices to see how much deviation from the ideal ones used in the security proofs. One can employ typical distance measures, like fidelity and trace distance, to quantify the deviation. In principle, Alice and Bob can perform some virtual measurement on the devices for each run in realtime to see whether it works the same as the ``ideal device" or its ``orthogonal case". Then, they can tag the sifted key bit as ``good" if the virtual measurement projects to the ideal case, and tag ``bad" if it is ``orthogonal case". Of course, in reality, Alice and Bob do not know the virtual measurement result. Instead, they know the ratio of these two cases. Both source and measurement imperfections can fit into this scenario. The GLLP security analysis essentially tells us how to extract secure bits when the good bits are mixed with bad ones. So far, the discussion is rather abstract, in the following discussions, we take the source imperfection as an example. The framework is generic. Here, let us take the BB84 protocol for example. In reality, a weak coherent state photon source is widely used as an approximate single photon source. With phase randomization, one can treat it as a mixture of Fock states \cite{lo2005decoy}. The vacuum and single-photon components are basis-independent, whereas the multiphoton components are not. In principle, Alice can measure the photon number to tag each encoded state being basis-independent or not (this is the aforementioned virtual measurement part). Denote the ratio of Bob's detected bits from the basis-independent source (good part, e.g. vacuum and single-photon component in the BB84 protocol) to be $1-\Delta$, and the rest (bad parte.g., multiphoton components in the BB84 protocol) is $\Delta$. Details of the source model and its security analysis will be presented in Section \ref{Sub:Source}. With Alice's tagging information (photon number in this example), she can sort the sifted key bit string $k_A$ into two sub-strings, $k_{good}$ and $k_{bad}$, where \begin{equation} \label{eq:kgoodbad} \begin{aligned} |k_{good}| &= (1-\Delta)N, \\ |k_{bad}| &= \Delta N. \\ \end{aligned} \end{equation} Following the phase error correction security proof, the underlying phase error rate of $k_{good}$ is $e_p$, which can be estimated accurately via complement measurements. The phase error rate of $k_{bad}$ is unknown. In the worst case scenario, the phase error rate of the string $k_{bad}$ could be as high as 1/2. The main idea of the GLLP security analysis is that if Alice and Bob employ linear privacy amplification, such as the matrix hashing introduced in Section \ref{sc:postprocessing}, they can still distill secure keys from $k_{good}$ by accessing $k_A$ only. Denote $k_{good}'$ to be the bit string if Alice modifies the sifted key bit string by setting the bad bit positions to be 0. Similarly, denote $k_{bad}'$ to be the bit string if Alice sets the good bit positions to be 0. Then, \begin{equation} \label{eq:kprime} \begin{aligned} |k_{good}'|=|k_{bad}'|&=|k_A|, \\ k_{good}'\oplus k_{bad}'&=k_A. \\ \end{aligned} \end{equation} Suppose a hashing matrix $T$ can distill secure bits from $k_{good}$. That is, $T k_{good}$ is a secure key. Then, it is not hard to show that $T' k_{good}'$ results in the same secure key if one extends the matrix $T$ to $T'$ by inserting new columns corresponding to the bad positions of $k_{good}'$. That is, $T$ is a sub-matrix of $T'$ by taking certain column vectors. Here comes the clever trick of GLLP: since $T' k_{good}'$ is private, the $XOR$ result \begin{equation} \label{eq:Tkprime} \begin{aligned} T' k_{good}'\oplus T' k_{bad}' &=T' (k_{good}'\oplus k_{bad}') \\ &=T' k_A \\ \end{aligned} \end{equation} is also private even Eve knows everything about $T' k_{bad}'$. Note that the new added columns from $T$ to $T'$ can be arbitrary. In practice, Alice can just pick up a universal hashing matrix $T'$ to do privacy amplification and its sub-matrix $T$ will automatically a smaller-size universal hashing matrix. Therefore, the secure key rate formula of GLLP is given by \begin{equation} \label{eq:GLLP} \begin{aligned} r \ge -H(E)+(1-\Delta)[1-H({e_p})]. \end{aligned} \end{equation} where $E$ is the total QBER. This key rate formula can be viewed as an extension to Eq.~\eqref{eq:ShorPreskillR}. Furthermore, we do not need to restrict ourselves to two tags case, good and bad. In principle, Alice and Bob can label sifted key bits with an arbitrary dimensional tag $g$ and for each $g$ they can derive its corresponding phase error rate $e_p^g$. With the same argument as above, we can extended the GLLP formula \cite{Ma2008PhD}, \begin{equation} \label{eq:GLLPext} \begin{aligned} r \ge -H(E)+\sum_g q_g[1-H(e_p^g)], \end{aligned} \end{equation} where $q_g$ is the ratio of sifted key bits with the tag $g$ and $\sum_g q_g=1$. Here, we assume Alice and Bob cannot access the tag $g$ in reality and hence they have to do the error correction part for all bits together. If they can really read out tags for each run, they can divide this error correction part as well. \subsubsection{Random sampling and finite-data size} \label{sec:randomsampling} The infinite data size limit ($N\rightarrow\infty$) is used for the key-rate formula Eq.~\eqref{eq:ShorPreskillR} and \eqref{eq:GLLPext}. When the data size is finite, the phase error rate, $e_p$, used for evaluate the amount of privacy amplification cannot be measured accurately. Instead, Alice and Bob can bound $e_p$ via certain complementary measurements. In the BB84 protocol, the phase error probability in the $Z$-basis is the same as the bit error probability in the $X$-basis. In the following discussions, we assume Alice and Bob obtained the sifted key in the $Z$-basis measurement and want to estimate the underlying phase error rate $e_p$. Thus, by sampling the qubits in the $X$-basis, Alice and Bob can bound $e_p$. This is a typical \emph{random sampling} problem. Given a certain number of phase error rates in $n_x+n_z$ positions, Alice and Bob randomly sample $n_x$ positions for phase error testing and find $n_xe_{bx}$ errors. The sampling problem lies on bounding the phase error rate $e_{pz}$ in the remaining $n_z$ positions. The (upper) bound is related to the failure probability by a hypergeometric function \cite{Fung2010Finite}. Specifically, the main objective is to evaluate the deviation $\theta$ of the phase error rate from the tested value, the bit error rate in the complementary basis, due to the finite-size effect. Here we recap the results from Section IX of \cite{Fung2010Finite} and list the variables in Table \ref{Tab:phaseest}. The phase error rate $e_{pz}$ is bounded by \begin{equation} \label{Finite:Phbound} \begin{aligned} e_{pz} \le e_{bx}+\theta, \end{aligned} \end{equation} with a failure probability of \begin{equation} \label{Finite:PhFailX} \begin{aligned} \varepsilon_{ph} &\le \frac{\sqrt{n_x+n_z}}{\sqrt{e_{bx}(1-e_{bx})n_xn_z}} 2^{ -(n_x+n_z)\xi(\theta) }, \\ \end{aligned} \end{equation} where $\xi(\theta)=H(e_{bx}+\theta-q_x\theta)-q_xH(e_{bx})-(1-q_x)H(e_{bx}+\theta)$. If we take the Taylor expansion of Eq.~\eqref{Finite:PhFailX}, one can obtain the first order approximation essentially the same as the Gaussian limit used in the Shor-Preskill security proof \cite{shor2000simple}. Another approach to deal with the problem of finite-size effect is by employing the \emph{smooth min-entropy}~\cite{renner2008security}, which is a valid measure of randomness in the non-asymptotic cases, and degenerates to Shannon entropy in the i.i.d.~limit. This approach has been applied to QKD to prove the finite-key security with almost tight bounds~\cite{tomamichel2011uncertainty,tomamichel2012tight}. Moreover, the smooth min-entropy approach is rather general to deal with non-i.i.d.~case and can be applied to other quantum information processing protocols, such as one-shot coherence resource theory~\cite{Zhao2018One} and device-independent QKD~\cite{Arnon2018Practical}. Note that for the security analysis of QKD systems with realistic devices, the finite data size effects are much more complicated. We shall review it in Section \ref{sec4:decoy}. \begin{table}[htb!] \caption{List of notations in phase error estimation.} \label{Tab:phaseest} \begin{tabular}{c|c} \hline \hline \textbf{Notation} & \textbf{Definition} \\ \hline $n_z$ & number of bits measured in the $Z$-basis \\ $n_x$ & number of bits measured in the $X$-basis \\ $e_{bx}$ & bit error rate in the $X$-basis \\ $e_{pz}$ & phase error rate in the $Z$-basis \\ $q_x$ & sampling ratio $n_x/(n_x+n_z)$ \\ $\theta$ & deviation of the phase error rate \\ $\varepsilon_{ph}$ & failure probability of phase error estimation \\ \hline \hline \end{tabular} \end{table} \section{QKD Implementation} \label{Sc:Implement} In practice, security of a QKD system is often related to its implementation. A QKD implementation is composed of three parts: source, channel, and detection. In a rigorous security proof, the channel is assumed to be under the full control of Eve, who can replace the channel with any quantum operation she desires. In the security proof model, no implementation assumption is required on the channel. As a result, the security of the system does not depend on the physical realization of the quantum channel. Therefore, the practical security for the channel is not an issue. For the quantum source and detection, on the other hand, a security proof normally requires some assumptions on practical realization. Photons are most widely used for communication, due to their robustness against decoherence due to noisy environment and fast traveling speed. Hence, we mainly focus on the quantum optical realization of QKD systems. We first discuss the encoding and decoding methods, then briefly introduce the practical source, channel and detection devices, and finally the classical postprocessing. Here, we primarily review the practical components of a discrete-variable QKD (DV-QKD) system, while the discussions for continuous-variable QKD (CV-QKD) can be found in Section~\ref{sec:CVQKD}. \subsection{Encoding and decoding} Different encoding and decoding methods are reflected on source, channel, and detection. For discrete-variable QKD schemes, Alice needs to figure out an efficient method to encoding her qubit (or qudit) into the quantum states. Accordingly, Bob needs to develop an efficient method to read out the quantum information encoded by Alice. In general, for qubit-based QKD, the quantum information can encoded into two quantum modes, $s$ and $r$, and their relative phases. Normally, the two modes are assumed to be orthogonal, say, using orthogonal polarizations or distinct time bins. Then, for a photon, the states $\{\ket{10}_{sr}, \ket{01}_{sr}\}$ form a Hilbert space, named $Z$ basis. Here, ``0" and ``1" refer to the photon number in a mode. Two complementary bases, $X$ and $Y$, are defined with the relative phases. The $X$ and $Y$ basis states can be written as, $\{ \ket{10}_{sr} \pm \ket{01}_{sr} \}$ and $\{ \ket{10}_{sr} \pm i \ket{01}_{sr} \}$. In reality, a widely applied method is polarization encoding, which utilizes the polarization modes. The horizontal and vertical polarizations of a photon, denoted by $\ket{10}_{HV}$ and $\ket{01}_{HV}$, are used for the $Z$ basis encoding. Then, the $X$ basis states, $\{\ket{10}_{HV} \pm \ket{01}_{HV} \}$, denote the linear polarization modes along the directions of $\pm 45^\circ$, respectively. The $Y$ basis states, $\{ \ket{10}_{HV} \pm i \ket{01}_{HV} \}$, denote the left- or right-handed circular polarizations. In the decoding process, the basis choice is realized by a polarization controller (Fig.~\ref{fig:encoding}), and the polarization measurement is realized by polarization beam splitter (PBS) connected with single photon detectors. Another common method is time-bin phase encoding, where Alice chooses two pulses, a signal pulse and a reference pulse, for two encoding modes, denoted by $s$ and $r$, respectively. Similar to the polarization encoding, for a single photon, the two time-bin modes form the $Z$ basis, $\{\ket{10}_{sr}, \ket{01}_{sr} \}$. Here, the qubit in the $Z$ basis determines whether the photon stays in the signal time bin or the reference time bin. The $X$ and $Y$ basis states, $\{ \ket{10}_{sr} \pm \ket{01}_{sr} \}$ and $\{ \ket{10}_{sr} \pm i \ket{01}_{sr} \}$, denote the photons with a relative phase $0,\pi$ and $\pi/2, 3\pi/2$ between the signal and reference pulses, respectively. In the decoding process, an interferometer (Fig.~\ref{fig:encoding}) is employed to extract the phase information. \begin{figure*}[hbt] \centering \includegraphics[width=14cm]{encoding.pdf} \caption{Illustration of optical device of \textbf{a}, polarization encoding, and \textbf{b}, relative phase encoding. The top shows the encoding, while the bottom shows the decoding. PolM: polarization modulator; PM: phase modulator; PBS: polarization beam splitter; SPD: single photon detector; BS: beam splitter.} \label{fig:encoding} \end{figure*} For qudit-based QKD, Alice and Bob need to find $d$ orthogonal modes and the encoding and decoding are similar. For example, the orbital angular momentum is the freedom of photons on spatial distribution, which contains a large Hilbert space. By encoding the high-dimensional key information into the orbital angular momentum, one can enhance the performance of QKD \cite{Simon2006,Cerf2002}. Another example is to encode with multiple time bins. In differential phase shift (DPS) QKD, the relative phase or each time-bin pulse is only $0$ or $\pi$, and the key is encoded into relative phase of two neighbouring pulses. Round-Robin-DPS QKD \cite{sasaki2014practical} encoding and decoding the phase difference circularly. \subsection{Photon sources} \label{Sub:Source} Here, we mainly discuss various practical photon sources for QKD: weak coherent-state source, thermal source, heralded single photon source, and entangled photon source. For most of prepare-and-measure QKD protocols, a single photon source is preferred. However, it is experimentally challenging to realize a high-quality and high-performance single photon source. We will discuss the photon sources according to different QKD schemes. \subsubsection{Prepare-and-measure} In a standard prepare-and-measure scheme like BB84, the common way is to employ other practical weak light sources to approximate the single photon source. In general, they are modulated to be a Fock state mixture, \begin{equation} \label{eq:FockSource} \begin{aligned} \rho &= \sum_{n=0}^{\infty} P(n) \ket{n}\bra{n}, \\ \end{aligned} \end{equation} where $P(n)$ is the photon number distribution and $\ket{n}$ is the $n$-photon number state. For different types of sources, the photon number distribution will also be different. Normally, the single photon component, $\ket{1}\bra{1}$, is required to be dominant comparing to higher order components. The weak coherent-state source is the most widely employed in QKD, which can be easily realized by attenuating laser lights. The light generated by a laser can be regarded as a coherent pulse $\ket{\alpha}$ within the coherence time, where $\alpha$ is a complex number, and $\mu = |\alpha|^2$ is the average photon number. The coherent state can be expanded in the Fock basis as \begin{equation} \label{eq:coherentstate} \begin{aligned} \ket{\alpha} = e^{-|\alpha|^2/2} \sum_{n=0}^{\infty} \dfrac{\alpha^n}{\sqrt{n!}} \ket{n}. \end{aligned} \end{equation} The phase of $\alpha$ reflects the relative phase between different photon number components. To realize a photon source in the form of Eq.~\eqref{eq:FockSource}, Alice can randomize the phase of coherent pulses, and make it a mixture of photon number states \cite{lo2005decoy}, \begin{equation} \label{eq:coherentFock} \begin{aligned} \rho_\mu &= \dfrac{1}{2\pi} \int_{0}^{2\pi} d\phi \ket{\alpha e^{i\phi}} \bra{\alpha e^{i\phi}} \\ &= \sum_{n=0}^{\infty} P_{\mu}(n) \ket{n}\bra{n}, \end{aligned} \end{equation} where the photon number follows a Poisson distribution, $P_{\mu}(n) = e^{-\mu} \dfrac{\mu^n}{n!}$. In many QKD protocols, such as BB84, only single photon component is secure for key distribution. Thus, the light intensity $\mu$ is typically in the single photon level, $\mu=O(1)$. The thermal source is a Fock state mixture, expanded by \begin{equation} \label{eq:thermalFock} \begin{aligned} \rho_{th} &=\sum_{n=0}^\infty P_{th}(n)\ket{n}\bra{n} \\ &=\sum_{n=0}^\infty \dfrac{ \mu^n }{ (\mu + 1)^{n+1} }\ket{n}\bra{n}, \end{aligned} \end{equation} where $\mu$ is the average photon number and the photon number follows a thermal distribution $P_{th}(n)$. Note that for a small average photon number $\mu\le2$, the single photon component ratio is bigger in a Poisson distribution than in a thermal distribution. This is the reason why the weak coherent state source normally can outperform the thermal one in QKD \cite{Curty2010PassiveDecoy}. \subsubsection{Entanglement-based} For the entanglement-based QKD protocol, such as BBM92 \cite{bennett1992quantum}, an entangled photon source via the parametric down-conversion (PDC) process is normally adopted. In a PDC process, a high frequency photon is converted to a pair of low frequency photons. A PDC source emits a superposition state of different number of photon pairs \cite{Walls2008Quantum,Ma2008PDC}, \begin{equation}\label{eq:SPDC} \begin{aligned} |\Psi\rangle=(\cosh\chi)^{-1}\sum_{n=0}^{\infty}(\tanh\chi)^n\ket{n,n}, \end{aligned} \end{equation} where $\chi$ is the nonlinear parameter for the down-conversion process, $\mu=\sinh^2\chi$ is the average photon pair number, and $\ket{n,n}$ represents $n$ photon pairs in two optical modes. The PDC process is widely used to generate photon pairs. In this case, four optical modes are used. For example, a typical PDC photon source emits photon pairs in two directions. In each direction, the photon can be in $H$ or $V$ polarization. The two optical modes are entangled in polarization. Comparing to Eq.~\eqref{eq:SPDC}, due to different collection means, the amplitudes of photon pair numbers are slightly different from the one in Eq.~\eqref{eq:SPDC} \cite{Kok2000Postselected,Ma2007Entangled}, \begin{equation}\label{eq:entPDC} \begin{aligned} \ket{\Psi}=(\cosh\chi)^{-2}\sum_{n=0}^{\infty}\sqrt{n+1}\tanh^n\chi\ket{\Phi_n}, \end{aligned} \end{equation} where $\chi$ is the nonlinear parameter for the down-conversion process, $\mu=2\sinh^2\chi$ is the average number of entangled photon pairs, and $|\Phi_n\rangle$ is the state of an $n$-entangled-photon pair, \begin{equation}\label{Model:PDCn} \begin{aligned} |\Phi_n\rangle=\frac{1}{\sqrt{n+1}}\sum_{m=0}^{n}(-1)^m|n-m,m\rangle_a|m,n-m\rangle_b. \end{aligned} \end{equation} In the aforementioned example, $a$ and $b$ represent two direction of the light, and $\ket{n-m,m}_a$ represents $n-m$ photons in the $H$ polarization and $m$ photons in the $V$ polarization. The number of the entangled-photon pairs follows a Super-Poissonian distribution, slightly different from the thermal distribution, \begin{equation}\label{eq:entPn} \begin{aligned} P(n)=\frac{(n+1)(\frac{\mu}{2})^n}{(1+\frac{\mu}{2})^{n+2}}. \end{aligned} \end{equation} Notice that the PDC source can also be used as a heralded photon source in the prepare-and-measure scheme. If we only focus on one of the optical modes (normally called signal mode), tracing out the other (normally called idle mode), the photon number follows the thermal distribution, $P_{th}(n)$ given in Eq.~\eqref{eq:thermalFock}. A typical usage of a PDC source for heralded photon involves measuring the idle optical mode locally as a trigger and encoding the signal mode for QKD. In this case, once Alice obtains a trigger locally, she can largely rule out the vacuum component in the signal mode. In fact, conditioned on whether or not a detection clicks on the idle mode, the photon number distribution is different on the signal mode. Such source can be used as a passive decoy-state source \cite{Adachi2007Simple,Ma2008PDC}. Note that when $\mu$ is very small, such heralded photon source can well approximate a single photon source, which is widely used in multi-photon processing \cite{Pan2000GHZ}. \subsection{Channel} Theoretically, we put no assumption on the quantum channel used for QKD. However, in the real-world implementation, we will build the QKD channel with mature optical communication technology to enhance the performance of QKD protocol. There are two widely adopted channels for QKD: fiber and free space. The most common channel used in QKD is built with commercial optical fiber. For a standard commercial single mode fiber (SMF), losses depend exponentially on the channel distance $l$ as $10^{-\alpha l/10}$, where the loss rate $\alpha$ is roughly $0.2$ dB/km for telecom wavelength at around 1550 nm. The loss rate can be remarkable if we extend the transmission distance to over $400$ km fiber~\cite{Yin2016,Fang2019surpassing}. Besides loss, a fiber-based QKD implementation should also solve several problems, such as chromatic dispersion, polarization mode dispersion, birefringence and so forth~\cite{gisin2002quantum}. Free space channel features some advantages compared to optical fiber. There are several atmospheric transmission windows, such as 780–850 nm and 1520–1600 nm, which have a low loss with an attenuation less than 0.1 dB/km in clear weather~\cite{bloom2003understanding}. More importantly, the attenuation is even negligible in the outer space above the Earth's atmosphere, which enables long-distance QKD over $1000$ km between ground and satellite~\cite{Liaosate}. Furthermore, the decoherence of polarization or of any other degree of freedom is practically negligible. However, there are also some drawbacks concerning the free space. For instance, the weather conditions influence the losses of free space heavily. The effective apertures of the sending and receiving telescopes, influenced by alignment, movements and atmospheric turbulence, contribute coupling losses and affect the performance of free space QKD. \subsection{Detection} For DV-QKD schemes, single photon detection is realized with threshold detectors which can only distinguish the vacuum (zero photon) from single photon or multi-photon cases. Besides, some imperfections may exist in the single-photon detector (SPD): the detector efficiency $\eta$ is not $100\%$, which means some non-vacuum signals will not cause a click on the SPD; there exists a dark count factor $p_d$, which means some vacuum signals will incorrectly cause a click. This will affect the performance of QKD systems. The measurement model is based on the threshold SPDs mentioned above. For the single-photon subspace, the detection here can be regarded as $X/Y$ basis qubit measurement. However, there is multi-photon component in the final signal, and the behavior of the measurement device will be different from the required $Z$-basis and $X$-basis measurement in DV-QKD. For example, there will be double-click signals caused by multi-photon component, which will not happen in the ideal $X/Y$ basis detection. To address this issue, the squashing model of measurement is proposed, combined with the random-assignment of double-clicked signals \cite{Beaudry2008squashing,Fung2011universal}. In 2012, the MDI-QKD scheme \cite{lo2012measurement} was proposed to fill the detection loophole. The design of measurement devices in MDI-QKD is similar to the one in point-to-point QKD protocol. In discrete variable MDI-QKD scheme, the measurement device, assumed to be manipulated by the adversary, can be divided into two categories, single detection and coincidence detection. The coincident detection MDI-QKD schemes \cite{Ma2012Alternative} is based on the schemes in which the two communication parties, Alice and Bob, encode their key information into a single photon, and build correlation between their key value by a Bell state projection. The single detection MDI-QKD scheme~\cite{Lucamarini2018TF,ma2018phase} can be regarded as the detection on the coherent states rather than the single photon. They both build correlations between Alice's and Bob's bit values by Bell state projections. \subsection{Postprocessing} \label{sc:postprocessing} Postprocessing is a procedure for Alice and Bob to distill a secure key from the raw data measured in quantum transmission with the help of public discussions. The flow chart of QKD postprocessing is shown in Figure \ref{fig:flowchart}. There are a few practical aspects need to be taken into consideration when the number of signals are finite, i.e., the finite-key effect. For example, the error correction efficiency may not reach to the Shannon limit; depending on the data size, normally a factor is applied. Also, on the privacy amplification side, there will be a small failure probability. Some public communication between Alice and Bob need to be authenticated and/or encrypted. Table~\ref{table:resources} summarizes the resource cost and the failure probabilities in the various steps. \begin{figure} [htbp] \includegraphics[scale=0.68,angle=0]{flowchart.pdf} \caption{Flow chart of data postprocessing procedures. The resource cost and the failure probabilities in encryption and authentication is listed in Table \ref{table:resources}. The encryption is optional for error correction depending on whether or not more privacy application is performed later. [Figure adapted from \cite{Fung2010Finite}].} \label{fig:flowchart} \end{figure} The first step is raw key assignment, which depends on different schemes. For example, in the commercial BB84 implementation, Alice and Bob discard all the no-clicks and randomly assign double-clicks. In the MDI-QKD, this step is based on the announcement of the measurement site. In DI-QKD, Alice and Bob can perform arbitrary assignment. Of course, any improper assignment will reduce the key rate. During the public discussions, some of the classical communication need to be authenticated, as listed in Table \ref{table:resources}. In the security proofs we reviewed, we assume the encryption of classical communication in error correction. Such encryption can be avoided in other security proofs. In this case, there may be some restriction on the error correction procedure and more follow-up privacy amplification is required. For example, in the original Shor-Preskill security proof, such encryption is not necessary when the CSS code is employed. In practice, there is an advantage to using error correction without encryption, since if Alice and Bob abort the QKD procedure after error correction, no pre-shared secret bits will be lost due to encryption. In the following discussions, we assume the number of bits communicated in error correction is counted for later privacy amplification. Thus, in privacy amplification, the extraction ratio will be $r$ given in Eq.~\eqref{eq:GLLPext} without considering the finite data size effects. If the one-time pad encryption is used for error correction, the privacy amplification ratio will be higher (by removing the error correction term in Eq.~\eqref{eq:GLLPext}). After that, certain amount of secure key bits need to return to the keep pool for encryption consumption. In the end, the final key rate will still be same in the encryption and non-encryption cases. \begin{table} [htbp!] \caption{List of resource cost and the failure probabilities in the various steps. The numbers of consumed secret key bits are denoted with $k$ while the failure probabilities with $\varepsilon$. Alice sends out $N$ signals and Bob detected $n$ of them in the $Z$-basis measurement. The final key output length is $l$. The tag length refers to authentication tag and - means no authentication is required. In error verification, no message but only an encrypted (authentication) tag is transmitted. The cost of error correction $k_{ec}$ is given by $n f H(E)$. No communication is required in phase error estimation. [Table adapted from \cite{Fung2010Finite}]} \label{table:resources} \begin{tabular}{ l|ccc} \hline \hline \textbf{Procedures} & \textbf{Message} & \textbf{Tag} & \textbf{Failure Prob} \\ \hline 1. Raw key assign & $N$ & - & - \\ 2. Basis sift & $n$ & $k_{bs}$ & $\varepsilon_{bs}$ [Eq.~\eqref{eq:failauthen}] \\ 3. Bit error correction & $k_{ec}$ & - & - \\ 4. Error verification & - & $k_{ev}$ & $\varepsilon_{ev}$ [Eq.~\eqref{eq:failauthen}] \\ 5. Phase error estimation & - & - & $\varepsilon_{ph}$ [Eq.~\eqref{Finite:PhFailX}] \\ 6. Privacy amplification & $(n+l-1)$ & $k_{pa}$ & $\varepsilon_{pa}$ [Eq.~\eqref{eq:failurePA}] \\ \hline \hline \end{tabular} \end{table} \subsubsection{Error correction} For practical error correction, normally an efficiency factor $f>1$ is put before the $H(E)$ term Eq.~\eqref{eq:GLLPext}, which means the actual cost is larger than the theoretical Shannon limit. Previously, a widely used error correction protocol for QKD is Cascade \cite{Cascade1994}. The Cascade protocol is simple and highly efficient that is able to achieve an error correction factor of around $1.1\sim1.2$ for a large QBER range from 0\% to beyond 11\%. In the Cascade protocol, Alice and Bob divide their sift key bit strings into blocks and compare parities of each block to look for errors. They perform a binary search to locate the error when the parity of a block is different. The process repeats for a few times with different block sizes and permutation to ensure all the error bits are corrected. The Cascade protocol is highly interactive because the binary search requires $1+\log_2(n)$ communications and successful error correction often requires several passes. Later, several improved protocols have been proposed to reduce the interaction rounds \cite{buttler2003fast,nakassis2004expeditious,elliott2005current}. Another family of error correcting codes is forward error correction, which only needs to send one syndrome from Alice to Bob. Due to its light classical communication load, the forward error correction is widely implemented in commercial QKD systems. One outstanding example is the low-density parity-check (LDPC) code \cite{mackay1996near}. The LDPC code works well for QKD due to its the high error correction efficiency and very limited communication requirement. The design and optimization of LDPC codes in QKD postprocessing is similar to the classical case, which can be divided into three steps. \begin{enumerate} \item The first step is to find a good degree distribution \cite{mackay1999comparison,richardson2001capacity} for the target error rate. \item The second step is to generate a good parity-check matrix. Like in classical communication, the small cycles may contribute to localized information transmitted in decoding. Thus, a good parity-check matrix generation algorithm should yield a relatively large girth. Progressive edge-growth is one of the most successful algorithms to generate parity-check matrix eliminating small cycles \cite{hu2005regular}. \item The third step is to decode with Bob's key string and the received syndrome. The brief-propagation algorithm \cite{fossorier1999reduced}, which is also known as the sum-product algorithm, is highly efficient in decoding. \end{enumerate} The standard LDPC algorithm is only optimum at its designed rate for the designed QBER. But the actual QBER is fluctuating from round to round. The rate compatible LDPC can solve this problem with puncturing and shorting \cite{ha2006rate}. The main technology here is to select a mother code close to the target rate, and then to adjust the code rate with puncturing. The puncturing operation can be done multi-times, in order to find a best code rate suitable for the actual error rate. This method has been employed in QKD \cite{elkouss2009efficient,martinez2010interactive}. Besides efficiency, another important factor of error correction is the throughput. The limitation for the Cascade code is the highly interactive communications, and for LDPC is the computational cost in iterative decoding. It was reported the throughput with both the Cascade \cite{pedersen2013high} and the LDPC \cite{dixon2014high} codes can be higher than 10 Mbps. Note that the computing in decoding LDPC is always assisted with GPU acceleration. \subsubsection{Error verification and authentication} Before error correction, Alice and Bob sample the sifted key bits to roughly estimate the error rates. Then they perform error correction. After error correction, Alice and Bob can perform error verification to make sure they share the same key \cite{lutkenhaus1999estimates,Finite:Short:11}. Then, the failure probability for error correction is reflected in error verification step when finite data size is considered. It is not hard to see that the two jobs, error verification and message authentication, are very similar. In both cases, Alice and Bob want to make sure the bit strings on the two ends are the same. The only difference is that authentication tag might reveal information about the message, but error verification should not. This difference can be overcome by encrypting the tag, which has already been done in most of information-theoretically secure authentication schemes. If we employ the LFSR-based Toeplitz matrix construction, the relation between the tag length (the same as the key cost) and the failure probability is given by \begin{equation} \label{eq:failauthen} \begin{aligned} \varepsilon &= n2^{-k+1}, \end{aligned} \end{equation} where $n$ is the message length and $k$ is the key cost. After error correction and error verification, Alice and Bob are almost sure that they have located all the errors. Then, they can accurately count the number of bit errors and hence the rate $e_b$ defined in Eq.~\eqref{eq:biterror}. If Alice and Bob choose not to encrypt error correction, they can count the amount of classical communication used in the error correction, $k_{ec}$. Then they perform addition amount of privacy amplification. For example, in the ideal devices case of Eq.~\eqref{eq:ShorPreskillR} and infinite data limit, $k_{ec}=nH(e_b)$. The final key output length is given by $l=r n$. \subsubsection{Privacy amplification} Practical privacy amplification turns out to be very efficient in terms of finite data size effect, once the necessary parameters, such as the phase error rates, are estimated as reviewed in Section \ref{sec:randomsampling}. Denote the error corrected bit strings for Alice and Bob to be $k_A=k_B$ with a length of $n$ and the output length to be $l$. In the infinite key limit, $l/n=r$ given in Eq.~\eqref{eq:GLLPext} if error correction is not encrypted. In the privacy amplification procedure, Alice randomly chooses a universal hashing matrix $T\in\{0,1\}^{l\times n}$ and sends it to Bob via a public classical channel. The final key will be given by $T k_A=T k_B$, with a small failure probability. Privacy amplification works for general classes of two-universal hash functions~\cite{tomamichel2011leftover}. In particular, the universal hashing function based on Toeplitz matrices is widely used for privacy amplification. An $l\times n$ Toeplitz matrix is a Boolean matrix with the structure of \begin{equation} \label{eq:Toeplitz} \begin{aligned} T= \begin{pmatrix} a_0 & a_{-1} & a_{-2} & \cdots & a_{-n+1}\\ a_1 & a_0 & a_{-1} & \ddots \\ a_2 & a_1 & \ddots && \vdots\\ \vdots & \ddots \\ a_{l-1} & & \cdots & & a_{l-n} \end{pmatrix}, \end{aligned} \end{equation} where $a_i\in\{0,1\}$ for $-n+1\le i\le l-1$. The Toeplitz matrix can also be concisely written as $T_{(i,j)}=a_{i-j}$ where $T_{(i,j)}$ is the $(i,j)$ element of $T$. Apparently, an $l\times n$ Toeplitz matrix can be specified by $N+K-1$ bits, as opposed to $N\times K$ bits for completely random matrices. The main advantage of Toeplitz matrix hashing is that the computational complexity for $T k_A$ is $O(n\log n)$ by the fast fourier transform (FFT). Following the security proofs reviewed in Section \ref{sub:SecurityProof}, the matrix $T$ should be related to the phase error correction. In order to ensure the phase error correction to commute with the final key measurement, we require the null space of $H$ to be capable of correcting the underlying phase errors. For universal hashing functions, such error correcting capability can be evaluated with certain failure probabilities. Details of derivation can be found in Section X of \cite{Fung2010Finite}. The failure probability for privacy amplification with Toeplitz hashing is given by, \begin{equation} \label{eq:failurePA} \begin{aligned} \varepsilon_{pa} &= 2^{-t_{pa}}. \\ t_{pa} &= n r-l, \\ \end{aligned} \end{equation} If Alice transmit the Toeplitz matrix to Bob, then she needs to authenticate that communication as well, which would add an extra term of Eq.~\eqref{eq:failauthen} to $\varepsilon_{pa}$. Here in privacy amplification, by sacrificing $t_{pa}$ extra bits in privacy amplification, one can obtain a failure probability of $2^{-t_{pa}}$. More general discussions for hash functions besides Toeplitz hashing can be seen in~\cite{tomamichel2011leftover}. Note that message authentication can be done more efficiently by piling up classical communication data and authenticate them at once. That is, the authentication terms listed in Table \ref{table:resources} can be done once with one authenticated tag and one failure probability. The main drawback of this saving data and authenticating approach is that it might require lots of local data storage. In QKD system design, it is normally preferred that each procedure of postprocessing is isolated. From the simulation results \cite{Fung2010Finite}, we learn that the failure probabilities for authentication, error verification, and privacy amplification are not the main contributions to the total system one. In fact, the one in phase error rate estimation, Eq.~\eqref{Finite:PhFailX}, is the dominate term. The summation of failure probabilities evaluated here can be converted to the trace-distance measure in Eq.~\eqref{eq:tracedis}. \subsubsection{Finite-key length} When the failure probability of the postprocessing procedure is $\epsilon$, the final key is $\sqrt{\epsilon(2-\epsilon)}$-secure in accordance with the composable security definition given in Eq.~\eqref{eq:tracedis}. Finally, by including the finite-data statistics for parameter estimation (see Sec.~\ref{sec:randomsampling}) and the postprocessing costs (see Table~\ref{table:resources}), we have the finite-key length $NR$ for the finite-size security of QKD, which can be written as~\cite{Fung2010Finite}, \begin{equation} \label{Finite:KeyFinal} \begin{aligned} NR &\ge l-k_{bs}-k_{ec}-k_{ev}-k_{pa} \\ \end{aligned} \end{equation} with a failure probability of \begin{equation} \label{Finite:FailFinal} \begin{aligned} \varepsilon\le\varepsilon_{bs}+\varepsilon_{ev}+\varepsilon_{ph}+\varepsilon_{pa} \\ \end{aligned} \end{equation} where $l$ is given by, \begin{equation} \label{Finite:PAkey} \begin{aligned} l &= n_x[1-H(e_{bz}+\theta_z)]+n_z[1-H(e_{bx}+\theta_x)], \\ \end{aligned} \end{equation} where the variables can be found in Table \ref{Tab:phaseest}. Notice that one can also utilize the smooth min-entropy approach to obtain the finite-key length~\cite{renner2008security,scarani2008quantum} or the tight bounds~\cite{tomamichel2012tight}. Note that for QKD systems with realistic devices, the finite-key length is slightly complicated, we refer to \cite{lim2014concise} for decoy-state QKD, \cite{curty2014finite} for measurement-device-independent QKD, \cite{lorenzo2019tight} for twin-field QKD, \cite{Arnon2018Practical} for device-independent QKD and~\cite{furrer2012continuous,leverrier2013security} for continuous-variable QKD. \section{Quantum Hacking}\label{sec:3} In theory, it is traditional to divide Eve's hacking strategy to three main classes: \emph{individual}, \emph{collective} and \emph{coherent} (or general) attack. Individual attack means that Eve interacts with each secure qubit in the channel separately and independently; Collective attack means that Eve prepares independent ancilla, interacts each qubit independently, but can perform a joint measurement on all the ancilla; Coherent attack means that Eve can prepare an arbitrary joint (entangled) state of the ancilla, which then interact with the qubits in the channel before being measured jointly. The last one does not limit Eve's capabilities beyond what is physically possible. Any QKD system aiming to implement an informational-theoretically secure protocol therefore has to be proven secure against coherent attacks. Another aspect which cannot be neglected is security in a finite size scenario. No key transmission session can be endless and the resulting statistical fluctuations have to be taken into account~\cite{scarani2009security}. In this section, different from the theory attacks, we focus on the practical attacks which exploit the device imperfections in QKD systems. Specifically, Eve may try to exploit the imperfections in real QKD systems and launch the so-called \emph{quantum hacking} not covered by the original security proofs. Researchers have demonstrated several quantum hacking attacks in practical QKD systems. An earlier review on quantum hacking attacks can be seen in~\cite{jain2016attacks}. Here we will provide a review for the quantum attacks to both the source and the detection. The detection attacks are similar to those reviewed in~\cite{jain2016attacks}, but we provide more details for the attacks at source which exploit the multiple photons, timing or phase information of the laser source. Also, some new attacks after~\cite{jain2016attacks} will also be mentioned. Table~\ref{Tab1} summarizes a list of the attacks developed from early 2000 to the present. \subsection{Attacks at source} In the standard QKD scheme, it is assumed that Alice (state preparation) is placed in a protected laboratory and she prepares the required quantum state correctly. Unfortunately, imperfect state preparation may leak information about the secret key. Indeed, practical preparation may introduce some errors due to imperfect devices or Eve's disturbance~\cite{lutkenhaus2000security,brassard2000limitations,Fred2007,xu2010,Sun2012,Tang2013,Sun2015}. To steal the information about the states, Eve can also actively perform the Trojan-horse attack~\cite{gisin2006trojan,jain2014trojan,jain2015risk} on intensity modulators and phase modulators. This section will review some examples of attacks at source. \subsubsection{Photon-number-splitting attack} \label{sec:3:PNS} The first well known kind of hacking strategy that was considered is the photon-number-splitting (PNS) attack~\cite{lutkenhaus2000security,brassard2000limitations} aiming at the imperfect photon source. As described in section~\ref{Sub:Source}, because of technological challenge, weak coherent pulses (WCPs) generated by a highly attenuated laser are widely used in QKD implementations. Since the photon number of a phase-randomized WCP follows the Poisson distribution (Eq.~\eqref{eq:coherentFock}), there is a non-zero probability for multiple-photon pulses, i.e., those pulses containing two or more photons. Consequently, Eve may exploit the multiple-photon pulses and launch the PNS attack. In this attack, for each WCP, Eve first utilizes a quantum non-demolition (QND) measurement to obtain the photon number information. Conditioned on the QND measurement result, Eve either blocks the one-photon pulse or splits the multiple-photon pulse into two. She stores one part of the multiple-photon pulse and sends the other part to Bob. Later on, during the basis-reconciliation process of the BB84 protocol, Eve can get the secret key information for the multiple-photon pulse without introducing any errors. By doing so, Alice and Bob could not notice Eve's attack. The PNS attack restricts the secure transmission distance of QKD typically below 30 km~\cite{gottesman2004security}. Actually, in early 2000s, there were not many research groups working on QKD experiments~\cite{Hughes2000quantum,ribordy2000fast,gobby2004quantum}. Researchers in the field had a doubt on the future of QKD, and they generally thought that QKD may be impractical with WCP source. This concern severely limits the development of QKD at that time. Fortunately, the discovery of the decoy state method perfectly resolved the problem of PNS attack and made QKD practical with standard WCP~\cite{hwang2003quantum,lo2005decoy,wang2005beating}. More details on the decoy state method will be discussed in section~\ref{sec:4}. \subsubsection{Phase-remapping attack} \begin{figure}[!t] \centering \resizebox{8cm}{!}{\includegraphics{remapping.pdf}} \caption{ Phase-remapping attack~\cite{xu2010}. \textbf{a}, Diagram of phase modulation signal. $t_{0}$ is the original time location where the signal pulse is properly modulated to have phase $\phi_{0}$. Eve can time-shift the signal pulse from $t_{0}$ to $t_{1}$, where the pulse will undergo a new modulated phase $\phi_{1}$. \textbf{b}, Implementation of the phase-remapping attack in a commercial IDQ QKD system. Original QKD system: LD, laser diode. Det1/Det2, single photon detector; PMB/PMA, phase modulator; C, circulator. PBS, polarization beam splitter; FM, Faraday mirror; CD, classical detector; DL, delay line. Eve's modifications: VODL, variable optical delay line; PC, polarization controller. [Figure reproduced from~\cite{xu2010}].} \label{Fig:remapping} \end{figure} Phase modulators are commonly used to encode random bits in the source of phase-coding QKD systems~\cite{gisin2002quantum}. In practice, a phase modulator has finite response time, as shown in Fig.~\ref{Fig:remapping}a. Ideally, the pulse passes through the phase modulator in the middle of the modulation signal and undergoes a proper modulation (time $t_{0}$ in Fig.~\ref{Fig:remapping}a). However, if Eve can change the arrival time of the pulse, then the pulse will pass through the phase modulator at a different time (time $t_{1}$ in Fig.~\ref{Fig:remapping}a), and the encoded phase will be different. This phase-remapping process allows Eve to launch an intercept-and-resend attack, i.e., phase-remapping attack~\cite{Fred2007}. The phase-remapping attack is a particular threat for the bidirectional QKD schemes, such as the plug-and-play QKD structure~\cite{stucki2002quantum}. In 2010, the phase-remapping attack was successfully demonstrated in a commercial ID-500 plug-and-play QKD system (manufactured by ID Quantique\footnote{https://www.idquantique.com/})~\cite{xu2010}, as shown in Fig.~\ref{Fig:remapping}b. In this experiment, Eve utilized the same setup as Bob to launch her attack. Eve modified the length of the short arm of her Mach-Zehnder interferometer by adding a variable optical delay line (VODL in Fig.~\ref{Fig:remapping}b) to shift the time delay between the reference pulse and the signal pulse. To remap the phase small enough into the low QBER range, Eve shifted the forward signal pulse out and only the backward signal pulse in the phase modulation range by using VODL, and properly aligned the polarization direction of the backward signal pulse orthogonal to the principal axis of the phase modulator by using a polarization controller (PC in Fig.~\ref{Fig:remapping}b). The experiment demonstrated that Eve could get full information and only introduce a QBER of 19.7\%, which is much lower than the well-known 25\% error rate for an intercept-and-resend attack in BB84. \subsubsection{Nonrandom-phase attack} Phase randomization is a basic assumption in most security proofs of QKD~\cite{gottesman2004security,hwang2003quantum,lo2005decoy,wang2005beating}. Although the security of QKD with non-random phase had been proven~\cite{Lo2007Nonrandom}, the performance is very limited in distance and key rate. By assuming that the overall phase is uniformly distributed in [0, $2\pi$], a coherent state with intensity can be reduced into a classical mixture of photon number states, i.e., Eq.~\eqref{eq:coherentFock}. This can greatly simplify the security proofs and allow one to apply classical statistics theory to analyze quantum mechanics. In practice, however, the phase randomization assumption may be violated in practice, thus resulting in various attacks~\cite{Sun2012,Tang2013,Sun2015}. The first example is the USD attack demonstrated in~\cite{Tang2013}. When the phase of WCPs is not properly randomized, the quantum state will be a pure state. Then in decoy state QKD~\cite{hwang2003quantum,lo2005decoy,wang2005beating}, it is possible for Eve to distinguish the signal state and decoy state with an unambiguous state discrimination (USD) measurement. Hence, Eve first measures each of Alice's WCPs to distinguish between signal state and decoy state by performing a USD measurement, which is combined with positive operator-valued measurement (POVM) operators without disturbing the quantum state sent by Alice. After the USD, Eve performs the PNS attack. Since Eve knows which state the pulse belongs to (signal or decoy), she could do different strategies for signal state and decoy state. As a result, the key assumption in decoy state QKD~\cite{lo2005decoy} -- a decoy state and a signal state have the same characteristics -- is violated. The second example is the laser seed-control attack which was proposed and demonstrated in~\cite{Sun2015}. Semiconductor laser diode (SLD) is normally used as a single-photon source in most commercial and research QKD systems. In the interdriven mode, the semiconductor medium of the SLD is excited from loss to gain by each driving current pulse. A laser pulse is generated from seed photons originating from spontaneous emission. The phase of the laser pulse is determined by the seed photons. Since the phase of the seed photons is random, the phase of each laser pulse is random inherently. However, if a certain number of photons are injected from an external source into the semiconductor medium, these photons will also be amplified to generate laser pulses. Consequently, the seed photons consist of two parts: one from spontaneous emission and the other part from the external source. Both parts will affect the phase of the resulting laser pulse. If the injected photons greatly outnumber the photons from spontaneous emission, the phase of the output laser pulse is largely determined by the phase of the injected photons. Therefore, Eve can control the phase of Alice’s signal laser by illuminating the SLD from an external control source and successfully violate the phase randomization assumption~\cite{gottesman2004security}. \subsection{Attacks at detection} The detection component is much more vulnerable to quantum hacking attacks than the source. Since Eve controls the channel and can send any signals (e.g strong optical pulses combined with X-ray and neutrinos) to Bob and Bob has no choice but to receive Eve's signal and any filters used by Bob may be imperfect, it may be hard for Bob to isolate his lab and avoid side channels or detector control from/by Eve. For instance, a significant number of attacks have been proposed to hack single-photon detectors (SPDs)\footnote{The vulnerabilities of SPDs are mainly due to their complex working mechanism: the detection is affected by incoming light and the control electronic circuits. Therefore, Eve can manipulate the intensity, the time, or the wavelength of incoming light to control the responses of SPDs.} \cite{makarov2009controlling,Lars2010, gerhardt2011full,gerhardt2011experimentally,wiechers2011after,Seb2011}. SPDs are regarded as the ``Achilles heel" of QKD by C. H. Bennett\footnote{C. H. Bennett, “Let Eve do the heavy lifting, while John and Won-Young keep her honest,” http://dabacon.org/pontiff/?p=5340}. This section will review some examples of attacks at detection. The first two examples, double-click attack and fake-state attack, were proposed only in theory. The last two examples, time-shift attack and detector-blinding attack, were successfully demonstrated in experiment. \subsubsection{Double-click attack} Since QKD systems require the detection of two different bit values, bit 0 and bit 1, they require at least two SPDs. The double-click event refers to the case where both SPDs detect signals. The double-click event will introduce a QBER of 50\% when either one of the two bits is selected. A naive strategy is to determine double-click events as abnormal events and discards these events so as to minimize the QBER. However, this strategy results in the problem of double-click attack. In this attack, Eve simply floods Bob's polarization beam splitter with multiple photons or a strong pulse of the same polarization. Then, when Bob makes a measurement using a conjugate basis different from that of Eve, a double-click event occurs and it is discarded; when the receiver makes a measurement using the same basis as Eve's, a normal event is detected. Consequently, Alice and Bob finally share the same information with Eve. To solve this problem, L\"utkenhaus has proposed that double-click events are not discarded and bit 0 or bit 1 is randomly allocated by Bob whenever a double-click event occurs~\cite{lutkenhaus1999estimates,lutkenhaus2000security}. \subsubsection{Fake-state attack} In 2005, Makarov et al. proposed a faked-state attack, which exploits the efficiency mismatch of two detectors in a practical QKD system~\cite{makarov2005faked,Makarov2006}. In practice, the standard SPDs such as Si/InGaAs APDs are often operated in a gated mode. Therefore, the detection efficiency of each detector is time-dependent. Since QKD systems require the detection of two different bit values, 0 and 1, they often employ at least two SPDs. It is inevitable that finite manufacturing precision in the detector and the electronics, and difference in optical path length will slightly misalign the two detector gates, and cause detector-efficiency mismatch. This is illustrated in Fig.~\ref{fig:efficiencymismatch}a. At the expected arrival time T, the detection efficiencies of the two detectors are identical. However, if the signal is chosen to arrive at some unexpected times (such as $t_{1}$ and $t_{2}$ in Fig.~\ref{fig:efficiencymismatch}a), it is possible that the detector efficiencies of the two detectors, $\eta_0$ and $\eta_1$, differ greatly. This problem often exists in practical QKD systems, and it will leave a back door for Eve to attack the system. The faked-state attack is an intercept-and-resend attack. For each signal, Eve randomly chooses one of the two BB84 basis to perform a measurement and obtain a measurement result. Then, she re-sends the opposite bit value from her measurement result in the opposite basis, at a time when the detector for the opposite bit has a lower detection efficiency than the other detector. As analyzed in~\cite{Makarov2006}, Eve introduces less than 11\% QBER if the detection efficiency $\eta\leq$6.6\%. The faked-state attack, while conceptually interesting, is hard to implement in a real-life QKD system. This is because it is an intercept-resend attack and as such involves finite detection efficiency in Eve's detectors and precise synchronization between Eve and Alice-Bob's system. A typical countermeasure against detector efficiency mismatch is the four-state QKD protocol~\cite{Makarov2006}. \begin{figure}[!t] \centering \resizebox{8.1cm}{!}{\includegraphics{mismatch.pdf}} \caption{Schematic of detection efficiency mismatch~\cite{Makarov2006} and time-shift attack~\cite{Qi2007,Zhao2008}. \textbf{a}, SPD, single-photon detector. At the expected arrival time T, the detection efficiencies of SPD$_{0}$, $\eta_{0}$, for the event of bit 0, and SPD$_{1}$, $\eta_{1}$, for the event of bit 1 are the same. However, at time $t_{1}$, SPD$_{0}$ is more sensitive to the incoming photon than SPD$_{1}$, while at time $t_{2}$, SPD$_{1}$ is more sensitive to the incoming photon than SPD$_{0}$. \textbf{b}, Real detector efficiencies of the two SPDs characterized on a commercial QKD system (manufactured by IDQ) by Zhao et al.~\cite{Zhao2008}. [Fig.b reproduced from~\cite{Zhao2008}]. } \label{fig:efficiencymismatch} \end{figure} \subsubsection{Time-shift attack} Motivated by the faked-state attack, in 2007, Qi et al. \cite{Qi2007} proposed the time-shift attack. This is also based on the detection-efficiency mismatch for gated SPDs in the time domain, but is much easier to implement. Let us suppose Fig.~\ref{fig:efficiencymismatch}a illustrates the detection efficiencies of the two gated SPDs in a real-life QKD system. Eve can simply shift the arrival time of each pulse sent from Alice by employing a variable optical delay line. For example, Eve randomly shifts the pulse from Alice to arrive at $t_{1}$ or $t_{2}$ through a shorter path or a longer path of optical line. This shifting process can partially reveal the bit value of Bob: if the pulse arrives at $t_{1}$ (or $t_{2}$) and Bob announces receipt, the bit value is more likely to be 0 (1). Moreover, Eve can carefully set how many bits should be shifted forward and how many should be shifted backward to ensure that the distribution of bit 0 and bit 1 received by Bob is balanced. Hence, the time-shift attack does not make any measurement on the quantum state, and quantum information is not destroyed. Since Eve does not need to make any measurement or state preparation, the time-shift attack is practically feasible with current technology. In 2008, it has been successfully implemented on a commercial QKD system by Zhao et al.~\cite{Zhao2008} as shown in Fig.~\ref{fig:efficiencymismatch}b. This is one of the first successful demonstrations of quantum hacking on a widely-used commercial QKD system. In their experiment \cite{Zhao2008}, Eve got an information-theoretical advantage in around 4\% of her attempts. The successful implementation of the quantum attack shows that a practical QKD system has non-negligible probability to be vulnerable to the time-shift attack. \subsubsection{Detector-control attack} \begin{figure}[hbt] \centering \resizebox{8.2cm}{!}{\includegraphics{blinding.pdf}} \caption{ Schematic illustration of the detector blinding attack~\cite{Lars2010}. \textbf{a}, Linear-mode and Geiger-mode APD operation. When the APD is reverse-biased above its breakdown voltage $V_{br}$, a single photon can cause a large current $I_{APD}$ to flow, and register this as photon detection (a `click'). After that, an external circuit quenches the avalanche by lowering the bias voltage below $V_{br}$, and then the APD goes into a linear mode. In the linear mode, $I_{APD}$ is proportional to the incident bright optical power $P_{opt}$. \textbf{b}, Eve sends Bob a tailored light pulse that produces a `click' in one of his detector only when Bob uses the same measurement basis as Eve. Otherwise, no detector `clicks'. [Figure reproduced from~\cite{Lars2010}].} \label{fig:blinding} \end{figure} The detector-control attack is the most powerful attack and it has been successfully demonstrated on several types of practical QKD systems~\cite{makarov2009controlling,Lars2010}. In general, the detector-control attacks can be divided into three categories: (i) detector-blinding attack \cite{makarov2009controlling,Lars2010,lydersen2011controlling,huang2016testing}, where Eve illuminates bright light to control detectors; (ii) detector-after-gate attack \cite{wiechers2011after}, where Eve just sends multi-photon pulses at the position after the detector gate; (iii) detector-superlinear attack~\cite{lydersen2011superlinear,qian2018hacking}, where Eve exploits the superlinear response of single-photon detectors during the rising edge of the gate. Most available SPDs are InGaAs/InP APDs operating in a Geiger mode \cite{Hadfield:2009}, in which they are sensitive to a single photon. The working principle of this type of APDs is shown in Fig.~\ref{fig:blinding}a. In the detector blinding attack, by sending a strong light to Bob, Eve can force Bob's SPDs to work in a Linear mode instead of Geiger mode, as shown in Fig.~\ref{fig:blinding}a. In the Linear mode, the SPD, such as the one based on InGaAs APD, is only sensitive to bright illumination. This detector operation mode is called ``detector blinding". After blinding the detectors, Eve sends a bright pulse with tailored optical power such that Bob's detector always reports a detection event from the bright pulse, but never reports a detection event from a pulse with half power. This is illustrated in Fig.~\ref{fig:blinding}b. Consequently, Eve can successfully launch an intercept-and-resend attack without increasing QBERs. For example, when Eve uses the same basis as Bob to measure the quantum state from Alice, Bob gets a detection event as if there were no eavesdropper. But if Eve uses the opposite basis from Bob to measure the quantum state from Alice, her bright pulse will strike each of Bob's detectors with half power, and neither detector will report a detection event. In practice, a simple detector blinding attack will introduces a 50\% total loss. However, Eve can place her intercept-unit close to Alice's laboratory while compensating the loss in the remaining fiber by re-sending brighter states. The detector-control attack is applicable to various types of SPDs, such as gated APDs~\cite{Lars2010,huang2016testing}, passively or actively quenched APDs~\cite{makarov2009controlling,Seb2011}, superconducting nanowire single-photon detectors (SNSPDs)~\cite{lydersen2011controlling}, and so forth. A full field implementation of the attacking strategy has been investigated in~\cite{gerhardt2011full}. The blinding attack was also demonstrated to fake the violation of Bell's inequality~\cite{gerhardt2011experimentally}. How to remove the detector-control attacks is a challenge in the field of QKD. One proposed countermeasure is carefully operating the single-photon detectors inside Bob's system and monitoring the photocurrent for anomalously high values~\cite{yuan2010avoiding2,yuan2011resilience}. This work also highlights that mis-operation of QKD devices allows the loophole to be exploited, which is related to the best-practice criteria for all QKD devices in QKD implementations~\cite{koehler2018best}. Recently, Qian et al. propose another countermeasure against the detector-control attacks by introducing a variable attenuator in front of the detector~\cite{qian2019robust}. However, these countermeasures may seem \emph{ad hoc} and lead away from provable security models of QKD and can often be defeated by advanced hacking technologies. A practical and promising solution is the MDI-QKD protocol, which will be reviewed in section~\ref{sec:5:mdi}. \subsection{Other attacks} \label{sec:3:otherattacks} \begin{figure}[hbt] \centering \resizebox{7.8cm}{!}{\includegraphics{trojan.pdf}} \caption{Schematic illustration of the Trojan-horse attack~\cite{jain2015risk}. Eve attacks Alice by sending bright Trojan-horse pulses to know Alice's selected basis information. This information is carried by the back-reflected pulses coming out of Alice. [Figure reproduced from~\cite{jain2015risk}].} \label{fig:trojan} \end{figure} Another well-know hacking strategy is the Trojan-horse attack (THA), as shown in Fig.~\ref{fig:trojan}, in which Eve sends a probe light to Alice or Bob and reads their information from the backscattered probe light. In 2001, Vakhitov et al. proposed the large pulse attack~\cite{vakhitov2001large} and Kurtsiefer et al. analyzed the possibility of THA by detecting the detector fluorescence of Si-based avalanche photodiodes~\cite{kurtsiefer2001breakdown}. Gisin \emph{et al.} studied the problem of THA in the QKD implementations where light goes two ways~\cite{gisin2006trojan}. Later, Jain \emph{et al.} performed a comprehensive analysis of the risk of THA against typical components in standard QKD systems~\cite{jain2014trojan,jain2015risk}. Recently, the backflash photons caused by detection events in single-photon detectors was exploited to realize the detector-backflash attack~\cite{pinheiro2018eavesdropping}. A countermeasure against the THA is to add proper isolations and consider the leaking information in privacy amplification, which will be reviewed in Section~\ref{sec:4:THA}. Besides the above attacks, Lamas-Linares and Kurtsiefer demonstrated that the timing information revealed during public communicating can be exploited to attack the entanglement-based QKD system~\cite{lamas2007breaking}. In two-way QKD system such as the ``plug-and-play" structure, Sun \emph{et al.} studied the imperfections of Faraday mirror and proposed the Faraday-mirror attack~\cite{Sun2011}; Jain \emph{et al.} experimentally demonstrated that the calibration routine of a commercial ``plug-and-play" system can be tricked into setting a large detector efficiency mismatch, and proposed an attack strategy on such a compromised system with a QBER less than 7\% \cite{Jain2011}. Moreover, Li et al.~\cite{Li2011} studied the imperfection of a practical beam splitter and demonstrated a wavelength-dependent beam-splitter attack on top of a polarization-coding QKD system. The detector dead-time issue was widely studied in~\cite{rogers2007detector} and demonstrated in~\cite{Wei2011}. Andun \emph{et al.} and Makarov \emph{et al.} demonstrated the laser damage attack by using a high-power laser to damage the SPDs~\cite{Audun2014,Vadim2016}. Recently, Huang \emph{et al.} show that the decoy states can be distinguishable if they were generated by modulating the pump current of a semiconductor laser diode~\cite{huang2018quantum}, and Wei \emph{et al.} exploited the efficiency mismatch in the polarization degree of freedom to hack SNSPD~\cite{wei2019implementation}. Most of the imperfections reviewed so far are in fiber-based QKD systems. There are also quantum attacks reported for free-space QKD systems~\cite{nauerth2009information,sajeed2015security,chaiwongkhot2019eavesdropper}. For instance, imperfect encoding methods result in side channels from which encoded states are partially distinguishable~\cite{nauerth2009information}. The imperfection due to non-single-mode quantum signals is a crucial issue in free-space QKD. Eve can exploit this imperfection and launch the spatial-mode attack against a free-space QKD system. This problem has been carefully studied in~\cite{sajeed2015security,chaiwongkhot2019eavesdropper}, following an earlier discussion on the origins of detection efficiency mismatch in~\cite{fung2009security}. Besides DV-QKD, the practical security of CV-QKD also deserves future investigations, which will be reviewed in Section~\ref{cvqkd:hacking}. More generally, as noted in~\cite{curty2017quantum}, in principle, there are simply too many side channels for Alice and Bob to close. This is because Eve might , in principle, attack Alice's and Bob's system via X-ray, neutrons, neutrinos or even gravitational waves. And, whatever detection systems Alice and Bob have will probably have limited ranges of responses. Moreover, classical post-processing units pose a serious threat to the security of QKD. Most QKD security framework assumes without proofs that classical post processing units are secure. However, in conventional security, it is well known that hardware Trojans and software Trojans are commonly used to compromise the security of conventional cryptographic system. It was proposed in~\cite{curty2017quantum} to use redundancies in QKD units and classical post-processing units to achieve security through e.g., verifiable secret sharing. \section{Source Security}\label{sec:4} In this section, we review various approaches to resolve the security issues of practical sources. On one hand, the imperfections in quantum state preparation, including multi-photon components of laser, nonrandomized phases, encoding flaws and so forth, need to be carefully quantified and taken into account in security analysis. In particular, we will discuss the decoy state QKD protocol in more details. On the other hand, practical countermeasures are required to prevent the Trojan horse attacks on the source. Note that we focus on the BB84 protocol, but most of techniques can be extended to other protocols. \subsection{Decoy-state method} \label{sec4:decoy} Decoy-state method is a common way to combat with source imperfection by introducing extra sources for better channal characterization. In the decoy-state method, the user randomly modify the source states during the quantum stage; after that, he reveal which state is used in each turn. Eve cannot modify her attack to different source states, but in the postprocessing the users can estimate their parameters conditioned on that knowledge. The decoy-state method is used mostly to bound the multi-photon components in a practical photon source. In practical photon sources, multi-photon components are inevitable. As reviewed in Section~\ref{sec:3:PNS}, Eve can split a multi-photon pulse and save one photon from it for later hacking. Since Alice and Bob cannot tell whether a detection comes from a single-photon component or multi-photon component and Eve controls the channel, they have to pessimistically assume all the multi-photon states cause clicks with 100\% efficiency. All the losses come from the single-photon states. In order to reduce the effects of multi-photon components, Alice has to use very low intensity optical pulses. In the case of coherent state photon source, it has been shown that the optimal intensity used is close to the channel transmittance $\eta$ \cite{lutkenhaus2000security,Ma2006low}, \begin{equation} \label{eq:nodecoyoptmu} \begin{aligned} \mu_{opt} \approx \eta, \end{aligned} \end{equation} where $\eta$ includes channel transmission and detection efficiency. Then, final key rate will quadratically depend on the transmittance $\eta$, $R=O(\eta^2)$. Various protocols~\cite{DPS:2002,hwang2003quantum,Scarani2004} have been proposed to over the key rate limit caused by the PNS attacks, among which the most effective one is the decoy state method~\cite{hwang2003quantum,lo2005decoy,wang2005beating}. In the decoy-state QKD scheme, instead of using one intensity for encoding, Alice employs a few additional intensities of optical pulses (as decoy states) in order to monitor the transmittance of different photon number components. After Bob detects the signals, Alice announces the intensities she uses for each pulse. With detection rate for decoy states, Alice and Bob can bound tightly the number of detections from single-photon components. If Eve simply changes the transmittance for different photon number states as adopting in the PNS attacks, she will inevitably change the detection rates for signal and decoy states differently. Without Alice's intensity information ahead, Eve has to let a significant amount of single-photon states passing in order to maintain the ratio of detection rates among signal and decoy states. The decoy state idea was first proposed by Hwang~\cite{hwang2003quantum} who considered using a strong decoy signal with intensity around 2 photons as a decoy state. The security proof of the decoy-state method is given later in \cite{lo2005decoy}, where a photon number channel model \cite{Ma2008PhD} is employed. With an infinite number of decoy state, Alice and Bob can estimate the detections from all photon number components accurately. After adopting the GLLP security analysis, reviewed in Section~\ref{sub:GLLP}, one can show that the optimal intensity of optical pulses can be increased to $O(1)$, which results in a key rate having a linear dependence of transmittance, $O(\eta)$ \cite{lo2005decoy}. The decoy-state method significantly booming the performance of practical QKD. The schematic diagram of the decoy-state method is shown in Fig.~\ref{fig:decoyimplement}. In the meantime, practical decoy-state methods with only a vacuum and \emph{weak} decoy states were proposed \cite{Lo2004Vacua,Ma2004PhD} and tight bounds were derived later \cite{wang2005beating,ma2005practical}. In the original security proof, continuous phase randomization is assume to decohere phases between different photon number components. As discussed later in Section \ref{Sc:NonRandom}, phase randomization is necessary but can be relaxed to discrete phase randomization \cite{Cao2015Discrete}. In fact, the uniformly discrete phase randomization with discrete phase number $m=10$ can already achieve a good approximation of continuous phase randomization. \begin{figure*}[hbt] \centering \resizebox{13cm}{!}{\includegraphics{decoyscheme.pdf}} \caption{Schematic diagram of decoy-state QKD. In a decoy-state BB84 transmitter, the optical pulses are normally generated with phase-randomized laser pulses. Decoy states are prepared using an amplitude modulator (AM). In the figure: (PM) phase modulator for phase randomization, (PolM) polarization modulation for encoding, (F) optical filter and (I) optical isolator. } \label{fig:decoyimplement} \end{figure*} \subsubsection{Theory} For the source with different photon number components, one can assume a photon number channel model~\cite{Ma2008PhD}. The decoy state method is a tomography to the photon number channel model, providing tighter estimations on single photon component~\cite{lo2005decoy}. In the decoy state method, the source is operated at different photon number distributions, leading to different measurement outcome statistics. The communication partners can estimate the channel parameters of yield $Y_n$ and QBER $e_n$ for each photon number component. One crucial assumption in the decoy-state QKD is that the signal state and decoy states are \emph{identical} except for their average photon numbers. This means after Eve's photon-number measurement, she has no way of telling whether the resulted photon number state is originated from the signal state or decoy states. Hence, the yield $Y_n$ and QBER $e_n$ can depend on only the photon number, $n$, but not which distribution (decoy or signal) the state is from. That is, \begin{equation} \begin{aligned} Y_n(signal) &= Y_n(decoy), \\ e_n(signal) &= e_n(decoy). \end{aligned} \end{equation} The implementations of decoy state method can be divided into active ones and passive ones. In the active decoy state method, the user perpares the source signals with different intensities to change the probability distributions of each photon number component. A simple solution for decoy state preparation, as shown in Fig.~\ref{fig:decoyimplement}, is to use an amplitude modulator (AM) to modulate the intensities of each WCP to the desired intensity level. This is indeed the implementations reported in most of decoy state QKD experiments. Another solution for decoy state implementation is to use multiple laser diodes of different intensities to generate different states~\cite{Peng2007}. In the passive decoy state method, heralded single photon source are often applied \cite{mauerer2007quantum,Adachi2007Simple,Ma2008PDC}. The probability distribution is changed by observing different measurement outcomes of the heralded photons. A popular source for decoy state method is the phase-randomized weak coherent state source, as shown in Eq.~\eqref{eq:coherentFock}. To apply the active decoy method, Alice randomly adjusts the intensity $\mu$ of the coherent state, which is related to different Poisson distribution $P_\mu(n)$. Alice estimates the single photon yield $Y_1$ and error $e_1$ by solving the equation provided by the observed gain $Q_\mu$ and quantum bit error rate (QBER) $E_\mu$ related to different intensity $\mu$ \begin{equation} \label{eq:QEmu} \begin{aligned} Q_\mu &= \sum_{n=0}^{\infty} P_\mu(n) Y_n, \\ E_\mu Q_\mu &= \sum_{n=0}^{\infty} P_\mu(n) e_n Y_n, \end{aligned} \end{equation} where $P_\mu(n)=\mu^n e^{-\mu}/n!$ for the coherent state case. Following the GLLP security analysis, Eq.~\eqref{eq:GLLP}, the key rate is given by \begin{equation} \label{eq:decoyR} \begin{aligned} R &\ge -Q_\mu H(E_\mu) + Q_1[1-H(e_1)], \end{aligned} \end{equation} where $Q_1=Y_1 \mu e^{-\mu}$. Here, the gain $Q_\mu$ and QBER $E_\mu$ can be directly obtained from experiment, and the signal intensity $\mu$ is set by Alice. With a tight estimation on $Y_1$ and $e_1$ by solving the linear equations in the form of Eq.~\eqref{eq:QEmu}, the key rate can be improved from $O(\eta^2)$ to $O(\eta)$. In practice, only several different intensities are enough to make an accurate estimation. The most popular practical decoy state method is vacuum and weak decoy state method \cite{Lo2004Vacua,Ma2004PhD}. That is, Alice randomly generates coherent states with three different intensities $\{0,\nu,\mu\}$, where states with intensity $\mu$ is the signal states for key generation, and states with intensity $\nu<\mu$ and vacuum state with intensity $0$ is for parameter estimation. The two parameters need to estimate in Eq.~\eqref{eq:decoyR} can be bounded by \cite{ma2005practical}, \begin{equation}\label{Practical:Y1Le1U} \begin{aligned} Y_1 &\ge Y_1^L = \frac{\mu}{\mu\nu-\nu^2}\big(Q_\nu e^{\nu}-Q_\mu e^{\mu}\frac{\nu^2}{\mu^2} -\frac{\mu^2-\nu^2}{\mu^2}Y_0\big) \\ e_1 &\le e_1^U = \frac{E_\nu Q_\nu e^\nu-e_0Y_0}{Y_1^L\nu}. \end{aligned} \end{equation} A similar result is also derived in \cite{wang2005beating}. For finite-data size effect, Ma \emph{et al.} took the first step to analyze the statistical fluctuations using standard error analysis, which essentially assumes i.i.d.~channel behavior \cite{ma2005practical}. The idea is that instead of directly using $Q_\mu$ and $E_\mu$ obtained from the experiment directly, one assumes these parameters fluctuates according to a normal distribution. Then, in Eq.~\eqref{Practical:Y1Le1U}, one can substitute the upper and lower bounds of $Q_\mu$ and $E_\mu$. The failure probability for this estimation would link to the number of standard deviations used for bounds. The finite data size effect has been discussed in a more rigorous manner in \cite{lim2014concise,Zhang2017Improved}. It turns out that the formulas used in the standard error analysis approach \cite{ma2005practical} can be directly applied with a different value of failure probabilities in parameter estimation, as presented in Table \ref{tab:deviations}. \begin{table} [htb] \centering \caption{The failure probability as a function of the fluctuation deviations, measured by the number of standard deviations, $(\chi-\mathbb{E}^L[\chi])/\mathbb{E}^L[\chi] = (\mathbb{E}^U[\chi]-\chi)/\mathbb{E}^L[\chi]=n\sigma$, where $\chi$ is counts in experiment. Here, $\varepsilon_{G}$, $\varepsilon_{\infty}$, $\varepsilon_{10000}$, and $\varepsilon_{70}$, respectively, denote failure probabilities for the bounds in the Gaussian approximate analysis, the rigourous method with large data size limit, and a data size of 10000 and 70. [A similar table is also presented in \cite{Zhang2017Improved}]}\label{tab:deviations} \begin{tabular}{cccccc} \hline \hline \textbf{Deviation} & \textbf{$\varepsilon_{G}$} & \textbf{$\varepsilon_{\infty}$} & \textbf{$\varepsilon_{10000}$} & \textbf{$\varepsilon_{70}$} \\ \hline $2\sigma$ & $10^{-1.34}$ & $10^{-0.57}$ & $10^{-0.57}$ & $10^{-0.57}$ \\ $3\sigma$ & $10^{-2.57}$ & $10^{-1.65}$ & $10^{-1.65}$ & $10^{-1.54}$ \\ $4\sigma$ & $10^{-4.20}$ & $10^{-3.17}$ & $10^{-3.17}$ & $10^{-2.65}$ \\ $5\sigma$ & $10^{-6.24}$ & $10^{-5.13}$ & $10^{-5.09}$ & $10^{-3.92}$ \\ $6\sigma$ & $10^{-8.70}$ & $10^{-7.52}$ & $10^{-7.43}$ & $10^{-5.36}$ \\ $7\sigma$ & $10^{-11.59}$ & $10^{-10.34}$ & $10^{-10.13}$ & $10^{-6.95}$ \\ $8\sigma$ & $10^{-14.91}$ & $10^{-13.60}$ & $10^{-13.18}$ & $10^{-8.67}$ \\ $9\sigma$ & $10^{-18.65}$ & $10^{-17.29}$ & $10^{-16.60}$ & $10^{-10.50}$ \\ $10\sigma$ &$10^{-22.82}$ & $10^{-21.41}$ & $10^{-20.36}$ & $10^{-12.38}$ \\ \hline \hline \end{tabular} \end{table} Other than weak coherent state photon sources, one can also use PDC source or thermal source, as reviewed in Section \ref{Sub:Source}. As long as the photon number distribution of the source is different from Poisson distribution, one can employ the passive decoy state scheme \cite{Adachi2007Simple,Ma2008PDC}, where Alice splits the pulses with a beam splitter, detects one arm as triggers, and uses the other arm for QKD encoding. Conditioned on the detection of triggering signals, the photon number distribution of the encoding arm is different. Alice can announce her local detection after Bob's detection. Then, they can have linear gains for photon pulses with different (conditional) photon number distributions for the decoy-state method analysis. It turns out that even with phase randomized coherent states, one can employ the passive decoy state method \cite{Curty2010PassiveDecoy}. \subsubsection{Experiment} \label{sec5:decoyexp} The decoy state methods have been widely implemented in different QKD systems. The decoy-state experiments are summarized in Table~\ref{Tab2}. Fig.~\ref{fig:decoy100} shows the four initial decoy-state QKD experiments. Zhao~\emph{et al.} reported decoy state experiments \cite{Yi2006,zhao2006simulation} up to 60-km fiber on top of a commercial plug-and-play QKD system; Peng \emph{et al.}~\cite{Peng2007} implemented decoy-state QKD over 102-km fiber using a one-way polarization-encoding QKD system; Rosenberg \emph{et al.}~\cite{rosenberg2007long} implemented decoy-state QKD over 107-km fiber using a one-way phase-encoding QKD system; Schmitt-Manderbach \emph{et al.} achieved 144 km decoy state QKD in free space~\cite{Tobias2007}. These experiments demonstrated that decoy state BB84 was secure and feasible under real-world conditions. \begin{figure*}[hbt] \centering \resizebox{16cm}{!}{\includegraphics{decoyexp.pdf}} \caption{(Color online) Decoy state QKD experiments. \textbf{a}, Experiment on a commercial plug-and-play QKD system~\cite{Yi2006}. CA, compensating AOM; CG, compensating generator; DA, decoy AOM; DG, decoy generator; LD, laser diode; $\phi$, phase modulator; PD, classical photo detector; DL, delay line; FM, faraday mirror. \textbf{b}, Phase-encoding experiment~\cite{rosenberg2007long}. DFB, distributed feedback laser; VOA, variable optical attenuator; AM, amplitude modulator; LP, linear polarizer. \textbf{c}, Polarization-encoding experiment~\cite{Peng2007}. FCN, fiber coupling network; FF, fiber filter; EPC, electric polarization controller; DAC, digital-to-analog converter. \textbf{d}, Free-space experiment~\cite{Tobias2007}. BS, beam splitter; PBS, polarizing beam splitter; HWP, half-wave plate; APD, avalanche photo diode. [Figures reproduced from~\cite{Yi2006,rosenberg2007long,Peng2007,Tobias2007}].} \label{fig:decoy100} \end{figure*} Since then, more and more experimental efforts have been made to QKD deployments in labs and field tests. In 2007, Yuan \emph{et al.} realized a stabilized one-way phase-encoding decoy state QKD system~\cite{yuan2007unconditionally}. Later, Dixon \emph{et al.}~implemented decoy state QKD with a high clock rate of 1 GHz~\cite{Dixon2008} and Liu \emph{et al.} extended decoy state QKD to long distance of 200-km fiber~\cite{liu2010decoy}. Importantly, a number of field QKD networks with the decoy-state implementation have been built in Europe~\cite{peev2009secoqc}, Japan~\cite{sasaki2011field}, China~\cite{chen2009field,chen2010metropolitan,wang2010field} and so forth. An illustration of the Tokyo QKD network is shown in Fig.~\ref{fig:tokyo}. \begin{figure}[hbt] \centering \resizebox{7.8cm}{!}{\includegraphics{tokyonetwork.pdf}} \caption{(Color online) Architecture of the Tokyo QKD Network~\cite{sasaki2011field}. [Figure reproduced from~\cite{sasaki2011field}].} \label{fig:tokyo} \end{figure} In the mean time, Wang et al.~experimentally implemented decoy state with a PDC source~\cite{Qin2008}. Also, the passive decoy-state method has also been demonstrated \cite{Sun2014Passive}. Recently, the decoy state experiment has been extended to a record-breaking distance of 1,200 km in free space~\cite{Liaosate} and 421 km in ultra low-loss optical fiber~\cite{boaron2018secure}. Due to its convenient implementation and remarkable enhancement on performance, the decoy-state method becomes a standard technique in current QKD implementations. \subsection{Source flaws} \subsubsection{Basis-dependent source} In practice, there is often some difference between $\rho_x$ and $\rho_z$, i.e., Eq.~\eqref{eq:basisindep} might not be fulfilled. Then, in the worst case scenario, we should assume Eve is capable of distinguishing the basis choice and hence she can attack two basis states separately. This kind of source is called basis-dependent source. Obviously, the more state dependence on the basis, the easier for Eve to distinguish the bases and hence a lower key rate. Without loss of generality, we take $Z$-basis as example. The general Shor-Preskill's key rate formula is \cite{shor2000simple} \begin{equation} \label{eq:ShorPreskillRbeta} \begin{aligned} r \ge 1-H(e_{Z})-H(e_{Z}^p), \\ \end{aligned} \end{equation} where $e_{Z}^p$ is the $Z$-basis phase error rate, defined in Eq.~\eqref{eq:biterror}. For a basis-dependent source, $e_{Z}^p \neq e_{X}$ since $\rho_Z \neq \rho_X$. However, if $\rho_X$ is close to $\rho_Z$, we can still bound $e_{Z}^p$ from measured $e_X$. In the GLLP security analysis framework~\cite{gottesman2004security}, the basis dependence is quantified by a bias, \begin{equation} \label{eq:SourceBias} \begin{aligned} \Delta=\frac{1-F(\rho_X,\rho_Z)}{2}, \\ \end{aligned} \end{equation} where $F(\rho_X,\rho_Z)=\sqrt{\sqrt{\rho_Z}\rho_X\sqrt{\rho_Z}}$ is the fidelity between the two states. Given this bias, the phase error rate used in the key rate formula can be bounded by \cite{koashi2009simple,Lo2007Nonrandom}, \begin{equation} \label{eq:ephasebound} \begin{aligned} e_z^p \le e_j^b+4\Delta(1-\Delta)(1-2e_x^b) \\+4(1-2\Delta)\sqrt{\Delta(1-\Delta) e_x^b(1-e_x^b)}. \\ \end{aligned} \end{equation} For the practical photon sources presented in Section \ref{Sub:Source}, Alice and Bob have more information than the bias in Eq.~\eqref{eq:SourceBias}. For example, in principle, they can measure the photon number $n$, with which they can tag each quantum signal. Then, in phase error correction of entanglement distillation process, which would be reduced to privacy amplification for prepare-and-measure schemes, they could take advantage of these tagging. With tagging, the GLLP key rate formula can be written as \cite{gottesman2004security}, \begin{equation} \label{eq:GLLPR} \begin{aligned} r \ge -H(E)+(1-\Delta)[1-H(e_z^p)], \\ \end{aligned} \end{equation} where $E$ is the total QBER, $\Delta$ is the ratio of tagged signals, and $e_z^p$ is the phase error rate of the untagged signals. Here, we use the same notation of the bias $\Delta$ in Eq.~\eqref{eq:SourceBias}. \subsubsection{Nonrandom phase}\label{Sc:NonRandom} A general example of source flaw is to use the weak coherent states with nonrandom phases to encode the basis and key information \cite{Lo2007Nonrandom}. Their difference is treated as source flaw, i.e, a basis-dependence of the source. The encoded state $\ket{\psi_{\beta\kappa}}_B$ is \begin{equation} \ket{\psi_{\beta\kappa}}_B = \ket{\alpha}_R \ket{\alpha e^{i\pi(\kappa + \frac{1}{2} \beta)}}_S, \end{equation} where $\alpha$ is constant and $\mu=2|\alpha|^2$ is the intensity. In this case, the basis dependence $\Delta$ is \begin{equation} \Delta = \dfrac{1}{2}\left( 1 - e^{\mu/2}( \cos(\mu/2)+\sin(\mu/2) ) \right) = \mu/8 + O(\mu^3). \end{equation} Note that in the practical QKD experiment, we will post-select the clicked signals. In this case, to calculate the basis dependence, we have to take the channel transmittance $\eta$ into account. In a worst-case scenario, the channel loss is caused by Eve's selection on the transmitted signals. To clarify this, we can consider Eve performs a unambiguous state discrimintation (USD) attack \cite{Duifmmode2000unambiguous}, where Eve performs USD to discriminate $\rho_X$ and $\rho_Z$. If the discrimination is successful, Eve can learn the basis and key, then he generates the same state $\rho_{\beta\kappa}$, and sends it to Bob; if the discrimination fails, Eve partially blocks the signal as loss. In this case, the basis dependence $\Delta^\prime$ of left signals will be amplified by $\eta$ \begin{equation} \label{eq:Deltaprime} \Delta^\prime = \Delta/(\eta\mu) \approx \mu/(8\eta). \end{equation} From Eqs.~\eqref{eq:ShorPreskillRbeta},~\eqref{eq:ephasebound},and~\eqref{eq:Deltaprime}, we can calculate the key rate. However, the achievable key generation rate scales only quadratically with the transmittance $\eta$ in the channel, i.e., $r = O(\eta^2)$. This question can be potentially solved using the scheme of discrete phase randomization~\cite{Cao2015Discrete}. \subsubsection{Encoding flaws} Another example of source flaw is the encoding flaws in the phase and polarization encoding due to the device imperfections in the encoding devices. This will also make the source basis-dependent. Although GLLP allows the security proof to consider th encoding flaws, the key rate drops dramatically~\cite{gottesman2004security}. This is because GLLP has a pessimistic consideration by assuming that the encoding flaws are in arbitrary dimensions. To address this issue, a loss-tolerant protocol was proposed in~\cite{tamaki2014loss}, which makes QKD tolerable to channel loss in the presence of source flaws~\cite{yin2014mismatched}. On the basis of the assumption that the single-photon components of the states prepared by Alice remain inside a two-dimensional Hilbert space, it was shown that Eve cannot enhance state preparation flaws by exploiting the channel loss and Eve’s information can be bounded by the rejected data analysis. The intuition for the security of loss-tolerant QKD protocol~\cite{tamaki2014loss} can be understood in the following manner. By assuming that the state prepared by Alice is a qubit, it becomes impossible for Eve to perform an unambiguous state discrimination (USD) attack. Indeed, in order for Eve to perform a USD attack, the states prepared by Alice must be linearly independent; but by having three or more states in a two-dimensional space, in general the set of states prepared by Alice is linearly dependent, thus making USD impossible. The above loss-tolerant protocol has been further developed and demonstrated experimentally for decoy-state BB84~\cite{xu2014experimental,boaron2018secure} and MDI-QKD~\cite{tang2016experimental}. \subsection{Leaky source} \label{sec:4:THA} As discussed in Section~\ref{sec:3:otherattacks}, the source is vulnerable under the Trojan-horse attack (THA). In particular, Eve could inject bright light pulses into Alice's transmitter and then measure the back-reflected light to extract information about Alice's state preparation process. This problem has been analyzed in~\cite{lucamarini2015practical}. The authors evaluated the security of a QKD system in the presence of information leakage from Alice's phase modulator (PM), which is used to encode the bit and basis information of the generated signals. A key observation is that, the joint state of Alice's transmitted signals and Eve's back-reflected light from her THA is not basis-independent but it depends on Alice's basis choice. The security of the system can be analyzed by quantifying Eve's information and considering this information in privacy amplification, based on the techniques introduced in~\cite{Lo2007Nonrandom}. Recently, these seminal results have been generalized to prove the security of decoy-state QKD in the presence of arbitrary information leakage from both the PM and the intensity modulator (IM)~\cite{tamaki2016decoy,WangNJP2018}. Here the IM is normally used to select the intensity setting for each emitted signal. Consequently, it is possible to quantify the amount of device isolation, against THA, so as to achieve a certain performance with a realistic leaky QKD system. \section{Detection Security}\label{sec:5} In this section, we review the various approaches to address the detection security of practical QKD. We will review the MDI-QKD protocol and its extensions in more details. \subsection{Countermeasures against detection attacks} Many approaches have been proposed to defeat the attacks at detection. The first one is \emph{security patch}. That is, once one discovers a new type of attack, a corresponding countermeasure against this attack can be proposed and realized in an existing QKD system. This approach usually only requires modifying the software or the hardware of a current system. For instance, the time-shift attack introduced in the previous section can be avoided by simply shifting the gating window of the detectors at random~\cite{Qi2007}. The detector blinding attack could, in principle, be avoided by monitoring the detector's photocurrent for anomalously high values~\cite{yuan2011resilience,da2012real} or by randomly varying the detector efficiency~\cite{lim2015random}. Although security patch can defeat certain attacks, the patched countermeasures themselves might open other loopholes. This could as a result, introduce one more layer of security risk~\cite{Shihan2015,huang2016testing,qian2019robust}. Furthermore, the major issue associate with the security patch is that they only prevent the known attacks. For potential and \emph{unknown} attacks, the countermeasures may fail. Therefore, security patch is only ad-hoc, which abandons the information-theoretic security framework of QKD. The second approach is to \emph{fully characterize} the devices using in a QKD system and precisely describe the devices in mathematical models. Then the models can be included in the security proof to estimate the real secure key rate based on an imperfect setup. A well-known example is the GLLP security proof~\cite{gottesman2004security}. While this approach seems straightforward, developing models to fully match the practical behavior of various QKD devices is rarely possible because the components are complex. Even so, there are several ongoing theoretical efforts to consider as many imperfections as possible into the security proof~\cite{fung2009security,maroy2010security,tamaki2014loss,lucamarini2015practical,tamaki2016decoy}. Nevertheless, this approach is limited by our understanding of the devices and a complete knowledge of the devices is rather challenging. Hence full characterization is still \emph{ad-hoc}. The third approach is device-independent QKD (DI-QKD), which will be reviewed in Section~\ref{sec:diQKD}. Note that there are also proposals for semi-device-independent QKD, where one party's measurements are fully characterized while the other's are unknown~\cite{PhysRevA.85.010301,pawlowski2011semi,smith2012conclusive}. The final approach is MDI-QKD protocol, which closes \emph{all} detection attacks and is practical with current technology. In below, we will review MDI-QKD in detail. \subsection{Measurement-device-independent scheme} \label{sec:5:mdi} MDI-QKD generates secret keys based on the ``time-reversed" entanglement protocol and leaves all the single photon detections to a public \emph{untrusted} relay (Eve). \subsubsection{Time-reversed EPR QKD} \begin{figure}[hbt] \begin{center} \includegraphics[scale=0.66,angle=0]{eprprotocol.pdf} \end{center} \caption{Einstein-Podolsky-Rosen (EPR) based QKD protocol~\cite{biham1996quantum}. One particle of each EPR correlated pair, denoted by dashed lines, is sent to the center, who performs a Bell state measurement (BSM). The second particles are sent to Alice and Bob, respectively, who project them onto the BB84 states. \textbf{a}, Original EPR QKD. The first measurement is done by the center and the particles arriving at Alice and Bob are therefore in the Bell state, which can be used to do QKD as in the EPR based protocol~\cite{ekert1991quantum}. \textbf{b}, Time-reversed EPR QKD. The first measurement is performed by Alice and Bob and each particle sent to the center is therefore in one of the BB84 states, which forms the concept behind MDI-QKD~\cite{lo2012measurement}. [Figure reproduced from~\cite{biham1996quantum}]. \label{Fig:eprprotocol}} \end{figure} The idea of measurement-device-independent (MDI) is inspired by the Einstein-Podolsky-Rosen (EPR) based QKD protocol~\cite{ekert1991quantum,bennett1992quantum}. This is illustrated in Fig.~\ref{Fig:eprprotocol}~\cite{biham1996quantum}. In the initial EPR-based protocol (Fig.~\ref{Fig:eprprotocol}a), Alice and Bob individually prepare an EPR pair at each side and send one photon from each pair to an untrusted center party, Charles. Charles then performs a Bell state measurement (BSM) for entanglement swapping. The measurement result is announced. Once the BSM is finished, Alice and Bob measure the other photon of the EPR pairs locally by choosing between the $X$ and $Z$ basis randomly. Comparing a subset of their measurement results allows Alice and Bob to know whether Charles is honest. Then Alice and Bob can generate the secret by using the BBM92 protocol \cite{bennett1992quantum}. Note that the EPR protocol can also work in a \emph{time-reversal} version, as shown in Fig.~\ref{Fig:eprprotocol}b. That is, Alice and Bob can measure their local photons first, instead of waiting for Charles' measurement results. This order of preparation and measurement is equivalent to that of the prepare-and-measurement QKD scheme, in which Alice and Bob prepare BB84 states and send them to Charles to perform the BSM. After that, the Charles' honesty can be checked by comparing a part of Alice's and Bob's results. Importantly, Charles' BSM is only used to check the \emph{parity} of Alice's and Bob's bits, and thus, it does not reveal any information about the individual bit values. This time-reversal EPR protocol forms the main concept behind MDI-QKD. This time-reversed EPR QKD protocol has been first proposed in~\cite{biham1996quantum}. Later, Inamori provided a security proof~\cite{inamori2002security}. Nevertheless, these two important works offered very limited performance and, therefore, they have been largely forgotten by the QKD community. For instance, the scheme in~\cite{biham1996quantum} requires perfect single-photon sources and long-term quantum memories, which renders it unpractical with current technology. Inamori's scheme~\cite{inamori2002security} uses practical weak coherent pulses (WCPs) but it does not include decoy states, since it was proposed long before the advent of the decoy-state protocol. Moreover, the two early papers~\cite{biham1996quantum,inamori2002security} were not specifically considering the side channel problem in QKD at all. In 2012, Braunstein and Pirandola performed a general security analysis of the time-reversed EPR QKD approach~\cite{braunstein2012side} and proved that detector side-channel attacks can be eliminated by using teleportation where any incoming quantum signals are excluded from access to the detectors. Note that the idea of using teleportation for the specific purpose of removing side channels was first discussed in footnote 21 of~\cite{lo1999unconditional}. \subsubsection{MDI-QKD protocol} \begin{figure*}[hbt!] \centering \resizebox{15cm}{!}{\includegraphics{mdiqkd2.pdf}} \caption{Schematic diagram of MDI-QKD proposed in~\cite{lo2012measurement}. Alice and Bob prepare BB84 polarization states using a decoy-state BB84 transmitter, same as the one illustrated in Fig.~\ref{fig:decoyimplement}. They send BB84 states to an \emph{untrusted} relay Charles/Eve. The relay is supposed to perform a Bell state measurement (BSM) that projects Alice's and Bob's signals into a Bell state. }\label{Fig:mdiqkd} \end{figure*} \begin{table}[!ht]\center \begin{tabular}{l|l|l} \hline \hline \ & \textbf{Singlet state $\ket{\psi^{-}}$} & \textbf{Triplet state $\ket{\psi^{+}}$} \\ \hline Coincident clicks & $D_{1H}$ \& $D_{2V}$ or & $D_{1H}$ \& $D_{1V}$ or \\ \ & $D_{2H}$ \& $D_{1V}$ & $D_{2H}$ \& $D_{2V}$ \\ \hline Rectilinear basis & Bit flip & Bit flip \\ \hline Diagonal basis & Bit flip & -- \\ \hline \hline \end{tabular} \caption{Post-selection for MDI-QKD~\cite{lo2012measurement}. Alice and Bob post-select the events where the relay outputs a successful result and they use the same basis in their transmission. Moreover, either Alice or Bob flips her/his bits, except for the cases where both of them select the diagonal basis and the relay outputs a triplet}~\label{Tab:mdipost} \end{table} The MDI-QKD proposal~\cite{lo2012measurement} [see also~\cite{braunstein2012side}] builds on the time-reversed EPR QKD. In particular, the main merits of the proposal, introduced by Lo, Curty and Qi~\cite{lo2012measurement}, are twofold: first, it identified the importance of the results in~\cite{biham1996quantum,inamori2002security} to remove all detector side-channels from QKD implementations; second, it significantly improved the system performance with practical signals by including decoy states. The protocol can be summarized in four steps: \begin{enumerate} \item Alice and Bob randomly and individually prepare one of four BB84 states using phase-randomized WCPs together with decoy signals. Then they send the states to an untrusted party, Charles. \item An honest Charles performs a BSM that makes Alice's and Bob's states interfere with each other, generating a Bell state. An example of a BSM implementation with linear optics in shown in Fig.~\ref{Fig:mdiqkd}: Charles interferes the incoming pulses at a 50:50 beam-splitter (BS), which has on each end a polarizing beam-splitter (PBS) that projects the photons into either horizontal ($H$) or vertical ($V$) polarization states. A ``click'' in the single-photon detectors $D_{\rm 1H}$ and $D_{\rm 2V}$, or in $D_{\rm 1V}$ and $D_{\rm 2H}$, indicates a projection into the singlet state $\ket{\psi^{-}}=(\ket{HV}-\ket{VH})/\sqrt{2}$, while a ``click'' in $D_{\rm 1H}$ and $D_{\rm 1V}$, or in $D_{\rm 2H}$ and $D_{\rm 2V}$, implies a projection into the triplet state $\ket{\psi^{+}}=(\ket{HV}+\ket{VH})/\sqrt{2}$. Other detection patterns are considered unsuccessful. \item Whether Charles is honest or not, he announces the outcome of his claimed BSM using a classical public channel when he claims to obtain a successful measurement. \item Alice and Bob keep the data that correspond to Charles' successful measurement events and discard the rest. Next, similar to the sifting in BB84 protocol, Alice and Bob announce their basis choices for sifting the events and keep the events using same bases. Based on Charles' measurement result, Alice flips part of her bits to guarantee the correct correlation with those of Bob. The post-selection strategy is illustrated in Table~\ref{Tab:mdipost}. Finally, they use the decoy-state method to estimate the gain and QBER of the single-photon contributions. \end{enumerate} In MDI-QKD, both Alice and Bob are senders, and they transmit signals to an untrusted third party, Eve, who is supposed to perform a Bell state measurement (BSM). Since the BSM is only used to post-select entanglement, it can be treated as an entirely \emph{black} box. Hence, MDI-QKD can remove all detection side-channels. The assumption in MDI-QKD is that the source should be trusted. The security assumptions of MDI-QKD, together with DI-QKD (see Section~\ref{sec:diQKD}), are summarized in Table~\ref{Tab:assumptions}. A comparison of practical security between MDI-QKD and DI-QKD, as commented by Charles H. Bennett in QCrypt 2018\footnote{See slide 6 of Charles H. Bennett's talk in Lightning Talks session of QCrypt 2018: http://2018.qcrypt.net/}, is summarized in Box~\ref{tab:bennett}. \begin{tcolorbox}[title = {Box VI.B.2: A security remark about MDI-QKD and DI-QKD by Charles H. Bennett.}] MDI-QKD at first sounds weaker than DI-QKD, but in fact it is stronger. In MDI-QKD, Eve's untrusted device remains outside Alice's and Bob's trusted enclosures. They need only trust themselves not to have inadvertently created a side channel to Eve through incompetent design of their do-it-yourself (DIY) light sources. By contrast, in DI-QKD they must trust Eve not to have deliberately created side channels from the untrusted devices to herself. \end{tcolorbox}~\label{tab:bennett} \begin{table}[!ht]\center \begin{tabular}{l|l|l} \hline \hline \ & \textbf{DI-QKD} & \textbf{MDI-QKD} \\ \hline True random number generators & Yes & Yes \\ \hline Trusted classical post-processing & Yes & Yes \\ \hline Authenticated classical channel & Yes & Yes \\ \hline No unwanted information leakage & \ & \ \\ from the measurement unit & Yes & \textbf{No} \\ \hline Characterized source & \textbf{No} & Yes \\ \hline \hline \end{tabular} \caption{Security assumptions in DI-QKD and MDI-QKD. While DI-QKD has the advantage of being applicable to an uncharacterized source, it demands no unwanted information leakage from the measurement unit. MDI-QKD applies to {\it any} measurement units. This means that the measurement unit in MDI-QKD can be an entire black-box, purchased from untrusted vendors.}~\label{Tab:assumptions} \end{table} \subsubsection{Theoretical developments} The decoy-state analysis is essential for MDI-QKD. The analysis is different from that of conventional decoy-state BB84 in that now both Alice and Bob send decoy signals to a common receiver (instead of only Alice sending decoy signals to Bob), which makes the mathematics slightly more complex. Fortunately, it has been shown that it is enough to obtain a tight estimation if Alice and Bob employ just a few decoy settings each. The authors of~\cite{ma2012statistical} and~\cite{wang2013three}, respectively, proposed a numerical method based on linear programming and an analytical approach based on Gaussian elimination. Both approaches assume that Alice and Bob can prepare a vacuum intensity. Following the similar analytical line, the authors of~\cite{Feihu:practical} studied the situation where none of the two decoy intensities are vacuum\footnote{A vacuum state is normally hard to realize in practice due to the finite extinction ratio of a practical intensity modulator}. A full parameter optimization method was proposed in~\cite{PhysRevA.89.052333}. Soon after, Yu et al proposed to use joint constraints for a better key rate~\cite{yu2015statistical}, and Zhou et al.~\cite{zhou2016making} proposed a four-intensity method, in which the key generation is conducted in $Z$ basis and the decoy analysis is performed only in $X$ basis. By doing so, the four-intensity method is efficient in the case of short data size. Recently, the four-intensity method was extend to seven-intensity method which can substantially enhance the key rate for MDI-QKD over asymmetric channels~\cite{wang2018enabling}. All these results provide experimentalists a clear path to implement MDI-QKD with a finite number of decoy states. For finite-key analysis, the authors of~\cite{ma2012statistical} provided an analysis that assumes a Gaussian distribution for the statistical fluctuations. Remarkably, Curty et al.~\cite{curty2014finite} presented a rigorous finite-key security proof against general attacks by using min-entropy analysis and Chernoff bound. In addition, this result satisfies the composable security definition. All these results confirm the feasibility of long-distance implementations of MDI-QKD within a reasonable time-frame of signal transmission. Simulations of the secret key rates with different kinds of decoy-state methods and finite-key analysis methods is shown in Fig.~\ref{Fig:mdirate}. \begin{figure*}[!ht] \centerline{\includegraphics[width=15cm]{mdirate.pdf}} \caption{(Color online) Simulation and experimental secret rates of MDI-QKD demonstrated in~\cite{Yin2016,Comandar}. \textbf{a}, 404-km MDI-QKD~\cite{Yin2016}. The experimental results (symbols) agree well with the theoretical simulations (solid lines). The dotted lines from upper to bottom show, respectively, the simulations for the balanced-basis passive BB84 protocol using ideal single-photon (SP) sources, the practical SP without the decoy-state method, the WCS with the decoy-state method, and the results of~\cite{yanlin2014}. \textbf{b}, 1-GHz MDI-QKD~\cite{Comandar}. Filled squares refer to key rates without the finite-size analysis. The star is the key rate obtained using two 25-km spools of fibre. The filled and open dots represent key rates with the finite-size analysis. The finite-size distillation methods are (1) standard error analysis~\cite{ma2012statistical}, and (2) composable security analysis~\cite{curty2014finite}. The dashed lines are simulations of the key rate for two different detector temperatures. [Figures reproduced from~\cite{Yin2016,Comandar}].}\label{Fig:mdirate} \end{figure*} Other practical aspects were also extensively analyzed in theory. Besides polarization encoding in the original MDI-QKD protocol~\cite{lo2012measurement}, alternative schemes including phase encoding~\cite{tamaki2012phase} and time-bin encoding~\cite{Ma2012Alternative} were proposed and analyzed. To extend the transmission distance further, one could include quantum memories~\cite{abruzzo2014measurement,panayi2014memory}, entanglement sources~\cite{xu2013long} or adaptive operations \cite{azuma2015all2}. Moreover, a key security assumption in MDI-QKD is the source should be trusted. Recently, there have been efforts to prove the security of MDI-QKD when Alice's and Bob's encoding devices are flawed~\cite{tamaki2014loss,xu2014experimental}, or when their apparatuses are not fully characterized~\cite{yin2014mismatched}. Furthermore, a plug-play type of MDI-QKD was proposed in~\cite{xu2015measurement,choi2016plug} and experimentally demonstrated in~\cite{tang2016experimental}. \subsubsection{Experimental developments} Table~\ref{Tab3} summarizes the MDI-QKD experiments after its invention. The main experimental challenge of MDI-QKD is to perform a high-visibility two-photon interference between photons from two (Alice's and Bob's) independent laser sources~\cite{lo2012measurement}, which is not required in conventional QKD schemes. To do so, Alice's photons should be indistinguishable from those of Bob. Importantly, if one implements MDI-QKD over telecom fibres, it is necessary to include feedback controls to compensate the time-dependent polarization rotations and propagation delays caused by the two separated fibres. Note that in standard BB84 QKD systems, the requirement of compensating polarization rotations and propagation delays can be relaxed by using phase encoding, because the two optical pulses, which interfere with each other at the receiver's end, pass through the same optical fibre and thus experience the same polarisation rotation and phase change. Therefore, one can achieve high interference visibility without performing any polarization control. Nevertheless, this advantage of phase encoding (in comparison to other encoding schemes) cannot be directly translated to MDI-QKD, because the two pulses pass through two \emph{independent} quantum channels. \begin{figure*}[ht!] \includegraphics[width=15.5cm]{mdi4exp.pdf} \caption{ (Color online) The four initial MDI-QKD experiments. \textbf{a}, Proof-of-principle MDI-QKD with time-bin encoding~\cite{Rub2013}. \textbf{b}, Full MDI-QKD implementation with random modulations of states and decoy intensities based on time-bin encoding~\cite{Liu2013}. \textbf{c}, Proof-of-principle MDI-QKD with polarization encoding~\cite{Silva2013}. \textbf{d}, Full MDI-QKD with random modulations of states and decoy intensities based on polarization encoding~\cite{Tang2014}. [Figures reproduced from~\cite{Rub2013,Liu2013,Silva2013,Tang2014}].} \label{Fig:exp4} \end{figure*} In 2013, several groups performed independent experimental study for MDI-QKD. Liu \emph{et al.}~\cite{Liu2013} reported the first demonstration of MDI-QKD with random modulation for encoding states and decoy states over 50 km fiber. Simultaneously, Rubenok \emph{et al.} were the first to demonstrate the feasibility of high-visibility two-photon interference between two independent lasers, passing through separate field-deployed fibres in a real world environment~\cite{Rub2013}. Later, Ferreira da Silva \emph{et al.} observed a similar interference using polarization encoding in the lab~\cite{Silva2013} and Tang \emph{et al.} reported a full demonstration of polarization encoding MDI-QKD with random modulation of encoding states and decoy states~\cite{Tang2014}. All these four initial experiments, when taken together, complete the cycle needed to demonstrate the feasibility of MDI-QKD using off-the-shelf optoelectronic devices. Their experiment diagrams are illustrated in Fig.~\ref{Fig:exp4}. MDI-QKD is attractive not only because of its security against detection attacks, but also due to its practicality. It can resist high channel loss and reach long distance. Tang \emph{et al.}, implemented MDI-QKD over 200 km fiber~\cite{yanlin2014} and in field environment~\cite{Yanlin2015} by increasing the system clock rate from 1 MHz to 75 MHz, by developing an automatic feedback system, and utilizing superconducting single photon detectors (SNSPDs). In 2016, two millstone MDI-QKD experiments that were subsequently reported. In the first one, Yin \emph{et al.}, extended the MDI-QKD distance to a record-breaking distance of 404 km by optimizing the implementation parameters and using a ultra-low loss fiber (0.16 dB/km)~\cite{Yin2016}. Importantly, the key rate achieved in the experiment at 100 km is around 3 kbps, which is sufficient for one-time-pad encoding of voice message. The results demonstrated in~\cite{Yin2016} are shown in Fig.~\ref{Fig:mdirate}a. In the second one, Comandar \emph{et al.}, increased the system clock rate of MDI-QKD to 1 GHz by exploiting the technique of optical seed lasers~\cite{Comandar}. The 1 GHz system demonstrated the feasibility for MDI-QKD to reach 1 Mbps key rate. The achieved secret rates in~\cite{Comandar} are shown in Fig.~\ref{Fig:mdirate}b. Besides long distance and high rate, several research groups have analyzed the practical aspects in the implementation of MDI-QKD. For instance, Valivarthi \emph{et al.}, analyzed the trade-offs among complexity, cost, and system performance associated with the implementation of MDI-QKD~\cite{valivarthi2015measurement} and implemented a cost-effective system~\cite{valivarthi2017cost}. Wang \emph{et al.}, demonstrated a reference-frame-independent MDI-QKD that requires no phase reference between Alice and Bob~\cite{Wang2015,wang2017measurement} and this scheme was recently improved to a clock rate of 50 MHz by Liu et al.~\cite{liu2018polarization}. Tang \emph{et al.}, demonstrated MDI-QKD with source flaws \cite{tang2016experimental}. Roberts \emph{et al.}, reported a reconfigurable system to switch between QKD and MDI-QKD~\cite{roberts2017experimental}. Aside from the MDI-QKD demonstration with WCP sources, Kaneda \emph{et al.} demonstrated MDI-QKD using heralded single-photon source~\cite{kaneda2017quantum}. Furthermore, a continuous-variable version of MDI-QKD was also proposed and studied~\cite{pirandola2015high}, which will be reviewed in Section~\ref{sec:CVQKD}. With all the above experimental efforts, MDI-QKD is ready for the applications in the future quantum network. Particularly, MDI-QKD is well suited to construct a centric star-type QKD network even with untrusted relays. Indeed, Tang \emph{et al.}, performed the first implementation of a field MDI-QKD network~\cite{Yanlin2016}, which has four nodes with one untrusted relay node and three-user nodes. Note that if the central relay is trusted, one can reconfigure the MDI-QKD network to allow many quantum communication protocols~\cite{roberts2017experimental}. Moreover, high-rate MDI-QKD over asymmetric fiber channels was demonstrated recently in~\cite{liu2018experimental}, based on the theoretical proposal in~\cite{wang2018enabling}. The asymmetric MDI-QKD is valuable to practical metropolitan network settings, where the channel losses are naturally asymmetric and the user nodes could be dynamically added or deleted. Furthermore, Wei et al., have implemented the first chip-based MDI-QKD at 1.25 GHz clock rate~\cite{wei2019high}. This is important to develop a low-cost and secure quantum network, where the expensive devices such as single-photon detectors can be placed in the central untrusted relay and each user requires only a simple Si chip. \subsection{Twin-field QKD} Fundamental bound~\cite{takeoka2014fundamental} and secret key capacity (SKC)~\cite{pirandola2017fundamental} have been obtained for the secure key rate vs distance of QKD. It was proven that, in the absence of relays, the key rate basically scales linearly with transmittance $O(\eta)$, where $\eta$ is the transmittance of the channel between Alice and Bob. This is called the linear bound (of the secret key rate of a lossy quantum channel). There are tremendous research interests towards developing a feasible scheme, known as quantum repeater~\cite{sangouard2011quantum}, to overcome the fundamental rate-distance limit. However, the deployment of quantum repeater is still beyond current technology. Remarkably, Lucamarini {\it et al.}~\cite{Lucamarini2018TF} have proposed a novel phase-encoding MDI-QKD protocol, called twin-field QKD (TF-QKD), which shows the possibility to overcome the SKC. In TF-QKD (see Fig.~\ref{Fig:tfqkd}), weak optical pulses are generated by two phase-locked laser sources, which are phase-randomized and then phase-encoded with secret bits and bases. The pulses are sent to Charlie for interference on a beam splitter; depending on which detector clicks, Charlie can infer whether the secret bits of the users (Alice and Bob) are equal or different, but cannot learn their absolute values. TF-QKD essentially uses single-photon interference~\cite{duan2001}, and the implementation requires only standard optical elements without the requirement of quantum memory~\cite{sangouard2011quantum}. The key goal of TF-QKD protocol is to achieve a quadratic improvement (i.e., scaling to $O(\sqrt{\eta})$ to key rate as a function of channel transmittance. Unfortunately, in the original paper~\cite{Lucamarini2018TF}, such a quadratic improvement was only proven for a restricted class of attacks by Eve. Following the TF-QKD scheme~\cite{Lucamarini2018TF}, Ma \emph{et al.}~\cite{ma2018phase} proposed a protocol named phase-matching QKD (PM-QKD) and proved its unconditional security, inspired by the previous phase-encoding MDI-QKD protocol~\cite{tamaki2012phase}, and the MDI version of the Bennett1992 protocol~\cite{ferenczi2013security}. PM-QKD employs coherent states as information carriers directly and uses the decoy state method in an indirect way. In a sense, PM-QKD adopts a (discrete-modulation) continuous-variable encoding and discrete-variable single-photon detection. The performance of PM-QKD is shown in Figure \ref{Fig:pmqkd}. One can clearly see that its key rate can go beyond the linear SKC with certain realistic parameter settings. On the other hand, with the BB84-type two-basis analysis, Tamaki \emph{et al.}~\cite{tamaki2018information} proposed a modified $X/Y$-basis protocol, and Wang \emph{et al.}~\cite{wang2018sending} proposed an $X/Z$-basis protocol, where the single-photon states are used regarded as the information carrier. Afterwards, the simplified coherent-state-based protocols without phase-randomization on the key generation mode have been proposed~\cite{curty2018simple,cui2018phase,lin2018simple} and analyzed in the infinite data size case. Later on, Maeda \emph{et al.}~\cite{maeda2019repeaterless} introduce an efficient parameter estimation method for the PM-QKD protocol and complete the finite-size analysis. All these recent theory works make the new TF-type MDI-QKD protocols important for the deployment of QKD over long distances. \begin{figure}[!ht] \includegraphics[width=7.5cm]{tfscheme.pdf} \caption{Schematic diagram of TF-QKD~\cite{Lucamarini2018TF}. The light sources (LSs) at Alice’s and Bob’s stations generate pulses that are varied randomly by intensity modulators (IMs) to implement the decoy-state method. Phase modulators (PMs) are combined with random number generators (RNGs) to encode each light pulse with phases, which include bit and basis information as well as the random phases. The variable optical attenuators (VOAs) set the average output intensity of the pulses to bright (classical regime) or dim (quantum regime). The pulses travel along independent channels to then interfere on Charlie’s beam splitter and be detected by the single-photon detectors D0 and D1. Charlie uses the bright pulses in the classical regime and the phase modulator in his station to phase-align the dim pulses that are emitted in the quantum regime, which provide the bits of the key. [Figure adopted from~\cite{Lucamarini2018TF}.]} \label{Fig:tfqkd} \end{figure} \begin{figure}[!ht] \includegraphics[width=8cm]{pmqkd.pdf} \caption{Key rate of PM-QKD comparing to the theoretical SKC~\cite{pirandola2017fundamental} and other protocols~\cite{ma2018phase}. The key rate is shown to surpass the linear key rate bound when the communication distance $l>230$ km. The simulation uses realistic parameters: detector efficiency $14.5\%$, dark count rate $7.2\times 10^{-8}$, error correction efficiency $1.15$, channel misalignment error $1.5\%$ and number of phase slices $16$. [Figure reprinted from~\cite{ma2018phase}.]} \label{Fig:pmqkd} \end{figure} Interestingly, the security of PM-QKD, different from the usual BB84-type two-basis protocol, is closely related to the symmetry of source state with respect to the encoding operation. To clearly establish the correlation between encoding symmetry and privacy, Zeng et al~\cite{Zeng2019PM} establish a symmetry-based security proof method for a general type of MDI-QKD protocols. For these MDI-QKD protocols, there exist symmetric source states, which promise perfect privacy, i.e., with no information leakage. Therefore, for a generic source state input, the privacy of the protocol only depends on the ratio of the symmetric component contained in it, irrelevant of the channel noise. As a result, this symmetry-based security proof allows higher error tolerance compared with the original complementarity-based proof. For example, PM-QKD is proved to be able to yield positive key even with high bit error rate up to $50\%$ and surpassing the linear key rate bound even with bit error rate of $13\%$~\cite{Zeng2019PM}. From a technical point of view, the replacement of two-photon detection to single-photon detection is the key reason for the quadratic improvement, but single-photon interference with two remote independent lasers requires subwavelength-order phase stability for optical channels~\cite{duan2001}, which is more demanding in long-distance communication than achieving two-photon interference which does not require phase stability between the two photons. Nonetheless, TF-QKD protocols are expected to be feasible with the current techniques of active phase randomization, optical phase-locking and so forth. Indeed, in 2019, four research groups have reported the experimental demonstrations on the feasibility of TF-QKD respectively~\cite{LiuTF2019,wang2019beating,ZhongTF2019,minder2019experimental}. Very recently, the PM-QKD experiment has been realized with the random modulations and the consideration of finite-key effect~\cite{Fang2019surpassing}, whose key rate surpass the linear key rate bound via $302$ km and $402$ km commercial fibre; through a $502$ km ultra-low loss fibre with $87.1$ dB total loss, PM-QKD can yield key rate of $0.118$ bps with unconditional security. By using the ultra-stable cavity and optical phase lacking, the sending-or-not-sending version of TF-QKD protocol was demonstrated over $509$ km ultra-low loss fibre~\cite{chen2019sending}, where the achieved secure key rate is even higher than that a traditional QKD protocol running with a perfect repeaterless QKD device. A proof-of-principle experiment has demonstrated TF-QKD over optical channels with asymmetric losses~\cite{zhong2020proof}. Table~\ref{Tab:tfqkd} summarizes the recent TF-QKD experiments. \section{Continuous-Variable QKD}\label{sec:CVQKD} Broadly speaking, QKD can be divided into two classes, namely discrete variable (DV) or continuous variable (CV). Different from DV or qubit-based QKD, the secret keys in CV-QKD are encoded in quadratures of the quantized electromagnetic field and decoded by coherent detections~\cite{weedbrook2012gaussian}. Coherent detection is a promising candidate for practical quantum-cryptographic implementations due to its compatibility with existing telecom equipment and high detection efficiencies without the requirement of cooling. CV-QKD protocols can be divided into several categories, according to the prepared state of coherent~\cite{gg02} or squeezed state~\cite{hillery2000quantum}, the modulation schemes of Gaussian~\cite{cerf2001} or discrete modulations~\cite{ralph1999continuous,hillery2000quantum,reid2000quantum}, the detection schemes of homodyne~\cite{gg02} or heterodyne detections~\cite{weedbrook2004quantum}, the error correction schemes of direct or reverse reconciliations~\cite{grosshans2003quantum}, and so forth. In this section, we will primarily review the simplest and the most widely developed CV-QKD protocol -- Gaussian-modulated coherent state (GMCS) protocol~\cite{gg02,grosshans2003quantum} -- which is believed as the core of today's implementations. We will briefly discuss the security analysis and experimental developments of CV-QKD, together with a focus on the practical security aspects in its implementations, including the side channels and the advanced countermeasures. We will not cover too much about other CV-QKD protocols~\cite{silberhorn2002continuous,pirandola2008continuous,weedbrook2010quantum,usenko2015unidimensional}, which can be found in an earlier review~\cite{weedbrook2012gaussian}. We also refer the readers to two recent CV-QKD reviews for the security analysis and practical issues~\cite{diamanti2015distributing}, the issues of trusted noise~\cite{usenko2016trusted} and the models of implementation and noise~\cite{laudenbach2018continuous}. \subsection{Protocol and security} \subsubsection{Gaussian-modulated protocol} The first Gaussian (continuous) modulated protocol ia a Gaussian-modulated squeezed state protocol~\cite{cerf2001}, where the key is encrypted in the displacement of a squeezed state. The random choice of the direction to squeeze is similar to the basis choice in the BB84 protocol. The squeeze state protocol was later extended to GMCS protocols~\cite{gg02,grosshans2003quantum} since coherent states are easier to prepare in practice. We summarize the prepare-and-measure version of a general GMCS protocol in Box~\ref{tab:gg02}. A difference from a DV-QKD protocol is that, in a coherent state protocol, the key information of both 'basis' is encrypted into the prepared state simultaneously per channel use. Therefore, Bob's measurement can be correlated with Alice's key in either basis or both bases. \begin{tcolorbox}[title = {Box VII.A.1: GMCS QKD protocol.}] (1) Alice produces two random numbers $x_A$ and $p_A$ from random numbers following a Gaussian distribution with a variance of $V_AN_0$, where $N_0$ is the vacuum noise unit. \\ (2) Alice prepares a coherent state $\ket{x_A+ip_A}$ and sends it to Bob through an untrusted quantum channel. \\ (3) Bob chooses homodyne (heterodyne) detection to measure $X$ and $P$ randomly (simultaneously) and obtains the outcomes $x_B$ and $p_B$. \\ (4) After repeating the above process $N$ times, Alice and Bob sift the measurement results by a classical channel, and obtain $N$ pairs of raw key, i.e, the correlated Gaussian variables, in the homodyne detection protocol ($2N$ pairs in heterodyne detection protocol).\\ (5) Alice and Bob perform the postprocessing on the raw key including parameter estimation, error correction and privacy amplification. \end{tcolorbox} \label{tab:gg02} In GMCS protocol~\cite{gg02,grosshans2003quantum}, the source is a mixture of coherent state $\ket{\alpha_j}=\ket{x_j+ip_j}$ with quadrature components $x_j$ and $p_j$ as the realizations for two i.i.d. random variables $X$ and $P$. These two random variables obey the same zero-centred Gaussian distribution $\mathcal{N}(0,V_m)$, where $V_m$ is the modulated variance. The total variance of the Gaussian modulated source is $V=V_s+V_m$, where $V_s$ is the intrinsic quadrature uncertainty of coherent state. Another type of GMCS scheme is coherent-state source mixed with trusted thermal noise~\cite{Weedbrook2010} whose total variance is $V=V_s+V_m+V_{th}$ with an additional thermal variance $V_{th}$. This type of protocol is also widely used due to its low cost in state preparation together with the feasibility for QKD in the wavelength longer than optical band. The decoding process is based on coherent detection measuring quadratures of optical fields. For CV-QKD schemes, coherent detection can be classified into homodyne detection and heterodyne detection, measuring quadratures of optical fields~\cite{weedbrook2012gaussian}. Note that the coherent state protocol with homodyne detection can be modified to a no-switching protocol using heterodyne detection, which enables the communication partners to extract secure key from both quadrature measurements \cite{weedbrook2004quantum}. Post-selection \cite{silberhorn2002continuous} and two-way communication \cite{pirandola2008continuous} can also be applied to improve the performance. So far, the GMCS protocol is believed to be the best understood protocol in terms of security and implementation. Its implementation is also relatively simple, as it requires only standard technology in telecommunication. \subsubsection{Discrete modulated protocol} Besides Gaussian-modulated protocol, there exists a different type of protocol using discrete modulation. Here the key is encoded into the random phases of coherent states, and the source is a $N$-discrete randomized coherent state mixture. In fact, discrete modulated protocol was proposed earlier than Gaussian modulated protocol \cite{PhysRevA.61.010303,PhysRevA.73.012330,PhysRevA.62.062308}. However, due to its non-Gaussian nature, a complete security proof of discrete modulated protocol that gives a good key rate in practice is challenging. In a discrete modulated protocol, Alice prepares an alphabet of $N$ coherent states $\ket{\alpha_k}=\ket{|\alpha|e^{i2k\pi/N}}$, where $k$ is the secret key. Bob uses either homodyne or heterodyne detection to estimate $k$. Discrete modulated protocol is more practical, because (i) a real Gaussian modulation can never be perfectly implemented, and (ii) it can simplify the crucial step of error correction. Early proofs of discrete modulated protocol restrict attacks to be a linear quantum channel between Alice and Bob \cite{PhysRevLett.102.180504}. Though there are proofs for specific protocols where $N=2$ \cite{PhysRevA.79.012307} or $N=3$ \cite{PhysRevA.97.022310}, the key rate is quite pessimistic and cannot be generalized into multiple state cases. Recently, the numerical method of security analysis has been proposed \cite{coles2016numerical}, where the security analysis is transformed into a convex optimization problem with the constraints that the statistics of certain observable should be compatible with experimental data. Following this line, there are two independent works analyzing the asymptotic security of the quadrature-phase-shift-keying (QPSK) protocol \cite{PhysRevX.9.021059,lin2019asymptotic}, i.e., $N=4$. With a photon number cut-off assumption on Bob's side, it is feasible to compute the target function and constraints as a function of Alice and Bob's two-mode state. Such a photon number cut-off assumption is valid since composable security proofs of CV-QKD usually require a projection onto a low-dimensional subspace of the Fock space, via some energy test \cite{PhysRevLett.102.110504}. These proofs can be generalized to multiple state cases, showing that the key rate converges to Gaussian modulated protocols when $N\rightarrow \infty$. Moreover, another security proof was reported lately by applying entropic continuity bounds and approximating a complex Gaussian probability distribution with a finite-size Gauss-Hermite constellation~\cite{kaur2019asymptotic}. Currently, to generalize those existing security proofs to the finite-size case remains an important open question. \subsubsection{Security analysis} Intuitively, the security of coherent state protocol comes from the fact that coherent states are non-orthogonal, which ensures the no-cloning theorem. To rigorously analyze the security, it is convenient to consider an entanglement-based protocol. Alice prepares an two-mode EPR state $\ket{EPR}_{AA^\prime}$. She keeps one mode $A$ in her lab and sends the other mode $A^\prime$ to Bob through a noisy channel $\mathcal{E}_{A^\prime\rightarrow B}$. Alice performs heterodyne detection on her mode and gets a coherent state output, which is identical to preparing a coherent state to Bob from Eve's point of view. We assume the worst case scenario where Eve holds a purification of $\rho_{AB}$, then the tripartite state shared by Alice, Bob and Eve is given by \begin{equation} \rho_{ABE}=(id_A\otimes\mathcal{U}_{A^\prime\rightarrow BE}(\ket{EPR}\bra{EPR}_{AA^\prime})), \end{equation} where $id_A$ denotes the identity map on Alice's mode $A$ and $\mathcal{U}_{A^\prime\rightarrow BE}$ is an isometry. Alice and Bob's secure information under collective attack in asymptotic limit, for reverse reconciliation, is given by the Devetak-Winter formula~\cite{Devetak2005Distillation}, \begin{equation} K=I(A:B)-\sup\chi(B:E), \end{equation} where $\chi(B:E)$ is the Holevo bound~\cite{holevo1973bounds}. The supremum is computed over all possible quantum channels compatible with the statistics obtained in the parameter estimation step in implementation. The secure key can be distilled as long as Alice and Bob's mutual information is larger than the maximum of Bob's classical information accessible to Eve through the quantum channel between Bob and Eve. Specifically, in the parameter estimation step, Alice and Bob exchange the statistics calculated from a subset of the sifted raw key, and estimate the covariance matrix of the two-mode state share by them, \begin{equation} \gamma_{AB}= \left( \begin{matrix} V_A I_2 & Z \sigma_z \\ Z \sigma_z & V_B I_2 \end{matrix} \right) \end{equation} where $V_A$ and $V_B$ are the variance of the quadratures, $I_2$ is the two-dimensional identity matrix, and $Z$ is the covariance calculated by the experimental data. Thanks to the Gaussian optimality proved in \cite{PhysRevLett.97.190502,PhysRevLett.97.190503,PhysRevLett.96.080502}, the optimal collective attack Eve can implement is the one based on Gaussian operations, which result in a two-mode Gaussian state. Owing to the one-to-one correspondence between Gaussian states and covariance matrix, we can directly calculate the secure key rate under collective attack by the covariance matrix. Suppose the optimal attack is characterized by a Gaussian channel of transmittance $T$ and excess noise $\xi$, then there will be the following relations \begin{equation} \begin{aligned} V_B&= T(V_A+\xi) \\ Z&=\sqrt{T(V_A^2-1)}. \end{aligned} \end{equation} And the mutual information between Alice and Bob is given by \begin{equation} I(A:B)=\frac{\omega}{2}\log\frac{V+\xi}{\xi+1}, \end{equation} where $\omega=1,2$ corresponds to Bob's homodyne detection and heterodyne detection, respectively. The Holevo bound is calculated by \begin{equation} \begin{aligned} \chi(B:E)=S(E)-S(E|b)=S(AB)-S(E|b), \end{aligned} \end{equation} where the second equation is because Eve holds a purification of $\rho_{AB}$ and $b$ is Bob's measurement result. Both $S(AB)$ and $S(E|b)$ can be calculated from the corresponding covariance matrix $\gamma_{AB}$ and $\gamma_{E|b}$ \cite{PhysRevLett.94.020505,PhysRevLett.94.020504}. The form of $V(E|b)$ depends on homodyne detection or heterodyne detection that Bob performs. Notice that to obtain a secret key, the two important parameters are the transmittance $T$ and excess noise $\xi$, which should be carefully estimated in the parameter estimation step. The above security analysis is restricted to collective attacks in the asymptotic limit of infinitely long keys. On one hand, one needs to generalize the collective attacks to coherent (or general) attacks, which is a challenging problem in CV-QKD. Fortunately, it turns out that collective attacks are as efficient as coherent attacks assuming the permutation symmetry of the classical postprocessing \cite{PhysRevLett.102.110504}. The phase-space symmetries and the postselection technique \cite{PhysRevLett.102.020504} can also be exploited to perform a reduction from general to collective attacks~\cite{leverrier2013security}. Recently, a new type of Gaussian de Finetti reduction was proposed which confirms the belief that proving the security against Gaussian collective attacks in CV-QKD is sufficient obtain the security against coherent attacks \cite{leverrier2017security}. On the other hand, the security analysis should be extend to \emph{finite-key case}. The finite-key rate will have a deviation from the asymptotic limit, which is due to the statistical fluctuations in parameter estimation. Moreover, other deviations arise when we assume Gaussian attacks and consider collective attacks instead of coherent attacks. These issues were well addressed in the literatures~\cite{furrer2012continuous,leverrier2013security,leverrier2015composable}. Based on postselection technique \cite{PhysRevLett.102.020504}, the security of GMCS CV-QKD was proven against general attacks in the finite-size regime~\cite{leverrier2013security}, but the security proof is not composable. For composable security proof, Furrer et al. provide the first proof for CV-QKD with squeezed states using the entropic uncertainty principle~\cite{furrer2012continuous}, whereas the analysis is only moderately tolerant to loss. For coherent-state protocols, Leverrier gives the first composable security proof against only collective attacks~\cite{leverrier2015composable}, and proposes a new type of Gaussian de Finetti reduction which shows the potential for finite-key security with small data sizes~\cite{leverrier2017security}. Nevertheless, the current proof techniques for composable security against coherent attacks still require rather large block sizes, e.g., $>10^{13}$~\cite{leverrier2017security}. The composable security of CV-QKD against coherent (or general) attacks in a realistic finite-size regime remains an outstanding open issue for the future study of improved proof techniques. Beside coherent state, the squeezed state protocol has also been widely studied for CV-QKD. In a squeezed state protocol~\cite{hillery2000quantum,cerf2001}, Alice squeezes the $X$ quadrature of a vacuum state and displaces it by an amount $a$, which follows a Gaussian distribution of variance $V_A$. Then she adds a random phase of $0$ or $\pi/2$ on it, which is equivalent to randomly choosing a direction to squeeze. Finally she randomly displaces the output state along the other direction (not the squeezing direction) following another Gaussian variable of variance $V$. The two variance $V_A$ and $V$ should satisfy $V_A+V^{-1}=V$ such that Eve cannot distinguish which quadrature is squeezed. Bob randomly measures $X$ or $P$ quadrature. Alice and Bob perform the postprocessing after a certain rounds of measurements. The squeezed state protocol is more similar to the DV-QKD protocols than the coherent state protocol. Its security is based on an entropic uncertainty principle~\cite{cerf2001}. The composable security of finite size analyses have also been given in~\cite{furrer2012continuous,PhysRevA.90.042325}, together with experimental verifications~\cite{gehring2015implementation}. \subsection{Experimental developments} \begin{figure} [htbp] \includegraphics[scale=0.62,angle=0]{cvsetup.pdf} \caption{An illustration of the implementation of GMCS CV-QKD. [Figure reproduced from~\cite{lodewyck2007quantum}].} \label{fig:cvexp} \end{figure} The widely implemented CV-QKD protocol is the GMCS protocol~\cite{gg02,grosshans2003quantum} (see Box~\ref{tab:gg02}) due to its simplicity in preparation, modulation and detection of coherent states. An illustration of the implementation is shown in Fig.~\ref{fig:cvexp}~\cite{lodewyck2007quantum}. Alice employs a laser diode to generate optical pulses, each of which is split into signal and local oscillator (LO) by a fiber-optic coupler. The signal pulses are modulated in amplitude and phase according to a Gaussian distribution, and attenuated to the desired modulation variance with a variable attenuator. The LO is time delayed and then combined with the signal at Alice's output. Bob passively demultiplexes the signal and LO using a coupler and then performs the measurement using a shot-noise-limited homodyne detector. Bob can select the quadrature to be measured by adjusting the measurement phase with a phase modulator placed in the LO path. An advanced feature of this implementation is that it consists entirely of standard fiber optics and telecommunication components. Reverse reconciliation was introduced to the GMCS protocol in 2003~\cite{grosshans2003quantum}, which allows GMCS to beat the 3-dB loss limit. Moreover, a free-space experiment at visible light wavelength was also performed there. With telecom wavelength, GMCS was performed over practical distance of optical fibers (of 25km and 5km respectively) in~\cite{qi2007experimental,lodewyck2007quantum}. Meanwhile, the heterodyne detection~\cite{lance2005no} and Gaussian post-selection~\cite{symul2007experimental} were also demonstrated. Later, the feasibility of GMCS CV-QKD was also extensively tested in field environments~\cite{fossier2009field}. The secure distance was substantially extended to 80 km based on the improved efficiency of the post-processing techniques~\cite{jouguet2013experimental}. By controlling the excess noise, the distance was further extended to 100 km standard fiber~\cite{huang2016long}. Recently, state-of-the-art CV-QKD implementations were sequentially reported, such as high-rate demonstrations with a secret key rate up to 3.14 Mbits/s in the asymptotic limit over 25-km fiber~\cite{wang2018high}, a 4-node field network~\cite{huang2016field}, a field test over 50-km commercial fiber~\cite{zhang2019continuous}, a long-distance CV-QKD over about 200-km ultralow loss fiber~\cite{zhang2020long}, a Si photonic chip-based CV-QKD implementation~\cite{zhang2019integrated} and so forth. Although we focus on the GMCS implementations, we note several other important experiments, such as the squeezed-state protocols~\cite{gehring2015implementation} and a CV-QKD experiment with entangled states over 50-km fiber~\cite{wang2018long}. Some recent developments of CV-QKD are shown in Table~\ref{Tab:cvqkd}. From a practical point of view, CV-QKD presents the key advantage that it only requires standard telecommunication technology which is compatible with classical optical communications, i.e., it uses the coherent detection techniques instead of single-photon detection technology as required in DV-QKD. Moreover, the LO in CV-QKD can serve as a built-in single mode filter, which makes it naturally resistant against background noises~\cite{qi2010feasibility}. This is particularly useful in the practical situations, including the coexistence of QKD with classical channels via DWDM~\cite{qi2010feasibility,kumar2015coexistence}, the daylight free-space CV quantum communication~\cite{heim2014atmospheric,peuntinger2014distribution,vasylyev2017free,zeng2019cv}. Nonetheless, CV-QKD systems are in general sensitive to losses, which restricts the secure distance, normally below 100 km fiber~\cite{jouguet2013experimental}. However, in theory, CV-QKD may provide higher key rates than DV-QKD at relatively short distances because of its high dimensionality~\cite{jouguet2014high}, while the exact rate in terms of bits/s depends on the technology of real implementation. High-rate CV-QKD requires high-speed and real-time implementations of several challenging techniques~\cite{wang2018high}, such as low-noise homodye detector, efficient error correction codes, precise parameter estimation etc. Also, the composable security proofs against general attacks still require a very large block sizes to allow a positive key in the finite-key regime~\cite{leverrier2015composable}, which cascades a challenge on the stability of the system. These issues are important subjects for future research. \subsection{Quantum hacking and countermeasures} \label{cvqkd:hacking} Similar to DV-QKD, the implementations of CV-QKD also suffer from side channels. On the source part, the Trojan horse attacks can probe Alice's modulators in CV-QKD systems~\cite{jain2014trojan}. Similar to DV-QKD, a countermeasure is to put an optical isolator and a monitoring detector at the output of Alice's setup. The imperfections in state preparation may also cause an increase of the excess noise and misestimate of the channel loss~\cite{liu2017imperfect}. On the detection part, the wavelength dependence of the beam splitter can be exploited by Eve to hack CV-QKD based on heterodyne detection~\cite{huang2013quantum,ma2013wavelength}. Qin et al. demonstrated the detector saturation attack~\cite{qin2016quantum} and blinding attack~\cite{qin2018homodyne} against homodyne detectors in CV-QKD by exploiting the nonlinear behavior of coherent detectors. A wavelength filter is effective against the first attack, and a proper monitor at detection may counter the second attack. A more general solution is to perform the real-time shot noise measurement as analyzed in~\cite{kunz2015robust,zhang2019one}. To completely remove the detection attacks, a CV version of MDI-QKD was proposed~\cite{pirandola2015high}. See also~\cite{li2014continuous,ma2014gaussian} for a security analysis against restricted attacks. The concept is similar to the MDI-QKD protocol discussed in Section~\ref{sec:5:mdi}, but here Alice and Bob prepare coherent states with a Gaussian modulation and send them to Charlie. Charlie then mixes them on a balanced beam splitter, measures a different quadrature for both output modes and publicly announces his measurement results. The security of CV MDI-QKD can be analyzed by considering the entanglement-based version of the protocol~\cite{lupo2018continuous}. A proof-of-principle CV MDI-QKD experiment was demonstrated in free space with advanced detection techniques in 2015 \cite{pirandola2015high}. Nonetheless, afull implementation with practical lengths of optical fibers is still a great challenge which has not been reported in the literature yet, partly because of the requirement of high-efficiency detection~\cite{xu2015discrete}. Even so, CV MDI-QKD has the potential to provide slightly higher key rates, and it might be interesting for network communication over relatively short distances~\cite{pirandola2015high}. Besides, an additional threat for CV-QKD is the transmission of LO, which can be manipulated by Eve. The attacks by controlling the transmitted LO were proposed in~\cite{ma2013local}. Eve can also exploit a subtle link between the local oscillator calibration procedure and the clock generation procedure employed in practical setups~\cite{jouguet2013preventing}. A countermeasure for the LO attacks consists of implementing a rigorous and robust real-time measurement of the shot noise~\cite{kunz2015robust,zhang2019one}. A better solution is the locally LO (LLO) CV-QKD scheme~\cite{qi2015generating,soh2015self} (see Fig.~\ref{fig:llocv}), which can completely avoid the transmission of the LO through the insecure channel. In this scheme, Bob uses a second, independent laser to produce LO pulses locally for the coherent detection. A challenge here is how to effectively establish a reliable phase reference between Alice's and Bob's independent lasers, which require a careful synchronization of the frequencies and phases. This can be achieved by sending the reference or pilot-aided pulse along with the signal pulse from Alice. Bob can use his LO pulse to perform coherent detection for the reference pulse to estimate the relative phase between Alice's and Bob's lasers. A phase correction can thus be established on Alice and Bob's signal data in order to generate the secret key. Due to the enhanced security of LLO CV-QKD, there has been great attention in this scheme. In 2015, three groups have independently demonstrated LLO CV-QKD~\cite{qi2015generating,soh2015self,huang2015high}. These experiments are shown in Fig.~\ref{fig:llocv}. Afterwards, a LLO CV-QKD experiment with pilot and quantum signals multiplexed in the frequency domain was reported in~\cite{kleis2017continuous}, a comprehensive framework to model the performance of LLO CV-QKD was reported in~\cite{marie2017self}, a pilot-assisted coherent intradyne reception methodology for LLO CV-QKD was proposed and demonstrated in~\cite{laudenbach2017pilot}, and a high rate LLO CV-QKD was demonstrated in~\cite{wang2018high}. \begin{figure} [htbp] \includegraphics[scale=0.28,angle=0]{llocv.pdf} \caption{(Color online) Initial local LO (LLO) CV-QKD experiments. \textbf{a}, LLO CV-QKD experiment with the pilot-aided feedforward data recovery scheme using commercial off-the-shelf devices~\cite{qi2015generating}. \textbf{b}, LLO CV-QKD experiment with self reference~\cite{soh2015self}. \textbf{c}, A high-speed LLO CV-QKD experiment~\cite{huang2015high}. [Figures reproduced from~\cite{qi2015generating,soh2015self,huang2015high}].} \label{fig:llocv} \end{figure} \section{Other Quantum Cryptographic Protocols}\label{sec:6} \subsection{Device-independent QKD} \label{sec:diQKD} A QKD protocol is device-independent if its security does not rely on trusting that the quantum devices used are truthful. A schematic illustration of device-independent QKD (DI-QKD) is shown in Fig.~\ref{Fig:diqkd}. DI-QKD~\cite{mayers1998quantum,barrett2005no,acin2007device} (hinted earlier by Ekert~\cite{ekert1991quantum}) relaxes all modelling assumptions about the quantum devices and allows the users to do QKD with uncharaterized devices. DI-QKD performs self-testing of the underlying devices, i.e., the devices cannot pass the test unless they carry out the QKD protocol securely. As a result, as long as certain necessary assumptions are satisfied, one can prove the security of DI-QKD based solely on a Bell nonlocal behaviour, typically the violation of a Bell's inequality, which certifies the presence of quantum correlations in a self-testing manner. Table~\ref{Tab:assumptions} lists a summary of the necessary assumptions of DI-QKD~\cite{pironio2009device}. The security proofs have required the assumption that the devices have no memory between trials, or that each party has many, strictly isolated devices~\cite{barrett2005no,acin2007device,pironio2009device,masanes2011secure}. If the devices have memory or the devices are reused, DI-QKD will suffer from the memory attacks~\cite{barrett2013memory} and covert channels~\cite{curty2017quantum}. \begin{figure}[hbt] \begin{center} \includegraphics[scale=0.64,angle=0]{diqkd.pdf} \end{center} \caption{Schematic diagram of DI-QKD~\cite{mayers1998quantum,acin2007device}. Entangled photon pairs are distributed to Alice and Bob, who are supposed to perform some measurements. Alice and Bob see their quantum devices as black boxes producing classical outputs, as a function of classical inputs $X_{A}$ and $X_{B}$. From the observed statistics, and without making any assumptions on the internal working of the devices, they should be able to conclude whether they establish a secret key. Alice and Bob assume giving the untrusted quantum devices tests that cannot be passed unless they carry out the QKD protocol securely, which can be check via the violation of a Bell's inequality~\cite{pironio2009device}.}\label{Fig:diqkd} \end{figure} The security proof for DI-QKD is a challenging task, because in DI-QKD, both the quantum state (generated by the source) and the measurement operators (by the detection devices) are untrusted or under Eve's control. Fortunately, recent theory efforts have significantly advanced the developments of DI-QKD to be possible in a large quantum system~\cite{reichardt2013classical}, secure for a large class of protocols by independent measurements~\cite{masanes2011secure}, secure against general attacks~\cite{vazirani2014fully,miller2016robust} and robust against noise~\cite{Arnon2018Practical}. In the asymptotic case against collective attacks, the key rate formula can be expressed as a function of the Bell violation value. For a protocol where Alice and Bob carry out a CHSH-type Bell test for self-testing privacy, the key rate formula can be given by~\cite{pironio2009device}, \begin{equation} r \geq 1-h(E)-h\left(\dfrac{1+\sqrt{(S/2)^2-1}}{2}\right), \end{equation} where the quantum bit error rate $E$ determines the amount of randomness consumed for error correction, and the violation value $S$ of the CHSH inequality determines the amount of randomness for privacy amplification. Though DI-QKD is remarkable in theory, unfortunately, it is hard to realize with current technology, because it needs almost perfect efficiency of single-photon detection~\cite{masanes2011secure}. In experiments, however, the emitted photons may not be detected due to the losses in the transmission or the limited detection efficiency of imperfect detectors. In addition, a faithful realization of DI-QKD requires that the Bell inequality is violated under the following two conditions: (i) The measurement settings are not correlated with the devices; (ii) The devices observe a no-signalling behaviour in generating the outcomes. To meet these conditions, a so-called loophole-free Bell test normally needs to be carried out. A key problem in a loophole-free Bell test is the limited detection efficiency which is referred to as efficiency loophole~\cite{pearle1970hidden}. It has been proved that for the simplest bipartite Bell inequalities with binary inputs and outputs, a detector efficiency of at least 2/3 is necessary for a faithful Bell inequality violation~\cite{eberhard1993background}. For the purpose of DI-QKD, a much higher efficiency is needed due to the requirement of information reconciliation. To guarantee the no-signalling behaviour between devices, i.e., the locality loophole, a space-like separated measurement set-up can be implemented~\cite{aspect1975proposed}. The requirement of uncorrelated inputs is referred to as the free-will loophole, which cannot be closed completely. Nonetheless, practical QRNGs can be used to overcome the problem to some extent. Recently, an exciting news is that researchers have demonstrated the Bell's inequality which simultaneously closed the locality loophole and the detection loophole in the same experiment~\cite{hensen2015loophole,shalm2015strong,giustina2015significant,rosenfeld2017event,liu2018device}. This is a milestone result towards the realization of DI-QKD. In future, the advanced technologies might make DI-QKD more practical, and ideas such as qubit amplification~\cite{gisin2010proposal} might also be proved useful to increase the key rate and distance of DI-QKD, though the key rate might be relatively low~\cite{curty2011heralded,seshadreesan2016progress}. Recent theoretical progresses have proposed the two-way classical communication to enhance the noise tolerance~\cite{tan2020advantage}, and provided the detailed analysis towards the realization of DI-QKD~\cite{murta2019towards}. Overall, we do believe that DI-QKD is an important subject for future research. \subsection{Some New QKD implementations} Besides the efforts in the security with imperfect devices, quite a few new QKD protocols have been proposed and implemented during the past ten years, which are summarized in Table~\ref{Tab:otherqkd}. \subsubsection{Round-robin DPS QKD} In general, there exists a threshold of the error rate for each scheme, above which no secure key can be generated. This threshold puts a restriction on the environment noises. Especially, in the key rate formula, the bit error can be directly computed from the experimental data, whereas the phase error needs to be estimated or bounded. In the BB84 protocol with strong symmetries, both error rate are approximately the same in the long key length limit. In other protocols, normally there is a relation between the two error rates. In the end, when the bit error rate goes beyond some threshold level, no secure key can be generated. For example, BB84 cannot tolerate error rates beyond 25\% considering a simple intercept-and-resend attack \cite{bennett1984quantum}. This threshold puts a stringent requirement on the system environment, which makes some practical implementations challenging. Round-robin differential phase shifted (RRDPS) QKD, proposed by Sasaki, Yamamoto and Koashi~\cite{sasaki2014practical} in 2014, essentially removes this restriction and can in principle tolerate more environment disturbance. In this protocol, Eve's information can be bounded \emph{only} by user's certain experiment parameters other than the error rates. In particular, the phase error rate $e_{p}$ is determined by the user's own settings rather than the channel performance, which makes the protocol fundamentally interesting and tolerate more errors. In the RRDPS QKD protocol, the sender Alice puts a random phase, chosen from $\{0, \pi\}$, on each of $L$ pulses, with an average photon number of $\mu$ in such an $L$-pulse signal. Upon receiving the block, the receiver Bob implements a single-photon interference with an Mach-Zehnder interferometer (MZI), as shown in Fig.~\ref{rrdps scheme}a. Bob can randomly adjust the length difference of the two arms of the MZI. After obtaining a detection click, Bob first identifies which two pulses interfere and then announces the corresponding indices $i$, $j$ to Alice. Alice can derive the relative phase between the two pulses as the raw key, and Bob can record the raw key from the measurement results. The phase error rate depends only on the number of photons in the $L$-pulse signal and $L$, not the bit error rate. By setting a larger $L$, the phase error tends to 0, and the scheme can tolerate a higher bit error rate. Triggered by the original protocol, an alternative passive type of RRDPS QKD is proposed by Guan et al. \cite{guan2015experimental}. As is shown in Fig.~\ref{rrdps scheme}b, when Bob receives a block from Alice, he prepares a local $L$-pulse reference in plain phases, i.e., all phases are encoded at phase 0. This $L$-pulse reference interferes with the $L$-pulse signal sent by Alice on a beam splitter. For each block, Bob records the status of his two detectors with time stamps, $i$ and $j$. If Bob's reference is in phase with Alice's signal, i.e. Bob has a phase reference, the whole setup is essentially a huge Mach-Zehnder interferometer. Any detection signal at time slot $i$ will tell the phase difference between $i$ and the phase reference. Then the encoding bit value can be figured out by Bob. If the phase Bob's reference is random relative to Alice's signal, the interference is a Hong-Ou-Mandel type interference \cite{HOM:1987} instead of a MZI. Bob post-selects the block where there are exactly two detections and announces their positions i and j (if i = j, the detection result is discarded). The raw key is the relative phase between these two pulses in the L-pulse signal. Alice can derive this phase difference from her record. While Bob can infer the bit value depending on that the coincidence happens between two different detectors or one detector at two different time slots. The security proofs of the two protocols are beyond the scope of the paper and we refer the readers to the original papers~\cite{sasaki2014practical,guan2015experimental}. \begin{figure}[!ht] \centering \resizebox{8cm}{!}{\includegraphics{rrdpsscheme.pdf}}\\ \caption{\textbf{a}, The original RRDPS scheme \cite{sasaki2014practical}. VDL means variable delay line. Bob splits the received signals into two paths and applies a variable delay $r$ to one of the paths. A click at $i$th place will indicate an interference between the pulses $i$ and $j=i+r$. \textbf{b}, The passive RRDPS scheme. Bob uses a local laser to generate an $L$-pulse reference, which interferes with Alice's $L$-pulse signal. Bob then records the coincidence clicks. [Figure reproduced from~\cite{guan2015experimental}].} \label{rrdps scheme} \end{figure} The first published experimental result is based on the passive protocol \cite{guan2015experimental}, as shown in Fig.~\ref{rrdps exp}a. Comparing to the original protocol, the passive one avoids randomly adjusting the length difference of the MZI. Based on the current technology, the main adjust-delay method is to utilize optical switches, which cannot provide both high speed and low insertion loss simultaneously. But meanwhile, it requires remote optical phase locking, which is challenging in the real deployment. \begin{figure} \centering \resizebox{8.2cm}{!}{\includegraphics{rrdpsexp.pdf}}\\ \caption{(Color online) The experimental setups for RRDPS QKD, reported in the references of \textbf{a},~\cite{guan2015experimental}; \textbf{b}, \cite{takesue2015experimental}; \textbf{c}, \cite{wang2015experimental}; \textbf{d}, \cite{li2016experimental}. [Figures reproduced from~\cite{guan2015experimental,wang2015experimental,li2016experimental}].} \label{rrdps exp} \end{figure} The key point for an active RRDPS is to realize the random time delay. Takesue et al. \cite{takesue2015experimental} exploits a one-input, four-output optical splitter followed by four silica waveguides based MZI with 0.5,1.0,1.5,2.0 ns temporal delays respectively, as is shown in Fig.~\ref{rrdps exp}b. Any two delays constitute a new MZI and the whole system realizes a $L=1-5$ variable delay. With this delay, the authors achieved secure key rate through 30 km fiber with an error rate of $18\%$. Later, Wang et al. \cite{wang2015experimental}, shown in Fig.~\ref{rrdps exp}c, combines a three-port circulator, a beam splitter, two $1\times8$ optical switches followed with two groups of fiber delays. The two optical switches shall actively choose different delays and achieve a $L$=1 to 64 bit variable-delay Faraday-Michelson interferometer. Based on the delay, Wang et al. distributed a secret key over a distance of 90 km fiber. In addition, Li et al. \cite{li2016experimental} exploited a different configuration, which can be seen in Fig.~\ref{rrdps exp}d. They put 7 MZI in series to achieve a 127-value variable delays. And each MZI is constructed of a Pockels cell, a fiber, or free-space link with specific length and two polarizing beam splitters (PBSs). The Pockels cell, controlled by a random number, may change the polarization of the photon and thus provide a delay. Very recently, the secure distance is extended to 140 km by increasing the bound on information leakage \cite{yin2018improved}. \subsubsection{High-dimensional QKD} Besides the qubit-based QKD, the secret keys can also be encoded with multi-level system, i.e. high-dimensional QKD (HD-QKD). HD-QKD can provide higher key rate per particle comparing to the qubit system \cite{Bourennane2001multilevel}, and it has a higher tolerance to noise \cite{Cerf2002}. A recent review on the subject can be seen in~\cite{xavier2020quantum}. The first experimental attempts of HD-QKD are using higher-order dimensional alphabets with spatial degrees of freedom of photons~\cite{walborn2006quantum} or energy-time entangled photon pairs \cite{Ali-Khan2007}. The later is shown in Fig.~\ref{HD QKD}a. With this setup, Ali-Khan et al. can generate a large-alphabet key with over 10 bits of information per photon pair, albeit with large noise. QKD with 5\% bit error rate is demonstrated with 4 bits of information per photon pair, where the security of the quantum channel is determined by the visibility of Franson interference fringes. \begin{figure} \centering \resizebox{8.2cm}{!}{\includegraphics{hdqkd.pdf}}\\ \caption{(Color online) The experimental setup for HD QKD. \textbf{a},HD QKD with time-energy entangled photon pairs \cite{Ali-Khan2007}. \textbf{b}, Dispersive-optics time-energy HD QKD \cite{lee2014entanglement}. \textbf{c}, A field test of OAM HD QKD in the city of Ottawa \cite{sit2017high}. [Figures reproduced from~\cite{Ali-Khan2007,lee2014entanglement,sit2017high}].} \label{HD QKD} \end{figure} Zhang et al. \cite{zhang2014unconditional} reported a complete security proof of time-energy entanglement QKD using dual-basis interferometry. Mower et al. \cite{mower2013high} suggested to utilize dispersive optics to replace the Franson interferometer and demonstrated its security against collective attack. In this scheme, as shown in~Fig.~\ref{HD QKD}b~\cite{lee2014entanglement}, Alice or Bob utilize normal or abnormal group-velocity dispersive element to measure the frequency basis. The absolute group delays of their dispersive elements are matched such that the group-velocity dispersion is nonlocally canceled. Alice and Bob use time basis measurements for generating keys and frequency basis measurements for bounding Eve's maximum accessible information about the time basis measurements. This is based on the fact that the dispersion cancelation only happens with entanglement and any reduced entanglement visibility due to eavesdropping will bring a broaden time correlation measurement. With the time energy entangled photon pairs, Zhong et al. observed a secure key rate of $2.7 Mbps$ after 20 km fiber transmission with a key capacity of $6.9$ bits per photon coincidence \cite{zhong2015photon}. Recently, high-rate QKD using time-bin qudits was reported in~\cite{islam2017provably}. Time energy type HD-QKD has advantage with a constant clock rate because it can utilize more time slots with high time resolution single photon detector. However, the advantage will be offset when the clock rate can be increased to the bandwidth of the single photon detector~\cite{Zhang:08}. One solution is to utilize a degree of freedom other than time, for example, the optical angular momentum (OAM). The first HD-QKD for OAM was published in 2006 \cite{Simon2006}. Qutrit entangled photon pairs were utilized to generate quantum key. In an E91-type protocol\cite{ekert1991quantum}, the violation of a three-dimensional Bell inequality verifies the security of the generated keys. A key is obtained with a qutrit error rate of approximately 10\%. Later, Etcheverry et al. report an automated prepare-and-measure HD-QKD with 16-dimensional photonic states~\cite{etcheverry2013quantum}; Mafu et al. exploit high-dimension OAM up to five dimensions for HD-QKD \cite{Mafu2013}; Mirhosseini et al. use the OAM of weak coherent state and the corresponding mutually unbiased basis of angular position~\cite{mirhosseini2015high}; Sit et al. \cite{sit2017high} implement a field test of OAM HD-QKD in the city of Ottawa, where 4-dimensional OAM HD-QKD was implemented and a QBER of 11\% was attained with a corresponding secret key rate of 0.65 bits per sifted photon (see Fig.~\ref{HD QKD}c). Recently, Cozzolino et al., demonstrated OAM HD-QKD over a 1.2-km-long multi-mode fiber~\cite{cozzolino2019orbital}. Different groups utilized spatial-division multiplexing optical fibres, such as multi-core fibres, to perform HD-QKD~\cite{canas2017high,ding2017high}. Naively, one might think that since a HD-QKD system offers a higher key rate per signal than a qubit-based QKD system. It seems always better to use a HD-QKD system. One has to be very careful in making such a comparison, because key rate per signal may not be the best measure when the signal size itself is big. Key rate per second (certain period of time) can be a better merit for applications. In fact, a HD-QKD protocol uses many e.g. time-bins/modes for each signal. Now, if one were to use the many time-bins/modes separately and in parallel (with many sets of high-speed single photon detectors), one would actually get a higher key rate in such a multiplexed QKD system. At the end of the day, the private capacity per \emph{mode} of a simple prepare-and-measure QKD system is limited by fundamental bounds~\cite{takeoka2014fundamental,pirandola2017fundamental}. The key rate of HD-QKD is still limited by those fundamental bounds. Nonetheless, HD-QKD may be useful in a practical situation, where the single photon detector has long dead time or resetting time and it can not operate at high speed~\cite{zhong2015photon}. Overall, the practical advantages of HD-QKD in real-life applications remain to be seen in future. \subsubsection{QKD with wavelength-division multiplexing} Except for the new protocols, reducing the cost of QKD system is another important topic in the field. Wavelength-division multiplexing (WDM) technology, which enables the coexistence of QKD and telecom communication in a single fiber, is exploited to reduce the cost of the channel. In order to protect ultra-weak QKD signals, most of previous QKD experiments are implemented in dark fibers. This implies dedicated fiber installations for QKD networks, which bears cost penalties in fiber leasing and maintenance, as well as limitations on the network scale. In classical optical communications, WDM technology has been widely exploited to increase the data bandwidth and reduce the requirement of fiber resource. Then, it is natural for QKD to coexist with classical optical communication based on WDM technology. The scheme of simultaneously transmitting QKD with conventional data was first introduced by Townsend in 1997 \cite{Townsend1997}. A series of QKD experiments integrating with various classical channels have been demonstrated \cite{Chapuran2009,Eraerds2010,Patel2012,patel2014quantum,Wang2015exp,dynes2016ultra}. Currently, by using spectral and temporal controls, state-of-the-art developments have been made to realize co-propagation of QKD with one 100 Gbps dense wavelength-division multiplexing (DWDM) data channel in 150 km ultra-low loss fiber at $-5 dBm$ launch power \cite{frohlich2017long}. By setting QKD wavelength to 1310 nm and inserting $100 GHz$ DWDM filters, Wang et al. implement QKD together with classical traffic with $11 dBm$ input power over 80 km fiber spools \cite{WANG2017}. A field trial of simultaneous QKD transmission and four 10 Gbps encrypted data channels was implemented over 26 km installed fiber at $-10 dBm$ launch power \cite{Choi:14}. Recently, the coexistence of QKD and commercial backbone network of $3.6 Tbps$ classical data over 66 km fiber at 21 dBm launch power has been demonstrated \cite{Mao:18}. The system provides 3 kbps secure key rate with a 2.5\% quantum bit error rate. Note that in current backbone networks, the data traffic is around Tbps and the launch power is around 20 dBm. In that sense, the recent work \cite{Mao:18} demonstrate the possibility of coexistence of QKD with backbone network. \subsubsection{Chip-based QKD} \label{sec:chipQKD} \begin{figure*}[!ht] \centering \resizebox{13cm}{!}{\includegraphics{inteqkd.pdf}}\\ \caption{(Color online) The experimental layout for \textbf{a}, Low cost and compact QKD setup~\cite{duligall2006low}, \textbf{b}, Silicon photonic QKD emitter~\cite{ma2016silicon}, \textbf{c}, A compact QKD transmitter, QCard~\cite{hughes2013network}, \textbf{d}, InP based QKD sender and SiOxNy receiver chip~\cite{sibson2017chip}. [Figures reproduced from~\cite{duligall2006low,ma2016silicon,hughes2013network,sibson2017chip}].} \label{Inte QKD} \end{figure*} Integrating QKD system has attracted more and more attention due to its advantage at compact size, low energy consumption and potential for low cost~\cite{orieux2016recent}. QKD, including optics and electronics, is a complicated system. Thus an integrated QKD system research should include integrating both optics and electronics. Fortunately, the integrated circuits (IC) is already commercialized and integrated optics is also well developed in industry. Table~\ref{Tab:chipqkd} summarizes a list of chip-based QKD experiments. In 2005, a commercial unbalanced Mach-Zahnder interferometer made of planar lightwave circuits (PLC) based on silica-on-silicon technology was exploited for the first time in a QKD system \cite{takesue2005differential} to replace the fiber based interferometer. Comparing to its fiber counterpart, PLC interferometer is more stable and can maintain its phase for several hours without any feedback~\cite{takesue2007quantum,nambu2008quantum}. Meanwhile, IC is exploited in a research towards compact and low cost QKD system \cite{duligall2006low}. As is shown in Fig.~\ref{Inte QKD}a. Alice module uses off-the-shelf IC components in a driver circuit to control four AlInGaP LEDs to emit four polarized BB84 states. The channel is a several-meter free space link, which is supposed to find application in a future quantum based Automated Teller Machine and even in a smart phone \cite{pizzi2012affordable} according to the authors. Along this direction of research, Gwenaelle Vest \emph{et al.} demonstrated an integrated QKD sender \cite{vest2015design}, where an array of four vertical cavity surface emitting lasers (VCSELs) emit synchronized picosecond optical pulses, which are coupled to micro-polarizers generating the polarization qubits. The final size of the QKD device can be as small as $25 mm \times 2 mm \times 1 mm$, which makes the system a strong candidate for short distance free-space QKD applications. On the metropolitan fiber network side, many individual users in the network will trust a central relay station. This is so called network-centric structure \cite{hughes2013network} or access network \cite{Bernd2013}. In such a structure, many users are all the senders and share only one central relay receiver. In that sense, the receiving station can have more space and expensive and bulky detection system can be used. Therefore, the community concentrates more integration efforts in the sending side. Hughes et al. provided a QCard~\cite{hughes2013network} in their pioneering paper, as is shown in Fig.~\ref{Inte QKD}b. The QCard has a similar size as an electro-optic modulator or a normal key. It incorporates a distributed feedback laser and modulator. The laser is attenuated into single photon level and modulated into BB84 polarization-state with decoy state. The repetition frequency is 10 MHz at the wavelength of 1550 nm, the telecom band. Recently, the size of the QKD sender has reduced dramatically. In 2014, Zhang \emph{et al.} put forward an on-chip LiNbO$_{3}$ polarization rotator and demonstrated the reference-frame-independent QKD protocol to overcome unstable fibre birefringence \cite{zhang2014reference}. In 2015,, the same group from University of Bristol implemented integration of QKD based on an indium phosphide transmitter chip and a silicon oxynitride receiver chip \cite{sibson2017chip}. This chip is shown in Fig.~\ref{Inte QKD}d. The authors exploited the chips in three different QKD protocols, namely BB84, coherent-one-way and differential-phase-shift QKD. Later, researchers from University of Toronto \cite{ma2016silicon} and Bristol \cite{sibson2017integrated} exploited Silicon photonics to build QKD sender system, respectively. As shown in Fig.~\ref{Inte QKD}c, Ma et al.~\cite{ma2016silicon} fabricated the QKD sender chip with a standard Si photonic foundry process and integrated two ring modulators, a variable optical attenuator and a polarization modulator in a $1.3 mm \times 3 mm$ die area. Meanwhile, Sibson \emph{et al.} \cite{sibson2017integrated} demonstrated coherent one-way QKD, polarization encoded BB84, and time-bin encoded BB84 based on Si photonic devices. The authors achieve estimated asymptotic secret key rates of up to 916 kbps and QBER as low as 1.01\% over 20 km of fiber. The clock rate of later experiment is much higher than the former one. However, Ma et al. integrated more components on the chip, i.e., the whole QKD emitter. Very recently, other research groups have demonstrated high-speed Si photonic chips for high dimensional QKD over multimode fiber~\cite{ding2017high}, transceiver circuit~\cite{cai2017silicon} and metropolitan QKD~\cite{bunandar2018metropolitan}. Moreover, CV-QKD is naturally suitable for photonic chip integration as its implementation is compatible with current telecom technologies (see Section~\ref{sec:CVQKD}). In particular, CV-QKD essentially uses the same devices as classical coherent communication, and only homodyne detector is required rather than the dedicated single-photon detector. Indeed, a recent experiment demonstrates Si photonic chips for CV-QKD, which integrates all the optical components (except the laser source)~\cite{zhang2019integrated}. Furthermore, based on the directly phase modulated light source~\cite{yuan2016directly,roberts2018direct}, a modulator-free QKD transmitter chip was demonstrated in~\cite{paraiso2019modulator}. This approach has the advantages that do not require conventional phase modulators, and it is versatile to accommodate several QKD protocols, such as BB84, COW and DPS, using the same optics. \subsection{Other quantum cryptographic protocols} \label{sec:8} So far, QKD is the most developed and mature subfield of quantum cryptography. Meanwhile, quantum cryptography has many other protocols~\cite{broadbent2016quantum}, which also have achieved quite remarkable progresses. A list of recent developments of other quantum cryptographic protocols is shown in Table~\ref{Tab5}. We will review a few examples. \subsubsection{Quantum bit commitment} \label{sec:qbc} Bit commitment is another important and fundamental cryptographic task that guarantees a secure commitment between two mutually mistrustful parties. Alice first commits her to a particular bit value $b$. After a period of time, Alice reveal the bit value to Bob. A success bit commitment requires that Bob can not learn $b$ before Alice reveals it, which is called concealing criterion. Meanwhile, Alice should not change $b$ once she made the commitment. This is called binding criterion. Bit commitment is a building block for many cryptographic primitives, including coin tossing~\cite{brassard1991quantum}, zero-knowledge proofs \cite{goldwasser1989knowledge,goldreich1986proofs}, oblivious transfer \cite{bennett1992practical,unruh2010universally} and secure two-party computation \cite{kilian1988founding}. In conventional cryptography, bit commitment is based on computational complexity assumptions similar to public key exchange protocols and might be vulnerable to quantum attacks. Unfortunately, it has been proven that information-theoretically secure bit commitment is impossible even if Alice and Bob are allowed to use quantum resources in the standard quantum circuit model by Mayers~\cite{mayers1996proceedings,mayers1997unconditionally} and by Lo and Chau~\cite{lo1997is}. Subsequently, such a no-go theorem has been further extended to case with superselection rules~\cite{kitaev2004superselection}. For an re-examination of this result, see e.g.~\cite{D'Ariano2007reexamination}. Furthermore, information-theoretic security of oblivious transfer and two-party secure computations are also proven to be impossible in~\cite{lo1997insecurity}. Interestingly, if we take into account the signalling constraints implied by the Minkowski causality in a relativistic context, it has been shown that there are bit commitment protocols offering unconditional security \cite{kent2012unconditionally,kaniewski2013secure}. On the experimental side, two groups implemented the secure relativistic quantum bit commitment simultaneously in 2013. One followed the original protocol and utilized decoy state method in free-space channel \cite{liu2014experimental} and the other exploited a revised protocol with a plug and play system, in fiber link \cite{lunghi2013experimental}. Both experiments were secure against any quantum or classical attack. The commitment time is defined as the maximal time during which the commitment can be held. The commitment time in these two experiments, however is limited to 21 ms if all attendees are located on Earth considering the relativistic constrains. Later, new protocols with weaker security but longer commitment time was proposed \cite{lunghi2015practical,chakraborty2015arbitrarily}. A 24-hour committed experiment \cite{verbanis2016hour} was presented, which is secure against only classical attacks. Alternatively, secure quantum bit commitment can be achieved with some additional physical assumptions as, for example, that the attacker’s quantum memory is noisy \cite{ng2012experimental}. \subsubsection{Quantum digital signature} \label{sec:QDS} Comparing to the previous two-party protocols, digital signature has one sender and multiple recipients, requiring that the messages cannot be forged or tampered with. Classical digital signature mainly exploits the Rivest-Shamir-Adleman protocol \cite{rivest1978method}, the security of which is based on the mathematical complexity of the integer factorization problem. Based on the quantum physics, quantum digital signature (QDS) protocol was provided~\cite{gottesman2001quantum}, which could provided information-theoretical security \cite{martin2001introduction}. Although novel, this protocol needs nondestructive state comparison, long-time quantum memory, and a secure quantum channel for real application. Thereafter, QDS has attracted a great deal of interest in both theory \cite{andersson2006experimentally,dunjko2014quantum,wallden2015quantum} and experiment \cite{clarke2012experimental,collins2014realization,donaldson2016experimental}. All the three requirements have been fixed sequentially \cite{clarke2012experimental,collins2014realization,donaldson2016experimental}. Later, more than 100 km QDS experiment has been demonstrated based on decoyed BB84 system \cite{yin2017experimental} and DPS QKD \cite{collins2016experimental}, which are also secure against PNS attack. Very recently, measurement-device-independent (MDI) QDS have been implemented in both lab \cite{roberts2017experimental} and field \cite{Yin2017}. \subsubsection{Other protocols} QKD has been assuming that the eavesdropper has unlimited power as long as it is not violated quantum physics. A protocol is said to be information-theoretically secure if it allows an adversary (e.g. an eavesdropper) to have unlimited quantum computing power as long as it does not violate quantum mechanics. As noted in Section~\ref{sec:qbc} above, information-theoretic security is not possible for quantum bit commitment, quantum oblivious transfer and two-party secure quantum computation. Naturally, restriction on adversary's power can expand the territory of quantum cryptography. Wehner et al. \cite{wehner2008cryptography} proposed one realistic assumption that quantum storage of qubits is noisy and demonstrated that an \emph{oblivious transfer} protocol is unconditionally secure for any amount of quantum-storage noise \cite{damgaard2008cryptography,konig2012unconditional}. Similar as bit commitment, oblivious transfer protocol is another primitive cryptograph protocol between two entrusted parties. The demonstration of the protocol was performed based on a modified entangled QKD system \cite{erven2014experimental}. The experiment exchanged a 1,366 bit random oblivious transfer string in 3 minutes and include a full security analysis under the noisy storage model, accounting for all experimental error rates and finite size effects. Similar to bit commitment and oblivious transfer, a quantum protocol for \emph{coin flipping} \cite{blum1981advances} can be unconditionally secure when considering relativistic constrains. This also means that without relativistic designs, no bias coin flipping could not be unconditionally secure~\cite{lO1998why}. Nevertheless, a quantum protocol can limit the cheating probability strictly lower than $1/\sqrt{2}$ \cite{aharonov2000quantum,kitaev1999quantum}. The first experimental demonstration was provided with OAM qutrit entangled photon pairs, which shows the quantum advantage in coin flipping for the first time \cite{molina2005experimental}. As a proof of principle demonstration, this experiment does not consider the channel loss. Theoretical and experimental efforts have been attempted towards this direction \cite{nguyen2008experimental,berlin2011experimental,pappa2014experimental}. For instance, an implementation of the loss-tolerant protocol using an entangled-photon source was provided \cite{berlin2011experimental}. The secure distance was extended to 15 km with a modified plug and play system \cite{pappa2014experimental}. Quantum \emph{data locking} \cite{divincenzo2004locking} allows one to lock information in quantum states with an exponentially shorter key, presenting an efficient solution to many resource-limited secure applications. However, the original quantum data-locking scheme may suffer from significant qubit loss. In 2013, Fawzi, Hayden, and Sen (FHS) developed a loss-tolerant quantum data-locking scheme \cite{Fawzi2013low}, in which the possible information leakage can be made arbitrarily small in a lossy environment while the unlocked information is significantly larger than the key size. This feature makes the protocol attractive also in secure communication \cite{seth2013quantum,lupo2014robust}. Two groups have implemented the loss tolerant protocols respectively \cite{liu2016experimental,lum2016quantum}. Quantum \emph{secret sharing} was proposed to share a secret quantum state among multiple parties~\cite{cleve1999share} or to use quantum states to share classical secrets~\cite{hillery1999quantum,cleve1999share}. Moreover, secure multi-party computing has been extended to quantum computation with quantum inputs and circuits~\cite{crepeau2002secure}. In \emph{distributed quantum computing}, quantum crypto protocols are still inevitable. Quantum computing is currently attracting tremendous interest from both academic and industry~\cite{mohseni2017commercialize}. However, due to its implementation complexity and cost, the future path of quantum computation is strongly believed to delegate computational tasks to powerful quantum servers on cloud~\cite{fitzsimons2017private}. Universal blind quantum computing (UBQC) \cite{broadbent2009universal} is an effective method for a common user, who has limited or no quantum computational power, to delegate computation to an \emph{untrusted} quantum server, without leaking any information about the user's input and computational task. The security or blindness of the UBQC protocol is unconditional, i.e., the server cannot learn anything about user's computation except its size. A proof of concept demonstration was reported in 2012~\cite{barz2012demonstration}. Recently, UBQC protocol with completely classical clients was proposed \cite{reichardt2013classical} and demonstrated in experiment~\cite{huang2017experimental}. UBQC with weak coherent states was proposed in~\cite{dunjko2012blind}, and adding the ingredient of decoy states, an efficient experimental demonstration with weak coherent states was made in~\cite{jiang2019remote}. Because of the developments in the field of quantum computing, we expect that BQC will play an important role in the future infrastructure of delegated quantum computation~\cite{fitzsimons2017private}. \section{Concluding Remarks}\label{sec:9} In this review, we have discussed the security aspects of practical QKD. These range from the security proofs of practical QKD (Section~\ref{Sc:Security}), the implementation (Section~\ref{Sc:Implement}), the practical vulnerabilities (Section~\ref{sec:3}), to the solutions of advanced QKD protocols (Sections~\ref{sec:4},~\ref{sec:5} and~\ref{sec:CVQKD}) and the advances of other quantum cryptographic protocols (Section~\ref{sec:8}). Historically, QKD has been a concrete playground for concepts in quantum mechanics. The study of QKD often leads to unexpected insights in other areas of quantum information. For instance, apparently the concept of quantum teleportation was invented during a search for a security proof of QKD~\cite{bennett1993teleporting}. We expect that in future the study of QKD will continue to lead to many new insights in other subfields of quantum information. Meanwhile, as a new technology stemming from the counterintuitive theory of quantum physics, QKD might not be easy to be understood and recognized by a general audience. For broad interest, in Appendix~\ref{app:general}, we summarize a few frequently asked questions/concerns on practical QKD, together with our views on how they can be overcome. Finally, we discuss the perspectives on the past, the present and the future about the developments of QKD. Overall, during the past three decades, the theory and practice of QKD have developed extensively. These developments can be divided into several stages, which can be summarized as follows (with a focus on DV-QKD). \begin{enumerate} \item \emph{Stage 1.} After the invention by Bennett-Brassard~\cite{bennett1984quantum} and Ekert~\cite{ekert1991quantum}, QKD was first demonstrated in the early 1990s~\cite{bennett1992experimental}, which started a series of theories and experiments~\cite{Townsend1993,doubleMZI:1994,Franson1995,muller1996quantum} to prove the possibility of QKD. \item \emph{Stage 2.} The implementation of QKD was extended from laboratory to outdoor environments, and various technical difficulties were studied~\cite{Townsend1997,buttler1998practical,Hughes2000quantum,ribordy2000fast,gobby2004quantum}. See ref.~\cite{gisin2002quantum} for a review on the developments in early experiments. Meanwhile, on the theory side, the security proof of QKD was a major challenge until a few papers appeared and solved the problem~\cite{mayers2001unconditional,lo1999unconditional,shor2000simple,biham2000proof}. These results put the security of QKD on a solid foundation. \item \emph{Stage 3.} With the security proofs for QKD under imperfect devices~\cite{gottesman2004security,hwang2003quantum,lo2005decoy,wang2005beating}, the feasibility of QKD was demonstrated from short range to long range, up to the scale of 100-km standard fiber~\cite{Yi2006,rosenberg2007long,Peng2007} and free space~\cite{Tobias2007}. \item \emph{Stage 4.} QKD was extensively deployed from point-to-point to small-scale metropolitan networks in field~\cite{elliott2005current,chen2009field,peev2009secoqc,sasaki2011field}. Meanwhile, the practical security loopholes, particularly those for detection devices, were identified~\cite{Lars2010} and then removed by the advanced MDI-QKD protocol~\cite{lo2012measurement} [see also~\cite{braunstein2012side}]. \item \emph{Stage 5.} The feasibility of QKD was extended to long distances and high rates, such as in a scale of 400 km~\cite{Yin2016,boaron2018secure} over ultra low-loss fiber and 1200-km over free space~\cite{Liaosate}, and a secret key rate of over 10 Mbits/s with GHz QKD system~\cite{Yuan:18}. \item \emph{Stage 6.} QKD was implemented from small scale to large scale that covers a wide area~\cite{Yuao2019}. See Fig.~\ref{Fig:backbone} for an example of the QKD network which has more than 700 QKD links, and covers more than 2,000~km area. New TF-QKD protocols~\cite{Lucamarini2018TF} were proposed to enable secure QKD over even longer distances~\cite{Fang2019surpassing,chen2019sending}. \end{enumerate} In future, towards the ultimate goal of a global QKD network, we expect that more and more QKD networks will be built in different countries. Besides physics, communities of computer science, engineering, optics, mathematics and so forth may work together to realize this goal. We do believe that a revolutionized global QKD network for secure communication stemming from quantum physics will be deployed and find widespread applications in the near future. This review is concluded with a discussion on a few directions for future research. \begin{enumerate} \item \emph{Quantum repeaters.} Quantum repeater can achieve an effective restoration of the quantum information without resorting to a direct measurement of the quantum state~\cite{BDCZ1998,duan2001}, enabling the realization of global quantum network in existing optical networks. Quantum repeater has received intense research efforts in recent years~\cite{kimble2008quantum,sangouard2011quantum,pan2012multiphoton,wehner2018quantum}. Nonetheless, the limited performance of quantum memory is still a major obstacle in realizing practical quantum repeaters without a future experimental breakthrough \cite{sangouard2011quantum,Yang2016nphoton}. New recent approaches manage to reduce the need for a quantum memory by using all-photonic quantum repeaters~\cite{azuma2015all,hasegawa2019experimental,li2019experimental}, but they require the resources of large-scale cluster states. Overall, we believe that quantum repeater is an important subject for future research. The first goal is to realize a practical quantum repeater that can beat the fundamental limits of direct quantum communication~\cite{takeoka2014fundamental,pirandola2017fundamental}. \item \emph{Standardization.} Towards the widespread applications, the commercial standards for QKD should be established. Important progress has been made in this direction, such as the efforts of ETSI, ISO/IEC, CCSA and ITU by several countries. One important direction is to include the practical security into the standardization process, by defining the best practices to operate QKD systems and standardizing those countermeasures to guarantee the security of a QKD setup. We encourage future research to establish the commercial standards for QKD. \item \emph{Battle-testing security.} We have provided a review on the practical vulnerabilities in Section~\ref{sec:3}, together with the solutions of advanced countermeasures and QKD protocols. However, the practical security issue has not been perfectly solved. For instance, as discussed in Section~\ref{sec:5:mdi}, a security assumption in MDI-QKD is that the source should be trusted without loopholes. It is important to verify this assumption in practice. Hence, the research in analyzing the practical security of QKD setup should continue. This includes the developments of practically-secure QKD systems building on the experience gained from the research on practical vulnerabilities and advanced countermeasures. It is highly important to battle-test existing QKD implementations, quantify and validate the security claims of real-world QKD systems, and design real-life QKD systems with testable security assumptions. \item \emph{Small-size, low-cost, long-distance system.} Recent developments of integrated QKD system have been reviewed in Section~\ref{sec:chipQKD}. These developments should continue to further reduce the costs and sizes of QKD, and to realize robust fully-integrated chip-based QKD systems. One important direction is to develop the star-type quantum access network~\cite{hughes2013network,Bernd2013}, in which the expensive devices such as single-photon detectors can be placed in the central relay and many users share this relay. Each user requires only a low-cost transmitter such as a compact QCard \cite{hughes2013network} or a simple Si chip~\cite{ma2016silicon,sibson2017chip}. Together with MDI-QKD, the central relay can be \emph{untrusted}. Wei et al., already implemented the first chip-based MDI-QKD at high secret key rates~\cite{wei2019high}. This is particularly valuable for star-type metropolitan QKD networks. Moreover, by using the new type of twin-field QKD~\cite{Lucamarini2018TF}, the distance can be further extended for intercity QKD. Therefore, we expect that MDI-type QKD networks will play an important role in the future global quantum network. \item \emph{QKD network with untrusted relays.} The previously deployed networks were based on trusted relays~\cite{elliott2005current,chen2009field,peev2009secoqc,sasaki2011field,Yuao2019}, which may raise the concern about the security properties of the relays. To remove this concern, it is important to develop QKD network with untrusted relays. In fact, MDI-QKD is naturally suitable for a star-type metropolitan network with an untrusted relay. Tang et al., already put forward the first implementation of a MDI-QKD network~\cite{Yanlin2016}. We expect that metropolitan MDI-QKD networks will be built soon. Besides, the TF-QKD can also be adopted to extend the transmission distance with an untrusted relay. Moreover, in entanglement-based QKD, the relay can be fully untrusted. A possible direction is to develop a entanglement-based QKD network, e.g., based on satellite~\cite{Yin2019}. For ultralong-distance QKD in fiber, it needs the quantum repeaters~\cite{sangouard2011quantum} to realize QKD networks with untrusted relays. We expect that with the technical improvements, quantum-repeater assisted QKD network may be achieved in the near future. \item \emph{Satellite-based QKD.} The reported satellite-based QKD was based on a low-earth-orbit (LEO) satellite of Micius~\cite{Liaosate,liao2018satellite}. To increase the coverage time and area for a more efficient satellite-based QKD network, one can launch higher-orbit quantum satellites and implement QKD in daytime. Important progress has been made in this direction~\cite{hughes2002practical,liao2017long}. An ultimate goal is to realize a satellite-constellation-based global quantum network. \end{enumerate} \section*{Acknowledgement} We thank enlightening discussions with and helpful comments from the reviewers and the numerous colleagues including, K. Azuma, D. Bacco, C. H. Bennett, Y.-A. Chen, A. Huang, D.~Huang, M. Lucamarini, Y.~Liu, S. Pirandola, B. Qi, K. Tamaki, K.~Wei, G. B. Xavier, Z. Yuan P.~Zeng and H.~Zhou. F.~Xu, X.~Ma, Q.~Zhang and J.-W.~Pan were supported by the National Key Research and Development (R\&D) Plan of China, the National Natural Science Foundation of China, the Anhui Initiative in Quantum Information Technologies and the Chinese Academy of Sciences. H.-K. Lo was supported by NSERC, US Office of Naval Research, CFI, ORF, MITACS Accelerate, Royal Bank of Canada (RBC) and Huawei Canada.

\section*{Introduction} For $d$-metal compounds with localized magnetic moments, basic guidelines to soothsay the sign of the nearest-neighbor (NN) magnetic exchange interactions, i.~e., the Anderson-Goodenough-Kanamori rules~\cite{Goodenough1958,Kanamori1959,anderson1959}, were laid down back in the 1950's. With one single bridging anion and half-filled $d$ states these rules safely predict antiferromagnetic (AF) exchange interactions, as is indeed encountered in numerous magnetic Mott insulators. It is however much harder to anticipate the sign of the couplings for geometries with two bridging ligands and bond angles close to $90^{\circ}$. Illustrative recent examples are the $5d$ honeycomb systems Na$_2$IrO$_3$ and Li$_2$IrO$_3$. The signs of the NN Heisenberg $J$ and of the additional symmetric Kitaev anisotropy $K$ are intensely debated in these iridates, with both $J\!<\!0$, $K\!>\!0$ \cite{Ir213_KH_chalopka_12,Ir213_andrade_14,Ir213_kee_14} and $J\!>\!0$, $K\!<\!0$ \cite{Ir213_jkj2j3_kimchi_2011,Ir213_jkj2j3_singh_2012,Ir213_choi_2012,Ir213_KH_mazin_2013,Ir213_rachel_14,Ir213_trebst_14} sets of parameters being used to explain the available experimental data. The sizable Kitaev interactions, that is, uniaxial symmetric terms $KS_i^{z}S_{i+\mathbf{r}}^{z}$ ($\mathbf{r}\!=\!\frac{\mathbf{x}+\mathbf{y}}{\sqrt{2}}$) that cyclically permute on the bonds of a particular hexagonal ring \cite{Kit_kitaev_06,Ir213_KH_jackeli_09,Ir213_KH_chaloupka_10}, are associated to strong frustration effects and unconventional magnetic ground states displaying, for example, noncollinear order, incommensurability, or spin-liquid behavior \cite{Ir213_KH_chalopka_12,Ir213_andrade_14,Ir213_jkj2j3_kimchi_2011,Ir213_jkj2j3_singh_2012,Ir213_KH_chaloupka_10,Na2IrO3_vmk_14,Li2IrO3_vmk_14,Ir213_trebst_14,Ir213_kee_14,Ir213_rachel_14}. Obviously, in the context of electronic-structure computational methods, such features cannot be thoroughly addressed by periodic total-energy calculations for a given set of predetermined spin configurations. A much more effective strategy is to first focus on individual pairs of NN $d$-metal sites and obtain reliable values for the associated effective magnetic couplings by using {\it ab initio} many-body quantum chemistry (QC) machinery (for a recent review, see Ref.~\citeonline{qc_review_j_2014}). The computed exchange parameters can be subsequently fed to effective spin Hamiltonians to be solved for larger sets of magnetically active lattice sites. Such an approach, earlier, allowed us to establish the signs plus the relative strengths of the Heisenberg and Kitaev interactions in both Na$_2$IrO$_3$ and Li$_2$IrO$_3$ and to additionally rationalize the qualitatively different types of AF orders in these two $5d^5$ honeycomb iridates \cite{Na2IrO3_vmk_14,Li2IrO3_vmk_14}. The related $4d^5$ honeycomb compound Li$_2$RhO$_3$ is even more puzzling because it features no sign of long-range magnetic order. Instead, an experimental study suggests the presence of a spin-glass ground state \cite{Li2RhO3_HK_luo_13}. While the spin-orbit couplings (SOC's) are still sizable for the $4d$ shell and may in principle give rise on the honeycomb lattice to compelling Kitaev physics, to date no conclusive evidence is in this respect available for Li$_2$RhO$_3$. To shed light on the nature of the essential exchange interactions in Li$_2$RhO$_3$ we here carry out detailed {\it ab initio} QC calculations. We show that large trigonal splittings within the Rh $t_{2g}$ shell, comparable with the strength of the SOC, dismiss a simple picture based on $j_{\rm eff}\!=\!1/2$ and $j_{\rm eff}\!=\!3/2$ effective states \cite{SOC_d5_thornley68,book_abragam_bleaney,IrO_mott_kim_08,Ir213_KH_jackeli_09}. The magnetic properties of the system can still be described, however, in terms of ${\tilde {S}}\!=\!1/2$ pseudospins. The calculations earmark Li$_2$RhO$_3$ as a 4$d$-electron system with remarkably large anisotropic magnetic couplings, in particular, FM Kitaev interactions of up to 10--15 meV. The isotropic Heisenberg exchange, on the other hand, features opposite signs on the two sets of structurally distinct links of Rh NN's. This sign modulation of the NN Heisenberg interactions, with strong ferromagnetic (FM) $J$'s for one type of Rh-Rh bonds and weaker AF couplings for the other pairs of adjacent Rh sites, enables the initial ${\tilde {S}}\!=\!1/2$ hexagonal network to be mapped onto an effective model of spin-1 dimers on a frustrated triangular lattice. We further address the issue of Rh-Li anti-site disorder in samples of Li$_2$RhO$_3$. By exact-diagonalization (ED) calculations for an extended spin model that also includes second and third neighbor couplings, we show that the experimentally observed spin-glass behavior can be rationalized as a superposition of different nearly-degenerate symmetry-broken states arising at finite concentration of in-plane spin vacancies. \section*{Results} \subsection*{Rh$^{4+}$ $4d^5$ electronic structure} The tetravalent rhodium ions in Li$_2$RhO$_3$ display a 4$d^5$ valence electron configuration, octahedral ligand coordination and bonding through two bridging ligands. In the simplest picture, i.e., for sufficiently large Rh $t_{2g}$--$e_g$ splittings and degenerate $t_{2g}$ levels, the ground-state electron configuration at each site is a $t_{2g}^5$ effective $j_{\rm eff}\!=\!1/2$ spin-orbit doublet \cite{SOC_d5_thornley68,book_abragam_bleaney,IrO_mott_kim_08,Ir213_KH_jackeli_09}. For $5d^5$ ions in a variety of three-dimensional, layered, and chain-like oxides, {\it ab initio} QC electronic-structure calculations yield excitation energies of 0.6--0.9 eV for the transitions between the $j\!\approx\!1/2$ and split $j\!\approx\!3/2$ levels \cite{Ir213_rixs_gretarsson_2013,Ir3116_rixs_liu_2012,Ir227_hozoi_14,RhIr_vmk_IC_14} and indicate values of 0.45--0.5 eV for the strength of the SOC $\lambda$, in agreement with earlier estimates \cite{SOC_d5_andlauer76}. Sharp features in the range of 0.6--0.9 eV are indeed found in the resonant x-ray scattering spectra \cite{Ir214_rixs_jkim_12,Ir227_hozoi_14,Ir213_rixs_gretarsson_2013,Ir3116_rixs_liu_2012}. The validity of the $j_{\rm eff}\!=\!1/2$ approximation for the ground state of Li$_2$RhO$_3$ is however questionable since the SOC is substantially weaker for $4d$ elements. Indeed our QC calculations (see Table~\ref{Rh213_dd}) indicate Rh $t_{2g}$ splittings $\delta\!\approx\!0.11$ eV, close to values of 0.14--0.16 eV estimated for $\lambda$ in various Rh$^{4+}$ oxides \cite{Sr2RhO4_haverkort_2008,RhIr_vmk_IC_14}. For the {\it ab initio} QC investigation we employed multiconfiguration complete-active-space self-consistent-field (CASSCF) and multireference configuration-interaction (MRCI) calculations \cite{book_QC_00}, see Supplemental Material (SM) and Refs.~\citeonline{RhIr_vmk_IC_14,Ir213_rixs_gretarsson_2013}. With $\delta$ and $\lambda$ parameters of similar size, the $j_{\rm eff}\!=\!1/2$ and $j_{\rm eff}\!=\!3/2$ states are strongly admixed, as discussed in earlier work \cite{SOC_d5_thornley68,SOC_d5_hill71} and illustrated in Table~\ref{Rh213_dd}. In particular, for the relativistic ground-state wave function the $t_{2g}$ hole is not equally distributed among the three Rh $t_{2g}$ levels as for the ``true'' $j_{\rm eff}\!=\!1/2$ ground state \cite{SOC_d5_thornley68,book_abragam_bleaney} but displays predominant $d_{xy}$ character, close to 60\%. \subsection*{Magnetic couplings between two adjacent Rh$^{4+}$ ions} Interestingly, while the results for the on-site $4d^5$ excitations are quite different as compared to the $5d^5$ excitation energies \cite{Ir213_rixs_gretarsson_2013}, the computed intersite effective interactions are qualitatively similar to those obtained for the $5d^5$ honeycomb iridate Li$_2$IrO$_3$ \cite{Li2IrO3_vmk_14}. The intersite exchange couplings were estimated by MRCI+SOC calculations for embedded fragments having two edge-sharing IrO$_6$ octahedra in the active region. As described in earlier work \cite{Na2IrO3_vmk_14,Li2IrO3_vmk_14,Ba214_vmk_14} and in SM, the {\it ab initio} QC data for the lowest four spin-orbit states describing the magnetic spectrum of two NN octahedra is mapped in our scheme onto an effective spin Hamiltonian including both isotropic Heisenberg exchange and symmetric anisotropy. Yet the spin-orbit calculations, CASSCF or MRCI, incorporate all nine triplet and nine singlet states that arise from the two-site $t_{2g}^5$--$t_{2g}^5$ configuration. For on-site Kramers-doublet configurations, the most general symmetry-allowed form of the effective spin Hamiltonian, for a pair of NN ions, is \begin{equation} {\mathcal H}_{ij}^{{C_{2h}}} = J\,{\tilde {\bf S}}_i\cdot {\tilde {\bf S}}_j + K\,{\tilde S}_i^z {\tilde S}_j^z + \displaystyle\sum\limits_{\alpha<\beta} \Gamma_{\alpha\beta}({\tilde S}_i^{\alpha}{\tilde S}_j^{\beta} + {\tilde S}_i^{\beta}{\tilde S}_j^{\alpha})\,, \label{eqn:Eqn1} \end{equation} where ${\tilde {\bf S}}_i$, ${\tilde {\bf S}}_j$ are 1/2 pseudospin operators, $J$ is the isotropic Heisenberg interaction, $K$ the Kitaev coupling, and the $\Gamma_{\alpha\beta}$ coeeficients are offdiagonal elements of the symmetric anisotropic exchange matrix with $\alpha,\beta\!\in\!\{x,y,z\}$. The antisymmetric anisotropic term vanishes since the crystallographic data reported in Ref.~\citeonline{Li2RhO3_structure_2011} indicate overall $C_{2h}$ point-group symmetry for one block of NN RhO$_6$ ocathedra, green (B1) bonds in Fig.\,\ref{213_structure}, and only slight deviations from $C_{2h}$ for the other type of NN's, blue (B2 and B3) bonds in Fig.\,\ref{213_structure}. For $C_{2h}$ symmetry of the Rh-Rh link, $\Gamma_{zx}\!=\!-\Gamma_{yz}$. We note that in (\ref{eqn:Eqn1}) $\alpha$ and $\beta$ stand for components in the {\it local}, Kitaev bond reference frame. The $\vec{z}$ axis is here perpendicular to the Rh$_2$O$_2$ plaquette (see SI and Refs.~\citeonline{Ir213_KH_jackeli_09,Na2IrO3_vmk_14,Li2IrO3_vmk_14}). Relative energies for the four low-lying states describing the magnetic spectrum of two NN octahedra and the resulting effective coupling constants are listed in Table~\ref{Rh213_mgn_coupl}. For the effective picture of ${\tilde S}\!=\!1/2$ pseudospins assumed in Eq.\,(\ref{eqn:Eqn1}), the set of four eigenfunctions contains the singlet $\Phi^{ij}_{\mathrm{S}}\!=\!(\uparrow\downarrow\!-\!\downarrow\uparrow)/\sqrt2$ and the triplet components $\Phi^{ij}_{\mathrm{1}}\!=\!(\uparrow\downarrow\!+\!\downarrow\uparrow)/\sqrt2$, $\Phi^{ij}_{\mathrm{2}}\!=\!(\uparrow\uparrow\!+\!\downarrow\downarrow)/\sqrt2$, $\Phi^{ij}_{\mathrm{3}}\!=\!(\uparrow\uparrow\!-\!\downarrow\downarrow)/\sqrt2$. In $C_{2h}$ symmetry, the ``full" spin-orbit wave functions associated to $\Phi^{ij}_{\mathrm{S}}$, $\Phi^{ij}_1$, $\Phi^{ij}_2$ and $\Phi^{ij}_3$ transform according to the $A_g$, $B_{u}$, $B_{u}$ and $A_u$ irreducible representations, respectively. Since two of the triplet terms may interact, the most compact way to express the eigenstates of (\ref{eqn:Eqn1}) is then $\Psi^{ij}_1\!=\! \Phi^{ij}_1\cos\alpha\!+\!i \Phi^{ij}_2\sin\alpha$, $\Psi^{ij}_2\!=\!i\Phi^{ij}_1\sin\alpha\!+\! \Phi^{ij}_2\cos\alpha$, $\Psi^{ij}_3\!=\! \Phi^{ij}_3$ and $\Psi^{ij}_{\mathrm{S}}\!=\!\Phi^{ij}_{\mathrm{S}}$. The angle $\alpha$ parametrizes the amount of $\Phi^{ij}_1$--$\Phi^{ij}_2$ mixing, related to finite off-diagonal $\Gamma_{\!\alpha\beta}$ couplings. This degree of admixture is determined by analysis of the full QC spin-orbit wave functions. The effective parameters provided in Table\,\ref{Rh213_mgn_coupl} are obtained for each type of Rh-Rh link by using the $E_1$, $E_2$, $E_3$, $E_{\mathrm S}$ MRCI relative energies and the $\Phi^{ij}_1$--$\Phi^{ij}_2$ mixing coefficients (see SM). For B1 links, we find that both $J$ and $K$ are FM. While by MRCI calculations $K$ always comes FM in spin-orbit coupled honeycomb systems \cite{Na2IrO3_vmk_14,Li2IrO3_vmk_14}, the FM $J$ for the B1 bonds has much to do with the peculiar kind of dependence on the amount of trigonal squashing of the oxygen octahedra and consequently on the variation of the Rh-O-Rh angles of the Rh$_2$O$_2$ plaquette. The latter increase to values larger than 90$^{\circ}$ for finite trigonal compression. This dependence of the NN $J$ on the Rh-O-Rh bond angles is illustrated in Fig.\,2 for a simplified structural model of Li$_2$RhO$_3$ where the Rh-O bond lengths are all the same, set to the average bond length in the experimental crystal structure \cite{Li2RhO3_structure_2011}. It is seen that $J$ displays a parabolic behavior, with a minimum of about --5 meV in the interval 92--93$^{\circ}$ and a change of sign to AF couplings around 96$^{\circ}$. For the B1 Rh-Rh links, the Rh-O-Rh bond angle is 93.4$^{\circ}$, close to the value that defines the minimum in Fig.\,2. The difference between the $\approx$--5 meV minimum of Fig.\,2 and the $\approx$--10 meV result listed in Table\,\ref{Rh213_mgn_coupl} comes from additional distortions of the O octahedra in the actual structure (see the footnotes in Table\,\ref{Rh213_mgn_coupl} and Ref.\,\citeonline{Li2RhO3_structure_2011}), not included in the idealized model considered for the plot in Fig.\,2. An even stronger FM $J$ was computed for the B1 type bonds in the related compound Li$_2$IrO$_3$ \cite{Li2IrO3_vmk_14}. In Na$_2$IrO$_3$, on the other hand, the Ir-O-Ir bond angles are $>$97$^{\circ}$ and the NN $J$ turns AF on all short Ir-Ir links \cite{Na2IrO3_vmk_14}. For the B2 and B3 links, we derive a FM Kitaev term and an AF Heisenberg interaction, again qualitatively similar to the QC data for Li$_2$IrO$_3$ \cite{Li2IrO3_vmk_14}. We assign the AF value of the NN $J$ on the B2/B3 units to the slightly larger Rh-O-Rh bond angle, which as shown in Fig.\,2 pulls the $J$ towards a positive value, and most importantly to additional distortions that shift the bridging ligands on the Rh-O$_2$-Rh B2/B3 plaquettes in opposite senses parallel to the Rh-Rh axis \cite{Li2RhO3_structure_2011}. The role of these additional distortions on the B2/B3 units was analyzed in detail in Ref.\,\citeonline{Na2IrO3_vmk_14} and shown to enhance as well the AF component to the intersite exchange. \subsection*{Effect of longer-range exchange interactions and occurence of spin-glass ground state} For further insights into the magnetic properties of Li$_2$RhO$_3$, we carried out ED calculations for an extended spin Hamiltonian that in addition to the NN terms of Eq.(1) incorporates longer-range second- and third-neighbor Heisenberg interactions $J_2$ and $J_3$ \cite{Ir213_choi_2012,Ir213_jkj2j3_singh_2012,Ir213_jkj2j3_kimchi_2011,Ir213_andrade_14}. We used clusters of 24 sites with periodic boundary conditions \cite{Ir213_KH_chalopka_12,Na2IrO3_vmk_14,Li2IrO3_vmk_14} and the quantum chemically derived NN coupling constants listed in Table\,\ref{Rh213_mgn_coupl}. The static spin-structure factor $S({\bf Q})=\sum_{ij} \langle {\tilde {\bf S}}_i\cdot {\tilde {\bf S}}_j \rangle \exp[i {\bf Q}\cdot ({\bf r}_i-{\bf r}_j)]$ was calculated as function of variable $J_2$ and $J_3$ parameters. For a given set of $J_2$ and $J_3$ values, the dominant order is determined according to the wave number {\bf Q} = ${\bf Q}_{max}$ providing a maximum value of $S({\bf Q})$. The resulting phase diagram is shown in Fig.\,3(a). Given the similar structure of the NN magnetic interactions, it is somewhat similar to that obtained in our previous study on Li$_2$IrO$_3$ \cite{Li2IrO3_vmk_14}. Six different regions can be identified for $|J_2|,|J_3|\!\lesssim\!6$ meV: FM, N\'eel, Kitaev spin liquid, stripy, diagonal zigzag and incommensurate ${\bf Q}$ (ICx) phases. The Kitaev spin liquid, stripy and incommensurate phases in strongly spin-orbit coupled honeycomb $5d^5$ systems were analyzed in a number of earlier studies \cite{Ir213_KH_chalopka_12,Ir213_jkj2j3_singh_2012,Ir213_jkj2j3_kimchi_2011,Ir213_kee_14,Na2IrO3_vmk_14}. The detailed nature of the diagonal zigzag and incommensurate ICx ground states for large FM $J$ on one set of NN links was described in Ref.\,\citeonline{Li2IrO3_vmk_14}. Remarkably, for $J({\mathrm{B1}})$ much larger than $K({\mathrm{B1}})$ and $J({\mathrm{B2}})$, the initial hexagonal $\tilde S\!=\!1/2$ lattice can be mapped onto an effective triangular model of {\it triplet} dimers on the B1 bonds \cite{Li2IrO3_vmk_14}. Since $J_2$ and $J_3$ are expected to be AF in honeycomb $d^5$ oxides \cite{Ir213_choi_2012,Ir213_jkj2j3_singh_2012}, the most likely candidate for the magnetic ground state of ``clean'' crystals of Li$_2$RhO$_3$, according to our results, is the diagonal zigzag state (see Fig.\,3) and is found to be stable in a wide region of $J_2\!\gtrsim\!0$ and $J_3\!\gtrsim\!0$. Experimentally, however, a spin-glass ground state was determined, with a spin freezing temperature of $\sim\!6$\,K \cite{Li2RhO3_HK_luo_13}. As possible cause of the observed spin-glass behavior in Li$_2$RhO$_3$ we here investigate the role of Li-Rh site intercalation. Significant disorder on the cation sublattice is a well known feature in Li$_2${\it M}O$_3$ compounds. A typical value for the degree of Li$^+$--{\it M}$^{4+}$ site inversion in these materials is 10--15$\%$ \cite{kobayashi_95,Ir213_ryoji_2003}. Partial substitution of the ``in-plane'' Rh$^{4+}$ ions by nonmagnetic Li$^+$ species introduces spin deffects in the $\tilde S\!=\!1/2$ honeycomb layer. On the 24-site cluster employed for our ED calculations, 10--15$\%$ site inversion translates in replacing two $\tilde S\!=\!1/2$ centers by vacancies. Hereafter, we denote the two spin defects as $p_1$ and $p_2$. The effect of spin vacancies on the static spin-structure factor in the diagonal zigzag phase ($J_2\!=\!J_3\!=\!3$) is shown in Fig.\,3(d),(e). For comparison, the static spin-structure factor is also plotted in Fig.\,3(c) for the ideal case without spin defects. In the absence of ``defects'', the ground state is characterized in the bulk limit by symmetry-broken long-range order with either ${\bf Q}_1\!=\!(\pm\pi,\pm\frac{\pi}{\sqrt{3}})$ or ${\bf Q}_2\!=\!(\pm\pi,\mp\frac{\pi}{\sqrt{3}})$ wave vectors. Since the two symmetry-breaking states are degenerate, the structure factor displays four peaks, at ${\bf Q}\!=\!{\bf Q}_1$ and ${\bf Q}\!=\!{\bf Q}_2$ [see Fig.\,3(c)]. However, by introducing spin vacancies, the degeneracy may be lifted via impurity pinning effects. For example, when the two defects occupy positions 17 and 20 [$(p_1,p_2)\!=\!(17,20)$, see Fig.\,3(b)] the spin structure defines one of the symmetry-breaking states with ${\bf Q}\!=\!{\bf Q}_1$ [Fig.\,3(d)]; likewise, defects at $(p_1,p_2)\!=\!(17,18)$ yield a state with ${\bf Q}\!=\!{\bf Q}_2$ [Fig.\,3(e)]. In other words, two different kinds of dominant short-range order can be obtained with anti-site disorder. The ``locally'' favored symmetry-breaking direction depends on the relative positions of the spin vacancies. In a macroscopic system, such ``local'' domains displaying different symmetry-breaking ordering directions are randomly distributed. Additional {\it frustration} is expected to arise because it is not possible to match two differently ordered domains without an emerging ``string''. It is therefore likely that by creating some amount of spin defects the long-range zigzag order disappears and the resulting state is perceived as a spin glass at low temperature. A similar mechanism was proposed for the isotropic Heisenberg-Kitaev and $J_1$-$J_2$-$J_3$ models \cite{Ir213_andrade_14}. An early well known example of frustration induced through the competition between two different, degenerate spin configurations is the two-dimensional Ising model on a square lattice with randomly distributed, competing FM and AF bonds~\cite{Vannimenus_77}. To investigate how the diagonal zigzag state is destroyed by increasing the concentration of spin defects, we also studied a simplified Ising model with $J\!=\!-\infty$ for the B1 bonds, $J_2\!=\!J_3$ and all other interactions set to zero. This is a reasonable approximation for the honeycomb layers of Li$_2$RhO$_3$ since the diagonal zigzag phase essentially consists of alternating spin-up and spin-down chains [see sketch in Fig.\,3(a)]. Spin structures obtained this way for various spin-defect concentrations $x$ are shown in Fig.\,3(f)-(i). For $x\!=\!0$ the symmetry-breaking diagonal zigzag state is realized, with degenerate ${\bf Q}\!=\!{\bf Q}_1$ and ${\bf Q}\!=\!{\bf Q}_2$ spin structures. At finite, low concentration $x\!\sim\!2\%$ those two configurations are no longer degenerate since one of them features slightly lower ground-state energy. We still have in this case a ``macroscopically stable'' ground state. At intermediate defect concentration $x\!\sim\!7\%$ the long-range order is in a strict sense destroyed. However, the large domain walls with either ${\bf Q}\!=\!{\bf Q}_1$ or ${\bf Q}\!=\!{\bf Q}_2$ seem to survive. At higher concentration $x\!\sim\!12\%$ the long-range order disappears completely. Moreover, we can now identify a mixture of local structures with different symmetry-breaking directions [see Fig.\,3(i)]. \section*{Discussion} In sum, we have calculated the microscopic neareast-neighbor magnetic interactions between effective 1/2 spins in Li$_2$RhO$_3$ and uncovered a substantial difference between the two types of bonds that are present: one is dominated by Heisenberg and the other by Kitaev types of couplings. The latter give rise to strong frustration, even if the interactions are predominantly ferromagnetic. In this setting we have additionally considered the effect of the presence of anti-site disorder. Experimentally the in-plane spin-defect concentration in Li$_2$RhO$_3$ has been estimated as $x\!=$10--15$\%$ \cite{kobayashi_95,Ir213_ryoji_2003}. Based on our theoretical findings it is likely that the observed spin-glass behavior arises from the combination of such anti-site disorder and strongly frustrating magnetic interactions, in particular, the different Kitaev/Heisenberg dominated magnetic bonds and the Ising-like physics associated with the triplet dimer formation that results from there. Our combined {\it ab initio} and effective-model calculations on both Li$_2$RhO$_3$ and related $d^5$ honeycomb iridates \cite{Ir213_rixs_gretarsson_2013,Na2IrO3_vmk_14,Li2IrO3_vmk_14} indicate that a description in terms of on-site 1/2 pseudospins can well account for the diverse magnetic properties of these systems. While alternative models rely on the formation of delocalized, quasimolecular orbitals \cite{Mazin_PRL_Na213_2012,213_hexagMO_2013} and for Li$_2$RhO$_3$ downplay the role of spin-orbit interactions \cite{213_hexagMO_2013}, here we show that the latter give rise in Li$_2$RhO$_3$ to anisotropic Kitaev interactions the same magnitude as in $5d$ iridates \cite{Na2IrO3_vmk_14,Li2IrO3_vmk_14,Ir213_KH_gretarsson_2013}. That happens in spite of having a Rh $t_{2g}$ splitting $\delta$ and a spin-orbit coupling $\lambda$ of similar magnitude, the same way similar sets of Ir $\delta$ and $\lambda$ parameters in CaIrO$_3$ \cite{Ir113_bogdanov_12} still generate symmetric anisotropic exchange terms in the range of 10 meV (work is in progress). \section*{Methods} The {\sc molpro} QC package was employed for all {\it ab initio} calculations~\cite{MOLPRO_WIREs}. To analyze the electronic ground state and the nature of the $d$-$d$ excitations, a cluster consisting of one reference RhO$_6$ octahedron plus three NN RhO$_6$ octahedra and 15 nearby Li ions was used. The magnetic spectrum for two Rh$^{4+}$ ions was obtained from calculations on a cluster containing two reference and four NN RhO$_6$ octahedra plus the surrounding 22 Li ions, see SI for details. The farther solid-state environment was in both cases modeled as a finite array of point charges fitted to reproduce the crystal Madelung field in the cluster region. The spin-orbit treatment was carried out according to the procedure described in Ref.~\citeonline{SOC_molpro}, using spin-orbit pseudopotentials for Ir. \section*{Acknowledgements} We thank V.~Yushankhai, Y.~Singh, N.~A.~Bogdanov, and U.~K.~R\"{o}{\ss}ler for useful discussions. L.~H. acknowledges financial support from the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG). \section*{Author contributions} V.M.K. carried out the {\it ab initio} calculations and subsequent mapping of the {\it ab initio} results onto the effective spin Hamiltonian, with assistance from L.H., H.S. and I.R. S.N. performed the exact-diagonalization calculations. V.M.K., S.N., J.v.d.B. and L.H. analyzed the data and wrote the paper, with contributions from all other coauthors. \section*{Additional information} \textbf{Competing financial interests:} Authors have no competing financial interests.

1. Introduction {#sec1-molecules-25-00136} =============== In order to obtain a compound with specific practical applications, supramolecular chemistry is centered at modification of the properties of macrocycles through structural modification. Typical modifications include changing the ring size of the macrocyclic ligand, the types of donor atoms and the kinds of substituents \[[@B1-molecules-25-00136],[@B2-molecules-25-00136],[@B3-molecules-25-00136]\]. Lariat ethers are still attractive research objects due to the possible creation of interesting structures of complexes the both with cations as well as anions. The terms "lariat ethers" refer to crown ethers having side chains attached to the crown moiety by the so-called pivot atoms (C, N, or P). The side arms provide the possibility of incorporating atoms or groups with a lone pair of electrons enabling additional cooperation with the heteroatoms of the macrocycle ring, thus providing three-dimensional coordination of the guest cation \[[@B4-molecules-25-00136],[@B5-molecules-25-00136],[@B6-molecules-25-00136]\]. Intensive development of the lariat ethers concept has led to the preparation of the number of crown ethers, which can be used in various applications such as classical (polymer-supported catalysts PTC, reagents for separation and extraction, etc.) and highly sophisticated (the redox switches for membrane transport, synthetic cation-conducting channels, nucleotide-based molecular boxes and enzyme mimics) \[[@B7-molecules-25-00136],[@B8-molecules-25-00136],[@B9-molecules-25-00136],[@B10-molecules-25-00136],[@B11-molecules-25-00136],[@B12-molecules-25-00136]\]. There are three known main types of lariat ethers: C-pivot lariat ethers \[[@B13-molecules-25-00136],[@B14-molecules-25-00136],[@B15-molecules-25-00136]\], where the side arm is attached to the carbon atom, N-pivot lariat ethers \[[@B16-molecules-25-00136],[@B17-molecules-25-00136],[@B18-molecules-25-00136]\], and finally P-pivot lariat ethers \[[@B19-molecules-25-00136],[@B20-molecules-25-00136]\]. The subjects of this research are the P-pivot lariat ether derivatives. The versatile precursors for this type of P-pivot lariat ethers have been obtained by the reaction of hexachlorocyclotriphosphazene with tetraethylene glycol in the presence of sodium hydride ([Scheme 1](#molecules-25-00136-sch001){ref-type="scheme"}). The numerous studies described in the literature on introduction different side groups indicated that oxygen and nitrogen nucleophiles cause regioselective substitution of chlorine atoms adjacent to the macrocycle. While the relevant sulphur bases such as sodium-sulphur thiolates, thiophenols, mercaptoalcohols, and dithiols, strongly prefer geminal substituting of the two chlorine atoms on the exomacrocyclic phosphorus atom from PCl~2~ group \[[@B21-molecules-25-00136],[@B22-molecules-25-00136],[@B23-molecules-25-00136],[@B24-molecules-25-00136],[@B25-molecules-25-00136],[@B26-molecules-25-00136]\], the introduction of additional "soft" nitrogen ligands derived from amino substituent (aziridine, pyrrolidine) increases the affinity of the ligand to the larger cations (potassium, rubidium, caesium, and silver). Ligands with amino substituents are also complexing with small transition metal ions like nickel and cobalt, and cadmium cation in the case of aziridinyl derivatives \[[@B25-molecules-25-00136]\]. Lariat ethers with alkylenediamines substituents have the complexing ability with both alkali metals, alkaline earth metals as well as transition metals. Such ligands, despite their good complexing abilities, do not exhibit high selectivity \[[@B27-molecules-25-00136],[@B28-molecules-25-00136],[@B29-molecules-25-00136]\]. In our research group, the studies of the complexing properties of macrocyclic compounds with particular focus on P-pivot lariat ethers have been conducted for years \[[@B30-molecules-25-00136],[@B31-molecules-25-00136],[@B32-molecules-25-00136],[@B33-molecules-25-00136],[@B34-molecules-25-00136]\]. These studies show that changes of substituents, as well as solvents used during the complexation process, affect the formation of the various structures of the complexes, especially with silver (I) cation. The studies for other cations comprise of the complexing behavior of PNP-lariat ether with other metal cations (Cd^2+^, Cu^2+^, Ca^2+^, and Pb^2+^) using ESI-MS and MS/MS techniques \[[@B35-molecules-25-00136],[@B36-molecules-25-00136]\]. Macrocyclic derivatives of cyclotriphosphazene and their complexes were found to reveal antitumor activity as well as anti-AIDS activity \[[@B37-molecules-25-00136],[@B38-molecules-25-00136],[@B39-molecules-25-00136]\]. Furthermore, these types of compounds are potentially useful as starting materials for the preparation of 'pH-controlled active ion carriers' in liquid membranes \[[@B40-molecules-25-00136]\]. This article describes the new P-pivot lariat ether formed by substitution of chlorine atoms in tetrachloro-PNP-macrocycle with the N-benzyl derivative of (S)-prolinamine as side groups ([Scheme 2](#molecules-25-00136-sch002){ref-type="scheme"}). This derivative was obtained from the enantiomerically pure (S)-proline, which is widely used in organic synthesis \[[@B41-molecules-25-00136],[@B42-molecules-25-00136]\] and is the convenient substrate in the preparation of a number of derivatives with potential biological activity \[[@B43-molecules-25-00136],[@B44-molecules-25-00136]\]. The resulting ligand L has much more binding sites in its structure that can potentially attract metal ions: (1) Oxygen donor atoms derived from the macrocyclic backbone, (2) endocyclic nitrogen atom from the cyclotriphosphazene ring, (3) exocyclic nitrogen atoms from attached substituents, and (4) benzyl groups. To date, macrocyclic cyclotriphosphazene derivatives have not been modified by this kind of side group. Introducing them into the system will allow us to obtain compounds with a stereogenic center, which is important from the point of view of application as potential pharmacological as well as chiral recognition agents. For the obtained ligand L, the complexing capabilities towards Ag^+^, Cu^2+^, Co^2+^, Ni^2+^, and Zn^2+^ ions have been examined. The choice of such a series of ions was imposed by the possibilities of subsequent applications. In particular silver (I) complexes are proposed for application in radioimmunotherapy \[[@B45-molecules-25-00136],[@B46-molecules-25-00136],[@B47-molecules-25-00136],[@B48-molecules-25-00136]\], as well as an antimicrobial agent \[[@B49-molecules-25-00136],[@B50-molecules-25-00136]\] especially for pulmonary infections related to cystic fibrosis \[[@B51-molecules-25-00136],[@B52-molecules-25-00136]\]. Furthermore, coordination polymers based on silver (I) cations with different ligands attract attention \[[@B53-molecules-25-00136],[@B54-molecules-25-00136]\] due to the possibility of constructing new functional materials \[[@B55-molecules-25-00136],[@B56-molecules-25-00136],[@B57-molecules-25-00136],[@B58-molecules-25-00136]\] based on functionalized cyclotriphosphazene core. Complexes of transition metals such as Cu^2+^, Co^2+^, Ni^2+^, and Zn^2+^ are interesting because of their use as hydrolytic metalloenzyme models \[[@B48-molecules-25-00136]\]. Hydrolases play an important role in the biochemical processes, which can catalyze the hydrolysis of various compounds like carboxylic esters, phosphoesters, amides, nucleic acids, and peptides. The design of artificial hydrolases leads to vast application prospects in many aspects, e.g., molecular genetic tools, the degradation of toxic phosphate esters, the design of medicine, and functional materials \[[@B59-molecules-25-00136],[@B60-molecules-25-00136],[@B61-molecules-25-00136],[@B62-molecules-25-00136]\]. Recent reports present metal complexes such as Cu^2+^, Co^2+^, Ni^2+^, Zn^2+^ with macrocycles having in its molecule nitrogen and oxygen atoms behave perfectly as an artificial hydrolase. This basic research for new derivatives of P-pivot lariat ethers will allow us to determine how stable complexes give these compounds and if possible, their use as artificial hydrolases \[[@B63-molecules-25-00136]\]. 2. Results and Discussion {#sec2-molecules-25-00136} ========================= In this paper, we reported the preparation of a new macrocyclic derivative of cyclotriphosphazene by the substitution of its reactive chlorine atoms with (S)-(1-methylpyrrolidin-2-yl) methanamine ([Scheme 3](#molecules-25-00136-sch003){ref-type="scheme"}). 2.1. Investigations of Silver Ions Complexes with the Ligand {#sec2dot1-molecules-25-00136} ------------------------------------------------------------ Simple direct titration of the ligand with AgNO~3~ solution was applied for the determination of formation constants of Ag^+^ complexes with investigated compounds in acetonitrile and methanol. The analyses were performed at 298.1 ± 0.1 K with the use of 0.1 M (C~2~H~5~)~4~NNO~3~ as the supporting electrolyte. Before each measurement Nernst's equation was verified in the following system: The specific measurements were performed in the system: where cL^0^ is the initial concentration of the ligand, and cAg is the concentration of silver ions. The experiments were triplicate for three different ligand concentrations cL^0^, which ranged from 1͘ × 10^−3^ to 1 × 10^−2^ M. The obtained reproducibility of the results was ± 0.20 mV. The activity of the unbound silver ions was calculated on the basis of the emf values applying the Nernst equation. The stability constants were calculated using HYPERQUAD 2003 program \[[@B64-molecules-25-00136]\]. In the case of our investigation, we made the assumption for different models of complexation equilibria and the resulting fitting parameters are presented in the final models. The presumption of the other models gave a very poor chi-square parameter or gave no reliable results at all. The determined values of stepwise stability constants were summarized in [Table 1](#molecules-25-00136-t001){ref-type="table"}. As can be seen from [Table 1](#molecules-25-00136-t001){ref-type="table"}, almost in all cases stability constants of investigated complexes were higher in acetonitrile solution. This tendency to form complexes with higher stability constants with macrocyclic cyclotriphosphazene derivatives occurs for the first time. Usually, other derivatives were characterized by higher stability constants in methanol \[[@B31-molecules-25-00136]\]. The much easier formation of the complexes in acetonitrile than in methanol was confirmed by MS analysis ([Figure 1](#molecules-25-00136-f001){ref-type="fig"}). On MS spectra in methanol ([Figure 1](#molecules-25-00136-f001){ref-type="fig"}b), we did not see signals indicating the various forms of the complex, whereas in acetonitrile ([Figure 1](#molecules-25-00136-f001){ref-type="fig"}a) ligand signals with silver ions were immediately visible (*m*/*z* = 1192). This was probably related to the strong solvation of the ligand by methanol. Solvent molecules will more easily penetrate into the polyether part, to which access is facilitated, as they are not heavily hindered by large substituents. Thus, the complex formed with a 1: 1 stoichiometry in methanol, which is formed with exocyclic nitrogen atoms from substituents, has a higher stability constant (log β = 8.61) in this solvent. The formation of further forms of complexes is probably hampered by solvent molecules. It is also worth noting that Fyles and Zeng, during the study of the formation of supramolecular complexes of cations with crown ethers, noted the electrospray ionization process can result in differential production of ions from the equilibrium mixture and comparisons between ions of differing charges or stoichiometries may not correspond to the inherent selectivity of the host in solution \[[@B65-molecules-25-00136]\]. In the case of investigated ligand, we observed four different types of complexes: AgL (1:1), Ag~2~L (2:1), Ag~3~L (3:1), and Ag~4~L (4:1). In both solvents, the most stable form of the complex is a system where one molecule of the ligand binds the first silver ion. Contrary to the previously-described PNP-lariat ethers derivatives \[[@B31-molecules-25-00136],[@B32-molecules-25-00136],[@B33-molecules-25-00136],[@B34-molecules-25-00136]\] the sites and way of successive binding of silver ions (as confirmed by NMR and MS studies) were completely different. In the AgL system, the silver ion is bound by exocyclic nitrogen atoms derived from substituents ([Scheme 4](#molecules-25-00136-sch004){ref-type="scheme"}). In the NMR spectra taken after the addition of the amount of silver ions corresponding to the system in a 1:1 ratio ([Figure 2](#molecules-25-00136-f002){ref-type="fig"}b), we see changes in the range of 3.10--2.65 ppm what implies that the first silver ion is binding by nitrogen atoms derived from the substituents attached to the exocyclic phosphorus atom. The second silver ion is bound by the endocyclic nitrogen atom from the cyclotriphosphazene ring. The method of binding two silver ions through ligand L is shown in [Scheme 5](#molecules-25-00136-sch005){ref-type="scheme"}. Also here, the confirmation of the binding of the second silver ion is the NMR spectra ([Figure 2](#molecules-25-00136-f002){ref-type="fig"}c). The shifts of signals from the CH~2~OP group in the range of about 4 ppm are clearly visible and are similar to the ones previously described for such place of the silver ion binding \[[@B31-molecules-25-00136],[@B32-molecules-25-00136],[@B33-molecules-25-00136],[@B34-molecules-25-00136]\]. In this stage of complexation, one can observe the shift of the signals located at about 3.30 ppm assigned to CH~2~-O fragments of the polyether ring. Such behavior suggests the additional interaction of the oxygen atoms of polyether moiety in the process of binding of the second silver ion. Subsequent silver ions in the Ag~3~L and Ag~4~L systems are bound by interaction with the π electrons of the benzyl rings in the attached substituents. In the NMR spectra ranging from 1 ppm to 5 ppm for both acetonitrile ([Figure 3](#molecules-25-00136-f003){ref-type="fig"}), no significant shifts are observed which would indicate the other site of silver ions binding. On the other hand, there are significant shifts above 7 ppm, which indicate the interaction of cation with π electrons of benzene rings ([Figure 4](#molecules-25-00136-f004){ref-type="fig"}). For the studied ligand L, we do not observe the sandwich complex, which is related to the high steric hindrance caused by the spacious substituents. 2.2. Mass Spectrometry Investigations of Ligand and Its Complexes with Ag^+^ Cu^2+^, Co^2+^, Ni^2+^, and Zn^2+^ Ions {#sec2dot2-molecules-25-00136} -------------------------------------------------------------------------------------------------------------------- The ESI-MS results obtained for the basic measurements of the studied complexes are presented in [Table 2](#molecules-25-00136-t002){ref-type="table"}. As one can see, the spectra of the investigated ligand L revealed the presence of the protonated molecule, at *m*/*z* = 1084.568 (\[L + H\]^+^), and a less abundant signal at *m*/*z* = 1106.582 corresponding to the sodium adduct of compound L+Na^+^ ([Figure 5](#molecules-25-00136-f005){ref-type="fig"}). The MS/MS fragmentation spectrum for the protonated ligand at *m*/*z* 1084 is presented in [Figure 6](#molecules-25-00136-f006){ref-type="fig"}. The analysis of the observed fragmentation ions shows two different series. The first fragmentation pathway passes through fragment ions with *m*/*z* 894, 706, 618, 547, 459 and 355 ([Scheme 6](#molecules-25-00136-sch006){ref-type="scheme"}). In the first phase, two substituents attached to endocyclic phosphorus atoms are detached, then the ring is broken down and the next fragment is removed from the substituent. The second path ([Scheme 7](#molecules-25-00136-sch007){ref-type="scheme"}) proceeds with the incomplete detachment of the two substituents attached to the endocyclic nitrogen atoms and further cleavage of polyether chain until a fragment of only one substituent is obtained (*m*/*z* 912, 722, 399, and 191). After the determination of the fragmentation pattern of the protonated ligand, we took the ESI-MS spectra for the mixtures of the investigated compound with the nitrate salts of the following ions: Ag^+^, Co^2+,^ Cu^2+,^ Ni^2+^, and Zn^2+^. Analysis of ESI-MS spectra showed that all metal ions are bound by the investigated ligand. For a monovalent silver ion, some diverse forms of complexes were obtained, whereas, in the case of divalent cations, only complexes with a nitrate anion were observed. The confirmation of the obtained stoichiometry and ascribing to the presented forms were done on the basis of isotopic pattern comparison of the observed and simulated spectra. We can observe four different forms of complexes for Ag^+^ ions ([Figure 7](#molecules-25-00136-f007){ref-type="fig"}). The most abundant form of the complex at *m*/*z* 1192 is derived from the \[Ag^(I)^L\]^+^ system. The second intensive signal is from the \[Ag^(I)^~2~L\]^2+^ system with two silver ions bound by the ligand. The Ag~3~L and Ag~4~L complexes were much less abundant than the ones with Ag:L ratio equal 1:1 or 2:1. In each of the observed complexes, the characteristic doubled signals with the difference of 2 Da can be noticed. Such shape of the spectrum is typical of silver complexes due to its existence in 2 isotopic forms (^107^Ag and ^109^Ag in comparable amounts). Binding of silver ions, as confirmed by NMR studies provided here, takes place through nitrogen atoms derived from substituents attached to the exocyclic phosphorus atom, an endocyclic nitrogen atom and through benzene rings. Fragmentation of the isolated signal *m*/*z* 1192 ([Scheme 8](#molecules-25-00136-sch008){ref-type="scheme"}) proceeds in the first stage with detaching of the substituents attached to the endocyclic phosphorus atoms (*m*/*z* 1002 and 814), followed by separation of the substituent and polyether ring fragments. The final signal on the spectrum at *m*/*z* 1192 comes from a system in which there is a silver ion, which indicates a strong binding of this ion by the investigated ligand. All divalent metal cations Co^2+,^ Cu^2+,^ Ni^2+^, and Zn^2+^ form only one type of complexes in which the nitrate anion is involved. The lack of other forms of complexes with divalent cations ([Table 1](#molecules-25-00136-t001){ref-type="table"} and [Figure 8](#molecules-25-00136-f008){ref-type="fig"}, [Figure S1](#app1-molecules-25-00136){ref-type="app"}, [Figure S2](#app1-molecules-25-00136){ref-type="app"}, [Figure S3](#app1-molecules-25-00136){ref-type="app"}) for Ni^2+^, Co^2+^, Cu^2+^, and Zn^2+^ complexes, respectively) probably is related to very high steric hindrance of the parent molecule strictly connected with the branchy substituents attached to the cyclotriphosphazene subunit. Analyzing the ESI-MS spectra as well as tandem mass spectra and the subsequent fragmentation patterns of the complexes with individual divalent cations [Figure 8](#molecules-25-00136-f008){ref-type="fig"} and [Scheme 9](#molecules-25-00136-sch009){ref-type="scheme"} for \[Ni^(II)^L + NO~3~\]^+^ complex, and [Figures S1--S6](#app1-molecules-25-00136){ref-type="app"} and [Schemes S1--S4](#app1-molecules-25-00136){ref-type="app"} for the other investigated divalent metal ions complexes, one can see a similar tendency in the observed fragmentation routes. At the first step, the fragmentation of individual forms of complexes \[ML + NO~3~\]^+^ begins with the detachment of the nitrate anion, and then as a result of the ionization, the molecule loses the individual fragments of the substituents. For all doubly charged metal ions complexes, in all complexes of the ligand with divalent metal ions, the fragmentation patterns clearly show that the binding of the metal ion occurs via the endocyclic nitrogen atom. Due to the fact that in case of Ni(II), Zn(II), and Co(II) complexes the tandem mass spectra do not show the peaks related to the breaking of the polyether macrocyclic moiety of the ligand one can suppose that binding of the metal ion is also supported by additional interaction with oxygen atoms of the ligand cavity. Only in the case of Cu(II) ion, which in the experimental conditions is reduced to Cu(I), has the cleavage of the macrocyclic ring been observed, suggesting that in this case the binding of the metal ion is realized solely by endocyclic nitrogen atom of the cyclotriphosphazene subunit. Furthermore, the interaction is probably strengthened by additional stabilization by oxygen atoms from crown ether moiety (there is no cleavage of the macrocycle skeleton under the influence of the applied voltage during the experiment). Similar fragmentation pathways were also noticed in previous studies of these complexes with morpholinyl and piperidinyl substituents \[[@B35-molecules-25-00136],[@B36-molecules-25-00136]\]. 3. Materials and Methods {#sec3-molecules-25-00136} ======================== 3.1. Materials {#sec3dot1-molecules-25-00136} -------------- ### 3.1.1. For the Synthesis {#sec3dot1dot1-molecules-25-00136} (S)-(1-methylpyrrolidin-2-yl) methanamine was synthesized according to the procedure presented earlier \[[@B66-molecules-25-00136]\]. We used Benzene (Sigma Aldrich) CHROMASOLV^®^ Plus "for HPLC", and Hexane (POCh Gliwice, 99%). The reactions were carried out under argon atmosphere (dry). ### 3.1.2. For the NMR Measurements {#sec3dot1dot2-molecules-25-00136} For all NMR measurements, we used deuterated solvents from Sigma Aldrich: Chloroform-d (99.8 atom % D), Methanol-d4 (99.8 atom % D) and Acetonitrile-d3 (99.8 atom % D). ### 3.1.3. For the Potentiometric Measurement {#sec3dot1dot3-molecules-25-00136} Acetonitrile (Sigma Aldrich) for HPLC and Methanol (J. T. Baker) for HPLC were used. AgNO~3~ (Sigma Aldrich, 99.0%) and tetraethylammonium nitrate (Sigma Aldrich, 98.0%, NT) were used. ### 3.1.4. For the MS Measurements {#sec3dot1dot4-molecules-25-00136} AgNO~3~, Cu(NO~3~)~2~, Co(NO~3~)~2~, Ni(NO~3~)~2~, and Zn(NO~3~)~2~ were analytical grade compounds from Sigma Aldrich. We used Acetonitrile for HPLC "Gradient grade" (Sigma Aldrich) and Methanol (J. T. Baker) for HPLC "Gradient Grade". All solutions in methanol and acetonitrile (10^−3^ M each) were prepared daily prior to dilution to 10^−4^ M for mass spectrometric investigation (ligand/metal salt ratio 1:1). In each case, the freshly prepared solutions were used. 3.2. Characterization Techniques {#sec3dot2-molecules-25-00136} -------------------------------- ### 3.2.1. NMR Measurements {#sec3dot2dot1-molecules-25-00136} NMR spectra were recorded with Bruker (Avance III 600) instrument in CDCl~3~; chemical shifts are reported relative to solvent residual peak (^1^H: δ = 7.26 ppm \[CDCl~3~\]; ^13^C: δ = 77.0 ppm \[CDCl~3~\]; for ^31^P: δ = 0.0 ppm external reference 85% H~3~PO~4~). ### 3.2.2. Elemental Analysis {#sec3dot2dot2-molecules-25-00136} Elemental analysis (C, H, N) was recorded using the Elementar Super Vario Micro Cube apparatus. ### 3.2.3. Potentiometric Measurements {#sec3dot2dot3-molecules-25-00136} The potentiometric measurements were carried out with the use of Cerko microburette controlled by computer with silver electrode Ag^0^/0.01 M AgNO~3~ + 0.09 M (C~2~H~5~)~4~NNO~3~ in methanol solution as a reference electrode. The activity of the uncomplexed silver ions was calculated from the emf values by using the Nernst equation verified at the first step of the measurement procedure \[[@B34-molecules-25-00136]\]. The stability constants were computed by using HYPERQUAD 2003 program \[[@B64-molecules-25-00136]\]. To determine the stability constants of the complexes of investigated ligands with the other selected metal ions, the competitive potentiometric method was used \[[@B30-molecules-25-00136]\]. ### 3.2.4. ESI-MS^n^ Analyses {#sec3dot2dot4-molecules-25-00136} The structural characterization of the complexes obtained was performed using Thermo Scientific LCQ Fleet ion trap mass spectrometer with electrospray ionization (Thermo Fisher Scientific Inc., San Jose, CA, USA). Samples of the complexes were dissolved in acetonitrile or methanol, and then the respective solutions were introduced into the ESI-MS source by continuous infusion by means of the instrument syringe pump at a rate of 10 μL min^−1^. Nitrogen was used as the nebulizing gas. The ESI source was operated at the range 3 to 3.5 kV (depending on the experiment), and the capillary heater was set to 200 °C. For ESI-MS/MS experiments, the ions of interest were isolated monoisotopically in the ion trap and were collisionally activated and the helium damping gas that was present in the mass analyzer acted as a collision gas. The RF amplitude was set to a value that caused to decrease in the peak height of the fragmented molecular ion in the range if 70% to 30% (depending on the experiment). All experiments were performed in positive-ion mode. All the investigated fragmentation patterns were based on the results obtained from MS-Fragmenter and Spectrus Processor 2016 software (Advanced Chemistry Development, Inc., Toronto, Ontario, M5C 1B5, Canada) by ACD/Labs. ### 3.2.5. TLC Analyses {#sec3dot2dot5-molecules-25-00136} Silica Gel 60-F254 plates (Sigma Aldrich) were used for TLC analyses. 3.3. Synthesis {#sec3dot3-molecules-25-00136} -------------- 1,3-(oxytetraethylenoxy)-1,3,5,5-tetra-\[(S)-(1-methylpyrrolidin-2-yl)methanamino\]-cyclotriphosphazene (L). 1,3-(Oxytetraethylenoxy)-1,3,5,5-tetrachlorocyclotriphosphazen (0.235 g, 0.0005 mole) and (S)-(1-methylpyrrolidin-2-yl) methanamine (0.760 mL, 0.004 mole) in dry benzene (50 mL) were placed in a 3-necked round-bottomed flask (100 mL) equipped with argon inlet tube. The reaction has been carried out at r.t. with continuous stirring for 72 h. TLC analysis was applied to monitor the progress of the reaction. In this case, hexane-THF (2:3) mixture was used as eluent. Then the reaction mixture was filtered and the solvent was removed under reduced pressure. The yield of the obtained ligand L: 52.6%. ^1^H NMR (600 MHz, CDCl~3~), δ(ppm): 7.32--7.17 (20H, m, HC(arom.)); 4.10 (2H, m, CH~2~OP); 3.99 (2H, m, CH~2~OP); 3.98 (4H, d, CH~2~-Ph); 3.78--3.50 (12H, m, CH~2~OC); 3.30 (4H, d, CH~2~-Ph); 3.10--3.00 (8H, m, CH~2~-NH); 2.85--2.75 (4H, m, P-NH); 2.99--2.89 (4H, m, H~2~C(5)); 2.85--2.75 (4H, m, HC(2)); 2.24--2.17 (4H, m, H~2~C(5)); 1.97--1.83 (4H, m, H~2~C(3)); 1.85--1.75 (4H, m, H~2~C(3)); 1.73--1.63 (8H, m, H~2~C(4)); ^13^C NMR (151 MHz, CDCl~3~), δ(ppm): 139.4, 128.7, 128.3, 126.9, 70.7, 70.3, 64.0, 62.1, 58.9, 58.4, 54.3, 54.1, 43.2, 42.3, 28.1, 23.0; ^31^P NMR (242 MHz, CDCl~3~), δ (ppm): 22,0 (d, P(OCH~2~)NNBn); 19.7 (t, P(NNBn)~2~, JPP = 50.0 Hz, A2B system); ESI-MS *m*/*z* = 1084.7 \[M + H^+^\], 1107.6 \[M + Na^+^\]; elemental analysis: calculated for C~56~H~84~N~11~O~5~P~3~: C 62.03; H 7.81; N 14.21; found: C 62.15; H 7.59; N 14.14. 4. Conclusions {#sec4-molecules-25-00136} ============== In this work, the synthesis of a new macrocyclic derivative with optically active substituents in the form of a benzyl derivative of (S)-prolineamine is presented. Potentiometric and NMR studies were presented for the investigated ligand complexes with silver ions in acetonitrile and methanol. Electrochemical, NMR, and MS studies have confirmed the unexpected and promising possibilities of complex formation by a new compound. Ligand with silver ions forms four types of complexes with different stoichiometry-AgL, Ag~2~L, Ag~3~L, and Ag~4~L, with the most stable 1:1 form observed in both solvents. Unlike similar derivatives the obtained stability constants, are higher in acetonitrile, which results from the stronger solvation of our ligand by methanol molecules. Furthermore, silver ion in the AgL complex is preferably attracted to nitrogen atoms from substituents which are attached to the exocyclic phosphorus atom. In the previously described PNP lariat ethers derivatives, binding of the first silver ion was realized by the endocyclic nitrogen atom from the cyclotriphosphazene subunit \[[@B30-molecules-25-00136],[@B31-molecules-25-00136],[@B32-molecules-25-00136],[@B33-molecules-25-00136],[@B34-molecules-25-00136],[@B35-molecules-25-00136],[@B36-molecules-25-00136]\]. The second silver ion is bound by the endocyclic nitrogen atom of the cyclotriphosphazene ring. However, in Ag~3~L and Ag~4~L systems, there are interactions between silver ions and π electrons of benzyl rings. MS studies confirmed the existence of complexes with Ag^+^, Cu^2+^, Co^2+^, Ni^2+^, and Zn^2+^ ions. For a monovalent silver cation, the existence of four types of complexes has been confirmed. The presence of the AgL^+^ complex with unusual binding through the nitrogen atoms from the exocyclic substituents may provide the structural unit to build a new coordination polymer. Divalent cations form only one type of complexes additionally supported by nitrate anion. Binding of Co^2+^, Ni^2+^, and Zn^2+^ ions takes place via the endocyclic nitrogen atom from the cyclotriphosphazene ring supported by oxygen atoms of polyether macrocyclic moiety. Only the case of Cu^2+^ ions reduction from copper (II) to copper (I) ion has been observed during experimental conditions and the obtained complex is only formed by endocyclic nitrogen atom from cyclotriphosphasen subunit. The MS/MS results show that all types of complexes that are formed are stable under the conditions of applied voltage and do not undergo fragmentation with metal ion detachment. **Sample Availability:** Samples of the compound **L**. are available from the authors. ###### Click here for additional data file. The following are available online. Figure S1: The ESI-MS spectrum of L with Co(NO~3~)~2~, Figure S2: The ESI-MS spectrum of L with Cu(NO~3~)~2~, Figure S3: The ESI-MS spectrum of L with Zn(NO~3~)~2~, Figure S4: The ESI-MS/MS^2^ spectrum of the complex with Co^2+^ ion at *m*/*z* = 1204, Figure S5: The ESI-MS/MS^2^ spectrum of the complex with Cu^2+^ ion at *m*/*z* = 1208, Figure S6: The ESI-MS/MS^2^ spectrum of the complex with Zn^2+^ ion at *m*/*z* = 1209, Scheme S1: Fragmentation pathways of the \[Co^(II)^L+NO~3~\]^+^ complex ion *m*/*z* 1204, Scheme S2: The first fragmentation pathway of the \[Cu^(II)^L+NO~3~\]^+^ complex ion *m*/*z* 1204, Scheme S3: The second fragmentation pathway of the \[Cu^(II)^L+NO~3~\]^+^ complex ion *m*/*z* 1204, Scheme S4: Fragmentation pathway of the \[Zn^(II)^L+NO~3~\]^+^ complex ion *m*/*z* 1204. Conceptualization, N.G. and P.S.; Methodology, G.A., M.K., N.G. and P.S.; Synthesis, N.G., J.R.; Formal Analysis, N.G., M.Z. and P.S.; Investigation, N.G., M.Z. and P.S.; Resources, N.G.; Data Curation, N.G. and P.S.; Writing---Original Draft Preparation, N.G. and P.S.; Writing---Review and Editing, G.A. and J.R.; Visualization, N.G. and J.R.; Supervision, G.A., M.K.; Project Administration, N.G.; Funding Acquisition, N.G., and J.R. All authors have read and agreed to the published version of the manuscript. The authors would like to thank the Polish National Centre of Science for financial support (grant no. PRELUDIUM UMO-2015/19/N/ST5/00694). The authors declare no conflict of interest. Figures, Schemes and Tables =========================== {#molecules-25-00136-sch001} {#molecules-25-00136-sch002} {#molecules-25-00136-sch003} {#molecules-25-00136-f001} {#molecules-25-00136-sch004} {#molecules-25-00136-f002} {#molecules-25-00136-sch005} {#molecules-25-00136-f003} {#molecules-25-00136-f004} {#molecules-25-00136-f005} ![The ESI-MS/MS^2^ spectra in acetonitrile of protonated ligand \[L + H\]^+^ at *m*/*z* = 1084, and ESI-MS/MS^3^ of signals at *m*/*z* = 912 and 706.](molecules-25-00136-g006){#molecules-25-00136-f006} {#molecules-25-00136-sch006} {#molecules-25-00136-sch007} {#molecules-25-00136-f007} ![Fragmentation pathways of the \[Ag(I)L\]^+^ complex ion *m*/*z* 1192.](molecules-25-00136-sch008){#molecules-25-00136-sch008} ![The ESI-mass spectrum of the complex obtained by mixing acetonitrile solutions of L with Ni(NO~3~)~2~ and the ESI-MS/MS^2^ spectrum of the \[Ni^(II)^L + NO~3~\]^+^ complex form (inset).](molecules-25-00136-g008){#molecules-25-00136-f008} ![Fragmentation pathways of the \[Ni^(II)^L + NO~3~\]^+^ complex ion *m*/*z* 1203.](molecules-25-00136-sch009){#molecules-25-00136-sch009} molecules-25-00136-t001_Table 1 ###### The stability constants of the complexes of the investigated ligand with Ag^+^ ions determined by potentiometric measurements. Complex Form L --------------------------------- ------------- ------------- Ag^+^ + L → AgL^+^ 7.12 ± 0.01 8.61 ± 0.01 AgL^+^ + Ag^+^ → Ag~2~L^2+^ 5.94 ± 0.01 3.83 ± 0.01 Ag~2~L^2+^ + Ag^+^ → Ag~3~L^3+^ 4.28 ± 0.01 3.41 ± 0.01 Ag~3~L^3+^ + Ag^+^ → Ag~4~L^4+^ 3.06 ± 0.01 1.99 ± 0.01 molecules-25-00136-t002_Table 2 ###### Complexes confirmed by electrospray mass spectrometry and tandem mass spectrometry methods (ESI-MS/MS). ---------------------------------------------------------------------------------------------------------------------------- L + M L + 2M L + 3M L + 4M L + M + NO~3~ -------- ------------------ ---------------------- ---------------------- ---------------------- --------------------------- Ag^+^ \[Ag^(I)^L\]^+^\ \[Ag^(I)^~2~L\]^2+^\ \[Ag^(I)^~3~L\]^3+^\ \[Ag^(I)^~4~L\]^4+^\ \-\-- 1192.388 648.966 468.395 378.340 Co^2+^ \-\-- \-\-- \-\-- \-\-- \[Co^(II)^L + NO~3~\]^+^\ 1204.090 Cu^2+^ \-\-- \-\-- \-\-- \-\-- \[Cu^(II)^L + NO~3~\]^+^\ 1208.414 Ni^2+^ \-\-- \-\-- \-\-- \-\-- \[Ni^(II)^L + NO~3~\]^+^\ 1203.062 Zn^2+^ \-\-- \-\-- \-\-- \-\-- \[Zn^(II)^L + NO~3~\]^+^\ 1209.260 ---------------------------------------------------------------------------------------------------------------------------- [^1]: Current affiliation: Łukasiewicz Research Network-Institute of Biopolymers and Chemical Fibres, Marii Skłodowskiej-Curie 19/27, 90570 Łódź, Poland.

Background {#Sec1} ========== Single nucleotide polymorphism (SNP) arrays allow for a whole genome view of complex copy number changes (CNCs) and loss of heterozygosity (LOH), Recent microarray developments have incorporated molecular inversion probe (MIP) technology, which requires a relatively small molecular footprint (\~40 bp) and is advantageous for assessing highly cross-linked and degraded DNA samples obtained from formalin-fixed, paraffin-embedded (FFPE) tissues \[[@CR1], [@CR2]\]. This ability to optimally assess copy number changes in FFPE tissue is significant to studying genomic alterations in solid tumors because solid tumors have traditionally been characterized through their histology and through FFPE tissue FISH. As a result, FFPE-preserved solid tumor samples have been archived for decades, awaiting improved technologies to better analyze them. MIP arrays, therefore, provide opportunities to characterize genomic profiles from archived FFPE samples that are already available for such studies. This study focused on assessing the genomic signatures of gliomas--brain tumors derived from the glial cells. Gliomas are the most common form of primary brain tumors comprising approximately 30 % of all brain tumors. More importantly, about 80 % of all malignant brain tumors are gliomas \[[@CR3]\]. The World Health Organization (WHO) has classified gliomas from I-IV based on their degree of aggressiveness, with WHO class I being relatively benign and WHO class IV as the most aggressive gliomal tumors \[[@CR4]\]. The objective of this study was to investigate the genomic changes that occur in gliomas. We performed microarray analysis on 2 types of glial tumors to identify and compare the respective genomic landscapes. Results and discussion {#Sec2} ====================== The two categories of gliomas analyzed in this study were initially identified by FISH testing and subsequently assessed for CNCs using the OncoScan® array (Affymetrix, Santa Clara, CA). Global CNCs were analyzed to identify consistent patterns across each cohort. The first group consisted of nine oliogodendriolglial tumors that had 1p19q co-deletions identified by FISH analysis. The second group contained 8 glioblastomas (GBM), WHO class IV, initially characterized by EGFR amplification, also identified by FISH. 1p19q Co-deleted {#Sec3} ---------------- Microarray results confirmed the 1p19q co-deletions. One strength of array technology is the broader coverage obtained in the assay. Significantly, FISH probes yield very focal results, identifying only the 200--700 kb region where the probe hybridizes. Array results show deletions of the entire 1p and 19q arms in all 9 samples (Fig. [1a](#Fig1){ref-type="fig"}). Typically, 1p19q co-deletions are consistent with a favorable prognosis, whereas partial 1p deletions are indicative of a poor prognosis \[[@CR5]\]. Although FISH is often the first line of testing for 1p19q deletions, the technology is limited and cannot distinguish between full chromosomal arm losses as opposed to partial losses that span the probe sites.Fig. 11p19q co-deletion cohort. **a** Highlights the 1p and 19q co-deletions (deletions marked by indicated region). Case 1p19q-03 had two copies of 1p and 19q, but was tetraploid throughout the remainder of the genome (blue arrow). **b** Highlights whole chromosomal changes observed in this cohort including gains (blue arrows) and deletions (red arrows) One case (1p19q-03), with diploid copy number along 1p and 19q arms, was tetraploid across the remainder of the genome (Fig. [1a](#Fig1){ref-type="fig"}). Genotyping data from the array indicate that 1p and 19q were co-deleted prior to a doubling of the genome, so that although diploid in number, the 1p and 19q arms are still deleted in the context of the entire genome. This is consistent with the FISH findings which compare the 1p or 19q probe to a control probe found on the opposing chromosome arm. Overall the 1p19q cohort had a relatively quiet genome, limited to the 1p and 19q co-deletions. Additional genomic changes were observed in the 1p19q cohort, the majority of which were relatively large scale, tending to encompass entire chromosomes. Genomic gains included trisomy 11, 17 & 21, trisomy 22 and one case that presented a mixed gain of chromosome 7, with 3 copies of 7p and 4 copies of 7q. Genomic losses included monosomy 14, monosomy 15 & 18, and a second case of monosomy 18 (Fig. [1b](#Fig1){ref-type="fig"}). These changes occurred case by case and not across the entire cohort, suggesting that additional genomic changes are not specifically associated with the 1p19q co-deletions. EGFR Amplified {#Sec4} -------------- The second cohort contained 8 GBM cases positive for EGFR amplification by FISH. Amplification of EGFR was confirmed in each case by OncoScan (Figs. [2a](#Fig2){ref-type="fig"} and [3a](#Fig3){ref-type="fig"}). Amplifications varied, ranging from 8--61 copies of the EGFR gene. EGFR copy number (CN) was called by the analysis software, which is designed to estimate the CN based on both the log2 ratio, as well as the estimated tumor content, giving an integer CN call for the tumor portion of the sample in most cases. In some cases, with significant tumor heterogeneity, the tumor content cannot accurately be determined by the software. In these cases the CN calls are binned and portrayed as an average CN change across the entire cell population.Fig. 2Copy number changes observed in GBMs. **a** A representative sample of EGFR amplification that was confirmed by array in each case (blue circle). **b** Additional CNCs observed across the GBM cohortFig. 3Additional copy number changes observed in the GBM cohort. **a** Whole genome overview of the glioblastoma cohort includes **b** Chromosome 7 complexity, **c** CDKN2A/B deletions and **d** Chromosome 10 deletions We compared the amplification calls made by FISH with those made by array (Table [1](#Tab1){ref-type="table"}). In addition, 3 samples were repeated on the array to determine reproducibility. FISH copy numbers were calculated as (signals counted)/(number of cells). Similar patterns were observed with respect to relative amounts of amplification. There was some variability in CN estimation---shown by the repeated arrays. However, the advantage of the CN call made by the array is that it is an estimation based on the entire sample, rather than subjectively called after assessing only regions of amplification. For example, EGFR-03 contained patches of high amplification amidst large regions of little or no amplification, within the tumor. The FISH report reflects a focus on scoring amplified regions while passing over "normal" regions. The array data, on the other hand, accounts for both, giving an overall count of EGFR amplification across the tumor. FISH analysis is open to more subjectivity and in cases of high copy gains where signals are clustered together it is often difficult to get an accurate signal count.Table 1CN comparison of EGFR by array and FISHSampleMIP arrayFISHEGFR CNRepeatEGFR CNEGFR-0137\>20EGFR-025137\>10EGFR-038\>13.20EGFR-0461\>20EGFR-0647\>15.98EGFR-0713.33^a^6.00EGFR-093427\>16.0EGFR-10221816.95^a^The tumor content could not be determined, likely due to the presence of more than 2 cell populations with differing degrees of CNCs. The CN reported here is therefore an average of the EGFR CN across the entire tissue biopsy After confirming EGFR amplification we analyzed each case to identify other CNCs found across this cohort. Recurrent CNCs identified in this group included complex gains across chromosome 7, CDKN2A/B deletions, loss of chromosome 10 and chromosomal changes of RB1 affecting 5 of 8 cases (Fig. [2b](#Fig2){ref-type="fig"}). In addition to EGFR amplification we observed a high degree of complex genomic changes involving chromosome 7 (Fig. [3b](#Fig3){ref-type="fig"}). All cases within this cohort had gains encompassing at least 39 % of chromosome 7. Complex changes included normal diploid segments, long stretches of single or double copy gains, and isolated regions of amplification not limited to EGFR. It is unclear whether EGFR amplification is a result of overall destabilization of the entire chromosome 7, or rather the initiating cause of it. Seven cases (88 %) contained homozygous deletions of CDKN2A/B (Fig. [3c](#Fig3){ref-type="fig"}), which is commonly associated with EGFR amplification in GBM \[[@CR6], [@CR7]\]. The last case (EGFR-07) had a normal diploid count along the 9p arm, including CDNK2A/B. The remainder of chromosome 9 presented a mosaic gain, indicating a relative loss of 9p (CDNK2A/B), although not to the same extent as the other cases. EGFR-07 was one of 2 cases that also contained a complete deletion of the RB1 gene locus. Disruption of the Rb pathway is a primary component of glioma development and can occur through loss of RB1 or CDKN2A/B expression or through amplification of CDK4 \[[@CR8], [@CR9], [@CR7]\]. It is possible that, in this case, deletion of the RB1 gene precluded the need for CDKN2A/B deletion. In addition to CDNK2A/B, the methylthioadenosine phosphorylase (MTAP) gene also falls within the deleted region, although in EGFR-06 only a hemizygous loss is observed at the MTAP position as compared to a very focal homozygous deletion of the CDNK2A/B genes. Although MTAP has not been extensively characterized, it is commonly deleted in a number of cancers and there is some evidence that is has tumor suppressive properties \[[@CR10]--[@CR12]\]. The third recurrent CNC observed across this cohort was a chromosome 10 deletion (Fig. [3d](#Fig3){ref-type="fig"}), observed in all eight cases (100 %). One case, EGFR-04, presented copy-neutral LOH, apparently due to a hemizygous loss of chromosome 10 followed by genome duplication. Deletion of 10q is commonly reported in GBM \[[@CR13]--[@CR16]\], and may be a mechanism for the inactivation of the tumor suppressor phosphatase and tensin homolog (PTEN), which has been associated with GBM \[[@CR17]\]. LOH resulting from 10q deletion may result in haploinsufficient expression, or exposure of inactivating PTEN mutations within the retained allele. However, such a large scale loss may indicate that additional tumor suppressor genes across chromosome 10 may also play a role in retaining cell-cycle equilibrium. Two putative tumor suppressor genes potentially impacted by chromosome 10 LOH are ANXA7 and PFKFB3. ANXA7 acts as a positive regulator of EGFR, and haploinsufficiency of ANXA7 reportedly results in stabilized EGFR protein \[[@CR18]\]. Fleischer and colleagues showed that LOH of the PFKFB3 gene results in the reduction of UB12K4 expression, a growth-inhibiting splice variant of PFKFB3. They concluded that this shift in UB12K4 expression tends toward more aggressive tumor growth \[[@CR19]\]. The final recurrent CNC observed in this cohort was a deletion overlapping the RB1 gene in 4 of 8 (50 %) cases. RB1 gene inactivation has been identified in numerous cancers \[[@CR20]\] either through inactivating mutations or RB1 deletions (reviewed in \[[@CR7]\]). Two cases (EGFR-02 & EGFR-06) contained partial deletions of RB1, whereas EGFR-07 & EGFR-10 had deletions spanning the entire gene locus. As mentioned above, EGFR-07 was the only case in which the CDKN2A/B locus was not homozygously deleted, which may be in part a result of the RB1 deletion---causing a prior disruption in the Rb pathway which is often accomplished by the loss of the CDKN2A/B gene locus. CDKN2A expression has therapeutic implications, which in turn may be blocked by the loss of RB1. For example, deletion of CDKN2A is a marker of increased sensitivity to the CDK4/6 inhibitor PD0332991 in melanoma. However, RB1 deletion leads to PD0332991 resistance, counteracting the CDKN2A-imbued sensitivity \[[@CR21]\]. Testing in GBM cell lines and GBM xenografts have shown similar results, indicating that PD0332991 could be effective in treating gliomas, which typically have CDKN2A/B deletions. Similarly, PD0332991 resistance was exhibited by cell lines with RB1 deletion, and also by cell lines lacking the expression of functional Rb1 protein \[[@CR22]\]. Conclusions {#Sec5} =========== Microarray enables us to gain a broader picture of the CNCs occurring in glioma cases. Gliomas with 1p19q co-deletion have a relatively quiet genome, apart from the obvious co-deletion. Additional CNCs are observed in isolated cases, but appear to occur primarily as larger aberrations---such as complete loss or gain of entire chromosomes. On the other hand, EGFR amplified cases tend to be more complex and have specific abnormalities associated with the EGFR amplification. Furthermore, 1p19q co-deletions and EGFR amplification associated CNCs appear to be mutually exclusive. Methods {#Sec6} ======= The samples used in this study were de-identified and the study was conducted under the University of Utah Institutional Review Board protocol 7275. Seventeen cases were selected retrospectively from confirmed glioma diagnoses based on positive FISH results for either 1p19q co-deletions (n = 9) or EGFR amplification (n = 8). All 17 cases were processed on the OncoScan array designed for copy number analysis of FFPE samples. Briefly, the OncoScan platform uses MIP probes---linear probes containing 2 genomic homology regions separated by a linker DNA. The probe forms an inverted loop and binds the gDNA with the homology regions next to each other flanking a SNP. A dinucleotide complementary to the SNP is incorporated into the probe resulting in circularization of the MIP probe, then the gDNA template and unutilized probe are removed via exonuclease degradation of linear DNA. The probe is cleaved between universal primer sequences encoded in the linker region of DNA, which allows for subsequent amplification of the probe. In the context of the OncoScan array, the linker region also encodes a unique molecular tag corresponding to the genomic region encoded in the probe. This unique tag is then separated from the probe and used to interrogate the array. The raw data was processed using OncoScan® Console (Affymetrix, Santa Clara, CA) and all cases were analyzed with Nexus Express software (BioDiscovery, Hawthorne, CA) to assess CNCs across the genome and identify consistent patterns across each class of gliomas. All calls made by the software were manually reviewed for probe performance. FFPE : Formalin-fixed, paraffin-embedded; SNP: Single nucleotide polymorphism; GBM: Glioblastoma CN : Copy number CNC : Copy number change LOH : Loss of heterozygosity CN-LOH : Copy neutral-loss of heterozygosity FISH : Fluorescent in situ hybridization WHO : World Health Organization **Competing interests** The authors declare that they have no competing interests. **Authors' contributions** All authors were instrumental in developing the concept and design of this study, and in reviewing and approving this manuscript. CNP drafted the manuscript and pre-reviewed all of the data that was acquired with the help of LRR. STS reviewed all of the data for clinical significance. We would like to thank Michelle Layton for extraction and preparation of the samples. This study represents a portion of an evaluation of the OncoScan product manufactured by Affymetrix. Affymetrix provided arrays and some reagents for this study. Other resources required for this study were provided by the ARUP Institute for Clinical and Experimental Pathology®.

Background {#Sec1} ========== Fabry disease (FD) (OMIM: 301500) is a rare X-linked hereditary disease (incidence 1:117.000 live births) caused by deficiency of the lysosomal alpha-galactosidase A enzyme (αGalA; E.C. 3.2.1.22) \[[@CR1]\]. Human αGal A is a homodimeric glycoprotein with a molecular weight of 46 kDa. GLA gene is located on Xq22.1, spans approximately 13 kb of genomic DNA and contains seven exons \[[@CR2], [@CR3]\]. As a consequence of the αGal A deficiency, neutral glycosphingolipids, mainly globotriaosylceramide (Gb3), accumulate in a variety of cells and tissues, leading to a wide clinical spectrum of clinical manifestations \[[@CR4]\]. Evolving knowledge suggests that FD has a wide spectrum of heterogeneous phenotypes, which ranges from the classical phenotype with kidney, cardiovascular and cerebrovascular disease \[[@CR5]\], to a seemingly asymptomatic disease, occasionally observed in females \[[@CR6]\]. To date, over 600 mutations are described in the human gene mutation database (<http://www.hgmd.org>) as causative of FD. Moreover, most mutations are "private", occurring only in single pedigrees, while mutations at CpG dinucleotides have been identified in unrelated families \[[@CR7]\]. Correlation between genotype and residual enzyme activity is not strong, depending on the mutation and additional genetic and non-genetic factors. Furthermore, very few "de novo" mutations have been detected with a frequency still unclear \[[@CR8]\]. Here we report a case of non-diagnosed FD patient until the age of 35 without a familial history of the disease; we performed both clinical and molecular diagnosis identifying a de novo mutation. Case report {#Sec2} =========== The patient, a white 35 years old man, had been healthy until the age of 31 years, except for mild articular pain and hypohidrosis since childhood. His father died of pulmonary enphysema at the age of 54 years, and had no renal or cardiovascular disease. His mother, 60 years old, has hypertension and a past history of ischemic heart disease. The patient has two brothers aged 38 and 33 years and a sister aged 27 years, all perfectly healthy. At the age of 31 years, he was hospitalized for proteinuria in nephrotic range (6 g/24 h), hypertension and a reduced creatinine clearance rate (eGFR 57 mL/min). He underwent renal biopsy, with a diagnosis of nephroangiosclerosis. The biopsy material contained 20 glomeruli: global glomerular sclerosis, due to ischemic damage, was demonstrated in fourteen glomeruli, and the remaining six glomeruli showed only minor abnormalities; arterial and arteriolar hyalinosis were also present. Signs of tubular atrophy, interstitial fibrosis, and interlobular sclerosis were noted. Immunofluorescence examination did not show deposition of IgA, IgG, IgM, C1q, C3, k, and lambda fragments. Electron microscopy was not performed. At that time, the echocardiography showed concentric left ventricular hypertrophy (LVH) (ejection fraction 64 %) and an abdominal ultrasound showed no abnormalities. The patient was discharged with diagnosis of hypertension-induced nephroangiosclerosis and started treatment with ACE-inhibitors. At the age of 35, because of the fast progression of renal disease and the need to start dialysis, the patient was referred to our hospital. On admission, his blood pressure was 120/85 mm Hg, pulse was 72 beats/min and body temperature was 36 °C. Physical examination of the head, face, abdomen and extremities showed no abnormalities; moreover, there were no abnormal neurological findings. Full blood count was normal. The values of blood biochemical parameters were urea 210 mg/dL, creatinine 9.4 mg/dL, total protein 8.0 g/dL and albumin 4.3 g/dL. Electrolytes and the result of liver function tests were normal. Twenty-four-hour protein excretion was 4.3 g, and his creatinine clearance was 9.3 mL/min. On admission, we also repeated abdominal ultrasound, which showed bilateral increased echogenicity, with a decrease in mean cortical thickness and loss of cortico-medullary definition. The echocardiography confirmed the concentric LVH with a normal ejection fraction (63 %). Cardiac magnetic resonance imaging (MRI) was subsequently performed to better evaluate the cardiac involvement, that showed a prolonged myocardial T2 relaxation time at the level of the inferolateral wall, secondary to myocardial fibrosis. On the basis of these evidences and of a more accurate anamnesis, typical symptoms of FD---neuropathic pain, hypohidrosis, LVH, proteinuria and CKD---were identified, although there was no apparent family history of FD. Therefore, extensive tests were subsequently undertaken: evaluation of levels of α-Gal A in the plasma and molecular analysis were performed. Despite the finding of Malta crosses in the urine is a very common finding in Fabry disease, a urinalysis was not performed. The patient had low levels of α-Gal A (0.5 nmol 4 MU/mL/h; normal range 8--19 nmol 4 MU/mL/h); sequence analysis of the GLA gene (Fig. [1](#Fig1){ref-type="fig"}) showed a single nucleotide point mutation in hemizygosis at nucleotide c.901 C\>T in exon 6 (p.Arg301X). We extended the genetic analysis to all family members of the patient (mother, sister and brothers) and none of them had any alteration in the GLA gene, suggesting a de novo mutation in the patient.Fig. 1Molecular characterization of the c.901 C\>T mutation in GLA gene. Sequence analysis of GLA exon 6 in our patient. The *top* electropherogram is from patient's mother (**a**), the *lower* shows the patient's mutation (**b**). The *vertical bar* indicates the position of the mutation Discussion {#Sec3} ========== FD is an X-linked disease in which mutations of the *GLA* gene result in a deficiency of the enzyme α-galactosidase A and subsequent progressive, intralysosomal deposition of undegraded glycosphingolipid products, primarily Gb3, in multiple organs \[[@CR1], [@CR2]\]. The initial signs and symptoms of FD emerge during childhood and adolescence; however, because these signs and symptoms are not specific of FD, these patients are frequently misdiagnosed, and the correct diagnosis may be delayed. Major organ involvement typically occurs between the age of 20 and 30 years and result in a significantly decreased quality of life for both heterozygous females and hemizygous males, with lifepan typically reduced by approximately 15--20 years. For the effective management of FD, an early diagnosis is required, and an early start of treatment remains essential in order to reduce the progression of the disease \[[@CR2]\]. Due to its rarity and heterogeneity of the symptoms, FD clinical diagnosis remains a diagnostic challenge. In addition, the diagnosis is even more complex when family history is absent. Here, we report a case of late diagnosis of FD in a patient of 35 years with progressive renal disease. The late diagnosis is mostly due to the absence of family history and of ultrastructural analysis of the renal biopsy. However, the clinical history and the finding of the characteristic cardiac abnormalities suggested the presence of FD, confirmed by the low α-Gal levels \[[@CR3], [@CR4]\] and the molecular analysis. In fact, the molecular diagnosis, performed by sequencing the seven exons, the exon--intron boundaries and the promoter region of the GLA gene in the patient, revealed in the patient the c.901 C\>T (p.R301X) mutation. This mutation, identified as a pathogenic mutation in 1995, inserts a stop codon in the aGalA protein and is reported to be responsible for the classical phenotype \[[@CR5]--[@CR9]\]. Moreover, it involves CpG dinucleotides, which are hot-spots for mutation \[[@CR7]\]. The study of the relatives of the patient showed, surprisingly, that none of his family members (mother, sister and brother) was affected by the disease; all of them showed normal enzymatic activity and the wild-type GLA gene. This indicated that the disease was caused by a de novo mutation, arisen spontaneously in this male patient. To confirm these findings, the analysis of the genetic profile of the patient and his relatives was carried out, proving that they had a family relationship. To date, more than 600 mutations have been identified in human GLA gene that are responsible for FD, including missense and nonsense mutations, small and large deletions (Human Gene Mutation Database, <http://www.hgmd.org>). Such mutations are usually inherited and cases of de novo onset, i.e. arisen spontaneously, occur rarely \[[@CR10]\]. De novo mutation may be due to maternal germline mosaicism or spontaneous mutation \[[@CR8], [@CR10]\] and are very difficult to diagnose. In the literature, only few cases of de novo mutations in FD have been described \[[@CR8]--[@CR11]\], for the difficulty to recognize them. In one, the mutation c.493 G\>C in the third exon of the GLA gene was detected in a 44-year old male patient with classical clinical manifestations of the disease and negative family history and in his 11-year old daughter \[[@CR8]\]; another case described a patient with a de novo diagnosis of point mutation (R301X) in the GLA gene \[[@CR9]\]. Other two studies described the cases of a 32-year old female and a 41-year old male, in which the de novo point mutation at position 691 of exon 5 and G373D, respectively, were detected \[[@CR10], [@CR11]\]. Conclusions {#Sec4} =========== Concluding, a strategy for timely FD diagnosis is recommended and genetic testing would be performed in suspicious patients even in absence of a typical family history to make an early diagnosis. Consent status {#Sec5} ============== Written informed consent was obtained from the patient for publication of this case report and any accompanying images. FD : Fabry disease αGalA : α galactosidase A Gb3 : globotriaosylceramide eGFR : estimated glomerular filtration rate LVH : left ventricular hypertrophy MRI : magnetic resonance imaging AP and ER took in care the patient; CDD, EN and AD carried out the genetic analysis, ER and EN have been involved in the draft of the manuscript; AP and AD revised critically the manuscript; FS has given the final approval of the version to be published. All authors read and approved the final manuscript. Competing interests {#FPar2} =================== The authors declare that they have no competing interests.

All relevant data are within the manuscript and its Supporting Information files. Introduction {#sec001} ============ People are warned of the effects of changing weather on TV and through online messages every day. Every day, some people report their health deterioration or some new symptoms they associate with changes in weather---such as temperature, humidity, cloudiness or atmospheric pressure changes. This usually applies to women and older or middle-aged people, although it may also apply to children \[[@pone.0232725.ref001]\]. It is estimated that approximately 55% of the German population and 69% of the Canadian population over the age of 60 years seem to be weather sensitive---that is, they consider the weather affects their health \[[@pone.0232725.ref002]\]. Behavioral data suggest environmental factors such as climate or weather may impact human emotional states and health conditions. For example, it was demonstrated that emotional distress decreases when exposure to the sun increases \[[@pone.0232725.ref003]\] and during day-to-day weather variations (solar exposition, wind speed and air pressure influences self-reported life satisfaction \[[@pone.0232725.ref004]\] and low and high temperature levels, precipitation, humidity and cloud cover worsens the expression of feelings among social media users \[[@pone.0232725.ref005]\]. It was also demonstrated that environmental stressors---such as considerable increase in temperature or increased precipitation produced by climate change, especially during natural disasters---may cause mental, emotional and bodily stress among humans \[[@pone.0232725.ref006], [@pone.0232725.ref007]\]. Weather conditions are also related to seasonal affective disorder, which is predominantly found in women \[[@pone.0232725.ref008]\], and panic anxiety attacks \[[@pone.0232725.ref009]\]. Some authors also pointed out the relationship between weather changes and cluster headache \[[@pone.0232725.ref010]\] (Lee et al. 2014), migraine \[[@pone.0232725.ref011]\], neuropathic pain \[[@pone.0232725.ref012]\] and human sleep disturbance \[[@pone.0232725.ref013]\]. Any disorders resulting from the impact of climatic conditions can be described as meteoropathy \[[@pone.0232725.ref001]\]---that is, a phenomenon of worsening existing diseases or the emergence of a new specific disease as a consequence of climate change \[[@pone.0232725.ref014]\]. Meteoropathy includes "a group of symptoms and pathological reactions in response to gradual or sudden changes in meteorological factors in a specific area interacting, presumably, through natural electromagnetic influences covering a wide range of frequencies and amplitudes" \[[@pone.0232725.ref001], p. 46\]. In general, people vary in their sensitivity to weather changes---although the most vulnerable to meteoropathy are women, middle-aged people and anxious and depressed individuals \[[@pone.0232725.ref001]\]. It is worth noting that women are usually indicated as more sensitive to weather conditions, as well as more easily susceptible to developing symptoms of physical and mental disorders in response to environmental changes \[[@pone.0232725.ref015], [@pone.0232725.ref016], [@pone.0232725.ref007]\]. Meteorosensitivity is the key concept here. Meteorosensitivity may be defined as "biological susceptibility to feel the effect of particular atmospherical events on \[the\] mind and body" \[[@pone.0232725.ref017], p. 103\]. Meteorosensitivity is also related to the human psychophysical capabilities linked with everyday stress management. Although the biological mechanisms of meteoropathy are unknown, in a recent study, it was suggested that superior vestibular nucleus activity may affect meteoropathy regulation in mice and probably in humans \[[@pone.0232725.ref018]\]. Furthermore, the activities of the hypothalamus and amygdala nucleus are suggested as significant in meteoropathy development \[[@pone.0232725.ref019]\]. The authors suggested electromagnetic waves directly affect the hypothalamus. Therefore, they indirectly enhance the secretion of the stress hormone, the adrenocorticotropic hormone, and decrease endorphin secretion---which may lead to increased anxiety, headache and other meteoropathic symptoms. The authors assumed both human psychophysical instability and susceptibility to weather changes causing stress through changes in the brain regulation of emotions becomes a factor that increases human susceptibility to somatic and mental disorders. From a psychological point of view, cerebral mechanisms of emotion regulation and autonomic nervous system (ANS) functioning are useful for understanding individual differences in meteorosensitivity levels and meteoropathy symptom development. Meteorological factors and other stressors may cause a challenge to the human physiological balance. When the balance is difficult to maintain, the development of somatic symptom disorder may be observed \[[@pone.0232725.ref020]\]. Taking account that many meteoropathy symptoms---such as lability of mood, depression, anxiety or an indefinite feeling of uneasiness---are related to mood disorder symptoms \[[@pone.0232725.ref017]\], it is worth considering the role of affective temperaments to explain the differences in meteoropathy among women. Affective temperaments (depressive, cyclothymic, hyperthymic, irritable and anxious) refer to a genetically determined, stable-across-lifespan and trait-related manifestations playing a fundamental role in the predisposition to mood disorders---such as depression, anxiety and bipolar disorder \[[@pone.0232725.ref021], [@pone.0232725.ref022], [@pone.0232725.ref023]\]. Several studies have showed women have higher levels of the depressive, cyclothymic and anxious temperaments than men \[[@pone.0232725.ref024], [@pone.0232725.ref025], [@pone.0232725.ref026]\]. These results are in line with the observation that mood disorders are approximately twice as prevalent in women compared to men \[[@pone.0232725.ref027]\]. Molecular genetics research has provided interesting data on the strong association between central serotonergic (the depressive, cyclothymic, irritable and anxious temperaments) and dopaminergic (the hyperthymic temperament) regulation in affective temperament development \[[@pone.0232725.ref028]\]. It is worth noting that several authors have pointed out the role of serotonin in human mood regulation \[[@pone.0232725.ref029], [@pone.0232725.ref030], [@pone.0232725.ref031]\]. Moreover, serotonin is involved in the hypothalamic--pituitary--adrenal axis reaction to acute and chronic stress, which contributes to the onset of anxiety and depression \[[@pone.0232725.ref032]\]. Affective temperaments may share a common biological disposition with various other psychiatric or somatic symptoms, including meteoropathic disturbances. For example, the cyclothymic/irritable temperament appears to be associated with elevated stress reactivity in daily life \[[@pone.0232725.ref033]\] and somatic symptoms with no organic explanation \[[@pone.0232725.ref034], [@pone.0232725.ref035]\]. In an earlier study, a strong association between the cyclothymic temperament and hypertension was demonstrated \[[@pone.0232725.ref036]\]. Certain affective temperaments, mainly the cyclothymic and anxious temperaments, have also been associated with sleep disturbances present among meteoropathic symptoms \[[@pone.0232725.ref037], [@pone.0232725.ref038]\]. In addition, the anxious temperament may be related to meteoropathy symptoms---such as anxiety, irritability, hypervigilance, inability to relax, insecurity, tension and gastrointestinal distress \[[@pone.0232725.ref028]\]. The main goal of this study was to assess the relationship between affective temperaments and meteoropathy among women and examine meteorosensitivity as a mediator in this relationship. We hypothesized (a) the depressive, cyclothymic, irritable and anxious temperaments are positively correlated with meteoropathy, while the hyperthymic temperament is negatively correlated with meteoropathy and (b) the meteorosensitivity dimension mediates the relationship between affective temperaments and meteoropathy. Materials and methods {#sec002} ===================== Participants {#sec003} ------------ The study sample consisted of 450 white women gathered from the general population via the University of Warsaw online recruitment platform. The participants' ages ranged from 18 to 70 years (M = 30.01; SD = 9.10). Among the participants, 303 had received higher education, 146 had received secondary education and 1 had received primary education. Fifty-nine women lived in rural areas, 65 in small towns and 326 in large cities. This was an anonymous cross-sectional study, participation was voluntary and the participants did not receive any compensation. The data were collected via online self-report questionnaires. All participants provided informed consent to participate in the study after reading the rules of the study. Information about the purpose and procedure of the study was displayed on a computer screen. Starting the questionnaires was tantamount to agreeing to participate in the study on the given terms. The research project including aforementioned procedure and all questionnaires used was approved by the local Research Ethics Commission at the Faculty of Psychology, University of Warsaw (ref: 5-03-2019). Measures {#sec004} -------- Affective temperaments were assessed using the Polish version of the Temperament Evaluation of Memphis, Pisa, Paris and San Diego Autoquestionnaire (TEMPS-A) \[[@pone.0232725.ref021], [@pone.0232725.ref039], [@pone.0232725.ref040]\]. TEMPS-A is a self-report instrument that is composed of 110 items (109 for men) with a yes/no response format. TEMPS-A is composed of five temperament scales (Cronbach's alphas for the current sample are given in parentheses): depressive (α = .74), cyclothymic (α = .83), hyperthymic (α = .81), irritable (α = .84) and anxious (α = .88). For each answer, a "yes" was scored as 1 and a "no" as 0. These scores were added and then divided by the number of items that belonged to each affective temperament scale (21 for each of the first four temperament categories and 26 for the anxious temperament). Meteorosensitivity and meteoropathy were assessed using the Polish adaptation of the METEO-Q questionnaire \[[@pone.0232725.ref017]\] (Polish adaptation by Włodzimierz Oniszczenko). The questionnaire consists of 11 items that measure meteorosensitivity (5 items) and meteoropathy (6 items) and a structured checklist aiming to identify the 21 physical and psychological symptoms related to climate variations (see [Table 3](#pone.0232725.t003){ref-type="table"}). All items are rated on a 4-point Likert response scale ranging from 0 (absent) to 3 (severe). Cronbach's alphas in the current sample are given in parentheses: meteorosensitivity scale (α = .79) and meteoropathy scale (α = .79). Statistical analysis {#sec005} -------------------- The statistical analysis was performed using IBM SPSS Statistics 25 \[[@pone.0232725.ref041]\]. Descriptive statistics, such as mean and standard deviation, of the main variables were recorded. Data normality was checked based on the skewness and kurtosis values (ranged from −1.5 to 1.5), following the application of criteria specified by Tabachnick and Fidell \[[@pone.0232725.ref042]\]. Relationships among variables were examined with Pearson product--moment coefficients. An absolute value of *r* of .1 was classified as small, .3 as medium and .5 as large per Cohen \[[@pone.0232725.ref043]\]. The mediation analyses were conducted using the PROCESS Model 4 macro for SPSS v. 3.3 \[[@pone.0232725.ref044]\]. In addition, the bootstrapping procedure with 5,000 sample draws and bias-corrected standard errors was used to estimate the direct and indirect effects \[[@pone.0232725.ref045]\]. Results {#sec006} ======= [Table 1](#pone.0232725.t001){ref-type="table"} provides descriptive statistics, as well as skewness and kurtosis values, for meteorosensitivity, meteoropathy and affective temperaments for the whole sample. The skewness and kurtosis analyses show that all the variables were normally distributed. 10.1371/journal.pone.0232725.t001 ###### Descriptive statistics for the meteorosensitivity and meteoropathy and Temperament Evaluation of Memphis, Pisa, Paris, and San Diego Autoquestionnaire (n = 450). {#pone.0232725.t001g} Variable Range Mean Standard deviation Skewness Kurtosis -------------------------- ----------- ------ -------------------- ---------- ---------- *METEO-Q* Meteorosensitivity 0--15 9.49 3.11 -.54 .18 Meteoropathy 0--18 8.81 3.74 -.07 -.23 *Affective temperaments* Depressive .00--.95 .44 .18 .26 -.52 Cyclothymic .00--.95 .42 .22 .20 -.69 Hyperthymic .00--1.00 .44 .21 .10 -.73 Irritable .00--.95 .31 .21 .53 -.39 Anxious .00--1.00 .42 .23 .26 -.58 [Table 2](#pone.0232725.t002){ref-type="table"} presents the correlation coefficients between age, meteorosensitivity, meteoropathy and affective temperaments in the studied sample. A small positive correlation was found between age and the hyperthymic temperament. Furthermore, small negative correlations were found between age and the cyclothymic, irritable and anxious temperaments. A large positive correlation was found between meteorosensitivity and meteoropathy. Medium positive correlations were found between meteorosensitivity and the cyclothymic and anxious temperaments, as well as between meteoropathy and the cyclothymic and anxious temperaments. Both the meteorosensitivity and meteoropathy scales positively correlated with the depressive and irritable temperaments (all the correlation coefficients were small). No correlation was found between mateorosensitivity and meteoropathy and the hyperthymic temperament. 10.1371/journal.pone.0232725.t002 ###### Pearson r correlations between age and meteorosensitivity and meteoropathy and Temperament Evaluation of Memphis, Pisa, Paris and San Diego Autoquestionnaire scales (n = 450). {#pone.0232725.t002g} 2\. 3\. 4\. 5\. 6\. 7\. 8\. ------------------------ ----- ---------------------------------------------- ---------------------------------------------- ----------------------------------------------- ----------------------------------------------- ----------------------------------------------- ----------------------------------------------- 1\. Age .03 .08 -.07 -.19[\*\*\*](#t002fn002){ref-type="table-fn"} .11[\*\*](#t002fn001){ref-type="table-fn"} -.15[\*\*](#t002fn001){ref-type="table-fn"} -.12[\*\*](#t002fn001){ref-type="table-fn"} 2\. Meteorosensitivity .70[\*\*\*](#t002fn002){ref-type="table-fn"} .17[\*\*\*](#t002fn002){ref-type="table-fn"} .33[\*\*\*](#t002fn002){ref-type="table-fn"} .02 .26[\*\*\*](#t002fn002){ref-type="table-fn"} .31[\*\*\*](#t002fn002){ref-type="table-fn"} 3\. Meteoropathy .18[\*\*\*](#t002fn002){ref-type="table-fn"} 36[\*\*\*](#t002fn002){ref-type="table-fn"} .05 .27[\*\*\*](#t002fn002){ref-type="table-fn"} 36[\*\*\*](#t002fn002){ref-type="table-fn"} 4\. Depressive .56[\*\*\*](#t002fn002){ref-type="table-fn"} -.60[\*\*\*](#t002fn002){ref-type="table-fn"} .44[\*\*\*](#t002fn002){ref-type="table-fn"} .67[\*\*\*](#t002fn002){ref-type="table-fn"} 5\. Cyclothymic -.21[\*\*\*](#t002fn002){ref-type="table-fn"} .67[\*\*\*](#t002fn002){ref-type="table-fn"} .61[\*\*\*](#t002fn002){ref-type="table-fn"} 6\. Hyperthymic -.20[\*\*\*](#t002fn002){ref-type="table-fn"} -.43[\*\*\*](#t002fn002){ref-type="table-fn"} 7\. Irritable .50[\*\*\*](#t002fn002){ref-type="table-fn"} 8\. Anxious \- \*\* p \< .01; \*\*\* p \< .001. Two separate analyses were performed using a bootstrapping procedure, with meteorosensitivity as the mediator between the cyclothymic and anxious temperaments and meteoropathy. The first analysis indicated a significant indirect effect of the cyclothymic temperament on meteoropathy through meteorosensitivity \[effect = 3.65, SE = .53, 95% CI = (2.64, 4.73)\]. The second analysis found a significant indirect effect of anxious temperament on meteoropathy through meteorosensitivity \[effect = 3.25, SE = .50, 95% CI = (2.30, 4.26)\]. The individual pathways in the mediation analysis are displayed in [Fig 1](#pone.0232725.g001){ref-type="fig"}. {#pone.0232725.g001} [Table 3](#pone.0232725.t003){ref-type="table"} presents means and standard deviations for each meteoropathy symptoms provided by the METEO-Q questionnaire. As observed in [Table 3](#pone.0232725.t003){ref-type="table"}, the most severe meteoropathy symptoms in the studied sample were asthenia, an indefinite feeling of malaise and irritability (75th percentile). 10.1371/journal.pone.0232725.t003 ###### Means and standard deviations for each meteoropathy symptom measured by the METEO-Q (n = 450). {#pone.0232725.t003g} Range M (SD) ------------------------------------------------- ------- ------------- Lability of mood 0--3 1.51 (.81) Extreme reactivity to external events 0--3 1.37 (.93) Depression 0--3 .94 (.98) Anxiety 0--3 .78 (.91) Asthenia 0--3 1.98 (.85) Anhedonia 0--3 1.33 (.92) Irritability 0--3 1.81 (.90) Indefinite feeling of malaise 0--3 1.93 (.86) Vague pain, articular pain, muscular pain 0--3 1.20 (1.09) Vertigos 0--3 .83 (.94) Headache 0--3 1.54 (1.05) Nausea 0--3 .42 (.73) Alterations of the cardiac rhythm (tachycardia) 0--3 .63 (.88) Difficulties with concentration 0--3 1.46 (.93) Insomnia 0--3 1.00 (1.03) Excessive sleepiness 0--3 1.72 (1.00) Lack of appetite 0--3 .55 (.78) Excessive appetite 0--3 .87 (.98) Digestion's dysfunctions 0--3 .70 (.91) Alterations of sexuality 0--3 .86 (.99) Weakness during working activities 0--3 1.60 (.90) Discussion {#sec007} ========== The results of the present research may extend the knowledge of the functional significance of affective temperaments and indicate the importance of affective temperaments for women's meteoropathy symptoms. As expected, four affective temperaments (depressive, cyclothymic, irritable and anxious) were positively correlated with meteoropathy among women, although the correlation coefficients were small to medium. The medium correlations between the cyclothymic and anxious temperaments and meteoropathy were revealed. These results indicate the cyclothymic and anxious temperaments may be associated with some meteoropathy symptoms similar to mood disorder symptoms. It is worth noting that the most severe symptoms of meteoropathy in the studied sample were asthenia, an indefinite feeling of malaise and irritability. Meteorosensivity was strongly correlated with meteoropathy and served as a mediator between affective temperaments and meteoropathy among the studied women. This result is not surprising because the level of sensitivity to climate or weather changes (i.e., meteorosensitivity) seems to be the most important factor triggering the symptoms of meteoropathy. Therefore, one of the most important questions concerns the link between affective temperaments and meteorosensitivity and meteoropathy. The analysis of mediation shows the cyclothymic and anxious temperaments affect meteoropathy both directly and indirectly through meteorosensitivity as a mediator. The lack of sufficient data on the biological basis of both affective temperaments and meteoropathy does not allow the formulation of a clear hypothesis about the common biological mechanism of these phenomena. Nevertheless, taking account of the hypothesis on the biological basis of affective temperaments, as well as the hypothetical biological foundations of meteorosensitivity, it can be assumed that some biological associations between these variables are related to brain activity and ANS functioning. The serotonergic regulation of the cyclothymic and anxious temperaments may connect affective temperaments with the so-called somatic awareness \[[@pone.0232725.ref026]\]. It is a physical discomfort that is medically unexplained, is associated with reduced levels of serotonin \[[@pone.0232725.ref046]\] and present in meteoropathy as, for example, headaches, fatigue or difficulty sleeping. Serotonin, involved in the functioning of the hypothalamic--pituitary--adrenal axis, can contribute to the development of two important meteoropathy components: anxiety and depression \[[@pone.0232725.ref032]\]. In a recent work, Di Nicola et al. \[[@pone.0232725.ref047]\] indicated that meteorological factors such as sunlight or atmospheric pressure are associated with changes in the level of serotonin in the brain, which---in addition to dysfunction in the dopaminergic system---may be important for the process of emotion dysregulation. In turn, Mazza et al. \[[@pone.0232725.ref048]\] suggested the increase in sympathetic nervous system tension resulting from a deficiency of serotonin metabolism in patients with fibromyalgia results in fatigue and weakness, sleep problems, headache and migraine, anxiety or depression and mood disorders. It is worth noting that these symptoms coincide with those that are present in meteoropathy. The probable impact of cerebral mechanisms of emotion regulation and ANS functioning on meteorosensitivity and meteoroptahy is also visible in affective temperaments traits, such as an intense experience of all emotions or rapid changes in mood and energy typical for the cyclothymic temperament \[[@pone.0232725.ref021]\] or tension and gastrointestinal distress typical for the anxious temperament \[[@pone.0232725.ref039]\]. The ANS may be of particular importance for meteorosensitivity and the development of meteoropathy associated with the affective temperaments. The ANS functions penetrate the functioning of many other internal systems and organs, and dysfunctions of this system can manifest as various disorders of the body. It is worth noting, however, that the mechanisms combining affective temperaments with meteorosensitivity and meteoropathy and differentiating women from men can be much more complex and involve sex-dependent neurobiology and genetic, hormonal and immune functions, as well as sex-environment interactions, as suggested by some authors \[[@pone.0232725.ref049], [@pone.0232725.ref050]\]. Several limitations of this study should be acknowledged. This research is cross-sectional in nature, making it impossible to draw any definite conclusions regarding the direction of the relationship between affective temperaments and meteorosensitivity and meteoropathy. Affective temperaments and other personality traits may play a significant role in the development of meteoropathy symptoms, but further longitudinal studies are required to corroborate this hypothesis. Another important limitation of our study is the covariability of factors treated as predictors of meteoropathy symptoms. We only used data based on self-reports, and the presence of comorbid physical and mental health disorders or substance use was not documented. Therefore, the interpretation of our results is restricted. In addition, we observed only women, thus limiting the generalizability of the results. It should be added that the study sample was not too large when considering the survey in the general population. Conclusions {#sec008} =========== Regardless of the methodological limitations of our study, however, our results may contribute to the understanding of how psychological factors influence weather-related changes in women\'s well-being. Our findings highlight the significant overlap between affective temperaments and meteoropathic symptoms among women. High levels of cyclothymic and anxious temperaments may be responsible for an increase in negative physical and mental health symptoms among women. To improve women's well-being, meteorosensitivity and meteoropathy should be taken into account when assessing medically difficult-to-explain symptoms in women. Our results suggest that meteoropathic symptoms may be associated with moods and the dominant type of affective temperament in women. Thus, women who show mood lability and anxiety symptoms should be a focus of medical and psychological care. Supporting information {#sec009} ====================== ###### (SAV) ###### Click here for additional data file. The author thanks Przemysław Święcki from the Faculty of Psychology, University of Warsaw, for his tremendous overall assistance in this study. [^1]: **Competing Interests:**The authors have declared that no competing interests exist.

Structures comprising 1,2-diamino functional groups are recurrent in natural products and synthetic materials and find utility across disparate fields of molecular science that include catalysis, supramolecular chemistry, agrochemical and pharmaceutical research.^[@ref1]^ The prevalence of 1,2-diamine derivatives in high-value targets has inspired considerable effort to devise expedient methods for crafting such compounds from simple precursors.^[@cit1f],[@ref2]^ Among the most direct and general approaches for synthesizing 1,2-diamines is the diamination of olefins.^[@ref3],[@ref4]^ The development of such a process, however, is complicated by several factors, including functional group compatibility with redox reaction conditions and product inhibition of catalyst turnover. In considering basic design criteria, an optimal process for 1,2-diamine assembly would enable differentiation of the two amino groups. Herein, we describe a Rh-catalyzed aziridination/rearrangement method for 1,2-diamine synthesis that offers a solution to these challenges (Figure [1](#fig1){ref-type="fig"}). {#fig1} Previous developments in our lab have provided methods for the facile aziridination of alkenes using dirhodium-based catalysts and hypervalent iodine oxidants.^[@ref5],[@ref6]^ Following this work and related studies on C--H amination with urea, sulfamide, and hydroxylamine-derived sulfamate esters, we recognized the potential of utilizing a bifunctional nitrogen source that could engage in an initial aziridination reaction followed by a simple isomerization to give a diaza-heterocyclic product (Figure [1](#fig1){ref-type="fig"}).^[@ref7]^ Ring opening of the heterocycle would then furnish the desired diamine. Reports describing the rearrangement of *N*-acyl aziridines to the corresponding oxazolines serve as useful precedent for this plan.^[@ref8]^ A series of urea, sulfamate, and sulfamide derivatives were tested in initial exploratory reactions with styrene as a model substrate (Table [1](#tbl1){ref-type="other"}).^[@ref9]^ Reactions with *N*-sulfonylated urea and hydroxylamine-derived sulfamate reagents showed little promise and resulted primarily in decomposition of both the nitrogen source and styrene (entries 1--4). By contrast, alkene aziridination occurred smoothly using *N*-carbamoyl sulfamides (entries 5--7). Experiments with *N*-benzylsulfamide, however, failed to give any of the desired product (entry 8). In the presence of PhI(OAc)~2~, MgO, and 2 mol % Rh~2~(esp)~2~, reactions with H~2~NS(O)~2~NHBoc (entry 6) and the corresponding methylcarbamate derivative (entry 7) performed best and gave the desired aziridine products in yields exceeding 75%.^[@ref10]^ Both aziridine derivatives are stable to isolation by silica gel chromatography (the Troc compound, entry 5, is not) and were used in subsequent experiments to optimize the desired rearrangement to the cyclic sulfamide. The Boc-derivative (entry 6) was ultimately selected given the ease of preparation of this reagent and the convenience of the *tert*-butoxycarbonyl as an amine protecting group. As depicted in Figure [2](#fig2){ref-type="fig"}, the successful rearrangement of the aziridine product furnishes the corresponding cyclic sulfamide in high yield. These heterocycles are broadly useful materials that appear in pharmacologic molecules and serve as precursors to *N*-Boc-1,2-diamines.^[@ref11]^ ###### Styrene Aziridination Using Bifunctional Nitrogen Sources {#fx1} {#GRAPHIC-d146e219-autogenerated} Conversion estimated from integration of the ^1^H NMR spectrum of the reaction mixture. Parentheses denote isolated yields. Product could not be isolated due to its instability on SiO~2~. Tces = 2,2,2-trichloroethoxysulfonyl. {#fig2} Isomerization of aziridine **1** to sulfamide **2** can, in principle, be induced by a Lewis acid catalyst or with an appropriate nucleophilic reagent.^[@ref8]^ While the former failed to give any of the desired cyclic sulfamide, treatment of aziridine **1** with NaI in DMF proceeded at room temperature and furnished the desired product in 74% yield as a single isomer.^[@ref12]^ The efficiency and selectivity of this process and the preceding aziridination step are sufficiently high that isolation of the intermediate aziridine is superfluous. Accordingly, the optimized protocol for conversion of styrene to **2** is performed in a single flask; once the aziridination reaction is complete, a solution of NaI in DMF is added and the desired isomerization ensues. Under these conditions, the product heterocycle is obtained in 73% yield as a white crystalline solid. Cyclic sulfamide formation occurs in modest to high yield with a range of mono- and disubstituted alkenes under conditions that employ limiting amounts of substrate (Table [2](#tbl2){ref-type="other"}). Reactions with styrene-type olefins are most efficient and products are furnished as single isomers, consistent with I^--^ opening of the intermediate aziridine at the benzylic carbon center (entries 1--4). Both acyclic and cyclic aliphatic olefins will engage in this diamination reaction (entries 5--12). In general, product yields from oxidation of terminal alkenes are elevated compared to those obtained when cis- or trans-disubstituted olefins are employed as substrates. The rearrangement reaction in these cases proceeds regioselectively to give the sulfamide products as single isomers. For alkenes that engage less effectively in the aziridination reaction, increasing the number of equivalents of the substrate can result in marked improvement in reaction performance.^[@ref13]^ Aziridination of both 1,1-disubstituted and trisubstituted olefins is also effective under these reaction conditions. Surprisingly, however, these products have proven recalcitrant to isomerization under a variety of conditions tested. ###### Cyclic Sulfamide Formation with Selected Alkenes {#fx2} Reactions were performed in *i*-PrOAc with 1 equiv of substrate, 2 mol % Rh~2~(esp)~2~, 1.1 equiv of BocNHSO~2~NH~2~, 1.1 equiv of PhI(OAc)~2~, and 2.3 equiv of MgO. After the completion of the aziridination reaction, 1.1 equiv of NaI and DMF were added to induce rearrangement. Isolated yields following chromatographic purification. Product isolated as a single diastereomer. dr = 1:1. See ref ([@ref14]). Reaction performed with 10 equiv of substrate. Although cyclic sulfamides generally require forcing conditions of heat and strong acid to induce ring opening, *N*-Boc-protected derivatives are considerably more labile. When these materials are heated in an aqueous solution of pyridine, the SO~2~-moiety is excised to give the mono-Boc diamine (Figure [3](#fig3){ref-type="fig"}).^[@cit9c]^ The product can be isolated following an extractive workup and purified as needed. These conditions have been employed on substrates bearing functional groups that include nitriles, esters, and even pinacolboron. The application of cyclic sulfamides as precursors to 1,2-diamines should enjoy greater use with the availability of this mild method for ring opening. {#fig3} Syntheses of (±)-enduracididine and (±)-allo-enduracididine, amino acid components of the naturally occurring macrocyclic polypeptide antibiotic enduracidin,^[@ref15],[@ref16]^ are made possible from allylglycine by employing our alkene diamination reaction (Scheme [1](#sch1){ref-type="scheme"}). Treatment of **3** with MgBr~2~·OEt~2~ promotes selective cleavage of the *N*-Boc sulfamide.^[@ref17]^ Chromatographic separation of the syn and anti diasteromers is followed by a one-step transformation of **4** to Tces-protected guanidine **6**. Deprotection of this material is achieved in a single operation that involves ester hydrolysis with LiOH and hydrogenolytic Tces-cleavage to furnish (±)-enduracididine. (±)-Allo-enduracididine can be prepared from sulfamide **5** following an analogous sequence. Overall, the preparation of both natural products requires just four steps from commercially available (±)-allylglycine *N-*Boc methyl ester, underscoring the effectiveness of the alkene diamination reaction. {#sch1} An efficient and selective process has been delineated for the synthesis of differentially protected 1,2-diamine derivatives through a two-step sequence involving alkene oxidation/isomerization and sulfamide ring opening. This chemistry is easily performed and applicable for use with a range of structurally disparate mono- and disubstituted alkenes. The effectiveness of such technologies for the construction of C--N bonds is displayed in a concise preparation of the guanidinium natural products, (±)-enduracididine and (±)-allo-enduracididine. Experimental details and characterization data are available in the supporting online material. This material is available free of charge via the Internet at <http://pubs.acs.org>. Supplementary Material ====================== ###### ja506532h_si_001.pdf ^†^ Dr. David E. Olson, Stanley Center for Psychiatric Research, Broad Institute of MIT and Harvard, Cambridge, Massachusetts 02142. ^‡^ Mr. D. Allen Roberts, Institute for Health Metrics and Evaluation, University of Washington, Seattle, Washington 98121. The authors declare no competing financial interest. We thank Arun Thottumkara for helpful discussions. D.E.O. gratefully acknowledges Eli Lilly for a graduate fellowship. This work was supported in part by the National Institute of Health (R01NS045684) and the National Science Foundation under the CCI Center for Selective C--H Functionalization (CHE-1205646).

Introduction ============ Descending aortic aneurysm has an incidence of approximately 3 in 100,000 people per year. The mortality within five years is about 50% for untreated aneurysms \[[@R1]\]. Acute aortic dissection affects 10 to 20 individuals per million in western populations \[[@R2]\], \[[@R3]\]. Treatment of patients with aneurysms of the descending aorta still remains a challenge. Open surgical repair with graft implantation has been the standard therapy for patients with asymptomatic descending aortic aneurysms larger than 5-6.5 cm, while aortic dissections without complications generally have been treated conservatively \[[@R4]\], \[[@R5]\]. Within the last years, new and much less invasive treatment modalities have been developed. In 1991 the first stent graft implantation was reported for a thoracic aortic lesion \[[@R6]\]. Since then there has been a major development in graft configuration and some clinical experiences have been published since the first clinical series were published by the Stanford group in 1994 \[[@R7]\]. We report our midterm results after stent graft repair of descending aortic aneurysms focusing on cerebral and spinal cord ischemia in high risk patients. Patients, material and methods ============================== Between February 2000 and July 2003 a total of 21 patients with lesions of the descending thoracic aorta were treated with endovascular stent grafting in general anaesthesia. Informed consent was obtained from all patients. General inclusion criteria were: adequate vascular access through the iliac arteries and the abdominal aorta, adequate landing zone free of thrombus with minimal calcification, aortic neck diameter \<44 mm, and a length of at least 15 cm. Indication for endovascular surgery in true aneurysms were a diameter \>50 mm and significant risk factors (Table 1 [(Tab. 1)](#T1){ref-type="fig"}) with a high risk for open surgery. In dissections endovascular stenting was indicated for acute dissection in patients with hypertension which were refractory to medical treatment of pain with clinical signs of malperfusion of the lower limbs. Patients in the study underwent preprocedural imaging with spiral computed tomography (CT) using 3-mm slices. In cases of dissection preoperative assessment included visualisation of the false lumen and re-entry sites. Within the 40 months period, 21 (17 men, 4 women; mean age 66.1 years, range 29-90 years) patients with aortic pathologies were evaluated at our institution with 15 true aneurysms and 6 type B-dissections (Stanford classification). The aneurysms had a mean diameter of 6.1 cm (min.-max. range: 3.2-7.1 cm). All of the patients with pathologies of the descending aorta were not suitable for open surgery because of their comorbidities and their clinical classification (ASA mean: 2.7). Patients with ASA score = 3 were defined as high risk patients. All 21 patients were evaluated electively. A team of two vascular surgeons performed the procedures using a Talent^TM^ Stent Graft System (World Medical, a division of Medtronic Vascular, Sunrise, FL, USA) in all cases with an oversize of 4-6 mm. 19 procedures were performed in the operating room (OR) under general anaesthesia and endotracheal intubation, 2 in local anaesthesia. Patients were placed supine and slightly (20-30 degrees) rotated to the right on a fluoroscopic image compatible operating table. Perioperative imaging was performed using a mobile C-arm machine. Intravenous heparin (10,000 IU) was administered as a bolus dose. In all cases we performed a single shot antibiotic prophylaxis by application of a 2^nd^ generation cephalosporine. If it was necessary to cover the left subclavian artery, neither a transposition nor a carotid-subclavian bypass was performed before overstenting the subclavian artery. The decision whether to reconstruct the left subclavian artery was made after overstenting according to the major clinical aspects. In one case a carotid bypass was performed before procedural occlusion of the left common carotid artery. Usually the right femoral artery was exposed by cutdown and controlled proximally and distally. Nearly in all cases a 0.035 inch Terumo guidewire (Terumo^®^, Frankfurt am Main, Germany) was inserted over a sheath and positioned in the ascending aorta. This was followed by the introduction of a 6-F pigtail catheter and an angiography (Figure 1 [(Fig. 1)](#F1){ref-type="fig"}) of the aortic arch and the descending aorta using a high pressure injection-pump with a non-ionic contrast medium (Imeron^®^, Altana Pharma, Germany). After marking the aortic arch with its branches as well as the descending aorta and the landing zone on the screen, a stiff guide wire was inserted and the angiography catheter removed. A delivery sheath (22, 24 or 25 F) was introduced with the punction of the common femoral artery. After pharmacologically controlled hypotension was induced (RR \<80 mmHg), the stent graft was deployed under X-ray guidance (Figure 2 [(Fig. 2)](#F2){ref-type="fig"}). The proximal and distal end of the device were dilated using a catheter to achieve optimal graft contact to the aortic wall. A final arteriogram verified sealing of the aortic lesion. The introducer sheath was removed and the artery sutured. Wound closure completed the procedure. Patients with cardiovascular risk factors (Table 1 [(Tab. 1)](#T1){ref-type="fig"}) were treated with 300 mg aspirin per day. Follow up with clinical examination and CT scans was performed at discharge, 3, 6, and 12 months after and then annually. Aortic size, perfusion of the false lumen, endogenous leakages and graft migration were evaluated. Results ======= For the aneurysms, the procedures took 97.8 minutes (average; range was 55-145 minutes). Mean blood loss was 163.3 ml (50-350 ml), mean contrast agent dosage 56.7 ml (45-70 ml). In one of the cases we found an early distal endoleak after the procedure. This patient with peripheral leakage remained without surgical intervention, and the leak was occluded after three months. The patients were discharged from the ward after 10.1 days in average. Within 30 days 2 patients died: one of myocardial infarction, and one of severe pneumonia. In one case the left common carotid artery had to be bypassed using a carotid-carotid graft before stenting the aorta. In 3 patients we had to overstent the left subclavian artery, thereof in two cases a subclavian steal syndrome with significant neurological deficit (markedly reduced vigilance, no peripheral neurological deficits) developed. After excluding cerebral haemorrhage using CT we had to perform a carotid-subclavian bypass in the two, and a subclavian transposition in the third case on the first postoperative day. After revascularisation both patients became asymptomatic immediately. For the B-dissections, the procedures took 75 minutes (average; range was 55-95 minutes). Mean blood loss was 141.7 ml (50-300 ml), mean contrast agent dosage 55 ml (45-70 ml). The patients were discharged from the ward after 10.7 days in average. 21 of 21 procedures were primarily technically completely successful without conversion to open surgery. Within the follow-up period (mean 25.4 months) another two patients (one with an aneurysm, the other with a B-dissection) died of myocardial infarction. There was no rupture of aneurysm or dislocation of stent graft during the observation period. The postprocedural data are summarized in Table 2 [(Tab. 2)](#T2){ref-type="fig"}. Discussion ========== Catheter-associated techniques have changed the face of modern vascular surgery. The treatment of aneurysms of the descending aorta using endovascular stents is one of the more recent advances in treatment, still receiving increasing attention as alternative to open surgery. Several reports have shown the feasibility of stent graft treatment of aneurysms of the descending thoracic aorta \[[@R8]\], \[[@R9]\], \[[@R10]\]. 30 day mortality ranges from 0%-20% in different reports \[[@R8]\], \[[@R9]\], \[[@R11]\], \[[@R12]\], \[[@R13]\] which is encouraging although long term results are unknown \[[@R14]\], \[[@R15]\]. For patients with high risk from concomitant diseases, for patients with previous thoracic operations, and for those with multiple traumas the endoluminal technique offers an alternative to open surgery although endoluminal surgery is accompanied by complications as well as open surgery \[[@R8]\], \[[@R9]\], \[[@R16]\], \[[@R17]\]. The often discussed 6 cm diameter threshold reflects the surgeon\'s reluctance to open repair because of serious risks of morbidity and mortality. In our opinion a diameter \>5 cm, or twice the size of an adjacent nonaneurysmal aortic segment seems to be more reasonable \[[@R18]\]. Our average procedure time for aneurysms and B-dissections together was 91.3 minutes (range 55-145 minutes), and correlates well with data of other groups \[[@R18]\], \[[@R19]\], \[[@R20]\], whereas much longer operating times are described in literature \[[@R21]\]. In our investigation the cumulative 30-day-mortality-rate was 9.5% and correlates well with comparable patient groups \[[@R21]\], \[[@R19]\], \[[@R20]\]. However, all deaths of 4 out of 21 patients were not associated with endovascular treatment: Two patients died of myocardial infarction and pneumonia within 30 days after stenting. Total mortality during follow-up (mean 25.4 months) was 19% (4/21). The 4.8% (1/21) incidence of endoleaks in our study is in the lower range of comparable studies \[[@R11]\], \[[@R22]\], \[[@R21]\], \[[@R19]\]. In the case with a small distal endoleak we waited 3 months for spontaneous sealing, and then performed a second procedure, overstenting the leak successfully. The deployment of endovascular stent grafts in the thoracic aorta may cause devastating neurological complications. In our patients we saw no stroke as described from other groups who explain those complications by severe hypovolaemia in those patients treated emergently or by alterations of the aorta by the guidewire during its passage through the arch and consecutive cerebral embolism \[[@R21]\]. We overstented the left common carotid artery after bypassing the vessel by a carotid-carotid bridge in one case, and occluded the left subclavian artery in 3 out of 21 cases (14.3%) and found a postoperative neurological deficit in 2 patients i.e. 9.5%. In those two patients whose blood pressure had been extremely stable during the operative procedure we found a psychological alteration which we thought had its origin in the occluded left subclavian artery. A consecutive subclavian steal syndrome could be documented using duplex sonography. The brain scintigrams were recorded to verify the supposed steal phenomenon in the patients suffering from a temporary organic brain syndrome. So we decided to perform a transposition in one and a carotid-subclavian bypass in the other patient on the first postoperative day, and both patients were asymptomatic indeed immediately after revascularisation. This experience correlates well with the data of other investigators \[[@R18]\]. Another dreaded complication, paraplegia, results from spinal cord ischemia secondary to intercostal artery occlusion, and is described in about 7% of the cases \[[@R9]\], \[[@R11]\], \[[@R12]\], \[[@R21]\]. In our patients no paraplegia occurred. As described by others \[[@R21]\] we routinely increase the mean blood pressure to about 90 mm Hg for 24 hours immediately after exclusion of the aneurysm. A drainage of the cerebrospinal fluid as described in literature \[[@R21]\] was performed in none of our patients because the benefit in respect to paraplegia is not yet definitively proven \[[@R23]\], \[[@R24]\], \[[@R25]\], \[[@R26]\]. Patients with acute or chronic type-B-dissection represent a cohort in whom the indication for stent graft treatment is still debated. In our study we treated 6 patients with acute dissection using endografts. Indications for endovascular treatment of B-dissections were: a progressive dissection, malperfusion of organs, therapy-refractory pain and/or hypertension. In all cases the entry site was closed. In two of our type B dissection patients (33.3%) four months after treatment the false lumen was still patent but the patients did not have any symptoms. The aorta of both patients did not show progressive dilatation. Because of a long dissection in both cases with perfusion of the celiac trunk, the superior mesenteric artery and the left renal artery through the false lumen, we decided not to cover the re-entry site but to treat it conservatively in order to avoid paraplegia after occlusion of some pairs of intercostal arteries \[[@R27]\], \[[@R28]\] as well as mesenteric ischemia. In the remaining cases the false lumen was obliterated within the first follow up examination period, and the patients did well. Dissections proximal to the endografts have not been observed during the follow-up. Conclusions =========== In summary our results demonstrate that endovascular treatment of the descending aorta is feasible with acceptable early complication rates and acceptable mean term results even in hight risk patients. Procedural overstenting of the subclavian artery requires subclavian revascularisation in the minority of cases. In selected patients endovascular treatment may be the most suitable therapeutic option. {#T1} {#T2} {#F1} {#F2}

\section{Introduction} Despite major advancements in contemporary neuroscience research, the mechanisms that underpin the pathophysiology of major psychiatric disorders such as schizophrenia remain elusive. A major challenge revolves around the development of an integrated framework, which can bridge the scales where the questions are asked and the answers are required: the microscopic level (molecular genetics, neurophysiology and neuropharmacology of the neuron) and the macroscopic level where one observes and measures changes in the behavioral and cognitive functions combined with alterations in the neurophysiology and anatomical-functional imaging of the whole brain \cite{Hig07, Cann15,Fat09,Fris16,Kub07,Ruiz13,Kel18}. Between these two scales, it extends the mesoscopic scale, where one attempts to describe the activation and communication of specific neuronal networks located in different brain areas. The coordinated activity of these networks gives rise to emerging/macroscopically-experimentally observed behavior and corresponding global brain activation patterns. The mesoscopic level of description has been the focus of many studies on the electrophysiology of rodents and primates using single neuron, and more recently multi-neuron recording techniques, combined with histology and more recently optogenetics \cite{Jen15,Krav12}. \par One way to shed light on the mesoscopic organization and functionality is through the use of detailed biophysical neuronal network models. In general, the biophysical modelling approach is described by the following steps (see Fig.~\ref{fig:experiment}): a) identification of a specific behavioral/cognitive function that is deviant in schizophrenia, b) construction of a detailed biophysical neuronal model based on the knowledge acquired from detailed experiments and/or from animal studies including the identification of the neuronal networks at specific brain areas that govern the corresponding behavioral/cognitive function, c) use of the biophysical network model to study structural and functional connectivity changes as predicted by existing neuropharmacological and sub-neuronal function hypotheses, d) test how these changes have an impact on the predicted changes in the cognitive function, and finally, e) validation of the results of the neuronal network model using experimental data to test whether the presence of the predicted changes of the network model result in the same cognitive/behavioral deviance as observed e.g. in schizophrenia.\par At the microscopic level, the dopamine hypothesis is a prominent theoretical framework that relates the mechanism of the antispychotic medications to psychopathology in schizophrenia \cite{Cann15, How09, Bris14}. The hypothesis suggests that hyper-dopaminergia or hyperactivity at D2-dopamine receptors in subcortical limbic system structures leads to the appearance of positive symptoms of the disorder (delusions, hallucinations) and that the hypo-dopaminergia or hypoactivity of D1-dopamine receptors in the prefrontal cortex leads to the appearance of negative symptoms and cognitive dysfunctions \cite{How09, Bris14}. Another prominent hypothesis, the glutamate hypothesis is linked with the functional modulation of N-methyl-D-aspartate (NMDA) receptors \cite{Fat09, Cann15, Lew09, Gon12, Sny13}. NMDA receptors which are located in cortical pyramidal neurons are activated by glutamate, the main excitatory neurotransmitter in the mammalian brain. Activation of pyramidal neurons in the deep cortical layer results in the activation of GABAergic inhibitory interneurons in the cortex forming a local functional network \cite{Lew09, Cann15, Sny13}. The glutamate hypothesis states that a hypoactivity at NMDA receptors results in an imbalance of this excitation-inhibition network. Substances that are NMDA antagonists, such as phencyclidine (PCP) and ketamine can produce both positive symptoms and cognitive dysfunctions in healthy individuals that resemble schizophrenia, as well as worsening of such symptoms in schizophrenic patients \cite{Laht14, Sny13}. Furthermore, postmortem studies in patients have shown significant changes in NMDA receptor protein expression in the prefrontal cortex \cite{Sny13, KRIS07}.\par Remarkably, these two phenomenological independent hypotheses might share the same pathophysiological pathway in the prefrontal cortex. The D1 receptor signaling in cortical neurons is regulated by the action of NMDA receptors and vice versa \cite{Scott02, Che04, How09}. Activation of D1 receptors in the prefrontal cortical pyramidal neurons by the agonist SKF81297 lead to an increase of the steady-state NMDA evoked current \cite{Che04} and this effect is canceled by incubation of neurons in $[\text{Ca}^{2+}]$ free medium. Inversely, activation of NMDA receptors by glutamate results in the recruitment of D1 receptors in cortical neurons while they don't have any effect on the distribution of D2 receptors \cite{Scott02}. The dysconnection hypothesis describes the above synaptic complex: it suggests that psychosis in terms of aberrant neuromodulation of synaptic efficacy mediates the influence of intrinsic and extrinsic (long-range) connectivity \cite{Fris16}. It suggests that the key pathophysiology lies in the interactions between NMDA receptor functions, the modulatory effect of neurotransmittion and the mediated changes in synaptic efficacy. The NMDAR-mediated plasticity affects the functionality at the level of neuronal circuits which in turn leads to an abnormal functional integration (dysconnection) of cognitive functions among brain regions in schizophrenia.\par Here, the cognitive function that we study is the inhibition of the response in healthy controls and patients with schizophrenia while performing an antisaccade task. In the particular task, subjects are instructed to look at the opposite direction of a visually presented stimulus, thus exerting inhibitory control over the natural tendency to look towards the visual stimulus \cite{Hall78}. There is a large body of literature that confirms a deficit in this task in patients with schizophrenia compared to healthy controls \cite{Pan19}. More specifically, it has been reported that patients produce more erroneous responses towards the visual target than healthy controls while the response latency for correct antisaccades for these patients is longer and more variable than that observed for healthy controls \cite{Bas08, Pan19}. A large body of animal and human studies has confirmed the critical role of the prefrontal cortical areas such as the frontal eye field, the supplementary eye field and especially the dorsolateral prefrontal cortex in the performance of the antisaccade task \cite{Mun04, Muri05, Brown07}. There is also evidence that deficits in the performance of the antisaccade task in schizophrenia may be related with hypo-activity in specific areas of the prefrontal cortex that are activated during this task such as the dorsolateral prefrontal cortex \cite{Mun04, Coe17,Pan19}, (see also Fig.~\ref{fig:experiment}).\par Here, building up on previous work \cite{Dur03,Cut07a, Cut07b, Kahr08}, we attempt to bridge the microscopic and macroscopic levels by addressing a multiscale neuronal network model of two brain regions, namely the prefrontal cortex (PFC) and the superior colliculus (SC) for the simulation of the antisaccade task. The model incorporates competing neuronal activity at the intermediate layer of the SC that is driven both by a planned input (the antisaccade command) and a reactive input (the erroneous prosaccade command). The output of the model are the reaction time distributions for the correct antisaccades and erroneous prosaccades, as well as the percentage of the error prosaccades. The PFC activity is approximated by leaky spiking integrate and fire pyramidal neuronal dynamics as described in \cite{Dur03}, while the connectivity of the PFC neurons is allowed to vary from a ring-like network to a completely random topology. Unlike previous modeling attempts \cite{Cut07a, Cut07b,Cut14}, we did not assume any a-priori knowledge of the biophysical parameter values and the PFC neuronal connectivity structure; these are estimated in a strict numerical way, by ``wrapping" around the detailed network model a numerical optimization structure, namely the Levenberg-Marquardt algorithm \cite{web:lm} to fit the experimental behavioral data. Then, we used the calibrated model to bridge microscopic and macroscopic scales, namely to study how changes in the dopamine levels acting on D1 receptors and the related hypo-activity of NMDA receptors on simulated prefrontal cortical neurons (microscopic level) give rise to antisaccade performance differences in patients with schizophrenia vs. healthy controls. In addition, the fact that the detailed connectivity of the biophysical neuronal network model of the PFC was also allowed to vary, gave us the opportunity to (qualitatively) study how changes at the microscopic level can give rise to changes at the mesoscopic level (namely how the topology of the specific network changes) that in turn give rise to macroscopic changes at the experimentally observed behavioral performance. Our hypothesis is that decreased DA levels acting on D1 receptors associated with hypofunction of NMDA receptors in the PFC would predict antisaccade performance differences as observed in schizophrenia, namely a large increase in the error rate as well as an increase in the mean and variance of the reaction times for correct antisaccades. Furthermore, in accordance to the dysconnection hypothesis, we predicted that the deviances in the microscopic level would mediate the macroscopic effects via a change at the mesoscopic level, namely a change in the functional connectivity of the underlying neuronal network. \begin{figure \centering \includegraphics[width=1.\linewidth]{figure1fin.eps} \caption{Schematic of the Experiment and different levels of inquiry ranging from the microscopic level: molecular genetics, neurophysiology and neuropharmacology of the neuron, to the mesoscopic level: activation and communication of specific networks of connected neurons located in different areas of the brain, for example of the PFC network (green triangles and red elipses are for pyramidal excitatory neurons and interneurons respectively), and finally to the macroscopic level: changes in the behavior and cognitive functions combined with changes in the neurophysiology and anatomical-functional imaging of the whole brain. Behavioral experiment: antisaccade task. Starting with a fixated cue centered on a screen, a peripheral stimulus (S) appears, the subject is instructed to look at the opposite direction (antisaccade) suppressing the saccadic motion (error prosaccade) exerting an inhibitory control over the natural tendency to look towards the visual stimulus.} \label{fig:experiment} \end{figure} \section{Material and Methods} \subsection*{The Prefrontal Cortex network model} We build up on a model proposed in \cite{Dur03} using spiking integrate and fire neurons whose dynamics are governed by the current balance equation: \begin{equation} C\frac{dV_i}{dt}=-I_\text{{LEAK}}-I_\text{{AHP}}-I_\text{ADP}-I_\text{NMDA}-I_\text{GABA}+I_\text{INJ}, \end{equation} where $C$ is the membrane capacity, $V_i$ is the membrane potential of the $i-$th neuron, $I_\text{LEAK}$ $I_\text{INJ}$, $I_\text{AHP}$ and $I_\text{ADP}$ are the leak, injected, after hyper polarizing and depolarizing currents respectively; $I_\text{NMDA}, I_\text{GABA}$ correspond to the excitatory and inhibitory synaptic currents, respectively.\\ If $V_i>V_\text{thres}$, a spike is being fired and the potential returns to $V_i=E_\text{LEAK}$. The leak current is given by: \begin{equation} I_\text{LEAK}=g_\text{LEAK}(V_i-E_\text{LEAK}), \end{equation} where $g_\text{LEAK}$ is the corresponding conductance. The after hyperpolarizing current is given by: \begin{equation} I_\text{AHP}=g_\text{AHP}e^{-(t-t_{sp})/\tau_\text{AHP}}(V_i-E_\text{AHP}), \end{equation} where $t_\text{sp}$ is the time of the last spiking and $\tau_\text{AHP}$ reflects the decay time scale of the current; $g_{AHP}$ is the conductance and $E_\text{AHP}$ is the reverse potential.\\ The after depolarizing current, $I_\text{ADP}$ is due to the $ [\text{Ca}^{+2}]_i$ influx in pyramidal neurons and is modeled according to the following equations \cite{Dur03}: \begin{equation} \begin{cases} I_\text{ADP}&=g_\text{ADP}m_{i}(V_i-E_\text{ADP}) \\ \frac{dm_i}{dt}&=\frac{(m_{\infty}-m_i)}{\tau_\text{ADP}}\\\ [\text{Ca}^{+2}]&=A_\text{{Ca}}\sum_{t_{sp}}e^{-(t-t_{sp})/t_{off}}-e^{-(t-t_{sp})/t_{on}}, \end{cases} \end{equation} where $m$ is the gating activation variable related to $[\text{Ca}^{+2}]$ influx, as: \begin{equation} m_{\infty}= (1+exp(\gamma(\theta-[\text{Ca}^{+2}]_i))^{-1}. \end{equation} $t_{\text{on}}, t_{\text{off}}$ are constant time scales marking the onset and offset of the $[\text{Ca}^{+2}]$ influx. $E_\text{{ADP}}$ is the reverse potential, $g_{\text{ADP}}$ is the conductance. The parameters $\gamma$ and $\theta$ control the sharpness of the equilibrium function $m_{\infty}$, while $A_{\text{Ca}}$ controls the calcium concentration resulting from the spike sequence \cite{Dur03}. In our model, the neurons are connected through a Watts and Strogatz (WS) small-world topology \cite{Wat98, Bull09, Stam07}. The network is reflected on the synaptic currents defined by the activation variable $s$ which for each single neuron is given by \cite{Lai02,Erm12,Com00}: \begin{equation} \frac{ds_i}{dt}=\alpha(1-s)H(V_i-\theta_{0})-\beta s_i, \end{equation} where $H(V)$ is the step function.\\ If $V_i<\theta_{0} \implies H(V_i-\theta_{0})=0 $ and the synaptic equation takes the form: \begin{equation} \frac{ds_i}{dt} = - \beta s_i \Rightarrow s_i = c{e^{ - \beta t}} \Rightarrow s\to 0, t \to +\infty. \end{equation} Hence, the synapse turns off with $\beta$ as time scale.\\ Instead, if $V_i>\theta_{0} \implies H(V_i-\theta_{0})=1 $ and the synaptic equation takes the form: \begin{equation} \frac{ds_i}{dt}= \alpha \left( {1 - s_i} \right) - \beta s_i. \end{equation} The above has the solution \begin{equation} s_i(t) = c\frac{{{e^{ - (\alpha + \beta )t}}}}{{\alpha + \beta }} + \frac{\alpha }{{\alpha + \beta }}. \end{equation} The above expression implies that synaptic jumps are given by $\frac{\alpha{}}{\alpha+\beta}$.\\ Usually, fast synapses have $\alpha ,\beta = \text{O}\left( 1 \right)$, while for slow synapses $\alpha=\text{O}\left(1 \right)$ and $\beta = \text{O}\left(\epsilon \right)$ meaning that fast synapses activate and deactivate on fast time scales, while the slow synapses activate on fast and deactivate on slow time scales \cite{Lai02,Erm12}. The form of the excitatory and inhibitory synaptic currents are given, respectively, by: \begin{equation} I_{\text{NMDA}}=g_{\text{NMDA}}(V_i-E_{\text{NMDA}})\sum{s_{j}}, \end{equation} for $E_{\text{NMDA}}=0$, and \begin{equation} I_{\text{GABA}}=g_{\text{GABA}}(V_i-E_{\text{GABA}})\sum{s_{j}}, \end{equation} for $E_{GABA}=-80$. The summation is taken over the presynaptic neurons.\\ Excitatory currents result from pyramidal neurons, while inhibitory currents result from inter-neurons. Out of the $N$ neurons, $\frac{3N}{4}$ are pyramidal neurons, while the rest are inter-neurons. Fig.\ref{fig:Simulations}(a) shows the dynamics of two neurons in the PFC network. The neurons although start with nearby initial conditions they exhibit slightly different firing rates (see Fig.\ref{fig:Simulations}(b)). \begin{figure \begin{center} \begin{picture}(400,240) \includegraphics[width=12.5cm]{fig2paper.eps \put(-320,230){\textbf{(a)}} \put(-140,230){\textbf{(c)}} \put(-320,110){\textbf{(b)}} \end{picture} \end{center} \caption{Simulation of neurons in PFC and SC. \textbf{(a)} Membrane potentials for two neurons in the PFC network. One of these neurons fires with higher frequency (black line). \textbf{(b)} The corresponding linear rising phase (see sec. \ref{sec:SC}) for the previous PFC neurons. A neuron fires with slightly higher frequency (black line) giving a bigger slope in the linear frequency-time approximation. \textbf{(c)} Firing rates in SC of buildup and burst neurons from a simulation run of the antisaccade task. Buildup neurons (black line and blue lines) encoding reactive and planned saccade tasks. When the activity of buildup neuron crosses a threshold (400HZ red dash line) an erroneous prosaccade (light green) is initiated followed by a correct saccade.} \label{fig:Simulations} \end{figure} \subsection*{The Superior Colliculus (Tectum) model} \label{sec:SC} The neuronal model is a classic on-center off-surround leaky competitive integrator \cite{Tra01,Tay99,Ara94,Kop95}. The internal state which represents the firing rate $x_i(t)$ of the $i-$th node is governed by the following differential equation: \begin{equation} \tau \frac{dx_i}{dt}=-x_i+\sum_{j}{A_j w_{ij}+I_p+I_r-u_0+I_n}, \end{equation} where $\textbf{w}=(w_{ij}), i, j=1,2,...N$ is the synaptic efficacy from neuron $i$ to neuron $j$, $A$ is the activity function of node $j$, $Ir$ and $Ip$ are the reactive and planned inputs, respectively, that the Tectum receives from other cortical areas, $u_{0}$ is a global inhibition term, and $I_{n}$ is the background noise.\par The value of $u_{0}$ is set to zero for the buildup neurons and to a large value for the burst neurons as burst neurons are shown to have a discharge activity only after the activity of the buildup neurons reaches a certain threshold \cite{Mun95}. The activity of the burst neurons is restored back to zero only when these neurons surpass an activity level equal to 80\% of their theoretical maximum discharge rate \cite{Tra01,Cut07a}. The activity function $A_j(t)$ of a neuron $j, j=1, 2,..N$ representing the average membrane potential is given by a sigmoid function \begin{equation} A_i(t)=\frac{1}{1+e^{-bx_i+\theta}}, \end{equation} where $b$ is the steepness and $\theta$ is the offset of the sigmoid function. The interaction matrix $w$ allows for lateral interactions between neurons in the same colliculus and between neurons located in opposite colliculi sites \cite{Mun98,Mer98}; it depends only on the spatial distance and it is positive (excitatory) i.e. $w_{ij}>0$ for short distances, and negative(inhibitory) i.e. $w_{ij}<0$ for long distancing neurons. Thus, the interaction matrix is described by \begin{equation} w_{ij}=a\frac{e^{-(j-i)^2}}{2\sigma_a^2}-b\frac{e^{-(j-i)^2}}{2\sigma_b^2}-c, \end{equation} where $a$, $b$, and $c$ are free parameters and $\sigma_a$ and $\sigma_b$ are spatial parameters.\par Two competing input signals are integrated in the Tectum: a planned and a reactive. In the model, the origins of these two input signals differ: the reactive signal is considered to originate from the Post Parietal Cortex (PPC), whereas the planned signal from the frontal executive centers of the brain (the PFC). The reactive input signal is governed by a simple differential equation reading: \begin{equation} \begin{cases} \frac{dI_r}{dt}=A\mid slope_r\mid, if \ t \geq t_{on}+t_r^{delay} \ and \ I_r\leq I_r^{max} \\ \frac{dI_r}{dt}=-a_rA \cdot I_r, if \ t \geq t_{on}+t_r^{delay} and \ I_r> I_r^{max} \\ \frac{dI_r}{dt}=-a_r A \cdot I_r, else\\ \end{cases} \end{equation} $a_r$ are the integration strengths, $I_r^{max}$ is a theoretical maximum allowed activity for the reactive input, $t_{on}$ is a constant indicating the onset of the incoming signal $I_r$. $A$ is the strength of the rising phase and $slope_r$ represents the linear rising phase of the reactive input and follows a bell like probability distribution function (normal with mean value $\overline{x}=5.27$ and standard deviation $std=2$) \cite{Cut07a,Cut07b}. The reactive input reflects the sensory information reaching the Tectum without extensive information processing and it is taken to follow closely the onset of a visual stimulus in the periphery with a delay, $t_r^{delay}$. \par The planned input signal has also a linear rising phase \cite{Red00} before it reaches its theoretical maximum value and is governed by: \begin{equation} \begin{cases} I_p=A\mid slope_p\mid, if \ t \geq t_{on}+t_p^{delay} \ and \ I_p\leq I_p^{max} \\ I_p=A \cdot I_p^{max}, if \ t \geq t_{on}+t_p^{delay} and \ I_r> I_p^{max} \\ I_p=0, else. \\ \end{cases} \end{equation} $I_{p}^{max}$ is the theoretical maximum activity of the planned input, $A$ is the strength of the planned input and $slope_p$ is the slope of the planned input rising phase and it is extracted from the neural integration (time evolution) of the PFC according to the following way: for each PFC neuron, we perform a linear approximation of the corresponding firing rate (which is called linear rising phase and it is calculated in the first 70 msec). The mean value of these slopes defines $slope_p$. The planned input reflects the processing of the planned antisaccade by higher processing centers such as the frontal eye fields (FEF), the supplementary eye fields (SEF), and the dorsolateral prefrontal cortex (DLPFC) to determine the behavioral response that would be appropriate for the given task instruction and it is considered to take longer, i.e. $ t_p^{delay}>t_r^{delay}$ for processing than the reactive input due to additional cortical processing. \par The strength of both planned and reactive signals is given by: \begin{equation} A(i,j)=A_0 e^{\frac{-(j-i)^2}{2\sigma_A^2}}, \end{equation} where $i$ and $j$ are the indices of nodes and $\sigma_A$ is the standard deviation of the Gaussian. The width of the Gaussian function was derived from the shape of movement fields of saccade related neurons in the monkey SC \cite{Mun95, Tra01}. In the simulations, the threshold level (horizontal line at about 400 Hz) was appropriately adjusted, so that both buildup neurons encoding the reactive and planned inputs cross the threshold and an erroneous prosaccade (error burst) was initiated followed by a correct saccade (correct burst). We considered a correct antisaccade response, if the movement of the eyes went in the opposite direction of the stimulus; all other cases were considered to be error prosaccades. Saccade reaction times were estimated to be the time interval from the onset of peripheral stimulus until the time the activity of the burst neurons deviated from zero plus 20 ms (approximate time required for burst neuron signal to reach the eye muscles) \cite{Cut07a,Cut07b}. Fig.\ref{fig:Simulations}(c) shows the firing rates in SC of two buildup and two burst neurons from a simulation run of the antisaccade task. \section{Experimental Data Analysis} \label{sec:Experimental} \subsection*{Experimental Data} The experimental data used in this study are derived from a previous study of the group \cite{Smy09}. Briefly the antisaccade performance data of 30 male patients (age span 18-30 years) with DSM-IV schizophrenia formed the patient group. Patients were evaluated at the Psychosis Unit of the Psychiatry Department of the National and Kapodistrian University of Athens at Eginition Hospital and the diagnosis of schizophrenia was confirmed by a trained psychiatrist with the use of the Mini International Neuropyschiatric Interview (M.I.N.I., version 5.0.0., DSM IV) \cite{She97}. Exclusion criteria consisted of the following: neurological disorder (epilepsy, multiple sclerosis etc), mental retardation and drug abuse within the last year before evaluation. All patient participants were receiving antipsychotic medication and were in a stable phase of the disorder during testing. The control sample was derived from the ASPIS (Athens Study for Psychosis proneness and Incidence of Schizophrenia) data base \cite{Smy02,Smy03,Evd02}. For the purposes of ASPIS oculomotor task data (smooth eye-pursuit, saccade, antisaccade, visual fixation) were collected from a population of 2120 conscripts of the Greek Air Force aged 18-25 years. Valid antisaccade data were obtained from 2006 individuals.\par Oculomotor tasks were performed in a set up that has been described in detail in our previous studies \cite{Smy02,Smy03,Evd02}. The antisaccade task was preceded by a calibration procedure, with saccades at 5 and 10 degrees to the left and to the right of a central fixation point. Each antisaccade trial started with the appearance of a central fixation stimulus (white cross 0.3 x 0.3 degrees of visual angle). After a variable period of 1-2 sec, the central stimulus disappeared and a peripheral stimulus (same white cross) appeared randomly at one of the 9 prescribed distances (2-10 degrees at 1 degree intervals) either to the left or to the right of the central fixation stimulus. The subjects were instructed to make an eye movement to the opposite direction from that of the peripheral stimulus as quickly as possible. Each subject performed 90 trials. \par An interactive PC program (created using the TestPoint CEC) was used for the detection and measurement of saccades from the eye movement record \cite{Smy02,Smy03,Evd02}. We excluded trials with artifacts (blinks, etc.) in the period extending from 100ms before the appearance of the peripheral target to the end of the first saccade as well as trials for which an eye movement occurred within the 100-ms period before the appearance of the peripheral target. In addition, in order to avoid including predictive movements or too slow responses, we excluded trials with reaction times that were not within the window of 80-600ms. Using the above criteria, the minimum number of valid trials retained for each subject was 30 and the maximum was 90. \begin{figure}[t] \begin{center} \begin{picture}(400,190) \includegraphics[width=12.5cm]{figexpeall.eps \put(-300,170){\textbf{(a)}} \put(-180,170){\textbf{(b)}} \put(-75,170){\textbf{(c)}} \end{picture} \end{center} \caption{Healthy controls. Distributions of reaction times (RT) during the experiment (see sec.~\ \ref{sec:Experimental}). \textbf{(a)} Distribution of RT which includes correct antisaccades and wrong prosaccades; the main statistical moments are found to be: mean value $\bar{t}_{RT}=253$, standard deviation $s=60.83$. \textbf{(b)} Distribution of antisccade only RT: mean value $\bar{t}_{RT}=270.84$ and standard deviation $s=55.53$. \textbf{(c)} Distribution of error prosaccases RT: mean value $\bar{t}_{RT}=208.94$ and standard deviation $s=42.59$.} \label{fig:CON_EXP} \end{figure} \begin{figure}[t] \begin{center} \begin{picture}(400,200) \includegraphics[width=12.5cm]{disttotSZ.eps \put(-300,170){\textbf{(a)}} \put(-180,170){\textbf{(b)}} \put(-75,170){\textbf{(c)}} \end{picture} \end{center} \caption{Distributions of reactions times (RT) of schizophrenic patients during the experiment (see sec.\ \ref{sec:Experimental}). \textbf{(a)} Distribution of RT which includes correct antisaccades and wrong prosaccades: mean value $\bar{t}_{RT}=265$, standard deviation $s=81$. \textbf{(b)} Distribution of antisccade only RT: $\bar{t}_{RT}=304$, standard deviation $s=77$. \textbf{(c)} Distribution of error prosaccases RT: mean value $\bar{t}_{RT}=215$, standard deviation $s=60$.} \label{fig:SZ_exp} \end{figure} \subsection*{Data Analysis} The error rate was computed for all patients and healthy controls as the percentage of erroneous prosaccades towards the peripheral target over all valid trials. The mean value and standard deviation for each group was then computed and compared with the error rate of the simulation data using the t-test at a 0.05 level of significance. \par The reaction times for valid trials for each subject were used to derive the cumulative reaction time distribution for each subject. Then, the reaction times for correct antisaccades and erroneous prosaccades were processed separately for each subject to derive the corresponding cumulative reaction time (RT) distributions. For all RT distributions, we also calculated the percentiles (from 5\% to 95\% with a step of 5\%). The percentiles were then averaged across the group to give the average group percentiles that are plotted in the average cumulative distribution. Ratcliff \cite{Ratc79} showed that this average distribution retains the basic shape characteristics of the individual distributions. Thus, we created experimental averages of the RT distributions for (a) all RT, (b) correct antisaccade RT only, and (c) error prosaccade RT. Fig.~\ref{fig:CON_EXP} depicts the resulting experimental RT distributions for the healthy controls, while Fig.~\ref{fig:SZ_exp} illustrates the RT distributions for the patients. The correct antisaccade cumulative RT distributions for each group were then compared with the corresponding average cumulative RT distributions derived from the model simulation using the Kolmogoroff-Smirnov test at a 0.05 significance level. \subsection*{Numerical Optimization} Our PFC neuronal network consisted of $N_1 = 75$ pyramidal neurons and $N_2 = 25$ inter-neurons. The SC neuronal network had $N_3 = 31$ neurons (15 buildup neurons, 15 burst neurons and 1 fixation neuron). Ensembles of the PFC network topology were constructed using the WS algorithm \cite{Wat98, Stam07, Bull09}, thus allowing the connectivity to vary between ring and random regular structures; the connectivity topology is adjusted by the switching probability $p$. In line with biological theories, variations in the receptors functionality (DA, NMDA) as well as alterations in the PFC network topology are reflected in the conductances of the synaptic currents $I_{\text{NMDA}}, I_{\text{GABA}}$ as well as in the ionic currents e.g. calcium-activated potassium $I_{\text{AHP}}$ \cite{Malenka86}. Here, we investigated these alterations in the conductances of currents $I_{\text{NMDA}}$ and $I_{\text{AHP}}$ by extracting the conductance $g_{\text{AHP}}$ from a normal distribution with mean value $r_1$ and standard deviation $r_2$, while the $g_{\text{NMDA}}$ was our third parameter $g_{\text{NMDA}} = r_3$ to be tuned.\par For each set of the model parameters $\textbf{r} = (r_1, r_2, r_3, p)$, the dynamic model produces a different PFC activity (defining a distribution of trials), thus resulting to different values of the $slope_p$. The resulting distributed values of $slope_p$ over all trials are then fed into the SC model in order to produce the reaction times for correct antisaccades as well error prosaccades RT. The number of simulation trials was set to $N = 100$. \par The values of the model parameters $r_i ,i=1,2,3$, as well as the switching probability $p$ for the construction of the network topology were estimated numerically by minimizing the residual of the objective function defined as the norm difference of the distribution functions $f(\cdot)$ of the RT (both antsaccade and error prosaccade) between experimental and those resulting from simulations, i.e.: \begin{equation} \argmin_{r\in \mathbb{R}_{+}^{4}} obj(\textbf{r}):= \left\lbrace r: min|| {f(x_i,\textbf{r})}_{model}- \ {f(x_i)}_{Data} ||,\ \textbf{r} \in \mathbb{R}_{+}^{4} \right \rbrace \end{equation} where the norm is the Euclidean distance between the $n$ points of approximate and experimental RT distributions functions. The procedure was implemented separately in the two groups of healthy controls and patients. The minimization problem was solved using the matlab implementation of the Levenberg-Marquardt algorithm (function nlqnonlin \cite{matlab})\cite{web:lm}. The step size tolerance was set to $\text{tol}(X)=0.001$ and the function tolerance was set to tolF=0.01. The maximum number of iterations was set to 100. \section*{Results} \subsection*{Healthy controls} The numerical optimization for the healthy controls resulted to the following values of the model parameters: $r_1 = 2.135$ (95\% \text{CI}: 1.986, 2.283), $r_2 = 0.631$ (95\% \text{CI}: 0.591, 0.672) ($g_{\text{AHP}}$) and $r_3 = 1.33$ (95\% \text{CI}: 1.197, 1.47), ($g_{\text{NMDA}}$), while for the network's topology, the switching probability was found to be $p=0.01 (95\% \text{CI}: 0.007, 0.0145)$, see also tab.~\ref{tab:norm}. The residual of the objective function was 0.0022. For the optimal parameter values, the model predicted an error rate of 34\% for the control group which was not significantly different from the experimentally observed one at a 0.05 significance level. Fig.\ref{fig:control_exp_opt} depicts the distribution function of reaction times for the experimentally observed and those derived by simulations. \begin{figure}[t] \begin{center} \begin{picture}(400,190) \includegraphics[width=12.5cm]{res1cont.eps \put(-300,170){\textbf{(a)}} \put(-180,170){\textbf{(b)}} \put(-75,170){\textbf{(c)}} \end{picture} \end{center} \caption{The computed distributions of reactions times (RT) for healthy controls (brown color), using the value parameters estimated by numerical optimization. The experimental distributions are also depicted for comparison purposes. \textbf{(a)} Distribution of all RT (correct antisaccades and error prosaccades): mean value $\bar{t}_{RT}=249.38$, standard deviation $s=57.14$ ($\bar{t}_{RT}=253.6$ and $s=60$ are the corresponding experimental values). \textbf{(b)} Distribution of RT only for the correct antisaccades: mean value $\bar{t}_{RT}= 269.48$, standard deviation $s=51.72$ ($\bar{t}_{RT}=270$ and $s=55$ are the corresponding experimental values).\textbf{(c)} Distribution of error prosaccades RT: mean value $\bar{t}_{RT} = 207.41$, standard deviation $s=42.76$ ($\bar{t}_{RT} =208.94$ and $s=42.59$ are the corresponding experimental values).} \label{fig:control_exp_opt} \end{figure} \begin{table}[h] \centering \begin{tabular}{ |p{4cm}||p{3cm}|p{3cm}| } \hline \hline Parameter & Estimated values & 95\% Confidence Intervals\\ \hline 1. $r_1$ & 2.135 & $r_1\in (2.23, 2.33)$ \\ 2. $r_2$ & 0.631 & $r_2\in (0.72, 0.79)$\\ 3. $r_3$ & 1.33 & $r_3\in (1.197, 1.47)$\\ 4. $p$ & 0.0101 & $p \in (0.007, 0.0145)$ \\ \hline \end{tabular} \newline \caption{Estimated values of model parameters along with their 95\% confidence intervals for the healthy controls.} \label{tab:norm} \end{table} In order to compare between the experimental and computational RT distributions, we performed a Kolomogorov-Smirnov test. The statistical test showed that the null Hypothesis that the RT data come from the same population cannot be rejected at a 0.05 significance level (p-value: 0.77). \subsection*{Patients with Schizophrenia} The resulting optimal values for the group of patients were: $r_1 = 2.45 (95\% \text{CI}:2.328, 2.58), r_2 = 0.885 (95\% \text{CI}:0.816, 0.954), r_3= 0.857 (95\% \text{CI}:0.805, 1.1), p=0.02 (95\% \text{CI}:0.008, 0.04)$ for the switching probability. The residual of the objective function was 0.003. With these parameter values, the model predicted an error rate of 44\% for the group of patients, which was not significantly different from the experimentally observed one. Fig.\ref{fig:SZ_exp_opt} depicts the distribution of RT antisaccades for the group of patients. Experimental and model-derived RT are in a good agreement. By applying the Kolmogorov Smirnov test, it results that the $H_0$ hypothesis that the two sets of data come from the same distribution cannot be rejected at the 0.05 significance level (p-value: 0.543). \begin{figure}[t] \begin{center} \begin{picture}(400,190) \includegraphics[width=12.5cm]{resschizo.eps \put(-300,170){\textbf{(a)}} \put(-180,170){\textbf{(b)}} \put(-80,170){\textbf{(c)}} \end{picture} \end{center} \caption{Computed distributions of reactions times (RT) for patients as derived using the parameter values estimated by numerical optimization (brown color). The experimental distributions are also depicted for comparison purposes (blue color). \textbf{(a)} Distribution of all RT (correct antisaccades and error prosaccades). Error prosaccades: mean value $\bar{t}_{RT}=266$, standard deviation $s=73.9$ (the corresponding experimental statistical moments are $\bar{t}_{RT}=266, s=81$). \textbf{(b)} Distribution of antisaccade only RT: mean values $\bar{t}_{RT}= 296.87$, standard deviation $s=70$ (the corresponding experimental statistical moments are: $\bar{t}_{RT}=304, s=77$). \textbf{(c)} Distribution of error prosaccades RT: mean value $\bar{t}_{RT}=212.63$, standard deviation $s=52$.} \label{fig:SZ_exp_opt} \end{figure} Figs.~\ref{fig:healthy_95per} and \ref{fig:schizo95_per} depict the mean and 95\% confidence interval estimations of the RT. Fig.~\ref{fig:Mechanism} shows a schematic of two characteristic PFC neural networks for the healthy controls (Fig.\ref{fig:Mechanism}(a)) and for the group of patients (Fig.\ref{fig:Mechanism} (b)); in the second case the resulting PFC network is found to be more random with an increased number of random remote connections. It has been suggested that such random structures act as communication-integration disruptors \cite{Hig07,Lo15}. In concurrent, Fig.\ref{fig:Mechanism}(c) depicts a schematic of a local circuit between a pyramidal excitatory PFC neuron and a GABAergic interneuron (microscopic scale). The optimization process manifested here a mechanism by which the dopamine D1 results to a hypofunction of NMDA neurons (lower $g_{\text{NMDA}}$ in patient case) disrupting this way the functionality of NMDA-GABAergic dipoles. In the higher level of the neural PFC network (mesoscopic scale) (see Fig.\ref{fig:Mechanism}(d)), the abnormal behavior of the pyramidal-interneuron circuit together with the network disruption leads to an ineffective behaviour performance at the macroscopic level. \begin{figure}[ht!] \begin{center} \begin{picture}(400,190) \includegraphics[width=12.9cm]{cand1conftotlastnorm2.eps \put(-300,170){\textbf{(a)}} \put(-180,170){\textbf{(b)}} \put(-75,170){\textbf{(c)}} \end{picture} \end{center} \caption{Healthy controls. The computed distributions of reaction times (solid line); the shaded area depicts the 95\% confidence intervals. \textbf{(a)} Distribution of all reaction times (RT) (correct antisaccades and wrong prosaccades). \textbf{(b)} Distribution of antisaccades only RT. {(c)} Distribution of error prosaccades RT.} \label{fig:healthy_95per} \end{figure} \begin{figure}[ht!] \begin{center} \begin{picture}(400,190) \includegraphics[width=12.5cm]{cand1conftotlastscizo2.eps \put(-300,170){\textbf{(a)}} \put(-180,170){\textbf{(b)}} \put(-75,170){\textbf{(c)}} \end{picture} \end{center} \caption{Schizophrenic patients. Computed distributions of reaction times (solid line); the shaded area depicts the 95\% confidence intervals. \textbf{(a)} Distribution of all reaction times (RT) (correct antisaccades and wrong prosaccades). \textbf{(b)} Distribution of antisccade only RT. {(c)} Distribution of error prosaccades RT.} \label{fig:schizo95_per} \end{figure} \begin{table}[t] \centering \begin{tabular}{ |p{4cm}||p{3cm}|p{3cm}| } \hline \hline Parameter & Estimated value & 95\% Confidence Intervals\\ \hline 1. $r_1$ & 2.456 & $r_1\in ( 2.328, 2.584)$ \\ 2. $r_2$ & 0.885 & $r_2\in (0.816, 0.954)$\\ 3. $r_3$ & 0.957 & $r_3\in (0.805, 1.11)$\\ 4. $p$ & 0.02 & $p \in (0.0075, 0.04)$ \\ \hline \end{tabular} \newline \caption{Estimated values of model parameters along with their 95\% confidence intervals for the group of patients with schizophrenia.} \label{tab:schiz} \end{table} \section*{Discussion} Numerical simulations of the proposed PFC-SC biophysical-based network model that was developed to assess the antisaccade performance in healthy controls and patients with schizophrenia predicted the hypothesis about the effects of the changes in the dopamine levels, NMDA functionality and simultaneously the alteration of the PFC network topology. In particular, decreased DA levels in the PFC, combined with the NMDA hypofunction and the connectivity topology alteration predicted all basic antisaccade performance differences as have been observed experimentally in patients with schizophrenia (see introduction), namely a large increase in the error rate as well as an increase in the mean and variance of the reaction times for the correct antisaccades. Here, it is important to stress that the numerical optimization for the calibration of the model parameters was performed based on the total experimental distribution which contains both antisaccade and error prosaccade RT. Simulation results were in a very good agreement with the experimentally observed ones (both antisaccades and error prosaccades), thus providing evidence that the proposed model is able to capture the essential aspects of the observed data.\par The estimated values of the model parameters for the group of patients suggest a deviant in the PFC activity. More specifically, the conductance of the inhibitory current $I_{\text{AHP}}$ current was found higher in the group of patients (2.456 in patients vs. 2.135 in healthy controls) suggesting an inhibition on the PFC neurons. The increment of $I_{\text{AHP}}$ decreases the firing rates of the PFC neurons which in turn results to a lower value of the $slope_p$ (see Fig.~\ref{fig:Simulations}(b)). This suggests that the buildup neurons trigger more slowly, i.e. they delay to reach the critical value that governs the RT, thus leading to larger reaction times. This result is consistent with the dopamine hypo-activity hypothesis \cite{Yi13, Malenka86, Ped95}. In \cite{Yi13}, using cell patch clamp recording in rat prefrontal neurons showed that the DA activation (under the D1 receptor) suppress the $I_{\text{AHP}}$ current, thus leading to a higher spiking activity. Pezedrani et al.\cite{Ped95} also reports similar results regarding the Hippocampus neurons.\par The conductance of the excitatory synaptic current $I_{\text{NMDA}}$ for the group of patients was found smaller with respect to the one for the group of healthy control (0.957 in patients vs 1.33 in healthy controls), which is consistent with the glutamate hypofunction theory. The reduction of the NMDA conductance weakens the current transmission, impairs the neurons interactions, thus resulting to smaller firing rates of the PFC neurons. The decreased PFC activity weakens the planned input to SC (see sec.\ref{sec:SC}). As a consequence, the planned antisaccade processing that would be appropriate for the given task instruction is defective and the SC produces longer and more variant reaction times. \begin{figure}[t] \centering \begin{picture}(400,270) \includegraphics[width=.98\linewidth]{mechcb.eps} \put(-340,240){\textbf{(a)}} \put(-120,240){\textbf{(b)}} \put(-330,120){\textbf{(c)}} \put(-120,120){\textbf{(d)}} \put(-360,110){D1 Hypoactivity} \put(-330,100){$\Downarrow$} \put(-350,90){NMDA } \put(-350,80){Hypofunction} \end{picture} \caption{A schematic of the PFC neural network. \textbf{(a)} A low number of remote connections is characteristic of the topology for the group of healthy controls. \textbf{(b)} A more random topology, i.e. an increased number of remote connections, resulting from larger values of the switching probability in the Watts and Strogatz \cite{Wat98,Bull09} model is characteristic for the group of schizophrenic patients. \textbf{(c)} A schematic of a disrupted circuit model in patients with schizophrenia. The circuit consists of cortical pyramidal neurons (green) and GABAergic neurons. A D1 hypofunction implies NMDA receptor hypofunction which results to a hypoactivity of the cortical pyramidal neurons. Due to this, the communication between pyramidal and interneurons is disrupted. \textbf{(d)} A schematic of the PFC network; an abnormal behavior of the pyramidal-interneuron circuit in conjunction with the network disruption (randomization) leads to ineffective performance of the PFC which is linked to the SC network.} \label{fig:Mechanism} \end{figure} Furthermore, the results of our study coincide with the disconnetivity hypothesis, that for the pathological case one would expect a disrupted neuronal-functional connectivity in the the PFC-SC network. In our model, the PFC was modeled by leaky integrate-and-fire dynamics evolving on a small-world network structure. In the case of the healthy control group, the value of the switching probability that determines the topology of the small-world network dictates a high clustered (similar to a ring) structure. Instead, in the group of patients, the switching probability was higher: $p \sim 0.01$ vs $p \sim 0.02$ for the group of healthy-control dictating a connectivity that is closer to a random structure. In the review article \cite{Hao07} (and the references therein), the authors report that under working memory tasks, schizophrenic patients involve a sparser network of cortical regions resulting to a reduced cortical signal to noise ratio (which can be interpreted as a loss of signal information). Disruptions in connectivity were also reported in auditory hallucinations, compared with healthy controls and patients who had schizophrenia but not hallucinations \cite{Hubl04,Day16}, suggesting that the auditory hallucinations originate from altered connectivity of the same regions that process normal hearing and speech \cite{Hig07}. Similar results have been reported by Rubinov et al. \cite{Rub09} who studied two groups (healthy controls and patients with schizophrenia) with resting state EEG. The authors concluded that cognitive disturbances may be due to the randomization of the underlying functional connectivity network. In Lo et al. \cite{Lo15} it is also reported significant randomization of global network metrics for patients with schizophrenia. These conclusions are in line with our model outcomes as changes in the network topology going from clustered and structured to more random ones result in a worse cognitive performance in the antisaccade task.\par Based one the above, our proposed model and numerical results may provide useful insights regarding the organization of the connectome and its alteration, especially on the hypothesis of the ``small-worldness" of key parts of the sensorimotor system (in particular of the PFC) and the study of its topological characteristics. Even though, functional neuroimaging studies have provided evidence of the small-world organization of the brain \cite{Bull09}, the connectome and its plasticity within and between areas of the cortex such as the PFC that possess a prominent role in the voluntary control of movements including saccades and the midbrain such as the SC is still not well understood \cite{Meer12}. Several neuroimaging studies that have focused toward this direction have revealed that there are certain differences in the functional connectivity of the PFC between healthy controls and subjects suffering from severe mental disorders such schizophrenia \cite{Yang12, Yu13,Gao2021}. The outcomes of our model was in line with the main findings of these studies.\par In conclusion, our PFC-SC biophysical network model for the approximation and assessment of the reaction times of the antisaccade task showed that the combination of three factors, namely the DA and NMDA hypofunction and the network alteration in the PFC explain the deficits in the antisaccade performance as observed experimentally in patients with schizophrenia. Furthermore, the connectivity changes, as the disconnection hypothesis suggests in the PFC connectome resemble (qualitatively) those observed in functional connectivity neuroimaging studies. Another direction as to which the proposed model could be exploited in future studies is the investigation of the so-called Stochastic Resonance mechanism \cite{Wies95, Kita06, Mend12}, which is related to improved reaction performance in the presence of a low level background noise within a certain interval. Future research could be also targeted on the study of the potential effects of DA-D2 receptors on the neuronal responses of pyramidal neurons in the cortical modules and subsequently on the behavioral data of the patient group. D2 receptors tend to have opposing effects relative to the action of the D1 component, as far as the enhancement or inhibition of cortical currents is concerned \cite{Seam04}. In addition, the extension of the model to include the cortico-striatal, the basal ganglia network constitutes another future direction. Studies in neurological inhibitory control, suggest that inhibition is achieved via a fronto-basal-ganglia network, which could intercept the Go process and thus decrease thalamocortical output \cite{Aron07,Wes16,Hest20}. The augmented model will be used to simulate the control override and inhibition of the habitual response selection mechanism. \cite{Wiec13,Coe19,Aron07}. \section*{Acknowledgements} C.S. acknowledges partial support by INdAM, through GNCS research projects. K.S. and J.S. thank the DFG for support through the Collaborative Research Center CRC 1270 (Deutsche Forschungsgemeinschaft, Grant/ Award Number: SFB 1270/1–299150580).

\section{Introduction} Topological properties for objects in 2D and 3D space is an important task in image processing. An interesting problem called hole counting is to count the number of holes in an 2D image. On the other hand, 3D computer graphics and computational geometry have usually used triangulation to represent a 3D object. It uses the marching-cube algorithm to transfer a digital object into the representation of simplicial complexes. It requires very large amount of space in computer memory. In recent years, the developments in medical imaging and 3D digital camera systems raise the problem of the direct treatment of digital 3D objects due to the speed concerns. In theory, digital topology will provide the method for such fast calculation. This paper will design fast algorithms for those calculation based on digital topology. This paper provides a complete process that deals with simulated and real data in order to obtain the topological invariants for 2D and 3D images. The algorithms are: (1) 2D hole counting, and (2) 3D boundary surface genus calculation. One of the most difficult parts in real image processing is to deal with some noises or pathological cases. This paper also gives detailed procedures for detecting those cases. And we will provide the reasons to modify the original image into the image where the mathematical formula could apply to. \section{Some Concepts of Digital Space } Digital topology was developed for image analysis. We especially interested in obtaining the topological properties for 2D and 3D images, e.g., topological invariants of images. In practice, hole counting in 2D and genus in 3D spaces are most popular problems for real world problems. Now, we first review some concepts. \subsection{Basic Concept of Digital Space} A digital space is a discrete space in which each point can be defined as an integer vector. Two-dimensional digital space $\Sigma_{2}$ first. A point $P (x, y)$ in $\Sigma_{2}$ has two horizontal $(x, y\pm 1)$ and two vertical neighbors $(x\pm 1, y)$. These four neighbors are called directly adjacent points of $p$ . $p$ has also four diagonal neighbors: $(x\pm 1, y\pm 1)$. These eight (horizontal, vertical and diagonal) neighbors are called general (or indirect) adjacent points of $p$ . Let $\Sigma_{m}$ be m-dimensional digital space. Two points $p=(x_1, x_2,...,x_m)$ and $q=(y_1,y_2,...,y_m)$ in $\Sigma_{m}$ are directly adjacent points, or we say that $p$ and $q$ are direct neighbor if \centerline {$ d_{D}(p,q)=\sum_{i=1}^m |x_i-y_i|=1 .$} \noindent $p$ and $q$ are indirectly adjacent points if \centerline {$ d_{I}(p,q)=\max_{1\le i\le m} |x_i-y_i|=1. $} \noindent For instance, in 2D, $d_{D}$ means 4-adjacency, and $d_{I}$ is 8-adjacency. In 3D, $d_{D}$ means 6-adjacency, and $d_{I}$ is 26-adjacency. \noindent (Note: ``Indirectly adjacent points'' include all directly adjacent points here. It may be the reason that we should change the word of ``indirectly'' to ``generally.'') In a three-dimensional space $\Sigma_{3}$, a point has six directly adjacent points and 26 indirectly adjacent points. Therefore, two directly adjacent points in $\Sigma_{3}$ are also called 6-connected, while two indirectly adjacent points are also called 26-connected. In this note, we mainly consider the direct adjacency. If we omit the word ``direct,'' ``adjacency'' means the direct adjacency. A point in $\Sigma_{m}$ is called a point-cell or 0-cell. A pair of points $\{p,q\}$ in $\Sigma_{m}$ is called a line-cell\index{Line-cell} or 1-cell\index{1-cell}, if $p$ and $q$ are adjacent points. A surface-cell\index{surface-cell} is a set of 4 points which form a unit square parallel to coordinate planes. A 3-dimensional-cell (or 3-cell)\index{3-cell} is a unit cube which includes 8 points. By the same reasoning, we may define a $k$-cell. Fig. 1(a)(b)(c)(d) show a point-cell, line-cell, a surface-cell and a 3-cell, respectively. Now let us consider to the concepts of adjacency and connectedness of (unit) cells. Two points $p$ and $q$ (point-cells, or 0-cells) areif there exists a simple path $p_0,p_1,...,p_n$, where $p_0=p$ and $p_n=q$, and $p_i$ and $p_{i+1}$ are adjacent for $i=1,...,n-1 $. Two cells are point-adjacent\index{Point adjacent} if they share a point. For example, line-cells $C1$ and $C2$ are point-adjacent in Fig. 1 (e), and surface-cells $s1$ and $s2$ are point-adjacent in Fig. 1(f). Two surface-cells are line-adjacent\index{Line adjacent} if they share a line-cell. For example, surface-cells $s1$ and $s3$ in Fig. 1(g) are line-adjacent. Two line-cells are point-connected\index{Point connected} if they are two end elements of a line-cells path in which each pair of adjacent line-cells is point-adjacent. For example, line-cells $C1$ and $C3$ in Fig. 1(e) are point-connected. Two surface-cells are line-connected\index{Line connected} if they are two end elements of a surface-cells path in which each pair adjacent surface-cells are point-adjacent. For example, $s1$ and $s2$ in Fig. 1(f) are line-connected. Two $k$-cells are $k'$-dimensional adjacent ($k'$-adjacent), $k>k'\ge 0$, if they share a $k'$-dimensional cell. A (simple) $k$-cells path with $k'$-adjacency is a sequence of $k$-cells $v_0,v_1,...,v_n$, where $v_i$ and $v_{i+1}$ are $k'$-adjacent and $v_0,v_1,...,v_n$ are different elements. Two $k$-cells are called $k'$-dimensional connected if they are two end elements of a (simple) $k$-cells path with $k'$-adjacency. Assume that $S$ is a subset of $\Sigma_{m}$. Let $\Gamma^{(0)}(S)$ be the set of all points in $S$, and $\Gamma^{(1)}(S)$ be the line-cells set in $S$,..., $\Gamma^{(k)}(S)$\index{$\Gamma^{(k)}(S)$} be the set of $k$-cells of $S$. We say two elements $p$ and $q$ in $\Gamma^{(k)}(S)$ are $k'$-adjacent if $p\cap q\in \Gamma^{(k')}(S)$, $k' < k$. Let $p\in \Sigma_{3}$, a line-neighborhood of $p$ is a set containing $p$ and its two adjacent points. A surface-neighborhood of $p$ is a (sub-)surface where $p$ is a inner point of the (sub-)surface. $\Sigma_{m}$ represents a special graph $\Sigma_{m}=(V,E)$. $V$ contains all integer grid points in the $m$ dimensional Euclidean space \cite{MR,Che04}. The edge set $E$ of $\Sigma_{m}$ is defined as $E = \{(a,b) | a, b \in V \& d(a,b)=1\}$ , where $d(a,b)$the distance between $a$ and $b$. In fact, $E$ contains all pairs of adjacent points. Because $a$ is an $m$-dimensional vector, $(a,b)\in E$ means that only one component, the $i$-th component, is different in $a$ and $b$, $|x_i - y_i|=1$, and the rest of the components are the same where $a=(x_1,...,x_m)$ and $b =(y_1,...,y_m)$. This is known as the direct adjacency. One can define indirect adjacency as $\max_{i} |x_i - y_i|=1 $. $\Sigma_{m}$ is usually called an $m$-dimensional digital space.basic discrete geometric element $n$-cells can be defined in such a space, such as 0-cells (point-cells), 1-cells (line-cells), and 2-cells (surface-cells). \begin{figure}[hbt] \begin{center} \input fig2_1.latex \end{center} \caption{Examples of basic unit cells and their connections : (a) 0-cells, (b) 1-cells, (c) 2-cells, (d) 3-cells, (e) point-connected 1-cells, (f) point-connected 2-cells, and (f) line-connected 2-cells.} \end{figure} \subsection{A Simple Lemma for 2D Digital Curve} We have proved some related theorem using Euler Characteristics and Gauss-Bonett Theorem. The first is about simple closed digital curves. $C$ is a simple closed curve in direct (4-) adjacency where each element in $C$ is a point in $\Sigma_{2}$ . We use $IN_{C}$ to represent the internal part of $C$. Since direct adjacency has the Jordan separation property, $\Sigma_2-C$ will be disconnected. We also call a point $p$ on $C$ a $CP_{i}$ point if $p$ has $i$ adjacent points in $IN_{C}\cup C$. In fact, $|CP_{1}|=0$ and $|CP_{i}|=0$ if $i>4$ in $C$. $CP_{2}$ contains outward corner points, $CP_{3}$ contains straight-line points, and $CP_{4}$ contains inward corner points. For example, the following center point is a outward corner point in array (Also see Fig. 2 ): \[ \begin{array}{lll} 0 & 0 & 0 \\ 0 & 1 & 1 \\ 0 & 1 & x \end{array} \] \noindent But in next array, the center point is an inward corner point: \[ \begin{array}{lll} 0 & 1 & x \\ 1 & 1 & x \\ x & x & x \end{array} \] \begin{figure}[h] \begin{center} \epsfxsize=3in \epsfbox{aji_iwow2d.eps} \end{center} \caption{The outward corner point and inward corner point} \end{figure} In ~\cite{Che04} (on page 20) , we used the Euler theorem to show a result for a simple closed curve $C$, \begin{lemma} \begin{equation} CP_2 = CP_4 + 4. \end{equation} \end{lemma} \subsection{Genus of Digital Surfaces in 3D} In this section, we review some existing work related to this paper including the genus of closed digital surfaces, homology groups of manifolds in 3D digital space, and a theoretical linear algorithm of finding Homology Groups in 3D~\cite {CR08}. Any continuous 3D object can be viewed as a collection of 3D voxels in digital or cubical space. Unless the sampling method is changed, any practical method of genus calculation must adapt to this fact. Medical imaging such as CT and MRI are such examples. Cubical space with direct adjacency, or (6,26)-connectivity in digital space\cite{Che04}, has the simplest topology in 3D digital spaces. It is also sufficient for the topological property extraction of digital objects in 3D. Two points are said to be adjacent in (6,26)-connectivity space if the Euclidean distance of these two points is 1, called direct adjacency. Let $M$ be a closed (orientable) digital surface in the 3D grid space in direct adjacency. We know that there are exactly 6-types of digital surface points (See Fig. 3). This was discovered by Chen and Zhang in~\cite{CZ93}. Relation to different definitions of digital surfaces can be found in ~\cite{CCZ99,Che04}. \begin{figure}[h] \begin{center} \epsfxsize=2in \epsfbox{six_type.eps} \end{center} \caption{ Six types of digital surfaces points in 3D} \end{figure} Assume that $M_i$ ($M_3$, $M_4$, $M_5$, $M_6$) is the set of digital points with $i$ neighbors. We have the following result for a simply connected $M$ ~\cite{CZ93}\cite{Che04}: \begin{equation} |M_3| =8 + |M_5| + 2 |M_6| . \end{equation} \noindent $M_4$ and $M_6$ has two different types, respectively. Gauss-Bonnet theorem states that if $M$ is a closed manifold, then \begin{equation} \int_{M} K_{G} d A = 2 \pi \chi(M) \end{equation} \noindent where $d A$ is an element of area and $K_{G}$ is the Gaussian curvature. Its discrete form is \begin{equation} \Sigma_{\{p \mbox{ is a point in } M\}} K(p) = 2 \pi \cdot (2- 2 g) \end{equation} \noindent where $g$ is the genus of $M$. Assume that $K_i$ is the curvature of elements in $M_i$, $i=$ 3,4,5,6. We have \begin{lemma}\label{l21} (a) $K_3 = \pi/2$, (b) $K_4= 0$, for both types of digital surface points, (c) $K_5 = - \pi /2$, and (d) $K_6 = - \pi$, for both types of digital surface points. \end{lemma} \noindent We obtained (see \cite{CR08}), \begin{equation} g = 1+ (|M_5|+2 \cdot |M_6| -|M_3|)/8. \end{equation} The three simple examples show that the above formula is correct \cite{CR08}. See Fig. 4. Different surface points can also be used to form a feature vector for 3D surfaces. We have used it in face modeling~\cite{CS}. \begin{figure}[h] \begin{center} \epsfxsize=2in \epsfbox{fig3.eps} \end{center} \caption{ Simple examples of closed surfaces with $g=0,1,2$} \end{figure} \begin{figure}[h] \begin{center} \epsfxsize=2.3in \epsfbox{face2_2D_both.eps} \end{center} \caption{Different surface points on human faces.} \end{figure} For a $k$-manifold, Homology group $H_{i}$, $i=0,...,k$ provides the information for the number of holes in each $i$-skeleton of the manifold. When the genus of a closed surface is obtained, we can then calculate the homology groups corresponding to its 3-dimensional manifold in 3D. The following result follows from standard results in algebraic topology \cite{Hat}. It also appears in \cite{Day98}. Let $b_i = \mbox{rank} H_i(M, Z)$ be the $i$th Betti number of $M$. The Euler characteristic of $M$ is defined by \[ \chi (M) = \sum_{i\geq 0} (-1)^i b_i \] If $M$ is a 3-dimensional manifold, $H_i(M)=0$ for all $i>3$ essentially because there are no $i$-dimensional holes. Therefore, $\chi (M) = b_o - b_1 + b_2 - b_3$. Furthermore, if $M$ is in $R^3$, it must have nonempty boundary. This implies that $b_3 = 0$. \begin{theorem}\label{Jordan2} Let $M$ be a compact connected 3-manifold in $S^3$. Then \begin{enumerate} \item [(a)] $H_0(M)\cong Z$. \item [(b)] $H_1(M) \cong Z^{\frac{1}{2} b_1(\partial M)}$, i.e. $H_1(M)$ is torsion-free with rank being half of rank $H_1(\partial M)$. \item[(c)] $H_2(M) \cong Z^{n-1}$ where $n$ is the number of components of $\partial M$. \item[(d)] $H_3(M)=0$ unless $M=S^3$. \end{enumerate} \end{theorem} A proof of above theorem is shown in \cite{CR08}. \section{Hole Counting Problem in 2D} In an image, hole counting is to count the number of holes in an 2D image. It was studied by some researchers before~\cite{Qia,LYS}. In this paper, we will get the simplest method to solve the problem. A line or curve in real world always have a thickness no matter how thin it is. However, a digital line could give human some wrong interpretation. The example in Fig. 6 will show you how similar digital objects produce different answers. \begin{figure}[h] \begin{center} \epsfxsize=4in \epsfbox{Aji2D_1.eps} \end{center} \caption{2D digital objects looks the same but topologically different: (1) an angle with $thickness=1$ by MS Paint Software, (2) an angle with $thickness=2$, (3) an angle with $thickness=3$, (4) an ellipse and a line with $thickness=1$, (5) an ellipse and a line with different thickness, and (6) an ellipse and a line with $thickness=2$. } \end{figure} Micro-software engineers for MS Paint probably never thought that there are some major difference except the width of a paint brash. The artistic effect are not very different. However, the images in digital world makes great deal of differentiality from Fig 6. (1) to (2) than the difference from (2) to (3). This is because that Fig 6. (1) can be interpreted as a (square-)dotted line for one of its leg for direct adjacency. A dotted line is a collection of several disconnected objects. This is much complicated issue than a connected line. Fig. 6 (4) may be no hole, one hole, or two holes. In 8-adjacency, there is no hole. In 4-adjacency, the points are not connected. The best way is to use 8-adjacency for "1"s (foreground) and use 4-adjacency for "0"s (background). Only Fig. 6 (6) will give the answer exactly the same as humans interpretation, i.e. two holes. If we use 8-adjacency for "1"s and use 4-adjacency for "0"s, this type of adjacency is called (8,4)-adjacency. It may cause another problem, for instance, if we have two parallel "1" lines with a "0" line in between in 45 degree. Each "0" point will be determined as a separated component. They are not formed a connected "0" line. That is also against the human interpretation. In this paper, our method will assume that $C$ does not contain the following cases (if there is any, we will modify the original image to remove them, we will discuss next): \[\begin{array}{ll} 1 & 0 \\ 0 & 1 \end{array} \] \noindent and \[\begin{array}{ll} 0 & 1 \\ 1 & 0 \end{array}\] \noindent These two cases are called the pathological cases (See Fig.7) \begin{figure}[h] \begin{center} \epsfxsize=2in \epsfbox{aji_pc2d.eps} \end{center} \caption{Two pathological cases } \end{figure} It is obvious that, our paper does not solve all problems. It is too strict. The advantage of our method is to get a simple treatment. In ~\cite{BNB}, there were detailed discussions about complex cases. Our algorithm will fill or delete some points in the original image to make the pathological cases go away. We also want to remove single point whatever it is black or white. it will be treated as noise. Topological invariants should maintain Jordan property. We shall only allow the direct adjacency in order to dealing with Topological invariants, at least in most of cases. The number of holes in a connected component in 2D images is a basic invariant. In this note, a simple formula was proven using our previous results in digital topology~\cite{Che04,CR}. The new is: $h =1+ (|C_4|-|C_2|)/4$ , where h is the number of holes, and $C_i$ indicate the set of corner points having $i$ direct adjacent points in the component. \subsection{The Simple Formula for the Number of Holes in $S$} An image segmentation method can extract a connected component. A connected component $S$ in a 2D digital image is often used to represent a real object. The identification of the object can be first done by determining how many holes in the component. For example, letter ``A'' has one hole and ``B'' has two holes. In other words, if $S$ has $h$ holes, then the complement of $S$ has $h+1$ connected components (if S does not reach the boundary of the image). \begin{theorem} Let $S \subset \Sigma_2$ be a connected component and its boundary $B$ is a collection of simple closed curves without pathological cases. Then, the number of holes in $S$ is \begin{equation} h= 1+(C_4-C_2)/4 \end{equation} \noindent $C_4, C_2 \subset B$. \end{theorem} In this section, we will give two methods to prove our result. First, use the 3D formula to get the theorem for holes. Second, we use 2D formula to directly prove the theorem.\\ {\bf \it The First Proof: } Let $S\subset \Sigma_2$ be a connected component and its boundary do not have the pathological cases. (We actually can detect those cases in linear time.) We can embed $S$ into $\Sigma_3$ to make a double $S$ in $\Sigma_3$. At $z=1$ plane, we have a $S$, denoted $S_1$, and we also have the exact same $S$ at $z=2$ plane, denoted $S_2$. Without loss generality, $S_1\cup S_2$ is a solid object. (We here omit some technical details for the strict definition of digital surfaces.) It's boundary is closed digital surfaces with genus $g=h$. We know \[g = 1+ (|M_5|+2 \cdot |M_6| -|M_3|)/8\] There will be no points in $M_6$. For each point $x$ in $C_2$ in $C\subset S$ ($C$ is the boundary of $S$), we will get two points in $M_3$ in $S_1\cup S_2$. In the same way, if a point $y$ is inward in $C_4\in C$, we will get two points in $M_5$ in $S_1\cup S_2$. There is no point in $M_6$, i,e., $|M_6|=0$. So $2|C_2|=|M_3|$, and $2|C_4|=|M_5|$. We have \[h=g = 1+ (|M_5|+2 \cdot |M_6| -|M_3|)/8= 1+ (2|C_4|-2|C_2|)/8 \] \noindent Thus, \[h= 1+ (|C_4|-|C_2|)/4 \] $\diamond$ We can also prove this theorem using Lemma 1 for digital curves: $CP_2=CP_4+4$ for a simple closed curve. \\ {\bf \it The Second Proof: } This can also be proved by the lemma in above section. \[CP_2=CP_4+4\] A 2D connected component $S$ with $h$ holes that contains $h+1$ simple closed curves in the boundary of $S$ Those curves do not cross each other. The $h$ curves corresponding to $h$ holes will be considered oppositely in terms of inward-outward. including one counts at inward and h is reversed outward with inward. It will get there. Let $CP^{(0)}$ the outside curve of $S$ and $CP^{(i)}, i=1,\cdots,h$, is the curve for the $i$-th hole. Inward points to $S$ is the outward points to $C^{(i)}, i=1,\cdots,h$. And vise versa. \[CP_2^{(0)}=CP_4^{(0)}+4\] \[CP_2^{(i)}=CP_4^{(i)}+4\] The total outward points in the boundary of $S$ is \[ CP_{2}=CP_2^{(0)} + \sum_{i=1}^{h} CP_4^{(i)}\] \noindent The inward points in the boundary of $S$ is \[CP_{4}=CP_4^{(0)} + \sum_{i=1}^{h} CP_2^{(i)}\] \noindent Thus, \[CP_{4}-CP_{2}=CP_4^{(0)} + \sum_{i=1}^{h} CP_2^{(i)} -CP_2^{(0)} - \sum_{i=1}^{h} CP_4^{(i)}\] we have $CP_{4}-CP_{2}=-4 + \sum_{i=1}^{h} 4 = -4 +4h$ Therefore, \[h = 1 + (CP_{4}-CP_{2})/4\] $\diamond$ We can see that this formula is so simple to get the holes (genus) for a 2D object without any little sophistic algorithm, just count if the point is a corner point, inward or outward. We could not get the similar simple formula in triangulated representation of the 2D object. This is the beauty of digital geometry and topology! {\bf Example 1} To test if this formula is correct, we select the following examples \begin{eqnarray} \left (\begin{array}{llllllll} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 1 & 1 & 1 & 1 & 1 & 0 & 0 \\ 0 & 1 & 1 & 1 & 0 & 0 & 0 & 0\\ 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 1 & 1 & 0 & 0 & 0\\ 0 & 0 & 1 & 1 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \end{array} \right ) \end{eqnarray} In order to see clearly, we use ``2'' to represent points in $CP_2$ and use ``4'' to represent points in $CP_4$. \begin{eqnarray} \left (\begin{array}{llllllll} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 2 & 1 & 1 & 2 & 0 & 0 \\ 0 & 2 & 4 & 4 & 1 & 2 & 0 & 0 \\ 0 & 2 & 4 & 1 & 0 & 0 & 0 & 0\\ 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\ 0 & 0 & 1 & 4 & 2 & 0 & 0 & 0\\ 0 & 0 & 2 & 1 & 2 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \end{array} \right ) \end{eqnarray} In this example $|CP_2| = 8$ and $|CP_4| = 4$. $h = 1 + (CP_{4}-CP_{2})/4 = 1 + (4-8)/4 =0$. Another example is the following \begin{eqnarray} \left (\begin{array}{llllllll} 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\ 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1 \\ 0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\ 0 & 1 & 1 & 1 & 0 & 0 & 1 & 1\\ 0 & 1 & 1 & 1 & 0 & 0 & 1 & 1 \\ 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1\\ 0 & 0 & 1 & 1 & 1 & 1 & 1 & 1 \\ 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \end{array} \right ) \end{eqnarray} In the second example $|CP_2| = 6$ and $|CP_4| = 6$. $h = 1 + (CP_{4}-CP_{2})/4 = 1 + (6-6)/4 =1$. \noindent When add a hole, we will add 4 more $CP_4$ points. That is the reason why this formula is correct. \subsection{Algorithms for Hole Counting} The key of the algorithm is to delete all the pathological cases. There two types of such cases. Sometimes it is hard to decide if we need to add a point or delete a point to a pathological case to be removed. The first method will based on the original grade space. Sometimes, delete a pixel for removing pathological case may add another pathological case. In such a case, we will add a pixel to make the original pathological case go away. and vice versa. In other case, if add and delete will not reduce the pathological case, use delete. since delete will eventually complete the job. It must be the case shown in Fig. 8. \begin{figure}[h] \begin{center} \epsfxsize=2in \epsfbox{aji_pc_del_2d.eps} \end{center} \caption{ Any deleting or adding a pixel will add a new pathological case } \end{figure} The second method is made to apply to the case that so many instances in the image as shown in Fig. 8. We need to consider a refinement method e.g. half grade space to avoid deletion and addition of a cell that will cause other pathological case. This method will increase the space need. We use half size cells to fill the space. Or we just delete a half cell to remove the pathological case.\\ {\bf Algorithm 3.1} Algorithm for calculating number of holes in 2D images. \begin{description} \item [Step 1.] Get connected components using direct-adjacency (4-adjacency). \item [Step 2.] Extract every component, do following. Fill all single "0" pixel or (unit-square or 2-cell) to be "1" if all its indirect neighbors (8-adjacency) are "1". Delete all single "1" pixel if all its indirect neighbors (8-adjacency) are "0". \item [Step 3.] Find all pathological cases, delete or add a pixel if this action does not create a new pathological case. If either action will create a new pathological case, use delete. Repeat this step until all pathological cases are removed. \item [Step 4.] Count all inward edge points $CP_4$ and outward points $CP_{2}$ . \item [Step 5.] $h = 1 + (CP_{4}-CP_{2})/4$ \end{description} Note that deleting a pixel means to change the single "0" pixel (unit-square or 2-cell) to "1" pixel. is for noise image. It is not very necessary for doing this step. The above calculation is for each component to get holes. Algorithm 3.1 is an $O(\log n)$ space algorithm without Step 1 since it is just count the number of point types. If we assume that the number of pathological cases are constant, this algorithm will be linear in time. \section {Algorithms for Genus on 3D Digital Surfaces } Basically, the topological properties of an object in 3D contains connected components, genus of its boundary surfaces, and other homologic and homotopic properties~\cite {Kaz03}. In 3D, the problem of obtaining fundamental groups is decidable but no practical algorithm has yet been found. Therefore, homology groups have played the most significant role~\cite{Day98} ~\cite{Kaz04} . Theoretical results show that there exist linear time algorithms for calculating genus and homology groups for 3D Objects in 3D space~\cite{Day98}. However, the implementation of these algorithms is not simple due to the complexity of real data samplings. Most of the algorithms require the triangulation of the input data since it is collected discretely. However, for most medical images, the data was sampled consecutively, meaning that every voxel in 3D space will contain data. In such cases, researchers use the marching-cubes algorithm obtain the triangulation since it is a linear time algorithm ~\cite{LC87}. However the spatial requirements for such a treatment will be at least doubled by adding the surface-elements (sometimes called faces). The theoretical work of calculating genus based on simple decomposition will turn into two different procedures: (1) finding the boundary of a 3D object and then using polygon mapping, also called polygonal schema, (2) cell complex reductions where a special data structure will be needed. In this paper, we look at a set of points in 3D digital space, and our purpose is to find homology groups of the data set. The direct algorithm without utilizing triangulation was proposed by Chen and Rong in 2008~\cite {CR08} However, this algorithm is based on the strict definition of digital surfaces. Many real 3D sets may not satisfy the definition. In other words, a set of connected points may not be able to be put into such a process without considerable associated theoretical and practical processes. In \cite {CR08}, we discuss the geometric and algebraic properties of manifolds in 3D digital spaces and the optimal algorithms for calculating these properties. We consider {\em digital manifolds} as defined in ~\cite {Che04}. More information related to digital geometry and topology can be found in ~\cite{KR} and ~\cite{KoR}. We presented a theoretical optimal algorithm with time complexity $O(n)$ to compute the genus and homology groups in 3D digital space, where $n$ is the size of the input data ~\cite {CR08}. The key in the algorithm in ~\cite{CR08} is to find the genus of the closed digital surfaces that is the boundary of the 3D object. However, the new algorithm is based on the strict definition of closed digital surfaces in \cite{Che04}, which means that there are many cases of real sampling of 3D objects that do not satisfy the definition of digital surfaces. In this paper, we will also deal with extreme situations. We have designed an adding and deleting method to make the 3D object into manifolds. first describe a theoretical procedure in this section for 3D genus for 3D objects. The implementation of the algorithm must consider all possible cases in practical data collection. We first need to find the boundary and then decide if the boundary is a 2D manifold. If the boundary data connecting voxel data sets are not purely defined digital surfaces, we will have three options: (1) we need to modify the data to meet the requirement before genus calculation, (2) if the change of the original data set is too great, we may need to stop the modification instead of outputting a result for reference, and (3) we make some limited changes, and then produce a result. The difference between the theoretical results and practical data processing is that we may not always get the input data we expected. In our case, the boundary of a solid object should be treated as a surface. However, practically, this might not always be the case. Some researchers also consider making real data sets ``well''-organized. Siqueira {\it et al} considered making a 26-connected data set well-composed~\cite{SLG05} \cite{SLTGG08}~\cite{BB08}. This means that two voxels will be connected by a sequence of voxels where each pair of two adjacent cubes share a 2D-cell (face-unit). The concept of well-composed is mathematically equivalent to 6-connected. An algorithm described in ~\cite{SLG05} \cite{SLTGG08} may generate new "none" well-composed cases, which are not good selections for genus calculation. Our new algorithm and implementation will perform: (1) pathological cases detection and deletion, (2) raster space to point space (dual space) transformation, (3) the linear time algorithm for boundary point classification, and (4) genus calculation. Some detailed considerations of recognition algorithms related to 3D manifolds can be found in ~\cite{BK08} where Brimkov and Klette made extensive investigations in boundary tracking. The discussions of 3D objects in raster space can be found in ~\cite{Lat}. \subsection{A Theoretical Procedure for Genus in 3D} Based on the results we presented in the above subsections, we now describe a linear algorithm for computing the homology group of 3D objects in 3D digital space ~\cite{CR08}. Assuming we only have a set of points in 3D. We can digitize this set into 3D digital spaces. There are two ways of doing so: (1) by treating each point as a cube-unit that is called the raster space, (2) by treating each point as a grid point, which is also called the point space. These two are dual spaces. Using the algorithm described in ~\cite{Che04}, we can determine whether the digitized set forms a 3D manifold in 3D space in direct adjacency for connectivity. The algorithm is in linear time. \\ \noindent {\bf Algorithm 4.1} Let us assume that we have a connected $M$ that is a 3D digital manifold in 3D. \begin{description} \item [Step 1.] Track the boundary of $M$, $\partial M$, which is a union of several closed surfaces. This algorithm only needs to scan though all the points in $M$ to see if the point is linked to a point outside of $M$. That point will be on boundary. \item [Step 2.] Calculate the genus of each closed surface in $\partial M$ using the method described in Section 2. We just need to count the number of neighbors on a surface. and put them in $M_i$, using the formula (5) to obtain $g$. \item [Step 3.] Using the Theorem \ref{Jordan2}, we can get $H_0$, $H_1$, $H_2$, and $H_3$. $H_0$ is $Z$. For $H_1$, we need to get $b_1(\partial M)$ that is just the summation of the genus in all connected components in $\partial M$. (See \cite{Hat} and \cite{Day98}.) $H_2$ is the number of components in $\partial M$. $H_3$ is trivial. \end{description} \begin{lemma} Algorithm 4.1 is a linear time algorithm. \end{lemma} Therefore, we can use linear time algorithms to calculate $g$ and all homology groups for digital manifolds in 3D based on Lemma 2, Formula (5) and Lemma 3. \begin{theorem} There is a linear time algorithm to calculate all homology groups for each type of manifold in 3D. \end{theorem} \subsection{3D Input Data Sets} This subsection will discuss the input data formats. We will focus on cubical data, for instance MRI and CT data. In cubical data samples, we assume that the sampling is contiguous, where each sample point is normally followed by another sample point in its neighborhood. It is important to know this because a random sampling can cause the problem of uncertainty. In this case, we usually cannot calculate the genus without making an assumption. For instance, we will not be able to know where a hole is. In order to get simplicial decomposition (usually triangulation), we usually need to use Voronoi or Delaunay decomposition with boundary information. That means the boundary must be assumed. A new technology is called persistent homology analysis that tells us how to find the best estimation for the location of holes, usually by multiscaling (the upscaling and downscaling methods). However, this method is not a precise analysis ~\cite{Car}~\cite{ZC}. Even though, our method can be modified to be used in persistent analysis, this paper mainly deals with the method of precise genus and homology group calculation. In summary, our assumption is that the digital object consists of cubical points (digital points, raster points). Each point is a cube, which is the smallest 3D object. The edge and point are defined with regards to the cube and an object may contain several connected components using a cube-linking path. Our purpose again is to calculate the topological properties of the object, or of each component, essentially. \subsection{Searching connected components of a cubical data set} Connected component search is an old task that can be done by using Tarjan's Breadth-first-search. Pavlidis was one of the first people to realize and use this algorithm in image processing. This problem is also known as the labeling problem. The complexity of the algorithm is $O(n)$ ~\cite{Pav}. The problem is what connectivity is based off of. In 3D, we usually have 6-, 18-, 26- connectivity. Since real data has noise, we have to consider all of those connectivities. So we must use 26-connectivety to get the connected components. Therefore, the connected component of the real processing is not a strictly 6-connected component. The topological theorem generated previously in ~\cite{CR08} is no longer suitable. So we need to transform a 26-connected component into a 6-connected component. This should be done by a meaningful adding or deleting process since optimization on the minimum number of changes could be an NP-hard problem. {\bf Problem of minimum modifications:} Given a set of points in 3D digital space, if this set is not a manifold, assume that the points are connected in a connectivity defined using adding or deleting processes to make the set a 3D manifold. The question becomes: is there a polynomial algorithm that makes the solution have minimum modifications where adding or deleting a data point will be counted as one modification? A similar problem was considered in ~\cite{SLG05} in which a decision problem of adding was proposed. This problem can be extended to a general $k$-manifold in $n$-D space. Even though we have the 6-connected component, there may still be cases that contain the pathological situation, which needs special treatment. We will discuss this issue in the next subsection. \subsection{Pathological Cases Detection and Deletion} In this paper, we only deal with the Jordan manifolds, meaning that a closed $(n-1)$-manifold will separate the $n$-manifold into two or more components. For such a case, only direct adjacency will be allowed since indirect adjacency will not generate Jordan cases. That is to say, if the set contains indirect adjacent voxels, we need to design an algorithm to detect the situation and delete some voxels in order to preserve the homology groups. It is known that there are only two such cases in cubical or digital space \cite{Che04}: two voxels (3-cells) share a 0-cell or a 1-cell. Therefore, we want to modify the voxel set to only contain voxels where two of these cases do not appear. Two voxels share exactly a 2-cell, or there is a local path (in the neighborhood) voxels where two adjacent voxels share a 2-cell ~\cite{Che04}. A special case was found in \cite{SLTGG08} that is the complement case of the case in which two voxels share a 0-cell (see Fig 9. (a)). The case may create a tunnel or could be filled. We will simplify it by adding a voxel in a $2\times 2\times 2$ cube. Such a case in point space is similar to the case (a) in Fig 9 since boundaries of these cases are the same. The problem is that many real data sets do not satisfy the above restrictions (also called well composed image). The detection is easy but deleting certain points (the minimum points deletion) to preserve the homology is a bigger issue. The following rules (observations) are reasonable: In a neighborhood $N_{27}(p)$ that contains 8 cubes and 27 grid points, a) if a voxel only shares a 0-cell with a voxel. This voxel can be deleted. b) if a voxel only shares a 1-cell with a voxel. This voxel can be deleted. c) if a boundary voxel $v$ shares a 0,1-cell with a voxel, assume $v$ also shares a 2-cell with a voxel $u$, $u$ must share a 0,1-cell with a voxel that is not in the object $M$. $u$ is on the boundary. Deleting $v$ will not change the topological properties. d) if in a $2\times 2\times 2$ cube, there are 6 boundary voxels and its complement (two zero-valued voxels) is the case (a) in Fig. 9. Add a voxel to this $2\times 2\times 2$ cube such that the new voxel shares as many 2-cells in the set as possible. This means that we want the adding voxel to be inside of the object as much as we can. \begin{figure}[h] \begin{center} \epsfxsize=3in \epsfbox{fig1.eps } \end{center} \caption{ Pathological Cases in 3D} \end{figure} In this paper, we implement or modify the above rules to fit the theoretical definition of the digital surfaces. We also design an algorithm based on these rules to detect and delete some data points while preserving the topology. This is essential to calculating the genus correctly. However, when the object becomes more complex, pathological situations may still exist. \begin{figure}[h] \begin{center} \epsfxsize=4in \epsfbox{pathe.eps} \end{center} \caption{ Without Pathological Case Process} \end{figure} The mathematical foundation of this above process that eliminates pathological cases is still under investigation. {\bf Mathematical foundation of Modifying a 3D object to be a 3D manifold:} Given a set of points in 3D digital space, how would we modify the data set into a manifold without losing or changing the topology (in mathematics)? \subsection{Boundary Search} In general, a point is on the boundary if and only if it is adjacent to one point in the object and one point not in the object (in 26-connectivity). A simple algorithm that goes through each point and tests the neighborhood will determine whether a point is on the boundary or not. This is a linear time and $O(log(n))$ space algorithm. The only thing special is that we use 26-connectivity to determine the boundary points. This is to take all possible boundary points into consideration in the next step. \subsection{Determination of the Configuration of Boundary Points} When all boundary points are found, we need to find their classifications. In other words, we need to determine whether a special point is in $M_3$, $M_4$, $M_5$, or $M_6$. Here is the problem: if we only have one voxel, is it a point (0-cell) or a 3D object (3-cell)? In this paper, we treat it as a 3-cell. The input data is in raster space, but the boundary surface will be in point space. We must first make the translation. Then, for each point on the surface, we count how many neighbors exist in order to determine its configuration category. After that, we use formula (5) to get the genus. If we still need to find homology groups, we can just use the simple calculations based on Theorem 1,3 to them. Using the program, we get the genus$=6$ for a modified real image (Fig. 12). \begin{figure}[h] \begin{center} \epsfxsize=3in \epsfbox{g_5.eps} \end{center} \caption{ After Pathological Case Process} \end{figure} \begin{figure}[h] \begin{center} \epsfxsize=3in \epsfbox{bone_1.eps} \end{center} \caption{ A Modified Real Image} \end{figure} \section{Summary and Discussion} In this paper, we have used digital topology to get a simple formula for calculating the number of holes in a connected component in 2D digital space. The formula is so simple and can be easily implemented. The author does not know if this formula was known or obtained already by other researchers. 3D images, we have practically get the genus by extracting the boundary surfaces and deleting the pathelogical case. Both algorithms are optimum in terms of time and space complexity (We usually assume that the number of the pathelogical cases are constant). This paper is modified based on one unpublished note~\cite{Che12} and a conference paper ~\cite{Che09}.

\section{Introduction} \label{Introduction} For $n > 0$, let ${\mathbb {P}}_{n+2}$ denote the regular $(n+2)$-gon. The \textbf{\textit{(classical) associahedron}} is the $(n-2)$-dimensional simplicial sphere whose faces are the dissections of ${\mathbb {P}}_{n+2}$. The polytopal dual to this complex was introduced by Stasheff \cite{Stasheff} to study nonassociative binary operations arising in algebraic topology. Given a Fu$\ss$ parameter $k \in {\mathbb {Z}}_{> 0}$, a dissection of ${\mathbb {P}}_{kn+2}$ is called \textbf{\textit{$k$-divisible}} if the number of vertices in each sub-polygon is congruent to $2$ modulo $k$. Przytycki and Sikora \cite{PS} proved that the number of $k$-divisible dissections of ${\mathbb {P}}_{kn+2}$ with $i$ diagonals equals $\frac{1}{n} {kn+i+1 \choose i} {n \choose i+1}$. Using this result and the Fomin-Zelevinsky cluster complexes \cite{FZ} as motivation, Tzanaki \cite{Tzanaki} studied the \textbf{\textit{generalized associahedron}} whose faces are the $k$-divisible dissections of ${\mathbb {P}}_{kn+2}$. She proved that this generalization of the associahedron is shellable. As a corollary, the generalized associahedron is homotopy equivalent to a wedge of $k$ spheres, all of dimension $n-2$. Rational Catalan combinatorics is a further generalization of Fu\ss-Catalan combinatorics which depends on a pair of coprime positive integers $a < b$, thought of as corresponding to the rational number $\frac{b-a}{a} \in {\mathbb{Q}}_{> 0}$. The choice $(a, b) = (n, n+1)$ recovers classical Catalan theory and the choice $(a, b) = (n, kn+1)$ recovers Fu\ss-Catalan theory. While some of the objects considered in rational Catalan theory date back at least to the 1950s (see \cite{Bizley}), its systematic study was initiated only recently. In particular, Armstrong, Williams, and the author \cite{ARW} defined and studied rational generalizations of Dyck paths, noncrossing and nonnesting partitions of $\{1, 2, \dots, n \}$, noncrossing perfect matchings on $\{1, 2, \dots, 2n \}$, and the associahedron. In a companion paper, Armstrong, Loehr, and Warrington \cite{ALW} defined rational parking functions and studied generalizations of the statistics `area', `dinv', and `iDes' to this setting. The motivation for this combinatorial program was Gordon's use of rational Cherednik algebras to give a generalization of the diagonal coinvariant ring to any reflection group $W$ \cite{Gordon}. The favorable representation theoretic properties of these algebras at parameter $\frac{b}{h}$, where $h$ is the Coxeter number of $W$ and $h < b$ are coprime, suggested the problem of studying the combinatorics of this `rational' case when $W = {\mathfrak{S}}_a$. Generalizing rational Catalan combinatorics beyond type A is almost entirely an open problem. This paper focuses on the rational analog of the associahedron, which gives a generalization of the classical and generalized associahedra. We prove a conjecture of Armstrong, Williams, and the author \cite[Conjecture 4.7]{ARW}, obtaining the conjecture \cite[Proposition 4.8]{ARW} as a corollary (the proof of \cite[Proposition 4.8]{ARW} was given in \cite{ARW} as an easy consequence of \cite[Conjecture 4.7]{ARW}). By verifying these conjectures, we will uncover some genuinely new features of rational Catalan combinatorics which are invisible at the classical and Fu\ss-Catalan levels of generality. This gives evidence that the rational level of generality is of combinatorial interest. Given coprime positive integers $a < b$, Armstrong, Williams, and the author \cite{ARW} defined two simplicial complexes ${{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(a,b)$ called `rational associahedra' whose faces are certain dissections of ${\mathbb {P}}_{b+1}$. When $a = n$ and $b = kn+1$, the complexes ${{\sf Ass}}(n, kn+1)$ and ${\widehat{\sf Ass}}(n, kn+1)$ coincide and both equal the generalized associahedron of $k$-divisible dissections of ${\mathbb {P}}_{kn+2}$. In general, we have that ${{\sf Ass}}(a,b)$ is a subcomplex of ${\widehat{\sf Ass}}(a,b)$, and the case of $(a, b) = (3, 5)$ shows that this inclusion can be strict. The complexes ${\widehat{\sf Ass}}(3,5)$ and ${{\sf Ass}}(3,5)$ are shown on the left and right of Figure~\ref{retract}, respectively. At the rational level of generality, many of the nice features of associahedra diverge between the two constructions ${{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(a,b)$. The complex ${\widehat{\sf Ass}}(a,b)$ has a simpler definition which is more intrinsically related to polygon dissections and is closed under the dihedral symmetries of ${\mathbb {P}}_{b+1}$. However, the complex ${\widehat{\sf Ass}}(a,b)$ is not in general pure and there does not appear to be a nice formula for the entires of its $f$- and $h$-vectors. The complex ${{\sf Ass}}(a,b)$ has a more complicated definition involving lattice paths and does not carry an action of the symmetry group of ${\mathbb {P}}_{b+1}$. However, the complex ${{\sf Ass}}(a,b)$ is pure and shellable, and there are nice product formulas for the $f$- and $h$-vector entries of ${{\sf Ass}}(a,b)$ given by rational analogs of the Kirkman and Narayana numbers. Needless to say, having two different models for rational associahedra is unfortunate and the different advantageous features of ${{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(a,b)$ make it difficult to choose which model is ``correct". We ameliorate this problem by proving that these complexes are equivalent up to combinatorial homotopy. \begin{theorem} \label{main} (\cite[Conjecture 4.7]{ARW}) Let $a < b$ be coprime positive integers. The simplicial complex ${\widehat{\sf Ass}}(a,b)$ collapses onto the simplicial complex ${{\sf Ass}}(a,b)$. \end{theorem} Theorem~\ref{main} is trivial up to the Fu$\ss$ level of generality because we have the equality of complexes ${\widehat{\sf Ass}}(n, kn+1) = {{\sf Ass}}(n, kn+1)$; the disagreement between the complexes ${\widehat{\sf Ass}}(a,b)$ and ${{\sf Ass}}(a,b)$ and the resulting nontrivial collapsing appears only at the rational level of generality. In the case $(a, b) = (3, 5)$, Theorem~\ref{main} can be proven by observing that in Figure~\ref{retract}, the complex on the right can be obtained from the complex on the left by collapsing the top and bottom triangles. To prove Theorem~\ref{main}, we will describe an explicit collapse of ${\widehat{\sf Ass}}(a,b)$ onto ${{\sf Ass}}(a,b)$. The proof of Theorem~\ref{main} will be given at the end of the paper after a series of intermediate results. The idea is to identify the obstructions which prevent a face of ${\widehat{\sf Ass}}(a,b)$ from being a face of the subcomplex ${{\sf Ass}}(a,b)$. These obstructions will be local in nature and can be encoded as the edges of an `obstruction graph'. Using a well chosen total order on these edges, we can use a sequence of collapses to eliminate these obstructions and prove Theorem~\ref{main}. As was noted in \cite[Proposition 4.8]{ARW}, Theorem~\ref{main} has a corollary which gives a topological relationship between different rational associahedra. Let $S$ be a sphere. A topological subspace $X$ of $S$ is said to be \textbf{\textit{Alexander dual}} to the complement $S - X$. Generalizing slightly, we also say that two topological subspaces $X$ and $Y$ of $S$ are Alexander dual to one another if $S-X$ deformation retracts onto $Y$ and $S-Y$ deformation retracts onto $X$. For any $a < b$ coprime, the faces of the complexes ${{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(a,b)$ are given by dissections of ${\mathbb {P}}_{b+1}$, so either of these complexes embeds in the $(b-3)$-dimensional simplicial sphere given by the classical associahedron ${{\sf Ass}}(b-1,b)$. We have the following involution on increasing pairs of coprime positive integers: \begin{equation} \label{integer-duality} a < b \longleftrightarrow b-a < b. \end{equation} At the level of rational associahedra, this ``categorifies" to Alexander duality. \begin{corollary} (see \cite[Proposition 4.8]{ARW}) \label{alexander} Let $a < b$ be coprime positive integers. The simplicial complexes ${\widehat{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(b-a,b)$ are Alexander dual within the sphere ${{\sf Ass}}(b-1,b)$. The simplicial complexes ${{\sf Ass}}(a,b)$ and ${{\sf Ass}}(b-a,b)$ are also Alexander dual within the sphere ${{\sf Ass}}(b-1,b)$. \end{corollary} Figure~\ref{alexanderdual} shows the complexes ${{\sf Ass}}(3,5)$ (in red) and ${{\sf Ass}}(2,5)$ (in blue) as Alexander duals within the $2$-sphere ${{\sf Ass}}(4,5)$ (in black). Since the `Catalan' pairs $\{(n, n+1)\}$ and the `Fu\ss-Catalan' pairs $\{(n, kn+1)\}$ are not closed under the above involution, Corollary~\ref{alexander} is another genuinely new feature of rational Catalan theory. Corollary~\ref{alexander} was proven for the ${\widehat{\sf Ass}}$ complexes in \cite{ARW} using a relatively elementary argument. By showing that ${{\sf Ass}}(a,b)$ is shellable and computing its $h$-vector, it is also shown in \cite{ARW} that ${{\sf Ass}}(a,b)$ is homotopy equivalent to a wedge of $\frac{1}{b}{b \choose a}$ spheres, all of dimension $a-2$. This implies that the homology groups of ${{\sf Ass}}(a,b)$ and ${{\sf Ass}}(b-a,b)$ have ranks predicted by Corollary~\ref{alexander}. Along the way of proving Theorem~\ref{main} and Corollary~\ref{alexander}, we will also give a structural result on the complexes ${{\sf Ass}}(a,b)$. \begin{proposition} \label{flag} The simplicial complex ${{\sf Ass}}(a,b)$ is flag. \end{proposition} It follows from its definition that ${\widehat{\sf Ass}}(a,b)$ is also a flag simplicial complex. The remainder of the paper is organized as follows. In \textbf{Section~\ref{Background}} we give background definitions related to simplicial complexes and recall the key combinatorial tool (Lemma~\ref{cone-vertex-lemma}) we will use to perform the collapsing in Theorem~\ref{main}. In \textbf{Section~\ref{Rational}} we review the definitions of ${{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(a,b)$ from \cite{ARW}. In \textbf{Section~\ref{Flag complexes}} we begin our analysis of the difference ${\widehat{\sf Ass}}(a,b) - {{\sf Ass}}(a,b)$ and prove Proposition~\ref{flag}. In \textbf{Section~\ref{Obstruction graphs}} we introduce the notion of an obstruction graph, which will serve as our gadget for keeping track of the obstructions which prevent a face of ${\widehat{\sf Ass}}(a,b)$ from being a face of ${{\sf Ass}}(a,b)$. In \textbf{Section~\ref{Collapsing}} we prove Theorem~\ref{main}. For the sake of completeness, we also recall from \cite{ARW} how Corollary~\ref{alexander} can be deduced from Theorem~\ref{main}. We make closing remarks in \textbf{Section~\ref{Open}}. \section{Background on simplicial complexes} \label{Background} Let $E$ be a finite set. A \textbf{\textit{simplicial complex}} $\Delta$ (on the \textbf{\textit{ground set}} $E$) is a collection of subsets of $E$ such that if $F \in \Delta$ and $F' \subseteq F$, we have that $F' \in \Delta$. Elements of $\Delta$ are called \textbf{\textit{faces}} and elements of the ground set $E$ are called \textbf{\textit{vertices}}. A \textbf{\textit{facet}} of $\Delta$ is a maximal face of $\Delta$. The \textbf{\textit{dimension}} $\dim(F)$ of a face $F \in \Delta$ is given by $\dim(F) := |F| - 1$. The \textbf{\textit{dimension}} $\dim(\Delta)$ of the complex $\Delta$ is the maximum $\max \{ \dim(F) \,:\, F \in \Delta \}$ of the dimensions of its faces. The complex $\Delta$ is called \textbf{\textit{pure}} if all of the facets of $\Delta$ have the same dimension. Let $\Delta$ be a simplicial complex with $\dim(\Delta) = d$. The \textbf{\textit{$f$-vector}} $f(\Delta) = (f_{-1}, f_0, \dots, f_d)$ is the vector obtained by letting $f_i$ be the number of $i$-dimensional faces of $\Delta$. The \textbf{\textit{$h$-vector}} is the vector $h(\Delta) = (h_0, \dots, h_d)$ defined by $h_k = \sum_{i = 0}^k (-1)^{k-i} {d - i \choose k - i} f_{i-1}$. Let $\Delta$ be a simplicial complex on the ground set $E$ and let $F$ be a subset of $E$ with $|F| \geq 2$. The set $F$ is called an \textbf{\textit{empty face}} if $F$ is not a face of $\Delta$, but every two-element subset $\{v, v'\} \subseteq F$ of $F$ is a face of $\Delta$. The complex $\Delta$ is called \textbf{\textit{flag}} if $\Delta$ does not contain any empty faces. If $\Delta$ is flag, then $\Delta$ is determined by its $1$-skeleton $\{ F \in \Delta \,:\, |F| \leq 2 \}$. Let $\Delta$ be a simplicial complex, let $F$ be a facet of $\Delta$, and let $F' \subset F$ be a face satisfying $\dim(F') = \dim(F) - 1$. The pair $(F, F')$ is called \textbf{\textit{free}} if $F$ is the unique face of $\Delta$ satisfying $F' \subset F$. When the pair $(F, F')$ is free, the set of faces $\Delta - \{F, F'\}$ is a subcomplex of $\Delta$ and we say that $\Delta - \{F, F'\}$ is obtained from $\Delta$ by an \textbf{\textit{elementary collapse}} (along the free pair $(F, F')$). Topologically, this corresponds to the deformation retraction given by crushing the facet $F$ through $F'$. Given an arbitrary subcomplex $\Delta'$ of $\Delta$, we say that $\Delta$ \textbf{\textit{collapses onto}} $\Delta'$ if there exists a sequence of subcomplexes $\Delta = \Delta_1 \supset \Delta_2 \supset \dots \supset \Delta_n = \Delta'$ such that $\Delta_{i+1}$ is obtained from $\Delta_i$ by an elementary collapse for $1 \leq i \leq n-1$. Collapses were introduced in 1938 by Whitehead \cite{Whitehead} and give a combinatorial model for certain deformation retractions. Collapses are transitive in the sense that if $\Delta$ collapses onto $\Delta'$ and $\Delta'$ collapses onto $\Delta''$, then $\Delta$ collapses onto $\Delta''$. Let $\Delta$ be a simplicial complex on the ground set $E$ and let $F$ be a face of $\Delta$. We define $\Delta(F) := \{ F' \in \Delta \,:\, F \subseteq F' \}$ to be the set of faces of $\Delta$ containing $F$. We say that a ground set element $c \in E - F$ which is not in $F$ is a \textbf{\textit{cone vertex for $F$ in $\Delta$}} if for any $F' \in \Delta(F)$ we have that $F' \cup \{c\} \in \Delta$. If $\Delta$ is a simplicial complex and $F$ is a face of $\Delta$, define the \textbf{\textit{deletion}} ${\mathrm {dl}}_{\Delta}(F)$ to be the subcomplex of $\Delta$ given by ${\mathrm {dl}}_{\Delta}(F) := \{ F' \in \Delta \,:\, F \nsubseteq F' \} = \Delta - \Delta(F)$. More generally, if $\Sigma$ is any set of faces of $\Delta$, define ${\mathrm {dl}}_{\Delta}(\Sigma) := \{ F' \in \Delta \,:\, \text{$F \nsubseteq F'$ for all $F \in \Sigma$} \} = \bigcap_{F \in \Sigma} {\mathrm {dl}}_{\Delta}(F)$. To prove that complexes collapse onto deletions, we will make repeated use of the following basic lemma. \begin{lemma} \label{cone-vertex-lemma} Let $\Delta$ be a simplicial complex on the ground set $E$ and let $F$ be a face of $\Delta$. Suppose that there exists a cone vertex $c \in E - F$ for $F$ in $\Delta$. Then the complex $\Delta$ collapses onto the deletion ${\mathrm {dl}}_{\Delta}(F)$. \end{lemma} \begin{proof} Since $c$ is a cone vertex for $F$ in $\Delta$, there is a fixed-point free involution $\phi: \Delta(F) \rightarrow \Delta(F)$ defined by \begin{equation} \phi(F') := \begin{cases} F' \cup \{c\} & \text{if $c \notin F'$,} \\ F' - \{c\} & \text{if $c \in F'$.} \end{cases} \end{equation} The map $\phi$ partitions $\Delta(F)$ into pairs $\{F, \phi(F)\}$. We let $\leq$ be any total order on these pairs such that whenever $F \subset \phi(F)$ and $F' \subset \phi(F')$ with $F \subseteq F'$, we have that $\{F, \phi(F)\} \leq \{F', \phi(F')\}$. If we decompose $\Delta(F)$ as $\Delta(F) = \biguplus_{i = 1}^n \{F_i, \phi(F_i)\}$ with $\{F_1, \phi(F_1)\} > \dots > \{F_n, \phi(F_n) \}$, we have that $\Delta - \{ F_1, \phi(F_1), F_2, \phi(F_2), \dots, F_i, \phi(F_i) \}$ is obtained from $\Delta - \{ F_1, \phi(F_1), F_2, \phi(F_2), \dots, F_{i-1}, \phi(F_{i-1}) \}$ by an elementary collapse along the free pair $(F_i, \phi(F_i))$ for all $1 \leq i \leq n$. It follows that $\Delta$ collapses onto ${\mathrm {dl}}_{\Delta}(F)$. \end{proof} \section{Rational associahedra} \label{Rational} Let $a < b$ be coprime positive integers. We label the extremal boundary \textbf{\textit{points}} of ${\mathbb {P}}_{b+1}$ clockwise with $0, 1, 2, \dots, b$ (see Figure~\ref{dyckfacet}). We call these ``points" rather than ``vertices" so that we do not get confused with the ground sets of our simplicial complexes. A \textbf{\textit{diagonal}} of ${\mathbb {P}}_{b+1}$ is a line segment other than a side which connects two boundary points. We use the shorthand $ij$ to refer to the diagonal connecting the boundary points $i$ and $j$ which satisfy $i < j$. Two diagonals $ij$ and $km$ of ${\mathbb {P}}_{b+1}$ are said to \textbf{\textit{cross}} if either $i < k < j < m$ or $k < i < m < j$. Following \cite{ARW}, we define a set $S(a,b)$ of positive integers as follows: \begin{equation} S(a,b) := \left\{ \left\lfloor \frac{ib}{a} \right\rfloor \,:\, i = 1, 2, \dots, a-1 \right\}. \end{equation} Let $d$ be a diagonal of ${\mathbb {P}}_{b+1}$ which separates $i$ boundary points from $j$ boundary points, where $1 \leq i, j \leq b-2$. The diagonal $d$ is said to be \textbf{\textit{$a,b$-admissible}} if $i, j \in S(a,b)$. \begin{defn} \cite{ARW} Let $a < b$ be coprime positive integers. The complex ${\widehat{\sf Ass}}(a,b)$ is the simplicial complex on the ground set of $a,b$-admissible diagonals in ${\mathbb {P}}_{b+1}$ whose faces are mutually noncrossing collections of $a,b$-admissible diagonals. \end{defn} \begin{figure} \centering \includegraphics[scale = 0.4]{retract.pdf} \caption{The rational associahedra ${\widehat{\sf Ass}}(3,5)$ and ${{\sf Ass}}(3,5)$.} \label{retract} \end{figure} The complex ${\widehat{\sf Ass}}(3,5)$ is shown on the left of Figure~\ref{retract}. In general, the complex ${\widehat{\sf Ass}}(a,b)$ carries an action of the symmetry group of ${\mathbb {P}}_{b+1}$. Figure~\ref{retract} shows that ${\widehat{\sf Ass}}(a,b)$ need not be pure in general. When $b = ka+1$ for $k \in {\mathbb {Z}}_{>0}$, we have that $S(a, ka+1) = \{k, 2k, \dots, (a-1)k \}$ and ${\widehat{\sf Ass}}(a, ka+1)$ is the complex of $k$-divisible dissections of ${\mathbb {P}}_{ka+2}$. The definition of the complex ${{\sf Ass}}(a,b)$ is more involved and uses lattice paths. An \textbf{\textit{$a,b$-Dyck path}} (or just a \textbf{\textit{Dyck path}} if $a$ and $b$ are clear from context) is a lattice path in ${\mathbb {Z}}^2$ consisting of north and east steps which starts at $(0, 0)$, ends at $(b, a)$, and stays above the line $y = \frac{a}{b} x$. A \textbf{\textit{vertical (horizontal) run}} of a Dyck path is a maximal contiguous sequence of east (north) steps. A \textbf{\textit{valley}} of a Dyck path is a lattice point on the path which is immediately preceded by an east step and immediately succeeded by a north step. We will sometimes write a Dyck path as a word of length $a + b$ in the alphabet $\{N, E\}$ consisting of a north and an east step. It will occasionally be convenient to think of $a,b$-Dyck paths in terms of partitions. A \textbf{\textit{partition}} (of length $a$) is a sequence $\lambda = (\lambda_1 \geq \dots \geq \lambda_a)$ of weakly decreasing nonnegative integers. We identify $\lambda$ with its \textbf{\textit{Ferrers diagram}} which consists of $\lambda_i$ left justified boxes in row $i$. We let $\subseteq$ denote the partial ordering of \textbf{\textit{Young's Lattice}} given by containment of Ferrers diagrams, so $\lambda \subseteq \mu$ if and only if $\lambda_i \leq \mu_i$ for all $i$. An $a,b$-Dyck path $D$ cuts out a partition $\lambda(D)$ whose Ferrers diagram lies to the northwest of $D$. Let $D$ be an $a,b$-Dyck path and let $P = (i, j)$ be a lattice point on $D$ other than $(0, 0)$ which lies at the bottom of a north step of $D$. The \textbf{\textit{laser}} $\ell(P) = \ell(i, j)$ of $P$ is the unique line segment of slope $\frac{a}{b}$ whose southwest endpoint is $P$, whose northeast endpoint lies on $D$, and whose interior does not intersect $D$. The northeast endpoint of $\ell(P)$ lies in the interior of an east step of $D$. If the right endpoint of this east step has coordinates $(k, m)$, we define $d(P) := ik$ to be the \textbf{\textit{diagonal}} of $P$, viewed as a diagonal in ${\mathbb {P}}_{b+1}$. Since the laser $\ell(P)$ has slope $\frac{a}{b}$, the diagonal $d(P)$ is $a,b$-admissible. Moreover, if $P$ and $Q$ are two bottom lattice points of north steps in $D$, we have that $d(P)$ and $d(Q)$ do not cross. We let ${\mathcal{F}}(D)$ be the set \begin{equation} {\mathcal{F}}(D) := \{ d(P) \,:\, \text{$P$ is the lattice point at the bottom of a north step of $D$} \}. \end{equation} We will occasionally use the more elaborate notation $\ell_D(P)$ and $d_D(P)$ to emphasize the role of the Dyck path $D$. The sets ${\mathcal{F}}(D)$ as $D$ varies over Dyck paths form the facets of ${{\sf Ass}}(a,b)$. \begin{defn} \cite{ARW} Let $a < b$ be coprime positive integers. The complex ${{\sf Ass}}(a,b)$ is the simplicial complex on the ground set of $a,b$-admissible diagonals of ${\mathbb {P}}_{b+1}$ whose facets are \begin{equation} \{ {\mathcal{F}}(D) \,:\, \text{$D$ is an $a,b$-Dyck path} \}. \end{equation} \end{defn} \begin{example} A $5,8$-Dyck path $D$ is shown in Figure~\ref{dyckfacet}. The line $y = \frac{5}{8} x$ is dotted. As a word in the alphabet $\{N, E\}$, the path $D$ is given by $NNENNEEENEEEE = N^2EN^2E^3NE^4$. The partition $\lambda(D)$ is given by $\lambda(D) = (4,1,1,0,0)$. The valleys of $D$ are the lattice points $(1, 2)$ and $(4, 4)$. Lasers have been fired from all possible lattice points of $D$. The lasers are shown in red. The corresponding facet ${\mathcal{F}}(D)$ of ${{\sf Ass}}(5,8)$ is shown on the right as a dissection of ${\mathbb {P}}_9$. \end{example} \begin{figure} \centering \includegraphics[scale = 0.6]{dyckfacet.pdf} \caption{A $5,8$-Dyck path and its facet in ${{\sf Ass}}(5,8)$.} \label{dyckfacet} \end{figure} Since the diagonals in the facets ${\mathcal{F}}(D)$ are automatically pairwise noncrossing, we have that ${{\sf Ass}}(a,b)$ is a subcomplex of ${\widehat{\sf Ass}}(a,b)$. The complex ${{\sf Ass}}(a,b)$ is pure of dimension $a-2$ and shellable. The number of facets of ${{\sf Ass}}(a,b)$ is the \textbf{\textit{rational Catalan number}} $\frac{1}{a+b}{a+b \choose a,b}$. More generally, the number $f_{i-2}$ of $(i-2)$-dimensional faces of ${{\sf Ass}}(a,b)$ is given by the \textbf{\textit{rational Kirkman number}} $\frac{1}{a} {a \choose i}{b+i-1 \choose i-1}$. The $h$-vector entry $h_{i-2}$ of ${{\sf Ass}}(a,b)$ is given by the \textbf{\textit{rational Narayana number}} $\frac{1}{a} {a \choose i}{b-1 \choose i-1}$. In contrast, there is no known formula for the number of faces of ${\widehat{\sf Ass}}(a,b)$. The complex ${{\sf Ass}}(3,5)$ is shown on the right of Figure~\ref{retract}. It is our aim to show that ${\widehat{\sf Ass}}(a,b)$ collapses onto its subcomplex ${{\sf Ass}}(a,b)$. At this point, it is not even clear that the difference ${\widehat{\sf Ass}}(a,b) - {{\sf Ass}}(a,b)$ contains an even number of faces. \section{Flag complexes} \label{Flag complexes} Let $a < b$ be coprime and let $F$ be a face of ${\widehat{\sf Ass}}(a,b)$. In order to prove that ${\widehat{\sf Ass}}(a,b)$ collapses onto ${{\sf Ass}}(a,b)$, we will need to develop a better understanding of when $F$ is contained in the subcomplex ${{\sf Ass}}(a,b)$. It will turn out that the obstructions preventing $F$ from being a face of ${{\sf Ass}}(a,b)$ can be encoded in a graph which will give an inductive structure allowing us to apply Lemma~\ref{cone-vertex-lemma} repeatedly to prove Theorem~\ref{main}. The main purpose of this section is to show that such a graph exists by proving Proposition~\ref{flag}. Given a face $F \in {{\sf Ass}}(a,b)$, there may be many Dyck paths $D$ such that $F$ is contained in the facet ${\mathcal{F}}(D)$. However, it will be useful to associate a `standard' Dyck path ${\mathcal{D}}(F)$ to $F$ such that $F$ is contained in the facet ${\mathcal{F}}({\mathcal{D}}(F))$. The Dyck path ${\mathcal{D}}(F)$ will be characterized by the facts that \begin{itemize} \item every valley of ${\mathcal{D}}(F)$ fires a laser corresponding to a diagonal in $F$, and \item $F \subseteq {\mathcal{F}}({\mathcal{D}}(F))$. \end{itemize} \begin{lemma} \label{valley-lemma} Let $F$ be a nonempty face of ${{\sf Ass}}(a,b)$. There exists a unique Dyck path ${\mathcal{D}}(F)$ whose vertical runs are at the $x$-coordinates \begin{equation*} \{ i \,:\, \text{there exists a diagonal of the form $ij$ in $F$} \} \end{equation*} such that the valley at the bottom of the vertical run on the line $x = i$ for $i > 0$ has laser diagonal $i j_0$, where $j_0$ is maximal such that $i j_0$ is a diagonal in $F$. \end{lemma} \begin{proof} Suppose that we can construct a Dyck path ${\mathcal{D}}(F)$ which satisfies the conditions of the lemma. The diagonals in $F$ determine the $x$-coordinates of the vertical runs of ${\mathcal{D}}(F)$. Since $a < b$ and our lasers have slope $\frac{a}{b}$, different lasers give rise to different diagonals. Therefore, the condition on the valley lasers implies that $F$ determines the valleys of ${\mathcal{D}}(F)$. This means that $F$ determines the path ${\mathcal{D}}(F)$ and ${\mathcal{D}}(F)$ is unique when it exists. We will show that a path ${\mathcal{D}}(F)$ satisfying the conditions of the lemma exists. To see that such a path ${\mathcal{D}}(F)$ exists, let $D$ be any Dyck path such that $F$ is a face of ${\mathcal{F}}(D)$ (we know that such a path $D$ exists because $F$ is a face of ${{\sf Ass}}(a,b)$). If $D$ does not satisfy the condition of the lemma, we will prove that there exists another Dyck path $D'$ with $\lambda(D') \subset \lambda(D)$ such that $F$ is also a face of ${\mathcal{F}}(D')$. Existence then follows by induction. Our construction of the path $D'$ is similar to a construction used in \cite{ARW} to prove that lexicographic order on the partitions $\lambda(D)$ induces a shelling order on the facets ${\mathcal{F}}(D)$. Suppose that $D$ does not satisfy the condition of the lemma. Then there exists a valley $P$ of $D$ such that $d(P)$ is not a diagonal in $F$. Let $N^{i_1}, \dots, N^{i_r}$ denote the vertical runs of $D$, read from east to west. There exists $1 < s \leq r$ such that $P$ occurs at the bottom of the vertical run $N^{i_s}$. There are two cases to consider. {\bf Case 1:} {\it There are no lattice points $Q$ on the vertical runs $N^{i_1}, \dots, N^{i_{s-1}}$ such that the laser $\ell(Q)$ hits $D$ east of $P$.} In this case, let $D'$ be obtained from $D$ by increasing the length of $N^{i_1}$ by one and decreasing the length of $N^{i_s}$ by one. This has the effect of lifting the horizontal runs of $D$ to the west of $P$ by one unit. We claim that $F$ is contained in ${\mathcal{F}}(D')$ and that $\lambda(D') \subset \lambda(D)$. We have $\lambda(D') \subset \lambda(D)$ because $D'$ is constructed from $D$ by lifting a nonempty collection of horizontal runs (namely, those between $N^{i_1}$ and $N^{i_s}$) up by one unit. To see that $F$ is contained in ${\mathcal{F}}(D')$, we will observe that any laser involved in the realization of a diagonal of $F$ in ${\mathcal{F}}(D)$ corresponds to a laser realizing that same diagonal in ${\mathcal{F}}(D')$. Indeed, let $R$ be a lattice point of $D$ such that $d_D(R) \in F$. If $R$ occurs to the northeast of $P$ in $D$, then $R$ is also a lattice point in $D'$ and we have the equality of lasers $\ell_D(R) = \ell_{D'}(R)$ since $D$ and $D'$ agree to the northeast of $P$. It follows that $d_D(R) = d_{D'}(R) \in {\mathcal{F}}(D')$. On the other hand, if $R$ occurs to the southwest of $P$ in $D$, then the vertical shift $R' = R + (0, 1)$ is a lattice point of $D'$ and the laser $\ell_{D'}(R')$ is obtained by shifting the laser $\ell_D(R)$ up by one unit. It follows that $d_D(R) = d_{D'}(R') \in {\mathcal{F}}(D')$. The following diagram illustrates the construction in Case 1 when $(a, b) = (5, 8)$. We have that $D = N^1 E^1 N^1 E^2 N^2 E^1 N^1 E^4$ and the lasers which give the face $F$ of ${\mathcal{F}}(D)$ are shown. The point $P$ is labeled and we have $s = 3$ and $i_s = 2$. None of the lasers which originate on lattice points to the west of $P$ hit $D$ to the east of $P$. To form $D' = N^2 E^1 N^1 E^2 N^1 E^1 N^1 E^4$, we increase the length of the first vertical run by one and decrease the length of the vertical run above $P$ by one. We can vertically translate the lasers involved in the face $F$ to see that $F$ is also a face of ${\mathcal{F}}(D')$. We have that $\lambda(D') = (4,3,1,0,0)$ and $\lambda(D) = (4,3,3,1,0)$, so that $\lambda(D') \subset \lambda(D)$. \begin{center} \includegraphics[scale = 0.4]{caseone.pdf} \end{center} {\bf Case 2:} {\it There are some lattice points $Q$ on the vertical runs $N^{i_1}, \dots, N^{i_{s-1}}$ such that the laser $\ell(Q)$ hits $D$ to the east of $P$.} In this case, let $1 \leq m \leq s-1$ be maximal such that there exists a lattice point $P$ on $N^{i_m}$ such that $\ell(P)$ hits $D$ to the east of $P$. Form $D'$ by increasing the length of $N^{i_m}$ by one and decreasing the length of $N^{i_s}$ by one. This has the effect of lifting the horizontal runs of $D$ between $N^{i_m}$ and $P$ by one unit. We again claim that $F$ is contained in ${\mathcal{F}}(D')$ and that $\lambda(D') \subset \lambda(D)$. The argument is similar to Case 1, but slightly more complicated. We have $\lambda(D') \subset \lambda(D)$ because $D'$ is formed from $D$ by moving the nonempty set of horizontal runs between $N^{i_m}$ and $N^{i_s}$ up one unit. To see that $F$ is contained in ${\mathcal{F}}(D')$, let $R$ be a lattice point of $D$ such that $d_D(R) \in F$. We argue that $d_{D}(R) \in {\mathcal{F}}(D')$. If $R$ occurs to the northeast of $P$, then $R$ is also a lattice point of $D'$. Since $D'$ agrees with $D$ to the northeast of $P$, we have the equality of lasers $\ell_D(R) = \ell_{D'}(R)$, so that $d_{D}(R) = d_{D'}(R) \in {\mathcal{F}}(D')$. If $R$ lies to the southwest of $P$, our reasoning breaks up into further subcases. If $R$ lies strictly east of the vertical run $N^{i_m}$ (but southwest of $P$), then the shift $R' = R + (0, 1)$ is on the Dyck path $D'$. By our choice of $m$, we know that the laser $\ell_D(R)$ hits $D$ to the east of $P$. By our construction of $D'$, the laser $\ell_{D'}(R')$ is obtained by shifting the laser $\ell_D(R)$ up one unit, so that $d_D(R) = d_{D'}(R') \in {\mathcal{F}}(D')$. If $R$ lies strictly west of the vertical run $N^{i_m}$, then $R$ also lies on the lattice path $D'$. By our choice of $m$ and the fact that lasers from different lattice points do not cross, the laser $\ell_D(R)$ hits $D$ either to the west of the vertical run $N^{i_m}$ or to the east of the valley $P$. Since $D'$ is formed from $D$ by shifting $D$ up between these points, we get that $\ell_D(R) = \ell_{D'}(R)$ in either case, so that $d_D(R) = d_{D'}(R') \in {\mathcal{F}}(D')$. Finally, suppose that $R$ lies on the vertical run $N^{i_m}$. Depending on where on this run $R$ lies, the laser $\ell_D(R)$ may hit $D$ either to the west of $P$ or to the east of $P$. Suppose $\ell_D(R)$ hits $D$ to the west of $P$. The construction of $D'$ implies that the shift $R' = R + (0, 1)$ is a lattice point on $D'$ and that the laser $\ell_{D'}(R')$ is obtained by shifting the laser $\ell_D(R)$ up by one unit. It follows that $d_D(R) = d_{D'}(R') \in {\mathcal{F}}(D')$. Suppose that $\ell_D(R)$ hits $D$ to the east of $P$. By the construction of $D'$ we know that $R$ is also a lattice point of $D'$ and that we have the equality of lasers $\ell_D(R) = \ell_{D'}(R')$. This means that $d_D(R) = d_D(R') \in {\mathcal{F}}(D')$. The following diagram illustrates the construction in Case 2 when $(a, b) = (5, 8)$. We have that $D = N^1 E^1 N^1 E^1 N^1 E^2 N^2 E^4$ and the lasers which give the face $F$ of ${\mathcal{F}}(D)$ are shown. The point $P$ is labeled and we have $s = 4$ and $m = 2$. To form $D' = N^1 E^1 N^2 E^1 N^1 E^2 N^1 E^4$, we increase the length of the $m^{th} = 2^{nd}$ vertical run by one and decrease the length of the vertical run above $P$ by one. As in Case 1, we can vertically translate the lasers involved in the face $F$ to see that $F$ is also a face of ${\mathcal{F}}(D')$. We have that $\lambda(D) = (4,4,2,1,0)$ and $\lambda(D') = (4,2,1,1,0)$, so that $\lambda(D') \subset \lambda(D)$. \begin{center} \includegraphics[scale = 0.4]{casetwo.pdf} \end{center} \end{proof} The path ${\mathcal{D}}(F)$ in Lemma~\ref{valley-lemma} will be called the \textbf{\textit{valley path}} of $F$. The path ${\mathcal{D}}(F)$ can be characterized as the unique Dyck path among those Dyck paths $D$ such that $F \subseteq {\mathcal{F}}(D)$ which minimizes $\lambda(D)$ in Young's Lattice, but we will not need this characterization. By convention, we will set the valley path of the empty face to be the path $N^a E^b$ consisting of $a$ north steps followed by $b$ east steps. We return to the more general situation of a face $F \in {\widehat{\sf Ass}}(a,b)$. The following algorithm constructs the valley path ${\mathcal{D}}(F)$ if $F$ is a face of ${{\sf Ass}}(a,b)$, or returns the statement that $F$ is not a face of ${{\sf Ass}}(a,b)$. Roughly speaking, we attempt to construct ${\mathcal{D}}(F)$ ``backwards" by starting with the empty path at $(b, a)$ and working our way towards $(0, 0)$. To emphasize the backwards nature of this construction, we refer to the steps added to ${\mathcal{D}}(F)$ as south and west steps rather than north and east steps. Starting at $(b, a)$, we go west until we hit an $x$-coordinate $i$ such that $ij$ is a diagonal in $F$ for some $j$. We then go south until we pick up all of the diagonals in $F$ of the form $ij$, at which point we go west again until we hit another $x$-coordinate $i'$ such that $i'j'$ is a diagonal in $F$ for some $j'$, etc. If it ever happens that a diagonal in $F$ cannot be achieved by going south, we go south until we cross the line $y = \frac{a}{b}x$ and return the statement that $F$ is not a face of ${{\sf Ass}}(a,b)$. Otherwise, we will eventually reach the origin $(0, 0)$ and recover the path ${\mathcal{D}}(F)$ (and the fact that $F$ is a face of ${{\sf Ass}}(a,b)$). PATH BUILDING ALGORITHM: \\ \indent INPUT: A face $F$ of ${\widehat{\sf Ass}}(a,b)$. \\ \indent OUTPUT: The valley path ${\mathcal{D}}(F)$ or the statement that $F$ is not a face of ${{\sf Ass}}(a,b)$. \begin{enumerate} \item Initialize ${\mathcal{D}}(F)$ to be the empty lattice path starting and ending at $(b,a)$ and initialize $i = b$. While $i \geq 0$, do the following. \begin{enumerate} \item If $i > 0$ and there are no diagonals in $F$ of the form $i j$, add a west step to the southwest end of ${\mathcal{D}}(F)$, so that ${\mathcal{D}}(F)$ has southwest corner with $x$-coordinate $i-1$. \item If $i = 0$ and there are no diagonals in $F$ of the form $0 j$, add south steps to the southwest end of ${\mathcal{D}}(F)$ until ${\mathcal{D}}(F)$ reaches $(0, 0)$. \item If there are diagonals in $F$ of the form $i j$, express these diagonals as $i j_1, \dots, i j_r$, where $j_1 < \dots < j_r$. Add south steps to the southwest corner of ${\mathcal{D}}(F)$ until there exist lattice points $P$ on ${\mathcal{D}}(F)$ whose laser diagonals correspond to all of the diagonals $i j_1, \dots, i j_r$ or until ${\mathcal{D}}(F)$ goes below the line $y = \frac{a}{b}x$. If $i > 0$, add a single west step to the southwest corner of ${\mathcal{D}}(F)$. If $i = 0$, add south steps to the southwest corner of ${\mathcal{D}}(F)$ until ${\mathcal{D}}(F)$ is weakly below $(0, 0)$. \item If ${\mathcal{D}}(F)$ lies weakly above the line $y = \frac{a}{b}x$, replace $i$ by $i-1$. Otherwise, terminate the algorithm and return ``$F$ is not a face of ${{\sf Ass}}(a,b)$". \end{enumerate} \item Return ``$F$ is contained in the facet ${\mathcal{F}}({\mathcal{D}}(F))$ of ${{\sf Ass}}(a,b)$ and ${\mathcal{D}}(F)$ is the valley path of $F$". \end{enumerate} A couple comments on this algorithm are in order. At the start of every iteration of the loop (1), the path ${\mathcal{D}}(F)$ starts at $(b, a)$ and ends at a point with $x$-coordinate $i$ while staying above the line $y = \frac{a}{b} x$. In step (1c), we consider firing lasers of slope $\frac{a}{b}$ to the northeast from north steps with $x$-coordinate $i$. Since ${\mathcal{D}}(F)$ has northeast corner $(b, a)$ and stays above $y = \frac{a}{b} x$, the laser diagonals referenced in this step are well defined. \begin{example} We examine a case where the path building algorithm does not terminate in a Dyck path. Let $(a, b) = (5, 8)$ and consider the face $F = \{57, 24, 05, 04\}$ of ${\widehat{\sf Ass}}(5,8)$. We initialize the lattice path ${\mathcal{D}}(F)$ to be the empty path at the point $(8, 5)$. As the algorithm progresses, the lattice path ${\mathcal{D}}(F)$ changes as shown. For ease of reading, $x$-coordinates and laser diagonals involved in $F$ are given. \begin{center} \includegraphics[scale = 0.4]{algorithm.pdf} \end{center} We start at the lattice point $(8, 5)$ and build the path ${\mathcal{D}}(F)$ from northeast to southwest, working west until hitting an $x$-coordinate corresponding to the smaller endpoint of at least one diagonal in $F$ and then working south until we get every diagonal of $F$ with that smaller endpoint appearing as a laser (or, if this is impossible, working south until we go below the line $y = \frac{5}{8} x$). In particular, starting at $(8, 5)$, we go west until we hit the line $x = 5$. Then we go south one unit to pick up the diagonal $58 \in F$. Then we go west until we hit the line $x = 2$. We go south one unit to pick up the diagonal $24 \in F$. Then we go west until we hit the line $x = 0$. While we can pick up the diagonal $05 \in F$ by going south, it is impossible to pick up the diagonal $04 \in F$. As a result of this `failure', we continue going south until we cross the line $y = \frac{5}{8} x$ (this crossing step is shown in green) and conclude that $F \notin {{\sf Ass}}(5,8)$. \end{example} \begin{lemma} \label{valid} The path building algorithm is correct. \end{lemma} \begin{proof} When $F \in {{\sf Ass}}(a,b)$, the path building algorithm constructs the unique lattice path ${\mathcal{D}}(F)$ satisfying the conditions of Lemma~\ref{valley-lemma}. Therefore the algorithm never returns the statement that $F \in {\widehat{\sf Ass}}(a,b)$ is not contained in ${{\sf Ass}}(a,b)$ when in fact $F \in {{\sf Ass}}(a,b)$. On the other hand, if $F \in {\widehat{\sf Ass}}(a,b)$ and the path building algorithm terminates with a valid $a,b$-Dyck path ${\mathcal{D}}(F)$, it is obvious that $F$ is a face of ${\mathcal{F}}({\mathcal{D}}(F))$ so that $F \in {{\sf Ass}}(a,b)$. \end{proof} The path building algorithm gives a constructive way to decide whether a face $F \in {\widehat{\sf Ass}}(a,b)$ is contained in the subcomplex ${{\sf Ass}}(a,b)$. It can also be used to prove that ${{\sf Ass}}(a,b)$ is flag. \begin{proof} (of Proposition~\ref{flag}) Let $F$ be a collection of $a,b$-admissible diagonals which is not a face of ${{\sf Ass}}(a,b)$. We need to show that there exist two diagonals $d, d' \in F$ such that $\{d, d'\}$ is not a face of ${{\sf Ass}}(a,b)$. If there exist two crossing diagonals $d, d' \in F$ this is clear, so we may assume that $F$ is a face of ${\widehat{\sf Ass}}(a,b)$. We apply the path building algorithm to $F$. Since $F$ is not a face of ${{\sf Ass}}(a,b)$, the loop in step (1) of this algorithm must break for some value of $i$ such that $0 \leq i \leq b$. Since for any input $F$ the path building algorithm starts by adding $\lfloor \frac{b}{a} \rfloor + 1$ west steps to ${\mathcal{D}}(F)$, we must have that $0 \leq i \leq b - \lfloor \frac{b}{a} \rfloor - 1$. We call a diagonal $jk$ in $F$ \textbf{\textit{redundant}} if any of the following conditions hold: \begin{itemize} \item $j < i$, \item $j > i$ and there exists a diagonal $jk'$ in $F$ with $k' > k$, or \item $j > i$ and there exists a diagonal $j'k'$ in $F$ with $i < j' < j$ and $k' \geq k$. \end{itemize} We define a subset $\overline{F}$ of $F$ by \begin{equation} \overline{F} := \{ d \in F \,:\, \text{$d$ is not redundant} \}. \end{equation} We define a further subset $\overline{\overline{F}}$ of $\overline{F}$ by \begin{equation} \overline{\overline{F}} := \{ d \in \overline{F} \,:\, \text{$d$ is of the form $jk$ for $j > i$} \}. \end{equation} It is easy to see that the loop in the path building algorithm breaks at the same value $i$ for $\overline{F}$ as it did for $F$. When this loop breaks, the path ${\mathcal{D}}(\overline{F})$ is a lattice path starting at $(b,a)$ and ending at the northernmost lattice point of the line $x = i$ which is strictly below $y = \frac{a}{b} x$. Reading ${\mathcal{D}}(\overline{F})$ `from left to right', we can factor the portion of ${\mathcal{D}}(\overline{F})$ starting with an east step at $x = i$ into nonempty horizontal and vertical runs as $E^{i_1} N^{i_1} \dots E^{i_r} N^{i_r} E^{i_{r+1}}$. By the definition of a redundant diagonal, there are lasers fired off of this portion of ${\mathcal{D}}(\overline{F})$ only at the valleys $P_s$ between $E^{i_s}$ and $N^{i_s}$ for $s = 1, 2, \dots, r$. Moreover, the path building algorithm shows that $\overline{\overline{F}}$ is contained in ${{\sf Ass}}(a,b)$ and that ${\mathcal{D}}(\overline{\overline{F}})$ has the form $N^k E^i E^{i_1} N^{i_1} \dots E^{i_r} N^{i_r} E^{i_{r+1}}$. Let $Q$ denote the westernmost point of the horizontal step sequence $E^{i_1}$ (so that $Q$ is a lattice point on either of ${\mathcal{D}}(\overline{F})$ or ${\mathcal{D}}(\overline{\overline{F}})$). Since $b > a$, by slope considerations there are a total of $\lfloor \frac{a(b-i)}{b} \rfloor - 1$ distinct $a,b$-admissible diagonals of the form $i j$. Since $Q$ is below the vertical runs $N^{i_1}, \dots, N^{i_r}$ of lengths $i_1, \dots, i_r$, the correctness of the path building algorithm implies that there are exactly $\lfloor \frac{a(b-i)}{b} \rfloor - 1 - (i_1 + \dots + i_r)$ values of $j$ such that $i j$ is admissible and $\overline{\overline{F}} \cup \{i j\}$ is contained in ${{\sf Ass}}(a,b)$. (These values of $j$ correspond to the lattice points on the interior of the vertical segment between $Q$ and the line $y = \frac{a}{b}x$.) On the other hand, by coprimality, for $s = 1, 2, \dots, r$ the laser emanating from the valley $P_s$ gives exactly $i_s$ $a,b$-admissible diagonals of the form $i j$ such that $\overline{\overline{F}} \cup \{i j\}$ is not contained in ${{\sf Ass}}(a,b)$. To see this, consider the vertical line segment $L_s$ of length $i_s$ underlying the vertical run $N^{i_s}$ above the valley $P_s$. Let $L'_s$ be the vertical line segment on $x = i$ obtained by translating the points of $L_s$ southwest along a line of slope $\frac{a}{b}$. By coprimality, the interior of the line segment $L'_s$ contains exactly $i_s$ lattice points $P$ on the line $x = i$. The diagonals $d_{{\mathcal{D}}(\overline{\overline{F}})}(P)$ of these points (constructed with respect to the lattice path ${\mathcal{D}}(\overline{\overline{F}})$) cannot be added to $\overline{\overline{F}}$ to get a face in ${{\sf Ass}}(a,b)$. Graphically, the laser $\ell(P_s)$ `cuts off' these diagonals. By the definition of redundant diagonals, if $s \neq s'$ are the indices of two different valleys, the two sets of diagonals cut off by $\ell(P_s)$ and $\ell(P_{s'})$ are disjoint. This means that the set of all lasers $\ell(P_s)$, as $P_s$ varies over the valleys of ${\mathcal{D}}(\overline{F})$, cut off $i_1 + \cdots + i_r$ diagonals of the form $ij$. By the pigeonhole principle and the reasoning of the last two paragraphs, if $\overline{\overline{F}} \cup \{ij \}$ is in the complement ${\widehat{\sf Ass}}(a,b) - {{\sf Ass}}(a,b)$ for some diagonal $i j$ in $F$, there exists $1 \leq s \leq r$ such that the laser $\ell(P_s)$ cuts off the diagonal $ij$. It is straightforward to check using the path building algorithm that the $2$-element subset $\{ i j, d(P_s) \}$ of $F$ is not contained in ${{\sf Ass}}(a,b)$, so $F$ is not an empty face. \end{proof} \begin{example} As an example of the argument in the proof of Proposition~\ref{flag}, we let $(a, b) = (5,8)$ and $F = \{57, 48, 24, 04\}$. Since the diagonals in $F$ do not cross, we have that $F$ is a face of ${\widehat{\sf Ass}}(5,8)$. To check whether $F$ is a face of ${{\sf Ass}}(5,8)$, we apply the path building algorithm to $F$. The loop in step 1 of the path building algorithm applied to $F$ breaks at the $x$-coordinate $x = 0$, at which point ${\mathcal{D}}(F)$ is the following lattice path. We conclude that $F$ is not contained in ${{\sf Ass}}(5,8)$. \begin{center} \includegraphics[scale = 0.5]{flaglemmaone.pdf} \end{center} As before, the green lattice step which goes below the line $y= \frac{5}{8}x$ indicates that the path building algorithm breaks on the line $x = 0$. To form $\overline{F}$ from $F$, we remove all redundant diagonals from $F$. The only redundant diagonal is $57$, so we have $\overline{F} = \{48, 24, 04\}$. In terms of lattice paths, the redundancy of $57$ corresponds to the fact that the laser corresponding to $57$ in the above lattice path is entirely to the northwest of the laser corresponding to $48$. Applying the path building algorithm to $\overline{F}$ again leads to the loop in step $1$ breaking at $x = 0$, at which point ${\mathcal{D}}(\overline{F})$ is the following lattice path. Observe that the only lasers fired from ${\mathcal{D}}(\overline{F})$ which give rise to diagonals in $\overline{F}$ are fired from valleys. \begin{center} \includegraphics[scale = 0.5]{flaglemmatwo.pdf} \end{center} To form $\overline{\overline{F}}$ from $\overline{F}$, we remove all diagonals in $\overline{F}$ with smaller vertex $0$. Therefore, we have that $\overline{\overline{F}} = \{48, 24\}$. We have that $\overline{\overline{F}}$ is a face of ${{\sf Ass}}(5,8)$ and ${\mathcal{D}}(\overline{\overline{F}})$ is the above lattice path without the green step. We claim that there exist diagonals $d \in \overline{\overline{F}}$ and $d' \in \overline{F}$ such that the smaller boundary point of $d'$ is $0$ and $\{d, d'\}$ is not a face of ${{\sf Ass}}(5,8)$ Since $d, d' \in F$, this would imply that $F$ is not an empty face. To see this, let $P_1 = (2,2)$ and $P_2 = (4,3)$ be the valleys of the above lattice path ${\mathcal{D}}(\overline{F})$, read from right to left. By the definition of a redundant diagonal, the only lasers fired from ${\mathcal{D}}(\overline{F})$ are fired from the valleys of ${\mathcal{D}}(\overline{F})$ and the laser $\ell_{{\mathcal{D}}(\overline{F})}(P)$ fired from a valley $P$ hits ${\mathcal{D}}(\overline{F})$ on the first horizontal run to the northeast of $P$. We consider the vertical line segments $L_1$ and $L_2$ given by the vertical runs of ${\mathcal{D}}(\overline{F})$ which lie above the valleys $P_1$ and $P_2$. In our example, the segment $L_1$ has length $1$ and the segment $L_2$ has length $2$. We translate these segments along the line $y = \frac{5}{8}x$ until they lie on the line $x = 0$, obtaining the line segments $L_1'$ and $L_2'$. This is shown in the diagram below (where the line $x = 0$ is drawn in green). This diagram also shows that the various $L_s'$ line segments need not be disjoint - indeed, $L_1'$ is contained in $L_2'$. By coprimality, none of the $L_s'$ line segments has a lattice point on its boundary. \begin{center} \includegraphics[scale = 0.5]{flaglemmathree.pdf} \end{center} For any valley $P_s$, the laser $\ell(P_s)$ cuts off a single diagonal of the form $0j$ for each lattice point in the interior of $L_s'$, where $j$ is an $x$-coordinate of the first horizontal run to the northeast of $P_s$. Looking at the above diagram, we see that the laser $\ell(P_1)$ cuts of the diagonal $04$ (corresponding to the lattice point $(0,1)$ in the interior of $L_1'$) and $\ell(P_2)$ cuts off the diagonals $05$ and $07$ (corresponding to the lattice points $(0,2)$ and $(0,1)$ in the interior of $L_2'$, respectively). While the line segments $L_s'$ may intersect for various values of $s$, the lack of redundant edges in $\overline{F}$ forces the sets of diagonals cut off by the lasers $\ell(P_s)$ to be disjoint. This means that the lasers for all of the valleys $P_s$ cut off a total of $2 + 1 = 3$ diagonals. On the other hand, there are $\lfloor \frac{5(8-0)}{8} \rfloor - 1 = 4$ total $5,8$-admissible diagonals of the form $0j$ and only $\lfloor \frac{5(8-0)}{8} \rfloor - 1 - (2+1) = 1$ of these diagonals (namely, the diagonal $02$) is not cut off by one of the lasers. This single diagonal $02$ can be realized as a laser diagonal on ${\mathcal{D}}(\overline{F})$ by firing a laser from the $\lfloor \frac{5(8-0)}{8} \rfloor - 1 - (2+1) = 1$ lattice point (namely, the lattice point $(0,1)$) on the line $x = 0$ which lies on ${\mathcal{D}}(\overline{F})$. Since the path building algorithm breaks on $\overline{F}$ on the line $x = 0$, this means that $\overline{F}$ must contain a diagonal $0j$ cut off by one of the lasers $\ell(P_s)$. In fact, the face $\overline{F}$ contains the diagonal $04$, which is cut off by the laser $\ell(P_1)$. We can use the path building algorithm to see that $\{ d(P_1), 04 \} = \{24, 04\}$ is not a face of ${{\sf Ass}}(5,8)$, so that $F$ is not an empty face. \end{example} \section{Obstruction graphs} \label{Obstruction graphs} Proposition~\ref{flag} tells us that the obstructions preventing a face $F$ of ${\widehat{\sf Ass}}(a,b)$ from being a face of ${{\sf Ass}}(a,b)$ are local in nature. To keep track of these obstructions, we make use of a graph. \begin{defn} Let $a < b$ be coprime positive integers. The \textbf{\textit{obstruction graph}} ${\mathcal{OG}}(a,b)$ has vertex set given by the collection of $a,b$-admissible diagonals in ${\mathbb {P}}_{b+1}$ and an edge connecting two distinct $a,b$-admissible diagonals $d$ and $d'$ if $\{d, d'\}$ is contained in the complement ${\widehat{\sf Ass}}(a,b) - {{\sf Ass}}(a,b)$. \end{defn} By Proposition~\ref{flag}, a face $F \in {\widehat{\sf Ass}}(a,b)$ is contained in ${{\sf Ass}}(a,b)$ if and only if $F$ does not contain any edges of ${\mathcal{OG}}(a,b)$. In the language of deletions, we have that ${\mathrm {dl}}_{{\widehat{\sf Ass}}(a,b)}(E({\mathcal{OG}}(a,b))) = {{\sf Ass}}(a,b)$, where $E({\mathcal{OG}}(a,b))$ denotes the set of edges of the obstruction graph ${\mathcal{OG}}(a,b)$. \begin{example} When $b \equiv 1$ modulo $a$, the graph ${\mathcal{OG}}(a,b)$ has no edges and ${\widehat{\sf Ass}}(a,b) = {{\sf Ass}}(a,b)$. The obstruction graph ${\mathcal{OG}}(3,5)$ has as its vertex set the collection of $3,5$-admissible diagonals of ${\mathbb {P}}_6$ and edges $\{04 - 24, 15 - 35\}$. Figure~\ref{obstruction} shows the obstruction graph ${\mathcal{OG}}(5,8)$. The rectangles in this diagram correspond to a decomposition of ${\mathcal{OG}}(a,b)$ given in Lemma~\ref{same-larger-endpoint}. \end{example} \begin{figure} \centering \includegraphics[scale = 0.6]{obstruction.pdf} \caption{The obstruction graph ${\mathcal{OG}}(5,8)$.} \label{obstruction} \end{figure} Our strategy for proving that ${\widehat{\sf Ass}}(a,b)$ collapses onto ${{\sf Ass}}(a,b)$ is to give a total order $E({\mathcal{OG}}(a,b)) = \{e_1 \prec \cdots \prec e_N\}$ on the edge set of the obstruction graph. For $0 \leq r \leq N$, we define complexes ${{\sf Ass}}(a,b)_r$ inductively by ${{\sf Ass}}(a,b)_N := {\widehat{\sf Ass}}(a,b)$ and ${{\sf Ass}}(a,b)_r := {\mathrm {dl}}_{{{\sf Ass}}(a,b)_{r+1}}(e_{r+1})$ for $0 \leq r \leq N-1$. It follows that we have a chain of subcomplexes ${{\sf Ass}}(a,b)_0 \subset \cdots \subset {{\sf Ass}}(a,b)_N$ with ${{\sf Ass}}(a,b)_0 = {{\sf Ass}}(a,b)$ and ${{\sf Ass}}(a,b)_N = {\widehat{\sf Ass}}(a,b)$. We will prove that ${{\sf Ass}}(a,b)_r$ collapses onto ${{\sf Ass}}(a,b)_{r+1}$ for all $1 \leq r \leq N$. Before we define this order and to help us realize these collapsings, we will prove several results on the structure of obstruction graphs. A glance at Figure~\ref{obstruction} shows that if $\{ij, km\}$ is an edge of ${\mathcal{OG}}(5,8)$, then $j = m$. This is not a coincidence. Our main structural result about obstruction graphs is as follows. \begin{lemma} \label{same-larger-endpoint} Let $i j$ and $km$ be two $a,b$-admissible diagonals which form an edge of ${\mathcal{OG}}(a,b)$. We have that $j = m$. \end{lemma} \begin{proof} If $i = k$, we claim that the set $\{ ij, im\}$ is a face of ${{\sf Ass}}(a,b)$. Let $D$ be the $a,b$-Dyck path shown below which has a vertical run of maximal length at $x = i$. \begin{center} \includegraphics[scale = 0.5]{samestartingcontradiction.pdf} \end{center} Since $ij$ and $im$ are $a,b$-admissible, there must be points $P$ and $Q$ on $D$ with $x$-coordinate $i$ whose laser diagonals are $d(P) = ij$ and $d(Q) = im$. We conclude that $\{ij, im\} \subseteq {\mathcal{F}}(D)$ and $\{ij, im\}$ is a face of ${{\sf Ass}}(a,b)$. This is a contradiction. We conclude that $i \neq k$. Without loss of generality, we may assume that $i < k$. Since the diagonals $i j$ and $k m$ do not cross, we have that either $j \leq k$ or $m \leq j$. We argue first that $j \leq k$ is impossible. Suppose that $i < k$ and $j \leq k$. We claim that $\{ij, km\}$ is a face of ${{\sf Ass}}(a,b)$, which is a contradiction. Since $km$ is admissible, there exists an $a,b$-Dyck path $D_1$ of the form $D_1 = N^r E^k N^{a-r} E^{b-k}$ (for some $r > 0$) such that the laser fired from the unique valley of $D_1$ gives rise to the diagonal $km$. Similarly, there exists another $a,b$-Dyck path $D_2$ of the form $D_2 = N^s E^i N^{a-s} E^{b-i}$ such that the laser fired just after the $E^i$ horizontal run (which is empty if $i = 0$) gives rise to the diagonal $ij$. Let $D_3$ be the lattice path $D_3 = N^{r+s-a} E^i N^{a-s} E^{k-i} N^{a-r} E^{b-k}$. Geometrically, the path $D_3$ is obtained from the path $D_1$ by lowering the easternmost $i$ steps of the horizontal run $E^k$ by $a-s$ units. This procedure gives rise to a valid $a,b$-Dyck path $D_3$ (i.e., the lattice path $D_3$ stays above the line $y = \frac{a}{b} x$) because the laser fired from the point just after the horizontal run $E^i$ on $D_3$ must hit $D_3$ in the step on the horizontal run $E^{k-i}$ whose right endpoint has $x$-coordinate $j$ (here we use $j \leq k$ and the construction of $D_3$). Since $D_1$ is an $a,b$-Dyck path, we conclude that this laser is above the line $y = \frac{a}{b} x$ and $D_3$ is also an $a,b$-Dyck path. But now $ij$ and $km$ are both laser diagonals arising from $D_3$, so that $\{ij, km\} \subseteq {\mathcal{F}}(D_3)$, which is a contradiction. By the above two paragraphs, we have that $i < k$ and $m \leq j$. We will argue that $j \leq m$ as well. Let $P'$ be the unique lattice point with $x$-coordinate $i$ such that the laser $\ell(P')$ of slope $\frac{a}{b}$ hits the horizontal line $y = a$ at the horizontal step whose right endpoint has $x$-coordinate $j$. Similarly, let $Q'$ be the unique lattice point with $x$-coordinate $j$ such that the laser $\ell(Q')$ hits the horizontal line $y = a$ at the horizontal step whose right endpoint has $x$-coordinate $m$. If the $y$-coordinate of $P'$ is greater than or equal to the $y$-coordinate of $Q'$, then clearly $j \leq m$. So we may assume that the $y$-coordinate of $P'$ is strictly less than the $y$-coordinate of $Q'$. Let $D'$ be the $a,b$-Dyck path shown below with lasers fired from and valleys at $P'$ and $Q'$. We extend the laser $\ell(P')$ in a dashed fashion to the line $y = b$. \begin{center} \includegraphics[scale = 0.5]{sameending.pdf} \end{center} We know that the right endpoint of the step containing $R$ has $x$-coordinate $j$ and that the right coordinate of the step containing $S$ has $x$-coordinate $m$. Since $\{ij, km\}$ is not a face of ${{\sf Ass}}(a,b)$, the laser $\ell(P')$ must hit the horizontal run of $D'$ which lies to the right of $Q'$. This means that the laser $\ell(P')$ (and its extension) must lie to the left of $\ell(Q')$. This implies that $j \leq m$. \end{proof} Lemma~\ref{same-larger-endpoint} implies that the obstruction graph ${\mathcal{OG}}(a,b)$ breaks up into components ${\mathcal{OG}}(a,b)_0 \uplus {\mathcal{OG}}(a,b)_1 \uplus \dots \uplus {\mathcal{OG}}(a,b)_b$, where ${\mathcal{OG}}(a,b)_m$ has edges consisting of two diagonals whose common larger endpoint is $m$. Figure~\ref{obstruction} gives these components as rectangles. Observe that these components may be disconnected (for example, the component ${\mathcal{OG}}(5,8)_6$ is disconnected) and may be empty (for example, the components ${\mathcal{OG}}(5,8)_0$ and ${\mathcal{OG}}(5,8)_1$ are empty). Although we will not need this result, we remark that the obstruction graph ${\mathcal{OG}}(a,b)$ is determined by the component ${\mathcal{OG}}(a,b)_b$. If $d = ij$ is a diagonal in ${\mathbb {P}}_{b+1}$ and $k \geq 0$ is such that $i-k \geq 0$, we let $d - k$ denote the diagonal $(i-k)(j-k)$ in ${\mathbb {P}}_{b+1}$. If $i-k < 0$, we leave $d - k$ undefined. We define $d + k$ similarly. \begin{lemma} \label{determined-by-largest-component} For $a < b$ coprime and $0 \leq m \leq b$, the component ${\mathcal{OG}}(a,b)_m$ of the obstruction graph ${\mathcal{OG}}(a,b)$ has vertex set given by the diagonals \begin{equation*} \{ d - (b-m) \,:\, \text{$d$ a vertex of ${\mathcal{OG}}(a,b)_b$ such that $d - (b-m)$ is defined} \}. \end{equation*} The edge set of ${\mathcal{OG}}(a,b)_m$ is given by \begin{equation*} \{ \{d, d'\} \,:\, \text{$\{d + (b-m) , d' + (b-m)\}$ is an edge of ${\mathcal{OG}}(a,b)_b$} \}. \end{equation*} \end{lemma} In other words, to form ${\mathcal{OG}}(a,b)_m$ from ${\mathcal{OG}}(a,b)_{m+1}$, we subtract $1$ from all the boundary points involved in the vertex diagonals and erase any vertex diagonals (together with their incident edges) which involve negative boundary points. This subtraction and deletion process can be seen from right to left in Figure~\ref{obstruction}. \begin{proof} (Sketch.) The statement about the vertex set of ${\mathcal{OG}}(a,b)_m$ follows because the set of $a,b$-admissible diagonals is closed under rotation of ${\mathbb {P}}_{b+1}$. The statement about the edge set of ${\mathcal{OG}}(a,b)_m$ follows by considering `east-west translations' of Dyck paths of the form \begin{equation*} N^{i_1} E^{j_1} \dots N^{i_r} E^{j_r} \mapsto N^{i_1} E^{j_1 - 1} N^{i_2} E_{j_2} \dots N^{i_{r-1}} E^{j_{r-1}} N^{i_r} E^{j_r + 1}. \end{equation*} \end{proof} When drawn on ${\mathbb {P}}_{b+1}$, the edges $\{im, km\}$ belonging to ${\mathcal{OG}}(a,b)_m$ look like wedges whose common point is $m$ such that $m$ is the largest boundary point on the wedge. The next two lemmas concern the more general situation of a wedge $\{im, km\}$ of $a,b$-admissible diagonals with $i < k < m$, whether or not $\{im, km\}$ is an obstructing edge. We want to develop some sufficient conditions which guarantee that the completion $ik$ of the triangle is an $a,b$-admissible diagonal. A schematic of this situation is shown below. \begin{center} \includegraphics[scale = 0.5]{schematic.pdf} \end{center} We are interested in whether $ik$ is $a,b$-admissible because these diagonals will be used to locate cone vertices as in Lemma~\ref{cone-vertex-lemma}. It is not always the case that $ik$ is admissible. For example, if $(a,b) = (5,8)$, then $\{05, 15\}$ is a wedge of $5,8$-admissible diagonals whose common point is the maximal point but $01$ is not even a diagonal in ${\mathbb {P}}_9$, let alone $5,8$-admissible. Our first lemma states that if $\{im, km\}$ is an edge of ${\mathcal{OG}}(a,b)$, then $ik$ is an $a,b$-admissible diagonal, so that the dotted edge is an $a,b$-admissible diagonal in the diagram above. \begin{lemma} \label{completing-wedges} Let $i m$ and $k m$ be two $a,b$-admissible diagonals such that $i < k < m$ and $\{i m, km\}$ is an edge of ${\mathcal{OG}}(a,b)$. We have that $ik$ is an $a,b$-admissible diagonal. \end{lemma} \begin{proof} We consider the second $a,b$-Dyck path $D'$ drawn in the proof of Lemma~\ref{same-larger-endpoint}. As was deduced in the proof of Lemma~\ref{same-larger-endpoint}, we know that $R$ and $S$ lie on the same horizontal step (namely, the horizontal step with right endpoint having $x$-coordinate $m$). In particular, the points $R$ and $S$ are $< 1$ unit apart. This means that the horizontal distance between the lasers $\ell(P')$ and $\ell(Q')$ is $< 1$ unit. In particular, the (solid portion of the) laser $\ell(P')$ hits $D'$ at the horizontal step with left endpoint $Q'$. Since the $x$-coordinate of $P'$ is $i$ and the $x$-coordinate of $Q'$ is $k$, we conclude that $ik$ is an $a,b$-admissible diagonal. \end{proof} While Lemma~\ref{completing-wedges} will give us some of the cone vertices we will need to perform our collapsing, we will need more. These additional vertices will be provided by the following result. \begin{lemma} \label{completing-half-wedges} Let $i m$ and $j m$ be two $a,b$-admissible diagonals such that $\{i m, j m\}$ is an edge of ${\mathcal{OG}}(a,b)$. Suppose that $i < k < j$ and $k m$ is an $a,b$-admissible diagonal such that $\{k m, j m\}$ is \textbf{not} an edge of ${\mathcal{OG}}(a,b)$. Then $\{i m, km\}$ is an edge of ${\mathcal{OG}}(a,b)$ and $i k$ is an $a,b$-admissible diagonal \end{lemma} \begin{center} \includegraphics[scale = 0.5]{schematictwo.pdf} \end{center} The situation in Lemma~\ref{completing-half-wedges} is described in the schematic above. We assume that the big wedge $\{i m, j m\}$ \textbf{is} an edge in ${\mathcal{OG}}(a,b)$ and that the smaller left wedge $\{km, jm\}$ \textbf{is not} an edge in ${\mathcal{OG}}(a,b)$. The conclusion is that the dashed segment is an $a,b$-admissible diagonal. \begin{proof} The idea is to consider a Dyck path $D$ with valleys at $i, j,$ and $k$. As in the proof of Lemma~\ref{completing-wedges}, we will get that $i k$ appears as a laser diagonal, and is therefore admissible. More precisely, let $D$ be the following Dyck path. \begin{center} \includegraphics[scale = 0.7]{halfwedge.pdf} \end{center} The valleys $P, Q,$ and $R$ lie on the vertical lines $x = i$, $x = j$, and $x = k$, respectively. The point $R$ is chosen so that $d(R) = jm$. The point $Q$ is chosen so that $d(Q) = km$. (We know that $\ell(Q)$ does not hit $D$ in the horizontal run to the left of $R$ because $\{km, jm\}$ is not an edge of ${\mathcal{OG}}(a,b)$.) Finally, the point $P$ is chosen so that the `extended laser' from $P$ (shown here as a dashed line) hits the line $y = a$ at the lattice step whose right endpoint has $x$-coordinate $m$. Since $\{i m, jm \}$ is an edge of ${\mathcal{OG}}(a,b)$, we know that $\ell(P)$ must hit $D$ strictly below the line $y = a$. Since lasers are parallel, this forces $\ell(P)$ to hit $D$ on a step of the horizontal run to the left of $Q$. We claim that the right endpoint of this step is $Q$. So see this, let $r$ denote the total distance along the line $y = a$ between the `extended' version of the laser $\ell(P)$ and the laser $\ell(R)$. Since the lasers $\ell(Q)$, $\ell(R)$, and the extended version of $\ell(P)$ intersect the horizontal line $y = a$ at a lattice step whose right endpoint has $x$-coordinate $m$, we have $r < 1$. But $r$ is also the distance along the horizontal line going through $Q$ between the point where the unextended laser $\ell(P)$ hits $D$ and $Q$. This implies that $\ell(P)$ hits $D$ at the lattice step with left endpoint $Q$ and that $ik$ is $a,b$-admissible. To see that $\{i m, km\}$ is an edge of ${\mathcal{OG}}(a,b)$, consider the path $D'$ whose unique valleys are at the lattice points $P$ and $Q$. That is, the path $D'$ looks like the path $D$ above with the valley at $R$ `folded out'. The laser $\ell(P)$ hits $D'$ in the horizontal run to the left of $Q$. By the path building algorithm, we conclude that $\{i m, km\}$ is an edge of ${\mathcal{OG}}(a,b)$. \end{proof} \section{Collapsing} \label{Collapsing} In this section we describe how to collapse the complex ${\widehat{\sf Ass}}(a,b)$ onto the complex ${{\sf Ass}}(a,b)$. By Proposition~\ref{flag}, we know that a face $F$ of ${\widehat{\sf Ass}}(a,b)$ is also a face of ${{\sf Ass}}(a,b)$ if and only if $F$ does not contain any edges of the obstruction graph ${\mathcal{OG}}(a,b)$. A first approximation of our collapse is performed by putting a total order on the edges of ${\mathcal{OG}}(a,b)$ and repeatedly using Lemma~\ref{cone-vertex-lemma} to collapse all faces $F$ of ${\widehat{\sf Ass}}(a,b)$ containing a given edge but none of the previous edges. However, it will turn out that some of these `subcollapses' indexed by edges of ${\mathcal{OG}}(a,b)$ must be further broken up as a sequence of smaller collapses when the wedges on ${\mathbb {P}}_{b+1}$ corresponding to the edges of ${\mathcal{OG}}(a,b)$ are `too wide'. We begin by describing our total order on the edges of ${\mathcal{OG}}(a,b)$. This is essentially an iterated version of lexicographical order on the $a,b$-admissible diagonals. \begin{defn} Given two $a,b$-admissible diagonals $d = i j$ and $d' = i' j'$, we say that $d \sqsubseteq d'$ if $j < j'$ or if $j = j'$ and $i \leq i'$. Given two edges $e_1 = \{d_1 \sqsubset d'_1\}$ and $e_2 = \{d_2 \sqsubset d_2'\}$ of ${\mathcal{OG}}(a,b)$, we say that $e_1 \preceq e_2$ if $d_1 \sqsubset d_2$ or $d_1 = d_2$ and $d'_1 \sqsubseteq d'_2$. \end{defn} \begin{example} \label{orderexample} If $(a,b) = (5,8)$, the order $\prec$ on the edges of ${\mathcal{OG}}(a,b)$ is given by \begin{equation*} \{04, 24\} \prec \{15, 35\} \prec \{26, 46\} \prec \{07, 27\} \prec \{07, 57\} \prec \{37, 57\} \prec \{18, 38\} \prec \{18, 68\} \prec \{48, 68\}. \end{equation*} \end{example} Let $\{e_1, \dots, e_N\}$ be the complete set of edges of ${\mathcal{OG}}(a,b)$ indexed so that $e_1 \prec \dots \prec e_N$. The order $\preceq$ can be used to define a family of simplicial complexes. \begin{defn} For $0 \leq r \leq N$, inductively define a family of complexes ${{\sf Ass}}(a,b)_r$ by ${{\sf Ass}}(a,b)_N := {\widehat{\sf Ass}}(a,b)$ and ${{\sf Ass}}(a,b)_r := {\mathrm {dl}}_{{{\sf Ass}}(a,b)_{r+1}}(e_{r+1})$ for $0 \leq r \leq N-1$. \end{defn} In other words, we have that \begin{equation*} {{\sf Ass}}(a,b)_r = \{F \in {\widehat{\sf Ass}}(a,b) \,:\, \text{$F$ does not contain any edges $e_s$ of ${\mathcal{OG}}(a,b)$ for $s > r$} \} \end{equation*} We are ready to prove Theorem~\ref{main}. \begin{proof} (of Theorem~\ref{main}) By the definition of the complexes ${{\sf Ass}}(a,b)_r$, we have that ${{\sf Ass}}(a,b)_0 = {{\sf Ass}}(a,b)$ and ${{\sf Ass}}(a,b)_N = {\widehat{\sf Ass}}(a,b)$. Fix $1 \leq r \leq N$. To prove that ${\widehat{\sf Ass}}(a,b)$ collapses onto ${{\sf Ass}}(a,b)$ it is enough to show that ${{\sf Ass}}(a,b)_r$ collapses onto ${{\sf Ass}}(a,b)_{r-1}$. Let $e_r = \{d \sqsubset d'\}$, where $d = ik$ and $d' = jk$ are the two $a,b$-admissible diagonals in ${\mathbb {P}}_{b+1}$ which form the edge $e_r$. By the definition of $\sqsubset$ we have $i < j < k$. Since $e_r$ does not contain any other edges of ${\mathcal{OG}}(a,b)$, we know that $e_r$ is a face of ${\mathcal{OG}}(a,b)_r$. Moreover, Lemma~\ref{completing-wedges} guarantees that the segment $ij$ is an $a,b$-admissible diagonal, so that $e_r \cup \{ij\} = \{ij, ik, jk\}$ is a face of ${\widehat{\sf Ass}}(a,b)$. In fact, since the only edge of ${\mathcal{OG}}(a,b)$ contained in $e_r \cup \{ij\}$ is $e_r$, we have that $e_r \cup \{ij\}$ is a face of ${{\sf Ass}}(a,b)_r$. It is tempting to hope that $ij$ is a cone vertex for $e_r$ in ${{\sf Ass}}(a,b)_r$, so that we could apply Lemma~\ref{cone-vertex-lemma} to collapse ${{\sf Ass}}(a,b)_r$ onto ${\mathrm {dl}}_{{{\sf Ass}}(a,b)_r}(e_r) = {{\sf Ass}}(a,b)_{r-1}$. Unfortunately, the diagonal $ij$ is not necessarily a cone vertex for $e_r$ in ${{\sf Ass}}(a,b)_r$ because there may exist faces $F$ of ${{\sf Ass}}(a,b)_r$ containing $e_r$ which contain diagonals which cross $ij$. To get around this problem, we will have to be more delicate and realize the collapsing of ${{\sf Ass}}(a,b)_r$ onto ${{\sf Ass}}(a,b)_{r-1}$ as a sequence of smaller collapsings indexed by a carefully chosen sequence of these `crossing faces' $F$. Let $F$ be a face of ${{\sf Ass}}(a,b)_r$ containing $e_r$ and suppose that $F$ contains an $a,b$-admissible diagonal $s m$ which crosses the diagonal $ij$. Since $F$ contains the wedge $\{ik, jk\}$ and the diagonals in $F$ do not cross, we have that $m = k$ and $i < s < j$. If $\{sk, jk\}$ were an edge of ${\mathcal{OG}}(a,b)$, we would have that $\{ik, jk\} \prec \{sk, jk\} \subset F$, which contradicts the assumption that $F$ is a face of ${{\sf Ass}}(a,b)_r$. We conclude that $\{sk, jk\}$ is not an edge of ${\mathcal{OG}}(a,b)$. Let $s_1 < s_2 < \dots < s_p$ be a complete list of the indices $i < s_q < j$ such that \begin{itemize} \item the segment $s_q k$ is an $a,b$-admissible diagonal and \item the pair $\{s_q k, jk\}$ is not an edge of ${\mathcal{OG}}(a,b)$. \end{itemize} It is possible that $p = 0$, in which case the rest of the argument simplifies. For $1 \leq q \leq p$, define a face $F_q$ of ${\widehat{\sf Ass}}(a,b)$ by $F_q = \{ik, s_q k, jk\}$. For all $1 \leq q \leq p$ the only edge of ${\mathcal{OG}}(a,b)$ contained in $F_q$ is $e_r$, so $F_q$ is a face of ${{\sf Ass}}(a,b)_r$. For $1 \leq q \leq p+1$, we define a subcomplex ${{\sf Ass}}(a,b)^{(q)}_r$ of ${{\sf Ass}}(a,b)_r$ by \begin{align*} {{\sf Ass}}(a,b)_r^{(q)} &:= {\mathrm {dl}}_{{{\sf Ass}}(a,b)_r}( \{F_1, F_2, \dots, F_{q-1}\} ) \\ &= \{ F \in {{\sf Ass}}(a,b)_r \,:\, \text{$F$ does not contain $F_{q'}$ for any $q' < q$} \}. \end{align*} In particular, we have that ${{\sf Ass}}(a,b)_r^{(1)} = {{\sf Ass}}(a,b)_r$ and that any face of ${{\sf Ass}}(a,b)_r^{(p+1)}$ containing $e_r$ does not contain any of the diagonals $s_q k$ for $1 \leq q \leq p$. Moreover, we have that ${{\sf Ass}}(a,b)_r^{(q+1)} = {\mathrm {dl}}_{{{\sf Ass}}(a,b)_r^{(q)}}(F_q)$ for $1 \leq q \leq p$. Fix $1 \leq q \leq p$. We claim that ${{\sf Ass}}(a,b)_r^{(q)}$ collapses onto ${{\sf Ass}}(a,b)_r^{(q+1)}$. To see this, observe that $F_q = \{ik, s_q k, jk\}$, where $i < s_q < j < k$, $\{ik, jk\}$ is an edge of ${\mathcal{OG}}(a,b)$, and $\{s_q k, jk\}$ is not an edge of ${\mathcal{OG}}(a,b)$. Applying Lemma~\ref{completing-half-wedges}, we get that $i s_q$ is an $a,b$-admissible diagonal in ${\mathbb {P}}_{b+1}$. We claim that the $a,b$-admissible diagonal $i s_q$ is a cone vertex for $F_q$ in the complex ${{\sf Ass}}(a,b)_r^{(q)}$. Since $i s_q \notin F_q$, by the definition of cone vertices we need to show that if $F$ is any face of ${{\sf Ass}}(a,b)_r^{(q)}$ containing $F_q$, we have that $F \cup \{i s_q\}$ is also a face of ${{\sf Ass}}(a,b)_r^{(q)}$. We begin by showing that $F \cup \{i s_q\}$ is a face of ${\widehat{\sf Ass}}(a,b)$, that is, that the diagonals in $F \cup \{i s_q\}$ are noncrossing on ${\mathbb {P}}_{b+1}$. Indeed, since $F$ contains $F_q = \{ik, i s_q, jk\}$, the only $a,b$-admissible diagonals in $F$ which could cross $i s_q$ would be of the form $i s_{q'}$ for $q' < q$. If any of these diagonals were contained in $F$, then $F$ would contain $F_{q'} = \{ik, i s_{q'}, jk\}$, which contradicts the assumption that $F$ is a face of ${{\sf Ass}}(a,b)_r^{(q)}$. We conclude that the diagonals in $F \cup \{i s_q\}$ are noncrossing and that $F \cup \{i s_q\}$ is a face of ${\widehat{\sf Ass}}(a,b)$. We still need to show that $F \cup \{i s_q\}$ is a face of the subcomplex ${{\sf Ass}}(a,b)_r^{(q)}$ of ${\widehat{\sf Ass}}(a,b)$. This amounts to showing that $F \cup \{i s_q\}$ does not contain \begin{enumerate} \item any edge $e_{r'}$ of ${\mathcal{OG}}(a,b)$ with $r' > r$, or \item any face $F_{q'}$ with $q' < q$. \end{enumerate} For (1), since $F$ is a face of ${{\sf Ass}}(a,b)_r^{(q)}$, we know that $F$ does not contain any edge $e_{r'}$ of ${\mathcal{OG}}(a,b)$ with $r' > r$. What possible edges of ${\mathcal{OG}}(a,b)$ could be added to $F$ by introducing the diagonal $i s_q$? Only edges of the form $e_{r'} = \{i' s_q, i s_q\}$ which belong to the component ${\mathcal{OG}}(a,b)_{s_q}$. But by the definition of $\prec$ and the fact that $s_q < k$, we know that any such edge would satisfy $e_{r'} \prec e_r$, so $r' < r$. This implies (1). For (2), we make the simple observation that $i s_q$ is not of the form $s_{q'} k$, so $F \cup \{i s_q\}$ does not contain $F_{q'}$ for $q' < q$ because $F$ does not contain $F_{q'}$ for $q' < q$. By the last two paragraphs, we have that $i s_q$ is a cone vertex for $F_q$ in ${{\sf Ass}}(a,b)^{(q)}_r$. By Lemma~\ref{cone-vertex-lemma} and induction, we get that ${{\sf Ass}}(a,b)_r = {{\sf Ass}}(a,b)^{(1)}$ collapses onto ${{\sf Ass}}(a,b)_r^{(p+1)}$. To finish the proof of Theorem~\ref{main}, we will show that ${{\sf Ass}}(a,b)_r^{(p+1)}$ collapses onto ${{\sf Ass}}(a,b)_{r-1}$. We begin with the observation that $e_r$ is a face of ${{\sf Ass}}(a,b)_r^{(p+1)}$ and that we have ${{\sf Ass}}(a,b)_{r-1} = {\mathrm {dl}}_{{{\sf Ass}}(a,b)_r^{(p+1)}}(e_r)$. By Lemma~\ref{completing-wedges} we get that $ij$ is an $a,b$-admissible diagonal in ${\mathbb {P}}_{b+1}$. We claim that $ij$ is a cone vertex for $e_r$ in ${{\sf Ass}}(a,b)_r^{(p+1)}$. To see this, let $F$ be any face of ${{\sf Ass}}(a,b)_r^{(p+1)}$ containing $e_r$. Since $ij \notin e_r$, it suffices to show that $F \cup \{ij\}$ is also a face of ${{\sf Ass}}(a,b)_r^{(p+1)}$. As before, we begin by showing that $F \cup \{ij\}$ is a face of ${\widehat{\sf Ass}}(a,b)$, i.e., that the diagonals in $F \cup \{ij\}$ do not cross. Indeed, since $F$ contains the wedge $e_r = \{ik, jk\}$, the only diagonals in $F$ which could cross $ij$ would be $a,b$-admissible diagonals of the form $sk$ for $i < s < j$. But since $F$ is a face of ${{\sf Ass}}(a,b)_r^{(p+1)} \subseteq {{\sf Ass}}(a,b)_r$, the definition of $\prec$ would force the pair $\{sk, jk\}$ to be a non-edge in ${\mathcal{OG}}(a,b)$. This means that $s = s_q$ for some $1 \leq q \leq p$ and $F$ would contain $F_q = \{ik, s_qk, jk\}$. But this contradicts that fact that $F$ is a face of ${{\sf Ass}}(a,b)_r^{(p+1)}$. We conclude that none of the diagonals in $F$ cross $ij$ and $F \cup \{ij\}$ is a face of ${\widehat{\sf Ass}}(a,b)$. We still need to show that $F \cup \{ij\}$ is a face of the subcomplex ${{\sf Ass}}(a,b)_r^{(p+1)}$ of ${\widehat{\sf Ass}}(a,b)$. This amounts to showing that $F \cup \{ij\}$ does not contain \begin{enumerate} \item any edge $e_{r'}$ of ${\mathcal{OG}}(a,b)$ with $r' > r$, or \item any face $F_q$ for $1 \leq q \leq p$. \end{enumerate} For (1), observe that the addition of $ij$ to $F$ can only add edges of ${\mathcal{OG}}(a,b)$ of the form $\{i'j, ij\}$ which belong to the component ${\mathcal{OG}}(a,b)_j$ and since $j < k$ any such edge is $\prec e_r$. For (2), observe that $ij \notin F_q$ for $1 \leq q \leq p$. By the last two paragraphs, we have that $ij$ is a cone vertex for $e_r$ in ${{\sf Ass}}(a,b)_r^{(p+1)}$. By Lemma~\ref{cone-vertex-lemma}, we have that ${{\sf Ass}}(a,b)_r^{(p+1)}$ collapses onto ${{\sf Ass}}(a,b)_{r-1}$. We already showed that ${{\sf Ass}}(a,b)_r$ collapses onto ${{\sf Ass}}(a,b)_r^{(p+1)}$, so we have the overall collapse of ${{\sf Ass}}(a,b)_r$ onto ${{\sf Ass}}(a,b)_{r-1}$. By induction, we have that ${\widehat{\sf Ass}}(a,b) = {{\sf Ass}}(a,b)_N$ collapses onto ${{\sf Ass}}(a,b) = {{\sf Ass}}(a,b)_0$. This completes the proof of Theorem~\ref{main}. \end{proof} \begin{example} \label{collapseexample} To better understand the argument used in the proof of Theorem~\ref{main}, we show how to collapse ${\widehat{\sf Ass}}(5,8)$ onto ${{\sf Ass}}(5,8)$. The obstruction graph ${\mathcal{OG}}(5,8)$ has $N = 9$ edges and the order presented in Example~\ref{orderexample} labels these edges as $e_1 \prec \dots \prec e_9$. We have that ${{\sf Ass}}(5,8)_9 = {\widehat{\sf Ass}}(5,8)$ and ${{\sf Ass}}(5,8)_0 = {{\sf Ass}}(5,8)$. To show that ${{\sf Ass}}(5,8)_9$ collapses onto ${{\sf Ass}}(5,8)_8$, we use the following diagram. \begin{center} \includegraphics[scale = 0.5]{collapseone.pdf} \end{center} The diagonals in $e_9 = \{48, 68\}$ are solid. The dotted diagonal $46$ (which is guaranteed to be $5,8$-admissible by Lemma~\ref{completing-wedges}) is a cone vertex for $e_9$ in ${{\sf Ass}}(5,8)_9$ because no face of ${{\sf Ass}}(5,8)_9$ containing $e_9$ contains a diagonal which crosses $46$. Applying Lemma~\ref{cone-vertex-lemma}, we see that ${{\sf Ass}}(5,8)_9$ collapses onto ${{\sf Ass}}(5,8)_8$. The argument that ${{\sf Ass}}(5,8)_8$ collapses onto ${{\sf Ass}}(5,8)_7$ is more complicated. If we try to use the above strategy directly, we encounter the following diagram. \begin{center} \includegraphics[scale = 0.5]{collapsetwo.pdf} \end{center} The diagonals in $e_8 = \{18, 68\}$ are solid. Lemma~\ref{completing-wedges} guarantees that the dotted diagonal $16$ is $5,8$-admissible. However, the diagonal $16$ is not a cone vertex for $e_8$ in ${{\sf Ass}}(5,8)_8$ because it crosses the diagonal $38$. (The diagonal $48$ is not a problem because any face of ${{\sf Ass}}(5,8)_8$ containing $e_8$ cannot contain $48$, as such a face would then contain $e_9$. In other words, the diagonal $48$ forms an obstructing edge with the diagonal $68$.) To get around this problem, we define a face $F_1$ of ${{\sf Ass}}(5,8)_8$ by $F_1 := \{18, 38, 68\}$. We have the following diagram. \begin{center} \includegraphics[scale = 0.5]{collapsethree.pdf} \end{center} The diagonals in $F_1$ are solid. Lemma~\ref{completing-half-wedges} guarantees that the dotted diagonal $13$ is $5,8$-admissible and it is the case that $13$ is a cone vertex for $F_1$ in ${{\sf Ass}}(5,8)_8$. By Lemma~\ref{cone-vertex-lemma}, the complex ${{\sf Ass}}(5,8)_8 = {{\sf Ass}}(5,8)_8^{(1)}$ collapses onto the complex ${\mathrm {dl}}_{{{\sf Ass}}(5,8)_8}(F_1) = {{\sf Ass}}(5,8)_8^{(2)}$. Working inside the complex ${{\sf Ass}}(5,8)_8^{(2)}$, we can revert to our previous argument. We have the following diagram. \begin{center} \includegraphics[scale = 0.5]{collapsefour.pdf} \end{center} The diagonal $38$ is no longer present because no face of ${{\sf Ass}}(5,8)_8^{(2)}$ contains $F_1$. We can apply Lemmas~\ref{completing-wedges} to conclude that $16$ is $5,8$-admissible and observe that $16$ is a cone vertex for $e_8$ in ${{\sf Ass}}(5,8)_8^{(2)}$. We conclude that ${{\sf Ass}}(5,8)_8^{(2)}$ collapses onto ${{\sf Ass}}(5,8)_7$. The proof that ${{\sf Ass}}(5,8)_7$ collapses onto ${{\sf Ass}}(5,8)_6$ is easier and relies on the following diagram. \begin{center} \includegraphics[scale = 0.5]{collapsefive.pdf} \end{center} Since the obstructing edge $e_7$ forms a sufficiently narrow wedge $\{18, 38\}$ in ${\mathbb {P}}_9$, there is no room for a $5,8$-admissible diagonal to get in the way of $13$ being a cone vertex. At this point, we have collapsed all the faces of ${\widehat{\sf Ass}}(5,8)$ which contain an obstructing edge involving the wedge boundary point $8$. We move along to those obstructing edges involving the wedge boundary point $7$. This process begins by considering the following diagram to give the collapse of ${{\sf Ass}}(5,8)_6$ onto ${{\sf Ass}}(5,8)_5$. \begin{center} \includegraphics[scale = 0.5]{collapsesix.pdf} \end{center} The dotted vertex $35$ plays the role of the cone vertex and we repeat the sequence of arguments above, rotated one boundary point counterclockwise. The collapses corresponding to the obstructing edges with wedge boundary points $6,5,4$ are easier and complete the collapsing of ${\widehat{\sf Ass}}(5,8)$ onto ${{\sf Ass}}(5,8)$. Observe that the dotted diagonals considered when collapsing from ${{\sf Ass}}(5,8)_r$ to ${{\sf Ass}}(5,8)_{r-1}$ never create an obstructing edge with index $> r$, so really are cone vertices inside the complex ${{\sf Ass}}(5,8)_r$. \end{example} We reiterate the deduction of Corollary~\ref{alexander} from Theorem~\ref{main} presented in \cite{ARW}. \begin{proof} (of Corollary~\ref{alexander}) An argument using the coprimality of $a$ and $b$ and the sets $S(a,b)$ and $S(b-a,b)$ shows that any diagonal of ${\mathbb {P}}_{b+1}$ is either $a,b$-admissible or $(b-a,b)$-admissible, but not both. In other words, the vertex sets of ${\widehat{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(b-a,b)$ partition the vertex set of the simplicial sphere ${{\sf Ass}}(b-1,b)$. By the definition of ${\widehat{\sf Ass}}$, a subset $F$ of ${\widehat{\sf Ass}}(a,b)$ is a face of ${\widehat{\sf Ass}}(a,b)$ if and only if $F$ is also a face of ${{\sf Ass}}(b-1,b)$ and similarly for ${\widehat{\sf Ass}}(b-a,b)$. This means that the complement of ${\widehat{\sf Ass}}(a,b)$ within ${{\sf Ass}}(b-1,b)$ deformation retracts onto ${\widehat{\sf Ass}}(b-a,b)$. In other words, we have that ${\widehat{\sf Ass}}(a,b)$ and ${\widehat{\sf Ass}}(b-a,b)$ are Alexander dual within the sphere ${{\sf Ass}}(b-1,b)$. By Theorem~\ref{main}, we also have that ${{\sf Ass}}(a,b)$ and ${{\sf Ass}}(b-a,b)$ are Alexander dual within the sphere ${{\sf Ass}}(b-1,b)$. \end{proof} \section{Closing remarks} \label{Open} The construction $(a, b) \leadsto {{\sf Ass}}(a,b)$ attaches a simplicial complex to every pair of coprime positive integers $a < b$. In this paper we proved that the duality $(a, b) \longleftrightarrow (b-a,b)$ on pairs of coprime integers manifests itself topologically as Alexander duality of simplicial complexes. Any increasing pair of coprime positive integers an be obtained from $(1, 2)$ by repeated applications of the duality $(a,b) \mapsto (b-a, b)$ and the map $(a, b) \mapsto (b, 2b-a)$. Alternatively, we could use repeated applications of the duality $(a,b) \mapsto (b-a,b)$ and the map $(a,b) \mapsto (a, a+b)$. It may be interesting to find a topological manifestation of the maps $(a,b) \mapsto (b, 2b-a)$ and $(a,b) \mapsto (a, a+b)$ in terms of rational associahedra. These maps are closely related to the the monoid endomorphisms of the set $\{x, y\}^*$ of words on the two-letter alphabet $\{x, y\}$ which generate the so-called `Christoffel morphisms' (see \cite{BLRS}). In the notation of \cite{BLRS}, the duality $(a,b) \mapsto (b-a,b)$ corresponds to the morphism \textbf{E}, the map $(a,b) \mapsto (b, 2b-a)$ corresponds to the morphism $\widetilde{\textbf{D}}$, and the map $(a,b) \mapsto (a, a+b)$ corresponds to the morphism \textbf{G}. Since Christoffel words can be identified with a subset of rational Dyck paths, it may be interesting to develop a deeper connection between Christoffel theory and rational associahedra. The main topological tool used in proving Theorem~\ref{main} was the elementary result of Lemma~\ref{cone-vertex-lemma}. Discrete Morse Theory was introduced by Forman \cite{Forman} and is a powerful and ubiquitous tool in geometric combinatorics for proving deformation retraction type results. More precisely, consider the \textbf{\textit{face poset}} $\widehat{P}$ of ${\widehat{\sf Ass}}(a,b)$ (that is, the set of faces of ${\widehat{\sf Ass}}(a,b)$ partially ordered by inclusion). The poset $\widehat{P}$ contains the face poset $P$ of ${{\sf Ass}}(a,b)$ as an order ideal. The collapse of ${\widehat{\sf Ass}}(a,b)$ onto ${{\sf Ass}}(a,b)$ in Theorem~\ref{main} gives a perfect matching on the Hasse diagram of the difference $\widehat{P} - P$. It follows that this perfect matching is Morse (see \cite[Chapter 11]{Koslov} for the definition of a Morse matching). Conversely, any Morse matching on $\widehat{P} - P$ gives rise to a collapse of ${\widehat{\sf Ass}}(a,b)$ onto ${{\sf Ass}}(a,b)$. It may be interesting to prove Theorem~\ref{main} by exhibiting a Morse matching on $\widehat{P} - P$. It remains an open problem to extend most of the constructions of rational Catalan theory to reflection groups $W$ other than the symmetric group ${\mathfrak{S}}_n$. The $W$-analog of the associahedron is given by the \textbf{\textit{cluster complexes}} of Fomin and Zelevinsky \cite{FZ} and a Fu$\ss$ analog of this construction is given by the \textbf{\textit{generalized cluster complexes}} of Fomin and Reading \cite{FR}. A rational analog of cluster complexes would have as its input data a pair $(W, b)$, where $W$ is an irreducible reflection group with Coxeter number $h$ and $b > h$ is an integer coprime to $h$. The `W-rational associahedron' ${{\sf Ass}}(W,b)$ should be pure of dimension $h-2$, have ${{\sf Cat}}(W, b) := \prod_{i = 1}^n \frac{e_i + b}{e_i + 1}$ facets (where $e_1, \dots, e_n$ are the exponents of $W$), and have $f$- and $h$-vector entries given by rational analogs of the $W$-Kirkman and Narayana numbers. Unfortunately, it does not seem possible to extend the Alexander duality proven in this paper to the full generality of `type $W$'. Such a duality would exist between ${{\sf Ass}}(W,b)$ and ${{\sf Ass}}(W',b)$, where $W'$ is a reflection group with Coxeter number $b-h$. Indeed, within the BCD infinite series, the only possible choices for $h$ are even, forcing both $b$ and $b-h$ to be odd. \section{Acknowledgements} \label{Acknowledgements} The author thanks Drew Armstrong, Benjamin Braun, Steven Klee, and Nathan Williams for helpful conversations. The author was partially supported by NSF grant DMS - 1068861.

Introduction {#JVEv5-sec-0005} ============ Possible 'correlates of immune protection' in HIV-1 infection have been widely reviewed [@JVEv5-bib-0001]. However, from tens of millions of individuals living with HIV globally, there is only one documented case of apparent clearance of HIV-1, the Berlin Patient, who underwent allogeneic bone marrow transplantation from a CCR5 Δ32 homozygous donor [@JVEv5-bib-0005]. Earlier, after approximately 10 years of the epidemic, we originally observed a small group of individuals living with HIV-1, the Sydney Blood Bank Cohort (SBBC), who remained asymptomatic [@JVEv5-bib-0007]. These subjects were subsequently found to be infected with a *nef*/long terminal repeat (LTR)-deleted strain of HIV-1 [@JVEv5-bib-0008], and attenuation of the transmitted viral strain appeared to be the common factor in this asymptomatic cohort. This is consistent with data from experimental infection with *nef*-deleted simian immunodeficiency virus (SIV) [@JVEv5-bib-0009], but overall, *nef* deletion is not a common finding in cohorts of long-term non-progressors (LTNPs) [@JVEv5-bib-0010]. Alternatively, LTNP cohorts exhibit over-representation of CCR5 Δ32 heterozygous individuals [@JVEv5-bib-0013], implicating reduced viral co-receptor expression in limiting viraemia. In the case of the Berlin Patient, it is plausible that ablative therapy and transplantation with progenitors that lacked expression of CCR5 were the main reason for the dramatic reduction of viral burden, although it cannot be ruled out that a low-grade graft versus host immune response could also have played a role [@JVEv5-bib-0005], since other studies found that such protocols reduce HIV DNA in peripheral blood mononuclear cells (PBMCs) to below detectable levels, but ultimately recipients' HIV plasma viraemia can rebound in the absence of antiretroviral therapy (ART) from an unknown reservoir [@JVEv5-bib-0016]. However, there is clear evidence for immune control of viral burden, as indicated by the prevalence of HLA-B27 and HLA-B57 genotypes amongst LTNP [@JVEv5-bib-0018]. Up to half of LTNPs with undetectable plasma viral load, in the absence of ART, referred to as Elite Controllers, are HLA-B57 [@JVEv5-bib-0019], consistent with the consensus that HIV-specific CD8+ cytotoxic T lymphocytes (CTLs), particularly Gag-specific CTL, contribute to control of viraemia [@JVEv5-bib-0003]. Another select group of LTNP subjects includes those with unusually strong CD4 lymphoproliferative responses to p24 [@JVEv5-bib-0022], which may also exert anti-p24 cytotoxicity [@JVEv5-bib-0024]. Furthermore, one report suggested that individuals with HLA-*DRB1\*13* genotype have higher CD4 lymphoproliferative responses to p24 and associated control of viraemia [@JVEv5-bib-0026], while HLA-*DRB1\*13* was over-represented in elite controllers [@JVEv5-bib-0027]. HLA-*DRB1\*1303* and HLA-*DRB1\*1302* were also found to be associated with lower plasma viral loads in two other studies [@JVEv5-bib-0028]. We have investigated immunological and virological parameters in one particular individual, subject C135, who was transfused with the SBBC strain of *nef*/LTR-deleted attenuated HIV-1, as detailed previously [@JVEv5-bib-0030], and maintained a steady CD4 T cell count, unlike some other subjects in this cohort [@JVEv5-bib-0031]. Subject C135 is also CCR5 Δ32 heterozygous [@JVEv5-bib-0031], has HLA-B57 and HLA-*DRB1\*13:02:01* alleles [@JVEv5-bib-0023], and a vigorous PBMC proliferative response to HIV-1 p24 [@JVEv5-bib-0023]. Therefore, in this individual, five well-defined factors associated with LTNP are combined. We now show that C135 has a potent CD4+ T cell response to an HLA-DR13-restricted epitope in HIV-1 p24 *in vitro*, which in turn appears to assist CD8+ T cells to proliferate in response to HLA-B57-restricted epitopes of p24. These responses are associated with complete control over replication of HIV-1 *in vivo* for over 30 years, in the absence of ART, to the point where cells latently infected with HIV-1 DNA can no longer be detected in PBMC, or in biopsy samples of either lymph node or gut-associated lymphoid tissue. Methods {#JVEv5-sec-0006} ======= Subjects {#JVEv5-sec-0007} -------- Subject C135, a 72-year-old male, is a member of the SBBC who was transfused in 1981 with a unit of erythrocytes from donor D36, later shown to be infected with *nef*/3′ LTR-deleted virus [@JVEv5-bib-0030]. Other SBBC transfusion recipients from this donor were also LTNPs [@JVEv5-bib-0007], consistent with attenuation of the transmitted viral strain. C135 was identified in 1996 by tracing of recipients of transfusions from donor D36 and remains completely asymptomatic after a total of 37 years. Detailed serological findings for this subject are described in the results section. C135 is CCR5 Δ32 heterozygous [@JVEv5-bib-0031] and HLA A1,33; B50,57; DR7,13 [@JVEv5-bib-0023]. The HLA-DRB1 alleles, by high-resolution typing, are 07:01:01 and 13:02:01. Healthy adult controls were recruited from university and hospital staff. The study was approved by the Australian Red Cross Blood Service--New South Wales Institutional Ethics Committee, and written informed consent was obtained. HIV-1 Western blots {#JVEv5-sec-0008} ------------------- Prospectively collected longitudinal serum samples from C135 were examined for antibodies to HIV-1 using the BioRad New LAV Blot I (BioRad, Marnes La Coquette, France), according to the manufacturer\'s instructions. Amplification of HIV-1 proviral DNA {#JVEv5-sec-0009} ----------------------------------- Proviral DNA was amplified from patient PBMC DNA using an ultrasensitive PCR method based on a biphasic booster amplification protocol and nef5′ and LTR3′ primers, and the *nef*/LTR product was cloned and sequenced, as previously described [@JVEv5-bib-0030]. In addition, five other assays for HIV-1 proviral DNA have been used, including (1) the diagnostic Roche Amplicor Monitor HIV DNA PCR, with a reported sensitivity of 5 copies per 10,000 PBMCs (Roche Diagnostics, Basel, Switzerland); (2) a published PCR assay for HIV-1 *gag* DNA with an estimated sensitivity of 10 copies per 1,000,000 cells [@JVEv5-bib-0035]; (3) a published sensitive real-time PCR for HIV-1 DNA using *gag* primers, capable of detecting \<10 copies per 300,000 cells [@JVEv5-bib-0036]; (4) a published real-time PCR for HIV-1 DNA using *pol* primers, capable of detecting \<3 copies per 100,000 cells [@JVEv5-bib-0038] [@JVEv5-bib-0040]; and (5) nested PCR for HIV-1 *gag* [@JVEv5-bib-0041] with a sensitivity \<\<10 copies per 300,000 cells [@JVEv5-bib-0042]. DNA for assays 3--5 was extracted using the Qiagen DNeasy Blood and Tissue kit (Qiagen, Hilden, Germany). The single-copy assay was used to quantify plasma viral load as previously described [@JVEv5-bib-0043]. Altogether, these assays were performed in four independent laboratories. Lymphocyte immunophenotyping and cell sorting {#JVEv5-sec-0010} --------------------------------------------- Lymphocyte immunophenotyping of fresh whole-blood samples from C135 between 2009 and 2014 was performed as previously described [@JVEv5-bib-0044]. Memory CD45RO+ CD4+ cells for HIV DNA analysis were purified from PBMC using a FACSAria (BD Biosciences, San Jose, CA, USA), as previously described [@JVEv5-bib-0037]. Similarly, CD4+ T cells were also purified from gut and lymph node biopsy samples (see further) by cell sorting. Lymphocyte function assays {#JVEv5-sec-0011} -------------------------- From 1996 to 2006, PBMCs from C135 were assayed for proliferation in response to HIV-1 p24 in a 6-day ^3^H-thymidine uptake assay, as previously described [@JVEv5-bib-0033]. PBMCs were also labelled with either carboxyfluorescein diacetate succinimidyl ester (CFSE) or Cell Trace Violet (Thermo Fisher Scientific, Waltham, MA USA), and cultured in the presence or absence of recombinant HIV-1 p24, or with pools of overlapping peptides or individual peptides from the sequence of HIV-1 Gag, as previously described [@JVEv5-bib-0023]. Proliferation was optimised by supplementation at day 5 of culture with low doses of exogenous interleukin (IL)-15 (0.5 ng/mL; R&D Systems, Minneapolis, MN, USA) and IL-21 (0.1 ng/mL; BioSource International, Camarillo, CA, USA), followed by a further 5 days of culture. This culture system has been extensively validated to provide an optimal ratio of antigen-specific proliferation to background proliferation [@JVEv5-bib-0048]. At the end of the culture, PBMCs were stained with CD3-PerCP-Cy5.5 or -Pacific Blue, CD4-Alexa Fluor 700 and CD8-APC-Cy7, CD25-APC and CD71-PE (BD Biosciences, San Jose, CA, USA) and analysed on a four-laser LSR II (BD Biosciences), as previously described [@JVEv5-bib-0048]. All proliferation experiments contained negative control cultures in the absence of added HIV-1 Gag antigenic peptides and positive control cultures containing phytohaemagglutinin (PHA), respectively, and were performed on at least three occasions. Intracellular cytokine (ICC) assays, in response to HIV-1 Gag peptides, were performed using fresh sodium heparin-anticoagulated whole blood, as previously described [@JVEv5-bib-0036], with a validated cut-off for positive results of 0.08% (mean from HIV uninfected healthy adult controls (*n* = 20) plus 3× SD) [@JVEv5-bib-0050]. All ICC experiments included negative control cultures and positive control cultures containing Staphylococcal enterotoxin B (SEB), respectively, and were performed on at least two occasions. Lymph node and gut biopsies {#JVEv5-sec-0012} --------------------------- Ultrasound-guided fine needle biopsy (FNB) was performed in 2014 under local anaesthesia, as previously described [@JVEv5-bib-0051], using a 25-gauge needle inserted into a single lymph node in the inguinal chain. Cells were transferred into RPMI/10% FCS, pelleted by centrifugation and resuspended, and CD4 T cells were accurately counted using immunophenotyping in Trucount^TM^ tubes (BD Biosciences), as previously described [@JVEv5-bib-0051]. The remaining cells were cell sorted for CD4+ T cells for HIV DNA analysis. C135 underwent routine endoscopy and colonoscopy screening in 2009 and provided 10 pinch biopsy samples taken from each of three sites: left colon, right colon and terminal ileum. No abnormalities were detected during the procedure. Single-cell suspensions were prepared from these biopsy samples, as previously described [@JVEv5-bib-0044]. CD45+ CD4+ lymphocytes and CD45--epithelial cell adhesion molecule (EpCAM)+ epithelial cells were counted using Trucount^TM^ tubes. Remaining cells were cell sorted for CD4+ T cells for HIV DNA analysis. CCR5 expression on memory CD4+ and CD8+ T lymphocytes {#JVEv5-sec-0013} ----------------------------------------------------- Fresh whole-blood samples from C135 and healthy adult controls, respectively, were stained exactly in parallel within 1 hour of venepuncture using an optimised indirect immunofluorescence method for CCR5, as previously described [@JVEv5-bib-0052]. Briefly, 100 μl of whole blood was incubated with 1 μg of purified anti-CCR5 monoclonal antibody (mAb), clone 2D7 (BD Biosciences) for 1 hour on ice, washed twice with PBS, incubated with 1/50 dilution of PE-labelled affinity purified Fab′2 goat anti-mouse IgG (H + L chain) (Jackson ImmunoResearch, West Grove, PA), washed once, incubated with 1/10 dilution of normal mouse serum for 10 minutes at RT, then incubated for a further 15 minutes at RT with directly conjugated CD3, CD4, CD8 and CD45RO-ECD (Beckman Coulter, Hialeah, FL) mAb. Red cells were lysed with Optilyse C (Beckman Coulter), washed once with PBS, fixed with 0.5% paraformaldehyde in PBS and analysed on the LSR II flow cytometer, as described earlier. Statistical analysis {#JVEv5-sec-0014} -------------------- Results are expressed as medians and interquartile ranges, analysed using Prism versions 6 and 7 (GraphPad, La Jolla, CA, USA). Results {#JVEv5-sec-0015} ======= HIV serology and CD4 T cell counts {#JVEv5-sec-0016} ---------------------------------- C135 presented in 1996 (Figure [1](#JVEv5-fig-0001){ref-type="fig"}a), 15 years after infection, as HIV antibody positive by routine diagnostic enzyme immunoassay, but with an indeterminate Western blot (WB) pattern [@JVEv5-bib-0030], as shown in Figure [1](#JVEv5-fig-0001){ref-type="fig"}b, according to standard Australian interpretative guidelines [@JVEv5-bib-0054]. An indeterminate WB serology was maintained for all available samples, with weak p18, p24 and gp160 bands between 1996 and 2007. These results would be interpreted as HIV positive, according to Centers for Disease Control and Prevention guidelines [@JVEv5-bib-0055], and also in combination with positive nucleic acid testing for HIV-1 on at least two occasions (see further), by revised Australian criteria [@JVEv5-bib-0056]. {#JVEv5-fig-0001} Later WB results from five samples between 2011 and 2017 have been weakly positive for gp160, with all other bands negative. In contrast, the other SBBC members had full WB reactivity [@JVEv5-bib-0007]. The WB pattern for C135 is uncommon and can be seen in some primary HIV subjects early during seroconversion [@JVEv5-bib-0054]. CD4 T cell counts for C135 over the 21 years between 1996 and 2017 remained steady (median: 505 cells/μl), as did the CD4 : CD8 ratio (median: 1.03), as shown in Figure [1](#JVEv5-fig-0001){ref-type="fig"}c. Activated CD38+ HLA-DR+ cells were 0.6%, 1.9% and 1.2% of CD8 T cells on three occasions, respectively, between 2009 and 2014 \[compared to a median of 1.5% (interquartile range: 1.0%--2.0%) for 54 healthy adult controls in a contemporaneous study \[49\]\]. Therefore, C135\'s PBMCs have shown no change in CD8 activation levels since he had a median of 2.0% from nine samples between 1996 and 1998, in our earlier report in 1999 [@JVEv5-bib-0057]. Thus, C135 differs from other LTNP subjects with *nef*-deleted HIV-1, whose CD4 cell counts have eventually declined [@JVEv5-bib-0058], including all but two other members of the SBBC [@JVEv5-bib-0031]. HIV-1 *nef*/LTR DNA PCR assays and sequencing {#JVEv5-sec-0017} --------------------------------------------- Proviral HIV-1 DNA was originally detected in PBMC from C135 in 1996, and the number of copies was estimated to be 2--5 copies per 100,000 PBMCs [@JVEv5-bib-0030]. Longitudinal PBMC samples from C135 were screened for the presence of the *nef*/LTR region of proviral HIV-1 using a boosted PCR (Table [1](#JVEv5-tbl-0001){ref-type="table"}), as previously described [@JVEv5-bib-0030]. Four samples from 1996 and 1997 were positive; however, all subsequent samples tested were negative by the boosted PCR (Table [1](#JVEv5-tbl-0001){ref-type="table"}). ###### PCR amplification of proviral DNA encoding the *nef*/LTR region Sample date *nef*/LTR PCR result Length of *nef*/LTR bp ------------- ---------------------- ------------------------ 13.02.96 − na 11.03.96 − na 09.05.96 \+ 660 25.09.96 \+ 664 18.11.96 \+ 664 24.03.97 \+ 660 14.07.97 − na 04.11.98 − na 19.10.99 − na 28.02.00 − na 23.05.01 − na 08.05.02 − na 05.06.04 − na na: not applicable. Amplification of samples from 1996 to 1997 with primers nef5′ and LTR3′ gave 660, 664-, 664- and 660-bp products, respectively, on the occasions for which a positive signal was obtained (Table [1](#JVEv5-tbl-0001){ref-type="table"}), compared with an 882-bp product for HIV-1~NL4-3~. Cloning and sequencing of the *nef*/LTR amplimer (25.09.96) showed deletions of 1, 72, 4 and 11 bp from the *nef*-alone region relative to HIV-1~NL4-3~, and 139 bp from the *nef*/LTR overlap. This 139-bp deletion includes the deletion common to all SBBC strains [@JVEv5-bib-0030], equivalent to nucleotides 9281--9435 of HIV-1~NL4-3~, and includes the 5′ nuclear factor (NF)-κB binding sequence. Other features common to the SBBC strain of viruses include a GA-rich sequence block at a position equivalent to HIV-1~NL4-3~ 9258--9270 bp, together with duplicated NF-κB and Sp1 sequences, which differ in arrangement between cohort viruses [@JVEv5-bib-0008]. The sequence of C135 includes an additional Sp1 and a partial NF-κB binding sequence [@JVEv5-bib-0030]. The C135 sequence encodes a truncated Nef open reading frame of 19 amino acids due to the introduction of a termination codon (TAA) brought into phase by the deletion of a single nucleotide equivalent to nucleotide 8842 of HIV-1~NL4-3~ [@JVEv5-bib-0030]. However, translation beginning at a second ATG codon (at nucleotides 8822--8824 with respect to HIV-1~NL4-3~), in a different reading frame, would encode a protein containing 96 amino acids of Nef, although, of the conserved blocks, only a complete block B would be retained. HIV DNA PCR assays of peripheral blood mononuclear cells and purified CD4 T cells {#JVEv5-sec-0018} --------------------------------------------------------------------------------- Table [2](#JVEv5-tbl-0002){ref-type="table"} summarises PCR results from C135 from 2002 onwards, from samples including PBMC, whole blood, and purified CD4 T cells from PBMC, gut and lymph node biopsies. ###### PCR results for C135 samples Sample Date PCR Amount of sample Result Limit of detection Method reference ------------------------------ ---------------- -------------------------------- -------------------- ---------- ----------------------------- ---------------------- PBMC 2002--2004 *gag* DNA PCR 5 × 10^6^cells Negative \<10 copies/10^6^cells [@JVEv5-bib-0035] Whole blood 2006--2007 Roche Qualitative Monitor v1.5 0.5 mL whole blood Negative 5 copies/10,000 cells Roche product insert Gut biopsy CD4 T cells August 2009 *gag* DNA PCR 156,000 cells Negative \<10 copies/300,000 cells [@JVEv5-bib-0036] PBMC March 2010 *gag* DNA PCR 83,000 Negative \<10 copies/300,000 cells [@JVEv5-bib-0036] Memory CD4 T cells from PBMC March 2010 *gag* DNA PCR 2 × 10^6^cells Negative \<10 copies/300,000 cells [@JVEv5-bib-0036] PBMC September 2014 *pol* DNA PCR 6.23 × 10^6^cells Negative \<3 copies/100,000 cells [@JVEv5-bib-0040] Lymph node CD4 T cells September 2014 nested *gag* PCR 76,000 cells Negative \<\<10 copies/300,000 cells [@JVEv5-bib-0042] Plasma April 2014 *gag* RNA PCR 14 ml Negative \<0.3 copy/mL [@JVEv5-bib-0043] PBMC: peripheral blood mononuclear cell. HIV DNA PCR testing of fresh whole blood, using samples from 2006 and 2007, were negative by the qualitative diagnostic Roche Amplicor HIV-1 DNA test. Negative results were also obtained for fresh PBMC samples by three other assays. The first PCR assay for HIV-1 *gag* DNA in PBMC had an estimated sensitivity of 10 copies per 1,000,000 cells [@JVEv5-bib-0035]. This assay readily detected HIV DNA in PBMC from other members of the SBBC (NK Saksena, unpublished data). The second assay, a real-time PCR for HIV-1 DNA in PBMC using Gag primers, capable of detecting \<10 copies per 300,000 cells [@JVEv5-bib-0036], was also negative. This assay was also negative on 2 × 10^6^ highly purified memory CD45RO+ CD4 T cells, but this assay readily detected HIV DNA in PBMC and cell-sorted CD4 T cells from other HIV-positive subjects [@JVEv5-bib-0036]. The third assay was a real-time PCR for HIV-1 DNA using *pol* primers, capable of detecting \<3 copies per 100,000 cells [@JVEv5-bib-0038]. A total of 53 replicates were assayed using this assay, from an equivalent of 6.23 × 10^6^ PBMCs, and all replicates were negative. Furthermore, plasma samples have always remained negative by routine diagnostic HIV-1 plasma viral load assays [@JVEv5-bib-0031] and were also negative in 2014 using the single-copy assay [@JVEv5-bib-0043]. The latest diagnostic HIV viral load result from 2017 was \<20 copies/mL. Importantly, HIV-1 has never been isolated from PBMC samples from C135, despite numerous attempts [@JVEv5-bib-0030], using coculture techniques that resulted in successful isolation of viral strains from other members of the SBBC [@JVEv5-bib-0007]. Lymph node and gut biopsies {#JVEv5-sec-0019} --------------------------- Lymph node cells were obtained from C135 by ultrasound-guided FNB of an inguinal lymph node in 2014. The number of CD4 T cells recovered was 148,043 cells, which was towards the low end of the observed normal range for inguinal lymph node FNBs from 10 healthy adult controls (73,701--1,666,592), as previously described [@JVEv5-bib-0051]. HIV DNA in cell-sorted FNB CD4 T cells was negative using a nested *gag* PCR [@JVEv5-bib-0041], recently optimised for assays of antigen-specific CD4 T cells [@JVEv5-bib-0042]. CD4+ T lymphocytes were also isolated from gut biopsy samples in 2009. The number of CD4 T cells in pooled single-cell suspensions derived from biopsy samples from terminal ileum, left colon and right colon for C135 was 28,415 CD4 T lymphocytes/10^6^ EpCAM+ epithelial cells, similar to the median for results for these three sites, from 21 HIV-uninfected adults of 26,381 CD4 T lymphocytes/10^6^ EpCAM+ epithelial cells, as previously described [@JVEv5-bib-0045]. HIV DNA was negative in the 156,000 cell-sorted C135 gut biopsy CD4 T cells (Table [2](#JVEv5-tbl-0002){ref-type="table"}) using real-time PCR [@JVEv5-bib-0037]. Lymphocyte function assays {#JVEv5-sec-0020} -------------------------- PBMC from C135 consistently showed a strong p24-specific proliferative response from 1996 to 2006 (Figure [2](#JVEv5-fig-0002){ref-type="fig"}a), as previously described [@JVEv5-bib-0023]. We now confirm that this response was due to proliferation of CD4+ T lymphocytes (Figure [2](#JVEv5-fig-0002){ref-type="fig"}b), which is characteristic of LTNP [@JVEv5-bib-0022]. Furthermore, when responses to individual HIV-1 Gag peptides were studied previously [@JVEv5-bib-0023], the PBMC proliferative response was found to be directed to the 15-mer WMTNNPPIPVGEIYK, which we now also confirm as a CD4 proliferative response (Figure [2](#JVEv5-fig-0002){ref-type="fig"}b). This sequence contains an epitope previously reported to be presented on HLA-*DRB1\*1301* [@JVEv5-bib-0026], and C135, by high-resolution typing, has the HLA-DRB1 alleles 07:01:01 and 13:02:01. ###### CD4 and CD8 T cell responses to HIV-1 Gag antigens for subject C135. (a) Stimulation indices for C135\'s PBMC proliferation in response to HIV p24, for longitudinal samples from 1996 to 2006, using the ^3^H-thymidine uptake assay. (b) Flow cytometric analysis showing proliferating CD4+ T lymphocytes in CFSE-labelled PBMC from C135, in response to culture with HIV-1 Gag peptide pool (middle histogram) or with peptide \#63, WMTNNPPIPVGEIYK (right histogram), compared with culture in the absence of antigen (left histogram). (c) Flow cytometric analysis showing proliferation of CD8+ T lymphocyte in CFSE-labelled PBMC from C135, in response to culture with the CD8 Gag peptide \#60 (AGTTSTLQEQIGWMT) in the presence (right histogram) and absence (left histogram) of the CD4 peptide \#63 (WMTNNPPIPVGEIYK). (d) Flow cytometric analysis showing proliferation of CD8+ T lymphocyte in CFSE-labelled PBMC from C135, in response to culture with the CD8 Gag peptides \#40 (KVVEEKAFSPEVIPM) and \#41 (EKAFSPEVIPMFSAL) in the presence (right histogram) and absence (left histogram) of the CD4 peptide \#63 (WMTNNPPIPVGEIYK). CFSE: carboxyfluorescein diacetate succinimidyl ester; PBMC: peripheral blood mononuclear cell   Moreover, PBMC also showed a CD8 T-lymphocyte proliferative response to the pool of 123 HIV-1 Gag overlapping peptides (see Figure [2](#JVEv5-fig-0002){ref-type="fig"}b, middle histogram, bottom left quadrant). Such a CD8 proliferative response is characteristic of HLA-B57 LTNP [@JVEv5-bib-0062]. We attempted to map CD8+ T-lymphocyte responses to individual peptides, but only minimal CD8+ responses were found in the absence of a concurrent CD4 response (see left histograms, Figures [2](#JVEv5-fig-0002){ref-type="fig"}c,d), even culturing optimally with the addition of low doses of exogenous IL-15 and IL-21 [@JVEv5-bib-0048]. However, a clear CD8 proliferative response was observed when PBMC were incubated with the 15-mer peptide containing the well-characterised HLA-B57 p24 epitope, TSTLQEQIGW [@JVEv5-bib-0063], in the presence of the CD4 response to the CD4-specific p24 peptide, WMTNNPPIPVGEIYK (Figure [2](#JVEv5-fig-0002){ref-type="fig"}c). Another well-described HLA-B57 epitope, KAFSPEVIPMF [@JVEv5-bib-0063], also gave a very weak CD8 proliferative response, again only in the presence of the CD4 antigenic peptide WMTNNPPIPVGEIYK (Figure [2](#JVEv5-fig-0002){ref-type="fig"}d). Previous experiments with PBMC from C135 have shown weak interferon (IFN)-γ responses and absent CTL responses immediately *ex vivo* [@JVEv5-bib-0023]. Using the ICC assay, we found a small, but distinct, CD4 IFN-γ+ response to HIV-1 Gag peptides (Figure [3](#JVEv5-fig-0003){ref-type="fig"}), which represented around 0.1% of CD4+ T cells in fresh whole blood. Even though this is a relatively low level of antigen-specific CD4+ T cells, it still reaches our validated cut-off for positive ICC results (0.08%), determined for HIV-uninfected controls as part of a vaccine trial [@JVEv5-bib-0050]. The population of IFN-γ+ CD4+ T cells clearly also made IL-2 and were CD127 positive (IL-7Rα+), but were CTLA-4 negative (Figure [3](#JVEv5-fig-0003){ref-type="fig"}), consistent with their proliferative phenotype. {#JVEv5-fig-0003} Cell surface CCR5 expression {#JVEv5-sec-0021} ---------------------------- C135 is CCR5 Δ32 heterozygous [@JVEv5-bib-0031], and we examined expression of CCR5 on T cells in fresh whole blood, as shown compared with T cells from healthy adult controls, stained and analysed in parallel (Figure [4](#JVEv5-fig-0004){ref-type="fig"}). The results show that C135 has a demonstrably lower level of CCR5 expression on his CD4+ T cells (Figure [4](#JVEv5-fig-0004){ref-type="fig"}a), but not a dramatically lower percentage of CCR5+ CD4+ T cells. The difference is even clearer when CD8+ T cells are examined (Figure [4](#JVEv5-fig-0004){ref-type="fig"}b), again with a lower expression level seen as a lower fluorescence intensity, but not as a lower CCR5+ percentage. {#JVEv5-fig-0004} It should be noted that PBMC from C135 were able to be infected with laboratory strains of HIV-1 *in vitro*, although this required depletion of inhibitory CD8+ lymphocytes [@JVEv5-bib-0023]. Discussion {#JVEv5-sec-0022} ========== Subject C135 is distinct from other members within the highly unusual SBBC, infected with the same *nef*/LTR-deleted HIV-1, and compared with other elite controllers. For over 37 years, C135 has maintained undetectable plasma viral load, an uncommon indeterminate WB pattern, and steady CD4 cell count and CD4 : CD8 ratio within the normal range, without CD8 activation. What really sets C135 apart as unique is having had detectable HIV DNA in PBMC samples from around 15 years after infection, but subsequent negative results on a range of assays of HIV DNA in PBMC, purified memory CD4 T cells and lymph node and gut biopsy samples. In contrast, two of the other SBBC recipients, C49 and C64, with consistently undetectable plasma viral loads without ART, have a readily detectable reservoir of HIV-1 DNA in PBMC, and C64 has had virus isolated on one occasion from PBMC [@JVEv5-bib-0031] (and NK Saksena, unpublished data). A leukapheresis sample has not been available from C135 but could possibly be tested in future, more exhaustive studies of viral outgrowth assays of replication-competent HIV-1. One study of PBMC from patients on ART reported that two out of 30 were negative for HIV DNA, but all patients had HIV DNA detected in rectal CD4 cells [@JVEv5-bib-0066]. Three early studies of elite controllers reported only one negative HIV DNA PCR result out of a total 25 elite controllers [@JVEv5-bib-0011], but analysis of tissue samples was not reported. One large study of 46 controllers found only one subject negative using a sensitive assay for HIV RNA in plasma, and this subject was positive for proviral DNA in PBMC [@JVEv5-bib-0067]. Another study of four highly selected elite controllers, with weak indeterminate HIV WBs similar to C135, has also found very low quantitative levels of viral burden, but with at least one positive assay of HIV DNA from PBMC, or from biopsy samples of colon or ileum, as well as one instance of viral outgrowth in high input coculture assay [@JVEv5-bib-0068]. In these latter four cases, proliferative CD8 responses to HIV-infected autologous CD4 T cells and granzyme B-associated cytotoxicity were a feature, but CD4 proliferative responses were not reported. Only two of these cases were HLA-B57 and one had some evidence of a *nef* deletion but still had replication-competent viral outgrowth [@JVEv5-bib-0068]. Furthermore, another study of elite controllers reported significantly elevated activation of circulating CD8 T cells [@JVEv5-bib-0069]. Altogether, subject C135 differs from other elite controllers in the literature. The combination of the antibody results and the T-lymphocyte HIV Gag peptide-specific proliferation and cytokine responses, which are consistent with this subject\'s HLA typing, and together with the early PCR results, convincingly argues that C135 has been weakly infected with a poorly replicating HIV-1 strain. C135 has no other risk factors, and it is believed that he has not been exposed to HIV-1 except in the case of the transfusion in 1981 with an HIV-infected unit from the SBBC donor D36 [@JVEv5-bib-0031]. Furthermore, consistent longitudinal results argue strongly against any mistaken identity or laboratory error in a single sample. Additionally, the sequencing data from the early PCR positive samples from C135 yielded a sequence distinct from, but still closely related to, sequences obtained from other SBBC cohort samples, further arguing against simple contamination or mistaken samples. We and others have found that primary HIV-1 infection in subjects who undergo treatment initiation very early can demonstrate a weakening of their HIV-specific antibody levels [@JVEv5-bib-0070]. More recently, it has been reported that HIV-positive patients who underwent allogeneic stem cell transplantation and subsequently exhibited undetectable HIV DNA showed declining titres of HIV-1 antibodies [@JVEv5-bib-0017]. Similarly, the Berlin Patient has exhibited a decline in HIV-specific titres [@JVEv5-bib-0006], and while it was interpreted that such antibody titre declines may have been due to lack of antigen, in these allogeneic stem cell transplant cases, this was possibly confounded by the conditioning regimens. Altogether, in these cases, and in elite controllers [@JVEv5-bib-0067], low antibody levels correlated with unusually low levels of viral burden. The limited development of the anti-HIV-1 antibody response in C135 and the very low HIV DNA proviral load suggest that HIV-1 replication was restricted very early after infection, consistent with previous observations of low viral load in HLA-B57 individuals during acute infection [@JVEv5-bib-0064]. Also, the limited antibody response to HIV envelope proteins suggests that neutralising antibodies played a minor role, if any, in limiting infection. In fact, plasma samples from C135 showed little or no activity when tested for neutralisation of a range of clade B isolates [@JVEv5-bib-0059]. C135 has a CD4 cell count and a CD4 : CD8 ratio in the normal range, and within those CD4+ T cells, along with other SBBC members, has a high proportion of circulating memory phenotype cells [@JVEv5-bib-0057]. We tested purified memory CD4 T cells from C135\'s PBMC and obtained a negative PCR result. We also tested CD4+ T cells from lymphoid tissue and gut biopsy samples from C135, which were negative. Using the same methods, for samples from other subjects living with HIV, we have previously readily measured HIV DNA in memory CD4 T cells from PBMCs [@JVEv5-bib-0036], gut biopsy samples [@JVEv5-bib-0044] and lymph node fine needle biopsy samples [@JVEv5-bib-0051]. Another reservoir that may harbour infected cells over such a long period could be central nervous system (CNS), but sampling of this compartment is problematic, and we have evidence that even mild symptomatic CNS involvement is reflected in HIV DNA levels in circulating CD4 T cells [@JVEv5-bib-0040]. The transfusion donor, D36, developed HIV-associated dementia after 18 years of being infected with the same strain of *nef*/LTR-deleted HIV-1 [@JVEv5-bib-0034]. However, donor D36 had a detectable plasma viral load and declining CD4 cell count [@JVEv5-bib-0031] and elevated CD8 T cell activation, in contrast with C135 [@JVEv5-bib-0057]. The undetectable viraemia, steady CD4 cell count and lack of activation of CD8 T cells, over more than three decades, argues against a smouldering infection in a non-lymphoid reservoir in C135. C135 has a CCR5 Δ32 heterozygous genotype and reduced cell surface expression of CCR5, which is known to be associated with LTNP status [@JVEv5-bib-0013]. Furthermore, the *nef*/LTR-deleted strain of HIV-1 with which he was transfused has clearly been shown to be attenuated *in vivo* [@JVEv5-bib-0008]. Nef is believed to reduce cell surface expression of HLA class I molecules, and its lack may have helped the Gag-specific CD8 CTL in C135 to rapidly identify infected cells and limit viral burden early in infection [@JVEv5-bib-0064]. One report has described that, amongst rhesus macaques immunised with *nef*-deleted SIV, those macaques with major histocompatibility complex class I alleles previously known to be associated with control of SIV infection exhibited better subsequent control of acute infection with heterologous strains of SIV [@JVEv5-bib-0076], suggesting synergy between *nef* deletion and CD8 responses, analogous to subject C135. The present studies of PBMC proliferation demonstrate that CD4 T cells from C135 may provide help for potent immunological HLA-B57-restricted CD8 T cell control of HIV-1 infection. C135 had responses to two of the best-studied HLA-B57 CD8+ CTL epitopes in p24. The proliferative CD4 response to the sequence WMTNNPPIPVGEIYK adjacent to helix 7 has previously been reported to be associated with viral control in HLA-*DRB1\*1301* individuals [@JVEv5-bib-0026]. While much effort has been made to delineate protective CD8 T cell responses, proliferative Gag-specific CD4 T cell responses are even more highly associated with LTNP status [@JVEv5-bib-0022]. Previously we found that PBMCs from C135 have shown weak IFN-γ responses and absent CTL responses immediately *ex vivo* [@JVEv5-bib-0023]. In contrast with the IFN-γ and CTL results, we had previously found that PBMCs from C135 exhibited strong *in vitro* CD8 antiviral suppressive activity [@JVEv5-bib-0023], which has also been reported for other HIV-positive LTNP and elite controllers [@JVEv5-bib-0023]. Furthermore, PBMCs from C135 protected reconstituted severe combined immunodeficiency mice from superinfection with HIV-1, in an adoptive transfer model [@JVEv5-bib-0080]. C135\'s CD4 response augmented his CD8 response to HLA-B57-restricted HIV-1 Gag peptides. The first CD8 epitope, TSTLQEQIGW, has been described to mutate relatively readily, but with a fitness cost [@JVEv5-bib-0081]. The other small C135 CD8 response is to the epitope, KAFSPEVIPMF, which is located in the highly conserved helix 2 of the N-terminal domain of p24, and mutation of a proline in this helix rendered HIV-1 non-infectious [@JVEv5-bib-0082]. We, and others, have described LTNP subjects with very large populations of proliferative CD4+ CTL that also targeted a nearly identical epitope in helix 2 [@JVEv5-bib-0024]. These results are similar to a previous study where IL-2 produced by CD4 T cells stimulated with an HIV-1 Nef peptide dramatically enhanced a CD8 response to a different Nef peptide [@JVEv5-bib-0084], but the CD4 enhancing activity was lost after resolution of acute infection. For subject C135, the CD4 proliferative activity has been maintained for over 30 years, as also reported for memory CD4 T cell responses to vaccinia virus [@JVEv5-bib-0085]. While it would be desirable to design vaccines to stimulate appropriate targeting of conserved residues in helices 2, 6 and 7 of HIV p24 [@JVEv5-bib-0086], it is difficult to conceive how non-HLA-B57, -B27 or -DR13 subjects could be vaccinated against exactly the same epitopes. Yet helices 2, 6 and 7 and the cyclophilin binding loop may represent a potential therapeutic target for a small molecule [@JVEv5-bib-0087]. However, a therapeutic approach to block Nef function is completely unclear at present, even though *nef* deletion appears to be the best described attenuation for either HIV-1 or SIV *in vivo*, and inhibition of Nef may be critical to T cell control of the HIV-1 reservoir [@JVEv5-bib-0088]. We interpret the results for C135 as a probable case of clearance of HIV-1 infection, resulting from a combination of plausible immune, host genetic and viral attenuation mechanisms. However, with only a single individual subject, caution is still required yet provides some hope that a similar combination of favourable pressures can be more generally applied to HIV-1 infection with a similar outcome. The authors thank the subject C135 for his co-operation in providing samples, Julie Yeung and Kate Merlin for PBMC cryopreservation, Leon McNally and Alex Carrera for diagnostic PCR assays, as well as the NIH Reference Reagents Program for provision of HIV-1 overlapping Gag peptides. This project was partly funded by Australian National Health and Medical Research Council Fellowships (1063422 to JZ, 351041 to CMLM), Project Grants (101085 to JZ) and Program Grants (510448 and 1052979 to AK, DAC). WBD was supported by The Australian Red Cross Blood Service, an NHMRC project grant, and the Australian Centre for HIV/HCV Virology Research. SP and BH were supported by the Delaney AIDS Research Enterprise (DARE) to Find a Cure (1U19AI096109 and 1UM1AI126611-01), Australian Centre for HIV and Hepatitis Virology Research (ACH^2^), and the Australian National Health and Medical Research Council (Project Grant AAP1061681). {#JVEv5-sec-0023} Authors' contributions {#JVEv5-sec-0024} ---------------------- JZ, WBD, MC, CMLM, PHC, KS, BW, SP, PRG, NS, NKS and AK designed experiments, and JZ, WBD, MC, CMLM, PC, KS, BW, KMcB, WH-N, BH, MB and YX performed experiments. MD recruited and performed endoscopy. SR provided clinical care for subject C135 and identified the clinical possibility of viral clearance. AK, KK, DAC, JSS and JL supervised the recruitment and management of LTNP cohorts. JZ, WBD, MC, CMLM, PHC, KK, SP and AK analysed the data. JZ, WBD, MC, CMLM and AK vouch for the data and analysis and wrote the paper. {#JVEv5-sec-0025} Conflict of interest {#JVEv5-sec-0026} -------------------- The authors do not disclose any financial conflicts of interest. [^1]: JJZ, WBD, MC and CMLM contributed equally to this paper.

1. Introduction {#sec1-jcm-08-01569} =============== Iron deficiency (ID) is a common comorbidity in patients with heart failure \[[@B1-jcm-08-01569]\]. Prevalence is similar in patients with heart failure with reduced ejection fraction (HFrEF), heart failure with mid-range ejection fraction (HFmrEF), and heart failure with preserved ejection fraction (HFpEF), at 50%, 61%, and 64% respectively \[[@B2-jcm-08-01569]\]. In these patients, ID can arise not only from low whole body ferritin stores, termed absolute ID and which is multifactorial including intestinal dysfunction with reduced absorption due to reduced gastric emptying and mucosal edema, along with reduced consumption of iron rich foods due to anorexia, and increased blood loss due to anticoagulants and gastric ulceration \[[@B3-jcm-08-01569],[@B4-jcm-08-01569]\]; but also from a rise in the regulatory protein hepcidin, which prevents ferritin release from enterocytes and macrophages in response to an inflammatory state, causing functional ID \[[@B5-jcm-08-01569]\]. Mounting evidence supports a key role of ID in heart failure outcomes. In patients with HFrEF, there is comprehensive data illustrating that ID is related to worse dyspnea class, higher biomarker levels, lower exercise capacity, lower quality of life, and greater risk of death and heart transplantation \[[@B1-jcm-08-01569],[@B6-jcm-08-01569],[@B7-jcm-08-01569],[@B8-jcm-08-01569]\]. Treatment of iron deficiency, in keeping with these findings, has focused on patients with HFrEF. Intravenous ferric carboxymaltose administration is associated with improvement of quality of life, reduced hospitalizations, increased exercise tolerance, and maximal oxygen consumption (VO~2~ max) in HFrEF \[[@B9-jcm-08-01569],[@B10-jcm-08-01569],[@B11-jcm-08-01569],[@B12-jcm-08-01569]\]. Of note, this is not the case with oral iron repletion \[[@B13-jcm-08-01569]\]. There is less data supporting a role of ID in HFpEF pathogenesis and prognosis. ID is nevertheless associated with lower VO~2~ max in all three phenotypes of heart failure \[[@B2-jcm-08-01569]\]. This lends weight to previous findings suggesting that ID is related to functional outcomes in advanced HFpEF \[[@B14-jcm-08-01569]\]. Furthermore, absolute ID has been associated with increased risk of 30-day readmission in a population of patients of whom HFpEF represented just over half \[[@B15-jcm-08-01569]\]. However, this association has not been investigated for HFpEF and HFrEF separately. A role for ID in the pathogenesis and severity of HFpEF could indicate potential therapeutic benefit of ferric carboxymaltose in this population for whom pharmacological options are lacking \[[@B16-jcm-08-01569]\], and which is currently under investigation in the FAIR-HFpEF trial (NCT03074591). Furthermore, whilst the prevalence of iron deficiency in acute decompensated heart failure is demonstrably high \[[@B17-jcm-08-01569]\], data is lacking for the significance of ID in this cohort. In this study we sought to identify whether there is an association between ID and clinical entities such as dyspnea class, length of stay, biomarker levels, and echocardiographic indices of diastolic function in patients with acute decompensated heart failure (ADHF) in both HFrEF and HFpEF. 2. Methods {#sec2-jcm-08-01569} ========== 2.1. Study Design and Population {#sec2dot1-jcm-08-01569} -------------------------------- Patients were recruited in a prospective registry on acute heart failure between November 2014 and August 2017 at the University Hospitals of Geneva, ClinicalTrials.gov NCT02444416. The cohort comprised consecutive patients admitted to the departments of General Internal Medicine with ADHF, according to the European Society of Cardiology definition; exhibiting symptoms of heart failure including dyspnea, ankle swelling, and fatigue, accompanied by signs of heart failure such as elevated jugular venous pressure, pulmonary crackles, and peripheral edema due to structural or functional cardiac abnormality \[[@B18-jcm-08-01569]\]. In addition, patients were required to have elevated brain natriuretic peptide (BNP) levels \>100 ng/L, or pro-BNP levels \>300 ng/L to be included in the study. This study complies with the Declaration of Helsinki, and was approved by the local medical ethics committee, and all cohort participants signed an informed consent form. 2.2. Study and Laboratory Measurements {#sec2dot2-jcm-08-01569} -------------------------------------- Demographic information, including age, sex, weight, smoking status, and a full medical history was recorded. Treatment history at presentation and medical therapy through the course of admission was noted, focusing on heart failure therapies. Clinical status at presentation, including New York Heart Association (NYHA) dyspnea class, was noted. Routine investigations as part of the admission with ADHF were registered for each patient, including full blood examination, urea and electrolytes, C-reactive protein (CRP), iron studies, liver function tests, troponin levels, electrocardiography, and echocardiography. BNP levels were recorded until March 2015, followed by recording of N-terminal prohormone of BNP (NT-proBNP) afterwards. Echocardiography data were analyzed by staff cardiologists. Participants were classified as having HFpEF if they had a left ventricular ejection fraction (LVEF) ≥50%; HFrEF if they had a LVEF \<40%; and heart failure with mid range ejection fraction (HFmrEF) if they had a LVEF ≥40% and \<50%. Iron deficiency was defined as either absolute iron deficiency, with a ferritin level \<100 ng/mL; or functional iron deficiency, with a ferritin 100--200 ng/mL accompanied by a transferrin saturation \<20%; in accordance with current consensus in the literature for the diagnosis of iron deficiency in heart failure. 2.3. Statistical Analysis {#sec2dot3-jcm-08-01569} ------------------------- Categorical variables are presented as numbers with percentages of patients; continuous variables are presented as means and standard deviations if approximately normally distributed, or with medians with interquartile ranges otherwise. Differences in LVEF and age were compared using the independent samples t-test, whilst the difference in gender, comorbidities, and NYHA class were compared with the chi-square test for independence. The differences in iron status, biomarkers including CRP and BNP/NT-proBNP, and echocardiographic markers of diastolic function were compared using the independent samples t-test or the Mann--Whitney U test. Linear regression analyses were used to examine the effect of iron deficiency on outcomes independent of other differences in baseline characteristics. A *p*-value of \<0.05 was considered statistically significant for all of the analyses. All statistical analyses were performed with R statistics. 3. Results {#sec3-jcm-08-01569} ========== 3.1. Patient Characteristics {#sec3dot1-jcm-08-01569} ---------------------------- A total of 503 patients were included during the recruitment period; 217 (43%) were female. The mean age of patients was 78 ± 11 years. The majority of patients (55%) had NYHA class IV dyspnea. Complete echocardiographic data was available for 487 out of 503 patients. A total of 160 (33%) had HFrEF, 270 (55%) had HFpEF, and 57 (12%) had HFmrEF. Given the small proportion of patients with HFmrEF in our cohort, they were not included in further analyses. Differences between HFrEF and HFpEF patients are detailed in [Table 1](#jcm-08-01569-t001){ref-type="table"}. Patients with HFpEF were significantly older than patients with HFrEF, with mean age 80 ± 10 and 75 ± 12 years respectively. A significantly higher proportion of patients with HFpEF were female (52%) as compared to those with HFrEF (30%). The length of stay (LOS) was not significantly different between the categories of heart failure, nor was the NYHA class. There was no difference in prevalence of coronary artery disease, diabetes mellitus, obesity, chronic kidney disease (CKD), or chronic obstructive pulmonary disease (COPD) between HFrEF and HFpEF. Patients with HFpEF had a higher prevalence of hypertension and atrial fibrillation or flutter; HFrEF patients were more likely to be smokers or have a history of smoking. 3.2. Biomarker Levels {#sec3dot2-jcm-08-01569} --------------------- There was no significant difference between groups with regard to CRP levels, although this trended higher in HFpEF patients. Patients with HFrEF had significantly higher BNP and NT-proBNP than the HFpEF cohort. 3.3. Prevalence of Iron Deficiency {#sec3dot3-jcm-08-01569} ---------------------------------- Both ferritin and transferrin saturation results were available for 416 patients. Of this, a total of 239 patients had iron deficiency, representing 54% of patients with HFrEF and 56% of patients with HFpEF. There were no statistically significant differences in absolute and functional iron deficiency between HFrEF and HFpEF. Haemoglobin level was significantly lower in patients with HFpEF, at 121 ± 23 g/mL, than those with HFREF, at 129 ± 20 g/mL. 3.4. Interaction between Iron Deficiency and HFpEF {#sec3dot4-jcm-08-01569} -------------------------------------------------- [Table 2](#jcm-08-01569-t002){ref-type="table"} details patients with HFpEF according to iron deficiency status. In patients with HFpEF, females were significantly more likely to be ID than males, across both categories of ID, representing 64% of ID patients, compared to 40% of iron replete patients. There were no significant differences in age or NYHA class determined by ID status. LOS was significantly longer amongst those with ID, irrespective of type (9 days vs. 11 days, *p* = 0.036), as depicted in [Figure 1](#jcm-08-01569-f001){ref-type="fig"}. HFpEF patients with absolute ID were less likely to have CKD than those with functional ID or iron replete patients. Conversely, those with absolute ID were more likely to have COPD compared to iron replete patients. There were no differences across iron status groups in other comorbidities. CRP was significantly higher in those with functional ID, and was lowest in those with absolute ID. BNP and NT-proBNP levels did not differ significantly across ID categories. In multivariate linear regression analysis for LOS with iron deficiency as a binary variable, iron deficiency remained an independent predictor of LOS in HFpEF (*b* = 2.74, 95% CI 0.32--5.17), after adjusting for CKD, CAD, hypertension, DM, obesity, atrial fibrillation or flutter, COPD, CRP, and age. The adjusted R2 for this model was 6.5%, indicating significant unmeasured variables accounting for differences in LOS. 3.5. Interaction between Iron Deficiency and HFrEF {#sec3dot5-jcm-08-01569} -------------------------------------------------- [Table 3](#jcm-08-01569-t003){ref-type="table"} details patients with HFrEF according to iron deficiency status. In patients with HFrEF, females were not more likely to be ID than males, but there were proportionally more females within the total and absolute ID groups than in the iron replete group, representing 40% of ID patients, compared to 23% of iron replete patients. There were no significant differences in age, NYHA class, LOS, or comorbidities according to ID. Biomarkers CRP, BNP, and NT-proBNP did not differ according to iron status. In multivariate linear regression analysis for LOS with iron deficiency as a binary variable along with comorbidities CKD, CAD, hypertension, DM, obesity, atrial fibrillation or flutter, COPD, and age, COPD was the only independent predictor of length of stay (*b* = 42, 95% CI 7.44--76.7). The adjusted R^2^ for this model was 7.6%, indicating significant other active factors accounting for the difference in LOS. 4. Discussion {#sec4-jcm-08-01569} ============= 4.1. Prevalence of Iron Deficiency {#sec4dot1-jcm-08-01569} ---------------------------------- Our study of ADHF found a high prevalence of ID in both HFrEF and HFpEF, at 54% and 56% respectively. Furthermore, in both patients with HFpEF and HFrEF, women were overrepresented amongst iron deficient patients. Finally, ID was associated with prolongation of hospital stay in HFpEF, but not HFrEF. With respect to prevalence of ID, these results mirror similar studies of the prevalence of ID in different HF categories. Martens et al. found the prevalence of ID to be 53% and 64% for HFrEF and HFpEF respectively in a cohort of 1197 patients. A higher prevalence of ID was found in a study by Nanas et al., with 73% showing ID on bone marrow aspirate \[[@B19-jcm-08-01569]\]; however these patients were categorized as advanced heart failure, and they only recruited patients with anemia, as compared to our study, which included all patients with ADHF. Regarding gender differences in our cohort, with HFpEF women more likely to be ID, this reflects a higher prevalence of ID in women across Europe \[[@B20-jcm-08-01569]\]. Given the link between ID and LOS, along with rehospitalization and functional outcomes in HFpEF in other studies, ID may contribute to the relative over expression of HFpEF in women. Our findings are concordant with those previously reported in the literature: just over half (55%) of patients had HFpEF \[[@B21-jcm-08-01569]\]; those with HFpEF were more likely to be female and older, with a higher prevalence of hypertension and atrial fibrillation, than those with HFrEF \[[@B21-jcm-08-01569]\]; patients with HFrEF were more likely to be active or past smokers \[[@B22-jcm-08-01569]\]; and natriuretic peptide levels were higher in patients with HFrEF \[[@B23-jcm-08-01569]\]. Our study expands on the limited existing literature regarding the interaction between ID and heart failure outcomes in HFpEF, and notably ID patients with HFpEF had a longer LOS. We found that in the HFpEF cohort, the LOS was approximately 2 days longer in those who were ID; a finding that was not seen in the HFrEF cohort. LOS might be a surrogate finding for a more prominent relationship between ID and HFpEF than HFrEF and this may suggest a prognostic role for ID in HFpEF. ID affects cardiovascular function through multiple mechanisms. It reduces oxygen carrying capacity and oxygen storage; and in heart and skeletal muscle, it causes mitochondrial dysfunction, reduced energetic efficiency, and anaerobic metabolism \[[@B17-jcm-08-01569]\]. ID is closely linked to exercise intolerance given this role in energy storage and production in both cardiac and skeletal muscle, resulting in reduced work productivity \[[@B24-jcm-08-01569]\]. This mirrors the complex nature of exercise intolerance in HFpEF, which derives from the interaction of pulmonary, cardiac, hematological and skeletal muscle impairments \[[@B25-jcm-08-01569]\]. Given exercise intolerance is a cardinal symptom in HFpEF, this relationship between ID and exercise tolerance may explain the impact of ID on LOS in our study, however data on cardiopulmonary exercise testing or 6-min walk tests would be required to confirm this hypothesis. ID also plays direct role in cardiac remodeling, as demonstrated by Martens et al. in a recent study showing that cardiac resynchronization therapy (CRT) is less effective at improving left ventricular geometry in those who are ID. Furthermore, ID predicted attenuated symptomatic improvement with CRT \[[@B26-jcm-08-01569]\]. These adverse effects of ID on cellular function and cardiac remodeling were independent of anemia \[[@B17-jcm-08-01569],[@B26-jcm-08-01569]\]. However, remodeling in patients with HFpEF differs from that of HFrEF, and the applicability of these findings to HFpEF patients is not clear. Naito et al. found that long-term ID with anemia in rats resulted in left ventricular hypertrophy with preserved ejection fraction, with associated myocardial interstitial fibrosis \[[@B27-jcm-08-01569]\], suggesting that ID may also play a pathophysiologic role in HFpEF; however, replication of these results in human subjects is required. We did not see any changes to echocardiographic parameters in our HFpEF cohort according to ID category; however, this may have been result of an inadequate sample size. A pathophysiological link between ID and HFpEF may arise from their shared inflammatory origins; HFpEF is driven by a pro-inflammatory state secondary to comorbidities \[[@B28-jcm-08-01569]\] as demonstrated by higher levels of circulating inflammatory markers compared to HFrEF \[[@B29-jcm-08-01569]\]; and ID is associated with systemic inflammation which causes elevated levels of hepcidin, leading to functional ID \[[@B5-jcm-08-01569]\]. This was highlighted in our cohort by higher levels of CRP in ID patients with HFpEF. Iron plays a role in regulation of the immune system and response to pathogens \[[@B30-jcm-08-01569]\], however there is no direct evidence to date suggesting that iron deficiency itself causes inflammation. The observation of increased LOS in ID patients with HFpEF could be a reflection of mediating comorbidities that cause iron deficiency and that could lengthen the hospital stay. CKD is commonly associated with HFpEF \[[@B31-jcm-08-01569]\], and causes ID through increased hepcidin, bleeding due to uremic platelet dysfunction, impaired dietary iron absorption, and loss through dialysis. Obesity and insulin resistance are associated with altered iron homeostasis, and particularly in females ID is a common feature \[[@B32-jcm-08-01569]\]. We analyzed the contribution of these comorbidities, along with CRP given a possible role of infections or inflammation, and age, in a multivariate linear regression model, in which ID remained an independent predictor of LOS. Whilst the model accounted for only a small portion of the variation in LOS in HFpEF, it remains a contributor to prolonged hospitalization. Another possible confounder for our finding is that increasing heart failure severity increases the likelihood of iron deficiency \[[@B6-jcm-08-01569]\], and therefore the prolonged LOS in ID patients could be driven by severity of heart failure rather than iron deficiency. However, there was no difference in NYHA class across ID groups in our study, suggesting that this may not explain our finding. 4.2. Clinical Implications {#sec4dot2-jcm-08-01569} -------------------------- Our study highlights the importance of diagnosing iron deficiency, as it plays a role in determining short-term prognosis. However, iron studies appear to be underperformed in the clinical setting. A study from Silverberg et al. in 2013 found that 72.4% of patients with a primary diagnosis of heart failure had no iron workup \[[@B33-jcm-08-01569]\]. In our study, 82.7% of patients had complete iron studies, representing a significant improvement on the aforementioned study. Intravenous iron therapy with ferric carboxymaltose is established as an important treatment for patients with HFrEF, improving functional capacity, symptoms of heart failure, quality of life, and reducing heart failure hospitalizations \[[@B34-jcm-08-01569],[@B35-jcm-08-01569]\]. Given that treatment for HFpEF is limited by a lack of efficacy of neurohormonal therapies \[[@B16-jcm-08-01569]\], and treatment currently focuses on comorbidities and diuresis \[[@B18-jcm-08-01569]\], novel therapeutic strategies are in demand. The effect of ID in prolonging LOS in HFpEF patients in our study may suggest potential efficacy of intravenous iron treatment in this population to reduce hospitalization, and to improve functional outcome. Our findings support the hypothesis of the FAIR-HF trial, which is currently investigating the efficacy of ferric carboxymaltose in the treatment of patients with HFpEF (NCT03074591). The elevated CRP in functional ID in this study may highlight an alternative therapeutic approach in targeting inflammation in HFpEF. Anti-inflammatory therapies have already proven effective in reducing myocardial infarction, nonfatal stroke, and cardiovascular death \[[@B36-jcm-08-01569]\]. The aforementioned relationship between inflammation, ID, and HFpEF could indicate a role of anti-inflammatory therapies in reducing ID and accordingly improving outcomes in HFpEF patients. 4.3. Limitations of the Study {#sec4dot3-jcm-08-01569} ----------------------------- Limitations as well as strengths of our study are its prospective observational design which on the one hand reduces our ability to assess causal relationships, notably because of residual confounding, but on the other hand reduced both selection and recall bias. Further limitations include lack of data on certain parameters of clinical management. For example, data on exercise capacity, such as maximal oxygen consumption testing, was not available. In addition, data on administration of iron infusions was not collected. This limitation may be particularly relevant to patients with HFrEF, who are treated with intravenous iron therapy if ID in accordance with heart failure guidelines \[[@B18-jcm-08-01569]\], unlike patients with HFpEF. This may explain our failure to find any association between ID and LOS in HFrEF, which may have been mitigated by iron infusion treatment, sometimes benefiting patients symptomatically within days of their infusion. Ideally, our study should be replicated with more sophisticated prognostic measures, such as cardiopulmonary exercise testing and invasive hemodynamic measures. 5. Conclusions {#sec5-jcm-08-01569} ============== In this cohort of patients with ADHF, there is a high prevalence of ID in both HFrEF and HFpEF, with a significantly higher proportion of women with ID and HFpEF. Furthermore, ID was associated with prolongation of hospital stay in HFpEF, but not HFrEF. This may indicate an important role for iron deficiency in HFpEF, which could therefore also represent a therapeutic target. Further studies in larger cohorts and more varied populations are warranted to increase the generalizability of these results. The authors would like to acknowledge the work of the Internal medicine research residents for their contribution to this work, in particular Sylvan Budry who did significant work on managing the data prior to analysis. Conceptualization A.B., P.M., S.C., D.K.; Methodology A.B., D.C., N.G., T.A., J.S. (Jerome Stirnemann), P.M., S.C.; Formal analysis, A.B., J.S. (Jerome Stirnemann); Investigation, D.C., J.S. (Jacques Serratrice), N.G., T.A., J.S. (Jerome Stirnemann), S.C.; Writing---Original Draft Preparation, A.B.; Writing---Review and Editing, S.C.; P.M., D.K. The authors declare no conflict of interest. {#jcm-08-01569-f001} jcm-08-01569-t001_Table 1 ###### Patient characteristics, iron status, biomarker levels, and diastolic function indices on transthoracic echocardiography (TTE), according to heart failure category. --------------------------------------------------------------------------------------------------- Total Cohort HFrEF,\ HFpEF,\ *p*-Value 160 (33%) 270 (55%) ------------------------------------------- -------------- ------------- ------------- ------------ Baseline characteristics   LVEF (%) 48 ± 16 27 ± 6 60 ± 6 \<0.001 \*   Age, year 78 ± 11 75 ± 12 80 ± 10 \<0.001\*   Female sex, *n* (%) 217 (43%) 48 (30%) 141 (52%) \<0.001 \*   NYHA class, I/II/III/IV (%) 2/6/37/55 1/7/40/52 1/7/38/54 0.97   Length of stay, days 11 ± 9 11 ± 8 10 ± 8 0.34   Coronary artery disease, *n* (%) 177 (36%) 65 (42%) 83 (32%) 0.054   Hypertension, *n* (%) 393 (79%) 111 (70%) 223 (84%) \<0.001 \*   Diabetes mellitus, *n* (%) 157 (31%) 47 (30%) 87 (32%) 0.61   Obesity, *n* (%) 231 (47%) 69 (45%) 130 (49%) 0.49   Active/past smoker, *n* (%) 288 (60%) 106 (68%) 145 (56%) 0.019 \*   Chronic kidney disease, *n* (%) 175 (36%) 49 (31%) 105 (40%) 0.081   Atrial fibrillation or flutter, *n* (%) 222 (45%) 58 (36%) 132 (50%) 0.01 \*   COPD, *n* (%) 73 (15%) 21 (14%) 38 (16%) 0.61 Iron studies   Iron deficiency, *n* (%) 239 (57%) 73 (54%) 123 (56%) 0.88   Absolute iron deficiency, *n* (%) 165 (38%) 49 (37%) 82 (41%) 0.64   Functional iron deficiency, *n* (%) 74 (18%) 24 (18%) 30 (15%) 0.61   Haemoglobin level, g/mL 123 ± 23 129 ± 20 121 ± 23 \<0.001 \* Biomarker levels   CRP, mg/L 11.4 ± 27.6 9.4 ± 19.8 12.1 ± 31.5 0.064   BNP, ng/L 857 ± 885 1279 ± 974 648 ± 687 \<0.001 \*   NT-proBNP, pg/mL 3926 ± 6763 7340 ± 8020 2695 ± 5171 \<0.001 \* --------------------------------------------------------------------------------------------------- COPD, chronic obstructive pulmonary disease; LVEF, left ventricular ejection fraction; NYHA, New York Heart Association; BMI, body mass index; LV, left ventricle; LVMI, left ventricular mass index; LAVI, left atrial volume index; RVSP, right ventricular systolic pressure; TTE, transthoracic echocardiography; RVSP, right ventricular systolic pressure; CRP, C-reactive protein, BNP, brain natriuretic peptide, NT-proBNP, N-terminal prohormone of BNP. \* denotes *p*-value less than the significance threshold of \<0.05. jcm-08-01569-t002_Table 2 ###### Patient characteristics, biomarker levels and diastolic function indices on TTE according to iron status in the HFpEF cohort. Iron Deficient ------------------------------ ------------- ---------------- ------------- ------------- ------------ Female sex, *n* (%) 39 (40%) 79 (64%) 59 (67%) 20 (57%) 0.0012 \* Age 79.1 ± 9.2 79.1 ± 10.2 78.7 ± 10.2 82.8 ± 9.6 0.038 \* NYHA class, I/II/III/IV (%) 2/6/34/44 1/9/45/57 1/3/32/44 0/6/13/13 0.19 Length of stay, days 9 ± 6 11 ± 7.7 11 ± 7 11 ± 9.5 0.036 \* CAD, *n* (%) 36 (38%) 35 (29%) 27 (31%) 8 (24%) 0.30 Hypertension, *n* (%) 84 (87%) 96 (81%) 69 (81%) 27 (82%) 0.58 Diabetes mellitus, *n* (%) 35 (36%) 37 (31%) 26 (30%) 11 (32%) 0.67 Obesity, *n* (%) 46 (47%) 60 (50%) 47 (54%) 13 (38%) 0.28 Active/past smoker, *n* (%) 53 (58%) 71 (59%) 51 (59%) 20 (59%) 0.98 CKD, *n* (%) 50 (53%) 44 (37%) 26 (31%) 18 (53%) 0.008 \* Atrial fibrillation, *n* (%) 50 (52%) 61 (50%) 41 (47%) 20 (59%) 0.50 COPD, *n* (%) 11 (12%) 26 (22%) 22 (27%) 4 (12%) 0.024 \* CRP 13.3 ± 51.6 9.9 ± 26 7.5 ± 20.1 24 ± 48.6 \<0.001 \* BNP 768 ± 710 585 ± 463 525 ± 461 1090 ± 236 0.46 NT-proBNP 3239 ± 6814 2389 ± 3174 2318 ± 2294 2515 ± 7058 0.12 EF (%) 60 ± 5 60 ± 7.5 60 ± 7.62 60 ± 7.5 0.97 CAD, coronary artery disease; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; NYHA, New York Heart Association; CRP, C-reactive protein, BNP, brain-natriuretic peptide; NT-proBNP, N-terminal prohormone of BNP; EF, ejection fraction. \* denotes *p*-value less than the significance threshold of \<0.05. jcm-08-01569-t003_Table 3 ###### Patient characteristics, biomarker levels and diastolic function indices on TTE according to iron status in the heart failure with reduced ejection fraction (HFrEF) cohort. Iron Deficient ------------------------------ ------------- ---------------- --------------- ------------- ---------- Female sex, *n* (%) 14 (23%) 29 (40%) 24 (49%) 5 (22%) 0.007 \* Age, HFpEF 74.7 ± 11.1 76.4 ± 12.4 77.3 ± 13.1 74.5 ± 11 0.32 NYHA class, I/II/III/IV (%) 0/4/23/34 0/6/32/32 0/3/24/20 0/3/8/11 0.74 Length of stay, days 12 ± 8 12 ± 9 11 ± 10 12 ± 6.5 0.78 CAD, *n* (%) 25 (42%) 28 (41%) 19 (40%) 8 (38%) 0.96 Hypertension, *n* (%) 45 (74%) 46 (64%) 32 (65%) 13 (59%) 0.39 Diabetes mellitus, *n* (%) 18 (30%) 20 (28%) 10 (20%) 9 (41%) 0.20 Obesity, *n* (%) 28 (48%) 29 (40%) 16 (33%) 12 (55%) 0.14 Active/past smoker, *n* (%) 37 (64%) 51 (72%) 32 (67%) 18 (82%) 0.30 CKD, *n* (%) 24 (39%) 17 (24%) 10 (21%) 7 (32%) 0.12 Atrial fibrillation, *n* (%) 23 (38%) 24 (33%) 14 (29%) 9 (41%) 0.49 COPD, *n* (%) 9 (15%) 10 (14%) 8 (17%) 2 (9%) 0.74 CRP 11.4 ± 24.6 8.7 ± 19.1 7.9 ± 12.8 11.4 ± 35.6 0.33 BNP 1294 ± 983 1292 ± 1052 1340 ± 678 926 ± 737 0.41 NT-proBNP 7928 ± 7201 6853 ± 10,327 6998 ± 10,787 3938 ± 7196 0.45 EF (%) 27.5 ± 10 27.5 ± 10 27.5 ± 7.5 32.5 ± 8.8 0.083 CAD, coronary artery disease; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; NYHA, New York Heart Association; CRP, C-reactive protein, BNP, brain-natriuretic peptide; NT-proBNP, N-terminal prohormone of BNP; EF, ejection fraction. \* denotes *p*-value less than the significance threshold of \<0.05.

INTRODUCTION {#S1} ============ Although many pathways that are required for healing have been discovered, little is known about how or whether healing is synchronized with general processes that regulate an animal's homeostasis and behavior. The circadian clock is an ancient molecular pathway that synchronizes organisms with daily environmental cues (zeitgebers) such as light intensity and temperature oscillations ([@R6]; [@R12]). Circadian rhythms are repeated over a 24 hr cycle, yet this chronological aspect of cell state has received little attention in the field of regenerative biology. For instance, many of the pathways that regulate intestinal regeneration and intestinal stem cells (ISCs) have been the subject of important studies ([@R5];[@R7]), but most of these studies did not consider whether results obtained during one part of the day occur at all times. Circadian rhythms are thought to influence the cell cycle ([@R6]), and there is some evidence that the clock plays a role in regeneration and proliferation. Hepatocyte cell division exhibits rhythms and is delayed following hepatectomy if circadian rhythms are disrupted ([@R17]). Earlier studies in the intestine indeed found a daily rhythmicity in cell number and villus length ([@R29]; [@R35]), as well as proliferation ([@R3];[@R28]), although clock mutants were not examined and ISCs were not specifically identified in those reports. Further, it was reported that metabolic processes display time-of-day variation in the intestine ([@R25]; [@R31]; [@R33]), and *per* mutation hastens tumorigenesis in Wnt pathway-driven colorectal cancer in mice ([@R37]). Finally, the degree of intestinal mucositis displays time-of-day variability in cancer patients treated by radiation ([@R34]). This suggests that circadian rhythms may influence the intestine's regenerative response, although the reasons for this remain a mystery. RESULTS {#S2} ======= The *Drosophila* Intestine Has a Circadian Clock {#S3} ------------------------------------------------ The intestinal biology of *Drosophila* parallels that of mammals ([@R5]; [@R7]) and allows for functional in vivo analyses to elucidate regenerative processes. *Drosophila* ISCs divide to produce progenitors called enteroblasts (EBs) that differentiate directly into absorptive enterocytes (ECs) or secretory enteroendocrine cells ([Figure 1A](#F1){ref-type="fig"}). We performed a transgenic RNAi screen for transcription factors required in *Drosophila* ISCs during regeneration (see Experimental Procedures). It was previously shown that after damage occurs, ISCs regenerate the intestine by proliferating to replace dying cells ([@R5]; [@R18]). Here we discovered that among the \~600 genes tested, *period* (*per*) was required for proliferation of adult ISCs following damage by dextran-sodium sulfate (DSS), a chemical that models inflammatory bowel diseases in flies and mice ([@R4]). The *Drosophila* circadian pacemaker comprises the transcription factor partners *clock* (*clk*) and *cycle* (*cyc*), which are negatively regulated by *per* and *timeless* (*tim*; [@R12]). One transcriptional target of CLK/CYC is *per* itself, which represses its own production and causes the cyclical transcriptional rhythms that underlie circadian rhythms. The existence of independent clocks throughout *Drosophila* tissues is known ([@R27]), and we confirmed the cyclical accumulation and loss of *per* in the intestine when flies were kept on a 12 hr light/12 hr dark (LD) regimen (all of the experiments described below were performed under LD and chemical damage unless otherwise noted). Quantitative RT-PCR (qRT-PCR) confirmed that *per* mRNA accumulates in the early evening (zeitgeber time 12--18 \[ZT12--18\]; [Figures 1B](#F1){ref-type="fig"} and [S1A](#SD4){ref-type="supplementary-material"}), and staining for PER confirmed its nuclear accumulation in the late night/early morning ([Figure 1C](#F1){ref-type="fig"}, ZT0). PER is expressed in the epithelial cells of this tissue (the polyploid ECs as well as the diploid ISCs; [Figures 1D](#F1){ref-type="fig"} and [S1](#SD4){ref-type="supplementary-material"}). The Clock Gene *per* Regulates Rhythmic Intestinal Regeneration {#S4} --------------------------------------------------------------- The *per^01^* allele is a loss-of-function nonsense mutation ([@R13]). Although they are viable, *per^01^* mutant animals do not exhibit circadian gene expression or behavioral rhythmicity ([Figures 1B](#F1){ref-type="fig"}, [S1, and S2](#SD4){ref-type="supplementary-material"}). We assayed the regenerative response of *per^01^* ISCs following damage by DSS. Only the ISCs in the *Drosophila* intestine divide ([@R24]), and mitotic ISCs were scored by phosphorylated histone H3 positivity. Control (*ry^506^*) ISCs show a peak in mitoses occurring at dawn ([Figure 1E](#F1){ref-type="fig"}, ZT0), the transition between night and day when PER accumulates. This peak is absent in *per^01^* intestines, which show reduced mitoses at all time points ([Figure 1E](#F1){ref-type="fig"}). A *UAS-per* transgene, which restores circadian rhythms behaviorally when expressed in pacemaker neurons ([Figure S2](#SD4){ref-type="supplementary-material"}), partially restored the mitotic peak in *per^01^* when expressed in ISCs using *esg-Gal4* ([@R4]), but not in ECs using *myo1A-Gal4* ([@R15]; [Figure 1E](#F1){ref-type="fig"}). Importantly, the *esg-Gal4* and *myo1A-Gal4* drivers are not expressed in pacemaker neurons, and do not rescue *per^01^* arrhythmic behavior when driving *UAS-per* ([Figure S2](#SD4){ref-type="supplementary-material"}). A characteristic of circadian rhythms is their free-running nature ([@R12]), which we tested by shifting flies to constant darkness (DD) after LD entrainment. PER expression rhythms and intestinal mitotic rhythms perpetuate in DD, demonstrating their circadian nature (see [Figures S1F and S5A--S5C](#SD4){ref-type="supplementary-material"}). Together, these results show that ISCs divide according to a circadian rhythm in response to damage, and that this response is *per* dependent. Undamaged *per^01^* intestines do not show obvious deficiencies in epithelial cell types ([Figures S3A and S3B](#SD4){ref-type="supplementary-material"}) or rhythmic mitoses (see [Figure 4C](#F4){ref-type="fig"}). Both ISCs and ECs participate in regeneration ([@R5]), raising the question as to which cells are responsible for the inability of *per^01^* intestines to display mitotic rhythms. A second important question is whether mitotic rhythms in response to damage are linked to behavioral activity or feeding ([@R38]). We validated a *UAS-per RNAi* construct for its ability to reduce PER expression and abolish circadian behavior rhythms ([Figure S2](#SD4){ref-type="supplementary-material"}). PER knockdown in ISCs phenocopied the arrhythmic *per^01^* intestine ([Figure 1F](#F1){ref-type="fig"}) and, strikingly, PER depletion in ECs also abolished ISC proliferation rhythms ([Figure 1G](#F1){ref-type="fig"}). These phenotypes are not correlated with circadian behavior ([Figure S2](#SD4){ref-type="supplementary-material"}) or feeding ([Figures S3C and S3D](#SD4){ref-type="supplementary-material"}), which are rhythmic (although we do note an 1 hr circadian period lengthening in the *esg-Gal4* driver). Since \~ only ISCs divide in this tissue, *per RNAi* disruption in ISCs ([Figure 1F](#F1){ref-type="fig"}) accounts for the *per^01^* phenotype ([Figure 1E](#F1){ref-type="fig"}), whereas *per RNAi* in ECs simply abolishes a peak at ZT0 ([Figure 1G](#F1){ref-type="fig"}). These results suggest that PER is required separately in both ISCs and ECs to produce intestinal mitotic rhythms, and that these rhythms are separate from feeding and behavioral rhythms. Next, we generated *per*-deficient mutant clones to test whether the defect associated with PER loss was cell autonomous. Following damage, *per^01^* and *per RNAi* clones are slightly smaller ([Figure S4](#SD4){ref-type="supplementary-material"}) and show reduced size over long periods of time in the absence of acute damage. This suggests that PER has a weaker ISC-autonomous role in initiating or boosting proliferation following damage or stress, but that overall a stronger nonautonomous role is predominant. The Core Clock Functions during Intestinal Regeneration {#S5} ------------------------------------------------------- Because *per* and *tim* work together to inhibit *clk*/*cyc*, the outcomes of CYC activity would be expected to oppose those of PER. The *cyc^0^* and *tim^0^* loss-of-function mutants are also viable, and also display altered intestinal mitotic rhythms in response to damage ([Figures 2A and 2B](#F2){ref-type="fig"}). The expression of a *UAS-cyc* transgene in ISCs (*esg-Gal4*) in the *cyc^0^* background was able to partially rescue this phenotype, but expression in ECs (*myo1A-Gal4*) did not ([Figure 2A](#F2){ref-type="fig"}). Although the *cyc^0^* phenotype is the opposite of the *per^01^* phenotype, we note that the *tim^0^* phenotype is not the same as that of *per^01^*, suggesting that *tim* may have additional functions in this tissue. It is also possible that genetic background plays a role in the level of mitoses observed in these conditions. We tested the epistatic relationships between these genes. The *per^01^*;*tim^0^* double mutant displays the *per^01^* phenotype ([Figure 2C](#F2){ref-type="fig"}), and the *cyc^0^*;*per^01^* double mutant displays the *cyc^0^* phenotype ([Figure 2D](#F2){ref-type="fig"}), as would be predicted from the circadian clock transcriptional feedback loop, which undergoes circadian rhythms in this tissue ([Figure S1](#SD4){ref-type="supplementary-material"}). We further tested the requirement of CYC in the regenerative process by expressing a functionally validated *UAS-cyc RNAi* construct ([Figures S2 and S3](#SD4){ref-type="supplementary-material"}) in ISCs and ECs. CYC is required in both of these cell types to produce mitotic rhythms, and the loss of CYC in either ISCs ([Figure 2E](#F2){ref-type="fig"}) or ECs ([Figure 2F](#F2){ref-type="fig"}) abolished any rhythms observed. Light levels entrain the circadian clock, and when flies are exposed to light-only (LL) conditions, the rhythmic nature of mitoses is abolished and remains constant at all time points ([Figures S5E--S5G](#SD4){ref-type="supplementary-material"}). Altogether, these data confirm that the circadian clock is required in both ISCs and their EC neighbors for mitotic rhythms. Bleocin is a potent DNA-damaging chemical that causes apoptosis in the intestine ([@R4]), and it was applied to investigate the outcome of a circadian-deficient damage response. Following Bleocin-induced damage, mitoses in control versus *cyc^0^* and *per^01^* mutant flies show phenotypes similar to those observed under DSS ([Figure 2G](#F2){ref-type="fig"}). The *cyc^0^* mutants exhibit reduced survival on Bleocin ([Figure 2H](#F2){ref-type="fig"}) or DSS ([Figure S5](#SD4){ref-type="supplementary-material"}), and *per^01^* and *tim^0^* show similar reduced survival ([Figures 2H, 2I](#F2){ref-type="fig"}, and [S5](#SD4){ref-type="supplementary-material"}). The knockdown of CYC or PER within ISCs or ECs results in reduced survival on Bleocin ([Figures 2J and 2K](#F2){ref-type="fig"}). Hence, the disruption of the circadian clock either throughout the body or only in ISCs or ECs negatively impacts the survival of animals when the intestine is damaged. Clock-Deficient ISCs Lag in the Cell Cycle during Regeneration {#S6} -------------------------------------------------------------- The accumulation of mitotic *cyc^0^* ISCs ([Figure 2A](#F2){ref-type="fig"}) suggests that loss of *cyc* throughout the animal causes ISCs to overproliferate or stalls these cells in mitosis. An EdU uptake assay, which measures cells in S phase, revealed that control (*ry^506^*) ISCs show a peak in S phase at ZT6. The *cyc^0^* and *per^01^* mutants do not exhibit any peaks, and *cyc^0^* mutants do not exhibit increased S phase ([Figures 3A and 3B](#F3){ref-type="fig"}). Hence, it is unlikely that *cyc^0^* ISCs overproliferate, and *cyc RNAi* clones also did not show an over-proliferation phenotype ([Figure S4](#SD4){ref-type="supplementary-material"}). We applied the FUCCI cell-cycle reporter ([@R21]; [@R32]), which accumulates *mAG-Geminin* during S/G2/M phases (*Azami Green* positive), to determine cell-cycle states when circadian rhythms are absent in ISCs. We expressed the FUCCI reporter along with *cyc RNAi* or *per RNAi* with *esg-Gal4*, and identified ISCs using Dl+. The control RNAi lines show a gradual accumulation of S/G2/M-phase-positive ISCs up to ZT18, when these cells divide ([Figures 3C and 3D](#F3){ref-type="fig"}). However, not all ISCs are in S/G2/M phases, indicating that a significant reserve population of ISCs exists at all times. Irrespective of time, nearly all *cyc RNAi* ISCs are S/G2/M phase negative, whereas nearly all *per RNAi* ISCs are positive. Because its loss causes ISCs to accumulate in G1 (or G0), these results suggest that CYC promotes the G1 to S phase transition. Conversely, when PER is lost, movement through G1 is unopposed, but ISCs accumulate after S phase entry without entering mitosis (see [Figure 1F](#F1){ref-type="fig"}). Thus, we propose that the circadian clock regulates the G1 to S phase transition in ISCs following damage. The Clock Regulates the Transcription of Hundreds of Genes in the Intestine {#S7} --------------------------------------------------------------------------- More than 10% of all mammalian genes are regulated in a circadian fashion ([@R26]), and components of the clock directly regulate transcription in a tissue-specific manner ([@R1]; [@R2]), suggesting that a tremendous variety of cell states are outcomes of circadian processes. Since *per* RNA and protein oscillate in the midgut, and *per* was identified in our screen, we performed genome-wide expression analysis on *ry^506^* control intestines and *cyc^0^* mutants over 24 hr following damage ([Figure 4A](#F4){ref-type="fig"}; [Tables S1](#SD1){ref-type="supplementary-material"}, [S2](#SD2){ref-type="supplementary-material"}, and [S3](#SD3){ref-type="supplementary-material"}). We reasoned that clock target genes would show 24 hr rhythms and would be perturbed if CLK/CYC were disrupted. We found that 433 genes were rhythmic in controls, like *per*, but arrhythmic in *cyc^0^*, indicating that they are under clock regulation in this tissue ([Table S1](#SD1){ref-type="supplementary-material"}). For instance, *Connector of kinase to AP-1* (*Cka*), a scaffold protein required for signal transduction of the JNK stress-response pathway ([@R8]), peaks at ZT15 ([Figure 4B](#F4){ref-type="fig"}). Direct CLK/CYC targets would be expected to be strongly reduced in *cyc^0^* mutants, yet only 21 of 433 genes (including *per* and *tim*) fit this profile ([Table S2](#SD2){ref-type="supplementary-material"}); hence, most rhythmic genes are likely to be indirectly regulated. Two hundred rhythmic genes showed the opposite phase to that of *per*, suggesting they are regulated by the transcription factors *vrille* or *Pdp1*, which are part of the clock and together generate antiphasic transcript rhythms that peak in the early day ([@R12];[Table S1](#SD1){ref-type="supplementary-material"}). One of these, *Ipk2*, is an inositol phosphate kinase and a positive regulator of Jak/STAT signaling ([@R20]), a pathway that is critical during intestinal regeneration ([Figure 4B](#F4){ref-type="fig"}). Another one of these genes, *bazooka*, was recently reported to polarize ISCs ([@R10]), suggesting that the clock also regulates cell polarity. An additional 205 genes showed low expression in *cyc^0^* mutants but did not display rhythms ([Table S3](#SD3){ref-type="supplementary-material"}). This includes *Kmn1*, which enables chromosome segregation during anaphase ([@R36]), suggesting that mitosis could be disrupted ([Figure 4B](#F4){ref-type="fig"}). Overall, a great diversity of intestinal transcripts are thus influenced by the clock. DISCUSSION {#S8} ========== Circadian pathway mutants are viable and their cells readily proliferate during development. Unlike other tissues ([@R1]; [@R6]), cell-cycle regulators do not seem to be clock targets in the intestine ([Table S1](#SD1){ref-type="supplementary-material"}). Although they are readily detected, neither cyclins nor regulators such as *Wee1* ([@R17]) exhibit circadian rhythms in this tissue. In the absence of acute damage, clock mutant ISCs divide normally ([Figure 4C](#F4){ref-type="fig"}) and have no ISC-autonomous phenotypes ([Figure S4](#SD4){ref-type="supplementary-material"}). So it is quite surprising that PER and CYC are critical for adult ISC division during regeneration. The ISC-autonomous phenotypes that occur during regeneration are modest compared with those that arise when the clock is disrupted systemically or in all ISCs/ECs by RNAi. This suggests that the clock predominantly regulates nonautonomous functions and may be involved in the synchronization of cell states across this tissue during the damage response. Indeed, because *esg-Gal4* is expressed in both ISCs and their immediate progeny (the EBs) for some time while they differentiate, it is possible that the clock regulates EB-to-ISC signaling. Intriguingly, disruption of the circadian clock in different cells leads to the accumulation of ISCs in different cell states; for instance, the *cyc^0^* mutant stalls during mitosis when CYC is absent systemically ([Figure 2A](#F2){ref-type="fig"}), whereas it stalls during G1 if CYC is depleted in all ISCs ([Figures 3C and 3D](#F3){ref-type="fig"}). This G1 lag explains why *cyc RNAi* ISCs show reduced mitoses compared with the *cyc^0^* mutant; however, given that the mechanisms underlying these processes are unresolved, it is possible that these differences are due to genetic background. At present, we thus conclude that rhythmic cell proliferation normally occurs in the damaged intestine and that this is dependent on the clock. We also note that forced expression of *per* or *cyc* in ISCs is able to partially restore rhythmic divisions in their respective mutant backgrounds ([Figures 1E](#F1){ref-type="fig"} and [2A](#F2){ref-type="fig"}), whereas disruption of these genes in only ECs perturbs ISC rhythmic division ([Figures 1G](#F1){ref-type="fig"} and [2F](#F2){ref-type="fig"}). This highlights the complexity of clock-regulated processes and suggests that desynchrony between ISCs and their surrounding cells ([Figures S1G and S1H](#SD4){ref-type="supplementary-material"}) can have different outcomes. Circadian rhythms occur in many intertwined processes, including metabolism ([@R30]), post-transcriptional regulation ([@R16]), and oxidation-reduction cycles ([@R23]). The rhythmic expression of *Cka*, which brings together kinases and transcription factors to transduce JNK signal ([@R8]), and *Ipk2*, which may boost the activity of cytokines involved in regeneration ([@R20]), suggests that the clock sensitizes the intestine to engage the regenerative response at specific times. For instance, several of the genes that exhibit circadian rhythms during regeneration also show these rhythms prior to damage ([Figure 4D](#F4){ref-type="fig"}). An emergent function of the clock could be to coordinate stem cell states according to either local niche signals or systemic signals, each of which would be under autonomous circadian control ([Figure 4E](#F4){ref-type="fig"}). Although *per* mutation increases cancer incidence ([@R6]; [@R9]; [@R37]) and cancer cell proliferation ([@R6]; [@R14]), our work suggests it is not simply a tumor suppressor. Recently, the circadian clock was shown to influence mammalian blood and hair stem cell biology ([@R14]; [@R19]). In particular, hair stem cells are strikingly heterogenous in their circadian rhythm activity ([@R14]), for unknown reasons. The coordination of proliferation, by synchronizing internal with external rhythms, may thus represent an important difference between normal stem cells and neoplastic cells. EXPERIMENTAL PROCEDURES {#S9} ======================= Animals were maintained at 25°C under LD conditions and damaged by being fed 5% w/v DSS (MP Biomedicals) or 25 μg/mL Bleocin (Calbiochem). The flies were maintained under LD conditions as before, except for experiments in which the light conditions were changed to complete darkness or complete light. Female flies \< 14 days of age were used in all experiments, with the exception of the mosaic analysis. The following *Drosophila* lines were used: OreRry^506^y, wcyc^0^, ry^506^per^01^;; ry^506^per^01^; tim^0^; ry^506^per^01^;; cyc^0^, ry^506^y, w; tim^0^UAS-per16UAS-cyc6esg-Gal4esg-Gal4, UAS-eGFP, tub-Gal80^TS^myo1A-Gal4tim-Gal4hsFlp, FRT19A, tub-Gal80; act \< y+ \< Gal4, UAS-GFP / CyOhsFlp; act \> CD2 \> Gal4, UAS-nlsGFP / Cyow; UAS-dcr2 (II)w; UAS-dcr2 (III)UAS-S/G2/M-Green / CyO*cyc RNAi* (National Institute of Genetics \#8727R-1, Mishima, Shizuoka, Japan)*per RNAi* (TRiP \#JF01226, Harvard Medical School, Boston, USA).*Luc RNAi* (TRiP \#JF01355, Harvard Medical School, Boston, USA). Full details regarding the procedures are provided in Extended Experimental Procedures. Supplementary Material {#SM} ====================== Stocks and antibodies were provided by Drs. Paul Hardin, Michael Rosbash, and Stephen Hou, and the Bloomington *Drosophila* Stock Center. We thank members of the Perrimon laboratory, particularly Richard Binari and Akhila Rajan, for their assistance. This work was supported by the Human Frontier Science Program (P.K.) and the Harvard Stem Cell Institute. P.E. is supported by NIH grants GM66777 and GM79182. N.P. is an investigator of the Howard Hughes Medical Institute. SUPPLEMENTAL INFORMATION Supplemental Information includes five figures, three tables, and Extended Experimental Procedures and can be found with this article online at <http://dx.doi.org/10.1016/j.celrep.2013.03.016>. {ref-type="supplementary-material"} for schematic), control (*ry^506^*) intestinal mitoses peak at ZT0, in contrast to *per^01^*. A *UAS-per* rescue construct expressed in ISCs using *esg-Gal4* rescues this effect partially in the *per^01^* background.\ (F) Rhythms are present in Luciferase (*esg \> Luc RNAi* is *esg-Gal4/+; UAS-dcr2/UAS-Luc RNAi*) controls, but PER knockdown in ISCs (*esg \> per RNAi* is *esg-Gal4/+; UAS-dcr2/UAS-per RNAi*) phenocopies *per^01^*.\ (G) PER knockdown in ECs also disrupts circadian mitotic rhythms (genotypes as above but with *myo1A-Gal4/+*).\ See also [Figures S1, S2, S3, and S5](#SD4){ref-type="supplementary-material"}.](nihms-569196-f0001){#F1} {ref-type="fig"}.\ (C and D) *per^01^*; *tim^0^* double-mutant intestines resemble the *per^01^* mutant phenotype. *per^01^*; *cyc^0^* double-mutant intestines resemble the *cyc^0^* mutant phenotype. Control and mutant data are duplicated from [Figures 1E](#F1){ref-type="fig"}, [2A, and 2B](#F2){ref-type="fig"}.\ (E and F) CYC knockdown in ISCs (*esg \> cyc RNAi* is *esg-Gal4/+; UAS-dcr2/UAS-cyc RNAi*) or in ECs (*myo1A \> cyc RNAi* is *myo1A-Gal4/+; UAS-dcr2/UAS-cyc RNAi*) disrupts circadian mitotic rhythms. Control data are from [Figures 1F and 1G](#F1){ref-type="fig"}. All graphs show the average of two separate experiments (n = 10 guts/genotype/time point, error bars ± SEM, \*p \< 0.05 at ZT0).\ (G) Following Bleocin exposure, control (*ry^506^*) intestinal mitoses peak at ZT0, in contrast to *per^01^* and *cyc^0^*, similarly to what happens following DSS damage.\ (H and K) The survival rates of all circadian clock mutants as well as animals in which PER or CYC was knocked down by RNAi in either ISCs or ECs are reduced compared with controls on Bleocin (black lines). Graphs show representative experiments (n = 3 vials, 15 flies per vial; genotypes as above).\ See also [Figures S1, S2, S3, and S5](#SD4){ref-type="supplementary-material"}.](nihms-569196-f0002){#F2} {ref-type="supplementary-material"}.](nihms-569196-f0003){#F3} {ref-type="fig"}.\ (C) Flies maintained in LD conditions on regular media do not show a mitotic peak at ZT0, in contrast to when the intestine is damaged. Under these conditions the mitotic index is similar between *ry^506^* controls and *cyc^0^* or *per^01^* mutants.\ (D) In the absence of damage, the expression of *Cka* and *Ipk2* (qPCR) is rhythmic, similar to what is observed during regeneration. *Kmn1* (qPCR) also shows lower expression both before and after damage.\ (E) A model of how the clock synchronizes ISC division: CYC is important for the transition through G1, and the clock also initiates systemic signals and local niche signals originating from ECs. Together, these signals activate ISC divisions, most likely through nonautonomous mechanisms.\ See also [Figures S2, S3, and S5](#SD4){ref-type="supplementary-material"}.](nihms-569196-f0004){#F4}

\section{Introduction} Symmetry-protected topological states (SPTs) has attracted lots of interests in recent years. These states do not break any Hamiltonian symmetry and are fully gapped in the bulk. In the presence of boundary, SPTs are characterized by gapless boundary modes. Importantly, as long as the protection symmetry is unbroken the gapless boundary states are protected.\\ In the presence of spontaneous symmetry breaking the symmetry group of the Hamiltonian is broken down to a subgroup. Due to the protection by this subgroup, symmetry breaking phases can also be divided into different topological classes. Transitions between topologically inequivalent symmetry-breaking phases are also either first order or continuous quantum phase transitions. Moreover, the interface between different topological phases must also harbor gapless modes. All of these features are the same as SPTs.\\ In the rest of the paper we consider two models of symmetry-breaking topological insulators. The first model is a gapped version of the Hamiltonian introduced in Ref.\cite{Laughlin}, the second model is introduced in Ref.\cite{Hsu}. We will show, after low level of p-type doping doping, these models exhibit hole pockets centered along the Brillouin zone diagonals. Interestingly, they also have the potential of explaining the unusual thermal Hall effect reported in Ref.\cite{Taillefer-2019}. \\ In Ref.\cite{Taillefer-2019} it is shown that La$_{2-x}$Sr$_x$CuO$_4$ at $x=0.06$ exhibits an unusual thermal Hall conductivity ($\kappa_{xy}$). This sample is superconducting below $5 K$ and situates close to the boundary of the antiferromagnetic phase. At low temperatures $\kappa_{xy}/T$ is negative and the magnitude rises monotonically with the magnetic field strength. This thermal Hall conductivity is apparently not due to charge carriers. Because according to the Wiedemann-Franz law the latter contribution is negligible. Importantly, this unusal thermal Hall effect is also observed in other cuprate compounds including La$_{1.6-x}$Nd$_{0.4}$Sr$_x$CuO$_4$, La$_{1.8-x}$Eu$_{0.2}$Sr$_x$CuO$_4$, and Bi$_2$Sr$_{2-x}$La$_x$CuO$_{6+\delta}$ under restricted conditions. The conditions are (1) the doping concentrations exclude those exhibiting charge order, and (2) the values of temperature and magnetic field are such that superconductivity is suppressed. Most surprisingly, under a 15 T magnetic field the temperature dependence of $\kappa_{xy}/T$ for the undoped La$_2$CuO$_4$ is very close to that of La$_{2-x}$Sr$_x$CuO$_4$ at $x=0.06$, suggesting a similar anomalous thermal Hall effect in the parent compound of cuprates ! \\ \section{The model} \subsection{Model 1: a modified DDW\cite{Laughlin}} The Hamiltonian is given by \begin{eqnarray} &&H = H_0 + H_{\textrm{s-DDW}} \nonumber\\ &&H_0 = -t_1 \sum_{\avg{ij}} C^\dagger_{i\alpha} C_{j\alpha} -t_2 \sum_{\avg{\avg{ij}}} C^\dagger_{i\alpha} C_{j\alpha} + h.c. \nonumber\\ &&H_{\textrm{s-DDW}} =\sum_i(-1)^{i_x+i_y}\Big\{i~m_2\Big[C^\dagger_{i\alpha} C_{i+x\alpha} -C^\dagger_{i\alpha}C_{i+y\alpha}\Big]\nonumber\\&&+\vec{m}_1\cdot\vec{{\sigma}}_{\alpha\beta}\Big[C^\dagger_{i\alpha}C_{i+x+y\beta} - C^\dagger_{i\alpha} C_{i-x+y\beta}\Big] + h.c.\Big\} \nonumber\\ \label{model} \end{eqnarray} Here $C_{i,\alpha}$ annihilates a spin $\alpha$ electron on site $i$ of the square lattice, and ${\sigma}^{x,y,z}$ are the Pauli matrices. The repeated spin indices $\alpha,\beta$ imply summation. $H_0$ describes the dispersion of the Zhang-Rice singlet band. The hopping amplitudes between nearest-neighbor and next-nearest-neighbor sites are $t_1$ and $t_2$, respectively. In the rest of the paper we set $t_1 = 1$ and $t_2 = -0.1$, and denote the values of all other energy parameters in unit of $t_1$. In $H_{\rm s-DDW}$, the term proportional to $im_2$ induces a spin-independent checkerboard pattern of electric current. This explains the nomenclature ``s-DDW'', i.e., ``singlet DDW''. In the absence of $\vec{m}_1$ the energy spectrum is nodal, with the nodes centered along the Brillouin zone diagonals. In Ref.\cite{Laughlin} this feature is regarded as the signature of pseudogap. The order parameter $\vec{m}_1$ is absent in Ref.\cite{Laughlin}. It describes a spin-dependent second neighbor hopping. After fixing the direction of $\hat{m}_1$, the hopping amplitude has opposite sign in the (1,1)/(1,-1) directions and modulates with momentum $(\pi,\pi)$. In addition, the hopping amplitudes change sign when electron's spin polarization along $\hat{m}_1$ reverses. We schematically represent $H_{\rm s-DDW}$ in \Fig{fig1}(a). A spatially uniform $\vec{m}_1$ opens a gap in the energy spectrum. Moreover, as long as $|\vec{m}_1|$ is small compared with $m_2$, doping will create Fermi pockets around the nodes. \\ \begin{figure}[t] \includegraphics[scale=0.33]{fig1} \caption{ (a) The real-space representation of $H_{\rm s-DDW}$ in \Eq{model}. The black arrows represent positive imaginary hopping along the designated direction. The hopping amplitude associated with the opposite direction is the complex conjugate. The blue(red) dash line represents positive (negative) real hopping amplitudes. Note, the translation and rotation symmetries are broken. (b) The projected band structure of \Eq{model} with periodic boundary condition in the $\hat{x}$ and open boundary condition in the $\hat{y}$ directions, respectively. The number of rows in y-direction is $n_y=100$. The parameters used are $\vec{m}_1 = 0.15 \hat{z}, m_2 = 0.5$. Each of the two in-gap edge branch is two-fold degenerate. This degeneracy is due to the presence of two different edges. (c) The projected band structure of \Eq{model} in the presence of a magnetic field in the $\hat{z}$ direction. The associated Zeeman energy is set to $0.2$, which causes the edge modes to be Zeeman split. (d) Under an external magnetic field the magnitude of the spin up and spin down edge currents are no longer equal. This results in a boundary circulating current in the disk geometry. } \label{fig1} \end{figure} The above model should be viewed as the mean-field theory of certain interacting Hamiltonian similar to that discussed in Ref.\cite{Laughlin2}. The order parameter $\vec{m}_1$ breaks the translation and 4-fold rotation symmetries of the lattice. Moreover, it also breaks the SU(2) spin rotation symmetry down to U(1), namely, rotation around the $\hat{m}_1$ axis. The order parameter $m_2$ breaks the translation and time reversal symmetry. However, $H_{\rm s-DDW}$ respects the combined operation of time reversal and translation. \\ \subsubsection{The edge states} In \Fig{fig1}(b) we plot the energy spectrum of \Eq{model} with $m_1=0.15\hat{z}$ and $m_2=0.5$ in the cylindrical geometry, namely, open boundary condition along $\hat{y}$ and periodic boundary condition along $\hat{x}$. Here $\hat{x}$ and $\hat{y}$ are 45 degrees rotated from the principal axes of the square lattice, and $k_x$ is a momentum in the antiferromagnetic Brillouin zone. There is a pair of counter-propagating helical edge modes localized on each of the two edges, as shown in \Fig{fig1}(b). They are reminiscent of the edge modes in a quantum spin Hall insulator. These edge modes are protected from back scattering by the residual U(1) spin rotation symmetry, hence the system is a topological insulator. In \Fig{fig1}(c) we plot the energy spectrum in the presence of a $z$-diection magnetic field. Clearly, the edge modes are Zeeman split. \\ Like the quantum spin Hall insulator, the electric Hall conductance of model 1 is zero. However, due to the Zeeman splitting, a magnetic field induces a non-zero current on each edge. This is because the spin up and spin down edge electron density are no longer equal. However, in the cylindrical geometry, this magnetic-field-induced edge current cancels among the two edges. In the disk geometry, the magnetic-field-induced edge current circulates around the perimeter, as shown in \Fig{fig1}(d). This edge current implies the presence of a bulk orbital magnetization.\\ \begin{figure}[t] \includegraphics[height=1.1in]{fig2} \caption{ (a) The Fermi surface of the Hamiltonians in \Eq{model} and \Eq{model2} for doping level $p=0.06$. The parameters used are $\vec{m}_1 = 0.15 \hat{z}, m_2 = 0.5$ for \Eq{model} and $m_1 = 0.15, \vec{m}_2 = 0.5\hat{z}$ for \Eq{model2}. The blue solid line represents the Fermi surface and the yellow dashed line encloses the antiferromagnetic Brillouin zone. (b) The thermal Hall conductivity $\kappa_{xy}/T$ (in units of $k_B^2/\hbar$) as a function of temperature $T$ at several magnetic fields $B$ for doping level $p=0.06$. (c) $\kappa_{xy}/T$ (in units of $k_B^2/\hbar$) as a function of applied magnetic field $B$ at several temperatures for doping level $p=0.06$. } \label{fig2} \end{figure} \subsubsection{The thermal Hall effect and the Fermi pockets} Following Ref.\cite{Lee-2019} we show that upon doping \Eq{model} exhibits an unusual thermal Hall effect. Doping is achieved by adding a chemical potential term to $H_0$, namely $-\mu\sum_{i}C^+_{i\alpha}C_{i\alpha}$. In the first version of the manuscript we attribute the thermal Hall effect to the edge thermal conduction. This leads to the conclusion that the thermal conductivity is non-zero even in the insulating state. The authors of Ref.\cite{Lee-2019} pointed out to us that the thermal conduction due to the helical edge states should be negligible for weak fields. This is because despite the Zeeman shift, the energy current due to the particle-hole excitations near the chemical potential are the same for both spins (due to the cancellation between the density of states and the Fermi velocity in 1D). Thus the spin up and spin down electron's contributions to the thermal conductivity cancel. However, when the chemical potential lies within the bulk bands, and when the Berry curvature is non-zero in the energy range of [$E_f-B,E_f+B$] the bulk thermal Hall conductivity is non-zero. However, this bulk contribution requires finite doping. \\ It can be shown straightforwardly that for \Eq{model} with $(\vec{m}_1,m_2)=(m_1\hat{z},m_2)$ the energy dispersion and the Berry curvature $B_{n\alpha}(\v k)$ are given by \begin{eqnarray} &&E_{n\alpha}(\v k)= -2t_2 (c_{x+y}+c_{x-y})-\mu - 2 n R_{x,y}-\alpha\mu_B B\nonumber\\ &&B_{n\alpha}(\v k) = n\alpha\left(2m_2 m_1 t_1(s^2_x + s^2_y -s^2_x s^2_y\right)/R_{x,y}^3\nonumber\\&&R_{x,y}=\sqrt{t_1^2(c_x+c_y)^2+m_2^2(c_x-c_y)^2+4m_1^2s_x^2s_y^2} \label{disp}\end{eqnarray} Here, $n=\pm 1$ refers to the lower and upper band, and $\alpha=\pm 1$ are the spin polarization along the $\hat{z}$ (i.e., $\hat{m}_1$) direction. In addition, we have used the abbreviations $c_{x(y)}=\cos k_{x(y)}$, $s_{x(y)}=\sin k_{x(y)}$,$c_{x\pm y} = \cos (k_x \pm k_y)$. In terms of $E_{n\alpha}(\v k)$ and $B_{n\alpha}(\v k)$ the thermal Hall conductivity (in units of $k_B^2/\hbar$) is given by\cite{Niu}: \begin{eqnarray} &&\frac{\kappa_{xy}}{T}=\frac{1}{4T^3}\int d\epsilon \frac{(\epsilon-\mu)^2}{\cosh^2[\beta(\epsilon-\mu)/2]} (\sigma_{xy\uparrow}(\epsilon)+\sigma_{xy\downarrow}(\epsilon))\nonumber\\&&\sigma_{xy\alpha}(\epsilon) = -\sum_{nk} B_{n\alpha}(\v k)\theta(\epsilon - E_{n\alpha}(\v k)). \label{kappa} \end{eqnarray} In the following we adjust the chemical potential $\mu$ so that the doping level is $p=0.06$. In \Fig{fig2}(a) we show the Fermi surface for this doping level. It consists of hole pockets centered along the Brillouin zone diagonals. In \Fig{fig2}(b) we show $\kappa_{xy}/T$ as a function of temperature at several magnetic field values. First, the sign of $\kappa_{xy}$ is negative. Second, at a fixed magnetic field $|\kappa_{xy}|/T $ increases with decreasing temperature. In \Fig{fig2}(c) we show the dependence of $\kappa_{xy}/T$ as a function of magnetic field at different temperatures. The result monotonic increases with $B$. Features (a)-(c) are consistent with what's seen in Ref.\cite{Taillefer-2019}. \\ A more stringent test of the theory is the actual size of the predicted $\kappa_{xy}/T$. According to Fig. 1(b) of Ref.\cite{Taillefer-2019}, under a 15T magnetic field the $|\kappa_{xy}/T|$ at the lowest measurement temperature is about 0.7 $k_B^2/\hbar$ per copper-oxide plane. If we set $t_1\sim$ 200 meV, 15T corresponds to $B=0.0075 t_1/\mu_B$ and $T_{\rm min}=14K$ corresponds to $T= 0.007 t_1/k_B$. We have checked that for these parameters the largest $|\kappa_{xy}/T|$ obtainable by varying $|\vec{m}_1|$ at a fixed $m_2=0.5$ is 0.1 $k_B^2/\hbar$. \\ Thus \Eq{model} has the potential to explain the following two very unusual experimental features observed in the underdoped regime of the cuprates where there is no charge order. (1) Hole pockets centered along the Brillouin zone diagonals with area equal to the doping concentration. (2) The anomalous thermal Hall effect observed in Ref.\cite{Taillefer-2019}. \\ In addition, \Eq{model} also predicts the existence of a checkerboard pattern of staggered orbital magnetic moments. These moments have been experimentally searched for, but so far there is no convincing evidence for it. For this reason we proceed to consider the ``tripet-DDW'' model in the following section. \section{Model 2\cite{Hsu}: a modified triplet-DDW} The model introduced in Ref.\cite{Hsu} is given by \begin{eqnarray} &&H = H_0 + H_{\textrm{t-DDW}} \nonumber\\ &&H_0 = -t_1 \sum_{\avg{ij}} C^\dagger_{i\alpha} C_{j\alpha} -t_2 \sum_{\avg{\avg{ij}}} C^\dagger_{i\alpha} C_{j\alpha} + h.c. \nonumber\\ &&H_{\textrm{t-DDW}} =\sum_i(-1)^{i_x+i_y}\Big\{(i~\vec{m}_2\cdot\vec{{\sigma}}_{\alpha\beta})\Big[C^\dagger_{i\alpha} C_{i+x\beta} \nonumber\\&&-C^\dagger_{i\alpha}C_{i+y\beta}\Big]+m_1\Big[C^\dagger_{i\alpha}C_{i+x+y\alpha} - C^\dagger_{i\alpha} C_{i-x+y\alpha}\Big] + h.c.\Big\} \nonumber\\ \label{model2} \end{eqnarray} Here the term proportional to $i \vec{m}_2$ is a spin-dependent DDW order parameter (hence the nomenclature of ``t-DDW'', i.e., ``triplet DDW''). The important difference with the model in \Eq{model} is the cancellation of the orbital magnetic moments because the pattern of circulating current is opposite for spin up and spin down electrons. Thus it removes the unwanted feature of a predicted, but unobserved, orbital magnetic moment. The order parameter proportional to $m_1$ is a spin-independent second neighbor hopping. It also opens an energy gap at the nodes.\\ \begin{figure}[t] \includegraphics[scale=0.28]{fig3} \caption{ (a) The projected band structure of the Hamiltonian in \Eq{model2} with periodic boundary condition in x-direction and open boundary condition in y-direction. The number of rows in y-direction is $n_y=100$. The parameters used are $m_1 = 0.15, \vec{m}_2 = 0.5 \hat{z}$. The edge branches for $k_x>0$ and $k_x<0$ are associated with the spin up and spin down electrons, respectively. For each spin direction the counter propagating branches are localized on different edges. (b) The projected band structure of \Eq{model2} in the presence of a magnetic field in the $\hat{z}$ direction. The associated Zeeman energy is set to $0.2$. } \label{fig3} \end{figure} The term proportional to $m_1$ breaks the translation, 4-fold rotation and mirror symmetries along the x and y axes. In addition, the order parameter $i\vec{m}_2$ breaks the SU(2) spin rotation symmetry down to U(1). However, interestingly, $i\vec{m}_2$ preserves the time-reversal symmetry. This last statement explains why $\vec{m}_2$ does not generate any orbital magnetic moment. It is also the reason why $\vec{m}_2$ is not visible to experimental probes such as neutron scattering and NMR. \\ \\ \subsubsection{The edge states} In \Fig{fig3}(a) we plot the energy spectrum of \Eq{model2} with $m_1=0.15$ and $\vec{m}_2=0.5\hat{z}$ in the cylindrical geometry. Again, on each edge there is a pair of counter-propagating helical edge modes. These edge modes are protected against back scattering by the time reversal and/or the residual spin U(1) symmetries. In the absence of disorder it is also prevented from back scattering because the Fermi momenta of the right and left movers are different. In \Fig{fig3}(b) we plot the energy spectrum in the presence of a magnetic field. Here we have assumed $\hat{m}_2$ to lie in the magnetic field direction, namely, $\hat{z}$. Clearly, the edge modes are Zeeman split. Like model 1, this topological insulator shows zero electric Hall conductance. In the disk geometry there is a magnetic-field-induced circulating boundary current, which reflects the existence of a non-zero bulk orbital magnetization.\\ \subsubsection{The thermal Hall effect and the Fermi pockets} It turns out that for \Eq{model2}, with $(m_1,\vec{m}_2)=(m_1,m_2\hat{z})$, the band dispersion and the Berry curvature are exactly the same as in those for \Eq{model} with $(\vec{m}_1,m_2)=(m_1\hat{z},m_2)$. Therefore at the same doping level ($p=0.06$) and with the same $m_1$ (0.15) and $m_2$ (0.5), the Fermi surface and $\kappa_{xy}/T$ are identical to those shown in \Fig{fig2}. However, model 2 does not possess the staggered orbital magnetic moment.\\ \section{The pinning of $\vec{m}_1$ and $\vec{m}_2$ by the magnetic field} The vector order parameter $\vec{m}_1$ in \Eq{model} and $\vec{m}_2$ in \Eq{model2} are free to rotate without causing any energy. This implies the presence of Goldstone modes. In the presence of these soft modes one needs to worry about the disordering of these vector order parameters at non-zero temperatures (particularly in two spatial dimensions).\\ To address these issues, we focus on zero doping. The generalization to the doped case is straightforward. In the following we shall focus on \Eq{model2}. To obtain the corresponding statements for \Eq{model} one just need to exchange the roles of $m_1$ and $m_2$.\\ As discussed earlier, a non-zero magnetic field induces a bulk orbital magnetization. The latter is given by\cite{vanderbilt} \begin{eqnarray} {\cal M}=-\sum_{\alpha=\hat{m}_2\cdot\vec{{\sigma}}=\pm 1}{e\over h c}C_{\alpha} \Delta E_{Z\alpha}.\label{M}\end{eqnarray} Here $c$ is the speed of light, $h$ is the Planck constant, $e$ is the electron charge, and $C_{\alpha}$ is the Chern number of the spin $\alpha$ band. In addition, the Zeeman energy, $ \Delta E_{Z\alpha}$, is given by $-\alpha\mu_B\hat{m}_2\cdot\vec{B}$ where $\mu_B$ is the effective electron magnetic moment, and $C_{-1}=-C_{+1}$. Since the reversal of the sign of $\alpha$ causes both $C_\alpha$ and $ \Delta E_{Z\alpha}$ to change sign, \Eq{M} can be simplified to \begin{eqnarray} {\cal M}=-{2e\over h c}C_{+1} \Delta E_{Z+1}={2e\mu_B\over h c}C_{+1}\hat{m}_2\cdot\vec{B}.\end{eqnarray} Importantly, the sign of $C_{+1}$ is determined by that of $m_1$, namely, \begin{eqnarray} C_{+1}={m_1\over |m_1|}.\end{eqnarray} Putting these results together we have \begin{eqnarray} {\cal M}={2 e\mu_B\over h c}{m_1\over |m_1|}\hat{m}_2\cdot\vec{B}.\end{eqnarray} The above orbital magnetization interacts with the magnetic field via the Zeeman coupling to yield the following energy density \begin{eqnarray} \Delta {\cal E}_{\rm Zeeman}=-{\mu_B B^2\over \pi c}{m_1\over |m_1|}(\hat{m}_2\cdot\hat{B}).\label{int}\end{eqnarray} \Eq{int} implies that in the presence of a magnetic field it is energetically favorable for $m_1\hat{m}_2$ to point in the same direction as $\hat{B}$. This eliminates the Goldstone modes and fixes the sign of $\kappa_{xy}$. Thus the sign of $\kappa_{xy}$ should not be random among different cool downs.\\ In two space dimensions the SO(3) symmetry breaking in both \Eq{model} and \Eq{model2} are only present in a non-zero applied magnetic field. This provides examples where the zero field and finite field electronic states can be different. In zero magnetic field it is interesting to study the fate of the topological insulators when $\vec{m}_1$ or $\vec{m}_2$ is thermally disordered. This study reveals an important difference between model 1 and model 2. For model 1 the residual U(1) spin symmetry is broken by any disordered configuration of $\vec{m}_1$. Hence we expect the edge states to loose symmetry protection. In contrast, for model 2 the edge states stay protected (by the time reversal symmetry) even when the U(1) spin rotation symmetry is lost. This difference is confirmed by examining the thermal-averaged edge spectral function of model 1 and model 2 in the cylindrical geometry, namely, \begin{eqnarray} &&\overline{A}(k_x,\omega) = \frac{\sum_{\{\vec{m}_{a,i}\}}W[\{\vec{m}_{a,i}\}] A(k_x,\omega)_{\{\vec{m}_{a,i}\}}}{\sum_{\{\vec{m}_{a,i}\}} W[\{\vec{m}_{a,i}\}]}\nonumber\\&&W[\{\vec{m}_{a,i}\}] = e^{-\beta J \sum_{\avg{ij}} \vec{m}_{a,i}\cdot \vec{m}_{a,j}}. \label{thermal} \end{eqnarray} Here $\{\vec{m}_{a,i}\}$, with $a=1 {\rm ~or~} 2$, are the spatial configurations of the vector order parameter in \Eq{model} or \Eq{model2}, and $W[\{\vec{m}_{a,i}\}]$ is the Boltzmann weight. $A(k_x,\omega)_{\{\vec{m}_{a,i}\}}$ is the spectral function under a fixed configuration of $\{\vec{m}_{a,i}\}$. \\ Our calculation is performed after fixing the amplitude $|\vec{m}_1|$ or $|\vec{m}_2|$. We sample the directions of $\hat{m}_{1}$ or $\hat{m}_{2}$ according to the Boltzmann weight by the Metropolis algorithm, and the number of sampled configurations is 30000. As shown in \Fig{fig4}(b,c), the edge modes in \Eq{model} are disorder scattered at non-zero temperatures. In contrast, the edge modes in \Eq{model2} remain sharp as shown in \Fig{fig4}(e,f). We attribute this difference to the fact that for \Eq{model2} thermal disordering of $\vec{m}_2$ does not jeopardize one of the protection symmetry, namely, the time reversal symmetry. \\ \begin{figure}[t] \includegraphics[height=1.75in]{fig4} \caption{ The thermal averaged electron spectral function in models \Eq{model} and \Eq{model2} in a finite cylinder (open boundary condition in $\hat{y}$ and periodic boundary condition in $\hat{x}$). The parameters used are $\vec{m}_1 = 0.15 \hat{z}, m_2 = 0.5$ for \Eq{model} and $m_1 = 0.15, \vec{m}_2 = 0.5\hat{z}$ for \Eq{model2}. The linear dimension of the cylinder is is $n_x = 80$ and $n_y = 20$. The ensembles of $\{\vec{m}_{1,i}\}$ and $\{\vec{m}_{2,i}\}$ are generated with the Boltzmann weight given in \Eq{thermal}. Panels (a)-(c) are the results for \Eq{model} while panels (d)-(f) are for \Eq{model2}. The inverse temperatures used in the calculations are $\beta J = \infty$ (zero temperature) in panels (a) and (d), $\beta J = 16$ and in panels (b) and (e), and $\beta J = 8$ in panels (c) and (f). The total number of sampled configurations is 30000.} \label{fig4} \end{figure} \section{\bf The Neel ordered phase} The topological nature of the model 1 and model 2 survives the presence of the Neel long range order, \begin{eqnarray} H_{\textrm{Neel}} = \sum_i (-1)^{i_x+i_y} \vec{m}_s\cdot\vec{{\sigma}}_{\alpha\beta}C^\dagger_{i\alpha}C_{i\beta},\end{eqnarray} as long as $\vec{m}_s$ is not too strong. For example, in \Fig{fig5}(a) and (b) we show the edge modes dispersion of model 2 in the presence of a non-zero $\vec{m}_s=0.2 \hat{x}$. The parameters used are $m_1 = 0.15$ and $\vec{m}_2=0.5\hat{z}$ in panel (a) and $m_1 = 0.15$ and $\vec{m}_2=0.5\hat{x}$ in panel (b) . \\ Despite the persistence of the edge states, our models predict the absence of thermal Hall effect in the undoped limit, agreeing with the result of Ref.\cite{Lee-2019}. This is because when the sample is undoped, the chemical potential lies in the gap of the Zeeman shifted spin up and spin down spectrum (at least when the Zeeman energy is small compared to the gap energy). Under such condition \Eq{kappa} predicts zero thermal Hall conductance because the $\epsilon$-integrals for spin up and spin down electrons yield values with opposite sign but the same (quantized) magnitude, hence they cancel\cite{Lee-2019}. \begin{figure}[t] \includegraphics[height=1.1in]{fig5} \caption{ The projected band structure of $H_{\rm t-DDW}$ in the presence of Neel order. The number of rows in the open ($\hat{y}$) direction is $n_y=100$. In these plots $\vec{m}_s$ is set to $ 0.2\hat{x}$ and $m_1=0.15$. In panel (a) $\vec{m}_2=0.5\hat{z}$ and in panel (b) $\vec{m}_2=0.5\hat{x}$. } \label{fig5} \end{figure} \section{The effect of residual electronic correlation on the edge states} The main effect of the electronic correlation is to render the system in the mean-field state described by \Eq{model} or \Eq{model2}. In the following we discuss the effects of residual electronic correlation on the edge dynamics. The fact that this is necessary is because the edge modes are gapless. \\ The edge Hamiltonian is given by \begin{eqnarray} H_E = iv_F \int dx~\Big[\psi^\dagger_{L\downarrow}(x) \partial_x \psi_{L\downarrow}(x) - \psi^\dagger_{R\uparrow}(x) \partial_x \psi_{R\uparrow}(x)\Big]\nonumber\\ \label{freefermion} \end{eqnarray} where $\psi_{L\downarrow}$ and $\psi_{R\uparrow}$ are the annihilation operators of the left (spin down) and right (spin up) moving edge electrons, and $v_F$ is the mean-field edge velocity. Due to the time reversal and/or the residual spin U(1) rotation symmetry, the single-particle backscattering terms, $\psi^\dagger_{R\uparrow} \psi_{L\downarrow} + \psi^\dagger_{R\uparrow} \psi_{L\downarrow} $, and $i \psi^\dagger_{R\uparrow} \psi_{L\downarrow} - i\psi^\dagger_{R\uparrow} \psi_{L\downarrow}$, are not allowed.\\ The most relevant, symmetry-allowed, four fermions interactions is given by \begin{eqnarray} &&H_{\textrm{int}} = \int dx~\Big\{~g_2 \psi^\dagger_{R\uparrow}(x)\psi_{R\uparrow}(x) \psi^\dagger_{L\downarrow}(x) \psi_{L\downarrow}(x) \Big] \Big\}. \end{eqnarray} It renormalizes the edge velocity and the Luttinger liquid parameter: \begin{eqnarray} &&v=v_F\sqrt{(1-\frac{g_2}{v_F})(1+\frac{g_2}{v_F})}\nonumber\\&& K = \sqrt{\frac{v_F-g_2}{v_F+g_2}}.\end{eqnarray} \\ The usual process that opens the charge gap is the umklamp scattering $g_u\psi^\dagger_R \psi^\dagger_R \psi_L \psi_L$. It is forbidden, due to the Fermi statistics, in the present situation due to the spin-momentum locking of the edge electrons. Hence residual correlation does not affect the edge states qualitatively. \\ \noindent{{\bf Final discussions}\\ The topological insulators described by \Eq{model} and \Eq{model2} have the following attractive features. (1) Under low level of p-type doping they predict hole pockets centered along the Brillouin zone diagonals. This is consistent with the Hall coefficient measurement\cite{Taillefer2} which shows a carrier density $p$ rather than $1+p$ in the doping range where the anomalous thermal Hall effect is observed. (2) These models can explain the anomalous thermal Hall effect in all samples {\it except the undoped La$_2$CuO$_4$}. It is also important to point out we did not provide any microscopic justification for the models in \Eq{model} and \Eq{model2}. Whether there exists, e.g., one-band or three-band Hubbard-like models which realize \Eq{model} or \Eq{model2} as the stable mean-field solution is unclear to us at present.\\ Finally, we take note of several related experimental facts. (1) There is a report from thermal transport that the pseudogap temperature $T^*$ coincides with the onset of 90 degree rotation symmetry breaking\cite{Taillefer2}. Could this be due to the symmetry breaking induced by $\vec{m}_1$ in model 1 or $m_1$ in model 2 ? Ref.\cite{Hsieh} reports that in the pseudogap regime, YBCO exhibits inversion symmetry breaking below $T^*$. In addition, the polar Kerr effect suggests the breaking of time reversal symmetry\cite{Kerr}. Although model 1 breaks time reversal symmetry, it does not break inversion. Model 2 does not break time reversal nor inversion. Although it is possible to add inversion and time reversal breaking features to the two models (for example by making $m_1$ complex) we prefer not to do so for the sake of simplicity. Lastly, in an ARPES experiment on Bi2201 a small nodal gap is observed in the doping range close to the AFM phase boundary\cite{Zhou-2013}. Could it be the gap caused by $\vec{m}_1$ (or $m_1$)? \\ Before the end, we take note of three recent interesting theory papers\cite{Sachdev-2018,Lee-2019,Xu-2019} on the same subject. Our theory, in particular model 1, bears a strong resemblance to that in Ref.\cite{Lee-2019}. Our explanation of the thermal Hall conductance is the same as theirs. However there is an important difference between our theory and Ref.\cite{Lee-2019}, namely, the fermions in our theory are the physical electrons.\\ \noindent{{\bf Acknowledgement}} \\ We are in debt to Prof. Steve Kivelson for bringing Ref.\cite{Taillefer-2019} to our attention. In addition, he pointed out two references which eventually lead us to Ref.\cite{Hsu}. We thank Prof. Bob Laughlin for enlightening discussions. He raised the important question concerning the sign of $\kappa_{xy}$ upon different cool down, and told us about the possible existence of Ref.\cite{Hsu}. We thank Prof. Chandra Varma for enlightening discussions including the question on the meaning of the sign of $\kappa_{xy}$. Finally, we are very grateful to the authors of Ref.\cite{Lee-2019} for pointing out that the thermal conduction due to helical edge states should vanish. This work was primarily funded by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Materials Sciences and Engineering Division under Contract No. DE-AC02-05-CH11231 (the Quantum Materials program). We also acknowledge support from the Gordon and Betty Moore Foundation's EPIC initiative, Grant GBMF4545.

\section{Introduction} The two-dimensional $q$-Whittaker particle system on the torus was introduced by Corwin and Toninelli \cite{CT}. The state of this system is interlacing collections of particles on the two-dimensional torus. Particles jump right by one on their row according to exponential clocks whose rates are determined by certain nearest neighbor inter-particle distances -- see (\ref{eq:6prime}) -- as well as a parameter $q\in[0,1)$. As discussed in \cite{CT}, the dynamics can be seen also as a stochastic growth process for a random discrete $(2+1)$-dimensional interface. The mapping is based on the fact that interlacing particle configurations correspond to perfect matchings (dimer coverings) of the hexagonal lattice, and the associated integer-valued height function provides the discrete interface. In the present work, we adopt the point of view of the particle system rather than that of the growth model. These dynamics originated in Borodin and Corwin's study of Macdonald processes \cite{BorCor} wherein they were defined on certain triangular arrays of interlacing particles. Based on that inspiration as well as a recent treatment by Toninelli in \cite{Ton} of the $q=0$ case of this system, \cite{CT} proposed and then verified that certain local Gibbs measures are stationary for the dynamics -- see (\ref{eq:3prime}) and Proposition \ref{Prop1} for a summary of those results. When $q=0$, \cite{Ton} determined that stationary measures are simply uniform measures on particle configurations, which are known to enjoy a relation to determinantal point processes and to have Gaussian Free Field type fluctuations in the infinite volume limit for the torus. In \cite{Ton} it was also shown that the dynamics remain well-defined in the infinite volume limit and bounds on the scale of fluctuations of the associated height function for the system were determined. In particular, it was shown that height function fluctuations grow at a smaller rate than any polynomial in time $t$ (and that the bound can be improved to $O(\sqrt {\log t})$ in a certain range of particle densities). The $q=0$ model is predicted (cf. \cite{BorFer}) to be in the anisotropic $(2+1)$-dimensional Kardar-Parisi-Zhang universality class which is represented by the stochastic PDE \begin{equation}\label{AKPZ} \frac{\partial h}{\partial t}(t;x,y) = \frac{1}{2}\Delta h(t;x,y) + (\nabla h, Q \nabla h)(t;x,y) + \xi(t;x,y) \end{equation} where $h(t;x,y)$ represents a height function at time $t$ and position $(x,y)$, $\Delta$ is the Laplacian in $x$ and $y$, $Q$ is a $2\times 2$ matrix with signature $(+,-)$ and $\xi(t;x,y)$ is a space-time white noise. This is called anisotropic because of the mixed signature on the non-linearity, whereas when the signature of $Q$ is $(+,+)$ or $(-,-)$ the model is called ``isotropic''. In 1991, Wolf \cite{Wolf} predicted that the fluctuations of the anisotropic equation should grow like $\sqrt{\log t}$ and moreover that the non-linearity should be irrelevant and the long-time behavior of the system should be exactly as that of the two-dimensional additive stochastic heat equation (i.e., the equation with $Q$ set to zero). Wolf's prediction of $\sqrt{\log t}$ fluctuations was demonstrated numerically soon after by Halpin-Healy and Assdah \cite{HalpinHealyAssdah}. Pr\"{a}hofer and Spohn \cite{PS} considered a microscopic model related to the Gates-Westcott model, and demonstrated through exact calculation this fluctuation scaling. Borodin and Ferrari \cite{BorFer} studied a (non-periodic) triangular array variant of the $q=0$ case of the $q$-Whittaker particle system and, using the technology of Schur processes they proved the $\sqrt{\log t}$ scaling and further demonstrated convergence to a Gaussian free field as time goes to infinity. The occurrence of a Gaussian free field is consistent with the prediction of convergence to the additive stochastic heat equation since the Gaussian free field is stationary for the time evolution of that stochastic PDE. It should be noted that since \cite{BorFer} dealt with dynamics on a triangular array of particles, the Gaussian Free Field fluctuations only appear after a suitable coordinate change. This coordinate change is not visible from the renormalization group arguments of Wolff. In this paper we focus on translation invariant models for which no coordinate change is necessary. We should note that (\ref{AKPZ}) is not, a priori, well-defined because of the non-linearity and the fact that solutions are not function valued, but rather generalized functions (like the Gaussian free field). We are not aware of any rigorous treatment of this equation, though it may eventually fall into the class of systems which can be defined through Hairer's regularity structures \cite{HairerRegularity}. As such, none of Wolf's predictions have been proved for (\ref{AKPZ}) itself. Let us also note that the story is quite different when the model is isotropic -- see recent numerical studied of Halpin-Healy \cite{HHisotropic}. To our knowledge, the present paper, together with the forthcoming work of Borodin, Corwin and Ferrari \cite{BCF} in the context of these dynamics on triangular arrays, is the first instance in which the full space-time picture has been rigorously established for a model in the $(2+1)$-dimensional anisotropic KPZ class converging to the additive stochastic heat equation. To be upfront about things, we do not prove this convergence as a process (which would require some additional tightness estimates), but rather in terms of the correlation structure for Gaussian processes. The initial motivation for this paper was the desire to extend the study of \cite{BorFer,Ton} to the $q\neq 0$ case. In that case the Schur / determinantal point process structure is lost. To overcome this impediment, we decided to consider a Gaussian limit of the model, hoping that calculations there would become sufficiently doable without said structure. In particular, we consider the $q=e^{-\epsilon}\to 1$ limit of the particle system, as we simultaneously scale the torus width and height like $\epsilon^{-1}$. We start particles spaced on the $\epsilon^{-1}$ scale according to a certain crystalline configuration (see the beginning of Section \ref{sec:conv}) with smaller $\epsilon^{-1/2}$ scale fluctuations. Speeding time up by $\epsilon^{-1}$, we prove (Theorem \ref{th:SDE}) that particle positions (multiplied by $\epsilon$) have asymptotically a constant speed $v$ and that fluctuations (multiplied by $\epsilon^{1/2}$) converge (as a space-time process) to a limiting system of SDEs. Likewise, under this scaling the stationary measure on the $q$-Whittaker system converges to a Gaussian measure (Lemma \ref{th:locmax}). (Note: we do not prove that the stationary measure concentrates on the crystalline configuration, though it is certainly compelling to conjecture this). Once in the setting of SDEs with Gaussian stationary measures (in fact, the stationary measures are on gradients) we are able to use Fourier transforms to explicitly compute the space-time correlations as the number of particles goes to infinity (Theorem \ref{th:2}) as well as the correlations and Gaussian free field limit of the stationary measure (Theorem \ref{th:muinv}). Theorem \ref{th:2} has a number of corollaries. Corollary \ref{cor1} shows that for fixed, large time $t$, fluctuations grow like $\sqrt{\log t}$ and correlations decay in a spatial range of order $t^{1/2}$. Corollary \ref{cor2} considers the correlations along space-time lines. There exists a distinguished direction $U$ along which correlations exist in a temporal scale of order $t$ and a spatial scale of order $t^{1/2}$. In fact, Corollary \ref{cor3} shows that in this scale, the correlations converge to those of the $(2+1)$-dimensional additive stochastic heat equation -- thus validating Wolf's prediction for this model. On the other hand, Corollary \ref{cor2} also shows that for space-time direction not equal to $U$, the correlations decay to zero on a $t^{1/2}$ time-scale, thus much faster than along $U$. We call the direction $U$ the ``characteristic'' direction. The reason is that the lines $y=U t$ are the characteristic curves of a PDE that we conjecture to describe the hydrodynamic limit of our model when the initial condition is not crystalline, see Section \ref{sec:pde}. There is a close analogy with what happens in the context of $(1+1)$-dimensional particle systems. In that setting, the hydrodynamics are described by Hamilton-Jacobi conservation laws which can be solved by computing the characteristics and propagating initial data along them. Characteristics are computed as the derivative of the flux with respect to the local slope. At a more microscopic level, initial data fluctuations are propagated along characteristics. In particular, one has ``slow decorrelation'' \cite{Ferrari,CFPslowdec} along characteristics whereby fluctuations along these space-time directions decorrelate far slower than along other space-time directions. The phenomenon of slow decorrelation along characteristics was conjectured (with some supporting evidence) in \cite{BorFer} to hold for the two-dimensional $q$-Whittaker particle system at $q=0$. The aforementioned results concerning correlations of two-dimensional systems of SDEs are actually proved below in much broader generality. In particular, all results are proved provided that the SDEs take the form of \eqref{eq:2} with the matrix $A$ satisfying the conditions of Proposition \ref{prop:prop}. This could reflect the expected universality of the $(2+1)$-dimensional anisotropic KPZ class. It would be compelling to see if any of these universality results can be proved directly for the general $q$ system, without first taking the SDE limit. Without the Gaussian structure, though, it is unclear how to proceed in this goal. As mentioned above, in the $q=0$ case, there exists a triangular array variant of the $q$-Whittaker particle system which was studied at length in \cite{BorFer} using Schur processes. The triangular variant of the general $q$ case relates to $q$-Whittaker processes \cite{BorCor} and though the system is no longer determinantal, there are many useful formulas provided through the technology of Macdonald processes. In a parallel paper to this, Borodin, Corwin and Ferrari \cite{BCF} develop the analogous $q\to 1$ limit of this triangular variant of the particle system and explore the limits of the exact formulas and their applications (in particular, concerning the asymptotic behavior of correlations along certain space-time directions). \subsection{Acknowledgements} The authors wish to thank Patrik Ferrari for conversations on this work and the related work \cite{BCF}. A. B. was partially supported by the NSF grant DMS-1056390. I. C. was partially supported by the NSF DMS-1208998, by a Clay Research Fellowship, by the Poincar\'{e} Chair, and by a Packard Fellowship for Science and Engineering. F. T. was partially funded by Marie Curie IEF Action “DMCP- Dimers, Markov chains and Critical Phenomena”, grant agreement n. 621894. This work was initiated during the Statistical Mechanics, Integrability and Combinatorics program at Galileo Galilei Institute (Arcetri). We appreciate the hospitality and support of these institutes. \section{Model and notation} We start by recalling the definition of the $q$-Whittaker particle system on the torus, and its stationary measure as defined in \cite{CT}. We consider an interacting particle system in which particles live on the $L\times N$ discrete torus $\mathbb T_{L,N}=\mathbb Z/(L \mathbb Z)\times \mathbb Z/(N \mathbb Z)$. The horizontal size is $L$ and the vertical size is $N$. The particle configuration space will be denoted $\Omega_{L,N;m_1,m_2}$, and it depends on two integers $1< m_1< L$ and $1\le m_2< N$ such that \begin{gather} \label{eq:26} m_1/L+m_2/N<1. \end{gather} At each site $x=(x_1,x_2)\in \mathbb T_{L,N}$ there is at most one particle. On each row there are exactly $m_1$ particles. We exclude $m_1=1$ and $m_1=L$ to avoid trivialities. The parameter $m_2$ has a more topological nature and its meaning will be explained below. The horizontal position of particle $p$ is denoted $x_p\in \mathbb Z/(L\mathbb Z)$. Particle positions are interlaced, in the following sense. Given particle $p$ (say on row $i$), we let $p_1,p_4$ denote its right/left neighbor on the same row (note that if $m_1=2$ then $p_1=p_4$). Then, we require that in row $i-1$ there is exactly one particle, labeled $p_2$, whose position satisfies \begin{eqnarray}\label{eq:1prime} x_{p_2}\in \{x_{p}+1,x_p+2,\dots,x_{p_1}\} \end{eqnarray} and exactly one particle, labeled $p_3$, satisfying \begin{gather} \label{eq:22} x_{p_3}\in \{x_{p_4}+1,x_{p_4}+2,\dots,x_{p}\}. \end{gather} See Figure \ref{fig:particelle}. Note that, automatically, in row $i+1$ there are exactly one particle $p_5$ and one particle $p_6$ satisfying respectively \begin{gather} \label{eq:23} x_{p_5}\in \{x_{p_4},\dots,x_p-1\},\quad x_{p_6}\in \{x_{p},\dots,x_{p_1}-1\}. \end{gather} \begin{figure}[ht] \includegraphics[width=.6\textwidth]{particelle.pdf} \caption{ The neighbors $p_1,\dots,p_6$ of particle labeled $p$. Note that conditions \eqref{eq:1prime}, \eqref{eq:22} allow $C_p:=x_{p}-x_{p_3}=0$ but they impose $B_p+1:=x_{p_2}- x_p\ge1$. } \label{fig:particelle} \end{figure} We define non-negative integers $A_p,\dots,F_p$ as \begin{eqnarray} \label{eq:19} &A_p=x_{p_1}-x_{p}-1;\quad B_p=x_{p_2}-x_p-1;\quad C_p=x_p-x_{p_3}\\ \nonumber &D_p=x_{p}-x_{p_4}-1;\quad E_p=x_p-x_{p_5}-1;\quad F_p=x_{p_6}-x_p. \end{eqnarray} The particles $p_1,\dots,p_6$ are the six neighbors of $p$, labeled clockwise starting from the one on the right. The definition of the dynamics will be such that the labels of the neighbors of a particle $p$ do not change with time (particles will not jump over each other or change interlacements). Let $\Omega_{L,N;m_1}$ be the set of particle occupation functions, i.e. of functions $\eta:\mathbb T_{L,N}\mapsto \{0,1\}$, with $m_1$ particles (i.e. occupation variables equal to 1) per row, whose positions satisfy the constraints \eqref{eq:1prime}-\eqref{eq:23}. The set $\Omega_{L,N;m_1}$ decomposes into disjoint ``sectors'': \begin{gather} \label{eq:24prime} \Omega_{L,N;m_1}=\bigcup_{m_2}\Omega_{L,N;m_1,m_2} \end{gather} as follows. Given any particle $p$, connect $p$ to its up-right neighbor $p_6$, then $p_6$ with its own up-right neighbor and repeat the operation until the path $\Gamma$ thus obtained gets back to the starting particle $p$. Note that $\Gamma$ forms a simple loop: otherwise, there would be a particle $r$ which is reached along $\Gamma$ from two different particles $r',r''$. This is impossible, since both $r'$ and $r''$ would be the lower-left neighbor $r_3$ of $r$. Call $N_v\in\mathbb N\cup\{0\},N_h\in\mathbb N\cup\{0\}$ the vertical and horizontal winding numbers of $\Gamma$ around the torus $\mathbb T_{L,N}$. It is easy to see that $N_h,N_v$ are independent of the chosen initial particle $p$. As discussed in \cite[Remark 2]{CT}, \begin{gather} \label{eq:27} m_2:=m_1\frac{N_h}{N_v} \end{gather} is an integer and it satisfies \eqref{eq:26}. The set $\Omega_{L,N;m_1,k}$ is defined as the subset of $\Omega_{L,N;m_1}$ with $ m_2=k$. Each sector $\Omega_{L,N;m_1,m_2}$ will remain invariant under our dynamics. Let us briefly remark that the particle configurations we are considering can also be mapped onto dimers on the periodized $L\times N$ hexagonal lattice. This perspective is explained in \cite{CT} (see in particular Fig. 2 there) wherein $n_1:=m_1 N$ corresponds to the number of vertical dimers, $n_2:=m_2 L$ to the number of north-west dimers and $n_3:=NL-n_1-n_2$ the number of north-east dimers. We will not pursue this perspective any further here. Given a configuration $\eta\in \Omega_{L,N;m_1,m_2}$, draw a directed upward edge from any particle label $r$ to its up-right neighbor $r_6$ if $F_r=0$ (in which case $r$ and $r_6$ have the same horizontal position). For the particle labeled by $p$ let $V^+_p$ be the set that includes $p$ plus the particle labels that can be reached from $p$ by following upward oriented edges. The dynamics we consider is a continuous-time Markov chain on $\Omega_{L,N;m_1,m_2}$. For each $p$, there is an exponential clock of rate \begin{eqnarray} \label{eq:6prime} \frac{(1-q^{B_p})(1-q^{D_p+1})}{(1-q^{C_p+1})} \end{eqnarray} with $q\in[0,1)$. When said clock rings, all particles with label $r\in V^{+}_p$ shift by $(1,0)$. Note that the rate is zero if $B_p=0$. This prevents particles from overlapping after the move. Note also that after the move, the configuration is still in $\Omega_{L,N;m_1,m_2}$. This is discussed in more detail in \cite{CT}. Another way to understand the dynamics is that when particle $p$ moves, provided its up-right neighbor $p_6\in V^+_p$, then $C_{p_6}$ becomes $-1$ and the jump rate for $p_6$ becomes infinite, and hence it immediately moves (and so on for all other $r\in V^+_p$). These dynamics are called the $q$-Whittaker particle system on the torus. As a side remark, let us add that, in terms of dimer model, shift to the right by $+1$ of a family $V^+_p$ corresponds to increasing the height by $1$ in $|V^+_p|$ faces of the hexagonal lattice. Given $q\in [0,1)$, let $\pi$ be the probability measure on $\Omega_{L,N;m_1,m_2}$ defined as \begin{eqnarray} \label{eq:3prime} \pi(\sigma):=\frac1{Z_{L,N;m_1,m_2}}\prod_{p} \frac{(q;q)_{A_p}}{(q;q)_{B_p}(q;q)_{C_p}}{\bf 1}_{\{\sigma\in \Omega_{L,N;m_1,m_2}\}} \end{eqnarray} where $(q;q)_n=(1-q)(1-q^2)\dots (1-q^n)$ and $Z_{L,N;m_1,m_2}$ is the normalizing constant necessary to make this a probability measure. The main result of \cite{CT} (Theorem 1 therein) is: \begin{Proposition} \label{Prop1} The probability law $\pi$ is stationary in time for $q$-Whittaker particle system on the torus. \end{Proposition} Clearly, the measure $\pi$ is not reversible, since the process is totally asymmetric. \bigskip We will consider a certain $q\to 1$ limit of the particle system. To fix scalings, for $\epsilon>0$ let $q=\exp(-\epsilon)$. For simplicity (to avoid a plethora of $\lfloor \cdot \rfloor$), we will assume $\epsilon^{-1}$ is an integer, though all results hold for arbitrary $\epsilon$. Further, let $L=\epsilon^{-1}\ell$ for some integer $\ell$. On each row there are $m:=m_1=n_1/N$ particles, with $m$ an integer of order $\ell$. Also for simplicity we take $N=m$, so altogether we have $n_1=m^2=O(\ell^2)$ particles. Taking $N$ to be another multiple of $m$ would not change our results qualitatively. \begin{Remark} \label{rem:bcc1} On $\mathbb Z^2$, introduce the equivalence relation $\sim$ such that $p\sim p'$ iff $p=p'+(j_1m-j_2 m_2,j_2 m)$ for some $j_1,j_2\in\mathbb Z$. Observe that if we take $N=m$ steps along $\Gamma$ starting from $p$, we get a particle $p'$ that is the $j^{th}$ right neighbor of $p$ on the same row, for some $0\le j\le m$. Actually, one has $j=m_2$, since $N_vj=N_h m$. Particles will be given a label $p\in \mathbb Z^2$ such that $p_1=p+(1,0), p_2=p+(1,-1), p_3=p+(0,-1),p_4=p-(1,0),p_5=p+(-1,1),p_6=p+(0,1)$, with an arbitrary choice of which particle is labeled $p=(0,0)$. Thanks to the above observation, given integers $j_1,j_2$, particle $p+(j_1 m_1,j_2 N)$ is to be identified with $p+(j_2 m_2,0)$. In other words, particles are labeled by $p\in \mathcal R_m:=\mathbb Z^2/\sim$ (the quotient set of $\mathbb Z^2$ by $\sim$, which contains $m^2$ equivalence classes). \end{Remark} We will first take the limit $\epsilon\to0$ with $m$ and $m_2$ fixed, and then $m\to\infty$ with the ratio $m_2/m$ bounded away from 0 and 1 if we want to take an infinite-volume limit ($m_2/m\le1 $ by definition, recall \eqref{eq:26} and $m=N$). The average inter-particle distance along a row is \begin{eqnarray} \label{eq:D} \frac{D}{\epsilon}:=\frac Lm=\frac{\ell}{\epsilon m} \end{eqnarray} so that, with reference to \eqref{eq:19}, the average value of $D_p$ is $\epsilon^{-1}D$. The average value of $C_p$ is instead fixed by choosing $m_2$ suitably, as follows. Let $C\in (0,D)$ be defined by \begin{eqnarray} \label{eq:C} \frac CD =\frac{m_2}m \end{eqnarray} and let $B=D-C$. (Note that, as long as the integers $m,m_2$ are kept finite, $C$ can take only a finite set of values in $(0,D)$; such set becomes dense in $(0,D)$ if we allow $m,m_2$ to diverge with a fixed ratio). We claim that the average value of $C_p$ in any particle configuration is $\epsilon^{-1}C$. Indeed, take the path $\Gamma$ defined after \eqref{eq:24prime}. It visits exactly $N_v\times N=m N_v$ particles and its total displacement to the right equals $N_h \times L$. On the other hand, $N_h \times L$ is also the sum of the $C_p$ over all particles in $\Gamma$. From this, we see that the average value of $C_p$ is \begin{eqnarray} \label{eq:bo} \frac{N_h L}{N_v m}=\frac{m_2 L}{m^2}= \epsilon^{-1}\frac {m_2}{m }\frac{\epsilon L}{m}=\epsilon^{-1}\frac {m_2}{m}D \end{eqnarray} where in the first equality we used \eqref{eq:27}, and in the third \eqref{eq:D}. Eq. \eqref{eq:C} then allows to conclude. Similarly, we see that the average of $B_p$ is $\epsilon^{-1}B$. On the other hand, the averages of $D_p,E_p,F_p$ equal by definition the averages of $A_p, B_p,C_p$ respectively. \section{Convergence to a system of SDEs}\label{sec:conv} We will start the dynamics from an initial condition where each particle $p\in \mathcal R_m$ is within distance $O(\epsilon^{-1/2})$ from its ``ideal position'' $X_p$ in a perfect ``crystalline configuration'' where $D_p=\epsilon^{-1}D, B_p=\epsilon^{-1}B,C_p=\epsilon^{-1}C$ for every $p$. Assuming without loss of generality that $X_0=0$, we have for $p=(p^{(1)},p^{(2)})$ \begin{eqnarray} \label{eq:Xp} X_p=p^{(1)} \epsilon^{-1}D+p^{(2)}\epsilon^{-1}C \end{eqnarray} where the r.h.s. has to be taken modulo $\epsilon^{-1}\ell$. Our first result (proven in Section \ref{sec:t1}) says that particles move macroscopically with a deterministic speed $v(C,D)>0$ and that, in the time-scale of order $\epsilon^{-1}$, fluctuations around such hydrodynamic limit are of order $\epsilon^{-1/2}$ and converge to a system of linear SDEs. \begin{Theorem} \label{th:SDE} Fix $N=m$ and $\ell$, so that the lattice $\Lambda_{L,N}$ depends only on $\epsilon$. Let $\{\bar\xi_p\}_{p\in \mathcal R_m}\in \mathbb R^{\mathcal R_m}$. Consider an initial configuration $\sigma_0$ such that, defining \begin{eqnarray} \label{eq:etaas} \eta_p:=\eta_{p,\epsilon}=\sqrt{\epsilon}(x_p-X_p), \end{eqnarray} one has \begin{eqnarray} \label{eq:xi} \lim_{\epsilon\to0}\eta_p=\bar\xi_p \quad \forall p\in \mathcal R_m. \end{eqnarray} Define \begin{gather} \label{eq:1} v(C,D)=\frac{(1-\exp(-B))(1-\exp(-D))}{1-\exp(-C)} \end{gather} (recall that $B=D-C$) and \begin{eqnarray} \label{eq:xit} \eta_{p,t}=\sqrt\epsilon\left({x_p(t/\epsilon)-X_p-\epsilon^{-1}vt}\right), \quad t\ge0. \end{eqnarray} Then, the random process $\{\eta_{p,\cdot}\}_{p\in\mathcal R_m}$ converges weakly as $\epsilon\to0$ to the solution of the system of linear stochastic differential equations \begin{gather} \label{eq:2} \left\{ \begin{array}{ll} d\xi_{p,t}=\sqrt v\, {\rm d}W_{p,t}+\sum_{p'}A_{p,p'}\xi_{p',t} dt\\ \xi_{p,0}=\bar\xi_p \end{array} \right. \end{gather} with \begin{gather} \label{eq:A} A_{p,p'}=\delta_{p'=p}\left(\frac{e^{-D}(1-e^{-B})}{1-e^{-C}}-\frac{e^{-B}(1-e^{-D})}{1-e^{-C}}-\frac{e^{-C}(1-e^{-B})(1-e^{-D})}{(1-e^{-C})^2}\right)\\ +\delta_{p'=p+(1,-1)}\frac{e^{-B}(1-e^{-D})}{1-e^{-C}} -\delta_{p'=p-(1,0)}\frac{e^{-D}(1-e^{-B})}{1-e^{-C}}\\+ \delta_{p'=p-(0,1)}\frac{e^{-C}(1-e^{-B})(1-e^{-D})}{(1-e^{-C})^2} \end{gather} and where $dW_{p,t}$ are independent white noises indexed by $p$, and one-dimensional in time $t$. \end{Theorem} \begin{Remark} Note that the matrix $A$ is not symmetric, so that the diffusion \eqref{eq:2} is not reversible with respect to its stationary measure described by Theorem \ref{th:muinv}. In other words, the irreversibility of the microscopic dynamics survives also in the Gaussian limit. \end{Remark} \subsection{Properties of the matrix $A$} It is convenient to work in Fourier space. For this purpose, let \begin{gather} \label{eq:6} f_k:\mathcal R_m\mapsto \mathbb C, \quad f_k: p\mapsto \frac1m e^{-i p k}. \end{gather} The set $\{f_k\}_{ k\in \mathcal K_{m}}$, where \begin{gather} \mathcal K_{m}=\{((2\pi/m)r_1,(2\pi/m)\left(\frac CD r_1+r_2\right), r_1,r_2\in\mathbb Z, -m/2\le r_1,r_2<m/2\}, \end{gather} forms an orthonormal basis of $\mathbb C^{\mathcal R_m}$. \begin{Remark} Note that $\mathcal K_m$ was chosen such that, if we define $f_k(p)=(1/m)e^{-i p k}$ for every $p\in\mathbb Z^2$, then $f_k(p)=f_k(p')$ if $p\sim p'$ (use that $m_2/m=C/D$, see \eqref{eq:C}). Also, observe that $|\mathcal K_m|=m^2$. \end{Remark} Define \begin{eqnarray} \label{eq:hatxi} \hat \xi_{k,t}=\sum_{p\in \mathcal R_m} \xi_{p,t}f_k(p), \end{eqnarray} so that \begin{eqnarray} \label{eq:eta-eta} \xi_{p,t}=\sum_{k \in \mathcal K_{m}}\hat\xi_{k,t}\overline{ f_k(p)},\quad \text{ and}\quad \hat\xi_{-k,t}=\overline{\hat \xi_{k,t}}. \end{eqnarray} Let also \begin{eqnarray} \label{eq:Ak} \hat A(k)=\sum_p A_{p,0}e^{-i p k}, k\in\mathbb R^2. \end{eqnarray} In our specific case, one sees that \begin{eqnarray} \label{eq:9} \hat A(k)&=&A_{0,0}+A_{0,(1,-1)}e^{-i(k_1-k_2)}+A_{0,(-1,0)}e^{ik_1}+A_{0,(0,-1)}e^{i k_2}\\ \hat A(k)+\hat A(-k)&=&2[A_{0,0}+A_{0,(1,-1)}\cos(k_1-k_2)\\&+&A_{0,(-1,0)}\cos(k_1)+A_{0,(0,-1)}\cos(k_2)]. \end{eqnarray} Observe that we defined $\hat A(k)$ for any $k\in \mathbb R^2$ and not just for $k\in \mathcal K_{m}$. \begin{Proposition} \label{prop:prop} The matrix $A$ satisfies the following properties: \begin{enumerate} \item Translation invariance: $A_{p,p'}=A_{p+r,p'+r}$ for every $r$; \item $\hat A(0)=\sum_{p'} A_{p,p'}=0$; \item \begin{eqnarray} \label{eq:37} \hat R(k):=\hat A(k)+\hat A(-k)\le 0 \end{eqnarray} and the only zero of $\hat R(k)$ on $[-\pi,\pi]^2$ is at $k=0$; \item \begin{eqnarray} \label{eq:rkw} \hat R(k)= \hat W(k)+O(|k|^3)=(k, \hat W k)+O(|k|^3), \quad k\to0, \end{eqnarray} with $\hat W$ a strictly negative definite $2\times 2$ matrix and $\hat W(k)$ the associated quadratic form. \item The function $\hat A(\cdot)$ is $2\pi$-periodic and $C^\infty$ on $[-\pi,\pi]^2$. \end{enumerate} \end{Proposition} All properties are trivial to check, except for (3) which is proven in Appendix \ref{app:Rk}. The fact that \begin{gather} \hat W(k):=-A_{0,(1,-1)}(k_1-k_2)^2-A_{0,(-1,0)}k_1^2-A_{0,(0,-1)}k_2^2 \end{gather} is a strictly negative definite quadratic form follows from negativity of $\hat R(k)$, together with the fact that \begin{eqnarray} \label{eq:38} \det(\hat W)=\frac{e^{-D}(1-e^{-D})(1-e^{-B})^2}{(1-e^{-C})^2}=:w^2>0. \end{eqnarray} For later convenience, let the $2\times 2$ matrix $V$ be such that \begin{eqnarray} \label{eq:Vv} V \hat W V^T=-I. \end{eqnarray} Let us also define \begin{eqnarray} \label{eq:Uu} U:=i\nabla \hat A(0)=\sum_{p}p A_{0,p}= (A_{0,(1,-1)}-A_{0,(-1,0)},-A_{0,(1,-1)}-A_{0,(0,-1)})\in \mathbb R^2, \end{eqnarray} and note that $U\ne 0$ for any choice of $B,C,D=B+C$. \begin{Remark} A few comments are in order: \begin{itemize} \item Property (1) (translation invariance of $A$) originates from the fact that we are considering fluctuations around a ``trivial'', translation invariant hydrodynamic limit where particles have positions $X_p+\epsilon^{-1}v t$ and are therefore equi-spaced at all times. This is the property that makes it convenient to work in Fourier space (e.g., property (1) is behind the fact that Eq. \eqref{eq:dxik} is diagonal in $k$); \item Property (2) means that the drift of a particle $p$ is unchanged if all particles are globally shifted by the same amount. This is a consequence of the fact that transition rates \eqref{eq:6prime} of the microscopic particle process depend only on inter-particle distances, and would hold even if we studied fluctuations w.r.t. a non-translation invariant hydrodynamic limit (cf. Section \ref{sec:pde}); \item Property (3) guarantees that there exists a stationary measure for gradients $\xi_p-\xi_{p'}$, see Theorem \ref{th:muinv} and formula \eqref{eq:inv}. Negative-definiteness of $\hat A(k)+\hat A(-k)$ corresponds to the fact that in the hydrodynamic scaling the particle configuration is ``crystalline'' at all times, and that crystalline configurations are an (at least local) maximum of the stationary measure $\pi(\cdot)$, see Lemma \ref{th:locmax}. \end{itemize} \end{Remark} \subsection{A conjectural hydrodynamic equation} \label{sec:pde} We will briefly consider the hydrodynamic behavior of the $q$-Whittaker particle system with general initial data (which may not be close to the crystalline configurations considered above). We provide a heuristic derivation of the hydrodynamic (law of large numbers) PDE satisfied by the limit. For us, the purpose of this derivation is to justify and explain our use of the word "characteristic" to describe the direction $U$ of slow decorrelation. In this hydrodynamic limit we will first take $\epsilon\to0$ and then $m\to\infty$. With this in mind, let us rescale space and time by $m/\epsilon$ and define \[ h(\tau,y)=(\epsilon/m) x_{\lfloor y m\rfloor }(\tau m/\epsilon) \] with $y \in [0,1]^2$, $\tau\ge0$, and where the r.h.s. has to be taken modulo $(\ell/\epsilon)\times(\epsilon/m)=D$ because $x_p$ is defined modulo $L=\ell/\epsilon$. Note that, as a consequence of Remark \ref{rem:bcc1}, $h(\tau,\cdot)$ satisfies \[ h(\tau,y+(j_1,j_2))=h(\tau,y+(Cj_2/D,0)), \quad j_1,j_2\in \mathbb Z. \] If at time zero the configuration satisfies the conditions of Theorem \ref{th:SDE}, then $h(0,y)$ is close to a linear function: \[ H(0,y):=\lim_{m\to\infty}\lim_{\epsilon\to0}h(0,y)=D y_1+C y_2. \] In this case, it follows from Theorem \ref{th:SDE} that, for $\tau>0$, the limit \begin{eqnarray} \label{eq:hlim} H(\tau,y):=\lim_{m\to\infty}\lim_{\epsilon\to0}h(\tau,y) \end{eqnarray} exists and solves \begin{eqnarray} \label{eq:55} \partial_\tau H=v \end{eqnarray} with $v(C,D)$ as in \eqref{eq:1}. Given that $D=\partial_{y_1} H$ and $C=\partial_{y_2} H$, and using $B=D-C$, we see that \begin{eqnarray} \label{eq:56} v(C,D)=v(\nabla H)=\frac{(1-\exp(\partial_{y_2}H-\partial_{y_1} H))(1-\exp(-\partial_{y_1}H))}{(1-\exp(-\partial_{y_2}H))}. \end{eqnarray} If we assume instead that the initial condition satisfies \[ \lim_{m\to\infty}\lim_{\epsilon\to0} h(0,y)=H(0,y), \] with $H(0,\cdot)$ some smooth enough but non-linear function, then it is natural to conjecture that the limit \eqref{eq:hlim} exists and still satisfies \eqref{eq:55}, with $v=v(\nabla H)$ equal to the r.h.s. of \eqref{eq:56}. It is easy to see that the characteristic lines of the PDE \eqref{eq:55} are the straight lines $y=U t$ with $U$ as in \eqref{eq:Uu}. In fact, the characteristic lines are determined by \[ y_i(t)=t \nabla_i v,\quad i=1,2 \] with $\nabla_i v$ the derivative of $v=v(\nabla h)$ with respect to its $i$-th argument. Explicitly, \[\nabla_1 v=\frac{e^{\partial_{y_2}H-\partial_{y_1}H}(1-e^{-\partial_{y_1}H})}{1-e^{-\partial_{y_2}H}}+\frac{e^{-\partial_{y_1}H}(1-e^{\partial_{y_2}H-\partial_{y_1}H})}{1-e^{-\partial_{y_2}H}}=A_{0,(1,-1)}-A_{0,(-1,0)}=U_1 \] and similarly $\nabla_2 v=U_2$. We emphasize that the identity $\partial_i v=U_i$ is not a coincidence. Indeed, view the hydrodynamic speed $v$ as a function of the relative horizontal distances between particle, say, $0$ and the other particles $p$ (in our case, the dependence is only through the three neighbors $p_2,p_3,p_4$ of $0$). On one hand, since the SDEs \eqref{eq:2} describe a linearization of the stochastic dynamics around the hydrodynamic limit, $A_{0,p}$ is obtained taking the derivative of $v$ w.r.t. the position $x_p$ of particle $p$, with the others kept fixed. On the other hand, if the slope $\nabla_i H$ is changed by $\epsilon$, the distance between particles $0$ and $p=(p^{(1)},p^{(2)})$ changes by $\epsilon\times p^{(i)}$. Therefore, \begin{eqnarray} \label{eq:16} \nabla_{i} v=\sum_{p=(p^{(1)},p^{(2)})}p^{(i)}A_{0,p}=U_i. \end{eqnarray} This shows that the identity $U_i =\nabla_i v$ is not related to the particular form of the function $v(\cdot)$. If we had another interacting particle process for which we could prove convergence to a hydrodynamic equation and convergence of fluctuations to a system of linear SDEs with a matrix $A$ satisfying the conditions in Proposition \ref{prop:prop}, we would have automatically that the direction of slow decay of correlations would coincide with the characteristic lines of the PDE. \section{Correlations of the Gaussian system} In this section we study the space-time correlations of the SDE system \eqref{eq:2}. Since the equations are linear, they can be solved explicitly. We formulate the results in wider generality. Again, the solution to \eqref{eq:2} with initial condition $\bar \xi$ is denoted $\{\xi_{p,t}\}_{t\ge0, p\in \mathcal R_m}$ and we let $\mathbb P_{\bar \xi}$ be its law. The matrix $A=\{A_{p,p'}\}_{p,p'\in \mathcal R_m}$ is not necessarily given by \eqref{eq:A} but is required to satisfy properties (1)--(5) of Proposition \ref{prop:prop}. Theorems \ref{th:2} and \ref{th:muinv} hold in this generality. Let $\mathcal C^{\bar \xi}(t,s)$ be the $m^2\times m^2$ symmetric matrix \begin{gather} \label{eq:3} \mathcal C_{p,p'}^{\bar \xi}(t,s)=\mathbb E_{\bar \xi} [\xi_{p,t}\xi_{p',s}], \quad p,p'\in \mathcal R_m \end{gather} and $\mathcal M^{\bar\xi}(t)\in \mathbb R^m$ be the vector \begin{gather} \mathcal M_p^{\bar \xi}(t)=\mathbb E_{\bar \xi} [\xi_{p,t}], \quad p\in \mathcal R_m \end{gather} with of course $\mathcal C^{\bar \xi}_{p,p'}(0,0)=\bar\xi_p\bar\xi_{p'}$, $\mathcal M^{\bar \xi}_{p}(0)=\bar\xi_p$. By It${\rm \bar o}$'s lemma, \begin{gather} \label{eq:4} \frac d{dt}\mathcal M_p^{\bar \xi}=\sum_{p'} A_{p,p'}\mathcal M^{\bar \xi}_{p'}(t),\\ \nonumber \frac d{dt}\mathcal C^{\bar \xi}(t,t)=v\mathbb I +\mathcal C^{\bar \xi}(t,t)A^T+A \mathcal C^{\bar \xi}(t,t),\\ \nonumber \frac d{dt}\mathcal C(t,s)=A \mathcal C(t,s),\quad t>s, \end{gather} with $A^T$ the transpose of $A$ and $\mathbb I$ the identity matrix. Let also \begin{gather} \label{eq:21} \mathcal W^{\bar \xi}_{p,p'}(t,s):=\mathcal C_{p,p'}^{\bar \xi}(t,s)-\mathcal M^{\bar \xi}_p(t) \mathcal M^{\bar \xi}_{p'}(s). \end{gather} \begin{Theorem} \label{th:2} Let $A$ satisfy the properties of Proposition \ref{prop:prop}. Then, $\mathcal W^{\bar \xi}_{p,p+y}$ does not depend on $\bar \xi$ or $p$ and the limit \begin{eqnarray} \label{eq:WW} \mathbb W_{y}(t,s)= \lim_{m\to\infty}\mathcal W^{\bar\xi}_{p,p+y}(t,s) \end{eqnarray} exists for any given $y\in\mathbb Z^2$. Moreover, \begin{eqnarray} \label{eq:44} \mathbb W_{y}(t,s)= \frac v{4\pi w}\int_{1+(t-s)/2}^{1+(t+s)/2}\frac{e^{-|H|^2/(4a)}}a da + j(t,s,y), \end{eqnarray} where $w>0$ is defined in \eqref{eq:38}, $H=V y+(t-s)V U$, $V$ is defined in \eqref{eq:Vv}, $U=i\nabla \hat A(0)$, and $j$ satisfies \begin{eqnarray} \label{eq:condizionij} \sup_{t,s,y}|j(t,s,y)|<\infty, \quad \lim_{\max(t-s,|y|)\to\infty}j(t,s,y)=0. \end{eqnarray} \end{Theorem} If the matrix $A$ is the one given in \eqref{eq:A}, using \eqref{eq:38} and \eqref{eq:1}, one sees that \begin{eqnarray} \label{eq:sees} \frac v{ w} = {\sqrt{e^{D}-1}}. \end{eqnarray} From \eqref{eq:44} one can obtain all desired asymptotics. For instance, one can obtain equal-time correlations. \begin{Corollary}[Equal-time correlations] \label{cor1} For $y=0$ one has \begin{eqnarray} \label{eq:44bis} \limsup_{t\to\infty}\left|\mathbb W_{0}(t,t)- \frac v{4\pi w}\log t\right|<\infty. \end{eqnarray} For $y\ne0$ one finds, with $Y=Vy$ and after the change of variables $a=|Y|^2/(4x)$, \begin{gather} \mathbb W_{y}(t,t)= \frac v{4\pi w}\int_{|Y|^2/4(1+t)}^{|Y|^2/4}\frac{e^{-x}}x\;dx +j(t,t,y) . \end{gather} This implies that \begin{eqnarray} \label{eq:44ter} \lim_{\substack {|y|\to\infty, t\to\infty\\ |y|=O(\sqrt t)}} \left(\mathbb W_{y}(t,t)-\frac v{4\pi w}\log(4(t+1)/|Y|^2)\right)=0 \end{eqnarray} where we used the fact that \begin{eqnarray} \label{eq:c-c+} c_-\le \frac{|Y|}{|y|}\le c_+ \end{eqnarray} for some non-zero constants $c_\pm$ (the lower bound holds because the determinant of $V$ is not zero). If instead both $|y|,t$ diverge and $|y|\gg \sqrt t$, then $ \mathbb W_{y}(t,t)=o(1)$. The same results hold if $ t-s=O(1)$. \end{Corollary} As for correlations at different times, the following result shows that the behavior is special along the space-time lines $y=t U$ (that will be called ``characteristics'', in view of the discussion in Section \ref{sec:pde}), with $U=i\nabla \hat A(0)=\sum_p p A_{p,0} \in \mathbb R^2$. We will assume that $t-s\gg1$, since the case $t-s=O(1)$ is effectively covered by the previous Corollary. \begin{Corollary}[Correlations along the characteristics] \label{cor2} We deduce from \eqref{eq:44} \begin{gather} \label{eq:along} \limsup_{t-s\to\infty}\left(\mathbb W_{\lfloor U(t-s)\rfloor}(t,s) -\frac v{4\pi w}\log\frac{t+s}{t-s}\right)=0, \end{gather} in particular $\mathbb W_{\lfloor U(t-s)\rfloor}(t,s)$ is large if $t-s\ll t$. If instead $u\ne U$, we have \begin{gather} \limsup_{t-s\to\infty}\left(\mathbb W_{\lfloor u(t-s)\rfloor}(t,s) -\frac v{4\pi w}\int_{\frac{(t-s)^2|V(U-u)|^2}{2(t+s)}}^\infty \frac{e^{-x}}xdx\right)=0; \end{gather} from this we deduce that \begin{eqnarray} \label{eq:nantra} \limsup_{t\to\infty,(t-s)/\sqrt t\to\infty}\mathbb W_{\lfloor u(t-s)\rfloor}(t,s)=0, \end{eqnarray} while \begin{gather} \limsup_{t-s\to\infty, t-s=O(\sqrt t)}\left|\mathbb W_{\lfloor u(t-s)\rfloor}(t,s)- \frac v{4\pi w} [\log t-2\log(t-s)]\right| <\infty. \end{gather} \end{Corollary} Summarizing: along the characteristic correlations are large as soon as $(t-s)\ll t$ and grow proportionally to $\log t$ if $t-s\le t^a,a<1$. For all other space-time directions, correlations are large only if $t-s\ll \sqrt t$. Equation \eqref{eq:along} shows that, along the characteristic, there is a limit for the correlation as $t,s\to\infty$ with $t-s$ and $t+s$ of comparable size. This suggests that in this scaling the whole fluctuation field near the characteristic has a non-trivial limit process which, in the following statement, we identify as being related to the Stochastic Heat Equation. \begin{Corollary} \label{cor3} Define \begin{eqnarray} \label{eq:43} h_{x,t}=a\xi_{\lfloor t U+V^{-1}x\rfloor ,t},\quad a=\sqrt{\frac{4\pi w}{8v}}. \end{eqnarray} Then, for any given $x,y,$ $0<s<t$, from \eqref{eq:44} we get \begin{gather} \label{eq:45} \lim_{\delta\to0}\left[ \mathbb E_{\bar\xi}(h_{\delta^{-1/2}x,\delta^{-1}t}\;h_{\delta^{-1/2}y,\delta^{-1}s})- \mathbb E_{\bar\xi}(h_{\delta^{-1/2}x,\delta^{-1}t})\mathbb E_{\bar \xi}(h_{\delta^{-1/2}y,\delta^{-1}s})\right]\\= \frac 18\int_{(t-s)/2}^{(t+s)/2}\frac{e^{-|x-y|^2/(4a)}}a da . \end{gather} In other words, in this limit the randomly evolving height field $h_{\cdot,\cdot}$ has the same space-time correlations as the additive stochastic heat equation in $2$ spatial dimensions, \begin{eqnarray} \label{eq:53} \partial_t h=\Delta h+ \dot W \end{eqnarray} with $\dot W$ the $(2+1)$-dimensional space-time white noise (compare \eqref{eq:45} with the formal expression derived in \cite[Formula (2.8)]{Hairer} for the covariance of the stochastic heat equation). \end{Corollary} \begin{Remark} If the particle label $p$ is $d$-dimensional, $ d\ne 2$, and the matrix $A$ satisfies the analog the properties (1)--(5) of Proposition \ref{prop:prop} (with $[-\pi,\pi]^2$ replaced by $[-\pi,\pi]^d$), then one can check that Theorem \ref{th:2} still holds, with \eqref{eq:44} replaced by \begin{eqnarray} \label{eq:dnon2} \mathbb W_{y}(t,s)= \frac {\pi v}{(2\pi)^d w}\int_{1+(t-s)/2}^{1+(t+s)/2}\frac{e^{-|H|^2/(4a)}}{a^{d/2}} da + j(t,s,y), \quad y\in \mathbb Z^d. \end{eqnarray} We leave it to an interested reader to derive the analogs of Corollaries \ref{cor1}--\ref{cor3} for $d\ne 2$. \end{Remark} \subsection{Stationary measure} We cannot expect that there is a stationary measure for $\{\xi_{p}\}_p$: since there is invariance by global shifts on $\mathbb R$ of all positions $\xi_p$, the inverse covariance matrix will have a zero mode (like a Gaussian Free Field not pinned to zero anywhere). However, the stationary measure on gradients of $\xi$ will be well-defined. Not surprisingly, its scaling limit (which requires letting $m\to\infty$) is the standard massless Gaussian Free Field (see e.g. \cite{GFF} for definitions), up to an affine transformation of coordinates by the matrix $V$ (cf. \eqref{eq:varianza}). \begin{Theorem} \label{th:muinv} Let $A$ satisfy the properties of Proposition \ref{prop:prop}. There exists a unique stationary measure $\mu$ for the gradients $(\xi_p-\xi_{p'})_{p,p'\in\mathcal R_m}$. This measure is Gaussian and its mean and covariances are given by \begin{eqnarray} \label{eq:nomedia} \mathbb E_\mu(\xi_p-\xi_{p'})=0 \quad \forall p,p' \end{eqnarray} and \begin{eqnarray} \label{eq:covar} {\rm Cov}_\mu\left[ (\xi_{y_1}-\xi_{y_2});(\xi_{y_3}-\xi_{y_4})\right]&=&-\frac v{m^2}\sum_{k\in \mathcal K_m}\frac{(e^{i k y_1}-e^{i k y_2})(e^{-i k y_3}-e^{-i k y_4})}{\hat R(k)} \\& \stackrel{m\to\infty}\to& \label{eq:moreg} -\frac{v}{(2\pi)^2}\int_{[-\pi,\pi]^2}d k \frac{(e^{i k y_1}-e^{i k y_2})(e^{-i k y_3}-e^{-i k y_4})}{\hat R(k)}\\&\equiv&{\rm Cov}_{\mu_\infty}\left[ (\xi_{y_1}-\xi_{y_2});(\xi_{y_3}-\xi_{y_4})\right]. \end{eqnarray} Moreover, one has \begin{gather} \label{4punti} {\rm Cov}_{\mu_\infty}[(\xi_{y_1}-\xi_{y_2});(\xi_{y_3}-\xi_{y_4})]= \frac v{2\pi w}\log\frac{1+|V(y_1-y_4)||V(y_3-y_2)|}{1+|V(y_1-y_3)||V(y_2-y_4)|}+R_{y_1,y_2,y_3,y_4} \end{gather} where $V$ is the matrix in \eqref{eq:Vv} and\footnote{here, $|y_i-y_j|$ denotes the Euclidean distance between $y_i$ and $y_j$ on $\mathbb Z^2$ and not on the ``torus'' $\mathcal R_m$ (recall that the limit $m\to\infty$ has already been taken).} \begin{eqnarray} \label{eq:R4} R_{y_1,y_2,y_3,y_4}=O(1/(1+ \min(|y_1-y_3|,|y_2-y_4|,|y_1-y_4|,|y_2-y_3|))). \end{eqnarray} From this we deduce a convergence to a massless GFF on $\mathbb R^2$, of covariance proportional to \[ -\log |V(x-y)|, \] in the following sense: Let $\phi:\mathbb R^2\mapsto \mathbb R$ be a $C^\infty$, compactly supported function such that $\int_{\mathbb R^2}\phi(x)dx=0$. Then, the zero-average random function \begin{eqnarray} \label{eq:smooth} \xi_\phi:= \delta^2\sum_{p}\phi(\delta p)(\xi_p-\xi_0) \end{eqnarray} converges in law, in the limit $\lim_{\delta\to0}\lim_{m\to\infty}$, to a centered Gaussian random variable of variance \begin{eqnarray} \label{eq:varianza} - \frac v{2\pi w} \int \phi(x)\phi(y)\log |V(x-y)|dx dy. \end{eqnarray} \end{Theorem} Remark that, given $\phi^{(i)},i=1,2$ satisfying the same assumptions as $\phi$ above, the limit covariance $\lim_{\delta\to0}\lim_{m\to\infty}\mathbb E_\mu(\xi_{\phi^{(1)}}\xi_{\phi^{(2)}})$ can be simply deduced via \begin{eqnarray} 2 \mathbb E_\mu(\xi_{\phi^{(1)}}\xi_{\phi^{(2)}})=\mathbb E_\mu(\xi^2_{\phi^{(1)}})+\mathbb E_\mu(\xi_{\phi^{(2)}}^2) -\mathbb E_\mu((\xi_{\phi^{(1)}-\phi^{(2)}})^2). \end{eqnarray} In particular, \eqref{4punti} gives \begin{gather} \label{eq:39} \limsup_{|y|\to\infty} \left|{\rm Var}_{\mu_\infty}[\xi_0-\xi_y]- \frac v{\pi w}\log |V y|\right|<\infty \end{gather} (using \eqref{eq:c-c+}, one can replace $Vy$ with $y$). \section{Convergence to the SDEs: Proof of Theorem \ref{th:SDE}} \label{sec:t1} \subsection{A tightness estimate} Let \begin{eqnarray} \label{eq:G} \mathcal G=\left\{\sigma\in\Omega_{L,N;m_1,m_2}:\forall p, \left|B_p-\frac B\epsilon\right|\le \frac{\log (1/\epsilon)}{\sqrt\epsilon}, \left|D_p-\frac D\epsilon\right|\le \frac{\log (1/\epsilon)}{\sqrt\epsilon} \right\}. \end{eqnarray} Note that the initial condition $\sigma_0$ is well inside $\mathcal G$, cf. \eqref{eq:xi}. The crucial ingredient in the proof of Theorem \ref{th:SDE} is the following a-priori tightness estimate: \begin{Lemma} \label{lemma:apriori} Let $\partial \mathcal G\subset \mathcal G$ denote the inner boundary of $\mathcal G$ (the set of configurations from which the dynamics can exit $\mathcal G$ with a single update) and define the stopping time \begin{gather} \tau_{\mathcal G}=\inf\{t>0: \sigma(t)\in\partial \mathcal G\}. \end{gather} Then, \begin{gather} \label{eq:taulargo} \lim_{\epsilon\to0} \mathbb P_{\sigma_0}(\tau_{\mathcal G}\le \epsilon^{-2})=0. \end{gather} \end{Lemma} The time $\epsilon^{-2}$ could be replaced by any $\epsilon^{-a}$. What we need is that $a>1$, so that this time is much larger than $1/\epsilon$, the time-scale of convergence to the SDEs. \begin{proof}[Proof of Lemma \ref{lemma:apriori}] Let us start with the following: \begin{Lemma} \label{lemma:qq} Let $a=\epsilon^{-1}b+X$ with $b>0$ and $\sqrt \epsilon |X|\ll \epsilon^{-1/10}$. The following asymptotic expansion holds as $\epsilon\to0$: \begin{gather} (q;q)_a=\exp\left[ \frac1\epsilon\sum_{n\ge1}\frac1{n^2}e^{-bn}-\frac12\sum_{n\ge1}\frac1ne^{-bn}\right. \\ \left.- X\sum_{n\ge1}\frac1n e^{-bn}+\epsilon\frac{X^2}2\frac{e^{-b}}{1-e^{-b}}+C(\epsilon)+o(1) \right] \end{gather} with $C(\epsilon)$ independent of $b $ and $X$. \end{Lemma} See Appendix \ref{app:qq} for the proof. \begin{Lemma} \label{th:locmax} If $\sigma\in{\mathcal G}$ and if as in \eqref{eq:xi} we let $\eta_p:=\sqrt\epsilon(x_p-\epsilon^{-1}P_p)$, then \begin{gather} \label{eq:11} \pi(\sigma)=\frac1Z \exp\left(v\sum_{k\in\mathcal K_m} \hat R(k)|\hat {\eta_k}|^2+o(1)\right) \end{gather} where $\pi(\cdot)$ is defined in \eqref{eq:3prime} and, as in \eqref{eq:hatxi}, we set \begin{eqnarray} \label{eq:hateta} \hat {\eta_k}=\sum_{p\in\mathcal R_m}\eta_p f_k(p),\quad \eta_p=\sum_{k\in\mathcal K_m}\hat \eta_k\overline{f_k(p)}. \end{eqnarray} We recall from Proposition \ref{prop:prop} that $\hat R(k)\le 0 $ and vanishes only for $k=0$. \end{Lemma} \begin{proof}[Proof of Lemma \ref{th:locmax}] From Lemma \ref{lemma:qq} we see that \begin{eqnarray} \label{eq:46} \pi(\sigma)=\frac1Z \exp\left(-\frac12 (\eta,Q \eta)+o(1)\right) \end{eqnarray} with $Z$ a normalization constant independent of $\sigma$ and \begin{gather} - (\eta, Q \eta)=\sum_p (\eta_p-\eta_{p+(1,0)})^2\frac{e^{-D}}{1-e^{-D}}\\- \sum_p (\eta_p-\eta_{p+(1,-1)})^2\frac{e^{-B}}{1-e^{-B}}- \sum_p (\eta_p-\eta_{p+(0,-1)})^2\frac{e^{-C}}{1-e^{-C}}. \end{gather} This can be rewritten in Fourier space as \begin{gather} \label{eq:Fourier} -\frac12(\eta, Q \eta)=\sum_k |\hat \eta_k|^2\hat Q(k)\\ \hat Q(k)= \frac{e^{-D}}{1-e^{-D}}(1-\cos(k_1))-\frac{e^{-B}}{1-e^{-B}}(1-\cos(k_1-k_2)) \\-\frac{e^{-C}}{1-e^{-C}}(1-\cos(k_2)). \end{gather} Here, we use for instance that \begin{gather} \sum_p(\eta_p-\eta_{p+(1,0)})^2=\sum_p\sum_{k,k'\in\mathcal K_m}\hat \eta_k\hat \eta_{-k'}\overline{f_k(p)}f_{k'}(p)(1-e^{i k_1})(1-e^{-i k'_1})\\ =\sum_{k\in\mathcal K_m}|\hat \eta_k|^2(2-2\cos(k_1)) \end{gather} where we used orthonormality of $\{f_k(\cdot)\}_{k\in\mathcal K_m}$ and the second of \eqref{eq:eta-eta}. One checks that \begin{eqnarray} \label{eq:QAR} \hat Q(k)=v (\hat A(k)+\hat A(-k))=v\hat R(k) \end{eqnarray} and the proof is concluded. \end{proof} Given any Markov chain with stationary measure $\pi$, generator $\mathcal L$ and two states $x,y$, we have for any $t>0$ \begin{eqnarray} \label{eq:5} P_t(x,y)\le \frac{\pi(y)}{\pi(x)}. \end{eqnarray} To see this, write \begin{eqnarray} \label{eq:7} P_t(x,y)=[\delta_x e^{t\mathcal L}](y)\le \left[\frac{\pi}{\pi(x)}e^{t\mathcal L}\right](y)=\frac{\pi(y)}{\pi(x)} \end{eqnarray} where we used $\pi(\cdot)/\pi(x)\ge \delta_x(\cdot)$. For the initial configuration, one has from \eqref{eq:46} \begin{eqnarray} \label{eq:8} \pi(\sigma_0)\ge \frac1Z e^{-K_1} \end{eqnarray} for some finite constant $K_1$ depending on $\bar\xi$ (and, typically, of order $m^2$). If instead $\sigma\in\partial \mathcal G$, then one has \begin{gather} \label{eq:WR} \pi(\sigma)\le \frac1Z e^{-K_2(\log \epsilon)^2} \end{gather} for some strictly positive $K_2$ (this is proven below). Finally, observe that \begin{eqnarray} \label{eq:ng} |\partial \mathcal G| = O(\epsilon^{-K_3}) \end{eqnarray} for some constant $K_3$ depending on $m$. This is trivial: indeed, the total number of configurations is bounded by $(\ell/\epsilon)^{m^2}$ (recall the definition of the model, the fact that $\ell$ is independent of $\epsilon$ and that there are $m^2$ particles). Using \eqref{eq:5}, \eqref{eq:8}, \eqref{eq:WR} and \eqref{eq:ng}, we deduce \begin{eqnarray} \label{eq:PGc} \mathbb P(\sigma(t)\in\partial\mathcal G)\le \exp(-K_4(\log \epsilon)^2). \end{eqnarray} As a consequence, \begin{eqnarray} \label{eq:34} \mathbb E\left[\int_0^{\epsilon^{-2}+1}1_{\{\sigma(t)\in\partial\mathcal G\}}\,dt\right]\le (\epsilon^{-2}+1) \exp(-K_4(\log \epsilon)^2). \end{eqnarray} Next observe that, if $\tau_{\mathcal G}<\epsilon^{-2}$, then there exists a probability at least $\delta>0$ independent of $\epsilon$ that the total time spent in $\partial\mathcal G$ up to time $\epsilon^{-2}+1$ is at least $\delta\times\epsilon$. This is simply because the maximal transition rate of the Markov chain is of order $\epsilon^{-1}$ (this is the case when $C_p$ in \eqref{eq:6prime} is of order $1$). In conclusion, \begin{eqnarray} \label{eq:35} \mathbb E\left[\left.\int_0^{\epsilon^{-2}+1}1_{\sigma(t)\in\partial\mathcal G}dt\right|\tau_{\mathcal G}\le \epsilon^{-2}\right]\ge \delta^2\epsilon. \end{eqnarray} Together with \eqref{eq:34} we deduce \begin{eqnarray} \label{eq:36} \mathbb P(\tau_{\mathcal G}\le \epsilon^{-2})\le \frac1{\delta^2\epsilon}\mathbb E\left[\int_0^{\epsilon^{-2}+1}1_{\{\sigma(t)\in\partial\mathcal G\}}\,dt\right]\le \exp(-K_5(\log \epsilon)^2) \end{eqnarray} and \eqref{eq:taulargo} is proven. It remains to prove \eqref{eq:WR}. Recall that $(\hat A(k)+\hat A(-k))$ is negative for every $k\in\mathcal K_m$ and vanishes only for $k=0$, so that \begin{eqnarray} \label{eq:29} -\frac12(\eta, Q \eta)\le -\delta\sum_{k\in\mathcal K_m,k\ne0} |\hat \eta_k|^2 \end{eqnarray} for some positive $\delta$ that depends only on the number of particles $m^2$. If $\sigma\in\partial\mathcal G$ we must have either $|\eta_p-\eta_{p+(1,0)}|\ge (1/2)\log(1/\epsilon)$ or $|\eta_p-\eta_{p+(1,-1)}|\ge (1/2)\log(1/\epsilon)$ for some $p$. Assume to fix ideas that the former is the case. Then, from \eqref{eq:hateta} (and writing $\sum_k$ instead of $\sum_{k\in\mathcal K_m}$) \begin{gather} \label{eq:33} \frac12 \log(1/\epsilon)\le |\eta_p-\eta_{p+(1,0)}|= |\sum_{k}\hat \eta_k \overline{ f_k(p)}(e^{i k_1}-1)|\\= |\sum_{k\ne0}\hat \eta_k \overline{f_k(p)}(e^{i k_1}-1)|\le2 \sqrt{\sum_{k\ne0}|\hat \eta_k |^2} \sqrt{\sum_k|f_k(p)|^2}= 2 \sqrt{\sum_{k\ne0}|\hat \eta_k |^2}. \end{gather} Then, \eqref{eq:WR} immediately follows. \end{proof} \subsection{Proof of Theorem \ref{th:SDE}} Let $U(t)=\{U_p(t)\}_{p\in\mathcal R_m}$ be defined as \begin{eqnarray} \label{eq:Bp} U_p(t)=\frac1{\sqrt\epsilon}\int_0^t\left[-v+ r_p(\sigma(s/\epsilon)\right]ds \end{eqnarray} where $r_p(\sigma)$ is the rate at which particle $p$ jumps $+1$ to the right in the configuration $\sigma$ (this includes the event that it jumps because it is pushed by another particle, i.e. because clock of particle $p'$ rings and $p\in V^+_{p'}$). Note that $\{M_p(t)\}_t\equiv \{\eta_{p,t}-U_p(t)\}_t$ is a martingale, since we have \begin{eqnarray} \label{eq:29bis} \frac d{dt}U_p(t)=\left.\partial_s\mathbb E[\eta_{p,t+s}|\{\eta_{p',t}\}_{p'}]\right|_{s=0^+}. \end{eqnarray} Next, define $V(t)=\{V_{p,p'}(t)\}_{p,p'}$ as solution to \begin{eqnarray} \label{eq:Vpq} \frac d{dt} V_{p,p'}(t)=\left.\partial_s\mathbb E[M_{p,t+s}M_{p',t+s}|\{\eta_{r,t}\}_r]\right|_{s=0^+}\quad V_{p,p'}(0)=0, \end{eqnarray} so that $\{M_p(t)M_{p'}(t)-V_{p,p'}(t)\}_t$ is again a martingale for every $(p,p')$. We will apply \cite[Theorem 4.1, Chapter 7]{cf:EK}, that gives a set of sufficient conditions on the processes $\eta, U$ and $V$ that imply that $\{\eta_{p,t}\}_{t\ge0,p\in\mathcal R_m}$ converges weakly to the solution of \eqref{eq:2}\footnote{A warning on notations: in \cite{cf:EK}, $\eta$ is called $X$, $U$ is called $B$ and $V$ is called $A$}. In particular, conditions (4.3)--(4.5) in \cite{cf:EK} are trivial. For (4.6), we have to check that, for any $T>0$, \begin{eqnarray} \label{eq:4.6} \sup_{t\le T}\left| U_p(t)-\int_0^t\sum_{p'} A_{p,p'}\eta_{p',s}ds \right|\stackrel {\mathbb P} \to0 \end{eqnarray} as $\epsilon\to0$. On the event $\{\tau_{\mathcal G}>\epsilon^{-2}\}$, whose probability tends to $1$ as $\epsilon\to0$ thanks to Lemma \ref{lemma:apriori}, no particle can push any other in configuration $\sigma(s/\epsilon)$ for any $s/\epsilon\le \epsilon^{-2}$ (particles are far away from each other and all the families $V^+_p$ defined just before \eqref{eq:6prime}) include only the particle $p$ itself). Then, $r_p(\sigma(s/\epsilon))$ equals the r.h.s. of \eqref{eq:6prime} for any $s\le T$. We have \begin{gather} \label{eq:circav} r_p(\sigma(s/\epsilon))=\frac{(1-e^{-B-\sqrt\epsilon(\eta_{p+(1,-1),s}-\eta_{p,s})}) (1-e^{-D-\sqrt\epsilon(\eta_{p,s}-\eta_{p-(1,0),s})-\epsilon})} {1-e^{-C-\sqrt\epsilon(\eta_{p,s}-\eta_{p-(0,1),s})-\epsilon}}\\= v+\sqrt\epsilon\left[e^{-B}\frac{1-e^{-D}}{1-e^{-C}}(\eta_{p+(1,-1),s}-\eta_{p,s}) +e^{-D}\frac{1-e^{-B}}{1-e^{-C}}(\eta_{p,s/\epsilon}-\eta_{p-(1,0),s})\right.\\ \left.-e^{-C}\frac{(1-e^{-B})(1-e^{-D})}{(1-e^{-C})^2}(\eta_{p,s}-\eta_{p-(0,1),s}) \right]+O(\epsilon\log(1/\epsilon))\\= v+\sqrt\epsilon\sum_{p'} A_{p,p'}\eta_{p',s}+O(\epsilon\log(1/\epsilon)) \end{gather} where $O(\epsilon\log(1/\epsilon))$ is uniform on $s\le T$. We used that, on $\mathcal G$, $|\eta_{p-b}-\eta_p|\le \log(1/\epsilon)$, $b\in\{(-1,1),(1,0),(0,1)\}$. Plugging this into \eqref{eq:Bp} we get \eqref{eq:4.6}. Finally, condition (4.7) in \cite[Theorem 4.1, Chapter 7]{cf:EK} amounts in our case to requiring that \begin{eqnarray} \label{eq:4.7} \sup_{t\le T}\left|V_{p,p'}(t)-vt \delta_{p,p'}\right|\stackrel {\mathbb P} \to0. \end{eqnarray} Indeed, one has (using that $\{M_{p,t}\}_{t\ge0}$ is a martingale) \begin{gather} \frac d{dt} V_{p,p'}(t)= \partial_s\mathbb E\left[ N_{p,t+s}N_{p',t+s} |\{\eta_{r,t}\}_r \right]|_{s=0^+} \end{gather} where \[ N_{p,t+s}=\sqrt \epsilon x_p(\epsilon^{-1}(t+s))-\frac1{\sqrt \epsilon}\int_0^{t+s}r_p(\sigma(\epsilon^{-1} u))du. \] Note that, by definition of the jump rate $r_p(\sigma)$, \begin{gather} \partial_s\mathbb E\left[\left. x_p(\epsilon^{-1}(t+s))\int_0^{t+s}r_{p'}(\sigma(\epsilon^{-1} u))du\right|\{\eta_{r,t}\}_r\right]_{s=0^+}\\=\epsilon^{-1}r_p(\epsilon^{-1}t)\int_0^tr_{p'}(\sigma(\epsilon^{-1}u)du+x_p(\sigma(\epsilon^{-1}t)r_{p'}(\sigma(\epsilon^{-1}t)) . \end{gather} Therefore, one sees that \begin{gather} \label{eq:30} \frac d{dt} V_{p,p'}(t)=-x_p( t/\epsilon)r_{p'}(\sigma( t/\epsilon)) -x_{p'}(t/\epsilon)r_p(\sigma(t/\epsilon))\\ +\epsilon\left.\partial_s\mathbb E[x_p((t+s)/\epsilon)x_{p'}((t+s)/\epsilon)|\{\eta_{r,t}\}_{r}]\right|_{s=0^+}. \end{gather} If $p=p'$, this gives \begin{gather} \label{eq:31} \frac d{dt} V_{p,p'}(t)=-2x_p( t/\epsilon)r_p(\sigma( t/\epsilon)) +r_p(\sigma( t/\epsilon))(1+2x_p(t/\epsilon))\\=r_p(\sigma(t/\epsilon)) \end{gather} where in the first step we used the fact that when $x_p$ jumps $+1$ (which happens with rate $r_p$), $x_p^2$ increases by $(x_p+1)^2-x_p^2=1+2x_p$. Recall that, on the event $\{\tau_{\mathcal G}\ge \epsilon^{-2}\}$, we have $r_p(\sigma(t/\epsilon))=v+o(1)$, cf. \eqref{eq:circav}. Since the probability of $\{\tau_{\mathcal G}\ge \epsilon^{-2}\}$ tends to $1$, this implies \eqref{eq:4.7} for $p=p'$. As for $p\ne p'$, on the event $\{\tau_{\mathcal G}\ge \epsilon^{-2}\}$ the particles $p$ and $p'$ cannot jump simultaneously since all particles are well spaced all the time so that no particle can push any other. Then, on this event, \begin{gather} \label{eq:32} \epsilon\left.\partial_s\mathbb E[x_p((t+s)/\epsilon)x_{p'}((t+s)/\epsilon)|\{\eta_{r,t}\}_{r}]\right|_{s=0^+}\\=r_{p'}(\sigma(t/\epsilon))x_p(t/\epsilon)+ r_p(\sigma(t/\epsilon))x_{p'}(t/\epsilon) \end{gather} so that $\frac d{dt}V_{p,{p'}}(t)=0$ for every $t\le T$ and \eqref{eq:4.7} follows. \section{Space-time correlations of the SDEs} \subsection{Proof of Theorems \ref{th:2}} In Fourier space, Eqs. \eqref{eq:4} give \begin{gather} \label{eq:dxik} \frac d{dt}\mathbb E_{\bar \xi}(\hat \xi_{k,t})=\hat A(k) \mathbb E_{\bar \xi}(\hat \xi_{k,t}),\\ \label{eq:xixi} \frac d{dt} \mathbb E_{\bar \xi}(\hat \xi_{k,t} \hat \xi_{k',t})= v\delta_{k+k'=0}+\mathbb E_{\bar \xi}(\hat \xi_{k,t} \hat \xi_{k',t}) (\hat A(k')+\hat A(k))\\ \label{eq:xixits} \frac d{dt} \mathbb E_{\bar \xi}(\hat \xi_{k,t} \hat \xi_{k',s})=\hat A(k) \mathbb E_{\bar \xi}(\hat \xi_{k,t} \hat \xi_{k',s}), \quad t>s \end{gather} whose solution is (just differentiate w.r.t. $t$ to check) \begin{gather} \label{eq:47} \mathbb E_{\bar \xi}(\hat \xi_{k,t})=e^{\hat A(k)t}\hat\xi_{k,0}\\ \mathbb E_{\bar\xi}(\hat\xi_{k,t}\hat \xi_{k',t})=v \delta_{k=-k'}\frac{e^{\hat R(k)t}-1}{\hat R(k)}+\mathbb E_{\bar \xi}(\hat \xi_{k,t}) \mathbb E_{\bar \xi}(\hat \xi_{k',t}), \\ \label{aterza} \mathbb E_{\bar\xi}(\hat\xi_{k,t}\hat \xi_{k',s})=v \delta_{k=-k'}e^{\hat A(k)(t-s)}\frac{e^{\hat R(k)s}-1}{\hat R(k)}+\mathbb E_{\bar \xi}(\hat \xi_{k,t}) \mathbb E_{\bar \xi}(\hat \xi_{k',s}), \quad t\ge s. \end{gather} Note that it immediately follows that $\mathbb E_{\bar \xi}(\hat \xi_{k,t} \hat \xi_{k',s})-\mathbb E_{\bar \xi}(\hat \xi_{k,t}) \mathbb E_{\bar \xi}(\hat \xi_{k',s})$ (and therefore $\mathcal W^{\bar \xi}_{p,p'}(t,s)$) is independent of the initial condition $\bar\xi$. Using the first of \eqref{eq:eta-eta} together with \eqref{aterza} we find, for $t\ge s$, \begin{gather} \label{eq:24} \mathcal W^{\bar \xi}_{p,p+y}(t,s)=\frac v{m^2}\sum_{k\in\mathcal K_m} e^{-i k y}e^{\hat A(k)(t-s)}\frac{e^{\hat R(k)s}-1}{\hat R(k)}\\ \label{aseconda} \stackrel{m\to\infty}\to \frac v{(2\pi)^2} \int_{[-\pi,\pi]^2}dk \frac{e^{ \hat R(k)s}-1}{\hat R(k)} e^{(t-s)\hat A(k)}e^{-i k y}. \end{gather} Observe that, as $m\to\infty$, the set $\mathcal K_m$ fills the parallelogram \[\mathcal K_\infty:=\{(k_1,k_2)\in \mathbb R^2:| k_1|\le\pi,|k_2-(C/D)k_1|\le \pi\}.\] However, since the integrand of \eqref{aseconda} is $2\pi$-periodic, integrating over $\mathcal K_\infty$ or on $[-\pi,\pi]^2$ gives the same result. Recall property (3) in Proposition \ref{prop:prop}. The singularity at $(0,0)$ is integrable and the dominant contribution to the integral comes from $k\sim 0$. From \eqref{aseconda} one deduces \eqref{eq:44}, see Appendix \ref{app:326} for details. \subsection{Proof of Theorem \ref{th:muinv}} The gradients $\xi_{p,t}-\xi_{p',t}$ can be written via \eqref{eq:eta-eta} as linear combinations of the Fourier components $\{\hat \xi_{k,t}\}_{ k\in \mathcal K_m, k\ne 0}$. Since the random variables $\hat \xi_{k,t}$ solve a set of linear SDEs, their invariant measure $\mu$ (if it exists) is Gaussian. Stationarity of $\mu$ implies, via \eqref{eq:dxik} and \eqref{eq:xixi}, that for every $k,k' \in\mathcal K_m, k,k'\ne0$, \begin{eqnarray} \label{eq:inv} \mathbb E_\mu(\hat \xi_{k})&=&0\\ \mathbb E_\mu( \hat \xi_{k} \hat \xi_{k'})&=&-\delta_{k=-k'} \frac v{\hat A(k)+\hat A(-k)}. \end{eqnarray} where in the first equality we used that the fact that $\hat A(k)\ne0$ whenever $k\ne0$. Since a Gaussian measure is uniquely characterized by mean and variance, uniqueness of the invariant measure also follows. Using \eqref{eq:eta-eta} we see that \eqref{eq:nomedia} and \eqref{eq:moreg} hold. Again, the dominant contribution comes from $k\sim0$ and one obtains the asymptotics \eqref{4punti}, see Appendix \ref{app:322}. Next we prove the claim about the limit behavior of the random variable $\xi_\phi$. Since $\xi_\phi$ is Gaussian and centered for any $\delta,m$, it suffices to prove that its variance converges to \eqref{eq:varianza}. Note that the support $S_\phi$ of the function $\phi(\delta\cdot)$ has a diameter of order $\delta^{-1}\ll m$. Let $p_0$ be a point outside $S_\phi$, at distance of order $1/\delta$ from the origin. Write \begin{eqnarray} \mathbb E_\mu(\xi_\phi^2)=\delta^4\sum_{p,p'}\phi(\delta p)\phi(\delta p') \mathbb E_\mu((\xi_p-\xi_0)(\xi_{p'}-\xi_{p_0}))+o(1) \end{eqnarray} where we used \begin{eqnarray} \label{eq:sumd} \delta^2\sum_{p'}\phi(\delta {p'})=O(\delta) \end{eqnarray} (because the integral of $\phi$ is zero and the function is smooth) together with ${\rm Cov}_{\mu_\infty}((\xi_p-\xi_0);(\xi_0-\xi_{p_0}))\le C\log (\delta^{-1})$ for $p$ in the support of $\phi(\delta\cdot)$, which follows from \eqref{eq:39} and Cauchy-Schwarz. Using \eqref{4punti} we have \begin{gather} \mathbb E_\mu(\xi_\phi^2)=\frac{v}{2\pi w}\delta^4\sum_{p,{p'}}\phi(\delta p)\phi(\delta {p'})\log\frac{|V(p-p_0)||V q|}{|V(p-{p'})||V p_0|} +o(1) \\ \label{acosa}= \frac{v}{2\pi w}\delta^4\sum_{p,{p'}}\phi(\delta p)\phi(\delta {p'})\log\frac{|V\delta(p-p_0)||V\delta {p'}|}{|V\delta(p-{p'})||V \delta p_0|} +o(1) \end{gather} where in the first step we used \begin{gather} \|\phi\|_\infty^2\delta^4\sum_{p,{p'}\in S_\phi}\frac1{\min(1,|p-p_0|,|{p'}|,|p-{p'}|,|p_0|)}\stackrel{\delta\to0}=o(1) \end{gather} The sum in \eqref{acosa} can be written as \begin{gather} \label{asera} - \frac{v}{2\pi w}\delta^4\sum_{p,{p'}}\phi(\delta p)\phi(\delta {p'})\log|V\delta(p-{p'})| +o(1). \end{gather} Indeed, the terms proportional to $\log |V(p-p_0)|, \log|V {p'}| $ and $\log |V p_0|$ are independent of at least one of the two summation variables $p,{p'}$: then, using once more \eqref{eq:sumd} together with e.g. $|\log |V \delta {p'}||=O(\log(1/\delta))$, \eqref{asera} follows. The sum in \eqref{asera} is the Riemann approximation of the convergent integral \eqref{eq:varianza}.

\section{Introduction} \label{sec:int} \vspace{-2mm} Given the current COVID-19 pandemic, it is essential to enable contactless and smooth-running operations, especially in contact-sensitive facilities like airports. With the ever-enhancing performance of face recognition, the technology has been preferred as a contactless means of verifying identities in applications ranging from border control to logical access control on consumer electronics. However, wearing masks is now essential to prevent the spread of contagious diseases and has been currently forced in public places in many countries. The performance, and thus the trust in contactless identity verification through face recognition can be impacted by the presence of a mask \cite{DBLP:journals/corr/abs-2102-09258}. The effect of wearing a mask on face recognition in a collaborative environment is currently a sensitive issue. This competition is the first to attract and present technical solutions that enhance the accuracy of masked face recognition on real face masks and in a collaborative verification scenario. In a recent study, the National Institute of Standards and Technology (NIST), as a part of the ongoing Face Recognition Vendor Test (FRVT), has published a specific study (FRVT -Part 6A) on the effect of face masks on the performance on face recognition systems provided by vendors \cite{ngan2020ongoing}. The NIST study concluded that the algorithm accuracy with masked faces declined substantially. One of the main study limitations is the use of simulated masked images under the questioned assumption that their effect represents that of real face masks. The Department of Homeland Security has conducted an evaluation with similar goals, however on more realistic data \cite{DHS-Masks-2020}. They also concluded with the significant negative effect of wearing masks on the accuracy of automatic face recognition solutions. A study by Damer et al. \cite{DBLP:conf/biosig/DamerGCBKK20} evaluated the verification performance drop in 3 face biometric systems when verifying masked vs not-masked faces, in comparison to verifying not-masked faces to each other. The authors presented limited data (24 subjects), however, with real masks and multiple capture sessions. They concluded by noting the bigger effect of masks on genuine pairs decisions, in comparison to imposter pairs decisions. This study has been extended \cite{https://doi.org/10.1049/bme2.12040} with a larger database and evaluation on both synthetic and real masks, pointing out the questionable use of simulated masks to represent the real mask effect on face recognition. Recent work has evaluated the human performance in recognizing masked faces, in comparison to automatic face recognition solutions \cite{DBLP:journals/corr/abs-2103-01924}. The study concluded with a set of take-home messages that pointed to the correlated effect of wearing masks on both, human recognizers and automatic face recognition. Beyond recognition, facial masks showed to affect both, the vulnerability of face recognition to presentations attacks, and the detectability of these attacks \cite{DBLP:journals/corr/abs-2103-01546}. There were only a few works that address enhancing the recognition performance of masked faces. Li et al. \cite{li2021cropping} proposed to use an attention-based method to train a face recognition model on the periocular area of masked faces. This presented improvement in the masked face recognition performance, however in a limited evaluation. Moreover, the proposed approach essentially only maps the problem into a periocular recognition problem. A recent preprint by \cite{anwar2020masked} presented a relatively small dataset of 53 identities crawled from the internet. The work proposed to fine-tune FacenNet model \cite{DBLP:conf/cvpr/SchroffKP15} using simulated masked face images to improve the recognition accuracy. Wang et al. \cite{wang2020masked} presented three datasets crawled from the internet for face recognition, detection, and simulated masked faces. The authors claim to improve the verification accuracy from 50\% to 95\% on masked faces. However, they did not provide any information about the evaluation protocol, proposed solution, or implementation details. Moreover, the published part of the dataset does not contain pairs of not-masked vs masked images for the majority of identities. A work by Montero et al. \cite{DBLP:journals/corr/abs-2104-09874} proposed to combine ArcFace loss with a specially designed mask-usage classification loss to enhance masked face recognition performance. Boutros et al. \cite{DBLP:journals/corr/abs-2103-01716} proposed a template unmasking approach that can be adapted on the top of any face recognition network. This approach aims to create unmasked-like templates from masked faces. This goal was achieved on top of multiple networks by the proposed self-restrained triplet loss \cite{DBLP:journals/corr/abs-2103-01716}. On a related matter, a rapid number of works are published to address the detection of wearing a face mask \cite{app11052070,loey2021hybrid,wang2021wearmask,qin2020identifying}. These studies did not address the effect of wearing a mask on the performance of face recognition or present solution to improve masked face recognition. Besides the exclusive interest in face recognition accuracy, there is a growing interest in compact face recognition models \cite{martinez2021benchmarking}. This interest is driven by the demand for face recognition deployment on consumer devices and the need to enhance the throughput of face recognition processes. A major challenge has been organized in ICCV 2019 to motivate researchers to build lightweight face recognition models \cite{DBLP:conf/iccvw/DengGZDLS19}. MobileFaceNets are an example of such face recognition models \cite{DBLP:conf/ccbr/ChenLGH18}. MixedFaceNets \cite{mixedfacenets} are a recent example where mixed depthwise convolutional kernels, with a tailored head and embedding design and a shuffle operation, are utilized to achieve high recognition accuracies with extremely light models. Motivated by (a) the hygiene-driven wide use of facial masks, (b) the proven performance decay of existing face recognition solutions when processing masked faces, (c) the need to motivate novel research in the direction of enhancing masked face recognition accuracy, and (d) the requirement of light-weight models by various applications, we conducted the IJCB Masked Face Recognition Competition 2021 (IJCB-MFR-2021). The competition attracted submissions from academic and industry teams with a wide international representation. The final participation toll was 10 teams with valid submissions. These teams submitted 18 valid solutions. The solutions were evaluated on a database collected to represent a collaborative face verification scenario with individuals wearing real face masks. This paper summarises this competition with a detailed presentation of the submitted solutions and the achieved results in terms of masked vs masked face verification accuracy, masked vs not-masked face verification accuracy, and the compactness of the recognition models. In the next sections, we start by introducing the competition evaluation database, the evaluation criteria, and the participating teams. Then, in Section \ref{sec:alg}, short descriptions of the submitted solutions are listed. In Section \ref{sec:res}, we present and discuss the achieved results along with listing the winning submissions. We end the paper in Section \ref{sec:con} with a final general conclusion. \vspace{-3mm} \section{Database, evaluation criteria, and participants} \label{sec:data} \vspace{-1mm} \subsection{Database} \vspace{-1mm} The evaluation data, the masked face recognition competition data (MFRC-21), simulates a collaborative, yet varying scenario. Such as the situation in automatic border control gates or unlocking personal devices with face recognition, where the mask, illumination, and background can change. The database is collected by the hosting institute and not available publicly. The data is collected on three different, not necessarily consecutive days. We consider each of these days as one session. On each day, the subjects have collected three videos, each of a minimum length of 5 seconds (used as single image frames). The videos are collected from static webcams (not handheld), while the subjects are requested to look at the camera, simulating a login scenario. The data is collected by subjects at their residences during the pandemic-induced home-office period. The first session is considered a reference session, while the other two were considered probe sessions. Each day contained three types of captures, no mask, masked with natural illumination, masked with additional illumination. The database participants were asked to remove eyeglasses only when the frame is considered very thick. No other restrictions were imposed, such as background or mask type and its consistency over days, to simulate realistic scenarios. The first second of each video was neglected to avoid possible biases related to the subject interaction with the capture device. After the neglected one second, three seconds were considered. From these three seconds, 10 frames are extracted with a gap of 9 frames between each consecutive frame, knowing that all videos are captured at a frame rate of 30 frames per second. The final considered portions of the database in the competition are (a) the not-masked baseline reference from the first session (noted as BLR), (b) the masked reference from the first session (noted as MR), and (c) the masked face probes from the second and third sessions under both illumination scenarios (noted as MP). A summary of the used database is presented in Table \ref{tab:DB} and samples of the database are presented in Figure \ref{fig:samples}. The database contained 47 subjects, all of them participated in all the sessions. All the subject provided their informed consent to use the data for research purposes. Two evaluation setups are considered, (a) not-masked vs masked, where all images in BLR are compared to all images in MP (noted as BLR-MP), and (b) masked vs. masked, where all images in MR are compared to all images in MP (noted as MR-MP). \input{data_summary} \input{sample_images} \subsection{Evaluation criteria} \label{sec:eval_criteria} The solutions evaluation will be based on both, the verification performance and the compactness of the used mode/models. The verification evaluation will be based on the verification performance of masked vs. not-masked verification pairs, as this is the common scenario, where the reference is not-masked, while the probe is masked, e.g. in entry to a secure access area. This scenario will be noted as BLR-MP. However, the performance of masked vs. masked verification pairs is also be reported in this paper. This scenario is noted as MR-MP. The verification performance is evaluated and reported as the false non-match rate (FNMR) at different operation points FMR100, FMR1000, which are the lowest FNMR for a false match rate (FMR) $<1.0\%$ and $<0.1\%$, respectively. The verification performance evaluation of the submitted solutions is based on FMR100. To get an indication of generalizability, we also report a separability measure between the genuine and imposter comparison scores. This is measured by the Fisher Discriminant Ratio (FDR) as formulaed in \cite{poh2004study}. To consider the deployability of the participating solutions, we will also consider the compactness of the model (represented by the number of trainable parameters \cite{DBLP:journals/corr/abs-2103-06877}) in the final ranking. The participants are asked to report the number of trainable parameters and can be asked to provide their solutions to validate this number. The final teams ranking is be based on a weighted Borda count, where the participants will be ranked by (a) the verification metric as mentioned above (noted as Rank-a), and (b) by the number of trainable parameters in their model/models (notes as Rank-b). For Rank-a, the solutions with lower FMR100 are ranked first, and for Rank-b, the solutions with the lower number of trainable parameters are ranked first. In the final ranking, Rank-a will have 75\% weight and Rank-b will have 25\% weight. Each participant is given a Borda count (BC) for each ranking criteria (BC-a and BC-b). For example, if solution X is ranked first out of 10 participants in the verification performance rank-a (BC-a =9) and third out of 10 solutions in model compactness Rank-b (BC-b = 7) (this corresponds to BC = total number of solutions – rank). Then the weighted Borda count w-BC = 0.75x9+0.25x7= 8.5. Therefore, the final score of solution X is 8.5 and higher indicates a better solution. The solutions are ranked from the highest w-BC to the lowest w-BC. \subsection{Submission and evaluation process} Each of the teams was requested to submit their solutions as Win32 or Linux console applications. These applications should be able to accept three parameters, evaluation-list (text file), landmarks (text file), and an output path. The evaluation-list contains pairs of the path to the reference and probe images and a label for each of the compared images, indicating if the image is masked or not. The landmarks provided a bounding box and five landmark locations of the images as detected by the MTCNN solution \cite{zhang2016joint}. Only the pairs of images with valid detected faces are provided to the solutions in the evaluation-list. From the initial considered data, the face detector \cite{zhang2016joint} did not provide valid face detections. For the BLR-MP pairs, 4.42\% of the pairs contained invalid detections of faces and thus not considered in the evaluation. for the MR-MP pairs, 4.75\% of the pairs contained invalid detections of faces and thus not considered in the evaluation. The output of the solution application script is a text file containing comparison scores for each pair in the evaluation-list. \subsection{Competition participants} The competition aimed at attracting participants with a high geographic and activity variation. The call for participation was shared on the International Joint Conference on Biometrics (IJCB 2021) website, on the competition own website \footnote{https://sites.google.com/view/ijcb-mfr-2021/home}, on public computer vision mailing lists (e.g. CVML e-Mailing List), and through private e-Mailing lists. The call for participation has attracted 12 registered teams. Out of these, 10 teams have submitted valid solutions. These 10 teams have affiliations based in nine different countries. Seven of the 10 teams are affiliated with academic institutions, two are affiliated with the industry, and one team has both academic and industry affiliations. Only one of the participating teams has chosen to be anonymous. Each team was allowed to submit up to two solutions. The total number of validly submitted solutions is 18. A summary of the participating teams is presented in Table \ref{tab:team}. \input{Tables/Teams} \input{Tables/results_NvsM} \input{Tables/results_MvsM} \input{Tables/summary} \section{Submitted solutions} \label{sec:alg} Ten teams have been registered for MFR 2021 competition and submitted 18 valid solutions. Table \ref{tab:team} presents a summary of the registered team members and their affiliation, submitted solutions, and type of institution of each registered team (Academic, Industry, or mix of both academic and industry). In the following, we provide a brief description of the valid submitted solutions: \vspace{-2mm} \paragraph{A1\_Simple} employed ArcFace \cite{DBLP:conf/cvpr/DengGXZ19} to train a ResNet model. A1\_Simple applied MaskTheFace \cite{anwar2020masked} method to synthetically generate masked face images in the training dataset- MS1MV2. A1\_Simple is trained with cosine annealing LR scheduling to adjust the learning rate. In the evaluation phase, A1\_Simple used the provided landmark facial point and bounding box in the MFRC-21 to align and crop the face image to $112 \times 112$. The feature embedding of the presented solution is of size 512-D. The model is trained with ArcFace loss. During the training phase, three data augmentation methods are used- random resized crops, random horizontal flip, and color jittering. \vspace{-2mm} \paragraph{TYAI} solution uses Sub-center ArcFace \cite{DBLP:conf/eccv/DengGLGZ20} and ir-ResNet152 model to train a masked face recognition model on Glint360K dataset \cite{DBLP:journals/corr/abs-2010-05222}. The proposed solution randomly augmented half of the training dataset with a synthetic generated mask using five types of transparent masks. The input image size of the proposed model is $112 \times 112$ and the size of the output feature embedding is 512-D. During the training, additional four data augmentation methods are used: random crop by resizing the image to $128 \times 128$ and then randomly cropping it to $112 \times 112$, random horizontal flip, random rotation, and random affine. The model uses a Sub-center ArcFace loss to train the proposed solution. \vspace{-2mm} \paragraph{Mask aware ArcFace (MaskedArcFace)} opts to generate a masked twin dataset from MS1MV2 \cite{DBLP:conf/eccv/GuoZHHG16,DBLP:conf/cvpr/DengGXZ19} dataset and to combine them during the training process. Both datasets are shuffled separately using the same seed and, for every new face image selected for the input batch, MaskedArcFace decides whether the image is taken from the original (not-masked) or the masked dataset with a probability of 50\%. MaskedArcFace use ArcFace \cite{DBLP:conf/cvpr/DengGXZ19} as the baseline work. MaskedArcFace selects the dataset recommended by ArcFace (MS1MV2) \cite{DBLP:conf/eccv/GuoZHHG16,DBLP:conf/cvpr/DengGXZ19} as the training dataset, which contains 5.8M images and 85,000 identities. MaskedArcFace uses IResNet-50 as the backbone among all the network architectures tested in the ArcFace repository as it is it offers good trade-off between the accuracy and the number of parameters. For the generation of the masked version of the dataset, MaskedArcFace uses MaskTheFace \cite{anwar2020masked}. The types of masks considered are surgical, surgical green, surgical blue, N95, cloth, and KN95. The mask type is selected randomly with a 50\% probability of applying a random color and a 50\% probability of applying a random texture. During the evaluation phase, MaskTheFace uses the provided landmark points and the bounding box provided by the competition to align and crop face images. The feature embedding produced by MaskedArcFace solution is of the size 512-D and the input face image is of the size $112 \times 112$ pixels. \vspace{-2mm} \paragraph{Multi-task ArcFace (MTArcFace)} utilized the same training dataset, loss function, backbone, and mask generation method as in MaskedArcFace. MTArcFace adds another dense layer in parallel to the one used to generate the feature vector by IResNet-50, just after the dropout layer. The new dense layer generates an output with two floats, which correspond to the scores related to the probability that the face is masked or not, respectively. This way, MTArcFace aims to force the network to learn when a face is wearing a mask. This information will also be used by the layer that generates the feature vector. The data preprocessing steps and the size of the feature embedding are identical to the MaskedArcFace. \vspace{-2mm} \paragraph{Masked face recognition using non-masked region extraction and fine-tuned recognition model (MFR-NMRE-F)} Based on the 5-point face landmark detections, the proposed approach identifies a crop that corresponds to the upper facial region where masks are not visible. Then, MFR-NMRE-F fine-tuned a VGG2-SE-ResNet-50 face recognition model for the classification task on these crops using the VGGFace2 \cite{cao2018vggface2} training dataset processed with the RetinaFace \cite{retinaface} detector. For the evaluation, MFR-NMRE-F uses the provided face landmarks provided by MFRC-21, since they correspond closely to the RetinaFace results obtained on the training dataset. Using the landmark coordinates, the MFR-NMRE-F solution extracts the upper face region, extracts feature vectors using the fine-tuned VGG2-SE-ResNet-50 model, and compares features using the cosine similarity measure. The proposed method is trained using cross-entropy (CE) loss. The input size of the proposed model is $96 \times 192$ and the feature embedding size is 2048-D. \vspace{-2mm} \paragraph{Masked face recognition using non-masked region extraction and pre-trained recognition model (MFR-NMRE-B)} identifies a crop that corresponds to the upper facial region where masks are not visible based on the 5-point face landmark. MFR-NMRE-B utilized a VGG2-SE-ResNet-50 model pre-trained for the classification task using the VGGFace2 \cite{cao2018vggface2} training dataset. Different from MFR-NMRE-F, the MFR-NMRE-B solution did not fine-tune the feature extraction model with cropped images. For the evaluation, the proposed method uses the provided face landmarks provided by MFRC-21. Using the landmark coordinates, the proposed method crops the upper face region, extracts feature vectors using the VGG2-SE-ResNet-50 model, and compares features using the cosine similarity measure. MFR-NMRE-B is trained using Softmax cross-entropy loss. The input size of the proposed model is $112 \times 224$ and the feature embedding size is 2048-D. \vspace{-2mm} \paragraph{Masked-Unmasked Face Matching Net (MUFM Net)} utilizes Momentum Contrast (MoCo) \cite{DBLP:conf/cvpr/He0WXG20} to create an initial embedding using a ResNet-50 model trained on CelebA dataset \cite{DBLP:conf/iccv/LiuLWT15}. Then, synthetic masked versions of CelebA, Spectacles on Faces \cite{DBLP:journals/jvcir/AfifiA19}, Youtube Faces \cite{wolf2011face} and LFW \cite{LFWTech} are created as defined in \cite{ngan2020ongoing}. The initial model is fine-tuned using these dataset. For fine-tuning, MUFM Net uses a siamese network with shared weights with absolute differences taken at the last bottleneck layer. This difference is fed into a 512 fully connected layer followed by a single softmax node The model is fine-tuned with binary cross-entropy loss with 50\% of layers frozen. The input size of the presented model is $224 \times 224$ pixels. \vspace{-2mm} \paragraph{Ensemble MUFM Net (EMUFM Net)} builds upon MUFM to create an ensemble. First, the best-performing MUFM models are selected based on the validation accuracy. The selected models are M1 (obtained after 695K iterations) and M2 (obtained after 885K iteration) These models are fine-tuned on hard examples drawn from the training set. Three models are fine-tuned- E1 and E2 builds on M1 where 90\% and 80\% of the layers are frozen, respectively, and E3 builds on M2 where 50\% of the layers are frozen. All these models have an input of size $224 \times 224$ and an output embedding of size 2048-D. During the testing phase, the similarity scores of these three models (E1-3) are averaged to provide the final similarity score. \vspace{-2mm} \paragraph{VIPLFACE-M} adopted ResNet-100 \cite{DBLP:conf/cvpr/HeZRS16} and ArcFace loss \cite{DBLP:conf/cvpr/DengGXZ19} for face recognition. The proposed solution uses a refined version of MS1M dataset \cite{DBLP:conf/eccv/GuoZHHG16} for training the proposed solution. The number of face images in the training dataset is 3.8M of 50K identities. VIPLFACE-M uses the synthetic mask creation method defined in \footnote{\url{https://github.com/JDAI-CV/FaceX-Zoo/blob/main/addition_module/face_mask_adding/FMA-3D/README.md}} to add synthetic masks on part of the training dataset. The number of synthetically masked face images used in the training is 500K and the number of synthetically masked identities is 50K. During the training phase, the proposed solution uses random flipping as a data augmentation method. The input size of the presented solution is $ 112 \times 112$ and the output feature embedding size is 512-D. \vspace{-2mm} \paragraph{VIPLFACE-G} is based on training ResNet-100 model \cite{DBLP:conf/cvpr/HeZRS16} with ArcFace loss \cite{DBLP:conf/cvpr/DengGXZ19}. The input size of the presented solution is $ 112 \times 112$ and the feature embedding size is 512-D. The model is trained on a clean version of MS1M \cite{DBLP:conf/eccv/GuoZHHG16} that contains 5.8M of 80K identities. The presented solution uses random flip to augment the dataset during training. \vspace{-2mm} \paragraph{SiMiT Lab – Masked Face Recognition–1 (SMT-MFR-1)} employs LResNet-100E-IR model \cite{DBLP:conf/cvpr/HeZRS16} trained with ArcFace loss function \cite{DBLP:conf/cvpr/DengGXZ19}. The model is originally trained on MS1MV2 dataset \cite{DBLP:conf/eccv/GuoZHHG16,DBLP:conf/cvpr/DengGXZ19}. SMT-MFR-1 solution depends on fine-tuning LResNet100E-IR using two real world masked face datasets- Real World Occluded Faces (ROF) \footnotemark\footnotetext{\url{https://github.com/ekremerakin/RealWorldOccludedFaces}} and MFR2 dataset \cite{anwar2020masked}. MFR2 contains 296 images of 53 identities. ROF dataset is crawled from the internet and contains 678 masked face images and 1853 not-masked face images of 123 identities. The proposed solution is fine-tuned using the ROF dataset and a part of the MFR2 dataset (35 identities). The model process input image of size $112 \times 112$ to produce feature embedding of size 512-D. During the training, the training dataset is augmented using a horizontal flip augmentation method. \vspace{-2mm} \paragraph{SiMiT Lab – Masked Face Recognition–2 (SMT-MFR-2)} is conceptually identical to SMT-MFR-1. Different from SMT-MFR-1, the SMT-MFR-2 model is fine-tuned using the ROF dataset and the entire MFR2 dataset. \vspace{-2mm} \paragraph{LMI - SiMiT Lab - Masked Face Recognition - 1 (LMI-SMT-MFR-1)} is a combination of two solutions- MFR-NMRE-F and SMT-MFR-1. First, the features are extracted separately by each of the solutions- MFR-NMRE-F and SMT-MFR-1. Then, the comparison scores are calculated for each solution. To combine the scores, cosine similarity measures are converted to euclidean distance in MFR-NMRE-F. The output of SMT-MFR-2 is euclidean distance. After this, the scores are normalized separately for each solution. Then, both scores are multiplied to generate the ensemble score. \vspace{-2mm} \paragraph{LMI - SiMiT Lab - Masked Face Recognition - 2 (LMI-SMT-MFR-2)} is also a combination of two solutions-MFR-NMRE-B and SMT-MFR-1. LMI-SMT-MFR-2 follows the same scores fusion method described in the LMI-SMT-MFR-1 solution. \vspace{-2mm} \paragraph{Ignoring masks for accurate masked face recognition (IM-MFR)} approach consists of two different training processes. The first, which aims to build a classification model, uses 6000 training identities from the VGGFace2 dataset \cite{cao2018vggface2} to minimize the cross-entropy while classifying these images. Each image had a probability of 65\% of being masked. All training images are randomly resized and cropped to $224 \times 224$ In this solution, the masked creation method \cite{ngan2020ongoing} uses the open implementation \footnotemark\footnotetext{\url{https://github.com/fdbtrs/MFR/blob/master/FaceMasked.py}} by Boutros \etal \cite{DBLP:journals/corr/abs-2103-01716}. After achieving above 96\% accuracy in the classification on the validation set, the last fully-connected layer was replaced with a fully connected layer with 512 outputs units. All the layers, except the newest one, are now frozen. The last layer is trained with and joint Triplet Loss and MSE for metric learning. The backbone network is a ResNet-50 \cite{he2016deep}. The model is trained for 65k iterations. \vspace{-2mm} \paragraph{Ignoring masks for accurate masked face recognition (IM-AMFR)} follows the same training procedure, architecture, and loss function as in IM-MFR. The only difference is the number of training iteration where the IM-AMFR model is trained for 32k training iterations. \vspace{-2mm} \paragraph{anonymous-1 and anonymous-2} employed FaceNet \cite{DBLP:conf/cvpr/SchroffKP15} as base architecture pre-trained on VGGFace2 \cite{cao2018vggface2}. MaskTheFace \cite{anwar2020masked} is used to augment the LFW \cite{LFWTech,DBLP:conf/nips/HuangMLL12} dataset and create a masked-face dataset. A masked version of each image in LFW is created. The FaceNet model is then fine-tuned using the augmented dataset. In the anonymous-1 solution, the model is fine-tuned using only masked face images. In the anonymous-2 solution, the model is fine-tuned using pairs of unmasked and masked images. For inference, the last layer of FaceNet consists of 512-dimensional embeddings, while the input size for both solutions is $160 \times 160$ pixels. One must note that the presented approach is reasonable, however, the verification accuracy presented in Section \ref{sec:res} is extremely low, which might indicate an implementation error in the submission. \vspace{-2mm} \paragraph{Baseline} The baseline is chosen to put the submitted approaches in perspective of state-of-the-art face recognition model performance. The considered baseline is the ArcFace, which scored state-of-the-art performance on several face recognition evaluation benchmarks such as LFW $99.83\%$ and YTF $ 98.02\%$ by using Additive Angular Margin loss (ArcFace) to improve the discriminative ability of the face recognition model. We considered ArcFace based on ResNet-100 \cite{DBLP:conf/cvpr/HeZRS16} architecture pretrained on refined version of the MS-Celeb-1M dataset \cite{DBLP:conf/eccv/GuoZHHG16} (MS1MV2). \begin{figure}[h] \centering \includegraphics[width=0.99\linewidth]{ROCroc.pdf} \vspace{-4mm} \caption{ The ROC curve scored by the top 10 solutions in the BLR-MP experimental setting. } \label{fig:roc} \vspace{-4mm} \end{figure} \input{vis_comparison} \vspace{-2mm} \section{Results and analyses} \label{sec:res} This section presents comparative evaluation results of the submitted solution. We present first the achieved results on the BLR-MP evaluation setting and the model compactness. Then, we present the achieved results on the MR-MP evaluation setting. \subsection{Not-masked vs. masked (BLR-MP)} Table \ref{tab:blr_mp} presents comparative evaluation results achieved by the submitted solutions for BLR-MP evaluation setting and the model compactness. The results are reported and ranked based on the evaluation criteria described in Section \ref{sec:eval_criteria}. From the reported results in Table \ref{tab:blr_mp} we made the following observations: \begin{itemize} \vspace{-1mm} \item Based on the defined evaluation criteria in Section \ref{sec:eval_criteria}, the top-ranked solution based on the weighted Borda count is TYAI (rank 1), followed by MaskedArcFace and SMT-MFR-2 (rank 2) and then A1\_Simple and VIPLFACE-M (rank 3). \vspace{-1mm} \item Most of the presented solutions achieved a competitive verification performance, in comparison to the baseline. Ten out of 18 solutions achieved higher verification performance than the baseline solution for the BLR-MP evaluation setting as reported in Table \ref{tab:blr_mp} and Figure \ref{fig:roc}. Figure \ref{fig:roc} presented the achieved verification performances in term of Receiver operating characteristic (ROC) curves by the top 10 solution on the BLR-MP experimental setting. The best verification performance in terms of FMR100 is achieved by the TYAI solution, where the achieved FMR100 was 0.05095 (Table \ref{tab:blr_mp} and Figure \ref{fig:fmr100_BLR_MP}). \vspace{-1mm} \item By comparing the verification performances reported in Table \ref{tab:blr_mp} and the loss function utilized by each of the solution reported in Table \ref{tab:solutions_summary}, it is noted that the models trained with margin-based softmax loss (ArcFace or Sub-center ArcFace loss) achieved higher verification performance than the models trained with other loss functions including cross-entropy and triplet loss. This points out the generalizability brought by the nature of the marginal penalty that forces a better separability between classes (identities) and better compactness within classes. \vspace{-1mm} \item The solutions that achieved competitive FMR100 to the baseline solution have relatively higher separability between genuine and imposter scores (FDR) than other solutions that achieved relatively lower verification performance. \vspace{-1mm} \item Regarding model compactness, all solutions contain between 23M and 108M parameters as shown in Table \ref{tab:blr_mp} and Figure \ref{fig:param}. The top 3 ranked solutions have less than 87M parameters. This indicates that utilizing a larger and deeper deep learning model does not necessarily and solely lead to higher verification performance. \vspace{-1mm} \item The common strategy to improve the masked face recognition verification performance by the submitted solutions is to augment the training dataset with a simulated mask. All submitted solutions depended on training or fine-tuning face recognition model with masked face images (real or simulated). However, none of the presented solutions propose a solution that could be applied on top of the existing face recognition model, as in \cite{DBLP:journals/corr/abs-2103-01716}. Furthermore, none of the presented solutions has clearly benefited from the mask labels included in the evaluation list. Four of the five top-ranked solutions utilized synthetically generate masks to augment the training dataset with simulated masked images. Utilizing such a method is usually easier than other solutions, such as using a real masked training dataset. Collecting a large-scale training dataset with pairs of not-masked/masked face images is, however, not a trivial task. \end{itemize} \vspace{-1mm} \subsection{Masked vs. masked (MP-MR)} \vspace{-1mm} The verification performance of the experimental setting MR-MP for all submitted solutions is presented in Table \ref{tab:mr_mp}. The achieved verification performance is reported in terms of FMR100, FMR1000, and FDR. The presented results are ordered and ranked based on the achieved FMR100. It can be noted from the reported verification performance in Table \ref{tab:mr_mp} that ten out of 18 solutions achieved better verification performance than the baseline solution when comparing masked reference to masked probe (MR-MP). TYAI solution achieved the best verification performance followed by VIPLFACE-M and A1\_Simple. By comparing the reported verification performance of BLR-MP evaluation setting (Table \ref{tab:blr_mp}) and the reported one of MR-MP (Table \ref{tab:mr_mp}), we can observe the following: a) Most of the solutions have higher separability between genuine and imposter scores (higher FDR) when both reference and probe are masked (MR-MP) than the case where only the probe are masked (BLR-MP). b) The top-ranked solutions in the MR-MP evaluation setting are also ranked among the top solutions in the BLR-MP evaluation setting. \section{Conclusion} \label{sec:con} Driven by the pandemic-driven use of facial masks, the Masked Face Recognition Competitions (MFR 2021) was organized to motivate and evaluate face recognition solutions specially designed to perform well with masked faces. A total of 10 teams from 11 affiliations participated in the competition and contributed 18 solutions for the evaluation. The evaluation focused on not-masked vs. masked face verification accuracy, the masked vs. masked face verification accuracy, and the face recognition model compactness. Out of the 18 submitted solutions, 10 achieved lower verification error (FMR100) than the considered baseline. Most of the top-performing solutions used variations of the ArcFace loss and either real or simulated masked face databases in their training process. The lowest achieved FMR100 for the not-masked vs. masked evaluation was 5.1\%, in comparison to an FMR100 of 6.0\% scored by the baseline. \vspace{-2mm} { \paragraph{Acknowledgments:} This research work has been funded by the German Federal Ministry of Education and Research and the Hessen State Ministry for Higher Education, Research and the Arts within their joint support of the National Research Center for Applied Cybersecurity ATHENE. } {\small \bibliographystyle{ieee}