Datasets:

kjappelbaum

/

chemnlp-arxiv

like 1

# Introduction Network operators continuously upgrade their networks due to increasing demands for ubiquitous communications. As popularities of cloud computing, Internet of Things (IoT), and 5G mobile communications increase, network traffic is becoming heavy and extremely bursty. Thus, just enlarging the network capacity is not an economical choice, while neglecting it will strongly affect the QoS. Also, emerging applications, such as online gaming, data backups and virtual machine migrations, result in heterogeneous demands in telecom networks. Today's network users request more customized services, such as Virtual Private Network (VPN) and Video on Demand (VoD), and more differentiated demands with different prices. For some of the traffic which is delay-insensitive and can accept some compromise in bandwidth or other aspects, it can be preempted by more "important\" requests when the network becomes congested. Thus, to maintain cost-effectiveness and customers' loyalty, network operators can provide different grades of service besides sufficient bandwidth, instead of trying to support all the traffic without distinction. To address these problems, *degraded provisioning* is proposed to provide a degraded level of service when network congestion occurs instead of no service at all. Generally, degraded provisioning has two approaches: 1) keep total amount of transferred traffic constant by time prolongation or modulation-level adjustment with immediate service access (QoS-assured), or 2) degrade requested bandwidth without time or modulation compensation, or no guarantee for immediate access (QoS-affected). We focus on QoS-assured degraded provisioning in this study. In a multi-layer network, QoS-assured degradation has different implementation methods in different layers. In electric layer, for a delay-insensitive and degradation-tolerant request, when network bandwidth is scarce, we degrade its transmission rate to enlarge the available bandwidth and extend its holding time based on the premise that total traffic amount is unchanged. Note that a traffic request cannot be degraded arbitrarily, and it is constrained by a given deadline. In elastic optical layer, degradation refers to decreasing the number of occupied spectrum slots of a lightpath and raising the modulation level to guarantee the capacity. In OFDM-based elastic optical networks, modulation level can be dynamically reconfigured in DSP and DAC/ADC via software. But optical degradation has a constraint that higher-order modulation has shorter transmission reach. Due to the flexibility enabled by degraded provisioning, there exist many studies on this topic in different kinds of optical networks. In WDM networks, Roy *et al.* studied degraded service using multipath routing in a QoS-affected way. Zhang *et al.* studied reliable multipath provisioning, exploiting flexibility in bandwidth and delay. Andrei *et al.* proposed a deadline-driven method to flexibly provision services without immediate access. Savas *et al.* introduced a dynamic scheme to reduce blocking by exploiting degraded-service tolerance in a QoS-affected way, and they also applied this method to increase network survivability. In Mixed-Line-Rate (MLR) networks, Vadrevu *et al.* proposed a QoS-affected degradation scheme using multipath routing considering minimum-cost network design. But the ITU-T grid limit puts constraints on optical layer flexibility. As a major development of optical technology, elastic optical networking enables more flexibility in optical modulation and spectrum allocation. Distance-adaptive spectrum allocation is a similar approach as optical degradation, but its limitations are that the modulation format of a lightpath is configured at one time and cannot be adjusted based on the fluctuation of traffic. Gkamas *et al.* proposed a dynamic algorithm for joint multi-layer planning in IP-over-flexible-optical networks without considering dynamic adjustment of lightpath modulation. Recent progress in modulation format conversion enables all-optical OOK to 16QAM adjustment, and its advantages in elastic optical networking were demonstrated in Yin *et al.* with modulation-formats-convertible nodes. Thus, degraded provisioning can be extended to the elastic optical layer, and this important issue has not been fully understood, with no previous studies. We summarize our contributions as follows: 1) to the best of our knowledge, this is the first investigation on QoS-assured degraded provisioning problem in multi-layer networks with optical elasticity; 2) we propose an enhanced multi-layer network architecture and solve the complex dynamic degraded routing problem; and 3) we further propose novel dynamic heuristic algorithms in both layers to achieve degraded resource allocation. Results show that a significant increase of service acceptance can be achieved by multi-layer degradation. # Dynamic Degraded Provisioning Scheme We decompose the online degraded provisioning problem into two subroutines: 1) degraded routing, and 2) degraded resource allocation. ## Degraded Routing Degraded routing solves the subproblem of degraded-route computation when conventional routing cannot be performed due to resource shortage. Optical degraded routing acts similarly as electric degraded routing. The term *request* here refers to the *lightpath request* in optical layer, and the *service request* in electric layer. There are two concerns: route hops (RH) and potential degraded requests (PDR). RH denotes the amount of resources occupied by the new request, while PDR denotes how many existing requests may be affected. We define a link in any layer of the multi-layer network as a tuple: \(V_{ij,k}=(\mathbf{\Theta}, C)\), which is the \(k^{th}\) link from \(i\) to \(j\). \(\mathbf{\Theta}\) is a set that contains existing requests routed on this link, and \(C\) is the available capacity of this link. \(r_{n}\) is a request, and \(\mathcal{P}_{r_n}\) is a degraded route for \(r_{n}\) in electric or optical layer. \[\small \mathcal{P}_{r_n}=\{V_{ij,k}|r_n \in V_{ij,k}.\mathbf{\Theta}\} \vspace{-0.4em}\] We introduce two metrics to evaluate the route \(\mathcal{P}_{r_n}\). Note that the \(|\cdot|\) operation returns the number of elements in a set. \[\small \mathcal{N}_{RH}=|\mathcal{P}_{r_{n}}| \vspace{-0.3em}\] \[\small \mathcal{N}_{PDR}=|\bigcup^{|\mathcal{P}_{r_{n}}|}_{c=1}\mathcal{P}_{r_{n}}[c].\mathbf{\Theta}| \vspace{-0.5em}\] To calculate a route for minimizing RH, Dijkstra algorithm is applied. However, minimizing PDR is not that easy, because the minimizing-PDR problem aims to obtain a route that has the smallest PDR among all possible routes between a given source-destination \((s,d)\) pair. A straight-forward idea is to list all possible routes between a given \((s,d)\) pair and compare their PDR. But the complexity of this process is \(O((N-2)!)\) (\(N\) denotes the number of nodes). Here, we propose the enhanced multi-layer network architecture by introducing the auxiliary *service layer*, which lies directly above the electric layer, and we solve the problem in polynomial time. In the enhanced multi-layer architecture (Fig. 1), all nodes in the optical layer are mapped to upper two layers. There are two kinds of directional weighed links, i.e., request link and resource link. Request link weight is equal to a given large number times a binary that indicates whether there are existing requests between the node pair. Resource link weight is binary, which means if there is sufficient resource for this request. Note that, if the upper layer has isolated nodes (no connected edges), and the isolated node happens to be the source or destination of the request, then we replace the isolated node with the originating or terminating nodes of lightpaths (or requests) that bypass the isolated node. For example, in Fig. 1, \(e^*\) is bypassed by request \#2 (\(a^*\)-\(e^*\)-\(c^*\)) on service layer. Thus, if we want to compute a degraded route on electric layer with minimized PDR for a new request (\(a^*,e^*\)), we should replace \(e^*\) with \(c^*\), and execute a shortest-path algorithm on service layer for the auxiliary request (\(a^*,c^*\)). Therefore, the *minimizing-PDR problem* on one layer is transformed into a *weighed shortest-path problem* on the upper layer. Finally, an actual route will be acquired based on the computed shortest path on upper layer.

{'timestamp': '2017-01-05T02:05:47', 'yymm': '1607', 'arxiv_id': '1607.04695', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04695'}

# Introduction ## The main theorem Let \(F\) be a totally real field, \(p\) be an odd prime, and \({\mathcal O}\) the ring of integers of a finite extension of \(\Q_p\). In proving automorphy of geometric representations \(\rho:G_F \rightarrow GL_2({\mathcal O})\) which are residually automorphic, there is an assumption made (the Taylor-Wiles hypothesis) that the residual representation \({\overline{\rho}}\) is irreducible when restricted to \(F(\zeta_p)\). In, Thorne has recently weakened this assumption in many cases to simply asking that \({\overline{\rho}}\) itself is irreducible but making the auxiliary assumption that the quadratic subfield of \(F(\zeta_p)/F\) is totally real.\ \ Under other assumptions on \({\overline{\rho}}\), but still allowing that \({\overline{\rho}} |_{G_{F(\zeta_p)}}\) is reducible, we prove automorphy lifting results wherein we do not assume that the quadratic subfield \(K\) of \(F(\zeta_p)/F\) is totally real. The assumption on \({\overline{\rho}}\) is that there is a ("level raising") place \(v\) of \(F\) that splits in \(K\), such that the ratio of the eigenvalues of \({\overline{\rho}} ({\rm Frob}_v)\) is \(q_v\) with \(q_v\) not 1 modulo \(p\). This assumption is automatic when \(K/F\) is totally real, as exploiting the oddness of \({\overline{\rho}}\) we can take any \(v\) such that \({\overline{\rho}} ({\rm Frob}_v)\) is conjugate to the image of complex conjugation.\ The main theorem we prove is the following: The hypothesis that \(F(\zeta_p) \not \subset K(\bar{\gamma}\epsilon^{-1}) \cap K(\bar{\gamma}\epsilon)\) is equivalent to the existence of a place \(v\) of \(F\) such that \(q_v\) is not 1 modulo \(p\), \(v\) splits in \(K\), and \({\overline{\rho}} ({\rm Frob}_v)\) has eigenvalues with ratio \(q_v\). Furthermore, as remarked above, this hypothesis is automatic when \(K\) is totally real as one may, using the oddness of \({\overline{\rho}}\), take any \(v\) such that \({\overline{\rho}} ({\rm Frob}_v)\) is conjugate to \({\overline{\rho}} (c)\) and \(q_v\) is-1 modulo \(p\). Thus the theorem generalizes the main theorem of, and in fact its proof is essentially the same as. In loc. cit., the author synthesizes two methods: -- the patching method of Taylor-Wiles, which relies on using places \(v\) of \(F\) such that \(q_v\) is \(1\) modulo \(p^N\) and \({\overline{\rho}} ({\rm Frob}_v)\) has distinct eigenvalues to kill certain elements of the mod \(p\) dual Selmer, and --a method due to Khare, which proves automorphy of \(\rho\) by "\(p\)-adic approximation" using Ramakrishna places \(v\), such that \(q_v\) is not \(1\) modulo \(p\) and \(\rho_N({\rm Frob}_v)\) has eigenvalues with ratio \(q_v\), to perform the task of killing certain elements of mod \(p\) dual Selmer which cannot be killed with Taylor-Wiles primes. Here \(\rho_N\) denotes the mod \(p^N\) reduction of \(\rho\).\ The Ramakrishna places \(v\) that Thorne used were such that \(\rho_N({\rm Frob}_v)\) is conjugate to the image of complex conjugation. To ensure that these places could effectively kill the troublesome part of mod \(p\) dual Selmer, he had to impose that the quadratic subfield of \(F(\zeta_p)/F\) be totally real, which implies \(p \equiv 1 \Mod 4\).\ The central new remark of this paper is that if we assume there is a place \(v\) of \(F\) that splits in \(K\), such that the ratio of the eigenvalues of \({\overline{\rho}} ({\rm Frob}_v)\) is \(q_v\) with \(q_v\) not \(1\) modulo \(p\), then by choosing "Teichmuller liftings" one may find, for each \(N\), places \(v\) such that \(q_v\) is not \(1\) modulo \(p\), \(\rho_N({\rm Frob}_v)\) has eigenvalues with ratio \(q_v\), and the \(v\)'s perform the task of killing the troublesome elements of mod \(p\) dual Selmer which cannot be killed with Taylor-Wiles primes. We present the proof of this remark in Proposition 2.11, and it serves as a replacement for Proposition 5.20 in which needed that \(K\) be totally real. Note that this assumption we make is automatic when \(K\) is totally real, which explains why our result generalizes the main theorem of loc. cit.\ After this remark the proof of our theorem is identical to that of loc. cit. where patching arguments followed by level raising and lowing arguments are used to conclude the automorphy of \(\rho\).\ The case when there is no place \(v\) of \(F\) that splits in \(K\) such that the ratio of the eigenvalues of \({\overline{\rho}} ({\rm Frob}_v)\) is \(q_v\) with \(q_v\) not 1 modulo \(p\) cannot be addressed by the methods of nor by the modifications that we carry out in this paper. In a forthcoming work of Khare and Thorne, this case is addressed. Instead of patching and \(p\)-adic approximation, the authors use patching together with Wiles' numerical isomorphism criterion and calculation of \(\eta\)-invariants using monodromy arguments. ## Automorphy of elliptic curves We get a modest application to automorphy of elliptic curves over totally real fields. Current automorphy lifting theorems prove the modularity of elliptic curves \(E\) over \(F\) which satisfy the "Taylor-Wiles" hypothesis at some prime \(p \in \{3,5,7\}\), i.e., \(\bar{\rho}_{E,p}(G_{F(\zeta_p)})\) is absolutely irreducible for at least one \(p \in \{3,5,7\}\). Recently, in, the authors prove the modularity of all elliptic curves over real quadratic fields. Moreover, they show that over any totally real field \(F\), there are only finitely many potentially non-modular elliptic curves over \(F\), which necessarily do not satisfy the Taylor-Wiles hypothesis at any prime \(p \in \{3,5,7\}\).\ \ In, Thorne was able to cut down on this list of potentially non-modular elliptic curves by proving a new automorphy lifting theorem and applying it to the above situation of elliptic curves with \(p=5\). However, the results in that paper do not deal with the case \(p=7\) because the requirement was that the quadratic subfield of \(F(\zeta_p)/F\) be totally real (which implies that \(p\) is 1 modulo 4). In this note, we try to deal with the cases when this quadratic subfield is not totally real by examining different sorts of auxiliary primes. We use this theorem to prove: We further note that one may also deduce by similar arguments the automorphy of elliptic curves \(E\) defined over a totally real field \(F\), with a prime \(p\) such that \(E\) has no \(F\)-rational \(p\)-isogeny, \(\rho_{E,p}(G_F)\) normalizes a split Cartan subgroup of \(\GL_2(\F_p)\), and \(E\) has bad semistable reduction at a place \(v\) such that \(q_v\) is not 1 mod \(p\) and \(v\) splits in the quadratic subfield of \(F(\zeta_p)/F\). # Galois Theory ## Global Deformation Problem Let \(F\) be a totally real number field. Let \(p\) be an odd prime, and let \(E\) be a coefficient field. Fix a continuous, absolutely irreducible representation \(\bar{\rho} \colon G_F \to \GL_2(k)\) and a continuous character \(\mu: G_F \to \Sc{O}^{\times}\) lifting \(\det \bar{\rho}\). We assume \(k\) contains the eigenvalues of all elements in the image of \(\bar{\rho}\).\ \ Let \(S\) be a finite set of finite places of \(F\), containing the places dividing \(p\) and the places at which \(\bar{\rho}\) and \(\mu\) are ramified. For \(v \in S\), fix a ring \(\Lambda_v \in \CNL_{\Sc{O}}\). Define \(\Lambda = \hat{\otimes}_{v \in S} \Lambda_v\), so that \(\Lambda \in \CNL_{\Sc{O}}\).\ \ If \(v \in S\), we write \(\Sc{D}_v^{\Box} \colon \CNL_{\Lambda_v} \to \text{Sets}\) for the functor which takes \(R \in \CNL_{\Lambda_v}\) to the set of continuous homomorphisms \(r \colon G_{F_v} \to \GL_2(R)\) such that \(r \Mod \I{m}_R = \bar{\rho}|_{G_{F_v}}\) and \(\det r\) agrees with the composite \[G_{F_v} \stackrel{\mu|_{G_{F_v}}}{\longrightarrow} \Sc{O}^{\times} \longrightarrow R^{\times}.\] This functor \(\Sc{D}_v^{\Box}\) is represented by an object \(R_v^{\Box} \in \CNL_{\Lambda_v}\). ## Galois Cohomology Given a finite place \(v\) of \(F\) and a local deformation problem \(\Sc{D}_v\), we denote by \(\Sc{L}_v\) the tangent space of the local deformation problem as a subspace \(\Sc{L}_v \subset H^1(F_v,\ad^0 \bar{\rho})\). We also let \(\Sc{L}_v^{\perp} \subset H^1(F_v,\ad^0 \bar{\rho}(1))\) be the annihilator of \(\Sc{L}_v\) induced by the perfect pairing of Galois modules: \[\ad^0 \bar{\rho} \times \ad^0 \bar{\rho}(1) \to k(1), \quad (X,Y) \mapsto \tr(XY).\] Let us now take a look at the specific deformation problem we want to consider. ## Special Deformations, case \(q_v \not \equiv 1 \Mod p\) Let \(v \in S\) be a prime not dividing \(p\), and suppose that \(q_v \not \equiv 1 \Mod p\). Suppose further that \(\bar{\rho}|_{G_{F_v}}\) is unramified, and that \(\bar{\rho}(\Frob_v)\) has two distinct eigenvalues \(\alpha_v ,\beta_v \in k\) such that \(\alpha_v/\beta_v = q_v\). Let \(\Lambda_v = \Sc{O}\). We define a subfunctor \(\Sc{D}_v^{\St(\alpha_v)} \subset \Sc{D}_v^{\Box}\) directly. Let \(R \in \CNL_{\Sc{O}}\) and let \(r: G_{F_v} \to \GL_2(R)\) be an element of \(\Sc{D}_v^{\Box}\). If \(\phi_v \in G_{F_{v}}\) is a choice of geometric Frobenius, then by Hensel's lemma the characteristic polynomial of \(r(\phi_v)\) factors as \((X-A_v)(X-B_v)\), where \(A_v,B_v \in R^{\times}\) with \(A_v\) lifting \(\alpha_v\) and \(B_v\) lifting \(\beta_v\). We will say \(r \in \Sc{D}_v^{\St(\alpha_v)}\) if \(A_v = q_vB_v\) and \(I_{F_v}\) acts trivially on \((r(\phi_v)-A_v)R^2\), which is a direct summand \(R\)-submodule of \(R^2\). One checks that this condition is independent of the choice of \(\phi_v\). ## Existence of Auxiliary Primes Continue with the notation from the previous section, and assume further that \(\bar{\rho}\) is totally odd, i.e. that \(\mu(c) =-1\) for all choices of complex conjugation \(c \in G_F\). Write \(\zeta_p \in \overline{F}\) for a primitive \(p\)-th root of unity, and now fix a choice of complex conjugation \(c \in G_F\).\ \ We will assume that \(\bar{\rho}|_{G_{F(\zeta_p)}}\) is the direct sum of two distinct characters. By Clifford theory, we know that \(\bar{\rho}\) is induced from a continuous character \(\bar{\chi}: G_K \to k^{\times}\), where \(K\) is the unique quadratic subfield of \(F(\zeta_p)/F\). That is, \(\bar{\rho} \cong \Ind_{G_K}^{G_F} \bar{\chi}\). Write \(w \in G_F\) for a fixed choice of element with nontrivial image in \(\Gal(K/F)\). We can assume that, possibly after conjugation, that \(\bar{\rho}\) has the form: \[\bar{\rho}(\sigma) = \MatTwo{\bar{\chi}(\sigma)}{0}{0}{\bar{\chi}^w (\sigma)}, \quad \text{ for } \sigma \in G_K,\] \[\bar{\rho}(w) = \MatTwo{0}{\bar{\chi}(w^2)}{1}{0}.\] Now let \(\bar{\gamma} = \bar{\chi}/\bar{\chi}^w\). By assumption, \(\bar{\gamma}\) is nontrivial, even after restriction to \(G_{F(\zeta_p)}\). We have the following: From now on, we will let \(M_0 = k(\delta_{K/F})\) and \(M_1 = \Ind_{G_K}^{G_F} \bar{\gamma}\). Fix the standard basis for \(\ad^0 \bar{\rho}\): \[E = \MatTwo{0}{1}{0}{0},\quad H = \MatTwo{1}{0}{0}{-1},\quad F = \MatTwo{0}{0}{1}{0}.\] If \(M \in \ad^0 \bar{\rho}\), we write \(k_M \subset \ad^0 \bar{\rho}\) for the line that it spans.\ \ We have the following easy lemma: Let \(\Sc{S} = (\bar{\rho},\mu,S,\{\Lambda_v\}_{v \in S},\{\Sc{D}_v\}_{v \in S})\) be a global deformation problem, and let \(T \subset S\) be a subset containing \(S_p\) (the set of places above \(p\)). Suppose that for \(v \in S-T\), the local deformation problem \(\Sc{D}_v = \Sc{D}_v^{\St(\alpha_v)}\). The lemma implies we can decompose \[H^1_{\Sc{S},T}(\ad^0 \bar{\rho}(1)) = H^1_{\Sc{S},T}(M_0(1)) \oplus H^1_{\Sc{S},T}(M_1(1)).\] We now show that we can kill the \(M_1(1)\) portion of dual Selmer using the special deformation problem we defined in the previous section, and then kill the \(M_0(1)\) portion using traditional Taylor-Wiles primes. ## Killing the \(M_1(1)\) portion First, we show that lemmas 5.21, 5.22, and 5.23 of continue to hold even if \(K\) is not totally real. Indeed, the only one which requires proof is the second, since this is the only place where Thorne used this assumption. However, we will need to impose an additional restriction. We state the other two lemmas here for convenience. The following proposition is the analog of Proposition 5.20 of and is the only place where we argue differently from Thorne because of not having (in the case that \(K\) is not totally real) the luxury to choose places \(v\) such that \(\rho_N({\rm Frob}_v)\) is the image of complex conjugation under \(\rho_N\). The proof relies on the simple observation that given an element \(g\) in \(\GL_2({\mathcal O}/p^M)\), then for \(N>>0\), the element \(g^{q^N}\) has (ratio of) eigenvalues that are the Teichmüller lift of the (ratio of) the eigenvalues of the reduction of \(g\). ## Killing the \(M_0(1)\) portion Having killed the \(M_1(1)\) portion of dual Selmer, we can try and get auxiliary primes that take care of the remaining part of the group. Indeed, we have the following proposition: # The Main Theorem From this point, everything from sections 6 of carries over. We can now work towards proving the main theorem. We proceed as in the aforementioned paper, making the necessary modifications. First, some preliminary results. Using this theorem, we arrive at the main theorem. # Application to Elliptic Curves We can apply this theorem to elliptic curves. In a paper of Freitas, Le Hung, and Siksek (see ) the authors prove there are only finitely many non-automorphic elliptic curves over any given totally real field. We can use the theorem above to prove automorphy of elliptic curves \(E\) defined over totally real fields \(F\) in some cases not covered in the existing literature. This comes about by applying our theorem to the 7-adic representations arising from \(E\). Note that the assumptions of imply that the theorem in that paper can be applied only to \(p\)-adic representations of elliptic curves with \(p\) congruent to 1 modulo 4.\ \ The following theorem is, Theorems 3 and 4. We will call an elliptic curve \(p\)-bad if \(E[p]\) is an absolutely reducible \(\F_p[G_{F(\zeta_p)}]\)-module. Otherwise \(E\) is \(p\)-good. The theorem tells us the only elliptic curves \(E\) which are potentially non-modular are those which are \(p\)-bad for \(p = 3,5\), and \(7\). In, the author deals with some of these remaining cases: We can prove a similar theorem, but before doing so we recall, Proposition 9.1. We end by remarking that we can extend Theorem [\[Main\]](#Main){reference-type="ref" reference="Main"} to primes other than \(p=7\).

{'timestamp': '2016-07-19T02:04:41', 'yymm': '1607', 'arxiv_id': '1607.04750', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04750'}

# Introduction [\[sec:intro\]]{#sec:intro label="sec:intro"} speech recognition, music classification, audio indexing, and to a less visible extent biometric voice authentication, have now achieved some degree of commercial success in mass markets. A vast proportion of these applications are supported by PC platforms, cloud computing and powerful smartphones. Nevertheless, a new market is quickly emerging in the domain of the *Internet of Things* (IoT). In this market, Automatic Environmental Sound Recognition (AESR) has a significant potential to create value, e.g., for security or home safety applications. However, in this context, industrial reality imposes hard constraints on the available computing power. Indeed, IoT devices can be broadly divided into two classes: (a) devices which are perceived as "doing only one thing", thus requiring the use of low cost processors to hit a price point that users are willing to pay for what the device does; (b) embedded devices where sound recognition comes as a bolt-on to add value to an existing product, for example adding audio analytic capabilities to a consumer grade camera or to a set-top box, thus requiring the algorithm to fit into the device's existing design and price points. These two cases rule out the use of higher end processors. Generally speaking, the following features jointly define the financial cost of a processor and the level of constraint imposed on embedded computing: a) the *clock speed* is related to energy consumption; b) the *instruction set* is related to chip size and manufacturing costs, but in some cases includes special instruction sets to parallelise more operations into a single clock cycle; c) the *architecture* defines, e.g., the number of registers, number of cores, presence/absence of a Floating Point Unit (FPU), a Graphical Processing Unit (GPU) and/or a Digital Signal Processing (DSP) unit. These features affect applications in the obvious way of defining an upper limit on the number and the type of operations which can be executed in a given amount of time, thus ruling the possibility to achieve real time audio processing at the proportion of processor load allocated to the AESR application. Finally, *onboard memory size* is an important factor related to processor cost, as it affects both the computational performance, where repetitive operations can be cached to trade speed against memory, and the scalability of an algorithm, by imposing upper limits on the number of model parameters that can be stored and manipulated. If trying to work within these limitations, and given that most IoT embedded devices allow Internet connectivity, it could be argued that cloud computing can solve the computational power constraints by abstracting the computing platform and making it virtually as powerful as needed. However, a number of additional design considerations may rule out the use of cloud computing for AESR applications: the *latency* introduced by cloud communications can be a problem for, e.g., time critical security applications; regarding *Quality of Service (QoS)*, network interruptions introduce an extra point of failure into the system; regarding *privacy* and *bandwidth consumption*, respectively, sending alerts rather than streaming audio or acoustic features out of the sound recognition device both rules out any possibility of eavesdropping  and requires less bandwidth. Thus, the reality of embedded industrial applications is that at the price points acceptable in the marketplace, IoT devices will most likely be devoid of an FPU, will operate in the hundreds of Megahertz clock speed range, won't offer onboard DSP or specialised instruction sets, and may prefer to run AESR on board rather than in the cloud. At the same time, most AESR performance evaluations are still produced with limited interest for the computational cost involved: many experimental results are obtained with floating point arithmetics on powerful computing platforms, and under the assumption that sufficient computing power will sooner or later become available for the algorithm to be practicable in the context of commercial applications. This article aims at crossing this divide, by comparing the performance of three AESR algorithms as a function of their computational cost. In other terms, it seeks which AESR algorithm is able to make the most of a limited amount of computational power. The rest of the paper is organised as follows: Section [\[sub:prior\]](#sub:prior){reference-type="ref" reference="sub:prior"} reviews prior art in AESR. Section [\[sec:cost\]](#sec:cost){reference-type="ref" reference="sec:cost"} defines the measure of computational cost used in this study, and applies it to a range of popular machine learning algorithms. Section [\[sec:perf\]](#sec:perf){reference-type="ref" reference="sec:perf"} defines the other axis of this evaluation, which is the metric used to measure classification performance. The data sets and methodology are detailed in Section [\[sec:methodology\]](#sec:methodology){reference-type="ref" reference="sec:methodology"}, while results are discussed in Section [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}. # Background [\[sub:prior\]]{#sub:prior label="sub:prior"} Audio or sound recognition research has primarily been focused on speech and music signals. However in the past decade, the problem of classifying and identifying environmental sounds has received more attention. AESR has applications in content-based search, robotics , security and a host of other tasks. Recognition of environmental sounds is a part of the much broader field of computational auditory scene analysis (CASA). We are interested in the problem of identifying the presence of non-speech, non-music sounds or *environmental sounds* in a given audio recording. Environmental sound recognition involves the study of two closely related problems. Acoustic Scene Classification (ASC) involves selecting an appropriate set of labels that encompasses all the sound events which occur in a recording, e.g., "crowded train station". With regards to machine learning, ASC can be seen as a sequence classification task where a set of labels is assigned to a variable length input sequence. A related problem is the problem of Acoustic Event Detection (AED). In addition to identifying the correct label, AED aims to segment the audio recording into regions where the label occurs, e.g., "stationmaster blowing whistle" or "doors closing". Again with regards to machine learning, AED involves identifying the correct labels and the correct *alignment* of labels with the recording. In this study, we compare Gaussian mixture models (GMMs), support vector machines (SVMs) and deep neural networks (DNNs) in terms of their performance on an AED task as a function of their computational cost. GMMs are generative probabilistic models and have been extensively applied to many problems in speech recognition . In a number of cases, GMMs have therefore been considered a sensible starting point for research on the classification of environmental sounds. For example, they have been used to detect gunshot sounds  and to discriminate between gunshots and screams . Portêlo et al.  have used Gaussian Mixtures as part of a hidden Markov model (HMM) approach to detecting sirens in audio. Elsewhere, Radhakrishnan et al.  have used GMMs when attempting to develop a model for recognising new 'outlier' sounds against a GMM trained to model background sounds. Principi et al.  proposed another novelty detection system, with an embedded implementation on a Beagle Board . Atrey et al.  have used GMMs to classify audio frames into foreground or background sounds, then into vocal or nonvocal sounds and finally into 'excited' events (e.g. shouting, running) or 'normal' events (e.g. talking, walking). SVMs are non-parametric classifiers that can generalise well given a limited training set and have well defined generalisation bounds . Portêlo et al. have experimented with an SVM when attempting to detect events in film audio. They found it to perform well on some sounds (aeroplanes, helicopters) but poorly on others (sirens, cats). By adding new features and applying PCA they found better performance on siren sounds. Moncrieff et al. have also experimented with SVMs when attempting to classify sound events in film audio. Chen et al. have used SVMs to classify audio data into music, speech, environmental sounds, speech mixed with music and environmental sounds mixed with music. Rouas et al. have presented a comparison of SVMs and GMMs when attempting to detect shouting in public transport vehicles. Testing on audio recorded by actors on a train, they find that the SVM achieves a lower false alarm rate whilst the GMM seems better at making positive identifications. More recently, Deep Learning, i.e., artificial neural networks with more than one intermediate hidden layer, have gained much popularity and have achieved impressive results on several machine learning tasks . However, they do not seem to have been extensively applied to AESR so far. Toyoda et al.  have used a neural network to classify up to 45 different environmental sounds. They achieved high recognition accuracies, though it should be mentioned that both the training and test data sets used in their study may appear quite small in comparison to the data sets used nowadays as benchmarks for speech recognition and computer vision research. Baritelli et al.  used a DNN to classify 10 different environmental sounds. The authors found that the proposed system, which classified MFCCs directly, performed similarly to other AESR systems which were requiring more complex feature extraction. In , the authors used neural networks to estimate the presence or absence of people in a given recording, as well as the number of people and their proximity to the recording device. Gencoglu et al. applied a deep neural network to an acoustic event detection task for classifying \(61\) different classes of sounds. Their results showed this classifier to outperform a conventional GMM+HMM classification system. Recently, the problem of AESR has been receiving more attention. Challenges like CLEAR and DCASE have been proposed in order to benchmark the performance of different approaches and to stimulate research. But, if drawing a parallel between the evolution of AESR and the evolution of speech recognition, the CLEAR and DCASE data sets could be thought of as being at the stage of the TIMIT era: although the TIMIT data set was widely used for many years, it was limited in size to about 5.4 hours of speech data. The subsequent growth in the volume of available speech evaluation data sets has significantly contributed to bringing speech recognition to the levels of performance known today, where modern speech recognition systems are being trained over 1700 hours of data. AED research is following a similar trend where the domain is now facing a need to collect larger data sets in order to make larger scale evaluations possible and to foster further technical progress. Many studies explore feature learning to derive useful acoustic features that can be used to classify acoustic events. A limited number of publications apply k-Nearest Neighbours (k-NNs) or Random Forests to AESR. However, these models have not been included in our comparative study, which is restricted to GMM, SVM and DNN models: these three types of models can be considered mainstream in terms of having been applied more extensively to a larger range of audio classification tasks which include speech, speaker and music classification and audio event detection. In preliminary experiments, we observed that k-NN classifiers were clearly outperformed by the classifiers included in this study. # Computational cost [\[sec:cost\]]{#sec:cost label="sec:cost"} ## Measure of computational cost used in this article The fundamental question studied in this paper is that of comparing various machine learning algorithms *as a function of their computational cost*. A corollary to this question is to compare the rate of performance degradation that occurs when down-scaling various AESR algorithms by a comparable amount of computing operations: a given class of algorithms might be more resilient to down-scaling than another. While computational cost is related to computational complexity, both notions are distinct: computational cost simply counts the number of operations at a given model dimension, whereas computational complexity expresses the mathematical law according to which the computational cost scales up with the dimension of the input feature space. Computational cost is studied here at the sound recognition stage, known as acoustic decoding. Unless specifically required by the application, it is indeed fairly rare to see the training stage of machine learning being realised directly on-board an embedded device. In most cases, it is possible to design the training as an offline process realised on powerful scientific computing platforms equipped with FPUs and GPUs. Offline training delivers a floating point model which is then quantised according to the needs of fixed point decoding in the embedded system. It is generally accepted that the quantisation error introduced by this operation is unlikely to affect the acoustic modelling performance. The computational cost estimate used in this study simply counts four types of basic operations: addition, comparison, multiplication, and lookup table retrieval (LUT). *Multiply-add* operations are commonly found in dot products, matrix multiplications and FFTs. Correlation operations, linear filtering or the Mahalanobis distance, at the heart of many recognition algorithms, rely solely on multiply-adds. Fixed point precision is fairly straightforward to manage over this type of operation. *Divisions*, on the other hand, and in particular matrix inversions, might be more difficult to manage. Regarding matrix inversion, it can happen that quantisation errors add up and make the matrix inversion algorithm unstable. In many cases, it is possible to pre-compute the inversion of key parameters (e.g., variances), possibly offline and in floating point before applying fixed-point quantisation. *Non-linear operations* such as logarithms, exponentials, cosines, \(N^{th}\) roots etc. are required by some algorithms. Two approaches can commonly be taken to transform non-linear functions into a series of multiply-adds: either Taylor series expansion, or look-up tables (LUTs). LUTs are more memory consuming than Taylor series expansions, but may require a lower number of instructions to achieve the desired precision. Generally speaking, it is important to be aware of the presence of a non-linearity in a particular algorithm, before assessing its cost: algorithms which rely on a majority of multiply-adds are more desirable than algorithms relying more heavily on non-linearities. For all the estimates presented below, it is assumed that these four types of operations have an equal cost. While this assumption would be true for a majority of processors on the market, it may underestimate the cost of non-linear functions if interpolated LUTs or Taylor series are used instead of simpler direct LUT lookups. In addition to the 4 operations considered here, there are additional costs incurred with other operations (like data handling) in the processor. However, we use the simplifying assumption that the considered operations are running on the same core, thus minimizing data handling overhead. As a case study, let us assume an attempt at running the K-nearest neighbours algorithm on a Cortex-M4 processor running at a clock speed of 80MHz and with 256kB of onboard memory. This is a very common hardware configuration for consumer electronic devices such as video cameras. Assuming that 20% of the processor activity is needed for general system tasks means that a maximum of 64 million of multiply or adds per second can be used for sound recognition (\(80\mbox{MHz} \times 80\% = 64\mbox{MHz}\)). Assuming sound sampled at 16kHz and audio feature extracted at, e.g., a window shift of 256 samples, equivalent to 62.5 frames per second, means that the AESR algorithm should use no more than 1,024,000 multiply or adds per analysis frame (64MHz / 62.5Hz = 1 024K instructions). From there, assuming a K-Nearest neighbour algorithm with a 40 dimensional observation vector \(x\), mean vector \(\mu\) and a Mahalanobis distance \(( x-\mu ) \sigma^{-1} (x-\mu)\) with a diagonal covariance matrix \(\sigma\), there would be one subtraction and two multiplications per dimension for the Mahalanobis distance calculation, plus one operation for the accumulation across each dimension: four operations in total, thus entailing a maximum of 1,024,000 instructions / 4 operations / 40 dimensions = 6 400 nearest neighbours on this platform to achieve real-time recognition. However, assuming that each nearest neighbour is a 40-dimensional vector of 32 bits/2 Bytes values, the memory requirement to store the model would be 512kB, which is double the 256kB available on the considered platform. Again assuming that the code size occupies 20% of the memory, leaving 80% of the memory for model storage, a maximum of \(256\mbox{kB} \times 80\% \,/\, 40\,\mbox{features} \,/\, 2\mbox{Bytes} = 2\,560\) nearest neighbours only which could be used as the acoustic model. Depending on the choice of algorithm, memory is not always the limiting factor. In the case of highly non-linear algorithms, the computational cost of achieving real-time audio recognition might outweigh the model storage requirements. While only the final integration can tell if the desired computational load and precision are met, estimates can be obtained as illustrated in this case study in order to forecast if an algorithm will fit on a particular platform. ::: ## Computational cost estimates for the compared algorithms {#sub:cost} Most AED systems use a common 3-step pipeline. The audio recording is first converted into a time-frequency representation, usually by applying the short-time Fourier transform (STFT) to overlapping windows. This is followed by feature extraction from the time-frequency representation. Typically, Mel Frequency Cepstral Coefficients (MFCCs) have been used as standard acoustic features since the seventies, due to favourable properties when it comes to computing distances between sounds. Then the acoustic features are input into an acoustic model to obtain a classification score, which in some cases is analogous to a posterior probability. The classification score is then thresholded in order to obtain a classification decision. Below, we provide an estimate of the computational cost for each of these operations. **Feature extraction-** Engineering the right feature space for AESR is essential, since the definition of the feature space affects the separability of the classes of acoustic data. Feature extraction from sound recordings thus forms an additional step in the classification pipeline, which contributes to the overall computational cost. In this study, the GMMs, SVMs and DNNs are trained on a set of input features that are typically used for audio and speech processing (Section [\[sub:feature_extraction\]](#sub:feature_extraction){reference-type="ref" reference="sub:feature_extraction"}). Although recent studies demonstrate that neural networks can be trained to jointly learn the features and the classifier, we have found that this method is impractical for most AESR problems where the amount of labelled data for training and testing is limited. Additionally, by training the different algorithms on the same set of features, we study the performance of the various classifiers as a function of computation cost, without having to account for the cost of extracting each type of feature from the raw audio waveform[^1]. It also allows a fair comparison of the discriminative properties of different classification algorithms on the same input feature space. Therefore, we factor out the computational cost of feature extraction. **Gaussian Mixture Models-** Given a \(D\) dimensional feature vector \(x\), a GMM  is a weighted sum of Gaussian component densities which provides an estimate of the likelihood of \(x\) being generated by the probability distribution defined by a set of parameters \(\theta\): \[p(x|\theta) = \sum\limits_{i=0}^{M-1} w_{i} \cdot g(x | \mu_{i}, \Sigma_{i} ) \label{eq:gauss1}\] where: \[g(x | \mu_{i}, \Sigma_{i} ) = \frac{1}{(2\pi)^{\frac{D}{2}} |\Sigma_{i}|^{\frac{1}{2}}} \cdot e^{-\frac{1}{2} (x-\mu_i)' \Sigma_{i}^{-1} (x-\mu_i)} \label{eq:gauss2}\] where \(\theta = (\mu_{i}, \Sigma_{i}, w_{i})_{0 \leq i < M}\) is the parameter set, \(M\) is the number of Gaussian components, \(\mu_{i}\) are the means, \(\Sigma_{i}\) the covariance matrices and \(w_{i}\) the weights of each Gaussian component. In practice, computing the likelihood in the log domain avoids numerical underflow. Furthermore, the covariance matrices are constrained to be diagonal, thus transforming equations ([\[eq:gauss1\]](#eq:gauss1){reference-type="ref" reference="eq:gauss1"}) and ([\[eq:gauss2\]](#eq:gauss2){reference-type="ref" reference="eq:gauss2"}) into: \[\begin{array}{l} \log p(x|\theta) = \\ \;\; \textrm{logsum}_{i=0}^{M-1} \left\{ \sum_{d=0}^{D-1} \left( x_d-\mu_{i,d} \right)^2 \cdot \sigma_{i,d}^{-1} + w_{i} \right\} + K \end{array} \label{eq:gauss3}\] where \(d\) is the dimension index, constant \(K\) can be neglected for classification purposes, and \(\textrm{logsum}\) symbolises a recursive version of function \(\log(a+b) = \log a + \log(1 + e^{\left( \log b-\log a\right)})\), which can be computed using a LUT on non-linearity \(\log(1 + e^x)\). Calculating the log likelihood of a \(D\)-dimensional vector given a GMM with \(M\) components therefore requires the following operations: - \(D\) subtractions (i.e., additions of negated means) and \(2D\) multiplications per Gaussian to calculate the argument of the exponent. - \(1\) extra addition per Gaussian to apply the weight in the log domain. - \(M\) LUT lookups and \(M\) additions for the non-linear \(\textrm{logsum}\) cumulation across the Gaussian components. This leads to the computational cost per audio frame expressed in the first row of table [\[num_ops\]](#num_ops){reference-type="ref" reference="num_ops"}. **Support Vector Machines-** [\[SVMs\]]{#SVMs label="SVMs"} Support Vector Machines (SVMs) are discriminative classifiers. Given a set of data points belonging to two different classes, the SVM determines the optimal separating hyperplane between the two classes of data. In the linearly separable case, this is achieved by maximising the margin between two hyperplanes that pass through a number of *support vectors*. The optimal separating hyperplane is defined by all points \(\textrm{x}\) that satisfy: \[\textrm{x} \cdot \textrm{w} + b = 0\] where \(\textrm{w}\) is a normal vector to the hyperplane and \(\frac{|b|}{\|w\|}\) is the perpendicular distance from the hyperplane to the origin. Given that all data points \(\textrm{x}_{i}\) satisfy: \[\left\{ \begin{array}{ll} \textrm{x}_{i} \cdot \textrm{w} + b \geq 1 & \mbox{for labels \(y_{i} = +1\)} \\ \textrm{x}_{i} \cdot \textrm{w} + b \leq-1 & \mbox{for labels \(y_{i} =-1\)} \end{array} \right.\] It can be shown that the maximum margin is defined by minimising \(\frac{\|\textrm{w}\|^2}{2}\). This can be solved using a Lagrangian formulation of the problem, thus producing the multipliers \(\alpha_{i}\) and the decision function: \[f(\textrm{x}) = \textrm{sgn} \Big( \sum\limits_{i=0}^{N-1} y_{i} \alpha_{i} \textrm{x} \cdot \textrm{x}_i + b \Big)\] where \(N\) is the number of training examples and \(\textrm{x}\) is a feature vector we wish to classify. In practice, most of the \(\alpha_{i}\) will be zero, and the \(\textrm{x}_i\) for non-zero \(\alpha_{i}\) are called the *support vectors* of the algorithm. In the case where the data is not linearly separable, a non-linear kernel function \(K(\textrm{x}_{i},\textrm{x}_{j})\) can be used to replace the dot products \(\textrm{x} \cdot \textrm{x}_i\), with the effect of projecting the data into a higher dimensional space where it could potentially become linearly separable. The decision function then becomes: \[\label{SVM_classification} f(\textrm{x}) = \textrm{sgn} \Big( \sum\limits_{i=0}^{N-1} y_{i} \alpha_{i} K(\textrm{x},\textrm{x}_{i}) + b \Big)\] Commonly used kernel functions include: ---------------------------------------------------------------- ------------------------------------------------------------------------------- \(\vcenter{\hbox{\tiny\)\bullet\(}}\) Linear: \(K(\textrm{x},\textrm{x}_{i}) = \textrm{x} \cdot \textrm{x}_{i}\) \(\vcenter{\hbox{\tiny\)\bullet\(}}\) Polynomial: \(K(\textrm{x},\textrm{x}_{i}) = (\textrm{x} \cdot \textrm{x}_{i})^d\) \(\vcenter{\hbox{\tiny\)\bullet\(}}\) Radial basis function (RBF): \(K(\textrm{x},\textrm{x}_{i}) = \exp(-\gamma | \textrm{x}-\textrm{x}_{i}|^2)\) \(\vcenter{\hbox{\tiny\)\bullet\(}}\) Sigmoid: \(K(\textrm{x},\textrm{x}_{i}) = \tanh(\textrm{x} \cdot \textrm{x}_{i})\) ---------------------------------------------------------------- ------------------------------------------------------------------------------- A further refinement to the SVM algorithm makes use of a *soft margin* whereby a hyperplane can still be found even if the data is non-separable (perhaps due to mislabeled examples). The modified objective function is defined as follows: \[\label{soft_margin} \arg\min_{\textrm{w},\mathbf{ \sum_{i=1}^n\] subject to: \(y_i(\mathbf{w}\cdot\mathbf{x_i}-b) \ge 1- \vspace{0.5em} \noindent\) finds the best hyperplane that fits the data while minimising the error of misclassified data points. The importance of these error terms is determined by parameter \(C\), which can control the tendency of the algorithm to over fit or under fit the data. Given a support vector machine with \(\lambda\) support vectors, a new \(D\)-dimensional example can be classified using \(\lambda\) additions to sum the dot product results for each support vector, plus a further addition for the bias term. For each dot product, a support vector requires the computation of the kernel function, then \(\lambda\) multiplications to apply the \(\alpha_i\) multipliers. Some kernels require a number of elementary operations, whereas others require a LUT lookup. Thus, depending on the kernel, the final costs are: - SVM Linear--\((\lambda \cdot D) + \lambda + 1\) additions and \((\lambda \cdot D) + \lambda\) multiplications. - Polynomial--\(\lambda(D + d + 2) + 1\) additions and \(\lambda(D+2)\) multiplications, where \(d\) is the degree of the polynomial. - Radial Basis Function--\(2\lambda D + \lambda + 1\) additions, \(\lambda(D+2)\) multiplications and \(\lambda\) exponential functions. - Sigmoid--\(\lambda(D+2) + 1\) additions and \(\lambda(D+2)\) multiplications and \(\lambda\) hyperbolic tangent functions. The computational cost of SVMs per audio frame for these kernels is summarized in Table [\[num_ops\]](#num_ops){reference-type="ref" reference="num_ops"}. **Deep Neural Networks-** Two types of neural network architectures are compared in this study: Feed-Forward Deep Neural Networks (DNNs) and Recurrent Neural Networks (RNNs). A *Feed-Forward Deep Neural Network* processes the input by a series of non-linear transformations given by: \[\label{ffNN} {h}^i = f \left( {W}^i {h}^{i-1}+{b}^i \right) \; \mbox{for} \; 0 \leq i \leq L\] where \({h}^0\) corresponds to the input \(x\), \({W}^i,{b}^i\) are the weight matrix and bias vector for the \(i^{th}\) layer, and the output of the final layer \({h}^L\) is the desired output. The non-linearity \(f\) in Equation [\[ffNN\]](#ffNN){reference-type="ref" reference="ffNN"} is usually a sigmoid \(f(x) = 1/\left( 1 + e^{-x} \right)\) or a hyperbolic tangent function \(\tanh(x)\). However recent research in various domains has showed that the Rectified Linear Unit (ReLU)  non-linearity can lead to faster convergence of models during training. The ReLU function is defined as \(f(x) = max(0,x)\), which is simpler than the sigmoid or \(\tanh\) activation because it only requires a piecewise linear operator instead of a fixed point sigmoid or \(\tanh\) LUT. While feed forward networks are designed for stationary data, *Recurrent Neural Networks* (RNNs) are a class of neural networks designed for modelling temporal data. RNNs predict the output \(y_t\) given an input \(x_t\), at some time \(t\) in the following way: \[\label{RNN} {h}^i_t = f \left( {W}^i_f {h}^{i-1}_t+{W}^i_r {h}^{i}_{t-1}+{b}^i \right) \; \mbox{for} \; 0 \leq i \leq L\] where \({W}^i_f,{W}^i_r\) are the forward and recurrent weight matrices for the \(i^{th}\) recurrent layer, \(b_i\) is bias for layer \(i\), and \(\theta = \left\{{W}^i_f,{W}^i_r,{b}^i \right\}_0^L\) are the model parameters. The input layer \(h^{0}_{t}\) corresponds to the input \(x_t\) and the output layer, \(h^{L}_{t}\) corresponds to the output \(y_t\). In Equation [\[RNN\]](#RNN){reference-type="ref" reference="RNN"}, \(f\) is the hyperbolic tangent function. The choice of \(f\) for the output layer depends on the particular modelling problem. RNNs are characterised by the recurrent connections \({W}^i_r\) between the hidden units. Similar to feed forward neural networks, RNNs can have several recurrent layers stacked on top of each other to learn more complex functions . Due to the recursive structure of the hidden layer, at any time \(t\), the RNN makes a prediction conditioned on the entire sequence seen until time \(t\): \(y_t = f(x_0^t)\). For simplicity, every neural network in this study was constrained to have the same number \(H\) of hidden units in each of its \(L\) layers. The input dimensionality is denoted by \(D\). The forward pass through a feed-forward neural network therefore involves the following computations: - Multiplying a column vector of size \(n\) with a matrix of size \(n \times k\) involves \(n \cdot k\) additions and an equal number of multiplication operations. - Therefore the first layer involves \(\left( D \cdot H \right)\) multiplication operations and \(H \cdot \left( D +1 \right)\) additions, where the additional \(H\) additions are due to the bias term. - The remaining \(L-1\) layers involve the following computations: \(\left( L-1 \right) \cdot \left( H \cdot H\right)\) multiplications and \(\left( L-1 \right) \cdot \left( H \cdot (H+1) \right)\) additions. - The output layer involves \(H\) multiplications and \(H+1\) additions. - A non-linearity is applied to the outputs of each layer, leading to a total of \(L \cdot H+1\) non-linearities. - The RNN forward pass includes an additional matrix multiplication due to the recurrent connections. This involves \(H \cdot H\) multiplications and an equal number of addition operations. The computational cost estimates for DNNs and RNNs are summarised in Table [\[num_ops\]](#num_ops){reference-type="ref" reference="num_ops"}. # Performance metrics for sound recognition [\[sec:perf\]]{#sec:perf label="sec:perf"} Classification systems produce two types of errors: False Negatives (FN), a.k.a. Missed Detections (MD), i.e., target sounds which have not been detected, and False Positives (FP), a.k.a. False Alarms (FA), i.e., non-target sounds which have been mistakenly classified as the target sound . Their correlates are the True Positives (TP), or correctly detected target sounds, and True Negatives (TN), or correctly rejected non-target sounds. In our experiments, "sounds" refer to feature frames associated with a particular sound class label. The above metrics are usually expressed as percentages of the available testing samples. They can be computed *after an operating point or threshold* has been selected for a given classifier. Indeed, most classifiers output a real valued *score*, which can be, e.g., a log-likelihood ratio, a posterior probability or a distance. The binary target/non-target classification is then made by comparing the score to a threshold. Thus, a system can be made either more permissive, by choosing a threshold which lets through more True Positives at the expense of generating more False Positives, or the system can be made more conservative, if the threshold is set to filter out more False Positives at the expense of rejecting more sounds in general and thus generating more Missed Detections. For example, the operation point of a GMM classifier is defined by choosing a log-likelihood threshold. Ideally, AESR systems should be compared independently of the choice of threshold. This can be achieved through the use of Detection Error Tradeoff (DET) curves , which depict the FA/MD compromises achieved across a sweep of operation points. DET curves are essentially equivalent to Receiver Operating Curves (ROC) , but plotted into normal deviate scale to linearise the curve, under the assumption that FA and MD scores are normally distributed. Given that the goal is to minimise the FA/MD compromises globally across all possible operation points, a DET curve closer to the origin means a better classifier. DET curve performance can thus be summarised for convenience into a single figure called the Equal Error Rate (EER), which is the point where the DET curve crosses the %FA = %MD diagonal line. In other terms, the EER represents the operation point where the system generates an equal rate of FA and MD errors. For GMM based classification systems, the output score is the ratio between the likelihood of the target GMM and the likelihood of the universal background model (UBM): \(P(x|\theta_{\mbox{\tiny target sound}})/P(x|\theta_{\mbox{\tiny UBM}})\). For neural networks, the output layer consists in a single unit followed by a sigmoid function, thus yielding a real-valued class membership probability \(P(C=1|x)\) taken as the score value. For SVMs, the score can be defined as the argument of the \(\textrm{sgn}()\) function in Equation [\[SVM_classification\]](#SVM_classification){reference-type="ref" reference="SVM_classification"}: whereas the \(\textrm{sgn}()\) function always amounts to classifying against a threshold fixed to zero, using its argument as a score suggests a more gradual notion of deviation from the margin defined by the support vectors, or in other terms it suggests a more gradual measure of belonging to one side or the other of the margin. # Data sets and methodology [\[sec:methodology\]]{#sec:methodology label="sec:methodology"} ## Data sets For the problem of AED, the training data set consists of audio recordings along with corresponding annotations or ground truth labels. The ground truth labels are usually provided in the form of alignments, i.e. onset times, offset times and the audio event labels. This is in contrast to ASC problems where the ground truth does not include alignments. ASC is similar to the problem of speech recognition where each recording has an associated word-level transcription. The alignments are then obtained as a by-product of the speech recognition model. For the problem of AED, annotating large quantities of data with onset and offset times requires significant effort from human annotators. Historically, relatively small data sets of labelled audio have been used for evaluating AED systems. As a matter of fact, the two most popular benchmarks for AED algorithms known to date are the DCASE challenge and the CLEAR challenge. The DCASE challenge data provides 20 examples for each of the 16 event classes. While the data set for the CLEAR challenge contains approximately 60 training instances for each of the 13 classes. The typical approach to AED so far has been to extract features from the audio and then use a classifier or an *acoustic model* to classify the frames into event labels. Additionally, the temporal relationship between acoustic frames and outputs is modelled using HMMs. In this study, we investigate whether a data-driven approach to AED can yield improved performance. Inspired by the recent success of Deep Learning, we investigate whether neural network acoustic models can outperform existing models when trained on sufficiently large data sets. Deep Neural Networks (DNNs) have recently contributed to significant progress in the fields of computer vision, speech recognition, natural language processing and other domains of machine learning. The superior performance of neural networks can be largely attributed to more computational power and the availability of large quantities of labelled data. Neural network models with millions of parameters can now be trained on distributed platforms to achieve good generalisation performance. However, despite the strong motivation to use neural networks for various tasks, their application is limited by the availability of labelled data. For example, we trained many DNN architectures on the DCASE data, but were unable to beat the performance of a baseline GMM system. In order to train large neural network models, we perform experiments using three private data sets made available by Audio Analytic Ltd.[^2] to support this study. These three data sets are subsets of much larger data collection campaigns led by Audio Analytic Ltd. to support the development of smoke alarm and baby cry detection products. Ground truth annotations with onset and offset times obtained from human annotators were also provided. The motivation for using these data sets is twofold. Firstly, to provide a large number of training and test examples for training models that recognize baby cries and smoke alarm sounds. Secondly, a very large *world* data set of ambient sounds is used as a source of impostor data for training and testing, in a way which emulates the quasi infinite prior probability of the system being exposed to non-target sounds instead of target ones. In other words, there are potentially thousands of impostor sounds which the system must reject in addition to identifying the target sounds correctly. While the DCASE challenge evaluates how an AED system performs at detecting one sound in the presence of 15 impostors or non-target sounds, the world data set allows to evaluate the performance of detecting smoke alarms or baby cries against the presence of over 1000 impostor sounds, in order to reflect a more realistic use case. The *Baby Cry* data subset comprises a proportion of baby cries obtained from two different recording conditions: one through camcorder in a hospital environment, one through uncontrolled hand-held recorders in uncontrolled conditions (both indoors and outdoors). Each recording was from a different baby. The train/test split was done evenly, and set to achieve the same balance of both conditions in both the train and test set (about 3/4 hospital condition and 1/4 uncontrolled condition) without any overlap between sources for training and testing. The recordings in this data set sum up to \(4\,822\) seconds of training data and \(3\,669\) seconds of test data, from which \(224\,076\) feature frames in the training set and \(90\,958\) feature frames in the test set correspond to target baby cry sounds. The *Smoke Alarm* data subset comprises recordings of 10 smoke alarm models, recorded through 13 different channels across 3 British homes. While it may be thought that smoke alarms are simple sounds, the variability associated with different tones, different audio patterns, different room responses and different channel responses does require some generalisation capability which either may not be available in simpler fingerprinting or template matching methods, or would require a very large database of templates. The training set comprises recordings from two of the homes, while the test set is composed of recordings from the remaining home. Thus, while the same device can appear in both the training and evaluation sets, the recording conditions are always different. So the results presented over the smoke alarm set focus mainly on the generalisation across room responses and recording conditions. The data set provided for this study consists of \(15\,271\) seconds of training data and \(5\,043\) seconds of testing data, of which \(194\,142\) feature frames in the training set and \(114\,753\) feature frames in the test set correspond to target smoke alarm sounds. The *World* data set contains about 900 files covering 10 second samples of a wide variety of complex acoustic scenes recorded from numerous indoor and outdoor locations around the UK, across a wide and uncontrolled range of devices, and potentially covering more than a thousand sound classes. For example, a 10 seconds sample of "train station" scene from the world data set covers, e.g., train noise, speech, departure whistle and more, while a 10 second sample of "supermarket" covers babble noise, till beeps and children's voices. While it does not seem feasible to enumerate all the specific classes from the World set, assuming that there are more than two classes per file across the 410 files reserved for testing leads to an estimate of more than a thousand classes in the test set. The 500 other files of the world set are used to train a non-target model, i.e., a Universal Background Model (UBM)  in the case of GMMs, or a negative input for neural network and SVM training. It was ensured that the audio files in the both the training and testing world sets were coming from completely disjoint recording sessions. The World data set provided for this study consists of \(5\,000\) seconds of training data and \(4\,089\) seconds of test data, thus amounting to \(312\,500\) feature frames for training and \(255\,562\) feature frames for testing. Where necessary, the recordings were converted to 16kHz, 16 bits using a high quality resampling algorithm. No other pre-processing was applied to the waveform before feature extraction. ## Feature Extraction [\[sub:feature_extraction\]]{#sub:feature_extraction label="sub:feature_extraction"} For this experiment, Mel-frequency cepstral coefficient (MFCC) features were extracted from the audio. MFCC features are extensively used in speech recognition and in environmental sound recognition [@clavel2005events; @valenzise2007scream; @portelo2009non; @radhakrishnan2005audio]. In addition to MFCCs, the spectral centroid, spectral flatness, spectral rolloff, spectral kurtosis and zero crossing rate were also computed. With audio sampled at 16kHz, all the features were calculated with a window size of \(512\) samples and a hop size of \(256\) samples. The first \(13\) MFCCs were used for our experiments. Concatenating all the features thus lead to an \(18\) dimensional feature vector of real values. Temporal information was included in the input representation by appending the first and second order differences of each feature. This led to an input representation with \(18 \times 3 = 54\) dimensions. It should be noted that the temporal difference features were used as inputs to all models except the RNNs, where only the \(18\) dimensional features were used as inputs. This is due to the fact that the RNN is designed to make predictions conditioned on the entire sequence history (Equation [\[RNN\]](#RNN){reference-type="ref" reference="RNN"}). Therefore, explicitly adding temporal information to the inputs is unnecessary. The reduction in input dimensionality and related reduction in computational cost is offset by the extra weight matrix in the RNN (Table [\[num_ops\]](#num_ops){reference-type="ref" reference="num_ops"}). For all the experiments, the data was normalised to have zero mean and unit standard deviation, with the mean and standard deviation calculated over the entire training data independently for each feature dimension. ## Training Methodology [\[training\]]{#training label="training"} **Gaussian Mixture Models-** GMMs were trained using the expectation maximisation algorithm (EM). The covariance matrices were constrained to be diagonal[^3]. A grid search was performed to find the optimal number of Gaussians in the mixture, on the training set. The best model with number of components \(M \in \{ 1,2,4,8,16,32,64,128,256,512,1024 \}\) was used for evaluation. A validation set was created by randomly selecting 20% of the training data, while the remaining 80% of the data was used for parameter estimation. The model that performed the best on the validation set was later used for testing. All the Gaussians were initialised using the k-means++ algorithm. A single UBM was trained on the World training data set. One GMM per target class was then trained independently of the UBM. We also performed experiments where GMMs for each class were estimated by adapting the UBM by Maximum A Posteriori (MAP) adaptation. We observed that this method yielded inferior results. We also observed worse results when no UBM was used and class membership was determined by thresholding the outputs of individual class GMMs. In the following, results are therefore only reported for the best performing method, which is the likelihood ratio between independently trained UBM and target GMMs. **Support Vector Machines-** The linear, polynomial, radial basis function (RBF) and sigmoid kernels were compared in our experiments. In addition to optimising the kernel parameters (\(d,\gamma\)), a grid search was performed over the penalty parameter \(C\) (Equation [\[soft_margin\]](#soft_margin){reference-type="ref" reference="soft_margin"}). In practice, real-world data are often non-separable, thus a soft margin has been used as necessary. Model performance was estimated by evaluating performance on a validation set. The validation set was formed by randomly selecting 20% of the training data, while the remaining data was used for parameter estimation. Again, the best performing model on the validation set was used for evaluation. SVM training is known not to scale well to very large data sets : it has a time-complexity of \(O(T^3)\), where \(T\) is the number of training examples. As a matter of fact, our preliminary attempts at using the full training data sets led to prohibitive training times. Standard practice in that case is to down-sample the training set by randomly drawing \(T\) frames from the target set, and the same number \(T\) of frames from the world set as negative examples for training. Results presented in section [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"} with \(T \in \left\{ 500,2000 \right\}\) show that contrary to the intuitive thought that down-sampling might have disadvantaged the SVMs by reducing the amount of training data compared to the other machines, such data reduction actually improves SVM performance. As such, down-sampling can be thought of as an integral part of the training process, rather than as a reduction of the training data. **Deep Neural Networks-** All neural network architectures were trained to output a distribution over the presence or absence of a class, using the backpropagation algorithm and stochastic gradient descent (SGD). The training data was divided into a \(80/20\) training/validation split to find the optimum training hyper-parameters. All the target frames were used for each class, and an equal number of non-target frames was sampled randomly from the World data set, for both training and validation. For the DNNs, a grid search was performed over the following parameters: number of hidden layers \(L \in \left\{ 1,2,3,4 \right\}\), number of hidden units per layer \(H \in \left\{ 10,25,50,100,150 \right\}\), hidden activations \(act \in \left\{ sigmoid, ReLU \right\}\). In order to minimise parameter tuning, we used ADADELTA to adapt the learning rate over iterations. The networks were trained using mini-batches of size \(100\) and training was accelerated using an NVIDIA Tesla K40c GPU. A constant dropout rate of \(0.2\) was used for all layers. The training was stopped if the cost on the validation set did not decrease after \(20\) epochs. For the RNNs, a grid search was performed over the following parameters: number of hidden layers \(L \in \left\{ 1,2,3 \right\}\) and number of hidden units per layer \(H \in \left\{10,25,50,100,150 \right\}\). An initial learning rate of \(0.001\) was used and linearly decreased to \(0\) over \(1000\) iterations. A constant momentum rate of \(0.9\) was used for all the updates. The training was stopped if the error on the validation set did not decrease after \(20\) epochs. The training data was further divided into sub-sequences of length \(100\) and the networks were trained on these sub-sequences without any mini-batching. Gradient clipping was also used to avoid the exploding gradient problem in the early stages of RNN training. Clipping was triggered if the norm of the gradient update exceeded \(10\). # Experimental results [\[sec:results\]]{#sec:results label="sec:results"} In this section, classification performance is analysed for the task of recognising smoke alarms and baby cries against the large number of impostor sounds from the world set. Table [\[classifier_performance\]](#classifier_performance){reference-type="ref" reference="classifier_performance"} presents the EER for the best performing classifiers of each type and their associated computational cost. Figure [\[DET_curves\]](#DET_curves){reference-type="ref" reference="DET_curves"} shows the corresponding DET curves. In Figure [\[scatter_plots\]](#scatter_plots){reference-type="ref" reference="scatter_plots"} we present a more detailed comparison between the performance of various classifier types against the computational cost involved in classifying a frame of input at test time. **Baby Cry data set-** From Table [\[performance_baby_cry\]](#performance_baby_cry){reference-type="ref" reference="performance_baby_cry"} we observe that the best performing GMM has 32 components and achieves an EER of 14.0% for frame-wise classification. From Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"} (+ markers) we observe that there is no performance improvement when the number of Gaussian components is increased beyond 32. It should be noted that the best performing GMM (with \(32\) components) has lower computational cost compared to any SVM classifier, and a cost comparable on average with DNN classifiers. From Table [\[performance_baby_cry\]](#performance_baby_cry){reference-type="ref" reference="performance_baby_cry"}, we observe that a SVM with a linear kernel is the best performing SVM classifier, with an EER of \(12.9\%\). The SVM was trained on 2000 examples, resulting in 655 support vectors. From Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"}, we note that the performance of the linear SVM (blue triangles) increases as the computational complexity (number of support vectors) is increased. The number of support vectors can also be controlled by varying the parameter \(C\). For the best performing linear SVM, \(C=1.0\) yields the best results. We observe that SVMs with a sigmoid kernel (light blue triangles) yield similar results to the linear SVM (Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"}). The best SVM with a sigmoid kernel yields an EER of 13.5%, while the second best sigmoid SVM has an EER of 13.9%. As in the linear case, we observe an improvement in test performance as the computational cost or the number of support vectors is increased. From Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"}, we note that SVMs with RBF and polynomial kernels (green and red triangles) are outperformed by linear and sigmoid SVMs, both in terms of %EER and computational cost. There is also no observable trend between test performance and computational cost. For the polynomial SVM (red triangles), we found a kernel with \(d = 3\) yielded the best performance, while low values of gamma \(\gamma \in (0.005,0.01)\) provided the best results for the RBF kernel (Section [\[SVMs\]](#SVMs){reference-type="ref" reference="SVMs"}). The number of training examples \(T = 2000\) was empirically determined to be the optimal value: adding more training examples did not yield any improvement in performance. Conversely, we tried training the same classifiers with \(T = 500\) training examples, since the computational cost of SVMs is determined both by the type of kernel and the number of support vectors, the latter being controlled by varying the number of training examples and the penalty parameter \(C\). With \(T = 500\) and \(C=0.1\), we were able to achieve a minimum EER rate of \(13.7\%\) with the linear SVM, while halving the number of operations. From Table [\[performance_baby_cry\]](#performance_baby_cry){reference-type="ref" reference="performance_baby_cry"}, the best performing neural network architecture, which achieves an EER of \(10.8\%\), is a feed-forward DNN with sigmoid activations for the hidden units. Figure [\[DET_baby_cry\]](#DET_baby_cry){reference-type="ref" reference="DET_baby_cry"} shows that the neural network clearly outperforms all the other models. From Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"} we observe that the feed forward DNNs (circle and square markers) achieve similar test performance over a wide range of computational costs. This demonstrates that the network performance is not particularly sensitive to the specific number of hidden units in each layer. However, we did observe that networks which were deeper (\(>1\) hidden layer) yielded better performance. From Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"}, we observe that the RNN architectures (hexagonal markers) yield slightly worse performance and are computationally more expensive. An interesting observation from Figure [\[baby_cry_scatter\]](#baby_cry_scatter){reference-type="ref" reference="baby_cry_scatter"} is that feed-forward DNNs with both sigmoid and ReLU activations yield similar results. This is a very important factor when deploying these models on embedded hardware, since a ReLU net can be implemented with only linear operations (multiplications and additions), without the need for costly Taylor series expansions or LUTs. **Smoke Alarm data set-** From Table [\[performance_smoke_alarms\]](#performance_smoke_alarms){reference-type="ref" reference="performance_smoke_alarms"}, we observe that the best performing GMM yields an EER of 2.9% and uses 16 mixture components. From Figure [\[smoke_alarm_scatter\]](#smoke_alarm_scatter){reference-type="ref" reference="smoke_alarm_scatter"} (+ markers), we observe that the GMM performance improves as the computation cost (number of Gaussians) increases till \(M=32\). Beyond this, increasing the number of Gaussian components does not improve results. Again, we observe that the best performing GMM with \(M=32\) has much lower computational cost compared to SVMs, and a cost comparable on average with DNNs. Similar to the results on the Baby Cry data set, the linear and sigmoid kernel SVMs show the best performance, out of all four SVM kernel types. The best linear SVM has an EER of 3.0%--a small improvement over the best GMM approach, but with a small increase in the number of operations. The model used 2000 training examples and \(C=1.0\). From Figure [\[smoke_alarm_scatter\]](#smoke_alarm_scatter){reference-type="ref" reference="smoke_alarm_scatter"} we again observe an improvement in performance, with an increase in computation cost for both the linear and sigmoid kernels (blue triangles). The best sigmoid kernel SVM scored 3.5% with 2000 training examples, while another configuration scored 3.6% with \(T = 500\) and \(C = 1\), at half the number of operations. Again, the polynomial and RBF kernels (red and green triangles) yield lower performance, with no observable trend in terms of performance versus computational cost (Figure [\[smoke_alarm_scatter\]](#smoke_alarm_scatter){reference-type="ref" reference="smoke_alarm_scatter"}). From Table [\[performance_smoke_alarms\]](#performance_smoke_alarms){reference-type="ref" reference="performance_smoke_alarms"}, we observe that a feed-forward sigmoid DNN yields the best performance, with an EER of \(1.6\%\). From the DET curves (Figure [\[DET_smoke_alarms\]](#DET_smoke_alarms){reference-type="ref" reference="DET_smoke_alarms"}), we see that the neural network clearly outperforms the other models. From figure [\[smoke_alarm_scatter\]](#smoke_alarm_scatter){reference-type="ref" reference="smoke_alarm_scatter"} we note that the neural networks consistently perform better than the other models, over a wide range of computational costs, which correspond to different network configurations (number of layers, number of units in each layer). The ReLU networks perform similarly to the sigmoid networks, while the RNNs perform worse and are computationally more costly. It is worth noting that the performance of all classifiers is significantly better for the smoke alarm sounds, since the smoke alarms are composed of simple tones. On the other hand, baby cries have a large variability and are therefore more difficult to classify. # Conclusion [\[sec:conclusion\]]{#sec:conclusion label="sec:conclusion"} In this study, we compare the performance of neural network acoustic models with GMMs and SVMs on an environmental audio event detection task. Unlike other machine learning systems, AESR systems are usually deployed on embedded hardware, which imposes many computational constraints. Keeping this in mind, we compare the performance of the models as a function of their computational cost. We evaluate the models on two tasks, detecting baby cries and detecting smoke alarms against a large number of impostor sounds. These data sets are much larger than the popular data sets found in AESR literature, which enables us to train neural network acoustic models. Additionally, the large number of impostor sounds allows to investigate the performance of the proposed models in a testing scenario that is closer to practical use cases than previously available data sets. Results suggest that GMMs provide a low cost baseline for classification, across both data sets. The GMM acoustic models are able to perform reasonably well at a modest computational cost. SVMs with linear and sigmoid kernels yield similar EER performance compared to GMMs, but their computational cost is overall higher. The computational cost of the SVM is determined by the number of support vectors. Unlike GMMs, SVMs are non-parametric models which do not allow the direct specification of model parameters, although the number of support vectors can be indirectly controlled with regularisation. Finally, our results suggest that deep neural networks consistently outperform both the GMMs and the SVMs on both data sets. The computational cost of DNNs can be controlled by limiting the number of hidden units and the number of layers. While changes in the number of units in the hidden layers did not appear to have a large impact on performance, deeper networks appeared to perform better in all cases. Additionally, neural networks with ReLU activations achieved good performance, while being an attractive choice for deployment on embedded devices because they do not require expensive LUT lookup operations. In the future, we would like to expand the evaluations presented here to include more event classes. However, the lack of large data sets for AESR problems is a major limitation. We hope that studies like this one will encourage collaboration with industry partners to collect large data sets for more rigorous evaluations. We would also like to investigate the performance of acoustic models as a function of the memory efficiency, since memory is an important consideration when designing models for embedded hardware.

{'timestamp': '2016-07-18T02:08:46', 'yymm': '1607', 'arxiv_id': '1607.04589', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04589'}

# Introduction Recent years have witnessed a remarkable convergence of two broad trends. The first of these concerns information i.e. data--rapid technological advances coupled with an increased presence of computing in nearly every aspect of daily life, have for the first time made it possible to acquire and store massive amounts of highly diverse types of information. Concurrently and in no small part propelled by the environment just described, research in artificial intelligence--in machine learning, data mining , and pattern recognition, in particular--has reached a sufficient level of methodological sophistication and maturity to process and analyse the collected data, with the aim of extracting novel and useful knowledge . Though it is undeniably wise to refrain from overly ambitious predictions regarding the type of knowledge which may be discovered in this manner, at the very least it is true that few domains of application of the aforesaid techniques hold as much promise and potential as that of medicine and health in general. Large amounts of highly heterogeneous data types are pervasive in medicine. Usually the concept of so-called "big data" in medicine is associated with the analysis of Electronic Health Records, large scale sociodemographic surveys of death causes, social media mining for health related data  etc. Much less discussed and yet arguably no less important realm where the amount of information presents a challenge to the medical field is the medical literature corpus itself. Namely, considering the overarching and global importance of health (to say nothing of practical considerations such as the availability of funding), it is not surprising to observe that the amount of published medical research is immense and its growth is only continuing to accelerate. This presents a clear challenge to a researcher. Even restricted to a specified field of research, the amount of published data and findings makes it impossible for a human to survey the entirety of relevant publications exhaustively which inherently leads to the question as to what kind of important information or insight may go unnoticed or insufficiently appreciated. The premise of the present work is that advanced machine learning techniques can be used to assist a human in the analysis of this data. Specifically, we introduce a novel methodology based on Bayesian non-parametric inference that achieves this, as well as free software which researchers can use in the analysis of their corpora of interest. ### Previous work A limitation of most models described in the existing literature lies in their assumption that the data corpus is static. Here the term 'static' is used to describe the lack of any associated temporal information associated with the documents in a corpus--the documents are said to be exchangeable . However, research articles are added to the literature corpus in a temporal manner and their ordering has significance. Consequently the topic structure of the corpus changes over time : new ideas emerge, old ideas are refined, novel discoveries result in multiple ideas being related to one another thereby forming more complex concepts or a single idea multifurcating into different 'sub-ideas' etc. The premise in the present work is that documents are not exchangeable at large temporal scales but can be considered to be at short time scales, thus allowing the corpus to be treated as *temporally locally static*. # Proposed approach[\[s:proposed\]]{#s:proposed label="s:proposed"} In this section we introduce our main technical contributions. We begin by reviewing the relevant theory underlying Bayesian mixture models, and then explain how the proposed framework employs these for the extraction of information from temporally varying document corpora. ## Bayesian mixture models {#ss:mixModels} Mixture models are appropriate choices for the modelling of so-called heterogeneous data whereby heterogeneity is taken to mean that observable data is generated by more than one process (source). The key challenges lie in the lack of observability of the correspondence between specific data points and their sources, and the lack of *a priori* information on the number of sources . Bayesian non-parametric methods place priors on the infinite-dimensional space of probability distributions and provide an elegant solution to the aforementioned modelling problems. Dirichlet Process (DP) in particular allows for the model to accommodate a potentially infinite number of mixture components : \[\begin{aligned} p\left(x|\pi_{1:\infty},\phi_{1:\infty}\right)=\sum_{k=1}^{\infty}\pi_{k}f\left(x|\phi_{k}\right). \end{aligned}\] where \(\text{DP}\left(\gamma,H\right)\) is defined as a distribution of a random probability measure \(G\) over a measurable space \(\left(\Theta,\mathcal{B}\right)\), such that for any finite measurable partition \(\left(A_{1},A_{2},\ldots,A_{r}\right)\) of \(\Theta\) the random vector \(\left(G\left(A_{1}\right),\ldots,G\left(A_{r}\right)\right)\) is a Dirichlet distribution with parameters \(\left(\gamma H\left(A_{1}\right),\ldots,\gamma H\left(A_{r}\right)\right)\). Owing to the discrete nature and infinite dimensionality of its draws, the DP is a useful prior for Bayesian mixture models. By associating different mixture components with atoms \(\phi_{k}\), and assuming \(x_{i}|\phi_{k}\overset{iid}{\sim}f\left(x_{i}|\phi_{k}\right)\) where \(f\left(.\right)\) is the kernel of the mixing components, a Dirichlet process mixture model (DPM) is obtained . ### Hierarchical DPMs While the DPM is suitable for the clustering of exchangeable data in a single group, many real-world problems are more appropriately modelled as comprising multiple groups of exchangeable data. In such cases it is desirable to model the observations of different groups jointly, allowing them to share their generative clusters. This so-called "sharing of statistical strength" emerges naturally when a hierarchical structure is implemented. The DPM models each group of documents in a collection using an infinite number of topics. However, it is desired for multiple group-level DPMs to share their clusters. The hierarchical DP (HDP)  offers a solution whereby base measures of group-level DPs are drawn from a corpus-level DP. In this way the atoms of the corpus-level DP are shared across the documents; posterior inference is readily achieved using Gibbs sampling . ## Modelling topic evolution over time[\[ss:contrib\]]{#ss:contrib label="ss:contrib"} We now show how the described HDP based model can be applied to the analysis of temporal topic changes in a *longitudinal* data corpus. Owing to the aforementioned assumption of a temporally locally static corpus we begin by discretizing time and dividing the corpus into epochs. Each epoch spans a certain contiguous time period and has associated with it all documents with timestamps within this period. Each epoch is then modelled separately using a HDP, with models corresponding to different epochs sharing their hyperparameters and the corpus-level base measure. Hence if \(n\) is the number of epochs, we obtain \(n\) sets of topics \(\boldsymbol{\phi}=\left\{ \boldsymbol{\phi}_{t_{1}},\ldots,\boldsymbol{\phi}_{t_{n}}\right\}\) where \(\boldsymbol{\phi}_{t}=\left\{ \theta_{1,t},\ldots,\phi_{K_{t},t}\right\}\) is the set of topics that describe epoch \(t\), and \(K_{t}\) their number. ### Topic relatedness[\[ss:similarity\]]{#ss:similarity label="ss:similarity"} Our goal now is to track changes in the topical structure of a data corpus over time. The simplest changes of interest include the emergence of new topics, and the disappearance of others. More subtly, we are also interested in how a specific topic changes, that is, how it evolves over time in terms of the contributions of different words it comprises. Lastly, our aim is to be able to extract and model complex structural changes of the underlying topic content which result from the interaction of topics. Specifically, topics, which can be thought of as collections of memes, can merge to form new topics or indeed split into more nuanced memetic collections. This information can provide valuable insight into the refinement of ideas and findings in the scientific community, effected by new research and accumulating evidence. The key idea behind our tracking of simple topic evolution stems from the observation that while topics may change significantly over time, changes between successive epochs are limited. Therefore we infer the continuity of a topic in one epoch by relating it to all topics in the immediately subsequent epoch which are sufficiently similar to it under a suitable similarity measure--we adopt the well known Bhattacharyya distance (BHD). This can be seen to lead naturally to a similarity graph representation whose nodes correspond to topics and whose edges link those topics in two epochs which are related. Formally, the weight of the directed edge that links \(\phi_{j,t}\), the \(j\)-th topic in epoch \(t\), and \(\phi_{k,t+1}\) is \(\rho_\text{BHD}\left(\phi_{j,t},\phi_{k,t+1}\right)\) where \(\rho_\text{BHD}\) denotes the BHD. In constructing a similarity graph a threshold to used to eliminate automatically weak edges, retaining only the connections between sufficiently similar topics in adjacent epochs. Then the disappearance of a particular topic, the emergence of new topics, and gradual topic evolution can be determined from the structure of the graph. In particular if a node does not have any edges incident to it, the corresponding topic is taken as having emerged in the associated epoch. Similarly if no edges originate from a node, the corresponding topic is taken to vanish in the associated epoch. Lastly when exactly one edge originates from a node in one epoch and it is the only edge incident to a node in the following epoch, the topic is understood as having evolved in the sense that its memetic content may have changed. A major challenge to the existing methods in the literature concerns the detection of topic merging and splitting. Since the connectedness of topics across epochs is based on their similarity what previous work describes as 'splitting' or indeed 'merging' does not adequately capture these phenomena. Rather, adopting the terminology from biological evolution, a more accurate description would be 'speciation' and 'convergence' respectively. The former is illustrated in Fig [\[f:speciation\]](#f:speciation){reference-type="ref" reference="f:speciation"} whereas the latter is entirely analogous with the time arrow reversed. What the conceptual diagram shown illustrates is a slow differentiation of two topics which originate from the same 'parent'. Actual topic splitting, which does not have a biological equivalent in evolution, and which is conceptually illustrated in Fig [\[f:splitting\]](#f:splitting){reference-type="ref" reference="f:splitting"} cannot be inferred by measuring topic similarity. Instead, in this work we propose to employ the Kullback-Leibler divergence (KLD) for this purpose. This divergence is asymmetric can be intuitively interpreted as measuring how well one probability distribution 'envelops' another. KLD between two probability distributions \(p(i)\) and \(q(i)\) is defined as follows: \[\begin{aligned} \rho_\text{KLD} = \sum_i p(i) \log \frac{p(i)}{q(i)} \end{aligned}\] It can be seen that a high penalty is incurred when \(p(i)\) is significant and \(q(i)\) is low. Hence, we use the BHD to track gradual topic evolution, speciation, and convergence, while the KLD (computed both in forward and backward directions) is used to detect topic splitting and merging. ### Automatic temporal relatedness graph construction[\[ss:construction\]]{#ss:construction label="ss:construction"} Another novelty of the work first described in this paper concerns the building of the temporal relatedness graph. We achieve this almost entirely automatically, requiring only one free parameter to be set by the user. Moreover the meaning of the parameter is readily interpretable and understood by a non-expert, making our approach highly usable. Our methodology comprises two stages. Firstly we consider all inter-topic connections present in the initial fully connected graph and extract the empirical estimate of the corresponding cumulative density function (CDF). Then we prune the graph based on the operating point on the relevant CDF. In other words if \(F_\rho\) is the CDF corresponding to a specific initial, fully connected graph formed using a particular similarity measure (BHD or KLD), and \(\zeta \in [0, 1]\) the CDF operating point, we prune the edge between topics \(\phi_{j,t}\) and \(\phi_{k,t+1}\) iff \(\rho(\phi_{j,t},\phi_{k,t+1}) < F^{-1}_\rho (\zeta)\). # Evaluation and discussion We now analyse the performance of the proposed framework empirically on a large real world data set. ## Evaluation data {#sss:rawData} We used the PubMed interface to access the US National Library of Medicine and retrieve from it scholarly articles. We searched for publication on the metabolic syndrome (MetS) using the keyphrase"metabolic syndrome" and collected papers written in English. The earliest publication found was that by Berardinelli *et al.* . We collected all matching publications up to the final one indexed by PubMed on 10th Jan 2016, yielding a corpus of 31,706 publications. ### Pre-processing[\[sss:preprocessing\]]{#sss:preprocessing label="sss:preprocessing"} The raw data collected from PubMed is in the form of free text. To prepare it for automatic analysis a series of 'pre-processing' steps are required. The goal is to remove words which are largely uninformative, reduce dispersal of semantically equivalent terms, and thereafter select terms which are included in the vocabulary over which topics are learnt. We firstly applied soft lemmatization using the WordNet\(^\circledR\) lexicon  to normalize for word inflections. No stemming was performed to avoid semantic distortion often effected by heuristic rules used by stemming algorithms. After lemmatization and the removal of so-called stop-words, we obtained approximately 3.8 million terms in the entire corpus when repetitions are counted, and 46,114 unique terms. Constructing the vocabulary for our method by selecting the most frequent terms which explain 90% of the energy in a specific corpus resulted in a vocabulary containing 2,839 terms. ## Results We stared evaluation by examining whether the two topic relatedness measures (BHD and KLD) are capturing different aspects of relatedness. To obtain a quantitative measure we looked at the number of inter-topic connections formed in respective graphs both when the BHD is used as well as when the KLD is applied instead. The results were normalized by the total number of connections formed between two epochs, to account for changes in the total number of topics across time. Our results are summarized in Fig [\[f:common\]](#f:common){reference-type="ref" reference="f:common"}. A significant difference between the two graphs is readily evident--across the entire timespan of the data corpus, the number of Bhattacharyya distance based connections also formed through the use of the KLD is less than 40% and in most cases less than 30%. An even greater difference is seen when the proportion of the KLD connections is examined--it is always less than 25% and most of the time less than 15%. To get an even deeper insight into the contribution of the two relatedness measures, we examined the corresponding topic graphs before edge pruning. The plot in Fig [\[f:smooth\]](#f:smooth){reference-type="ref" reference="f:smooth"} shows the variation in inter-topic edge strengths computed using the BHD and the KLD (in forward and backward directions)--the former as the \(x\) coordinate of a point corresponding to a pair of topics, and the latter as its \(y\) coordinate. The scatter of data in the plot corroborates our previous observation that the two similarity measures indeed do capture different aspects of topic behaviour. We performed extensive qualitative analysis which is necessitated by the nature of the problem at hand and the so-called 'semantic gap' that underlies it. In all cases we found that our algorithm revealed meaningful and useful information, as confirmed by an expert in the area of metabolic MetS research. Our final contribution comprises a web application which allows users to upload and analyse their data sets using the proposed framework. The application allows a range of powerful tasks to be performed quickly and in an intuitive manner. For example, the user can search for a given topic using keywords (and obtain a ranked list), trace the origin of a specific topic backwards in time, or follow its development in the forward direction, examine word clouds associated with topics, display a range of statistical analyses, or navigate the temporal relatedness graph. # Summary and Conclusions In this work we presented a case for the importance of use of advanced machine learning techniques in the analysis and interpretation of medical literature. We described a novel framework based on non-parametric Bayesian techniques which is able to extract and track complex, semantically meaningful changes to the topic structure of a longitudinal document corpus. Moreover this work is the first to describe and present a method for differentiating between two types of topic structure changes, namely topic splitting and what we termed topic speciation. Experiments on a large corpus of medical literature concerned with the metabolic syndrome was used to illustrate the performance of our method. Lastly, we developed a web application which allows users such as medical researchers to upload their data sets and apply our method for their analysis; the application and its code will be made freely available following publication.

{'timestamp': '2016-07-19T02:00:36', 'yymm': '1607', 'arxiv_id': '1607.04660', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04660'}

# Introduction {#sec:intro} Many central problems in geometry, topology, and mathematical physics lead to questions concerning the long-time dynamics of solutions to ordinary and parabolic partial differential equations. Three basic scenarios are possible.\ First, under suitable dissipativity conditions, longtime existence is guaranteed and a low-dimensional global attractor of uniformly bounded eternal solutions exists. The structure of such global attractors details longtime behavior. Some of those structures are encoded in meanders--which open doors to a broad variety of algebraic structures far beyond the original PDE scope of the attractor question.\ Second, longtime existence without dissipativity features trajectories which escape to infinity in infinite time (grow-up). In some cases it is possible to construct compactifications of an unbounded attractor containing, for example, equilibria at infinity. As an example we show how equivariant bifurcations give rise to black hole initial data that is not spherically symmetric.\ Third, singularities may develop after finite time. In other words, some solutions escape to infinity within finite time (blow-up). Some of the ramifications are explored in complex time, here. Cosmologically, the most meaningful singularity is the big-bang. We address the Bianchi models: a system of ordinary differential equations which represent the homogenous, but anisotropic, Einstein equations. The Belinsky--Khalatnikov--Lifshitz (BKL) conjecture favors them as models for a spacelike singularity. We obtain stability results for a primordial universe tumbling between different Kasner states, each of which represents a self-similarly expanding universe. Furthermore we address the BKL question of particle horizons. To conclude, we embed general relativity in the broader context of Hořava--Lifschitz gravity, to study the influence on the tumbling behavior between the Kasner states and the resulting chaoticity. # Global attractors of parabolic equations {#sec:globalattractors} In this section we survey recent progress on the global dynamics, mostly of the scalar parabolic semilinear equation \[\label{C8:eq:2.1.1} u_t = u_{xx} +f\.\] The emerging global theory gives rise to surprising topological structures, as finite regular cell complexes, in Section [2.1](#subsec:sturmglobalattractors){reference-type="ref" reference="subsec:sturmglobalattractors"}, and to combinatorial algebraic structures like Temperley--Lieb algebras, in Section [2.2](#subsec:meanders){reference-type="ref" reference="subsec:meanders"}. An application to axisymmetric, but not spherically symmetric, initial data for black holes which satisfy the Einstein constraint is given in Section [2.4](#subsec:blackholedata){reference-type="ref" reference="subsec:blackholedata"}. The question of stability of stationary solutions is addressed in Section [2.3](#subsec:statsol){reference-type="ref" reference="subsec:statsol"}.\ For simplicity we consider the PDE [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} under Neumann \(\mathcal{N}\) or periodic \(\mathcal{P}\) boundary conditions, on the unit interval \(0<x<1\). Other separated linear boundary conditions can be treated analogously to the Neumann case. The quasilinear uniformly parabolic case may easily be believed to yield identical results, but the long necessary groundwork is not complete at this stage. ## Sturm global attractors {#subsec:sturmglobalattractors} We consider the PDE [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} for several classes of nonlinearities \(f\). In particular, we survey our results . We denote these classes by \(\mathcal{N}(x,u,u_x)\), \(\mathcal{N}(u,u_x)\), \(\dots\) for \(f=f(x,u,u_x)\), \(f=f(u,u_x)\), \(\dots\) under Neumann boundary conditions, and by \(\mathcal{P}(x,u,u_x)\), \(\mathcal{P}(u,u_x)\), \(\dots\) in the \(x\)-periodic case. Dirichlet or other separated boundary conditions of mixed type can be treated analogously to the Neumann case. We always assume \(f\in C^1\) to be dissipative, i.e., the set of bounded *eternal solutions* \(u(t,x),t\in\mathbb{R}\) is itself bounded in a suitable Sobolev type interpolation space \(u(t,\cdot)\in X \hookrightarrow C^1\). Explicit assumptions on \(f\) which guarantee dissipativity involve, for example, sign conditions \(f(x,u,u_x) \cdot u<0\), for large \(|u|\), and subquadratic growth in \(u_x\). The bounded set \(\mathcal{A}_f \subseteq X\) of spatial profiles \(u(t,\cdot)\) on bounded eternal solutions is called the *global attractor* of [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}. The concept originated with Ladyzhenskaya's work on the two-dimensional Navier--Stokes equation.\ For general surveys on global attractors \(\mathcal{A}\) see for example . The general theory establishes properties like existence, global attractivity (hence the name), compactness, finite-dimensionality in the fractal box counting sense, and sometimes connectivity. From such a general, and hence necessarily superficial, point of view the global attractors \(\mathcal{A}_f\) of our model class [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} are just an easy exercise.\ However, our objective is a much deeper global understanding of the detailed global structure of the global attractors \(\mathcal{A}_f\). Let us address the general Neumann case \(f\in\mathcal{N}(x,u,u_x)\) first. By results of Zelenyak and Matano, the PDE [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} then possesses a gradient-like Morse structure due to a decreasing Lyapunov, or energy, or Morse, functional \(V\). Throughout we assume all Neumann equilibrium solutions \[\label{C8:eq:2.1.2} 0=v_{xx}+f\] to be hyperbolic, i.e., to be quadratically nondegenerate critical points of \(V\). Then \[\label{C8:eq:2.1.3} \mathcal{A}_f =\mathcal{E}_f \cup\mathcal{H}_f\] consists of the finitely many equilibria \(v\in\mathcal{E}_f\) and *heteroclinic orbits* \(u(t,\cdot)\in\mathcal{H}_f\) *connecting* them: \[\label{C8:eq:2.1.4} u(t,\cdot)\rightarrow v_{\pm}\in\mathcal{E}_f \quad \text{for} \quad t \rightarrow \pm \infty\,,\] with equilibria \(v_+\neq v_-\). We call \(\mathcal{A}_f\) a *Sturm global attractor*, to emphasize the particular origin from the one-dimensional parabolic PDE class [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}. Information on the vertex pairs \((v_-,v_+)\) which do possess a heteroclinic edge [\[C8:eq:2.1.4\]](#C8:eq:2.1.4){reference-type="eqref" reference="C8:eq:2.1.4"}, and on the pairs which don't, can be encoded in the (directed, acyclic, loop-free) *connection graph* \(\mathcal{C}_f\); see  and Fig. [\[C8:fig:2.1.1\]](#C8:fig:2.1.1){reference-type="ref" reference="C8:fig:2.1.1"}(a).\ As was discovered by Fusco and Rocha , a permutation \(\sigma\) given by the boundary values of the equilibria \(\mathcal{E}_f= \{ v_1,\dots, v_N\}\) is the key to a global analysis of the Sturm global attractor \(\mathcal{A}_f\) and its heteroclinic connection graph \(\mathcal{C}_f\). The *Sturm permutation* \(\sigma=\sigma_f\) encodes the orderings of the equilibria, \[\label{C8:eq:2.1.5} \begin{aligned} v_1 \;&<\; v_2 \!\!\!\!&<\; \dots \;&< \; v_N &\quad \text{at}\quad x=0\,,\\ v_{\sigma (1)} \;&<\; v_{\sigma (2)} \!\!\!\!&< \;\dots \;&< \;v_{\sigma (N)} &\quad \text{at} \quad x=1\,, \end{aligned}\] at the respective Neumann boundaries. More precisely, the Sturm permutations \(\sigma_f\) determine all *unstable dimensions*, alias *Morse indices*, \(i=i(v_k)\) of the equilibria \(v_k \in \mathcal{E}\); see . This is a linear ODE question of Sturm--Liouville type; hence the name Sturm permutation for \(\sigma_f\). Moreover \(\sigma_f\) determines the (PDE) heteroclinic connection graph \(\mathcal{C}_f\), explicitly and constructively, in a combinatorial and algorithmic style . Global Sturm attractors \(\mathcal{A}_f\) and \(\mathcal{A}_g\) with the same Sturm permutation \(\sigma_f = \sigma_g\) are in fact \(C^0\) orbit equivalent, i.e. homeomorphic in \(X\) under a homeomorphism which maps eternal PDE orbits \(\{ u(t,\cdot),t\in\mathbb{R}\}\subseteq\mathcal{A}_f\) to eternal orbits in \(\mathcal{A}_g\); see .\ Not all permutations \(\sigma\) are Sturm permutations \(\sigma_f\) arising from the Neumann class \(f \in \mathcal{N}(x,u,u_x)\) of [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}. In fact \(\sigma=\sigma_f\) for some \(f\), if and only if the permutation \(\sigma\in S_N\) is a meander permutation which, in addition, is dissipative and Morse . Following Arnold, here, we call \(\sigma\) a *meander permutation* if \(\sigma\) describes the transverse crossings of a planar \(C^1\) Jordan curve, or "meandering river", oriented from south-west to north-east, with the horizontal axis, or "road". Labeling the crossings along the "river", sequentially, the meander permutation \(\sigma\) is defined by the labels \(\sigma(1),\sigma(2),\dots,\sigma(N)\) read along the road, left to right. We call a meander \(\sigma\) *dissipative* if \(\sigma(1)=1\) and \(\sigma(N)=N\). We call \(\sigma\) *Morse* if the integers \[\label{C8:eq:2.1.6} i_k:= \sum_{j=1}^{k-1}(-1)^{j+1}\operatorname{sign}\big(\sigma^{-1}(j+1)-\sigma^{-1} (j)\big)\] are nonnegative, for all \(k=1,\dots,N\). Geometrically, this amounts to a net right winding of the river, as counted from the first crossing; see Fig. [\[C8:fig:2.1.1\]](#C8:fig:2.1.1){reference-type="ref" reference="C8:fig:2.1.1"}(b) and Fig. [\[C8:fig:2.1.2\]](#C8:fig:2.1.2){reference-type="ref" reference="C8:fig:2.1.2"} for many examples. The PDE connection graphs \(\mathcal{C}_f\) then follow, in each case, without further PDE analysis.\ It is not difficult to see why Sturm permutations \(\sigma=\sigma_f\) satisfy these three properties. Dissipativeness of \(\sigma\) follows because \(\mathcal{E}_f\) must contain extreme equilibria \(v_1(x)<v_2(x),\dots,v_{N-1}(x)<v_N(x)\), for all \(0\leq x\leq1\), by dissipativeness of [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} and parabolic comparison. The Morse property \(i_k\geq0\) follows because \(i_k=i(v_k)\geq0\) are the nonnegative Morse indices; see . The meander property follows by ODE shooting for the equilibrium boundary value problem [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"} in the \((v,v_x)\) plane. Indeed, the "river" is the image of the horizontal Neumann \(v\)-axis \(\{ v_x =0\}\), at \(x=0\), propagated to \(x=1\) under the ODE-flow [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"}. The "road" is the horizontal Neumann \(v\)-axis \(\{v_x=0\}\), at \(x=1\). See Section [2.2](#subsec:meanders){reference-type="ref" reference="subsec:meanders"} for a deeper discussion of meanders. The converse claim, how all dissipative Morse meanders \(\sigma\) are actually Sturm permutations \(\sigma=\sigma_f\) for some \(f\), is beyond the limitations of this introductory survey. See. The deceptively elementary description of Sturm permutations \(\sigma_f\) should not conceal the formidable complications arising from the meander property. Already Gauss got intrigued by the related question of self-intersecting orders in planar curves; see . Rescaling \(x=1\) to become large, the evident complexities of chaotic stroboscope maps for the periodically forced pendulum or the forced van der Pol oscillator are also lurking here, as special cases.\ A key technical ingredient to the results indicated above are nodal properties, i.e. a refinement of the parabolic comparison principle in one space dimension, or in radially symmetric settings. Indeed, let \(u_1(t,x)\) and \(u_2(t,x)\) be any two non-identical solutions of [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}, in \(X \hookrightarrow C^1\), and let the *zero number* \(z(\varphi)\leq\infty\) count the strict sign changes of any continuous \(x\)-profile \(\varphi: [0,1]\rightarrow\mathbb{R}\). Then the zero number \[\label{C8:eq:2.1.7} t \mapsto z\big(u_1(t, \cdot)-u_2 (t,\cdot )\big)\] is finite, for any \(t >0\), nonincreasing with \(t\), and drops strictly at any multiple zeros of \(x\mapsto u_1(t,x)-u_2(t,x)\); see . Early versions of this, in the linear case, in fact date back to Sturm and convinced us to name the nonlinear ramifications after him; see .\ It is the zero number dropping property [\[C8:eq:2.1.7\]](#C8:eq:2.1.7){reference-type="eqref" reference="C8:eq:2.1.7"} which generates surprising rigidity in the global Sturm attractors \(\mathcal{A}_f\) and enables a detailed global analysis. For example, stable and unstable manifolds of hyperbolic equilibria \(v_+\) and \(v_-\) intersect transversely, automatically, along any heteroclinic PDE orbit \(u(t, \cdot)\rightarrow v_\pm\), for \(t \rightarrow\pm\infty\), in \(\mathcal{H}_f\); see . In particular, this implies local structural stability of Sturm attractors and, moreover, a strict dropping of Morse indices, \[\label{C8:eq:2.1.8} i(v_-) > z\big(u(t,\cdot)-v_-\big) \geq z(v_+-v_-) \geq z\big(v_+-u(t,\cdot)\big) \geq i(v_+)\.\] By a (standard) transitivity property of transverse heteroclinicity and a (Sturm specific) cascading principle it turns out, in fact, that the full heteroclinic connection graph \(\mathcal{C}_f\) is generated by the edges \((v_-,v_+)\) with adjacent Morse index. Saddle-saddle heteroclinic orbits \(i(v_-) = i(v_+)\) are also excluded by Morse dropping [\[C8:eq:2.1.8\]](#C8:eq:2.1.8){reference-type="eqref" reference="C8:eq:2.1.8"}. This precludes global topological ambiguities which, for example, homotopy-invariant Conley--Morse theory is not allowed to discard. Indeed the positive integer polynomial terms \((1+t)\,q(t)\) in the Morse (in)equalities, alias connecting homomorphisms or Conley--Franzosa connection matrices, always need to accommodate global reshufflings of the connection graph due to nongeneric saddle-saddle connections. This is where nodal arguments involving the zero number \(z(u_1-u_2)\) step in.\ We summarize more recent results on Sturm global attractors \(\mathcal{A}_f\) and their characterizing Sturm permutations \(\sigma_f\) for the parabolic PDE [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} under Neumann \((\mathcal{N})\) and periodic \((\mathcal{P})\) boundary conditions, next. We conclude with recent progress in our global geometric understanding of \(\mathcal{A}_f\) as regular topological cell complexes.\ Little information is provided, in our approach, on the global Sturm attractor \(\mathcal{A}_f\) of particular nonlinearities \(f\in\mathcal{N}(x,u,u_x)\). The reason is purely on the ODE pendulum side [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"}. It is not an easy task to determine the equilibrium boundary order of \(v\in\mathcal{E}_f\), as encoded in the Sturm permutation \(\sigma_f\), for particular nonlinearities \(f\). Indeed we address the whole class of dissipative nonlinearities \(f=f(x,u,u_x)\), rather than any particular instance.\ As a partial remedy we also consider the subclass \(\mathcal{N}(u)\) of nonlinearities \(f=f(u)\), for which the equilibrium ODE [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"} becomes \(x\)-autonomous Hamiltonian. See  and the survey  for an explicit combinatorial characterization of this subclass of Sturm permutations \(\sigma_f\). For a complete listing with \(N=9\) equilibria see again Fig. [\[C8:fig:2.1.2\]](#C8:fig:2.1.2){reference-type="ref" reference="C8:fig:2.1.2"}. The case \(f\in\mathcal{N}(u,u_x^2)\) of nonlinearities \(f(u,p) =f(u,-p)\) which are even in \(p=u_x\) is only superficially more general. The equilibrium ODE [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"} is then only reversible in \(x\), but not necessarily Hamiltonian. Still, reversibility implies integrability here. The ODE [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"}, but not the PDE [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}, can be smoothly transformed to become Hamiltonian. The original Sturm permutation \(\sigma_f\) becomes an involution, \(\sigma_f=\sigma_f^{-1}\). Moreover \(\sigma_f\) can be represented by a Hamiltonian nonlinearity \(g\in\mathcal{N}(u)\) such that \(\sigma_f=\sigma_g\). In particular, the Sturm global attractors \(\mathcal{A}_f\) in the larger \(x\)-reversible class \(\mathcal{N}(u,u_x^2)\) coincide with those of the \(x\)-reversible Hamiltonian class \(\mathcal{N}(u)\).\ Next, we turn to the case \(f\in\mathcal{P}(x,u,u_x)\) of periodic boundary conditions \(x\in S^1\). The nodal dropping property [\[C8:eq:2.1.7\]](#C8:eq:2.1.7){reference-type="eqref" reference="C8:eq:2.1.7"} of zero numbers prevails. It has long been known, however, that any planar autonomous ODE vector field embeds into this larger class of PDEs . In particular, saddle-saddle heteroclinic and homoclinic trajectories may arise, and automatic transversality of stable and unstable manifolds fails. In the \(S^1\)-equivariant case of \(x\)-independent nonlinearities \(f\in\mathcal{P}(u,u_x)\), on the other hand, automatic transversality does follow from mere (normal) hyperbolicity of equilibria and nonconstant rotating waves \(u(t,x) = U\, (x-ct)\). In fact the global attractor then decomposes into equilibria \(\mathcal{E}\), rotating waves \(\mathcal{R}\), and their heteroclinic orbits \(\mathcal{H}\) as \[\label{C8:eq:2.1.9} \mathcal{A}_f = \mathcal{E}_f \cup \mathcal{R}_f \cup \mathcal{H}_f\.\] In this periodic class, the PDE [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} therefore defines a Morse--Smale system. In the \(x\)-reversible, i.e. \(O(2)\)-equivariant, class \(f \in \mathcal{P}(u,u_x^2)\) an explicit Lyapunov function has been constructed; see . Then all "rotating" waves are frozen, i.e. of wave speed \(c=0\), and time-periodic solutions other than equilibria cannot arise.\ In , we have developed an explicit algorithm to determine the heteroclinic connection graph in the general \(S^1\)-equivariant case \(f_0\in\mathcal{P}(u,u_x)\). The basic procedure consisted in a homotopy \(f_\tau\) from \(f_0\) to \(f_1\in\mathcal{P}(u,u_x^2)\) which preserved (normal) hyperbolicity of all equilibria and rotating waves, freezing the latter to wave speed \(c=0\) at \(\tau =1\). We were then able to derive the heteroclinic orbits in \(\mathcal{P}\) from those in the reversible Neumann class \(f_1\in \mathcal{N }(u,u_x^2)\) or, equivalently, the Hamiltonian Neumann class \(f_1\in\mathcal{N}(u)\).\ As an explicit example we mention the *Sturm spindle* attractor \(\mathcal{A}\) of Fig. [\[C8:fig:2.1.3\]](#C8:fig:2.1.3){reference-type="ref" reference="C8:fig:2.1.3"}, which consists of three equilibria \(e_k\) of Morse indices \(i(e_1)=i(e_3)=0\), \(i(e_2)=3\) and one rotating wave \(w\) with (strong) unstable dimension one. Surprisingly, the same global attractor \(\mathcal{A}\) has also been encountered by Krisztin and Walther in the quite different setting of delay differential equations \[\label{C8:eq:2.1.10} \dot{u}(t) = g\big(u(t),u(t-1)\big)\,,\] for positive monotone feedback \(g\) in the delayed argument; see . There is a nodal property similar to the zero number \(z\) of [\[C8:eq:2.1.8\]](#C8:eq:2.1.8){reference-type="eqref" reference="C8:eq:2.1.8"} at work, here. The delay class is not exhaustively classified; it might turn out to be a subclass of the parabolic class, together with the closely related class of negative monotone delayed feedback. This striking analogy remains unexplored at present.\ At the end of this section we return to the general Neumann class \(f\in\mathcal{N}(x,u,u_x)\). We have a combinatorial description of this PDE class in terms of ODE shooting meanders, the Sturm permutations \(\sigma_f\), and the explicit derivation of the heteroclinic connection graph \(\mathcal{C}_f\). We did not yet succeed, however, to a priori characterize the class of all connection graphs \(\mathcal{C}_f\) or, more ambitiously, the topological properties of the global Sturm attractors \(\mathcal{A}_f\) themselves--other than by mindless successive enumeration of all \(\sigma_f\). However, some promising steps have been achieved.\ We recall the gradient-like structure which implies a general decomposition \(\mathcal{A}=\mathcal{E}\cup\mathcal{H}\) of the global attractor into (hyperbolic) equilibria \(\mathcal{E}\) and their heteroclinic orbits \(\mathcal{H}\). In other words, \[\label{C8:eq:2.1.11} \mathcal{A}=\bigcup_{v\in\mathcal{E}}W^u(v)\] decomposes into the disjoint union of the unstable manifolds \(W^u(v)\) of all equilibria \(v\). Here \(W^u(v)\) consists of all ancient trajectories \(u(t,\cdot)\) in \(\mathcal{A}\) which limit onto \(v\), for \(t\rightarrow-\infty\). In particular, \(v \in W^u(v)\). Note that \(\dim W^u(v)=i(v)\) is the Morse index of \(v\), and \(W^u(v)=\{ v\}\) for \(i(v) = 0\). We call [\[C8:eq:2.1.11\]](#C8:eq:2.1.11){reference-type="eqref" reference="C8:eq:2.1.11"} the *dynamic complex* of \(\mathcal{A}\). In general, the boundaries \(\partial W^u =(\operatorname{clos}W^u)\smallsetminus W^u\) need not be manifolds. In fact, [\[C8:eq:2.1.11\]](#C8:eq:2.1.11){reference-type="eqref" reference="C8:eq:2.1.11"} need not even be a cell complex, even though the \(W^u(v)\) are embedded balls of dimensions \(i(v)\). In the class of Sturm global attractors \(\mathcal{A}= \mathcal{A}_f\), \(f \in \mathcal{N}(x,u,u_x)\), however, the decomposition [\[C8:eq:2.1.11\]](#C8:eq:2.1.11){reference-type="eqref" reference="C8:eq:2.1.11"} is a *regular cell complex*. This means that the embedded \(i(v)\)-cells \(W^u(v)\) possess embedded sphere boundaries \(\partial W^u(v) \cong S^{i(v)-1}\), so that the attaching maps onto the boundary complexes are homeomorphisms, rather than just continuous maps. The proof requires a theorem of Schoenflies type for the embeddings of the closures \(\operatorname{clos}W^u(v)\); see . For the resulting regular dynamic cell complex [\[C8:eq:2.1.11\]](#C8:eq:2.1.11){reference-type="eqref" reference="C8:eq:2.1.11"} see .\ For a more concrete example, the case of a single closed Sturm 3-ball \[\label{C8:eq:2.1.12} \mathcal{A}_f = \operatorname{clos}W^u(0) = W^u(0) \cup S^2\.\] with trivial equilibrium \(i(0)=3\) has been addressed in . It has been shown how any finite regular cell decomposition of the Schoenflies boundary \(\partial W^u(0) = S^2\) can be realized, topologically, as the dynamic complex \[\label{C8:eq:2.1.13} S^2 = \bigcup_{v \in \mathcal{E} \smallsetminus \{0\}} W^u(v)\] of the remaining equilibria \(v\) with Morse index \(i(v) \leq 2\), in the Neumann class \(f \in \mathcal{N}(x,u,u_x)\).\ A more careful analysis of this example, however, also keeps track of the two-dimensional *fast unstable manifold* \(W^{uu}(0) \subset W^u(0)\) of those solutions \(u(t, \cdot)\) in \(W^u\) which do not decay to \(v\) at the slowest possible exponential rate in \(W^u(0)\), for \(t \rightarrow-\infty\). By our Schoenflies results \(\partial W^{uu}(0)=S^1\) decomposes \(S^2\), as an "equator", into two (closed) hemispheres \(S_\pm^2\). Each closed hemisphere, separately, is realizable as a planar Sturm attractor \(\mathcal{A}_\pm = \mathcal{A}_{f_\pm}\) for some nonlinearities \(f_\pm \in \mathcal{N}(x,u,u_x)\). To prove our Sturm realization [\[C8:eq:2.1.13\]](#C8:eq:2.1.13){reference-type="eqref" reference="C8:eq:2.1.13"} in  we took one of the hemispheres to be a single closed 2-cell. The general question of admissible, i.e. Sturm realizable, hemisphere decompositions is unexpectedly delicate; see our work in progress . As a caveat, we just mention that the stable extreme equilibria \(v_1, v_{27}\) in any dissipative Sturm realization of the 3d solid octahedron complex \(\mathbb{O}\) of 27 equilibria must be chosen as adjacent corners of \(\mathbb{O}\), rather than antipodally. For adjacent corners \(v_1, v_{27}\), the two hemispheres \(S_\pm^2\) are allowed to decompose \(S^2\) into 1+7 or 2+6 octahedral faces, but not into 3+5 or 4+4 faces, by the fast unstable manifold \(W^{uu}(0)\).\ To complete our ambitious goal of a complete description of the Sturm regular dynamic complexes [\[C8:eq:2.1.11\]](#C8:eq:2.1.11){reference-type="eqref" reference="C8:eq:2.1.11"}, an inductive process comes to mind. Suppose all Sturm global attractors \(\mathcal{A}_f\), \(f \in \mathcal{N}(x,u,u_x)\), have been understood for dimensions up to \(n-1\). Two steps remain. In a first step we have to understand the sphere boundary of single Sturm balls \(\mathcal{A}=\operatorname{clos}W^u(0)\), for \(i(0)=n\), as gluings of closed hemisphere complexes \(\mathcal{A}_\pm\) of dimension \(n-1\), which are assumed to be known by induction hypothesis. In a second step we have to understand how these new \(n\)-cells attach to each other, and to lower-dimensional cells, to form a general Sturm global attractor of dimension \(n\). Again, the respective hemisphere decompositions should play a crucial role here. At present, we have understood step 2 only up to the planar case \(n=2\); see . Step 1 for \(n=3\) is imminent . ## Meanders {#subsec:meanders} In the previous section, we introduced meander curves in the plane as shooting curves of parabolic PDEs [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} in one space dimension; see in particular [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"}, [\[C8:eq:2.1.5\]](#C8:eq:2.1.5){reference-type="eqref" reference="C8:eq:2.1.5"} and Figs. [\[C8:fig:2.1.1\]](#C8:fig:2.1.1){reference-type="ref" reference="C8:fig:2.1.1"}, [\[C8:fig:2.1.2\]](#C8:fig:2.1.2){reference-type="ref" reference="C8:fig:2.1.2"}. Quite independently from this ODE and PDE context, meanders also arise as trajectories of Cartesian billiards , and as representations of monomials in Temperley--Lieb algebras . For simplicity we now represent meander curves, or equivalently meander permutations, by \(\left[N/2\right]\) standardized upper and lower half-circular arches joining at vertices \(1,\dots,N\) along the horizontal axis. Also we replace the open meanders from south-west to north-east, of Section [2.1](#subsec:sturmglobalattractors){reference-type="ref" reference="subsec:sturmglobalattractors"}, by their closed Jordan curve counterparts. See Fig. [\[C8:fig:2.1.1\]](#C8:fig:2.1.1){reference-type="ref" reference="C8:fig:2.1.1"}(c) for one simple closing procedure.\ More generally, we now regard closed meanders \(\mathcal{M}\) as the pattern created by one or several disjoint closed smooth Jordan curves in the plane as they intersect a given horizontal line transversely. Let \(c(\mathcal{M}) \geq 1\) count the connected components of \(\mathcal{M}\). In view of Sturm global attractors, of course, the case \(c(\mathcal{M}) =1\) of connected meanders generated by a single closed Jordan curve is of primary interest.\ Several new results have been obtained which are based on rainbows. Here a *proper upper rainbow* consists of a number \(\alpha\) of concentric and adjacently nested upper arches in any meander. It is understood that \(\alpha\) is chosen maximal, in all cases. An *upper rainbow* consists of one or several neighboring proper upper rainbows. Lower rainbows are defined analogously. Note that upper and lower rainbows may themselves be nested, giving rise to a forest of their sizes \(\alpha\). A closed meander without rainbow nesting is called a *seaweed meander*. Symbolically a seaweed meander can be described as \[\label{C8:eq:2.2.1} \mathcal{M}(\alpha_1,\dots,\alpha_{m_+}\,|\, \beta_1,\dots,\beta_{m_-})\,,\] where \(\alpha_1,\dots, \alpha_{m_+}\) are the sizes of the neighboring proper upper rainbows, and \(\beta_1,\dots, \beta_{m_-}\) enumerate the proper lower rainbows, left to right. See Fig. [\[C8:fig:2.2.1\]](#C8:fig:2.2.1){reference-type="ref" reference="C8:fig:2.2.1"}(a) for an example. Trivially \[\label{C8:eq:2.2.2} \alpha_1+\dots+\alpha_{m_+}=\beta_1+\dots+\beta_{m_-}=N/2\] for any closed meander of \(N\) vertices. The special seaweed case \(m_-=1\) of a single lower rainbow of size \(\beta = N/2\) is called a *bi-rainbow meander*. The case where the upper rainbows may be arbitrarily nested, but there is only one single lower rainbow \(\beta = N/2\), is called *lower rainbow meander*; see Fig. [\[C8:fig:2.2.1\]](#C8:fig:2.2.1){reference-type="ref" reference="C8:fig:2.2.1"}(b). For example any closed meander with \(N\) vertices can be "opened" to a closed lower rainbow meander with \(2N\) vertices and the same number of connected components. Indeed we just duplicate the original vertices \(j \mapsto j,j'\) and open up the horizontal axis to accommodate the vertices in the order \(1,\dots, N,N',\dots, 1'\); see again Fig. [\[C8:fig:2.2.1\]](#C8:fig:2.2.1){reference-type="ref" reference="C8:fig:2.2.1"}. Recall how only connected meanders relate to an interpretation as ODE shooting curves for PDE Sturm global attractors. Lower rainbow meanders of the above general type were related to Cartesian billiards by Fiedler and Castañeda; see  and Fig. [\[C8:fig:2.2.2\]](#C8:fig:2.2.2){reference-type="ref" reference="C8:fig:2.2.2"}.\ In fact, there exists a plane with only horizontal and vertical billiard boundaries, with vertices all on the Cartesian integer grid and billiard slopes of \(\pm 45^\circ\) between half-integer boundary points, such that the connectivities of paths in the billiard and in the meander coincide. As an almost immediate consequence one obtains the following list for the number \(c(\alpha_1,\dots,\alpha_m)\) of connected components of bi-rainbows \(\mathcal{M}(\alpha_1,\dots, \alpha_n\,|\,\beta)\) in terms of greatest common divisors \(\gcd\): \[\label{C8:eq:2.2.3} \begin{aligned} c(\alpha_1) &= \alpha_1\,,\\ c(\alpha_1,\alpha_2) &= \gcd (\alpha_1, \alpha_2)\,,\\ c(\alpha_1, \alpha_2, \alpha_3) &= \gcd (\alpha_1 +\alpha_2, \alpha_2 + \alpha_3)\. \end{aligned}\] The dissertation of Karnauhova  establishes how such formulas cannot be extended to \(m\geq4\). In fact \(c\) cannot be written as the \(\gcd\) of any two polynomials \(f_1,f_2\) in \(\alpha_1,\dots,\alpha_m\) with integer coefficients. In  this obstacle is overcome, replacing the algorithm "\(\gcd\)" by an explicit recursion of related logarithmic complexity in terms of the data \(\alpha_1,\dots,\alpha_m\).\ Encouraged by suggestions of Matthias Staudacher (projects C4 "Structure of Quantum Field Theory: Hopf Algebras versus Integrability" and C5 "AdS/CFT Correspondence: Integrable Structures and Observables") and earlier hints by Eberhard Zeidler, we also began to explore relations between meanders and Temperley-Lieb Hopf algebras \(TL_N(\tau)\); see . Algebraically these are given by \(N\) multiplicative generators \(e_0=1,e_2,\dots,e_{N-1}\) with the three relations \[\label{C8:eq:2.2.4} \begin{aligned} e_i^2 &=\tau e_i\,,\\ e_i e_j &= e_j e_i\,,\\ e_i e_{i \pm 1} e_i &= e_i\,, \end{aligned}\] for all appropriate \(i,j >0\) with \(|i-j|\, >1\). See Fig. [\[C8:fig:2.2.3\]](#C8:fig:2.2.3){reference-type="ref" reference="C8:fig:2.2.3"} for a representation of the Temperley--Lieb generators \(e_i \in TL_N(\tau)\) by \(N\)-strand diagrams. Multiplication of generators is realized by horizontal concatenation of the generator diagrams, up to strand isotopy. Isolas effect multiplication by \(\tau\), each. As illustrated in Fig. [\[C8:fig:2.2.4\]](#C8:fig:2.2.4){reference-type="ref" reference="C8:fig:2.2.4"} for \(e\):\(=e_2e_1e_3 \in TL_4(\tau)\), Temperley--Lieb monomials \(e\) correspond to lower rainbow meanders \(\mathcal{M}[e]\). The number \(c\) of connected components of \(\mathcal{M}[e]\) is then related to the *Markov trace* \(\operatorname{tr}_F\) of \(e\): \[\label{C8:eq:2.2.5} \operatorname{tr}_F(e) = \tau^{c(\mathcal{M}[e])}\,,\] see . For further comments and a relation to the classical Young-Baxter equations see also the dissertation by Karnauhova  and the references there.\ The case [\[C8:eq:2.2.1\]](#C8:eq:2.2.1){reference-type="eqref" reference="C8:eq:2.2.1"}, [\[C8:eq:2.2.2\]](#C8:eq:2.2.2){reference-type="eqref" reference="C8:eq:2.2.2"} of seaweed meanders \(\mathcal{M}(\alpha_1,\dots, \alpha_{m_+}\,|\, \beta_1,\dots,\beta_{m_-})\) is related to the index of seaweed algebras, as studied by Dergachev and Kirillov . The seaweed algebra of type \((\alpha_1,\dots,\alpha_{m_+}\,|\, \beta_1,\dots,\beta_{m_-})\) is the \(\operatorname{gl}(N/2)\) subalgebra of matrices which preserve the vector spaces \(V_j\) spanned by unit vectors \(e_1,\dots,e_{\alpha_1+\ldots+\alpha_j}\) and \(W_j\) spanned by \(e_{\beta_1+\ldots+\beta_j+1},\dots,e_{N/2}\). In fact the seaweed meander corresponds to the Cartesian billiard on the domain of admissible nonzero entries of the seaweed matrices. The index of the algebra is the minimal kernel dimension of the map \(A \mapsto f[A,\cdot]\), for any linear functional \(f\) in the dual algebra. By  this index coincides with the number \(c(\mathcal{M})\) of connected components of the meander \(\mathcal{M}\).\ In her dissertation, Karnauhova considers the maximally disconnected case \(\alpha_j=\beta_j\) of seaweed algebras with maximal index \(N/2\), generated by trivial \(2\)-cycles of \(\sigma\), for the closed meander \(\mathcal{M}\); see . She then shifts the lower \(\beta\)-rainbows one step to the right. Ratcheting into the upper \(\alpha\)-rainbows, she obtains a dissipative Morse meander \(\sigma_f\), and hence a Sturm global attractor \(\mathcal{A}_f\). The Chafee--Infante attractors, for example, which arise for symmetric cubic nonlinearities \(f\) and have been the first Sturm global attractors studied historically, correspond to \(m_\pm=1\) and \(\alpha=\beta\). They are the Sturm global attractors of maximal dimension \(\alpha =N/2\) for any given (odd) number \(N+1\) of equilibria. The same construction in fact works for general, but identical, upper and lower arch configurations. Although the resulting global attractors are all pitchforkable, and thus have been known "in principle" by Fusco and Rocha , the new approach is elegant and geometric: the Sturm global attractors can be obtained explicitly by successive double cone suspensions and elementary gluing concatenations.\ In , Fiedler embarked on a different approach to explore the relation between the chaotic dynamics of (\(x\)-periodically forced) equilibrium equations [\[C8:eq:2.1.2\]](#C8:eq:2.1.2){reference-type="eqref" reference="C8:eq:2.1.2"} with nonlinearity \(f\) and the Sturm permutations \(\sigma_f\) of the resulting shooting meanders. Homoclinic tangles with their associated shift dynamics are taken as a paradigm. In this direction, meander permutations have been defined for standard hyperbolic linear \(SL_2(\mathbb{Z})\) torus automorphisms. The meander permutations compare the orderings of homoclinic points along the one-dimensional stable and unstable manifolds of their limits, respectively. As such, they are new topological invariants--with intriguing relations to continued fraction expansions and quadratic number fields. ## Stability, instability, and obstacles {#subsec:statsol} Consider the higher-dimensional analogue of [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}, that is, for semilinear equations \[u_t=\Delta u+f(u)\] on a domain \(\Omega\subseteq{\mathbb R}^N\) with \(N\ge1\). More generally consider systems \(u\colon\Omega\to{\mathbb R}^n\) \[\label{C8:eq:semilinear} u_t=Lu+f(u)\] with an elliptic differential operator \(L\) and associated Dirichlet, Neumann, or mixed boundary values. Then the naturally associated function spaces \(W_\Gamma^{1,2}(\Omega)\) and \(L^2(\Omega)\) are not continuously embedded into \(C(\overline\Omega)\). Therefore, differentiability of the corresponding superposition operator \(F(u)(x)=f(u(x))\) fails in \(L^2(\Omega)\) (except in the linear case). This imposes an unrealistically restrictive growth condition for \(f\) in case \(F\colon W_\Gamma^{1,2}(\Omega)\to L^2(\Omega)\). Moreover, no analogue of the zero number [\[C8:eq:2.1.7\]](#C8:eq:2.1.7){reference-type="eqref" reference="C8:eq:2.1.7"} with a property like [\[C8:eq:2.1.8\]](#C8:eq:2.1.8){reference-type="eqref" reference="C8:eq:2.1.8"} is available. Nevertheless, we are interested in the long-term behavior of the equation in these two function spaces.\ A particularly natural question is whether a "formally" linearized stability of some stationary solution \(u_0\) implies asymptotic stability in these spaces. In other words, suppose the spectrum of the "formal" linearization \(L+f'(u_0)\) lies in the complex left half-plane. Do all solutions starting close to \(u_0\) in the topology of the space \(W_\Gamma^{1,2}(\Omega)\) or \(L^2(\Omega)\) indeed remain close to \(u_0\) and eventually tend to \(u_0\) in the corresponding topology? Let \(A\) denote the operator naturally associated to \(-L\) and the boundary conditions. Suppose \(A\) solves Kato's square root problem, that is, if \(A^{1/2}\colon W_\Gamma^{1,2}(\Omega)\to L^2(\Omega)\) is an isomorphism. Then \(W_\Gamma^{1,2}(\Omega)=D(A^{1/2})\). The asymptotic stability in \(W_\Gamma^{1,2}(\Omega)\) is known for \(F\colon W_\Gamma^{1,2}(\Omega)\to L^2(\Omega)\); see e.g. . Gurevich and Väth  have relaxed this hypothesis to a subcritical growth condition on \(f\). Moreover, they obtained an asymptotic stability result in \(L^2(\Omega)\). The method of proof was to consider extrapolated spaces and extrapolated semigroups which were obtained by two different procedures. As a side result, they have shown that these extrapolation procedures coincide if and only if \(A\) solves Kato's square root problem.\ One particular motivation to consider the topology of the spaces \(W_\Gamma^{1/2}(\Omega)\) and \(L^2(\Omega)\) is that the following question so far can be addressed only in these topologies. Suppose that the problem [\[C8:eq:semilinear\]](#C8:eq:semilinear){reference-type="eqref" reference="C8:eq:semilinear"} is endowed with some obstacle, i.e., the values of \(u\) on some parts of \(\Omega\) or of its boundary are forced to stay within a certain cone by some constraint. One of our surprising results is that the principal of linearize asymptotic stability may then fail in \(W_\Gamma^{1,2}(\Omega)\). The proof is based on an instability criterion for variational inequalities and a rather general formula relating the fixed point index of the flow of the obstacle problem with the topological degree of a certain associated operator . For specific examples see . ## Symmetric and asymmetric black hole data {#subsec:blackholedata} The Cauchy initial-value problem for the relativistic Einstein equations requires initial data which satisfy the Einstein constraint. For black holes these constraints take the form of an equation for the metric \(g\) with prescribed scalar curvature \(R\) on an asymptotically flat Riemannian \(3\)-manifold \(M\); see Fig. [\[C8:fig:2.3.1\]](#C8:fig:2.3.1){reference-type="ref" reference="C8:fig:2.3.1"} for an illustration and  for a general background. For simplicity we consider a collar region \(M= [r_0, r_1] \times S^2\) with standard \(2\)-sphere cross section. In a suitable coordinate foliation, the metric can be expressed as \[\label{C8:eq:2.3.1} g=u^2dr^2 + r^2\omega\] with the standard metric \(\omega\) on \(S^2\). For example consider prescribed scalar curvatures \(R\) such that \(\lambda\):= \(r^2 R-2\) is a constant bifurcation parameter. Then \(u=u(r,x)\) can be rescaled to \[\label{C8:eq:2.3.2} u(r,x) = \big( \tfrac{\lambda}{2} (\tfrac{1}{r}-1) \big)^{-1/2} \big( v \big(-\tfrac{1}{2} \log (1-r),x\big) +1 \big)\,,\] and the Einstein constraints on the radial metric coefficient \(u\) take the quasilinear parabolic form \[\label{C8:eq:2.3.3} v_t = (1+v)^2 \big( \Delta v + \lambda f(v)\big)\.\] Here \(v = v(t,x)\), \(x \in S^2\), \(\Delta\) is the Laplace-Beltrami operator on \(S^2\), and \(f(v) = v-\tfrac{1}{2}v^2/(1+v)\) defines a cubic nonlinearity.\ In , we obtain spherically anisotropic solutions \(v\) for isotropically prescribed scalar curvature \(R\). Equilibria \(v(b,x)=v_*(x)\not\equiv\operatorname{const}\) provide self-similar anisotropic metrics \(g\). Heteroclinic trajectories \(u(t,x)\rightarrow v_\pm(x)\), for \(t \rightarrow\pm\infty\), are asymptotically self-similar for \(r \rightarrow 0\) and \(r \rightarrow 1\). The metric \(g\) remains smooth, in fact, in the latter limit. We obtain these solutions by an \(O(3)\) symmetry breaking bifurcation analysis of the scaled quasilinear PDE in a center manifold at the trivial isotropic equilibrium \(v_*\equiv0\).\ Symmetry breaking bifurcations occur at \(\lambda_\ell = \ell(\ell+1)\), for \(l=0,1,2,\dots\), in the spherical harmonics kernels \(V_\ell\) of dimension \(2\ell+1\). See Fig. [\[C8:fig:2.3.2\]](#C8:fig:2.3.2){reference-type="ref" reference="C8:fig:2.3.2"} for the bifurcating branches of nonconstant equilibria \(v\), and Fig. [\[C8:fig:2.3.3\]](#C8:fig:2.3.3){reference-type="ref" reference="C8:fig:2.3.3"} for their remaining reduced, nonspherical spatial isotropies. Heteroclinic trajectories follow, between \(v_*\) and \(0\), by standard bifurcation theory in the one-dimensional isotropy subspaces of each kernel \(V_\ell\). For heteroclinic trajectories between equilibria of different isotropies; see also .\ The results above are of a local nature, near the spherically isotropic trivial equilibrium \(v_* \equiv 0\). The fully isotropic case \(\ell=0\) was treated in  previously, by more direct methods. In practice specific computations, particularly of the unstable dimensions and the bifurcation directions, are limited to low \(\ell\) and have been carried out for \(\ell\leq4\), only. Axisymmetric solutions \(v=v(t,\vartheta)\), however, which depend only on the azimuthal angle \(\vartheta\) of \(x \in S^2\), are known to bifurcate for all \(\ell=0,1,2,\dots\). Clearly [\[C8:eq:2.3.3\]](#C8:eq:2.3.3){reference-type="eqref" reference="C8:eq:2.3.3"} becomes "spatially" one-dimensional, \[\label{C8:eq:2.3.4} v_t=(1+v)^2 \big(v_{\vartheta \vartheta} + \cot \vartheta\, v_\vartheta + \lambda f(v) \big)\,,\] with Neumann boundary on \(0 < \vartheta < \pi\). The dissertation of Lappicy explores these versions of quasilinear Sturm parabolic equations with a singular spatial coefficients, by the global methods of Section [2.1](#subsec:sturmglobalattractors){reference-type="ref" reference="subsec:sturmglobalattractors"}; see . In particular, this yields a plethora of axisymmetric, but not fully spherically isotropic, equilibrium solutions and their heteroclinic trajectories as parts of black hole initial data which satisfy the Einstein constraint. # Blow-up and Bianchi cosmology ## Dynamics at infinity and blow-up in complex time In the previous Section [2](#sec:globalattractors){reference-type="ref" reference="sec:globalattractors"}, a class of dissipative equations [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"} was studied, where longtime existence and boundedness of the solutions was guaranteed. In the present section we focus on equations which develop singularities in finite or infinite time. In other words, some trajectories escape to infinity. In order to study the dynamics at infinity, one possible approach uses *Poincaré "compactification"*: the solution space is projected centrally onto a tangent upper hemisphere. More precisely the hemisphere at its north pole is tangent to the solution space at its origin. Arbitrarily far points in the solution space are thereby mapped to the equator of the hemisphere, also called the *sphere at infinity*. For details see . We keep the name Poincaré "compactification" which originates from ODE theory, where the solution space is \(\mathbb{R}^n\) and the compactified space is a compact hemisphere. In the PDE context, the solution space is infinite dimensional and its "compactification" is a bounded hemisphere, which is not compact.\ As an elementary but instructive example, consider the linear equation: \[\label{C8:linearreacdiff} u_t=u_{xx}+bu\,,\\ \] for \(0<x<\pi\), with Neumann boundary conditions; see . The equation [\[C8:linearreacdiff\]](#C8:linearreacdiff){reference-type="eqref" reference="C8:linearreacdiff"} is not dissipative for positive \(b\): some trajectories escape to infinity, in infinite time. The class of equations at the boundary between dissipativity and finite time blow-up is called *slowly nondissipative* in : they admit at least one immortal trajectory escaping to infinity in forward time. The linear example [\[C8:linearreacdiff\]](#C8:linearreacdiff){reference-type="eqref" reference="C8:linearreacdiff"} admits only one bounded equilibrium, located at the origin, and generates infinitely many equilibria at infinity. Those are located at the intersections of the one-dimensional eigenspaces, associated to the eigenfunctions of the Laplacian, with the sphere at infinity. Let us denote the equilibria at infinity by \(\pm\Phi_m\), \(m\in\mathbb{N}\). For all \(b\neq n^2\) which are not square integers, all equilibria, both bounded and at infinity, are hyperbolic. Furthermore the dynamics within the sphere at infinity has a Chafee--Infante structure; see Section [2.1](#subsec:sturmglobalattractors){reference-type="ref" reference="subsec:sturmglobalattractors"}. More precisely the two equilibria \(\pm\Phi_0\) at infinity are stable. The equilibria \(\pm\Phi_{m}\) at infinity each possess an \(m\)-dimensional unstable manifold that consists of heteroclinic orbits to the equilibria \(\pm\Phi_{j}\), \(j\in\{0,\dots,m-1\}\). The unstable manifold of the origin is of finite dimension \([\sqrt{b}]+1\), where \([\sqrt{b}]\) denotes the integer part of \(\sqrt{b}\). This unstable manifold consists of heteroclinic orbits to the equilibria \(\pm \Phi_i\), \(i\in\{0,\dots,[\sqrt{b}]\}\). The stable manifold of the equilibrium at the origin is infinite dimensional and consists of heteroclinic orbits from the equilibria \(\pm\Phi_m\) at infinity, \(m>[\sqrt{b}]\), to the origin. This describes the global unbounded attractor for [\[C8:linearreacdiff\]](#C8:linearreacdiff){reference-type="eqref" reference="C8:linearreacdiff"}.\ When [\[C8:linearreacdiff\]](#C8:linearreacdiff){reference-type="eqref" reference="C8:linearreacdiff"} is augmented by a smooth bounded nonlinearity \(g=g(u)\), \[\label{C8:linreacdiff+bd} u_t=u_{xx}+bu+g\,,\\ \] the dynamics at infinity remains essentially unchanged. In her dissertation , Ben-Gal adapts the techniques of Section [2.1](#subsec:sturmglobalattractors){reference-type="ref" reference="subsec:sturmglobalattractors"} to study the global heteroclinic structure of [\[C8:linreacdiff+bd\]](#C8:linreacdiff+bd){reference-type="eqref" reference="C8:linreacdiff+bd"}. In particular, she proves the existence of a finite-dimensional unbounded inertial manifold containing the so called *unbounded* (or *noncompact*) *attractor*. This object consists of all bounded equilibria (generically hyperbolic), their finite dimensional unstable manifolds, the finitely many equilibria at infinity (also generically hyperbolic) to which those unstable manifolds connect, and the finite dimensional unstable manifolds of those equilibria at infinity. Theorems similar to the dissipative case reveal the heteroclinic structure of the unbounded global attractor of slowly nondissipative equations. Pimentel and Rocha generalized some of these results to bounded nonlinearities \(g(x,u,u_x)\); see .\ We consider blow-up in finite time for [\[C8:eq:2.1.1\]](#C8:eq:2.1.1){reference-type="eqref" reference="C8:eq:2.1.1"}, \(f=f(u)\), next. For smooth right hand side with polynomial growth, time can be rescaled such that trajectories do not run into equilibria at infinity in finite time. The resulting semiflow on the closed Poincaré hemisphere is nonsingular. Conley index theory then provides heteroclinic orbits between equilibria, or general isolated invariant sets. Isolated invariant sets are contained in isolating neighborhoods. An isolating neighborhood is a closed neighborhood whose boundary points eventually leave the neighborhood in forward or backward time direction. For details see  and the references there.\ It turns out that even for some polynomial ordinary differential equations, equilibria at infinity might not be isolated invariant sets. Fig. [\[homflower\]](#homflower){reference-type="ref" reference="homflower"}, left hand side, illustrates such an equilibrium at infinity. This equilibrium is isolated, invariant, and stable within the sphere at infinity. Towards the interior of the Poincaré hemisphere, however, it shows two petals of nested homoclinic loops. Therefore it does not possess small isolating neighborhoods. The complement of any small neighborhood of the degenerate equilibrium, however, is an isolating neighborhood. An invariant set with this property is called of *isolated invariant complement*. Although the complementary invariant set isolated by such a neighborhood is possibly very complex, its Conley index is well-defined. Poincaré duality then defines a meaningful Conley index for the degenerate equilibrium at infinity. An invariant set at infinity attracts some finite initial conditions, if and only if it is not a repeller. This can be characterized in terms of Conley index. Therefore this tool allows us to determine which part of the dynamics in the sphere at infinity is reachable by grow-up or blow-up solutions.\ Via a standard construction illustrated in the right hand side of Fig. [\[homflower\]](#homflower){reference-type="ref" reference="homflower"}, and described in detail in , it is possible to replace the invariant set of isolated invariant complement by a benign *ersatz* with a well-defined Conley index. This ersatz can be used for the analysis of the global heteroclinic structure via classical Conley index theory.\ Based on the idea of self-similar scaling and the Poincaré sphere at infinity, Fiedler has contributed to an unexpected application in a certain renormalization approach to the experimentally observed formation of spin liquids in carbon nanotubes; see . Renormalization has been suggested to overcome the limitations of mean-field descriptions. In its most simple form, homogeneously quadratic ODE systems \[\frac{dx}{dt}=-\nabla_{x}V\] were addressed. The variational structure features a cubic functional \[V(x)=-\frac13 \sum_{i=1}^{N} x_i (x^T A_i x)\] and real symmetric \(N\times N\)-matrices \(A_i\). As a result we obtain self-similar and asymptotically self-similar blow-up solutions, as well as their blow-up rates. These results provide a rational foundation for the renormalization group approach to large, strongly correlated electron systems.\ For fast nonlinear diffusion in several space dimensions the phenomenon of extinction of \(u\), alias blow-up of \(1/u\), has been addressed by Fila and Stuke . As a very simple paradigm for blow-up itself, consider the one-dimensional semilinear heat equation \[\label{C8:eq:nonlinear_heat} u_t=u_{xx}+u^2\] on a bounded or unbounded interval. For real \(u,x\) and *real time* \(t\), this equation has already been well-studied in the literature . The solution is analytic in time and admits a local analytic continuation into the *complex time plane*. For Dirichlet boundary conditions there exists a unique positive equilibrium \(u^*\) which is one-dimensionally unstable. Solutions in the one-dimensional analytic unstable manifold of \(u^*\) possess a real solution with complete blow-up in finite real time. Motivated by analyticity, Stuke studied extensions to complex, rather than real, time \(t\). Of course this entails complex \(u\). Following an idea of Masuda  and Guo at al. , Stuke attempts to bypass real time blow-up at \(t=T\) by an excursion into the complex plane; see Fig. [\[fig:complex_time_domain\]](#fig:complex_time_domain){reference-type="ref" reference="fig:complex_time_domain"}. Stuke proves that solutions on the complex unstable manifold of \(u^*\) stay bounded on time paths \(\Gamma_p^\pm\) parallel to the real axis. The solutions converge to zero: they are heteroclinic from \(u^*\) to zero, along \(\Gamma_p^\pm\), and locally foliate the nonsingular part of the complex one-dimensional unstable manifold \(W_*^u\) of \(u^*\). After some real time \(t=T+\delta>T\), at the latest, the solutions \(\Gamma^{\pm}_c\) extend back to the real axis from, both, the upper and the lower half-plane. Complete real blow-up, however, prevents these conjugate complex analytic branches to coincide, at real time overlap \(t>T+\delta\).\ Traveling pulse waves \(u=U(x-ct)\) with wave speed \(c\), and self-similar solutions, open a second approach to blow-up in complex time. Specifically, they correspond to ODE solutions which are homoclinic to zero. The standing wave \(u(t,x)=U(x)=-6/x^2\) with wave speed \(c=0\), for example, is an explicit solution of [\[C8:eq:nonlinear_heat\]](#C8:eq:nonlinear_heat){reference-type="eqref" reference="C8:eq:nonlinear_heat"}. In general, homoclinic traveling waves \(u=U(\psi)\), \(\psi=x-ct\), possess local expansions in \(\psi\) and \(\log\psi\).\ For the branched real blow-up on the complex unstable manifold \(W^u_*\) of \(u^*\), matters are not quite that simple. Indeed a \"center manifold\", rather than a fully hyperbolic structure, appears in rescaled variables. This heralds essential singularities beyond \(\log t\), and sectors with exponentially flat correction terms. For further details see the dissertation .\ ## The tumbling universe Arguably, the big-bang is the largest imaginable blow-up. For us it is the initial singularity of a purely gravitational cosmological model. We briefly describe the general setting, and the exact reduction to ODE models of Bianchi class. In the next section we present our own results.\ Cosmological models of general relativity are solutions of the full Einstein equations \[\label{C8:eq:Einstein} \mathrm{Ric}(g)-\tfrac{1}{2} \mathrm{Scal}(g) g \;\;=\;\; T\,,\] which relate the geometry of spacetime--the curvature of a 4-dimensional Lorentzian metric \(g\)--to the matter content encoded in the stress-energy tensor \(T\), with or without a cosmological constant. The Einstein PDEs are of hyperbolic type and are usually coupled to kinetic equations of matter. A complete understanding of the full PDE system [\[C8:eq:Einstein\]](#C8:eq:Einstein){reference-type="eqref" reference="C8:eq:Einstein"} is currently out of reach. Therefore, additional symmetries are often assumed to discuss special solutions, and their perturbations.\ The simplest model--and core of the standard cosmological model still in use--is the Friedmann model of spatially homogeneous and isotropic space-time. This assumption of a 6-dimensional symmetry group allows a reduction of [\[C8:eq:Einstein\]](#C8:eq:Einstein){reference-type="eqref" reference="C8:eq:Einstein"} to a single scalar ODE that determines the expansion rate of the universe. This expansion rate can be compared with measurements of the Hubble constant and with the consequences of large-scale thermodynamics of the matter part.\ The second simplest class are *Bianchi models* of spatially homogeneous but anisotropic spacetime. In other words, the spacetime is assumed to be foliated into spatial hypersurfaces given by the orbits of a three-dimensional symmetry group. Belinskii, Khalatnikov, and Lifschitz (BKL) described the dynamics of cosmological models near the big-bang singularity as vacuum dominated, local, and oscillatory . This suggests to approximate the early universe by the dynamics of homogeneous, but not necessarily isotropic, vacuum spacetime.\ The Einstein system [\[C8:eq:Einstein\]](#C8:eq:Einstein){reference-type="eqref" reference="C8:eq:Einstein"} with a non-tilted, perfect-fluid matter model can be reduced to a 5-dimensional ODE system in expansion-reduced variables, \[\label{C8:eq:Bianchi} \begin{array}{rcl} N_1' &=& ( q-4 \Sigma_+ ) N_1\,, \\ N_2' &=& ( q + 2 \Sigma_+ + 2\sqrt{3}\Sigma_-) N_2\,, \\ N_3' &=& ( q + 2 \Sigma_+-2\sqrt{3}\Sigma_-) N_3\,, \\ \Sigma_\pm' &=& (q-2) \Sigma_\pm-3 S_\pm\,, \end{array}\] with the abbreviations \[\label{C8:eq:BianchiAbbrv} \begin{array}{rcl} S_+ &=& \tfrac{1}{2}\left( \left( N_2-N_3 \right)^2-N_1 \left( 2 N_1-N_2-N_3 \right) \right)\,, \\ S_-&=& \tfrac{1}{2}\sqrt{3} \left( N_3-N_2 \right) \left( N_1-N_2-N_3 \right)\,, \\ q &=& 2 \left( \Sigma_+^2 + \Sigma_-^2 \right) + \tfrac{1}{2}(3\gamma-2)\Omega\,, \\ \Omega &=& 1-\Sigma_+^2-\Sigma_-^2-K\,, \\ K &=& \frac{3}{4} \Big( N_1^2 + N_2^2 + N_3^2-2\left( N_1 N_2 + N_2 N_3 + N_3 N_1 \right) \Big)\. \end{array}\] This system is due to Wainwright and Hsu . The initial big-bang singularity is approached in the rescaled time limit \(t\to-\infty\). The variables \(N_k\) describe the curvature of the spatial hypersurfaces. Their signs determine the Lie-algebra type of the associated spatial symmetry imposed by the homogeneity assumption. Due to Bianchi's classification of three-dimensional Lie algebras--the tangent spaces to the assumed symmetry group--these models are called Bianchi models, although they have been introduced by Gödel  and Taub . The shear variables \(\Sigma_\pm\) relate to the rescaled eigenvalues of the second fundamental form of the spatial hypersurfaces. The matter density \(\Omega\) is nonnegative, and the vacuum boundary \(\Omega=0\) is invariant. The coefficient \(\gamma<2\) determines the equation of state of the perfect fluid, e.g. \(\gamma=4/3\) for radiation and \(\gamma=1\) for dust. See also  for further details on this dynamics approach to cosmology, and  for a survey on Bianchi models and open questions.\ The most prominent features of system [\[C8:eq:Bianchi\]](#C8:eq:Bianchi){reference-type="eqref" reference="C8:eq:Bianchi"} are the *Kasner circle* \(\mathcal{K}\) of equilibria, \[\label{C8:eq:Kasner} \mathcal{K} \;=\; \{\; \Sigma_+^2+\Sigma_-^2=1, \; N_1=N_2=N_3=0 \;\}\,,\] and the invariant *Kasner caps* \[\label{C8:eq:Caps} \mathcal{H}_k^\pm \;=\; \{\; \Sigma_+^2+\Sigma_-^2=1-N_k^2, \; \pm N_k > 0, \; N_{k+1}=N_{k-1}=0, \; k \;\mathrm{mod}\; 3 \;\}\.\] The Kasner caps are filled with heteroclinic orbits between equilibria on \(\mathcal{K}\); see Fig. [\[C8:fig:Caps\]](#C8:fig:Caps){reference-type="ref" reference="C8:fig:Caps"}. Equilibria of [\[C8:eq:Bianchi\]](#C8:eq:Bianchi){reference-type="eqref" reference="C8:eq:Bianchi"} indicate self-similar blow-up in the big-bang limit which occurs at rescaled time \(t\to-\infty\). The projections of the heteroclinic orbits to the \(\Sigma\)-plane lie on straight lines through the corners of a circumscribed equilateral triangle; see Fig. [\[C8:fig:Kasner\]](#C8:fig:Kasner){reference-type="ref" reference="C8:fig:Kasner"}. In reversed physical time, these sequences define the *Kasner map* as an expansive map from the Kasner circle onto its double cover; see Fig. [\[C8:fig:Kasner\]](#C8:fig:Kasner){reference-type="ref" reference="C8:fig:Kasner"} and Section [3.3](#subsec:kasnerdynamics){reference-type="ref" reference="subsec:kasnerdynamics"}.\ Concatenation of the heteroclinic orbits in the Kasner caps, i.e. iteration of the Kasner map, defines a subshift of finite type. Misner  introduced the name *Mixmaster* for the heuristic picture of generic trajectories following these formal heteroclinic sequences close to the big-bang singularity: the universe tumbles between various close-to self-similar solutions given by the various Kasner equilibria.\ A crucial question is the *locality* of the BKL/Mixmaster picture. In the (non-vacuum) Friedmann model, the particle horizon (backward light cone) remains finite in expansion reduced variables, close to the big-bang singularity. This bounded particle horizon also bounds the domain of dependence by particle interaction, and therefore represents locality. In the Bianchi model [\[C8:eq:Bianchi\]](#C8:eq:Bianchi){reference-type="eqref" reference="C8:eq:Bianchi"}, the particle horizon corresponds to the *Mixmaster integral* \[\label{C8:eq:MixIntegral} I \;\;=\;\; \int_{-\infty}^0(\sqrt{N_1N_2}+\sqrt{N_2N_3}+\sqrt{N_1N_3}) \mathrm{d}t\,,\] see . Locality holds true, if and only if this integral remains finite. Obvious counterexamples are the solutions on the line of Taub equilibria \[N_1=0\ne N_2=N_3\,,\qquad \Sigma_-=0\,,\quad\Sigma_+=-1\,,\qquad \Omega=0\,,\] where the Mixmaster integral is \(+\infty\). It is currently not known, whether nontrivial counterexamples with \(I=\infty\) exist. The first nontrivial positive examples with \(I<\infty\) are quite recent . More precisely, the longstanding BKL/Mixmaster conjecture therefore claims that \(I<\infty\) for generic initial conditions \(N_j(0)\), \(\Sigma_{\pm}(0)\) in the respective Bianchi class . Here *genericity* can be either understood as *residual*, in the topological sense of Baire category, or as *almost everywhere* with respect to Lebesgue measure. In the Bianchi type IX of class A, where all \(N_j>0\), Ringström  has shown Baire and Lebesgue generic convergence to the Kasner caps. Of course, the Mixmaster integral \(I\) vanishes on the Kasner caps. The mere generic convergence to the Kasner caps, however, does not imply generic finiteness of \(I\) if this convergence is too slow. ## Kasner dynamics {#subsec:kasnerdynamics} In the vacuum model [\[C8:eq:Bianchi\]](#C8:eq:Bianchi){reference-type="eqref" reference="C8:eq:Bianchi"}, \(\Omega=0\), each non-Taub Kasner equilibrium possesses one trivial eigenvalue zero of the linearization in the direction of the Kasner circle and three nontrivial eigenvalues in the remaining directions of the three heteroclinic caps. At the Taub points, two nontrivial eigenvalues become zero and a *bifurcation without parameters* occurs, in the sense of the habilitation thesis  of Stefan Liebscher.\ To clarify the relation of generic trajectories to formal heteroclinic sequences, we ask for the set of points converging to the heteroclinic sequence attached to a given Kasner equilibrium. This is a difficult question because each heteroclinic step, by itself, already consumes infinite rescaled time \(t\). We might hope for an invariant codimension-one foliation transverse to the Kasner circle, which in particular contains the local strong unstable and strong stable invariant manifolds of each Kasner equilibrium. Each fiber would then become the stable set of the Kasner equilibrium, in backwards time \(t \to-\infty\), as the big-bang singularity is approached.\ However, there are topological obstructions: the sets of initial conditions converging to the Taub points, backwards, were classified in  and form manifolds of codimension two. Thus, if the heteroclinic sequence of a Kasner equilibrium terminates at a Taub point, this equilibrium cannot possess an attached stable fiber of codimension one. In particular, any transverse backwards stable foliation has to miss at least a dense subset of the Kasner circle.\ Stable foliations along Kasner sequences away from Taub points have first been addressed by Liebscher et al. in , as follows. Consider a Kasner equilibrium such that its attached (unique, backwards) heteroclinic sequence does not accumulate at a Taub point. Then this equilibrium possesses a unique, locally invariant, codimension-one, stable Lipschitz leaf. The leaves of equilibria in the same heteroclinic sequence form a backwards invariant continuous foliation. The distance in which the trajectories pass the equilibria in this foliation, tends to zero exponentially.\ The proof basically consists of three parts. First, a suitable eigenvalue property guarantees an arbitrarily strong contraction, transversely to the Kasner circle, and an arbitrarily small drift along the Kasner circle, locally. Second, the global passage along a heteroclinic orbit in the Kasner caps is modelled by a diffeomorphism close to the Kasner map. This provides a bounded expansion along the Kasner circle, away from Taub points. Third, the combination of both maps yields a hyperbolic map between suitably chosen cross sections near the Kasner circle: contraction transverse to the Kasner circle is inherited from the local passage, expansion along the Kasner circle is inherited from the global passage. The stable leaf is then obtained as the fixed point of a graph transform, for the above construction.\ In , Liebscher et al. generalize this approach to weaker eigenvalue conditions. This applies to non-vacuum Bianchi systems with arbitrary ideal fluids, and to Einstein--Maxwell equations of Bianchi type VI\({}_0\). In the Einstein--Maxwell system one of the heteroclinic caps is no longer induced by the geometry but rather by the dynamics of the magnetic field.\ An alternative approach in , and the dissertation  by Buchner, utilize Takens linearization  to study the local passage. Indeed, after linearization, contraction and absence of drift follow trivially. However, the linearization requires strong spectral non-resonance conditions. Non-resonance imposes severe limitations on this approach and on possible generalizations. In , for example, all formal sequences which are periodic or accumulate to periodic sequences (or the Taub points) are excluded to avoid resonances. In , Buchner has studied the non-resonance conditions more carefully, and certain periodic sequences are now admissible. His results on local passages extend to Bianchi class-B systems.\ The techniques described above require to avoid the Taub points. This is their main restriction. In fact, the Kasner equilibria which, under iteration of the Kasner map, do not accumulate at Taub points form an (albeit uncountable) meager set of measure zero. They are not generic, in any sense, but possess a finite Mixmaster integral [\[C8:eq:MixIntegral\]](#C8:eq:MixIntegral){reference-type="eqref" reference="C8:eq:MixIntegral"} and thus a finite particle horizon.\ Brehm, in his dissertation , is the first to answer the BKL/Mixmaster question in the Lebesgue sense. This problem had remained open for more than four decades. Brehm obtains two new results in the study of the Wainwright--Hsu system [\[C8:eq:Bianchi\]](#C8:eq:Bianchi){reference-type="eqref" reference="C8:eq:Bianchi"}, [\[C8:eq:BianchiAbbrv\]](#C8:eq:BianchiAbbrv){reference-type="eqref" reference="C8:eq:BianchiAbbrv"} for the vacuum case \(\Omega=0\).The first result extends Ringström's attractor result to Bianchi type [VIII]{.smallcaps} vacuum solutions. The second result asserts finiteness \(I<\infty\) of the Mixmaster integral [\[C8:eq:MixIntegral\]](#C8:eq:MixIntegral){reference-type="eqref" reference="C8:eq:MixIntegral"} for Lebesgue almost all initial conditions.\ Consider the Bianchi type [IX]{.smallcaps} first, where all \(N_j>0\) initially (and then for all times). The *Taub space* \(\{N_2=N_3,\, \Sigma_-=0\}\) is time invariant, together with its two rotated "cousins" given by cyclic permutations of the indices \(\{1,2,3\}\). Ringström  has shown convergence to the Kasner caps, for initial conditions outside the Taub spaces, in the time-rescaled big-bang limit \(t\to-\infty\). The proof proceeded by averaging over rotations around the invariant Taub-spaces, which are known explicitly.\ The Bianchi type [VIII]{.smallcaps} is defined by \(N_2,N_3>0>N_1\) (up to index permutations). The above Taub-space remains invariant, but its two cousins with \(N_1=N_2\) and \(N_1=N_3\) are now incompatible with \(N_2,N_3>0>N_1\) and have disappeared. Even worse: all solutions with \(N_3=0\) and \(N_2>0>N_1\) converge to the Kasner-cicle \(\mathcal K\) in *both* time-directions \(t\to\pm\infty\) and fail to lie on the Kasner-caps. This poses a serious obstacle to the attractivity of the caps. For these reasons, attractor results have remained elusive in the Bianchi [VIII]{.smallcaps} setting.\ For both Bianchi [VIII]{.smallcaps} (\(N_2,N_3>0>N_1\)) and Bianchi [IX]{.smallcaps} (\(N_1,N_2,N_3>0\)), the attractor problem for Bianchi class A vacuum (\(\Omega=0\)) has been settled by Brehm in his dissertation  as follows. The dynamics in the big-bang limit \(t\to-\infty\) falls into one of the following three mutually exclusive classes: - The solution converges to the three Kasner caps and has at least one \(\alpha\)-limit point on the Kasner-circle, which is distinct from the three Taub-points \(T_j\) of Fig. [\[C8:fig:Kasner\]](#C8:fig:Kasner){reference-type="ref" reference="C8:fig:Kasner"} and their antipodal points \(Q_j=-T_j\). - The solution is contained in the invariant Taub subspace \(\{N_2=N_3,\, \Sigma_-=0\}\), or, in the case of Bianchi [IX]{.smallcaps}, one of its two cousins. - The solution has exactly two \(\alpha\)-limit points on the Kasner-circle, which are \(T_2\) and \(Q_2=-T_2\). It has \(\limsup |N_1N_3|>0 = \liminf|N_1N_3|\), for \(t\to-\infty\), and \(\lim|N_1N_2|=\lim|N_2N_3|=0\). - The analogon of \((c_2)\), with the indices \(2\) and \(3\) interchanged. Case \((a)\) applies for an open set of initial conditions which are a neighborhood of the three caps minus the three Taub spaces. Furthermore, for any solution in case \((a)\), the following non-Mixmaster integral \(J\) remains finite: \[\label{C8:eq:brehm:weak-integral} J=\int_{-\infty}^0 \left(|N_1N_2|+|N_2N_3|+|N_3N_1|\right)\mathrm{d} t < \infty\.\] The cases \((c_2)\) and \((c_3)\) are impossible in the case of Bianchi [IX]{.smallcaps}; whether they can occur at all is currently unknown. Taken together with suitable measure-theoretic estimates, one can see that case \((a)\) applies in a set of initial conditions, both of fat Baire category and of full Lebesgue measure. In other words \((b),(c_2),(c_3)\) occur for a meager set of Lebesgue-measure zero, only. The estimate [\[C8:eq:brehm:weak-integral\]](#C8:eq:brehm:weak-integral){reference-type="eqref" reference="C8:eq:brehm:weak-integral"} for the non-Mixmaster integral \(J\) is novel also in the Bianchi [IX]{.smallcaps} case. It is insufficient to bound the Mixmaster integral \(I\) of [\[C8:eq:MixIntegral\]](#C8:eq:MixIntegral){reference-type="eqref" reference="C8:eq:MixIntegral"}, but nevertheless promising: from \(\lim_{t\to-\infty}|N_iN_j|= 0\), to \(\int_{-\infty}^0|N_iN_j|\mathrm{d} t <\infty\), and finally towards the goal \(\int^0_{-\infty}\sqrt{|N_iN_j|}\mathrm{d} t <\infty\). The proof in Bianchi [IX]{.smallcaps}, where only the \(J<\infty\) estimate [\[C8:eq:brehm:weak-integral\]](#C8:eq:brehm:weak-integral){reference-type="eqref" reference="C8:eq:brehm:weak-integral"} is novel compared to , proceeds along the lines of Ringström, via a refined quantitative averaging over rotations around the three invariant Taub subspaces combined with a more careful gluing of the different averaging regimes. In the Bianchi [VIII]{.smallcaps} case, the two missing invariant Taub subspaces are replaced by \(\{|N_1|=|N_3|,\, \Sigma_-=\sqrt{3}\Sigma_+\}\) and its index-exchanged cousin. These two replacement Taub subspaces are invariant only up to leading order; this suffices for the averaging arguments and yields the trichotomy \((a)\), \((b)\), \((c)\). The second main result in  establishes boundedness \(I<\infty\) of the Mixmaster integral [\[C8:eq:MixIntegral\]](#C8:eq:MixIntegral){reference-type="eqref" reference="C8:eq:MixIntegral"}, for Lebesgue almost every initial condition in Bianchi [VIII]{.smallcaps} and [IX]{.smallcaps} vacuum. In other words, particle horizons form towards the singularity and the blow-up is local in the sense of BKL.\ The above considerations reside within the general relativity (GR) framework of a purely gravitational Einstein universe. A larger framework is provided by the Hořava-Lifshitz models (HL). HL gravity has been proposed as a candidate for the theory of quantum gravity; see  and the references there. The motivation of Hell to study those models in , with Uggla, is the following. HL gravity breaks relativistic first principles and introduces anisotropic scalings between space and time. How do such modifications of the first principles affect the initial singularity and the structure of the BKL conjecture? As an attempt to answer this question, Hell and Uggla address the question of the dynamics on its first level of complexity, i.e. the discrete dynamics contained in the heteroclinic chains arising in HL. Similarly to the Bianchi models for general relativity, it is possible to formulate HL models as a system of ODEs with constraints that admit a circle of equilibria, which we call the Kasner circle by analogy. It corresponds to the Bianchi type I model for HL. On the next level of the symmetry hierarchy, Bianchi type II, we find three invariant hemispheres of heteroclinic orbits that intersect only at their boundary, the Kasner circle, for HL and GR alike. On each hemisphere, the projections of the heteroclinic orbits to the plane of the Kasner circle emanate from a single point, as a straightforward calculation shows. Those three points of emanation form an equilateral triangle whose center coincide with the center of the circle of equilibria: for the HL parameter \(v=1/2\) corresponding to GR, the Kasner circle of equilibria is inscribed in the triangle, while this is not the case any more when the parameter changes to \(v\neq 1/2\); see Fig. [\[HLKasnerdyn\]](#HLKasnerdyn){reference-type="ref" reference="HLKasnerdyn"}. Concatenation of heteroclinic orbits on different hemispheres form heteroclinic chains, just as they did in the GR case. However, the discrete dynamics on the Kasner circle induced by those heteroclinic chains change.\ The BKL conjecture for GR describes the asymptotic behavior towards the big-bang singularity. This conjecture can be seen as an extrapolation from the dynamics of the heteroclinic chains to the continuous dynamics near the invariant hemispheres. Studying the discrete dynamics for the HL Bianchi models reveals how the BKL conjecture should be modified in this context.\ Let us focus on the discrete dynamics induced by the heteroclinic chains, which map each equilibrium of the Kasner circle to the end of the heteroclinic leaving from it. Each point of emanation expands an arc of the Kasner circle enclosed by the two tangents to the circle through this point, by this HL Kasner map. Varying \(v\) amounts to changing the distance between the points of emanation of the heteroclinics and the circle of equilibria. This changes the size of the three expanding arcs and the factor of expansion; see Fig. [\[HLKasnerdyn\]](#HLKasnerdyn){reference-type="ref" reference="HLKasnerdyn"}. Two very different behaviors appear, depending on whether \(v\) is below or above \(1/2\).\ Let us first discuss the case \(1/2<v<1\). As can be seen in Fig. [\[HLKasnerdyn\]](#HLKasnerdyn){reference-type="ref" reference="HLKasnerdyn"}, right, the three expanding arcs are separated by three arcs of stable fix points at which heteroclinic chains end, due to the lack of heteroclinic orbits leaving from there. As proved in , generic initial conditions (both in the sense of Baire and Lebesgue) on the Kasner circle will lead to finite heteroclinic chains that end in a stable arc. Hence an analogue to the BKL conjecture should state that the asymptotic behavior of the system is, in general, convergence to a single equilibrium in a stable arc. This behavior is very different from the GR case where oscillatory chaos is present in the asymptotic behavior towards the big-bang singularity. However, by , not all initial conditions on the Kasner circle lead to finite heteroclinic chains. An invariant Cantor set of measure zero leads to heteroclinic chains that never enter the stable arcs, and the discrete dynamics on this Cantor set is chaotic. The proofs of these facts use symbolic dynamics and equivalence to a subshift of finite type, known to be chaotic. Hence chaos has not completely vanished from the system for \(1/2<v<1\), but it is not generic.\ Let us now discuss the opposite case \(0<v<1/2\). We see from Fig. [\[HLKasnerdyn\]](#HLKasnerdyn){reference-type="ref" reference="HLKasnerdyn"}, left, how the three expanding arcs now overlap, pairwise, while no stable arc is present. Therefore the continuation of a heteroclinic chain that hits the overlap region is not unique: there are two heteroclinics leaving from there. In other words, the map describing the discrete dynamics is not well-defined and the concept of iterated function systems (IFS) comes into play--details can be found in  by Hell. The IFS we consider is a collection of maps whose domains of definition cover the Kasner circle. The notion of chaotic discrete dynamical systems is generalized to IFS using the Hausdorff distance between sets. Iterations of a well-defined map are said to be chaotic if three conditions are fulfilled: sensitivity to initial conditions, topological mixing, and density of periodic orbits. In the case of an IFS, a single initial condition generates several trajectories. Therefore each iteration of an IFS is a set instead of a single state. The three above conditions for chaos translate to IFS when distance between iterations is translated to Hausdorff distance between sets of possible iterations. The IFS arising for the HL Bianchi models with \(0<v<1/2\) are chaotic in this sense, mostly due to the expansion property. Roughly speaking, this means that there are chaotic realizations of the IFS. This property only concerns the heteroclinic chains. The nearby flow will be investigated elsewhere. Even if the HL models with \(0<v<1/2\) are chaotic, like the GR case \(v=1/2\), their chaotic structure is quite different. The concept of a BKL era (i.e. phases of oscillations between two neighboring expanding arcs) and the growth of their length along heteroclinic chains shape the chaos in the GR case, via the continued fraction expansion. In the HL for \(0<v<1/2\), BKL erae do not play such a role: their lengths are bounded above, and the fine structure associated to their rate of growth disappears. The value \(v=1/2\) associated to GR can be seen as a critical transitional value where the dynamics bifurcates from nongeneric chaos to generic chaos. The importance of erae is a feature that does not survive below \(v=1/2\). Hence the perturbation of GR along the one-parameter family of HL will lead to a breakdown of the BKL conjecture.\ This concludes our extended survey of some global dynamics aspects, from geometric and algebraic aspects of parabolic global attractors and their blow-up to black hole data and exact chaotic solutions of the Einstein equations in the setting of Bianchi cosmology. Not only many of the particular problems but, much broader, our overall perspective on these problems we owe, deeply, to the unique framework provided by the Sonderforschungsbereich 647 "Space--Time--Matter" of the Deutsche Forschungsgemeinschaft.

{'timestamp': '2016-07-18T02:09:04', 'yymm': '1607', 'arxiv_id': '1607.04600', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04600'}

# Introduction In the past decade, a new field dubbed topological insulator (TI) has caused enormous attentions in the condensed matter physics. Different from the earliest topological states in quantum Hall effect which requires a strong magnetic field, TI is time-reversal-symmetry protected, and can be characterized by the \(Z_2\) topological invariant. In TIs, the bulk insulating states are accompanied by metallic helical Dirac-like boundary states which are related to its bulk topology and protected by the time-reversal symmetry. With the novel properties of the edge states, such as high mobility, absence of backscattering and spin-polarized conductivity channels, TIs provide a promising platform for realizing new electronics and spintronics applications. The discovery of TIs also results in the realization of quantum anomalous Hall effect. Moreover, the proximity effect between TI and superconductor hosts non-Abelian Majorana fermion and opens a new venue for topological quantum computations. On the materials front, all TI materials are narrow-gap semiconductors with inverted band gap at an odd number of time-reversal-invariant momenta (TRIMs). Two-dimensional (2D) TIs (also known as quantum spin Hall states) range from HgTe/CdTe and InAs/GaSb quantum well structures to the layered honeycomb lattice materials such as silicene, germanene, stanene, their halogenides, and ultrathin Bi films. The 3D TIs materials include Bi\(_{1-x}\)Sb\(_x\) alloys, Bi\(_2\)Se\(_3\)-class materials, half-Heusler compounds, TlBiSe\(_2\) family chalcogenides, strained HgTe, \(\alpha\)-Sn, and bismuth-based III-V semiconductors, etc. Additional ways could convert the normal insulators into TIs, such as external strain, chemical doping. The discovery of TIs stimulates the identification and searching for more topological quantum states. In TIs, the two branches of helical edge/surface states are related to each other by time-reversal symmetry, and they are degenerate at the TRIMs due to the Kramers' theorem thus forming the Dirac points. In principle, degeneracies can also come from other types of symmetries, such as particle-hole symmetry, crystalline symmetry, etc. Therefore, finding new topological phases protected by other symmetries is a new hot topic in this rapidly developing field. In 2011, Fu proposed that crystalline symmetry can protect new types of topological states, which are called topological crystalline insulators (TCIs). The first theoretically predicted and experimentally realized TCI materials are IV-VI semiconductors, with SnTe as a representative. The symmetry responsible for their topological character is the mirror symmetry. Due to the simple crystal structure, SnTe-class TCI materials have caused extensive concerns. Later, some other TCI materials are theoretically proposed in transition metal oxides with a pyrochlore structure and anti-perovskite materials; both of these two classes of materials are also topologically protected by mirror symmetry. Additionally, crystalline symmetry has been extended from the point group to non-symmorphic space group and a new class of TCIs named topological non-symmorphic crystalline insulator is proposed. Considering the richness and complexity of crystal operations, recently some researchers have made great efforts to classify the TCIs systematically. However, these proposals ask for the support of realistic materials, and to date the only TCI materials realized in experiments are the IV-VI semiconductors. In this article, we mainly review the recent research on the SnTe-class TCI materials, which may facilitate future researches to search for more TCI materials and particularly to explore their potential applications. In the research field of topological materials, the theoretical calculations play an important role in the material predictions and provide a reliable guidance for the experimental observations. So in the following (Sections II and III), we will first describe the topological identification of the bulk and surface states properties of SnTe from a computational point of view, and then we introduce the experimental observations. Due to the crystalline symmetry protection and multiple Dirac surface states, TCIs have more tunable properties than TIs. In Sections IV to VI, we will demonstrate the electronic properties of TCIs under various perturbations, carrier types control or superconductivity by chemical doping, strain-tuned surface states and topological phase transition, and thickness-dependent confinement effects on 2D topological states. Some novel quantum transport properties, such as valley-selective filtering and helicity-resolved functionalities for Dirac fermions, are also discussed in Sections V and VI. Finally, we give the conclusions and outlook in Section VII. # Bulk topology of SnTe class of materials SnTe-class IV-VI semiconductors have a rocksalt structure \[Fig. 1(a)\], and its Brillouin zone (BZ) is shown in Fig. 1(b). These materials are narrow-gap semiconductors, with their valence-band maximum (VBM) and conduction-band minimum (CBM) at four equivalent \(L\) points of BZ. In the vicinity of \(L\) point, the band structures of PbTe and SnTe with \(p\) orbital projections of cation and anion based on first-principles calculations are presented in Figs. 1(c) and (d), respectively. It is seen that PbTe has a normal band ordering, i.e. The VBM is primarily derived from Te atoms and the CBM from Pb atoms, suggesting that PbTe can be smoothly connected to the atomic limit. While SnTe has an inverted band structure: the VBM is dominated by the Sn atoms and the CBM by Te atoms. The band inversion of SnTe relative to PbTe indicates that the relative ordering of \(L_{6}^{+}\) and \(L_{6}^{-}\) states is switched, as shown in Figs. 1(c) and (d). Nonetheless, it should be noted that there are four \(L\) points in the whole BZ, leading to that band inversion occurs at an even number of points. So neither SnTe nor PbTe in the rocksalt structure is a TI with non-trivial \(Z_2\) topological invariant. However, the band inversions of SnTe can give new non-trivial topology if we take the crystalline symmetry into account. Face-centered-cubic structure has the mirror symmetry with respect to the {110} planes, one of which is shown by the green plane in Fig. 1(a), corresponding to the plane \(\Gamma L_1L_2\) in momentum space \[Fig. 1(b)\]. The Hamiltonian of this system under mirror operation \(M\) satisfies \(MH(k_1,k_2,k_3)M^{-1}=H(-k_1,k_2,k_3)\), in which \(k_1\) is along the direction perpendicular to the plane \(\Gamma L_1L_2\), and \(k_3\) is along the direction of \(\Gamma L_1\). On plane \(\Gamma L_1L_2\), crystal momenta have \(k_1=0\), and the Hamiltonian commutate with the mirror operator, i.e., \([H,M]=0\). Then these two operator have the common eigenstates. The mirror operation satisfies \(M^2=-1\) for spin 1/2 electrons, and its eigenvalues are \(M=\pm i\). So all the eigenstates of Hamiltonian can be labeled by the eigenvalues \(\pm i\) of \(M\), i.e., \(\Psi_{+i}\) and \(\Psi_{-i}\). The Hamiltonian can be written as the form of block diagonalization, \(H=\left(\begin{smallmatrix} H_{+i} & 0 \\ 0 & H_{-i} \\ \end{smallmatrix}\right)\). Following the definition of spin Chern number, here each class of eigenstates in subspace has an associated Chern number \(C_{\pm i}\). Though the total Chern number is zero, \(C=C_{+i}+C_{-i}=0\), due to the time-reversal symmetry, the mirror Chern number defined as \(C_M=(C_{+i}-C_{-i})/2\) can be nonzero. The crystal may have several mirror planes, and each can define an associated mirror Chern number. A nonzero mirror Chern number defines a topological crystalline insulator protected by the mirror symmetry. Hsieh *et al.* introduced a low-energy effective model near the \(L\) point to discuss the different topology between PbTe and SnTe, and they found that the band inversion of SnTe at \(L\) changes the mirror Chern number by one. Because there are two \(L\) points (\(L_1\) and \(L_2\)) in the \(\Gamma L_1L_2\) plane and these two \(L\) points have the same contributions to the mirror Chern number, finally the band inversion changes the mirror Chern number for the \(\Gamma L_1L_2\) plane by two. As a result, PbTe is topologically trivial, while SnTe is a topological crystalline insulator with the mirror Chern number \(C_M=-2\). Actually, there are six equivalent {110} mirror planes for SnTe, and each possesses two \(L\) points, so the mirror Chern number for every plane is \(-2\). # Surface states of SnTe-class TCIs The bulk-boundary correspondence indicates the existence of topological surface states. Since band inversion occurs at four \(L\) points in bulk BZ of SnTe, there should be four Dirac cones in the surface BZ. However, the gapless surface states of TCIs require the underlying crystalline symmetry to be preserved on the boundary. So not all crystal surfaces of SnTe have gapless surface states; only those surfaces that keep the mirror symmetry with respect to the (110) planes of rocksalt structure are gapless. Three common surface terminations are (001), (111), and (110), and they meet the corresponding symmetry. Interestingly, there are two types of TCI surface states with qualitatively different electronic properties, which depend on the surface orientation and are schematically shown in Fig. 2. The type-I surface states have the properties that their Dirac points are located at TRIMs, such as the (111) surface states; while the Dirac points of type-II surface states on (001) and (110) surfaces are deviated from TRIMs. The latter types of surface states exhibit a Lifshitz transition as a function of Fermi energy. The dependence of boundary states on surface orientations is an important property of TCIs. Next we will first introduce the (001) surface states, and then the (111) surface states. ## (001) surface states The first-principles calculated band structure of (001) surface of SnTe is shown in Fig. 3(a). The red lines depict the topological surface band and they are of Dirac-like linear dispersion. Interestingly, the Dirac point is not located at the time-reversal invariant \(\bar{X}\) point which is the projection of \(L\) point in the bulk BZ, but has a little deviation along \(\bar{\Gamma}\bar{X}\). As shown in Fig. 3(b), all four Dirac points in the whole surface BZ are deviated from \(\bar{X}\), and they are related to each other by the \(C_4\) rotation symmetry. Another important feature is that the change of Fermi surface topology exhibits a Lifshitz transition as a function of the Fermi energy. When the Fermi energy is slightly lower than the Dirac point, the Fermi surface consists of two disconnected hole pockets \[red or orange elliptical region in Fig. 3(c)\] on either side of the \(\bar{X}\) point along \(\bar{\Gamma}\bar{X}\); as the Fermi energy decreases, these two pockets enlarge and touch each other, then they reconnect to form a large hole pocket \[green or blue region in Fig. 3(c)\] and a small electron pocket \[gray elliptical region in Fig. 3(c)\], both centered at \(\bar{X}\). In addition, this transition of Fermi surface is accompanied by a Van Hove singularity in the density of states arising from saddle points in the band structure. This type of TCI surface states can be understood as follows. From Fig. 2, we can see that two \(L\) points (\(L_1\) and \(L_2\)) in bulk BZ of SnTe are projected onto the same time-reversal invariant \(\bar{X}\) point on the (001) surface. As a result of band inversions at these two \(L\) points, two coexisting gapless Dirac cones will be created at \(\bar{X}\). However, the sharp surface introduces intervalley scattering between these two \(L\) points, thus producing the hybridization between the two massless surface Dirac fermions at the surface, and forming the novel surface states at \(\bar{X}\). The essential properties of the (001) surface states can be captured by the following \(k\cdot p\) Hamiltonian: \[H_{\bar{X}}(\textbf{k})=(v_xk_xs_y-v_yk_ys_x)\otimes I+m\tau_x+\delta s_x\tau_y. \label{eq:H}\] Here, the first term describes two identical surface Dirac fermions associated with \(L_1\) and \(L_2\) (denoted by \(\tau_z=\pm 1\)); \(\vec{s}\) is a set of Pauli matrices associated with the spin components; the second and third terms describe all possible intervalley hybridizations to zeroth order in \(\textbf{k}\), which satisfy all the symmetries of the (001) surface. The \(m\) term induces the coupling between the two Dirac cones and produces an relative shift between them in energy; the \(\delta\) term turns the band crossing of the two Dirac cones into an anticrossing via hybridization and opens a gap, which is, however, strictly forbidden along \(\bar{\Gamma}\bar{X}\). Consequently, the bands are gapless along \(\bar{\Gamma}\bar{X}\), but gapped along other directions. The band structure of \(k\cdot p\) model of Eq. ([\[eq:H\]](#eq:H){reference-type="ref" reference="eq:H"}) with 2D momentum near \(\bar{X}\) is shown as an inset in Fig. 3(a). The Hamiltonian above can perfectly repeat the first-principles electronic structure of (001) surface, such as the deviation of Dirac point from \(\bar{X}\), the Lifshitz transition of Fermi surface, and the Van Hove singularity in the density of states. The Lifshitz transition is interesting, because it would accompany a marked change in the Dirac-carrier properties and provide another ingredient in the physics of topological materials. As the cleavage plane, the (001) surface of SnTe and its family materials has been successfully prepared and the corresponding type-II surface states have been confirmed by experiments. Tanaka *et al.* grew the high-quality single crystal of SnTe, and observed the signature of a metallic Dirac-cone surface band by the angle-resolved photoemission spectrum (ARPES), with its Dirac point slightly away from the edge of the surface BZ \[Fig. 4(a)\]. In fact, because of the hole-doped nature of SnTe crystal, the Dirac point of surface states is not detected in their work and its energy is estimated to be 0.05 eV above \(E_{\rm F}\) from a linear extrapolation of the two dispersion branches. In line with the theoretical prediction, they did not observe such gapless surface states in the cousin material PbTe. About the p-type conductivity of SnTe, we will explain its microscopic origin later in the article. For the Pb\(_{1-x}\)Sn\(_x\)Te and Pb\(_{1-x}\)Sn\(_x\)Se alloys, the chemical potential can easily be tuned to yield n-type or p-type conductivity by changing the crystals' growth condition, which makes them more suitable for experimental investigations on the TCI state. For example, Dziawa *et al.* prepared the Pb\(_{0.77}\)Sn\(_{0.23}\)Se single crystal, and the ARPES of its (001) surface shows a n-type nature of this material. Importantly, they demonstrated that this alloy will undergo a temperature-driven topological phase transition from a trivial insulator to a TCI \[Fig. 4(b)\]. Xu *et al.* utilized ARPES and spin-resolved ARPES to study the low-energy electronic structure below and above the band inversion topological transition of Pb\(_{1-x}\)Sn\(_x\)Te (\(x=0.2\) and \(x=0.4\)). They demonstrated the observation of spin-polarized surface states in the inverted regime and their absence in the non-inverted regime, in which the ARPES iso-energetic contour mapping in the first surface BZ and spin-resolved topological Dirac surface states of Pb\(_{0.6}\)Sn\(_{0.4}\)Te are shown in Fig. 4(c). The interplay between topology and crystal symmetry is the character of TCIs. If the mirror symmetry of SnTe is broken, the gapless surface states will be gapped, i.e., the massless Dirac fermion will acquire mass. Okada *et al.* reported high-resolution scanning tunneling microscopy studies of Pb\(_{1-x}\)Sn\(_x\)Se that reveal the coexistence of zero-mass Dirac fermions protected by crystal symmetry with massive Dirac fermions consistent with crystal symmetry breaking. ## (111) surface states Theoretically, the (111) surface states of SnTe are more simple than those of (001) surface. As shown in Fig. 2, four different \(L\) points of bulk BZ are projected to different momenta in surface BZ, i.e., one \(\bar{\Gamma}\) and three \(\bar{M}\) points. There are no interactions between them. In fact, due to the polarity of (111) surface, it is not so easy to achieve the ideal (111) surface and observe its topological surface states in experiments. On the theoretical computations, there are only some tight-binding calculations at first without regard to the surface dangling bonds. Fig. 5 shows the surface band structure with two different terminations of SnTe (111) surface calculated by Liu *et al.* It is seen that Dirac-like surface band crossing (denoted by the red lines) occurs at \(\bar{\Gamma}\) and \(\bar{M}\) for both two terminations, whereas the Dirac points are close to the top (bottom) of the valence (conduction) band for Sn (Te) termination. The Dirac fermion at \(\bar{\Gamma}\) is isotropic: the Fermi velocity along \(\bar{\Gamma}\bar{M}\) is identical to that along \(\bar{\Gamma}\bar{K}\); while at \(\bar{M}\), the Dirac fermion is anisotropic: the Fermi velocity along \(\bar{M}\bar{K}\) is larger than that along \(\bar{M}\bar{\Gamma}\). The \(k\cdot p\) Hamiltonians at \(\bar{\Gamma}\) and \(\bar{M}\) are given by \(H_{\bar{\Gamma}}(\textbf{k})=v(k_1s_2-k_2s_1)\) and \(H_{\bar{M}}(\textbf{k})=v_1k_1s_2-v_2k_2s_1\), where \(k_1\) is along the \(\bar{\Gamma}\bar{K}\) direction and \(k_2\) is along the \(\bar{\Gamma}\bar{M}\) direction. The ideal (111) surface of SnTe-class materials is a polar surface with unpaired surface electrons. In principle, the huge electrostatic potential induced by the dipole accumulation will cause the instability of the surface. Using first-principles calculations, Wang *et al.* studied the (111) surface morphology and the associated electronic structures of SnTe under different growth conditions. They found that three different surfaces can stably exist, with the increasing of chemical potential of Sn: unreconstructed Te-terminated surface, \((\sqrt{3}\times\sqrt{3})\)-reconstructed surface and \((2\times 1)\)-reconstructed surface. The electronic structure of unreconstructed (111) surface shows that the nontrivial surface states are affected by the unpaired electrons, however clean type-I surface states can be achieved by hydrogen adsorption. Importantly, they revealed the critical effect of surface reconstruction on the topological surface states: the reconstruction will result in the folding of surface BZ and probably induce the breaking of the mirror symmetries, both of which can change the characteristics of topological surface states. For example, in Fig. 6(a,b), the \((2\times 1)\)-reconstructed surface has the type-II surface states at \(\bar{\Gamma}\): the Dirac point is deviated from TRIM. The \((2\times 1)\) reconstruction folds the surface BZ (SBZ) from hexagonal SBZ1 to rectangular SBZ2, causing the \(\bar{M_1}\) point in SBZ1 to fold to the \(\bar{\Gamma}\) point in SBZ2. As a result, two different \(L\) points in the bulk BZ are projected to the same momentum \(\bar{\Gamma}\) for reconstructed surface, just like the (001) surface. The additional intervalley scattering and the only one symmetric (110) mirror plane remaining create the type-II surface states at \(\bar{\Gamma}\), which will also undergo a Lifshitz transition when tuning Fermi energy. Meanwhile the mirror symmetry breaking gaps the surface states at \(\bar{M_2}\). In \((\sqrt{3}\times\sqrt{3})\)-reconstructed surface, as shown in Fig. 6(c,d), the reconstruction only folds three different \(\bar{M}\) in SBZ1 to three new \(\bar{M^{\prime}}\) in SBZ3, so four Dirac cones still exist at four TRIMs belonging to type-I surface states. However, due to the different energy of Dirac points at \(\bar{\Gamma}\) and \(\bar{M}\), their conductive properties can be different when changing the Fermi energy. In experiments, the observation of (111) surface states is truly more challenging owing to the difficulty in the preparation of non-cleavage surface. To date, epitaxial layer growth is an effective method for obtaining the (111) film. Fig. 7(a) and (b) show the ARPES measurements on the (111) surface of SnTe and Pb\(_{1-x}\)Sn\(_x\)Te performed by Tanaka *et al.* and Yan *et al.*, respectively. For the pure SnTe, both two experiments demonstrate the p-type conductivity, and the spectrum of surface states is covered by the signals of bulk valence band. Tanaka *et al.* utilized the second derivatives of the momentum distribution curves and changed the photon energies to recognize the linear-dispersion surface band. While Yan *et al.* changed the Pb/Sn ratio and found a topological phase transition in Pb\(_{1-x}\)Sn\(_x\)Te. A clear observation of the (111) surface states is done by Polley *et al.* on the Pb\(_{1-x}\)Sn\(_x\)Se alloys. From Fig. 7(c), the Dirac-like band crossing is unchanged when varying the photo energy, revealing the character of surface band. They also observed that there is no relative binding energy difference between the \(\bar{\Gamma}\) and \(\bar{M}\) Dirac points. In addition, Wang *et al.* obtained the metastable rocksalt SnSe (111) films with the molecular beam epitaxy (MBE), and demonstrated that the epitaxial rocksalt SnSe is also a TCI through observing its (111) surface states. # Native defects and doping in SnTe-class materials As mentioned above and observed by experiments, the as-grown SnTe crystals are usually heavily hole doped, which hinders the detection and investigations of the surface states and possible applications. Meanwhile, PbTe, the cousin material, can be either n or p type, depending on the growth conditions. Wang *et al.* studied the role of native point defects in the doping behavior of these two materials and explained the microscopic origin of the p-type conductivity of SnTe. As shown in Fig. 8(a), the dominant defect in SnTe is the negatively charged cation vacancy (V\(_{\rm Sn}\[^{2-}\)) under both Sn-rich and Te-rich conditions, and V\(_{\rm Sn}\]^{2-}\) has a negative formation energy throughout the bulk gap. As a result, the Fermi level of SnTe has to be pushed under the VBM \[i.e., a minus value of Fermi level in Fig. 8(a)\] until the formation energy of V\(_{\rm Sn}\[^{2-}\) becomes positive. So SnTe is always hole doped. While in PbTe \[Fig. 8(b)\], the dominant defect is not always the cation vacancy (V\(_{\rm Pb}\]^{2-}\)), but could be either anion or cation vacancy V\(_{\rm Te}\[^{2+}\), V\(_{\rm Pb}\]^{2-}\) or antisite Te\(_{\rm Pb}\[^{2+}\), depending on the growth conditions. V\(_{\rm Te}\]^{2+}\) and Te\(_{\rm Pb}\)\(^{2+}\) are positively charged so that PbTe could be electron doped when these defects dominate. The large difference in the nature of defects between SnTe and PbTe originates from the different work functions or the distinction of the VBM of these two materials, as the formation energies of charged defects depend explicitly on the Fermi level referenced to the VBM. The VBM of SnTe is 0.5 eV higher than that of PbTe, inducing that SnTe has a lower formation energy of the negatively charged defects than PbTe. Meanwhile, alloying SnTe with Pb can continuously lower the VBM and CBM, so as to modify the conductivity type. So tuning the Pb/Sn ratio in Pb\(_{1-x}\)Sn\(_x\)Te alloy could obtain either n-type or p-type conductivity, which is a suitable way to realize bulk insulating and observe the topological surface states. When doping the magnetism into thin films of TIs, the topological electronic states with a Chern number \(\pm 1\) can be formed and the quantum anomalous Hall effect (QAHE) can be realized, which is experimentally observed in Cr doped Bi\(_2\)Te\(_3\) class of TIs. Here in TCIs, Fang *et al.* proposed that when an out-of-plane ferromagnetic order is introduced into a TCI film, the QAHE with a tunable large Chern number can be produced. However, magnetic doping in SnTe-class TCIs is not easily achievable. In the early years, some experimental efforts have been made to study various types of magnetism in SnTe crystals doped with 3d transition metals (TMs), which are found to be degenerate magnetic semiconductor with small magnetic moments for these TM atoms. These systems show complex magnetic properties: for example, doping with Mn, Cr and Fe act as ferromagnetic behavior, while doping with Co and Ni are magnetically ineffective. More peculiarly, a strong effect of carrier concentration on the magnetic properties of these diluted magnetic semiconductors has been demonstrated, and it can be understood on the basis of the Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction mechanism. Recently, some calculations investigated the electronic and magnetic properties of V, Cr or Mn-doped SnTe. However their doping concentration is very high, and it is likely to damage the TCI phase of SnTe. Another theoretical study comes to a conclusion that the doped TM atoms in SnTe have comparatively high formation energies, and predicts that the uniform TM doping with a higher concentration in SnTe will be difficult unless clustering. To date, the effective magnetic doping in SnTe-class TCI materials is still a challenge, and the further investigations of magnetic effect on TCIs are required. Another important doping is In-doped SnTe, which hosts the superconducting behavior. In particular, Sn\(_{1-x}\)In\(_x\)Te with an indium content of \(x=0.045\) preserving the topological surface states has been proved a topological superconducting state by the signals of surface Andreev bound states in the point-contact spectroscopy, and the critical temperature is 1.2 K. Furthermore it has recently been reported that the superconducting \(T_{\rm C}\) of Sn\(_{1-x}\)In\(_x\)Te continues to increase with a higher level of doping, such as 4.5 K with In content of \(x=0.45\). However, the carrier density does not vary monotonically with Indium content. In fact at high Indium content, Indium has a mixed oxidation state in the system, i.e. it is neither In\(^{1+}\) nor In\(^{3+}\). So superconducting Sn\(_{1-x}\)In\(_x\)Te can not be viewed as a simple hole doped semiconducting material. # Strain effect Protected by the crystal symmetry and possessing multiple branches of Dirac surface states, TCIs have a much wider range of tunable electronic properties than TIs under various perturbations. Strain is an important controllable method. Not only can strain drive topological phase transition in the bulk system, but it can also modulate the electronic properties of the Dirac surface states. ## Strain effect on bulk: topological phase transition Strain is an effective way to tune the band gap of semiconductors. For the narrow-gap semiconductors whose VBM and CBM have the opposite parity or bonding types, the strain will influence their levels differently, and even induce the band inversion. As has been proposed in Bi\(_2\)Se\(_3\) family compounds, strain can induce the topological phase transition from a normal insulator (NI) to a TI and vice verse. This manipulation can also be applied to TCIs. As mentioned by Hsieh *et al.*, with the lattice constant decreasing, PbTe can be changed from a NI to a TCI. Such topological phase transition also occurs for other lead chalcogenides under external pressure or volume compression; while for the volume expansion, SnS and SnSe can be transformed from an ambient pressure TCI to a topologically trivial insulator. As has been known to us, there are four \(L\) points in the bulk BZ; the isotropic strain can not distinguish them and will change the band orders of four valleys simultaneously. So the isotropic strain, such as hydrostatic pressure for the rocksalt structure, can only induce the topological phase transition between a NI and a TCI. However, an uniaxial or biaxial strain, which could be applied in the epitaxial films, can offer a diverse manipulation of the electronic properties. A systematic study of SnTe with various in-plane biaxial strains perpendicular to the \[001\] direction has been made by Qian *et al.* Since the unit cell fitting for this type of strain is along the \[001\] direction, there is a folding of bulk BZ. As depicted in Fig. 9(a), four \(L\) points in the original BZ (polyhedron encircled by black lines) are folded to \(R\) and \(R^{\prime}\) in the new BZ (blue tetragonal cell). The in-plane biaxial strain does not break the \(C_4\) rotation, so \(R\) and \(R^{\prime}\) are still equivalent. The calculated band gap at \(R\) as a function of biaxial strain is shown in Fig. 9(b). As the strain increases, the energy gap initially decreases, and then opens up again. Above 2.5\(\%\) strain, band inversion disappears, and the system is transformed from a TCI to a NI. A special strain which can distinguish the four \(L\) points is the \[111\]-oriented strain. Zhao *et al.* studied the band evolution of SnTe with the in-plane biaxial strain and out-of-plane uniaxial strain in the \[111\] direction. The unit cell and BZ are shown in Figs. 9(c) and (d). Due to the enlarged unit cell, the BZ shrinks from the face-centered-cubic case to a hexagonal prism. Consequently, the \(L_1\) point in the original BZ folds to the \(A\) point in the new BZ, and the band at \(A\) will change differently from those at other three \(L\) points under strain. The band-gap evolutions at the \(A\) and \(L\) points under the two types of strains (the in-plane biaxial strain and the out-of-plane uniaxial strain) are shown in Figs. 9(e) and (f), respectively. As the in-plane strain increases, though the band gaps at both \(A\) and \(L\) reduce to zero then reopen again, the gap closing will first occur at the \(A\) point. So except for TCI and NI phases, there is a TI phase for the in-plane strain level between +2\(\%\) and +2.5\(\%\). For the out-of-plane strain, the band-gap closing and reopening only occur at the A point, so below +10\(\%\), there are two topological phase: TCI and TI, where the TI phase could appear in the range of \(+5 \% < \varepsilon_{\perp} < +10 \%\). ## Strain effect on (001) surface Here, maintaining the TCI phase, we discuss the effect of strain on the (001) surface states. The Dirac points on the (001) surface of SnTe are not pinned to TRIMs. A mechanical strain can shift the Dirac point positions in the momentum space just like an electromagnetic field acts on an electron. Qian *et al.* studied the strain-dependent electronic properties of SnTe (001) nanomembranes. Here we take the band structure of a 51-layer (001) membrane under the in-plane biaxial strain as an example. As shown in Fig. 10, the compressive strain drives the Dirac points of surface states to move away from the \(\bar{X}\) points, while the tensile strain makes them close to the \(\bar{X}\) points. In addition, the strain will influence the penetration length of the surface states like the case of TIs. Qian *et al.* found that the compressive (tensile) strain can reduce (increase) the surface state penetration length, so the tensile strain will cause a hybridization gap for the SnTe thin film, such as the surface band for the 51-layer membrane under the biaxial strain of +1\(\%\) \[right panel in Fig. 10\]. Such shift of Dirac point position in k space under strain has been realized in experiments, which establishes a tunable platform for the strain-based applications. The biaxial strain will equally shift the four Dirac points of the (001) surface states. However, a structural distortion with a relative displacement of atoms along the \[110\] direction on the (001) surface can break one of two mirror planes, and generate a mass to the Dirac cone. As a result, the two Dirac points along \[\(1\bar{1}0\)\] direction are gapped, but the other two along \[110\] remain gapless, which has been demonstrated by the STM measurements of the coexistence of zero-mass Dirac fermions with massive Dirac fermions. ## Strain effect on (111) surface The Dirac surface states on the (111) surface are noninteracting and located at well-separated TRIMs (i.e., \(\bar{\Gamma}\) and \(\bar{M}\)) in the surface BZ. Most importantly, differing from the valleys on the (001) surface, the \(\bar{\Gamma}\) and \(\bar{M}\) valleys appear highly distinct for crystal symmetry, and thus their band topology is sensitively differentiated by mechanical deformation as mentioned above. In addition, strain effect could also alter the valley-based landscape in the (111) surface. Maintaining the TCI phase of the entire system, Zhao *et al.* studied the strain-dependent Dirac valley evolution on the SnTe (111) surface under different strains. Figs. 11(a--c) show the DFT-calculated surface band structures under no strain, in-plane biaxial strain of-2\(\%\) and out-of-plane uniaxial strain of-2\(\%\), respectively. In the absence of strain \[Fig. 11(a)\], the \(\bar{\Gamma}\) and \(\bar{M}\) valleys are degenerate in energy; while the compressive strain drives the surface bands at \(\bar{\Gamma}\) and \(\bar{M}\) to shift oppositely with reference to the Fermi level \[Figs. 11(b) and (c)\]. In particular, for the out-of-plane compressive strain, the opposite shift of the surface bands at \(\bar{\Gamma}\) and \(\bar{M}\) also applies with reference to the unstrained Fermi level. Such properties can be used to design the strain-engineered nanodevices for dynamic valley control. As depicted in Fig. 11(d), a local compressive strain in the SnTe (111) surface can be generated by applying perpendicular pressure using a piezoelectric-material-based actuator. Under out-of-plane compressive strain, the Dirac cone at \({\bar{\Gamma}}\) shifts downwards, while the cones at \({\bar{M}}\) shift upwards, as schematically illustrated in Fig. 11(e). Therefore, with increasing strain, the Fermi level in the strained region will approach the Dirac point at \(\bar{M}\), then goes across the valence band. The opposite shift in energy of Dirac cones at \({\bar{\Gamma}}\) and \(\bar{M}\) in the strained region will lead to different transport behaviors for the massless Dirac fermions at these four valleys. Note that the strain-driven shift of Dirac point in energy also appears in Bi\(_2\)Se\(_3\)-class TIs, but there is only one Dirac cone located at \(\bar{\Gamma}\) for them. The transmission probabilities \(T(\theta)\) through the strain-induced nanostructure with two different strain levels are shown in Figs. 11(f) and (g). It can be seen that for the \(\bar{\Gamma}\) valley, high transmission probabilities can always be maintained over a large angular range. This is because the Dirac point shifts downward and the scattering events always stay in the intraband tunneling regime. On the contrary, the Dirac points at the \(\bar{M}\) valley shift upward with increasing strain, and the Fermi level in the strained region will gradually approach the Dirac point of zero carrier concentration. Eventually, the scattering of Dirac fermions at the \(\bar{M}\) valley will be driven into the interband tunneling regime. Particularly in Fig. 11(g), when the Fermi level exactly crosses the Dirac points at \(\bar{M}\) in the strained region, strong beam collimation along three different directions (i.e., \(\theta=0\), \(\pm 18.2^{\circ}\)) can be clearly seen. Deviating from these specific directions, it is possible to achieve the pure current of the \(\bar{\Gamma}\) valley. So the strong valley-selective filtering for massless Dirac fermions can be accomplished by dynamically applying local external pressure, which pave the way for strain-engineered nanoelectronic and valleytronic applications in TCIs. # Thickness engineering When the thickness of SnTe film decreases, the top and bottom surface states will hybridize to open up an energy gap at the Dirac points. However, the hybridization is complex and multiple novel properties can emerge. ## 2D TCI in (001) thin films The Dirac point of SnTe (001) surface states is deviated from the \(\bar{X}\) point in the surface BZ; while at \(\bar{X}\), there is an inverted band gap, i.e., the valence band (conduction band) is derived from the cation (anion). When the thickness of SnTe (001) film is below the penetration length of surface states, the top and bottom surface states will hybridize and result in an energy splitting between the bonding and anti-bonding states. The conduction and valence bands of the 2D film at \(X\) will come from the bonding state of the anion at energy \(E_{\rm A+}(X)\) and anti-bonding state of the cation at energy \(E_{\rm C-}(X)\). However, the band ordering of \(E_{\rm A+}(X)\) and \(E_{\rm C-}(X)\) depends on the competition between the hybridization of the two surfaces and the inverted gap of each surface. For thick films, the hybridization is weak so that \(E_{\rm A+}(X) > E_{\rm C-}(X)\), which inherits the inverted band gap of the 3D limit. While the strong hybridization in the thin film drives \(E_{\rm C-}(X)\) higher than \(E_{\rm A+}(X)\), leading to a trivial phase. Considering the SnTe (001) films with an odd number of atomic layers, which are symmetric under the reflection \(z \rightarrow-z\) about the (001) plane in the middle, the band gap \[\(E_{\rm A+}(X)-E_{\rm C-}(X)\)\] at the \(X\) point as a function of film thickness is shown in Fig. 12(a). Above five layers, the gap of SnTe at \(X\) is inverted and increases with the thickness. Due to the (001) mirror symmetry of the film, one can define the mirror Chern number of 2D film as in the case of 3D TCI. Liu *et al.* found that the thick (001) film with an inverted gap at \(X\) is a 2D TCI with mirror Chern number \(|N_M|=2\) \[Fig. 12(a)\]. An important result of 2D TCI with mirror Chern number \(|N_M|=2\) is that there are two pairs of counter-propagating edge states in the band gap. As shown in the left panel of Fig. 12(b), in a strip structure of 11-layer SnTe thin film, the edge states with opposite mirror eigenvalues cross each other at the edge of BZ, but are not located at TRIM. Here the Dirac-like edge states are protected by the (001) mirror symmetry, rather than the time-reversal symmetry as in the quantum spin Hall insulator. So applying a perpendicular electric field, which breaks the mirror symmetry, will generate a band gap in these edge states, as is shown in the right panel of Fig. 12(b). This unique property that the conductance of edge states is easily and widely tunable by an electric-field-induced gap instead of carrier depletion can be utilized to propose a topological transistor device made of dual-gated TCI thin films, as depicted in Fig. 12(c). With high on/off speed, the device can control the coupled charge and spin transport by purely electrical means. A similar mechanism appears in PbTe/SnTe superlattices along the \[001\] direction. The folding of BZ due to the superlattice structure projects two \(L\) points in the original BZ to a single point \(L^{\prime}\), and their hybridizations form the bonding and anti-bonding states. The coupling strength between the states at equivalent \(L\) points depends on the relative thickness or the ratio \(m/2n\) in (PbTe)\(_m\)(SnTe)\(_{2n-m}\) superlattice. For a large range of this ratio, the strong coupling induces a band inversion between the bonding state of conduction band and anti-bonding state of valence band. Thanks to reduction in the number of \(L\) points from four to two, the band inversion at two \(L^{\prime}\) will lead to a weak TI phase in PbTe/SnTe superlattices. ## Quantum spin Hall states in (111) thin films Another important film is the (111) thin film, which has been grown epitaxially in recent experiments. Different from (001) surface states, four Dirac cones of (111) surface states are centered at TRIMs: one is at the \(\Gamma\) point and the other three are at the \(M\) points, and they are not all equivalent. When the thickness of (111) film decreases, the hybridization strength between the top and bottom surface states at \(\Gamma\) and \(M\) will be different. In fact, the penetration length of the surface states at \(M\) is much larger than that at \(\Gamma\), hence the hybridization-induced gap at \(M\) is correspondingly larger than the one at \(\Gamma\) by orders of magnitude. As a result of this inequivalence between \(\Gamma\) and \(M\), the properties of SnTe (111) films are mainly determined by the low-energy physics at the \(\Gamma\) point. The magnitude and sign of the gap at \(\Gamma\) will depend on the film thickness, thus leading to a topological phase transition as a function of the film thickness. Taking the films with an even number of layers as an example \[Fig. 13(a)\], when the thickness decreases, both the fundamental gap \(E_g\) and the gap at the \(\Gamma\) point \(E_g(\Gamma)\) exhibit a oscillatory behavior, indicating a sign change of the Dirac mass and the \(Z_2\) topological phase transition. As shown in the inset of Fig. 13(a), the conduction bands of the 12-layer thin film are mainly derived from Sn orbitals, while they are mainly from Te orbitals for the 14-layer thin film, suggesting that 14-layer thin film is a quantum spin Hall (QSH) state. This result can be confirmed by the edge state calculations. As shown in Fig. 13(b), for a 14-layer thin film, the linear-dispersion gapless states exist in the bulk gap with a Dirac point at \(\bar{\Gamma}\), while there are no such edge states in the 12-layer thin film. From Fig. 13(a), all the thin films with an even number of layers between 14 and 30 layers are QSH insulators. The similar results apply to the odd-layer thin film, and even the SnSe (111) film. The mechanism for realizing the QSH phase in TCI thin films through the intersurface coupling is similar to the thin films of Bi\(_2\)Se\(_3\)-class TIs. However, due to the much larger penetration length of the SnTe/SnSe surface states, the hybridization gap is much larger than that in Bi\(_2\)Se\(_3\) class of TIs. ## Electric field and helicity control The Dirac fermion of the surface states in the topological materials has the property of spin-momentum locking, which can also be defined as the helicity degree of freedom. As depicted in Fig. 14(a), in the film of topological materials, two surface Dirac cones located on the top and bottom surfaces are mirror images of each other obeying inversion symmetry and have opposite helicity. Specifically, the top surface has the left-handed (LH) electron and right-handed (RH) hole states, while the bottom surface has the RH electron and LH hole states. For the thinner film \[Fig. 14(b)\], the two Dirac cones interact with each other and open a hybridization gap. Meanwhile, the twofold helicity-degenerate band structures sustain both LH and RH electron (or hole) states of Dirac fermions. However, the helicity degeneracy can be removed in a controllable way by breaking the spatial inversion symmetry of the film, e.g., by applying an external electric field \[Fig. 14(c)\]. This feature allows the control over the helicity degree of freedom of Dirac fermions. Taking an odd-layer SnTe (111) film as examples, Zhao *et al.* studied the helicity-resolved band structures and transports. They found that a moderate electric field can induce a giant helicity splitting to the degenerate bands in the vicinity of the \(\bar{\Gamma}\) point. To explore helicity-dependent manipulation of Dirac fermions, a dual-gated heterostructure is exploited in SnTe (111) film, as shown in Fig. 15(a). For Fabry-Pérot-type model, when tuning the Fermi level in the dual-gated region to cross the upper Dirac point, a strong direction-dependent helicity-selective transmission is indicated in Fig. 15(b). The RH electron wave has very high transmission probability over a large angular range, whereas a sharp transmission peak appears at normal incidence for the LH electron wave. Such pronounced helicity-selective transmission stems from the fact that the intrahelical scattering dominates over the interhelical one in the transmission events. When changing the width of dual-gated region and tuning the Fermi level, more helicity-resolved functionalities, including helical negative refraction, birefraction, and electronic focusing, are demonstrated. Like the spin manipulation in thin film of magnetically doped TIs, here the electrically tunable helicity-resolved control of Dirac fermions could open up a promising field for helicity-based electronics. # Conclusions and outlook In this article, we have reviewed the electronic properties of SnTe-class TCI materials, from the intrinsic topology, orientation-dependent surface states, to tunable properties under various perturbations. Protected by the mirror symmetry and characterized by the mirror Chern number, SnTe-class TCI materials have multiple branches of Dirac surface states, whose types depend on the surface orientations and even the surface reconstructions. Chemical doping can realize the control of carrier types or concentrations, the topological superconductivity, and the large-Chern-number QAHE. Strain is an effective way to tune the electronic properties of IV-VI materials, including driving topological phase transition for the bulk system, shifting the Dirac point position of the surface states in the momentum space or energy space. The latter could be used for the valleytronic applications. For the thin film of SnTe-class materials, the thickness of film can influence the hybridization strength, and create diverse topological phases, for instance, 2D TCI, weak TI, QSH states, and NI. Among them, the (001) thin film with 2D TCI phase could be utilized to design a topological transistor device with a high on/off speed, and more generally, the helicity degree of freedom of Dirac fermions in the thin film can be controlled by an external electric field to realize the helicity-resolved functionalities. These novel properties and potential applications are not limited to the IV-VI materials. In fact, more TCIs are proposed theoretically, and a great deal of studies are needed to search for new TCI materials and explore their applications. Moreover, the crystal symmetry protection has been extended from the insulators to semimetals, such as the Dirac semimetal and nodal-line semimetal, which possess more exotic effects. The interplay between the electronic topology and crystal symmetry has broadened the field of topological phase, and paves a path for developing the quantum electronic and spintronic devices.

{'timestamp': '2016-07-19T02:04:37', 'yymm': '1607', 'arxiv_id': '1607.04743', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04743'}

# Introduction Employing the AdS/CFT correspondence it was shown that a holographic superconductor can be builded on a boundary of a gravity theory consisting of a black hole which acquires scalar hair at temperature below a critical temperature, while above the critical temperature there is no scalar hair. A condensate of the charged scalar field is formed through its coupling to a Maxwell field of the gravity sector. The study was carried out in the probe limit , in which in general the product of the charge of the black hole and the charge of the scalar field is held fixed while the latter is taken to infinity and this resulted to neither field to backreact on the metric. Considering fluctuations of the vector potential, the frequency dependent conductivity was calculated, and it was shown that it develops a gap determined by the condensate. This model was further studied beyond the probe limit. If the scalar charge is large then the backreaction of the scalar field to the spacetime metric has to be considered. It was found that all the basic characteristics of the dual superconductor were persisting even for very small charge. One of the characteristics of these models is the strong pairing mechanism which is in operation. This is due to strong bounding of Cooper pairs and this is manifest to the high values of the gap at low temperature. This strong pairing mechanism resulted to \(2 \Delta \approx 8.4 T_c\), where \(\Delta\) is the condensation gap, which has to be compared to the BCS prediction \(2 \Delta \approx 3.54 T_c\), which is much lower in real materials, due to impurities. The above proposals have inspired much effort on further research and on their possible extensions (see and references therein for a review). The effects of paramagnetic impurities on superconductors were studied in. It was found that the transition temperature decreases sharply with increasing impurity concentration and goes to zero at a critical concentration. Measurements of the energy gap and of the transition temperature \(T_c\) as a function of the concentration of paramagnetic impurities show that the gap decreases much more rapidly than does the critical temperature which agrees with the results of. This behaviour was also observed in and . Further calculations on the effects of paramagnetic impurities in superconductors were carried out in, taking full advantage of the information contained in the Green's function of the system. The density of states in energy has been computed for different values of the inverse collision time. The excitation energy gap \(\Omega_g\) is defined to be the energy at which the density of states vanishes. The temperature-dependent order parameter has been computed and the behaviour of \(\Omega_g (T)\) was determined. A comparison with tunneling experiments showed a disagreement of the two parameters. The real part of the conductivity at \(T=0\) is shown to be zero for frequencies less than \(2 \Omega_g\) and proportional to the square of the density of states for vanishingly small frequencies in the gapless region of impurity concentrations. These detailed calculations showed clearly that in real materials, the essential feature of superconductivity is the correlation of the electrons and the formation of Cooper pairs, rather than the existence of a gap in the excitation energy spectrum. For even in the absence of a gap, so long as correlations persist there is an ordered state below a certain critical temperature which displays the usual properties of a superconductor. A holographic realization of impurities was discussed in. Using the AdS/CFT correspondence two different types of impurities were calculated and the Nernst response for the impure theory was specified. A model of a gravity dual of a gapless superconductor was proposed in. The gravity sector was consisting of a charged scalar field which provided the scalar hair of an exact black hole solution  below a critical temperature \(T_c\), while an electromagnetic perturbation of the background determined the conductivity giving rise to a gapless superconductor. It was found that the normal component of the DC conductivity had a milder behaviour than the dual superconductor in the case of a black hole of flat horizon  in which \(n_{n}\) exhibited a clear gap behaviour. The behaviour it was observed in the boundary conducting theory was attributed to materials with paramagnetic impurities as it was discussed in. Impurities effects were studied in, while impurities in the Kondo Model were considered in. In this work we extend the model in by introducing a derivative coupling of the scalar field to Einstein tensor. This term belongs to a general class of scalar-tensor gravity theories resulting from the Horndeski Lagrangian. These theories, which were recently rediscovered, give second-order field equations and contain as a subset a theory which preserves classical Galilean symmetry [@Nicolis:2008in; @Deffayet:2009wt; @Deffayet:2009mn]. The derivative coupling of the scalar field to Einstein tensor introduces a new scale in the theory which on short distances allows to find black hole solutions with scalar hair just outside the black hole horizon, while if one considers the gravitational collapse of a scalar field coupled to the Einstein tensor then the formation of a black hole takes more time to be formed compared to the collapse of a scalar field minimally coupled to gravity . On large distances the presence of the derivative coupling acts as a friction term in the inflationary period of the cosmological evolution . Moreover, it was found that at the end of inflation in the preheading period, there is a suppression of heavy particle production as the derivative coupling is increased. This was attributed to the fast decrease of kinetic energy of the scalar field due to its wild oscillations. The above discussion indicates that one of the main effects of the kinematic coupling of a scalar field to Einstein tensor is that gravity influences strongly the propagation of the scalar field compared to a scalar field minimally coupled to gravity. We are going to use this behaviour of the scalar field to holographically simulate the effects of a high concentration of impurities in a material. The presence of impurities in a superconductor is making the pairing mechanism of forming Cooper pairs less effective and this is happening because the quasiparticles are loosing energy because of strong concentration of impurities. This holographic correspondence is supported by our finding. We found that as the value of the derivative coupling is increased the critical temperature is degreasing while the condensation gap \(\Delta\) is decreasing faster than the temperature. Also by calculating the perturbation of the scalar potential we found that the condensation gap for large values of the derivative coupling is not proportional to the frequency of the real part of conductivity which is characteristic of a superconducting state with impurities [^1]. The work is organized as follows. In Section [2](#review){reference-type="ref" reference="review"} we review the basic ingredients of the holographic model discussed in. In Section [3](#derscalar){reference-type="ref" reference="derscalar"} we extend the holographic model of Section [2](#review){reference-type="ref" reference="review"} adding a scalar field coupled to Einstein tensor and we give the equations of motion. In Section [3.1](#solution){reference-type="ref" reference="solution"} we solve numerically the equations of motion and we find the condensation gap and the critical temperature for various values of the derivative coupling. In Section [3.2](#conductivity){reference-type="ref" reference="conductivity"} we study the conductivity while in Section [4](#conc){reference-type="ref" reference="conc"} are our conclusions. # Holographic Superconductor with a scalar field minimally coupled to gravity {#review} In this section we review in brief the holographic model of. To introduce the minimal holographic superconductor model, we consider a Maxwell field and a charged complex scalar field coupled to gravity with the action \[\begin{aligned} \label{System} S=\int d^{4}x\sqrt{-g}\left[\frac{R+\Lambda}{16 \pi G}-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-|\nabla\psi-iqA\psi|^{2}-m^2|\psi|^2 \right] \. \end{aligned}\] If we rescale the fields and the coupling constant as \[\begin{aligned} A_{\mu} =\frac{\tilde{A}_{\mu}}{q}~,~~~ \psi =\frac{\tilde{\psi}}{q}~, \end{aligned}\] then the matter action in ([\[System\]](#System){reference-type="ref" reference="System"}) has a \(1/q^2\) in front, so the backreaction of the matter fields on the metric is suppressed when \(q\) is large and the limit \(q\rightarrow\infty\) defines the probe limit. In this limit the Einstein equations admit as a solution the planar Schwarzschild AdS black hole \[\begin{aligned} \label{schwar} ds^2=-f(r) dt^2 + \frac{dr^2}{f(r)} + r^2 (dx^2 + dy^2)\qquad,\qquad f(r)=r^2-\frac{r_h^3}{r}~. \end{aligned}\] Taking the ansatz \(\psi=|\psi|\), \(A=\phi dt\) where \(\psi\), \(\phi\) are both functions of \(r\) only, we can obtain the equations of motion for \(\psi, \phi\) \[\begin{aligned} \psi^{\prime\prime}+\left( \frac{f^\prime}{f}+\frac{d-2}{r}\right)\psi^\prime +\left(\frac{\phi^2}{f^2}-\frac{m^2}{f}\right)\psi=0\,, \label{Psi} \end{aligned}\] \[\begin{aligned} \phi^{\prime\prime}+\frac{d-2}{r}\phi^\prime-\frac{2\psi^2}{f}\phi=0~. \label{Phi} \end{aligned}\] It was argued in that there is a critical temperature below which the black hole acquires hair . The equations ([\[Psi\]](#Psi){reference-type="ref" reference="Psi"}) and ([\[Phi\]](#Phi){reference-type="ref" reference="Phi"}) can be solved numerically by doing integration from the horizon out to the infinity taking under consideration the right boundary conditions. The solutions behave like \[\begin{aligned} \psi=\frac{\psi_{-}}{r^{\lambda_{-}}}+\frac{\psi_{+}}{r^{\lambda_{+}}}\,,\hspace{0.5cm} \phi=\mu-\frac{\rho}{r}\,, \label{infinity} \end{aligned}\] with \[\begin{aligned} \lambda_\pm=\frac{1}{2}\Big{[}3 \pm\sqrt{(9+4m^{2}L ^2)}~\Big{]}\,,\label{LambdaZF} \end{aligned}\] where \(L\) is the scale of the AdS space and \(\mu\) and \(\rho\) are interpreted as the chemical potential and charge density in the dual field theory respectively. The coefficients \(\psi_{-}\) and \(\psi_{+}\) according to the AdS/CFT correspondence, correspond to the vacuum expectation values \(\psi_{-}=\langle\mathcal{O}_{-}\rangle/\sqrt{2}\), \(\psi_{+}=\langle\mathcal{O}_{+}\rangle/\sqrt{2}\) of an operator \(\mathcal{O}\) dual to the scalar field. We can impose boundary conditions that either \(\psi_{-}\) or \(\psi_{+}\) vanishes. Fluctuations of the vector potential \(A_x\) in the gravity sector gives the conductivity in the dual CFT as a function of frequency. The Maxwell equation for zero spatial momentum and with a time dependence of the form \(e^{-i \omega t}\) reads as \[\begin{aligned} \label{eq:Axeq} A_x'' + \frac{f'}{f} A_x' + \left(\frac{\omega^2}{f^2}-\frac{2 \psi^2}{f} \right) A_x = 0 \. \end{aligned}\] The above equation can be solved by imposing ingoing wave boundary conditions at the horizon \(A_x \propto f^{-i\omega/3r_0}\). On the other hand, the asymptotic behaviour of the Maxwell field at boundary is \[\begin{aligned} A_x = A_x^{(0)} + \frac{A_x^{(1)}}{r} + \cdots. \end{aligned}\] Then according to AdS/CFT correspondence dual source and expectation value for the current are given by \[\begin{aligned} A_x = A_x^{(0)} \,, \qquad \langle J_x \rangle = A_x^{(1)} \. \end{aligned}\] Finally, Ohm's law gives the conductivity \[\begin{aligned} \label{eq:conductivity} \sigma(\omega) = \frac{\langle J_x \rangle}{E_x} =-\frac{ \langle J_x \rangle}{\dot A_x} =-\frac{ i \langle J_x \rangle}{\omega A_x} =-\frac{i A_x^{(1)}}{\omega A_x^{(0)}} \. \end{aligned}\] # Holographic Superconductor with a scalar field kinematically coupled to Einstein Tensor {#derscalar} In this section, we will consider a complex scalar field which excepts its coupling to curvature it is also coupled to Einstein tensor with action \[\begin{aligned} \label{EGBscalaraction} I=\int d^4x\sqrt{-g}\left[ \frac{R+\Lambda}{16\pi G}-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}-(g^{\mu\nu}+\kappa G^{\mu\nu})D_\mu\psi (D_\nu\psi)^{*}-m^2|\psi|^2 \right]~, \end{aligned}\] where \[\begin{aligned} D_\mu = \nabla_\mu-i e A_\mu \end{aligned}\] and \(e\), \(m\) are the charge and the mass of the scalar field and \(\kappa\) the coupling of the scalar field to Einstein tensor of dimension length squared. For convenience we set \[\begin{aligned} \Phi_{\mu\nu} &\equiv& D_{\mu}\psi (D_{\nu}\psi)^*~,\\ \Phi &\equiv& g^{\mu\nu}\Phi_{\mu\nu}~,\\ C^{\mu\nu} &\equiv& g^{\mu\nu} + \kappa G^{\mu\nu}~. \end{aligned}\] The field equations resulting from the action ([\[EGBscalaraction\]](#EGBscalaraction){reference-type="ref" reference="EGBscalaraction"}) are \[\begin{aligned} G_{\mu\nu} +\Lambda g_{\mu\nu} = 8\pi T_{\mu\nu} \, \ \ \ \ T_{\mu\nu} = T_{\mu\nu}^{(\psi)} + T_{\mu\nu}^{(EM)} + \kappa\Theta_{\mu\nu}~, \label{einst}\end{aligned}\] where, \[\begin{aligned} T_{\mu\nu}^{(\psi)} & = & \Phi_{\mu\nu} + \Phi_{\nu\mu}-g_{\mu\nu}(g^{ab}\Phi_{ab} + m^2 |\psi|^2)~, \\ T_{\mu\nu}^{(EM)} & = & F_{\mu}^{\phantom{\mu} \alpha} F_{\nu \alpha}-\frac{1}{4} g_{\mu\nu} F_{\alpha\beta}F^{\alpha\beta}~, \end{aligned}\] and \[\begin{aligned} \Theta_{\mu\nu} = &-& g_{\mu\nu} R^{ab}\Phi_{ab} + R_{\nu}^{\phantom{\nu}a}(\Phi_{\mu a} + \Phi_{a\mu}) + R_{\mu}^{\phantom{\mu}a} (\Phi_{a\nu} + \Phi_{\nu a})-\frac{1}{2} R (\Phi_{\mu\nu} + \Phi_{\nu\mu}) \nonumber\\ &-& G_{\mu\nu}\Phi-\frac{1}{2}\nabla^a\nabla_\mu(\Phi_{a\nu} + \Phi_{\nu a})-\frac{1}{2}\nabla^a\nabla_\nu(\Phi_{\mu a} + \Phi_{a\mu}) + \frac{1}{2}\Box (\Phi_{\mu\nu} + \Phi_{\mu\nu}) \nonumber \\ & + & \frac{1}{2}g_{\mu\nu} \nabla_a\nabla_b (\Phi^{ab} + \Phi^{ba}) + \frac{1}{2}(\nabla_\mu\nabla_\nu + \nabla_\nu\nabla_\mu) \Phi-g_{\mu\nu}\Box\Phi~. \label{theta} \end{aligned}\] The Klein-Gordon equation is \[\begin{aligned} (\partial_\mu-i e A_\mu) \left[ \sqrt{-g}C^{\mu\nu}(\partial_\nu-i e A_\nu)\psi\right] = \sqrt{-g} m^2\psi~, \label{glgord} \end{aligned}\] while the Maxwell equations read \[\begin{aligned} \nabla_\nu F^{\mu\nu} + C^{\mu\nu} \left[ 2 e^2 A_\nu |\psi|^2 + i e (\psi^*\nabla_\nu\psi-\psi\nabla_\nu\psi^*)\right] =0~. \label{max}\end{aligned}\] ## Solution of equations of motion and phase transition {#solution} As in we also take the ansatz \(\psi=|\psi|\), \(A=\phi dt\) where \(\psi\), \(\phi\) are both functions of \(r\) only. In the probe limit, under the metric ([\[schwar\]](#schwar){reference-type="ref" reference="schwar"}), the above equations of motion for the matter fields become \[\left[1+\kappa\left(\frac{f}{r^2}+\frac{f'}{r}\right)\right]\psi^{\prime\prime} + \left[\frac{2}{r} + \frac{f^\prime}{f} + \kappa\left (\frac{3f'}{r^2}+\frac{f'^{2}}{r f}+\frac{f''}{r} \right)\right]\psi^\prime + \left[\frac{e^2\phi^2}{f^2}\left(1+\kappa\left(\frac{f}{r^2}+\frac{f'}{r}\right)\right)-\frac{m^2}{f}\right]\psi = 0 \label{newpsi}~,\] \[\phi^{\prime\prime} + \frac{2}{r}\phi^\prime-\frac{2e^2\psi^2}{f}\left[1+\kappa\left(\frac{f}{r^2}+\frac{f'}{r}\right)\right]\phi = 0~. \label{newphi}\] Note that if \(\kappa=0\) then we get the equations ([\[Psi\]](#Psi){reference-type="ref" reference="Psi"}) and ([\[Phi\]](#Phi){reference-type="ref" reference="Phi"}). The new equations depend now on the second derivative of the function \(f\) in contrast to the preview case where we had only its first derivative, due to the additional term of \(G^{\mu\nu}\) in the action ([\[EGBscalaraction\]](#EGBscalaraction){reference-type="ref" reference="EGBscalaraction"}). With the same study in the minimal case presented in previous section, we can numerically solve the above equations of motion. We especially study the effect of \(\kappa\) on the critical temperature and strength of condensation in the dual supercondutor. Without loss of generality, we set \(e=1\) and \(m^2=-2\), so that we have \(\lambda_-=1\) while \(\lambda_+=2\). We will choose \(\psi_-\) as the source which is set to be vanishing, while \(\psi_+\) as vacuum expectation values. The critical temperatures with different \(\kappa\) are summarized in Table [1](#tableTc){reference-type="ref" reference="tableTc"}. It shows that as the coupling becomes stronger, the critical temperature is lower, meaning that the phase transition is hard to occur as it is expected. This effect of coupling is very different from the influence of higher-derivative coupling between the matters studied in. As the temperature is decreased, the scalar field can have non-zero solution and the strength of condensation becomes higher. We show this phenomena in FIG. [\[fig-condensation\]](#fig-condensation){reference-type="ref" reference="fig-condensation"}. It is obvious that the strength of condensation is suppressed by the derivative coupling. And with enlarging the coupling[^2], the condensation decreases faster than the temperature, so that the condensation gap tends to be zero at low temperatures for the strong coupling. This behaviour from \(\kappa\) is coincident with the effect of paramagnetic impurities on superconductors observed in. We fit the curve near the critical temperatures. As \(T\rightarrow T_c\), the condensation is continuous and behaves as \[\langle\mathcal{O}_{+}\rangle\simeq C_1 T_c^2(1-T/T_c)^{1/2} \label{near}\] where \(C_1\) are also listed in Table [1](#tableTc){reference-type="ref" reference="tableTc"}. We see that the coefficient \(C_1\) decreases drastically to suppress the condensation when the coupling is increased. Observe that the exponent is always \(1/2\) which implies that \(\kappa\) does not modify the order of the phase transition and it is always second order. In the next subsection we will study the conductivity caused by the condensation below the critical temperature. ## Conductivity To compute the conductivity in the dual CFT as a function of frequency we need to solve the Maxwell equation for fluctuations of the vector potential \(A_x\). The Maxwell equation at zero spatial momentum and with a time dependence of the form \(e^{-i \omega t}\) gives \[\label{eq:Axeq} A_x'' + \frac{f'}{f} A_x' +\left[\frac{\omega^2}{f^2}-\frac{2 e^2\Psi^2}{f} \left(1+\kappa\left(\frac{f''}{2}+\frac{f'}{r}\right)\right)\right] A_x = 0~.\] We will solve the perturbed Maxwell equation with ingoing wave boundary conditions at the horizon, i.e., \(A_x \propto f^{-i\omega/3r_0}\). The asymptotic behaviour of the Maxwell field at large radius is \(A_x = A_x^{(0)} + \frac{A_x^{(1)}}{r} + \cdots\). Then, according to AdS/CFT dictionary, the dual source and expectation value for the current are given by \(A_x = A_x^{(0)}\) and \(\langle J_x \rangle = A_x^{(1)}\), respectively. Thus, similar to the equation ([\[eq:conductivity\]](#eq:conductivity){reference-type="ref" reference="eq:conductivity"}), the conductivity is also read as \[\sigma(\omega)=-\frac{i A_x^{(1)}}{\omega A_x^{(0)}}~.\] The numerical results of conductivity at low temperature with \(\frac{T}{T_c}\simeq 0.1\) are shown in FIG. [\[fig-sigma\]](#fig-sigma){reference-type="ref" reference="fig-sigma"}. Similarly, due to the Kramers-Kronig(KK) relations \[\label{eq:KK} \mathrm{Im}[\sigma(\omega)]=-\frac{1}{\pi}\mathcal{P}\int_{-\infty}^{\infty}\frac{\mathrm{Re}[\sigma(\omega')]}{\omega'-\omega}d\omega'~,\] where \(\mathcal{P}\) denotes the Cauchy principal value, the divergence of the imaginary part at zero frequency indicates a delta function of real part at \(\omega=0\). At very low temperature, the energy gap in the real part of the conductivity measured by the critical temperature (left part of FIG. [\[fig-sigma\]](#fig-sigma){reference-type="ref" reference="fig-sigma"}) decreases as the derivative coupling becomes stronger, which agrees well with the behaviour of the condensation as it is depicted in FIG. [\[fig-condensation\]](#fig-condensation){reference-type="ref" reference="fig-condensation"} in the last subsection. In order to show how the conductivity opens a gap as the temperature decreases, in FIG. [\[fig-sigmak=0.01\]](#fig-sigmak=0.01){reference-type="ref" reference="fig-sigmak=0.01"}, we show the frequency-dependent conductivity with \(\kappa=0.01\) at different temperatures. In the left plot, the horizontal line, which is frequency independent, corresponds to temperatures at or above the critical value and there is no condensate. As we lower the temperature in the subsequent curves, we see that a gap \(\omega_g\) opens in the real part of the conductivity and the gap becomes wider and deeper as the temperature becomes smaller. In the right plot, we plot the conductivity for temperature lower than \(T_c\) related the left plot by rescaling the frequency by the condensate. It is obvious that the curves tend to a limit in which the width of the gap is proportional to the size of the condensate. i.e., \(\omega_g \simeq \sqrt{<O_+>}\). The above features are similar with those in minimal coupling disclosed in . Furthermore, the frequency-dependent conductivity with \(\kappa=0.5\) at different temperatures is presented in FIG. [\[fig-sigmak=0.5\]](#fig-sigmak=0.5){reference-type="ref" reference="fig-sigmak=0.5"}. Similarly to the cases with \(\kappa=0.01\) and minimal coupling, lower temperature gives us wider energy gap. By comparing the left plots of FIG. [\[fig-sigmak=0.01\]](#fig-sigmak=0.01){reference-type="ref" reference="fig-sigmak=0.01"} and FIG. [\[fig-sigmak=0.5\]](#fig-sigmak=0.5){reference-type="ref" reference="fig-sigmak=0.5"}, we see that at the same scaled temperature \(T/T_c\), the gap is narrower for larger coupling, which is consistent with the results shown in FIG. [\[fig-condensation\]](#fig-condensation){reference-type="ref" reference="fig-condensation"} and FIG. [\[fig-sigma\]](#fig-sigma){reference-type="ref" reference="fig-sigma"}. However, if being scaled by the condensation, see the right plot for \(\kappa=0.5\), the behavior \(\omega_g \simeq \sqrt{<O_+>}\) at low \(T/T_c\) is violated and it is \(\omega_g \simeq 20\sqrt{<O_+>}\) for \(T/T_c\simeq 0.1\). The coefficient \(20\) is not out of expect. Because if we see carefully the condensation plot FIG. [\[fig-condensation\]](#fig-condensation){reference-type="ref" reference="fig-condensation"} and FIG. [\[fig-sigma\]](#fig-sigma){reference-type="ref" reference="fig-sigma"}, we will find \(\sqrt{<O_+>}\simeq 0.2 T_c\) and \(\omega_g \simeq 4T_c\), so we get the same relation as before \(\omega_g \simeq 20\sqrt{<O_+>}\). For the systems far away from the critical temperature (\(T/T_c\simeq 0.10\)), by careful study, we summarize the values of \(\frac{\omega_g}{\sqrt{<O_+>}}\) with some \(\kappa\) in Table [2](#tablewg){reference-type="ref" reference="tablewg"}. It is obvious that as \(|\kappa|\) shifts from the minimal coupling, the relation \(\omega_g \simeq \sqrt{<O_+>}\) is sharply violated, which is one of the main features of a superconducting state with impurities. Note that this effect of possible impurity parameter \(\kappa\) is different from that studied in where it can not introduce this kind of violation. We can now explore the behaviour of conductivity at very low frequency. When \(T<T_c\), the real part of the conductivity present a delta function at zero frequency and the imaginary part has a pole, which is attributed to the KK relations ([\[eq:KK\]](#eq:KK){reference-type="ref" reference="eq:KK"}). More specifically, as \(\omega\rightarrow 0\), the imaginary part behaves as \(Im(\sigma)\sim n_s/\omega\), and according to Kramers-Kronig relations, the real part has the form \(Re(\sigma)\sim\pi n_s\delta(\omega)\). Here the coefficient \(n_s\) of the delta function is defined as the superfluid density. By fitting data near the critical temperature, we find that with various couplings, the superfluid density has the behaviour \[\label{eq:ns} n_s\simeq C_2 T_c(1-T/T_c)~,\] which means that \(n_s\) will vanish linearly as \(T\) goes to \(T_c\). This is consistent with that happens in the minimal coupling. Also we find that the coefficient \(C_2\) is not very sensitive to the coupling and the value oscillates around \(24\) with \(\pm 2\) shifting for the samples of \(\kappa\). This insensitivity to the changes of the coupling is reasonable because in the region very close to the critical temperature, the system is marginally shifted from the normal state, so that varying \(\kappa\) has very small effect on the behaviour of the solutions. Meanwhile, as proposed in, the non-superconducting density can be defined as \(n_n=\lim_{\omega\rightarrow 0}Re(\sigma)\). Far from the critical temperature (with \(T/T_c\sim 0.1\)), we fit and obtain that \(n_n\) decays as \[\label{eq:nn} n_n\sim \exp[-\frac{\Omega_g}{T}]~,\] where \(\Omega_g\) can be explained as the energy gap for charged excitation at the corresponding temperature. The values of \(\Omega_g\) close to \(T/T_c\sim 0.1\) are summarized in Table [3](#tableWg){reference-type="ref" reference="tableWg"}. Rough comparison with the numerical energy gap \(\omega_g/T_c\) shown in the left plot of FIG.[\[fig-sigma\]](#fig-sigma){reference-type="ref" reference="fig-sigma"} gives us that \(\omega_g \simeq 2\Omega_g\), where the factor \(2\) suggests that the gaped charged quasiparticles are formed in pairs as addressed in. Analyzing the data in the table, we see that \(\kappa=0.5\) corresponds to \(\omega_g \simeq 2\Omega_g \simeq 4 T_c\) which is very near the prediction \(\omega_g \simeq 3.54 T_c\) in BCS theory. In this tendency, we believe that BCS prediction may be fulfilled by larger coupling once the numerics are controlable. Thus, in this sense, the derivative coupling somehow mimic the effect of the impurities in a real material. # Conclusions {#conc} We studied a holographic description of a superconductor in which the gravity sector consists of a Maxwell field and a charged scalar field which except its usual minimal coupling to gravity it is also coupled to Einstein tensor. Solving the equations of motion numerically in the probe limit, we found that as the strength of the new coupling is increased, the critical temperature below which the scalar field condenses is lowering, the condensation gap decreases faster than the temperature, the width of the condensation gap is not proportional to the size of the condensate and at low temperatures the condensation gap tends to zero for the strong coupling. Analysing the frequency dependence of conductivity we found that at strong coupling the relation \(\omega_g \simeq 2\Omega_g \simeq 4 T_c\) holds, where \(\Omega_g\) is the energy gap for charged excitation, which is very near the prediction \(\omega_g \simeq 3.54 T_c\) in BCS theory. We argued that these results suggest that the derivative coupling in the gravity bulk can have a dual interpretation on the boundary corresponding to impurities concentrations in a real material. This correspondence can be understood from the fact that the coupling of a scalar field to Einstein tensor alters the kinematical state of the scalar field a behaviour which it is also exhibited by the quasiparticles moving in a material with impurities. We believe that the probe limit captures all the essential features of the problem under study. Nevertheless, it would be interesting to extend this study beyond the probe limit. Assuming a spherically symmetric ansatz for the metric the field equations ([\[einst\]](#einst){reference-type="ref" reference="einst"}), ([\[glgord\]](#glgord){reference-type="ref" reference="glgord"}) and ([\[max\]](#max){reference-type="ref" reference="max"}) have to be solved. However, even their numerical solution is a formidable task mainly because of the presence of the energy-momentum tensor ([\[theta\]](#theta){reference-type="ref" reference="theta"}) resulting from the coupling of the scalar field to Einstein tensor. A more realistic approach would be to follow the perturbative methods employed in. [^1]: We note that in a holographic superconductor was considered containing a derivative coupling for a scalar field in the background of a regular phantom plane symmetric black hole. [^2]: There is an instability in our numerical solution of the equations of motion for large or small values of the derivative coupling when the temperature is far way from \(T_c\). Specially, it is difficult to reach a low temperature \(T/T_c\sim 0.1\) when \(\kappa>0.5\). For this reason in FIG. [\[fig-condensation\]](#fig-condensation){reference-type="ref" reference="fig-condensation"} we restricted the curves to the values of \(\kappa\) in the range \(-0.01\leqslant\kappa\leqslant0.5\). However, when we are near \(T_c\) our numerics work better, so we could fit relation ([\[near\]](#near){reference-type="ref" reference="near"}) for larger values of \(\kappa\). So in order to study the same shift from the critical temperature in the following study we focus on the range \(-0.01\leqslant\kappa\leqslant0.5\).

{'timestamp': '2016-08-30T02:02:20', 'yymm': '1607', 'arxiv_id': '1607.04928', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04928'}

# Introduction Heavy quarks (charm and beauty) are abundantly produced at the LHC in the early stages of hadronic collisions via hard scattering processes and experience the full evolution of the collision. The measurement of heavy flavours in \(\mathrm {p\kern-0.05emp}\) collisions can be used to test pQCD calculations. In addition, they provide an essential baseline for the studies in nuclear (\(\mbox{p--Pb}\) and \(\mbox{Pb--Pb}\)) collisions. Their production in nuclear collisions is modified by cold nuclear matter effects (CNM) such as shadowing and energy loss. The knowledge of these effects is fundamental for understanding the interactions of heavy quarks with the deconfined medium formed in heavy-ion collisions where the modified transverse momentum distribution can be used to infer the medium's characteristics. Heavy quark production is experimentally accessible through the measurement of heavy-flavour hadrons (D or B hadrons) via their decay products. The charged-particle multiplicity is defined as the average number of charged particles produced per event. It is a key observable for multi-parton interactions (MPI) in high-energy hadronic collisions. The charged-particle multiplicity may be used to test models of particle production such as PYTHIA and PHOJET. Finally, the charged-particle multiplicity dependence of open heavy-flavour yields provides insight into mechanisms of particle production in hadronic collisions. In this report we present the relative yield of D mesons as a function of the relative charged-particle multiplicity in \(\mathrm {p\kern-0.05emp}\) collisions at \(\sqrt{s}\) = 7 TeV and in \(\mbox{p--Pb}\) collisions at \(\sqrt{s_{NN}}\) = 5.02 TeV. At the LHC, ALICE comprises central detectors, used for the measurement of charged particles as well as photons and a forward muon spectrometer which measures muons. The L3 solenoid magnet provides a field of 0.5 T to the central barrel detectors covering the rapidity-0.9 \(<\) \(\textit{y}\) \(<\) 0.9. The Silicon Pixel Detector (SPD) is used for the determination of the interaction vertex as well as the measurement of charged-particle multiplicity. Particle identification information is given by the Time Of Flight (TOF) and the Time Projection Chamber (TPC). Two arrays of scintillator detectors (V0A and V0C) placed on both sides of the interaction point provide centrality in \(\mbox{Pb--Pb}\) collisions as well as trigger information. They are also used for multiplicity determination. The forward muon spectrometer covers the rapidity range-4 \(<\) \(\textit{y}\) \(<\)-2.5. It consists of a composite absorber, a dipole magnet, five tracking stations, a muon filter and two trigger stations. The absorber reduces background muons from the decays of pions and kaons. The dipole magnet provides a horizontal magnetic field perpendicular to the beam axis and is used for charge and momentum determination. The tracking stations are used to determine the trajectories of the muons traversing the detector. The muon filter is mounted at the end of the spectrometer and filters all particles with momentum \(\textit{p}\) \(<\) 4 GeV/c. The trigger stations are used for muon identification and triggering.The reconstruction of D mesons is done via their hadronic decay channels: D\(\textsuperscript{0}\) \(\rightarrow\) \({\rm K}^{-}\[\pi^{-}\) (branching ratio, BR = 3.88\(\pm\)`<!-- -->`{=html}0.05 %), D\(\textsuperscript{+}\) \(\rightarrow\) \({\rm K}^{-}\]\pi^{+}\)(BR = 9.13\(\pm\)`<!-- -->`{=html}0.19 %), D\(\textsuperscript{*+}\) \(\rightarrow\) D\(\textsuperscript{0}\[\pi^{+}\) (BR = 67.7\(\pm\)`<!-- -->`{=html}0.05 %) followed by D\(\textsuperscript{0}\) \(\rightarrow\) \({\rm K}^{-}\]\pi^{+}\) and their charge conjugates at central rapidity. Further details on D meson and charged-particle multiplicity measurements are reported in and. The measurement of prompt and non-prompt (produced via the decay of beauty hadrons) J/\(\psi\) \(\rightarrow\) e\(\textsuperscript{+}\)e\(\textsuperscript{+}\) (BR = 5.97\(\pm\)`<!-- -->`{=html}0.03 %) and J/\(\psi\) \(\rightarrow\) \(\mu\[\textsuperscript{+}\]\mu\[\textsuperscript{+}\) (BR = 5.96\(\pm\)`<!-- -->`{=html}0.03 %) at central and forward rapidity are reported in.   # Results ## \(\textbf{D-meson production in \)\textit{pp}\( collisions}\) The results of the relative yields (ratio between the yield in a given multiplicity interval normalised to the multiplicity-integrated one) of D mesons, D\(\textsuperscript{0}\), D\(\textsuperscript{+}\), D\(\textsuperscript{*+}\), measured at central rapidity as a function of relative charged-particle multiplicity (multiplicity of charged particles normalised to the average value) in \(\mathrm {p\kern-0.05emp}\) collisions at \(\sqrt{s}\) = 7 TeV are presented in Figure [\[results1\]](#results1){reference-type="ref" reference="results1"}. In Figure [\[fig:first_1\]](#fig:first_1){reference-type="ref" reference="fig:first_1"}, the prompt D-meson relative yield is compared with that of inclusive J/\(\psi\) as a function of the relative multiplicity. The yields of D mesons and inclusive J/\(\psi\) show a similar increase with the multiplicity, in particular for J/\(\psi\) measured at central rapidity. In Figure [\[fig:first_2\]](#fig:first_2){reference-type="ref" reference="fig:first_2"} the yield of D mesons is compared to that of non-prompt J/\(\psi\) (produced in the decay of beauty hadrons), which provides an indirect measurement of the multiplicity dependence of beauty hadrons. The trend is similar to that observed in Figure [\[fig:first_1\]](#fig:first_1){reference-type="ref" reference="fig:first_1"}. We observe no difference, within uncertainty, in the multiplicity dependence of charm and beauty hadron yields at central rapidity. ## \(\textbf{D-meson production in \ensuremath{\mbox{p--Pb}}\ collisions}\) Figure [\[results3\]](#results3){reference-type="ref" reference="results3"} shows the prompt D-meson yield as a function of the relative multiplicity measured at central rapidity (-0.9 \(<\) \(\textit{y}\) \(<\) 0.9) compared with J/\(\psi\) in \(\mbox{p--Pb}\) collisions at \(\sqrt{s_{NN}}\) = 5.02 TeV at forward (p going) and backward (Pb going) rapidity. The yield of D mesons increases faster than the J/\(\psi\) yield towards high multiplicity, particularly for J/\(\psi\) measured at forward rapidity. ## \(\textbf{Comparison of D-meson measurements in \)\textit{pp}\( and \ensuremath{\mbox{p--Pb}}\ collisions}\) The yield of D mesons in \(\mathrm {p\kern-0.05emp}\) collisions at \(\sqrt{s}\) = 7 TeV is compared to the yield in \(\mbox{p--Pb}\) collisions at \(\sqrt{s_{NN}}\) = 5.02 TeV as a function of charged-particle multiplicity in Figure [\[results5\]](#results5){reference-type="ref" reference="results5"}. The D-meson yield increases faster than linear with multiplicity for both \(\mathrm {p\kern-0.05emp}\) and \(\mbox{p--Pb}\) collisions. The increase is independent of the colliding system. ## \(\textbf{Comparison of results with theoretical models}\) The comparison of the measurements in \(\mathrm {p\kern-0.05emp}\) and \(\mbox{p--Pb}\) collisions with theoretical calculations is shown in Figure [\[results6\]](#results6){reference-type="ref" reference="results6"}. In \(\mathrm {p\kern-0.05emp}\) collisions the yield of D mesons as a function of the charged-particle multiplicity is compared with theoretical predictions based on percolation, EPOS with and without hydrodynamical effects and PYTHIA 8. All the models predict an increase of the relative yield of D mesons with the relative multiplicity for both \(\mathrm {p\kern-0.05emp}\) and \(\mbox{p--Pb}\) collisions. However, percolation and EPOS+hydro predict a faster than linear increase while PYTHIA 8 and EPOS are consistent with the data at lower multiplicity. In \(\mbox{p--Pb}\) collisions, the results are compared with EPOS+hydro and EPOS. Both models predict an increase of the relative yield with the multiplicity. EPOS+hydro predicts a faster than linear increase. # Conclusion The ALICE Collaboration measured the relative yield of D mesons as a function of the charged-particle multiplicity in \(\mathrm {p\kern-0.05emp}\) and \(\mbox{p--Pb}\) collisions and compared the results with analogous measurements of inclusive and non-prompt J/\(\psi\). In both colliding systems the relative yield of D mesons increases with multiplicity and the increase is faster than linear. The increase is independent of the \(\textit{p}\]_\text{T}\) interval considered. J/\(\psi\) yields show a similar increase in \(\mathrm {p\kern-0.05emp}\) collisions, however in \(\mbox{p--Pb}\) collisions D-meson yields increase faster than J/\(\psi\), particularly at forward rapidity. The results were also compared with theoretical calculations. All models considered in the study predict an increase of heavy-flavour yields with multiplicity. The multiplicity dependence of D-meson production is described better by calculations that include hydrodynamics, especially at higher multiplicity.

{'timestamp': '2016-07-18T02:07:58', 'yymm': '1607', 'arxiv_id': '1607.04554', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04554'}

# [\[sec:level1\]]{#sec:level1 label="sec:level1"}Introduction Contamination of a cylindrically or toroidally trapped electron plasma by a fractional population of ions can, under certain conditions, disrupt the equilibrium of the trapped electron cloud. The related instability, known as the ion resonance instability, manifests itself in the form of growing poloidal modes (diocotron modes) of the plasma. The primary cause for such contamination (either planned or inadvertent) in most electron plasma trap experiments is the process of ionization of the neutral background gas in the trap by the electrons. A brief history of analytical modelling, numerical analysis, and experiments performed to understand this destabilization process can be found in Sec I of Reference 3. Of the theoretical models for the ion resonance instability, the one developed by Davidson and Uhm aptly describes the collisionless dynamics of the instability in the present set of numerical experiments performed with an initially cold, partially neutralized electron cloud in cylindrical confinement. The model describes the ion resonance instability as a rotating two-stream instability between the electron component and the partially neutralizing ion component of the plasma. The differential rotation between the components in mutual equilibrium drives the instability; the rotation itself being a result of the radial self-electric field of the nonneutral plasma. Davidson and Uhm's linearised model works for any cold equilibrium of the two component plasma in cylindrical confinement. It shows that depending on the equilibrium parameters, one or more exponentially growing diocotron modes may be excited on the plasma profile with the equilibrium condition determining the fastest growing diocotron mode in the system. An abridged version of the analytical procedures for this model can be found in Sec III of Reference 3. In an earlier work we performed numerical experiments to investigate the full nonlinear evolution of the ion resonance instability in partially neutralized electron clouds in cylindrical traps. The numerical experiments were performed for a set of different initial unstable equilibria of the two component plasma (defined by the equilibrium parameters), and a well benchmarked 2D Electrostatic Particle-in-cell (PIC) code was utilized for these simulations. Results from the initial linear growth phase of the instability were in excellent agreement with the linearised analytical model of Davidson and Uhm which served as an added benchmark for the code. The simulations delved further in time beyond the linear growth phase and revealed several new and interesting nonlinear phenomena associated with the instability. For example, a process of simultaneous wave breaking of the excited poloidal mode on the ion cloud and pinching of the poloidal perturbations on the electron cloud occurring in the nonlinear phase of the instability was reported. This simultaneous nonlinear dynamics of the two components was found to be associated with an energy transfer process from the electrons to the ions. At later stages of the nonlinear phase, a heating induced cross-field transport of the heavier ions and tearing across the pinches on the electron cloud followed by an inverse cascade of the torn sections was also demonstrated. These ion resonance instability simulations of the previous work were performed in an ideal scenario with a perfect vacuum as a background. For the present work we have taken these simulations a notch closer to a usual experimental scenario by including the effect of elastic and excitation collisions of the electrons with background neutrals that are invariably present at experimental low pressures inside the trap. The first set of numerical experiments of this paper (Sec III) are PIC-with-MCC simulations that investigate how collisional relaxation of the electron cloud's profile can influence dynamics of the ion resonance instability happening due to the ion contamination of the cloud. In the second set of numerical experiments of this paper (Sec IV) we initiate electron clouds in stable \(l=1\) (\(l\) being the diocotron mode number) mode orbits of very small radius, and utilize the MCC to simulate non-ionizing elastic and excitation collisions of these electrons with a neutral background in course of the simulations. The objective of this set of numerical experiments is to test if such non-ionizing collisions can destabilize the stable motion of the pure electron clouds. It had been theorized by Davidson and Chao that elastic collisions between electrons and background neutrals will not only cause relaxation of the electron cloud's profile but can also destabilize any small-amplitude, stable \(l=1\) mode azimuthal asymmetry present in the cloud into a growing \(l=1\) mode by virtue of their capability of extracting potential energy from the cloud. Contrary to these theoretical predictions, our simulations show that such non-ionizing collisions between electrons and background neutrals can not by themselves destabilize a stable configuration of the cloud. We believe that this has got to do with the manner in which potentially energy of the cloud gets reduced by such collisions. A descriptive comparison of the potential energy extraction process between non-ionizing collisions and a resistive wall instability that is known to be capable of destabilizing the cloud, will make this point clear. This is presented in Sec IV of this paper. A brief outline of the following Sections of this manuscript is as follows. Sec. II is a description of the 2D3v PIC code with newly incorporated MCC facility. The PIC-with-MCC simulations of partially neutralized electron clouds and initially axially asymmetric pure electron clouds will be presented in Sec. III and Sec. IV respectively. The numerical MCC scheme used to produce the results of Sec. III and IV is validated statistically and numerically in Sec. V. The MCC verification procedure is carried out for one of the experiments of Sec. III. A summary of the simulation results and related discussions make up the last section of the paper, Sec. VI. # The 2D3v PIC code with MCC facility Different technical aspects of our PIC-with-MCC code, and their signifance are described in the following sub-sections. ## From the 2D PIC code to the 2D3v version A previously developed 2D Electrostatic PIC has been upgraded to a 2D3v PIC code with facility for Monte-Carlo-Collisions for the purpose of simulating the collisional interactions of nonneutral plasmas with background neutrals along with the collisionless dynamics of the plasma in cylindrical trap cross-sections. The 2D PIC code (earlier version) was developed in FORTRAN-90 using Cartesian co-ordinates, and parallelized with OPEN-MP. It can simulate cross sections of multi-component plasmas of varying neutrality, confined within any perfectly conducting closed boundary curve. The size, shape, and toroidal aspect ratio of the boundary can be manoeuvred as per requirements. The numerical schemes employed in the development of different parts of the code along with references to the sources of these numerical procedures is described Sec. II of Reference 3. The 2D code has passed rigorous benchmarking with analytical results for cylindrically trapped pure electron plasmas as well as partially neutralized electron plasmas. The 2D3v PIC-with-MCC version of the code retains all the features of the original 2D code. In subsections B to D we will discuss only those new features of the 2D3v PIC-with-MCC version that have been applied for the numerical experiments of this paper. For convenience of describing these numerical features of the code we will be explicitly referring to the computational particles of simulation as pseudo-electrons/ions throughout Sec. II instead of loosely calling them electrons/ions as in the rest of the paper. The last two subsections, E and F, describe the setup of the numerical experiments of Sec. III-IV and the numerical diagnostics used in the simulations respectively. ## Giving the pseudo-electrons an axial velocity component perpendicular to the 2D PIC plane In addition to the 2D Cartesian components of velocity the pseudo-electrons now have a \(3^{rd}\) velocity component representing their axial velocities. A common fixed magnitude of this component is initially given to all pseudo-electrons. The axially up or down direction for this component is selected randomly using a pseudo random number generator. The axial velocity does not get modified by the 2D PIC dynamics of the pseudo-electron. It can get modified only when the pseudo-electron is scattered by a collisional event. Thus the extra axial component, with a suitable choice of magnitude, will have a very similar effect on the collisional dynamics of the pseudo-electrons in simulation, as the axial bouncing of electrons between electrostatic end plugs have on the collisional dynamics of electrons in cylindrical trap experiments. ## The MCC procedure The MCC is a technique of making computational particles (here pseudo-electrons) undergoing guided motion in a simulation collide at suitable intervals (collision time step) with background matter that is not represented in any form in the simulation (here neutral Argon atoms). The method is useful when we are interested only in the effect (through collisions) of the latter on the former and not vice versa. To set up the MCC within the framework of a PIC code, the computational particles (here pseudo-electrons) of the simulation have to be treated as real particles (here electrons) inside the MCC algorithm. This implies that mass of a pseudo particle entering the MCC routine gets divided by its representation number. MCC can be implemented using a pseudo random number generator if the collision cross-section of a particle (here electrons) is available in analytical or numerical form as a function of its kinetic energy. Taking the average kinetic energy of a particle (particle here being a pseudo-electrons represented as an electron by scaling down its mass) in the collision time step its collision probability in that interval is determined from the available analytical/numerical collision cross-section. Then the pseudo random number generator churns out a pseudo random number that decides whether the particle undergoes a collision or not based on its collision probability (the Monte-Carlo method). If there are more than one kind of collision possible between the particle and the background then the generated random number can also decide between different possible types of collision and the possibility of no collision for the particle. Of course the probability of each type of collision in the collision time step should be available. For the present set of experiments only elastic collisions and the most abundant first level excitation collisions of background atom has been considered. If the fate of the particle is found to be no collision then its velocity components are left untouched else the code proceeds to the next step of the MCC which decides how the particle gets scattered by the collision. If the collision happens to be of the elastic type then using a set of numerical formulas relying on pseudo random numbers the scattered speed and direction of the particle from the collision is determined. These formulas are fed with the incident kinetic energy of the colliding particle and also the fixed temperature and pressure of the background gas to produce the required outputs. In our MCC code the set of formulas for elastic collision execution are very realistic. They implement the change in kinetic energy of the particle as a result of the elastic collision as well as the natural anisotropy of the scattering with respect to the incident direction of the particle. The scattering angle is determined from a very accurate formula for differential scattering cross section introduced in the year 2002. This formula is actually a correction on the differential scattering cross section formula that had been in use earlier. The colliding neutral though massive (in comparison with electron) is not approximated as a stationary body in the collision but its inherent thermal velocity is also taken into account in the scattering procedure through pseudo random numbers. If on the other hand, the collision happens to be a first excitation of the background (Ar atom), then the MCC follows a two step algorithm to simulate the collision. First the kinetic energy of the particle (pseudo electron with mass scaled down to electron mass) is reduced by an amount equal to the first excitation energy (= 11.55 eV for Ar atom). Next the particle with reduced velocity goes through an elastic scattering procedure as explained above. The first-level excitation collisions have been grouped together with, and treated as elastic collisions in the numerical experiments of this paper. This is because there is an elastic collision happening in every exciting collision adding to the relaxation of the electron component profile which is the subject of interest here. After the MCC procedure is complete the masses of the computational particles are again scaled up to their original values for the regular PIC part of the code. The procedure adopted for the MCC in the PIC-with-MCC code can be summarized as follows: - Scale down the mass of the pseudo electron to that of an electron - Now calculate the electron's kinetic energy - Calculate the probability of elastic collision, probability first level excitation collision, and probability of no collision for that kinetic energy of electron - Generate a pseudo random number to decide the fate of the particle - a) For elastic collision use pseudo random number based formulas to decide kinetic energy after scattering and scattered direction of the electron b) For first level excitation reduce energy of the electron by a magnitude equal to first level excitation energy of the neutral and then use the elastic collision formulas to determine final scattered speed and direction c) For no collision leave the velocity components of the particle untouched - Scale up the mass of the particle to the representative mass of pseudo electron for the regular PIC part of the code ## The collision time interval The interval of time representing the collision step is a critical parameter which determines how closely and how efficiently the MCC simulates the collisional interaction of particles with the background. The first criterion the collision time step should satisfy is that its value should be so small that possibility of simulation particles colliding more than once with the background matter in this interval is practically removed all through the simulation. Otherwise the MCC routine will overlook possible collisions leading to divergence of the simulation from a real experiment. It can be calculated that a collision time step in which the highest (among particles) probability of a collision remains below the value \(0.095\) will satisfy this first criterion. At the same time one should avoid diminishing the size of the collision time step beyond the point at which average (over particles) probability of collision in the interval becomes less than \(1/N\), where \(N\) is the size of the colliding pseudo-particle population. This is because such a choice will result in too many executions of the MCC routine in the simulation period with zero or very few pseudo particles actually colliding in each execution. Hence it will bring down the speed of execution of the PIC-with-MCC code. However improving efficiency of the MCC by increasing size of the collision time step should only be attempted as long as the primary criterion (related to accuracy of simulation) is satisfied for all pseudo particles. In our simulations the pseudo electrons are loaded in rigid rotor equilibrium with a fixed angular velocity and a common fixed magnitude of the axial velocity component. Hence a pseudo electron lying at the half-radius of the cylindrical, uniform density electron cloud will have the mean speed and the average collision cross section for electrons in the cloud. Hence in order to ensure that the first criterion is satisfied for our chosen collision time step we checked that the collision probability of the particles at half-radius of the cloud is well below the \(0.095\) limit. Also the size of the collision time step was adjusted so that on a average a fraction of \(1.197\times10^{-3}\) from the total pseudo electron population of \(87834\) collides in every collision step, thus satisfying the second criterion as well. ## The numerical experimental setup and its PIC parameters All simulation experiments described in sections III and IV of this paper were carried out for typical parameters found in experiments of cylindrical traps, in particular, a wall radius, \(R_w=0.125m\), and axial magnetic field, \(B_z=0.015T\), and radial extent of the uniform density plasma, \(r_p=0.5\times{R_w}\). The simulation time step, \({\delta}t=10^{-11}sec\) is chosen much smaller than the cyclotron time period, \(T_{ce}=2.38\times10^{-9}sec\) of electrons, such that the code can well resolve the cyclotron motion of both ions and electrons. The collision time step \({\Delta}t_c\) is adjusted according to the chosen pressure and temperature of the background gas. The magnitude of axial velocity with which the electrons are loaded is \(1.02727\times10^7 m s^{-1}\) which implies that for a typical cylindrical plasma column of length of \(60 cm\) the bounce frequency of electrons is roughly \(8.5\) MHz. This value of the axial bounce frequency is higher than the typical experimental bounce frequency (\(<0.2\) MHz) of a \(60 cm\) long cylindrical plasma column. By loading the electrons with high axial speeds we have made the electrons more energetic and thereby maintain the average total non-ionizing (elastic + excitation) collision cross section between the moderate orders of \(10^{-21}\,m^2\) and \(10^{-20}\,m^2\) (see Fig. 3 of Reference 30) in the numerical experiments of Sec III and IV. It must be re-emphasized here that the loaded value of the axial velocity of electrons in the simulations serves only to adjust the collisional probabilities and does not partake in any of the regular PIC dynamics of the plasma. PIC-parameters used in the simulations of Sec. III are \(87834\) pseudo particles for each plasma component, on a \(70\times70\) grid. As the actual population of the ions is a fraction, \(f\) of the electron population, the number of real ions represented by each ion pseudo-particle has also been scaled down to a fraction, \(f\) of the representation value of electron pseudo-particles. The ion species (H+ ion) is \(1836\) times heavier than an electron and the background atom (Ar atom) is \(72820.77\) times heavier than an electron. In Sec. IV, the simulations are for pure electron plasmas represented by \(87834\) pseudo electrons on a \(70\times70\) grid. The background species is again Argon. ## Diagnostics for the simulations The diagnostics that will be used in the experiments of Sec. III and IV are 4 potential probes which are basically 4 azimuthally equispaced Cartesian cell nodes inside the cylindrical trap that are all located at a common radius that is very close to the wall. As per their location on the trap cross-section the 4 potential probes have been named as left probe, right probe, top probe, and bottom probe. The potential probes signal are used to study the dynamics of the plasma, and find the frequencies and growth rates of diocotron modes excited on the plasma. Besides the potential probes the energetics (kinetic and potential energies versus time) of both components of the plasma, and the radius of the centre-of-mass of the electron component as a function of time have also been used as diagnostics in Sec. III and IV. # Ion resonance instability in presence of electron-neutral elastic collisions For the numerical experiments of this Section we loaded two uniform density 2-component (e-and H+) nonneutral plasma profiles in their respective rigid rotor equilibrium at different background gas pressures. The first profile was loaded with a value of electron component Brillouin ratio, \(f_b=0.02\) and a fractional neutralization, \(f=0.15\) provided by the uniformly mixed H+ ions. The second profile had values \(f_b=1.0\) and \(f=0.1\). In both cases the electron and the ion components of the cylindrical, uniform density nonneutral plasma are rotating in the slow mode of the mutual rigid rotor equilibrium. Hence the electron component is rotating with angular velocity \(\omega_{re}^{-}\) while the ion component is rotating with angular velocity \(\omega_{ri}^{-}\) in mutual equilibrium. Davidson and Uhm's linear analytical theory, when applied on the first equilibrium (\(f_b=0.02\), \(f=0.15\)) at zero background pressure (absence of neutrals) predicts growth of an unstable \(l=1\) azimuthal mode (\(l\) being the mode number) on both components with angular frequency \(2.827\times10^6 rad/sec\) and growth rate \(9.47\times10^5 rad/sec\). Our simulation of the same equilibrium sans neutrals are in excellent agreement with the linearised theory in the linear phase of the instability. We get an unstable \(l=1\) mode with angular frequency \(2.792\times10^6 rad/sec\) and growth rate \(9.6\times10^5 rad/sec\). A linear perturbation analysis of the second equilibrium (\(f_b=1.0\), \(f=0.1\)) at zero background pressure predicts that this equilibrium should be stable. However the corresponding simulation reveals that this equilibrium infact destabilizes very slowly in the the form of a nonlinear \(l=1\) mode accompanied by a few higher weaker modes. A linearly fitted growth rate of the impure (mixed with a few higher weaker modes) \(l=1\) mode is obtained as \(4.0\times10^5 rad/sec\) while its angular frequency comes out to be \(2.733\times10^7 rad/sec\). The slower nonlinear destabilization of the second equilibrium in absence of neutrals can be attributed to nonlinear terms in the perturbed equations of motion that have been neglected in its linear analysis. That the destabilization of the second equilibrium is not a numerical artefact has been verified in our previous work by repeating the experiment with greater number of pseudo particles for both components. The repeated simulations also gave results in perfect agreement with original simulation at zero background pressure. We now subjected these two unstable equilibria to three different values of neutral background pressures, \(P_{bg}\). The values of \(P_{bg}\) chosen are \(2\times10^{-8} Torr\), \(2\times10^{-7} Torr\), and \(2\times10^{-6} Torr\) at a common fixed background temperature \(T_{bg}=300 K\). The collision time step, \({\Delta}t_c\) for the lowest background pressure was chosen as \(6\times10^{-6} sec\) and it was reduced by an order of magnitude and two orders of magnitude for the \(2\times10^{-7} Torr\) and \(2\times10^{-6} Torr\) background pressures respectively. This kind of scaling of \({\Delta}t_c\) between the three background pressures helped in maintaining uniformity of the MCC routine's collision resolving capacity among the different experimental pressures. Ofcourse all the three values of \({\Delta}t_c\) satisfy the two criteria for collision time interval in their corresponding experiments. With the help of these PIC-with-MCC simulations we were able to study the effect of electron-neutral elastic (and exciting) collisions on two ion resonance instabilities that are growing via the fundamental \(l=1\) mode at different rates. A point to be noted here is that in these sets of experiments the loaded ions are not of the same species as the background gas while in typical electron plasma traps the destabilizing ions are created from the ionization of the background atoms by electrons. Simulations of the latter kind is beyond the scope of this paper and has been reserved for our next work. For the present work we chose the ion species H+ for which the ion resonance instability is a theoretically and numerically well understood phenomenon. The background gas with which the electrons make non-ionizing collisions is Argon. In our upcoming work we will be studying the effects of electron impact ionization of the Argon atoms on the dynamics of the plasma. As there are no ionizing collisions in the present set of numerical experiments the neutrals will only influence the instability through collisional relaxation of the electron cloud's profile. So the experiments of this Section can be described as a qualitative investigation of how such a process of profile relaxation of the electron cloud due to an arbitrary neutral species influences the ion-resonance instability due to an arbitrary ion contaminant mixed in the electron cloud. The detailed linear and nonlinear dynamics of the first (\(f_b=0.02\), \(f=0.15\)) equilibrium in the absence of any colliding neutrals has been described in detail in Sec. IV-subsection A of Reference 3 along with corresponding snapshots of the two-component plasma in course of the simulation (Fig 1 of Reference 3). In brief, this unstable equilibrium leads to an in-phase excitation of the \(l=1\) mode on both components of the plasma. The two components gradually get separated azimuthally (growing phase separation) as the orbital radius of the \(l=1\) mode increases for both components. During the nonlinear growth phase of the mode, ions gain potential energy and are also collisionlessly heated by the instability. The gain in potential energy and kinetic energy of the ion component comes at the expense of the potential energy of electrons. The heated ion component develops a filamentation structure on its surface which eventually makes contact with the grounded wall during the unstable \(l=1\) orbital motion. Soon after the entire ion population gets transported and lost to the grounded wall through this filamentation structure. The \(l=1\) orbit of the electron component however stabilizes very close to grounded wall with the electron component profile deforming into an elliptical shape. Hence the final state of the ion resonance instability for this particular unstable equilibrium is an elliptical pure electron cloud making stable orbits very close to the grounded wall. Now we load this equilibrium in the presence of an Argon neutral background at the three experimental background pressures. For the three PIC-with-MCC simulations the electrons are also loaded with an added axial velocity component enabling them for 3D elastic and exciting collisions with background neutrals. What we observe from our diagnostics comprising of potential probes (Fig 1), radius of centre-of-mass of the electron component, and the potential energy of the electron component, is that there is negligible variation in the dynamics of this instability brought about by the electron-neutral non ionizing collisions for the above set of background pressures. The diagnostics record values with progression of the simulation that are identical to the zero pressure experiment for all the three experimental background pressures. More specifically for all the experimental \(P_{bg}\) values including \(P_{bg}=0\), the potential probes record a similar exponential growth and saturation of \(l=1\) mode, the potential energy of the electron cloud decreases and saturates similarly with time, and the radial location of the electron cloud also increases and saturates similarly with time. As a illustrative example, we have shown in Fig. 1 the left potential probe's reading for the lowest and the highest background pressures (\(P_{bg}=0\) and \(P_{bg}=2\times10^{-6} Torr\)). It can bee seen that the the two signals lie nearly on top of one another. All these diagnostic results bring us to the conclusion that for this particular equilibrium at the chosen set of background pressures the electron-neutral elastic and exciting collisions can not influence the evolution of the ion resonance instability on the e-H+ cloud. We will get back to reason for this in a later comparison with the results for the second equilibrium. Next we come to the set of experiments performed at \(f_b=1.0\), \(f=0.1\). Loading this equilibrium at zero pressure produces a slower growing \(l=1\) mode than the \(f_b=0.02\), \(f=0.15\) equilibrium. Fig. 2 is a set of snapshots showing the evolution of the 2-component plasma profile for this unstable equilibrium at zero background pressure. The snapshots show that the \(l=1\) mode is eventually excited after the plasma profile goes through a few higher weaker modes (the \(l=3\) mode being the most visible in Fig. 2d). The \(l=1\) mode gradually takes over as the most dominant mode (Fig. 2h-i) and towards the end of the simulation it is the only excited mode on the profile executing stable orbits (Fig. 2j-l) after having saturated its growth. We analysed the growth and frequency of the \(l=1\) mode using our potential probes. Fig. 3a shows the signal recorded by the left potential probe for this simulation at zero background pressure. As explained earlier for this particular equilibrium the excited \(l=1\) mode is nonlinear and mixed with other modes from the outset. However it shows a slow growth rate right up to its saturation phase (Fig. 3b). So it is worthwhile to get an estimate of the growth of this impure \(l=1\) mode with a linear exponential fit on the growing part of the potential probe signal (Fig. 3c). This growth rate comes out to be \(4.0\times10^5 rad/sec\). We also performed a FFT on the truncated reading of Fig. 3b to obtain an estimate of the angular frequency of the mode which comes to be \(2.733\times10^7 rad/sec\) (Fig. 3d). The energetics of the simulation at zero background pressure is shown in Fig 4. Here we can see that the nonlinear growth of the mode is associated with collisionless heating of the electrons by the instability at the expense of their potential energy. A smaller fraction of the potential energy of electron component is also pumped into the potential energy of the ion component during the growth of the mode. The energy exchange process stops with the saturation of the mode. Now the \(f_b=1.0\), \(f=0.1\) equilibrium is loaded at the same experimental background pressures as the \(f_b=0.02\), \(f=0.15\) equilibrium. The electron now have cylindrical trap like axial velocities and are capable of 3D elastic and excitation collisions with the background neutrals. The collisions relax the profile of the electron cloud and the resultant changes in the electron density profile influences the ongoing collisionless ion resonance instability. Specifically collisional relaxation of the electron cloud reduces the average density of the electron cloud and thereby increases the fractional density of the ions mixed in it. The dynamically changing electron density and fractional density of ions feed back on the ongoing ion-resonance (two-stream) instability between the two components of the nonneutral cloud and produce deviations in the paths of progression of the instability that are uncorrelated at different background gas pressures. As the variations in the path of evolution of the plasma at the experimental background pressures are small in comparison to average evolution over time it is sufficient to identify and discuss these variations only qualitatively without quantifying their magnitudes with direct measurements. We will discuss these variations with the help of the diagnostic results of Fig. 5. First let us look at radial location of the electron component as a function of time which directly shows how the \(l=1\) mode grows in orbital radius and saturates at different background pressures for the electron component (Fig. 5a). It can be seen from Fig. 5a that the \(P_{bg}=0\) load has the slowest increase in orbital radius of the electron cloud with the growth rate increasing the in the order of \(P_{bg}=2\times10^{-7} Torr\), \(P_{bg}=2\times10^{-6} Torr\), and \(P_{bg}=2\times10^{-8} Torr\). After the growth period the \(l=1\) mode for the lower background pressures of \(P_{bg}=0\) and \(P_{bg}=2\times10^{-8} Torr\) saturate at slightly smaller orbital radii as compared to the two higher background pressures of \(P_{bg}=2\times10^{-7} Torr\) and \(P_{bg}=2\times10^{-6} Torr\) (see inset of Fig. 5a). The potential probe readings also mirror the same information about the growth of the mode (Fig. 5b). Next we have also compared the time evolution of two energy components between the different background pressures, viz the potential energy energy of the electron cloud and the potential energy of the ion cloud. We have already seen in Fig. 4 that for the \(P_{bg}=0\) load the potential energy of the electron component decreases while that of ion component increases during the growth of the instability. Fig. 5c and 5d show the potential energies of the electron component and ion component respectively upto the the growth phase of the instability for the experimental background pressures. It can be seen from Fig. 5c-d that the rates of decrease/increase in potential energy of electrons/ions increases with increase in background pressure. Hence we can conclude that changing the collisionality of the electron cloud by changing the background gas densities has brought about interesting subtle variations in the dynamics of the ion-resonance instability that are unique to each experimental background gas pressure. It is also to be noted that once the growth of the collisionless ion resonance instability saturates there is no further destabilization caused by the continuing collisions (see inset of Fig. 5a) which further emphasizes that the collisions are only influencing the dynamics of the cloud through feedback of electron cloud's profile relaxation on the growing ion resonance instability. We have seen that elastic and excitation collision of electrons with background neutrals influence the dynamics of the ion resonance instability for the \(f_b=1.0\), \(f=0.1\) equilibrium while they have no effect on the instability dynamics for the \(f_b=0.02\), \(f=0.15\) equilibrium. From the nature of the growth of instability in the two equilibria it is evident that this is a due to the different lengths of growth phase in the two equilibria. The excited \(l=1\) mode in the \(f_b=0.02\), \(f=0.15\) equilibrium grows faster and saturates quicker (Fig. 1) than the \(f_b=1.0\), \(f=0.1\) equilibrium (Fig. 3a). So at a particular background pressure there will be more elastic and excitation collisions occurring during the growth phase of the instability for the \(f_b=1.0\), \(f=0.1\) equilibrium as compared to the \(f_b=0.02\), \(f=0.15\) equilibrium. Hence there is more feedback on the instability by the collisional relaxation of the electron cloud for the \(f_b=1.0\), \(f=0.1\) equilibrium. The feedback process brings about the subtle changes in the evolution of this equilibrium at the experimental background pressures. Furthermore we have already seen that once saturation of the mode is achieved, the ongoing collisions can not further destabilize the cloud. Hence a smaller time window of influence of the electron-neutral collisions for the \(f_b=0.02\), \(f=0.15\) equilibrium is the precise reason why we do not observe any deviation in dynamics at different experimental background pressures for this equilibrium. # Dynamics of \(l=1\) perturbed pure electron clouds in presence of collisional background In Sec III we have explained that the elastic and exciting collisions can influence the dynamics of the plasma through a feed back of the profile relaxation on the ongoing instability. But they can not by themselves destabilize the stable orbital motion of the plasma once saturation of the instability is reached. In this section we will put this reasoning to the test and verify that these collisions are indeed incapable of exciting an instability on the plasma profile. For this purpose we loaded two uniform density cylindrical pure electron profiles with a very small \(l=1\) perturbation seeded from the outset *i.e.* the profiles' central axis were shifted from the trap's central axis by a very small distance of \(0.1\times{R_W}\). The Brillouin ratios of the profiles chosen were \(f_b=0.9\) and \(f_b=0.02\). It is well understood that in the absence of any other influence these profiles will continue to make stable \(l=1\) obits at their preset orbital radius with their characteristic \(l=1\) mode frequency. In our simulations we loaded these profiles at the highest experimental background pressure used in Sec III *i.e.* \(P_{bg}=2\times10^{-6} Torr\) to ensure maximum exposure of the profiles to elastic and excitation collisions during the simulations. However as expected there was no destabilization of the stable orbital motion caused by the collisions for both the high and the low Brillouin ratios. Fig. 6 shows the squared and logscaled plot of the reading of the left potential probe for these two simulations. The signals have been plotted in such a manner for accurate exponential fitting of any observable growth of the signals (see also Fig. 3d) and also for the purpose of depicting on the same graph the two signals with vast difference in their mean values as well as their amplitudes. Fig. 6 shows that both the signals remain perfectly stable throughout the simulations. Our simulation results also contradict a theoretical model by Davidson and Chao for the effects of electron-neutral elastic collisions on the dynamics of the cloud. This model predicts that the non-ionzing collisions should destabilize the initially stable \(l=1\) orbital motion in the above two simulations into a growing \(l=1\) mode with growth rate, \(\gamma=2{\pi}{\nu_{en}}{\omega_{0}}/{\omega_{ce}}\), where \({\nu_{en}}\), \({\omega_{ce}}\), and \({\omega_{0}}\) are the non-ionizing collision frequency in \(s^{-1}\), cyclotron frequency in \(rad/s\) and characteristic frequency of the \(l=1\) mode in the loaded electron cloud expressed in \(rad/s\). The actual theory considered only electron-neutral elastic collisions contributing to the collision frequency \({\nu_{en}}\). In applying the model to our numerical experiments we have suitably included the effectively elastic-like excitation collisions along with the regular elastic collisions in estimating an experimental value of \({\nu_{en}}\). The theoretical model also assumed a constant temperature of the plasma profile maintained by electron-electron collisions in the derivation of the above expression for the growth rate, \(\gamma\) due to electron-neutral collisions. We do not have electron-electron collisions happening in our simulations. But the plasma is loaded at zero temperature and does not undergo any collsionless heating process, although there is some amount of heating by the non ionizing electron-neutral collisions. So applying the above expression for \(\gamma\) to our numerical experiments will include the approximation of a constant temperature of the plasma profile. The procedure for obtaining the growth rates from Davidson of Chao's model when applied to our numerical experiments is as follows: From the initial cold loads of the plasma profiles it is simple to find an average velocity magnitude of electrons which we approximate as the average velocity magnitude all through the simulation. The total non-ionizing (elastic + excitation) collision cross-section can then be estimated using this average velocity magnitude, which leads to a value for the collision frequency \({\nu_{en}}\). This value comes out to be \(2.3985\times10^4\) per sec for the \(f_b=0.02\) load and \(1.7548\times10^4\) per sec for the \(f_b=0.9\) load. Plugging in the other fixed frequencies in the above expression for the growth rate, \(\gamma\), we can estimate the growth rate from Davidson and Chao's model to be \(1.8838\times10^2\) rad/sec for the \(f_b=0.02\) load and \(6.2021\times10^3\) rad/sec for the \(f_b=0.9\) load. These growth rates imply that at the end of the simulations period of \(5.4\times10^{-5}\) seconds, the \(l=1\) orbital radius should have increased by \(1.02\%\) for the \(f_b=0.02\) load and by \(39.78\%\) for the \(f_b=0.9\) load according to the model of Davidson and Chao. Even if we neglect the theoretical small increase in orbital radius for the \(f_b=0.02\) load, the theoretically predicted increase in orbtial radius is substantial for the \(f_b=0.9\) load and should be detected by the simulation diagnostics. However as we have discussed earlier using the potential probe diagnostic (Fig 6) the stability and initial amplitude of the \(l=1\) mode remains completely unaltered in both the simulations, \(f_b=0.02\) and \(f_b=0.9\). A new theoretical explanation in support of our experimental results and their contradiction with the model given by Davidson and Chao is provided below. The electron plasma in diocotron motion, *e.g.* a stable \(l=1\) orbit, can be thought of as having its electrostatic potential energy in two parts. The first part is the electrostatic energy of the profile without the surface wave/ diocotron mode and the other part which is a negative quantity is the electrostatic energy associated with the diocotron mode. Any process that further draws out energy from the mode *i.e.* makes the mode's electrostatic energy even more negative and thereby also reduces the total potential energy of the profile, will bring about growth of the mode. A process that only reduces the first part of the potential energy *i.e.* the profile's electrostatic energy by virtue of its internal configuration without any surface wave, can not cause growth of the dioctron mode. Through our simulations we found that the elastic and excitation collisions can only reduce this first part of the potential energy by means of profile relaxation. However these collisions can not directly influence the energy associated with diocotron mode. On the other hand a process like a resistive wall instability can directly takes away energy from the diocotron mode by changing surfaces charge configuration on the grounded wall (or alternatively the virtual image charge configuration) that interacts with the diocotron mode of the plasma. So while a resistive wall may cause growth of negative energy dioctron modes on the electron plasma the elastic and excitation collisions can not directly influence diocotron motion by virtue of their potential energy reduction capability. Hence contrary to the theory by Davidson and Chao which predicts that elastic collisions can reduce both parts of the potential energy of an electron cloud and destabilize them, our simulations show that elastic (and excitation) collisions can only cause reduction in the first part of the potential energy through relaxation of the profile. # Validation of the MCC numerical scheme

{'timestamp': '2016-07-19T02:06:51', 'yymm': '1607', 'arxiv_id': '1607.04843', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04843'}

# abstract {#abstract .unnumbered} Light possesses both spin and orbital angular momentum (AM). While spin AM is determined by helicity of circular-polarization, orbital AM is characterized by topological charge of vortex beam. Interaction of AM with optical beam orbit leads to optical spin Hall or orbital Hall effect, exhibited as spin-dependent or topological charge-dependent transverse shift of optical beam. Conservation of AM enables spin-to-orbital AM conversion, where circular-polarized Gaussian beam is converted to opposite-helicity circular-polarized vortex beam with topological charge \(\pm 2\), an example of controlling spatial beam profiling by spin flip. However, the resultant vortex beam has the beam center of gravity unchanged, the same as that of incident Gaussian beam, meaning a null transverse shift. Here we introduce a cyclic group symmetric metasurface to demonstrate generation of vortex beam exhibiting spin-dependent transverse shift, namely, spin-and orbital-Hall effect, attributed to an alteration of dynamical phase of scattered beam according to the order \(n\) of cyclic group while keeping geometric phase constant. Capability of spin-controlled spatial beam profiling with a transverse shift via spin-and orbital-Hall effect has important implications for spatial demultiplexing in optical communication utilizing orbital AM mode division multiplexing as well as for optical vortex tweezer and signal processing involving vortex beams. # Introduction {#introduction .unnumbered} Circularly polarized (CP) light possesses spin angular momentum (SAM) with a helicity defined as its projection to linear momentum of propagating optical wave. In addition to SAM, optical wave can be rendered to possess orbital angular momentum (OAM) by constructing a vortex beam having helical wavefront. Both SAM and OAM are utilized to carry informations in optical communication via polarization-division or OAM-division multiplexing. Differently from 2 modes of SAM with spins of \(\pm 1\), the number of modes of OAM is not restricted since the topological charge of vortex beam is determined by twisting number of the helical wavefront. Owing to the availability of a large number of modes, vortex beams have been widely employed in applications to increase the signal channels for optical communications. In those applications it is necessary to be able to manipulate and control the intensity and phase distributions of vortex beam. Anisotropic planar structures such as metasurface are important optical elements for such applications when SAM is utilized for shaping vortex beam, since interaction of SAM and OAM with optical beam orbit is enhanced. There are several well-known means to create vortex beams, for example, a pair of cylindrical lenses, spiral phase plate, fork-shaped grating, and spatial light modulator as well as photonic integrated devices. Another important means to create vortex beams is to make use of geometric phase by constructing space-variant Pancharatnam-Berry phase optical elements including subwavelength diffraction grating and liquid crystal q-plates in spin-to-orbital AM conversion (SOC) process. Related to metasurface, a high-efficiency SOC has been demonstrated to generate vortex beams with high topological charges in the visible wavelength based on dielectric metasurfaces. Generation of vortex beam by SOC is an example of spin-controlled spatial beam profiling since an incident Gaussian beam converts to a vortex beam with donut-shaped spatial beam profile by spin flip. In case of a rotational symmetric Pancharatnam-Berry phase optical element, SOC-generated vortex beams with opposite signs of topological charges are degenerate in spatial beam profiles, which results from conservation of angular momentum consisting of spin AM and orbital AM. In other words, the beam centers of gravity of vortex beams are the same at that of the incident Gaussian beam. Now we address the question how the degeneracy of spatial beam profiles of SOC-generated vortex beams can be lifted in SOC to provide spin-controlled spatial beam profiling such that the beam centers of gravity of vortex beams with opposite signs of topological charges are spatially separated by transverse-shifting in opposite directions. Optical Hall effect is one means to achieve spatial separation by use of spin-dependent transverse shift originating from SAM-or OAM-orbit interaction, which is also technically important when spatially demultiplexing OAM-mode division multiplexed superposition states. Here, we introduce a cyclic group symmetric metasurface (CGSM) to generate vortex beam exhibiting a spin-and orbital-AM dependent transverse shift by identifying SOC in terms of geometric phase acquired by an optical beam in cross-polarization scattering from ring structure of nano-antennas. By designing CGSMs belonging to the cyclic group \(C_n\), dynamical phase of cross-polarization scattered beam is altered according to the order \(n\) of cyclic group while keeping geometric phase constant. When \(n\)-fold rotational symmetry of azimuthal dynamical phase gradient is an odd integer \(n\), there results a spin-and orbital-Hall effect. # Geometric and dynamical phases in optical Hall effect {#sec:geometirc} Intrinsic optical spin or orbital Hall effect refers to spin-dependent or vortex-charge dependent transverse shift, a consequence of the coupling between SAM or OAM with optical beam orbit originating from transversality of electromagnetic wave.. In a general framework of physical description, it is an effect of fast degree of freedom (SAM or OAM) on slow degree of freedom (optical beam orbit), described by momentum-space Lorentz equation of motion in the presence of a Berry curvature of topological monopole. Spin-dependent transverse shift of optical beam is also achieved by geometric Pancharatnam-Berry (PB) phase in a linear array of nano-antennas, called extrinsic optical spin Hall effect, to distinguish from the transverse shift originating from topological monopole. The amount of shift is related to PB phase as shown in **Fig. [\[constantPB\]](#constantPB){reference-type="ref" reference="constantPB"} a, d, e**. In a circular array of nano-antennas with uniform width and length, even though not exhibiting spin-dependent transverse shift, spin-dependent PB phase is utilized to generate vortex beam by SOC in **Fig. [\[constantPB\]](#constantPB){reference-type="ref" reference="constantPB"} b, f, g**. Polarization states in Poincaré sphere go through a circular trajectory twice providing the solid angle \(\pm8\pi\) corresponding to geometric phase \(\pm4\pi\), generating vortex beam of \(l=\pm 2\), as shown in **Fig. [\[constantPB\]](#constantPB){reference-type="ref" reference="constantPB"} c**. Optical wave scattered from the circular array acquires a phase factor, entirely geometric phase coming from a circular closed path \(\textbf{C}\) in Stokes parameter space, leading to \(\oint d\gamma(\textbf{C}) = \pm 4\pi\) with a constant azimuthal PB phase gradient \(\nabla_\phi \Phi_{PB} = \pm 2\). Therefore, the intensity profiles of \(I_{-+}(r,\phi)=|{E}_{-+}|^2\) and \(I_{+-}(r,\phi)=|{E}_{+-}|^2\) are degenerate even though helical phases are opposite for \(l=\pm 2\), where \(\{+-\}\) stands for left-circular polarization (LCP, \(\sigma_{+}\)) scattering / right-circular polarization (RCP, \(\sigma_{-}\)) incidence and \(\{-+\}\) vice-versa. In fact, \({E}_{-+}(r,\phi) = J_2(k_0r)e^{-{k_0}^2r^2}\exp(i2\phi)\) and \({E}_{+-}(r,\phi) = J_{-2}(k_0r)e^{-{k_0}^2r^2}\exp(-i2\phi)\), where a constant \(\nabla_\phi \Phi_{PB}\) leads to \(\phi\)-independent amplitude \(J_{\pm 2}(r)e^{-{k_0}^2r^2}\) from the fact that each two neighboring nano-antennas provide an equal amount of PB phase during one full cycle, owing to rotational invariance of circular array of nano-antennas. It has been reported that a spin-dependent transverse shift of vortex beam in SOC is observed by lifting degeneracy in OAM though rotational symmetry breaking without clarifying the underlying role of azimuthal phase gradient in giving rise to extrinsic spin-and orbital-Hall effect. Now we examine how the azimuthal phase gradient is closely related to the existence of extrinsic spin-and orbital-Hall effect. Consider a circular array of nano-antennas with different sizes from \(1\) to \(16\) in the range of \(0 \le \phi \le 2\pi\), as shown in **Fig. [\[nonconstantPB\]](#nonconstantPB){reference-type="ref" reference="nonconstantPB"} a**. Scattering wave from the circular array acquires a phase factor originating from both geometric phase \(\gamma(\textbf{C})\) and dynamical phase \(\Phi_D\), where \(\gamma(\textbf{C})\) is wavelength-independent while \(\Phi_D\) has a frequency dispersion related to plasmonic resonances of each nano-antenna (see Supplementary Information). When rotational symmetry is broken, azimuthal dynamical phase gradient \(\nabla_\phi \Phi_{D}(\phi)\) is \(\phi\)-dependent, which results in \(\phi\)-dependent amplitude of scattered wave, generating additional vortex beams with \(l \neq 2\). \[\begin{aligned} E(r,\phi)_{-+} &=& BG_2(r)f(r,\phi) \exp\{i2\phi+ i\Phi_D^{-+}(\phi)\}\nonumber\\ &=& \sum_l^\infty \tilde{a}_l BG_l(r)e^{il\phi}\nonumber \end{aligned}\] where scattering wave is partial-wave expanded in term of vortex beams of topological charge \(l\) with \(BG_l (r) \equiv J_l(k_0r)e^{-{k_0}^2r^2}\), and expansion coefficients \(\tilde{a}_l={a}_le^{i\psi_l}\) are complex, and \(f(r,\phi)=1\) when \(\nabla_\phi \Phi_{D}^{-+}\) vanishes. A similar expression holds for \(E(r,\phi)_{+-}\). Owing to the difference in scattering amplitudes from nano-antennas with different sizes, polarization states of scattered waves form a spiral trajectory on Poincaré sphere as shown in **Fig. [\[nonconstantPB\]](#nonconstantPB){reference-type="ref" reference="nonconstantPB"} b**. In other words, solid angles subtended by polarization states between two neighboring nano-antennas are not constant. Rotational symmetry breaking introduced by nano-antennas with different sizes yields a spin-dependent non-constant azimuthal PB phase gradient, \(\nabla_\phi \Phi_{PB}\), permitting spin-dependent deflections in SOC as shown in **Fig. [\[nonconstantPB\]](#nonconstantPB){reference-type="ref" reference="nonconstantPB"} c**. Analysis of PB phase in rotation symmetry-broken circular array of nano-antennas shows that \(\phi\)-dependence of beam profiles \(I_{-+}(r,\phi)\) and \(I_{+-}(r,\phi)\) is determined by the total phase \(\Phi(\phi)^{-+} = 2\phi + \Phi_D^{-+}(\phi)\) and \(\Phi(\phi)^{+-} =-2\phi + \Phi_D^{+-}(\phi)\) acquired by cross polarization scattering beam, where non-constant azimuthal dynamical phase gradients \(\nabla_\phi \Phi_{D}\) are responsible for \(\phi\)-dependence behavior of beam profiles. The occurrence of spin-dependent transverse shift, equivalent to non-vanishing \(\Delta I \equiv I_{-+}-I_{+-}\), depends on the relation between \(\Phi_D^{-+}(\phi)\) and \(\Phi_D^{+-}(\phi)\) dictated by symmetry property of the circular array of nano-antennas with different sizes. # Sample design and measurement {#sec:sample} For a systematic study of relation between symmetry property of metasurface and spin-dependent transverse shift, tapered arc (TA) antennas are arranged in a circle to form metasurfaces belonging to cyclic group \(C_n\), called tapered arc cyclic group symmetric metasurface (TA-CGSM), where in-plane inversion symmetry is determined by the order \(n\) of cyclic group, as shown in **Fig. [\[samplestructure\]](#samplestructure){reference-type="ref" reference="samplestructure"} a**. Six different TA-CGSMs are fabricated, each TA-CSGM being composed of multiple TA antennas with varying width from 45\(nm\) to 150\(nm\) organized in \(8\) azimuthal segments of concentric rings repeated with 600\(nm\) radial spacing. We used 1\(mm\) thick borosilicate glass round substrates with diameter of one inch (UQG Optics). First, the substrates were cleaned in sequential bathes of acetone and isopropanol, assisted by ultrasonicating. Then the substrates were rinsed in deionized water and dried by nitrogen flow followed by oxygen radical plasma treatment in a barrel reactor (Nanoplas, France) to activate the glass surface. Second, an electron beam resist (PMMA, ARP 679) and a conductive polymer (SX AR PC 5000/90.2 from Allresist, Germany) layers were successively spin-coated. The thickness of the resulting layer was chosen to be about 70\(nm\) and it was measured by a contact profilometer (DektakXT, Bruker, Germany) after e-beam exposition and resist development. Third, we used an e-beam lithography (EBL) tool (Pioneer, Raith, Germany) for sample patterning. We used 20\(kV\) acceleration voltage of the electron gun, and the beam current was 0.016\(nA\) for the aperture of 7.5\(\mu m\). The typical working distance was about 5\(mm\), and the nominal exposition dose was chosen in the range from 100 to 180\(\mu C/cm^2\). Forth, two successive metal layers (Cr, 3\(nm\) for seeding and Au, 27\(nm\)) were deposited by thermal evaporation (Auto 306, Edwards, UK). The thickness was monitored by quartz crystal microbalance during the deposition and measured by a contact profilometer after lift-off process. The lift-off was done in ethyl lactate solution, followed by rinsing and nitrogen drying. Finally, the samples were observed by optical and scanning electron microscopy (SEM). For SEM observation, we used low acceleration voltage of 3\(kV\) in order to decrease the sample charging because the glass substrate is insulating. (For more information, see Supplementary Information) **Fig. [\[samplestructure\]](#samplestructure){reference-type="ref" reference="samplestructure"} b** shows optical microscope images of TA-CGSMs possessing \(C_\infty\), \(C_1\), \(C_2\), \(C_3\), \(C_4\), \(C_5\), and \(C_6\) symmetries, and SEM images of \(C_\infty\) and \(C_1\) TA-CGSMs are shown in **Fig. [\[samplestructure\]](#samplestructure){reference-type="ref" reference="samplestructure"} c**. Among experimental measurements in **Fig. [\[samplestructure\]](#samplestructure){reference-type="ref" reference="samplestructure"} e, f**, \(C_\infty\) TA-CGSM exhibits degenerate \(\phi\)-independent intensity profiles for \(I_{-+}\) and \(I_{+-}\), while \(C_n (n\neq\infty)\) TA-CGSMs exhibits \(\phi\)-dependent intensity profiles of \(I_{-+}\) and \(I_{+-}\), owing to non-constant azimuthal PB phase gradient \(\Phi_{PB}\) coming from plasmonic resonance of TAs. Difference \(\Delta I = I_{-+}-I_{+-}\) in **Fig. [\[samplestructure\]](#samplestructure){reference-type="ref" reference="samplestructure"} g** corresponds to spin-and orbital-Hall shift, observed only in \(C_1\), \(C_3,\) and \(C_5\) TA-CGSMs (odd order \(n\) of \(C_n\)), where degeneracy of geometric structure is lifted by in-plane inversion symmetry breaking. Vortex charge of cross-polarization scattering beam from \(C_n (n\neq\infty)\) TA-CGSM is identified by an interference pattern between scattering vortex and incidence Gaussian beams, and counter-clockwise and clockwise twisted fringes confirm topological charges of \(l=2\) and \(l=-2\) of \(I_{-+}\) and \(I_{+-}\), respectively. See **Fig. S2** in Supplementary Information. FDTD calculation is in a perfect agreement with experimental results. See **Fig. S1** in Supplementary Information. # Partial wave decomposition {#sec:partialwave} Vortex beam with topological charge \(l\) is described by Bessel-Gaussian beam, \[\begin{aligned} {{J}_{l}}(k_0r){{e}^{-{{{k_0}^2r}^{2}}}}\exp(i l \phi) = {BG}_{l}(r)\exp(i l \phi).\nonumber \end{aligned}\] In order to clarify how the addition of dynamical phase to geometric phase in SOC alters \(\phi\)-dependence of phases \(\Phi^{-+}(\phi)\) and \(\Phi^{+-}(\phi)\), experimentally measured intensity profiles in **Fig. [\[samplestructure\]](#samplestructure){reference-type="ref" reference="samplestructure"} e, f** are fit as a linear superposition of vortex beams, as shown in **Fig. [\[Results1cal\]](#Results1cal){reference-type="ref" reference="Results1cal"} a, b** along with phases in **Fig. [\[Results1cal\]](#Results1cal){reference-type="ref" reference="Results1cal"} d, e**. Spin-dependent vortex beams with asymmetric helical wave front are described by \[\begin{aligned} {{{E}}(r,\phi)}=\sum\limits_{l}{{{\tilde{a}}_{l}}}{{{BG}}_{l}(r)}{{{e}}^{i l\phi}}= A(r,\phi )e^{i{\Phi }_{\rm tot}},\nonumber \end{aligned}\] where \(\tilde{a}_l\) is the \(l\)-th order complex coefficient of the Fourier expansion, i.e., \(\tilde{a}_l={a}_l e^{i\psi_l}\), \(BG_l (r) = J_l(k_{0}r)e^{-{k_{0}}^2r^2}\), \(A(r,\phi)=\left|E(r,\phi)\right|\) is the amplitude, \({{\Phi }_{\rm tot}}= \gamma (\mathbf{C})+{\Phi}_{D} = \tan^{-1}(\textit{Im}(E)/\textit{Re}(E))\) is total phase. It shows that the far fields are determined by the local topological features of interfering fields. The fit values of partial-wave expansion coefficient \(\tilde{a}_l\) are listed in **Table 1**. We observe two important features. First, phases of \(E_{-+}(\phi)\) and \(E_{+-}(\phi)\) are opposite, i.e., \(\Phi^{-+}(\phi)=-\Phi^{+-}(\phi)\), which is from the opposite senses of rotation of polarization state trajectories on Poincaré sphere for \(\{-+\}\) and \(\{+-\}\). Second, intensity profile \(I_{+-}(\phi)\) is the same as that of in-plane inversion symmetry operated \(I_{-+}(\phi)\), i.e., \(I_{+-}(\phi) = I_{-+}(\phi+\pi)\), as a result of \(\Phi^{-+}(\phi)=-\Phi^{+-}(\phi)\). The two features immediately explain why spin-and orbital-Hall effect takes place only for \(C_n\) with odd \(n\). For even \(n\), \(\Delta I\) vanishes identically, since geometric structures possess in-plane inversion symmetry. Now we further clarify the contributions of dynamical phase to spin-controlled spatial beam profiling in SOC. First, we expect a correlation between intensity profile and \(\phi\)-dependence of phase. As shown in **Fig. [\[matlab\]](#matlab){reference-type="ref" reference="matlab"}** for \(C_1\) TA-CGSM, \(E_{-+}(\phi)\) and \(E_{+-}(\phi)\) have a local maximum at the azimuthal angles \(\phi\) where the absolute value \(|\nabla_\phi\Phi_D|\) is a local minimum. Helical wavefronts with azimuthal geometric phase gradient \(\pm2\) are altered by introduction of non-constant azimuthal dynamical phase gradient \(\nabla_\phi\Phi_D\), to build focused spots along azimuthal direction, similar to wavefront modification when a plane wave goes through a medium with spatially varying index. Refer to Supplementary Information for \(C_n\) TA-CGSMs with \(n\) other than \(1\). The intensity profile possesses a feature of OAM superposition states, useful for quantum information systems. Second, the phase \(\psi_l\) of partial-wave expansion coefficient \(\tilde{a}_l\) is related to dynamical phase \(\Phi_D(\phi)\), determining interference behavior between partial-wave vortex beams having difference topological charge \(l\)s. \(C_4\) and \(C_5\) TA-CGSMs are closely examined in the left panel (**Fig. [\[Results1cal\]](#Results1cal){reference-type="ref" reference="Results1cal"} f, g, h, i**) and right panel (**Fig. [\[Results1cal\]](#Results1cal){reference-type="ref" reference="Results1cal"} j, k, l, m**). In \(C_4\) TA-CGSM \[\begin{aligned} E_{-+} = BG_2 \exp\{i 2 \phi\}+ 0.24 BG_{-2}\exp\{-i2\phi\} \exp\{+i{3\pi\over 2}\}\nonumber \end{aligned}\] and \[\begin{aligned} E_{+-} = BG_{-2}\exp\{-i2\phi\} + 0.24 BG_{2}\exp\{i 2 \phi\} \exp\{-i{3\pi\over 2}\},\nonumber \end{aligned}\] where azimuthal angles for constructive and destructive interferences are the same for \(E_{-+}\) and \(E_{+-}\), resulting in \(\Delta I = 0\). In \(C_5\) TA-CGSM \[\begin{aligned} E_{-+} = BG_2 \exp\{i2\phi\}+ BG_{-3}\exp\{-i3\phi\} \exp\{+i\pi\}\nonumber \end{aligned}\] and \[\begin{aligned} E_{+-} = BG_{-2}\exp\{-i2\phi\} + BG_{3}\exp\{i3\phi\} \exp\{-i\pi\},\nonumber \end{aligned}\] where azimuthal angles for constructive and destructive interferences differ by \(0.2\pi\) for \(E_{-+}\) and \(E_{+-}\), and \(\Delta I \neq 0\) leading to non-vanishing spin-dependent transverse shift. Third, a frequency dispersion of dynamical phase due to the plasmonic resonance of TA renders intensity profiles to be dependent on the incidence beam wavelength, confirmed to be true. # Wavelength dependence of extrinsic spin-and orbital-Hall effect {#sec:wavelength} The role of dynamical phase can be identified by examining wavelength dependence of the spin-and orbital-Hall effect since plasmonic resonance of nano-antennas is wavelength-dependent. We measured the spin-and orbital-dependence of far field intensity \(C_1\) metasurface by employing lasers with different wavelengths as shown in **Fig. [\[Results2\]](#Results2){reference-type="ref" reference="Results2"}**. Single-mode fiber pigtailed laser diodes with wavelengths \(\lambda =1,310nm\), \(\lambda =730nm\), and \(\lambda =660nm\) are adopted as normal-incidence illumination sources. Direction of the transverse shift goes through clock-wise rotation as the wavelength gets shorter. While geometric phase is wavelength-independent, dynamical phases originating from multiple TA antennas with varying widths give rise to wavelength-and \(\phi\)-dependent far-field intensity distributions. On the other hand, \({C_1}'\) metasurface possessing \(\sigma_{yz}\) symmetry, i.e., reflection symmetry with respect to \(yz\)-plane, allows for splitting of vortex beam along \(x\)-direction only as shown in **Fig. [\[Results2\]](#Results2){reference-type="ref" reference="Results2"} g-i**, and direction of the transverse shift is fixed along \(x\)-direction as the wavelength get shorter, exhibiting no wavelength-dependence. We note that in \({C_1}'\) only \(y\)-direction dynamical phase gradient gets involved owing to breaking of \(y\)-direction inversion-symmetry. Normalized horizontal \((X)\) and vertical \((Y)\) spin-and orbital-Hall effect profiles, \(\Delta I (X)\) and \(\Delta I (Y)\) along the dashed coordinate axes, are displayed in two right panel, where \(\Delta I\)=\(I_{-+}(r,\phi)-I_{+-}(r,\phi)\). # Conclusion {#sec:conclusion} In summary we demonstrated spin-and orbital-Hall effect in metasurfaces possessing cyclic point group symmetry \(C_n\). Lifting degeneracy in orbital angular momentum is achieved by selectively controlling the symmetry order \(n\). By identifying the role of dynamical phase in giving rise to spin-dependent transverse shift in SOC, spatial separation and modification of vortex beam profile are made possible. This new approach provides significant advantages in applications such as vortex multiplexing in communication device and vortex beam analyzer as well as fundamental understating of interactions among angular momenta of light in metasurface.

{'timestamp': '2016-08-18T02:06:59', 'yymm': '1607', 'arxiv_id': '1607.04806', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04806'}

# Introduction {#sec:intro} The precise measurement of the polarization of the Cosmic Microwave Background (CMB) is an active area of experimental physics, motivating a variety of current and upcoming polarimeter projects. The CMB temperature anisotropy, measured to extreme precision over the last two decades, has provided evidence for the standard model of Big Bang cosmology and hints at an inflationary period. In the same way, the CMB polarization, and in particular the curl 'B-mode' pattern, contains a wealth of cosmological information on many scales and continues to be targeted by experimentalists. At arcminute scales on the sky, the dominant B-mode signal is primarily due to the gravitational lensing of the much brighter, curl-free 'E-mode' signal. This lensing B-mode signal probes the large scale structure of the universe that sits between the CMB surface and ourselves and has recently been detected by P[olarbear]{.smallcaps}, SPTpol, and /*Keck* *Array*. At larger, degree-scales on the sky, the B-mode signal due to inflationary gravitational waves (IGW) is predicted to peak with an as-yet-unmeasured amplitude. During inflation, quantum fluctuations are expected to produce both density and gravitational wave perturbations, the latter of which imprints a B-mode pattern on the CMB. The ratio of the amplitude of the primordial tensor spectrum to the scalar spectrum, parametrized as \(r\), is a measurable quantity that will constrain inflationary models. To this end, experiments aim to continually lower the upper-limit on, and eventually detect, the IGW B-mode signal and the resulting \(r\) parameter, a recent summary of which is given in Ref. . The B[icep]{.smallcaps}/*Keck* series of South Pole-based CMB polarimeters has focused efforts on this IGW B-mode signal, taking the unique approach of using compact, refracting telescope receivers that only resolve the larger degree-scale features necessary. This approach precludes science at the smaller scales but has several advantages for IGW signal sensitivity. The first experiment, , observed for three seasons from 2006--2008 using 98 feedhorn-coupled bolometers (split primarily between 100 and 150 GHz band centers) and found a 95% confidence upper limit of \(r<0.70\). followed with three observing seasons from 2010--2012, using 500 antenna-coupled transition edge sensor (TES) bolometers centered at 150 GHz. It detected on-sky B-mode signal, later found to be consistent with galactic dust foregrounds, for a combined limit \(r<0.12\) at 95% confidence with *Keck* *Array* and *Planck* data. While was still operating, *Keck* *Array* began observing on a nearby telescope mount, with first science observations in 2012. *Keck* *Array* is made up of 5 -class receivers sharing a common boresight. It is currently in its fifth observing season and has now included receivers centered on multiple frequencies (95, 150, and 220 GHz) in order to help distinguish polarized galactic foregrounds from the IGW signal. The latest results give \(r<0.09\) at 95% confidence, or \(r<0.07\) when combined with external datasets.  replaces  and is now in its second observing season (and first science season) during the 2016 austral winter, with a full complement of 2560 TES bolometers at 95 GHz. With its larger aperture, faster optics, and higher detector count,  acts as a pathfinder receiver for the 3rd-generation of the B[icep]{.smallcaps}/*Keck* series. Its relatively compact design with on-axis optics again lends itself to multi-frequency observations, in the same way that  was evolved into the *Keck* *Array*. Such an expansion to an array is highly motivated by the continued need for increasing observation depth combined with discrimination of galactic foregrounds, and is easily realized by upgrading the *Keck* *Array* with -class receivers. In the following sections, these proceedings first give an overview of the instrument (Sec. [2](#sec:inst){reference-type="ref" reference="sec:inst"}) and its South Pole observing location and strategy (Sec. [3](#sec:obs){reference-type="ref" reference="sec:obs"}). It then summarizes 's current, second-season status and performance (Sec. [4](#sec:perf){reference-type="ref" reference="sec:perf"}) before finally outlining plans for a future multi-frequency array of -class receivers (Sec. [5](#sec:array){reference-type="ref" reference="sec:array"}). # Instrument Overview {#sec:inst} Like its B[icep]{.smallcaps}/*Keck* predecessors described in Sec. [1](#sec:intro){reference-type="ref" reference="sec:intro"},  is composed of a single cryostat, also called the receiver, which contains all of the detectors and optical components required for its measurement of the CMB. This compact design, which keeps all optics at cryogenic temperatures, maintains the practical and scientific advantages of the previous B[icep]{.smallcaps}/*Keck* experiments. Namely, - the cryostat lends itself to complete assembly and disassembly by a small crew of experienced people; - the compact, on-axis optics allow for symmetric rotation of the telescope as a whole, an important method for probing systematics in the polarization-based measurement; - the compact optics allow for ground-based characterization in the optical far-field; - the compact receiver size allows installation of a comoving absorptive forebaffle that terminates beam sidelobes and reduces stray pickup; - and the ability to cryogenically cool the optical elements and internal baffling minimizes the thermal photon noise seen by the detectors and leads to excellent stability. This section provides a short overview of the instrument in its current, 2016 observing season state, offered here as both a recap before giving a status update in Sec. [4](#sec:perf){reference-type="ref" reference="sec:perf"}, and to highlight features that will transfer over to a future -class array as outlined in Sec. [5](#sec:array){reference-type="ref" reference="sec:array"}. A more detailed overview before the first (2015) season deployment is found in Ref. , and an in-depth design and operation summary during the 2015 season is provided in Ref. . ## Cryostat Receiver The receiver is a custom aluminum dry cryostat, 2.4 m tall along the optical axis with a 0.73 m outer diameter and totalling roughly 550 kg without attached readout electronics and forebaffle. Fig. [\[fig:b3_rx_cutaway\]](#fig:b3_rx_cutaway){reference-type="ref" reference="fig:b3_rx_cutaway"} shows a cutaway view of the CAD model in this state. The top of the receiver is capped by a 31.75 mm thick high-density polyethylene (HDPE) window through which light is collected, while the bottom end is used to interface to electronic subsystems. From the side protrudes a vacuum shell extension that houses the single PT415[^1] two-stage pulsetube cryocooler, which provides continuous cooling to \(<\)`<!-- -->`{=html}4 K. The cryogenic design can be simply described as concentric cylindrical aluminum shells, each an increasingly cold and thermally isolating radiation shield that together protect the sub-Kelvin detectors from warm radiation. Moving inward from the room-temperature vacuum shell are the nominal '50 K' and '4 K' stages, respectively cooled by the 1st and 2nd stages of the pulsetube. The pulsetube is thermally linked to the 50 K and 4 K stages in the main body of the cryostat via oxygen-free high-conductivity (OFHC) copper heat straps, including flexible foil and braid sections to minimize transmission of vibrational energy from the pulsetube. Within the 4 K shield is a series of stacked planar 'sub-Kelvin' stages (2 K, 350 mK, 250 mK) culminating in the 250 mK focal plane (Fig. [\[fig:subk\]](#fig:subk){reference-type="ref" reference="fig:subk"}). The various temperature stages are structurally supported and thermally isolated from each other by trusses of low thermal-conductivity materials (G10 fiberglass and carbon fiber). A three-stage (\(^4\)He/\(^3\)He/\(^3\)He) helium sorption fridge[^2] is mounted on the 4 K stage near the pulsetube and provides non-continuous cooling for all three sub-Kelvin stages. The three stages of the fridge contain 55 L of \(^4\)He and 33 L/4 L of \(^3\)He at standard temperature and pressure, able to cool 's detector modules and large copper focal plane from 4 K to 250 mK and then maintain operating temperatures for more than 48 hours before needing to re-cycle. The 250 mK stage is separated into two planar copper stages to provide thermal stability for the detectors. The lower level, closest to the 350 mK stage, is actively temperature-controlled to 269 mK by a resistive heater and acts as a buffer for the more isolated focal plane. Stainless steel blocks separate the two levels and provide low-pass thermal filtering. The focal plane itself is additionally temperature-controlled to 274 mK. ## Detectors and Readout 's focal plane is populated by planar arrays of polarization-sensitive slotted antennas, coupled to transition edge sensor (TES) bolometers and centered at 95 GHz, the same technology also fabricated by JPL/Caltech for /*Keck* *Array* and detailed in Ref. . Each silicon detector tile contains an \(8\times8\) array of pixels with each pixel made up of two co-located, orthogonally-polarized sub-antenna networks and two TES bolometers. A Gaussian taper of each pixel's beam is achieved by tuning the summing tree that couples each bolometer to its sub-antenna network. A major difference from /*Keck* *Array* is the installation method of detector tiles on the focal plane. In /*Keck* receivers, four detector tiles are directly wire bonded to the focal plane, and the supporting readout electronics sit adjacent at the focal plane's outer radius.  instead packages each detector tile and its electronics in a more easily installed compact detector module. The concurrently presented Ref.  describes these detector modules in more detail. Twenty such detector modules are tiled onto the focal plane, each containing 128 TES bolometers, for a total of 2560. Each detector tile contains 8 'dark' TES bolometers that are intentionally left uncoupled to antennas, 4 of which are read out as non-optical probes, giving a total of 2400 optically-coupled bolometers or, equivalently, 1200 polarization-sensitive detector pairs centered at 95 GHz. This contrasts to a single /*Keck* *Array* receiver at 150 GHz, which has four detector tiles for a total of 248 polarization-sensitive detector pairs. At 95 GHz, a single *Keck* *Array* receiver provides an even sharper comparison with 136 polarization-sensitive detector pairs, owing to its slower optics and smaller aperture than . continues to use time-domain multiplexed (TDM) readout of the TES bolometers, via SQUID MUX chips developed by NIST, in order to reduce the total number of lines that must enter the cryostat and their associated heat load. The signal current through each voltage-biased TES is inductively coupled to its own 1st-stage SQUID series array, forming one column of one multiplexed 'row'. A pair of 11-row MUX chips is wire bonded in series in order to allow multiplexing through 22 rows. Each detector module contains 6 pairs of MUX chips, forming 6 columns of 22 rows, or 132 total 1st-stage channels. Each column's multiplexed output is inductively coupled to a separate SQUID series array on the 4 K stage for additional amplification before exiting the cryostat. uses MUX11d chips, a newer generation than those used by /*Keck* *Array* that has a different row-switching mechanism. Details of the new design are found in the concurrently presented Ref. . Control of the MUX system and feedback-based readout of the TES data are via room temperature Multi-Channel Electronics (MCE) systems developed by the University of British Columbia, again similar to those used in /*Keck* *Array*. ## Optics and IR Filtering {#subsec:op} Following the other B[icep]{.smallcaps}/*Keck* experiments, targets degree-scale polarization features by using a compact, on-axis, two-refractor optical design. Its \(f/1.7\) system with a 27.4° field of view (FOV) is faster and wider than the \(f/2.2\), FOV\(\sim\)`<!-- -->`{=html}15° optics of /*Keck* *Array*. It has a significantly larger aperture diameter of 520 mm, vs. 264 mm for /*Keck* *Array*, defined by a millimeter-wave absorbing aperture stop. Both refracting lenses and the aperture stop are heatsunk to the 4 K stage, with the objective lens (warmest and furthest from the focal plane) operating at 4.9 K. The lenses are made of 99.6% pure alumina ceramic, which allows a thinner design and smaller thermal gradients than the HDPE plastic used for previous B[icep]{.smallcaps}/*Keck* optics, crucial advantages for 's faster system and increased infrared (IR) thermal loading. The much larger diameter of the receiver window (\(\sim\)`<!-- -->`{=html}670 mm clear), necessitated by the increase in aperture size, creates a large IR radiative heat load. IR power entering the cryostat is in excess of 100 W, more than \(2\times\) the cooling power of the 1st stage of the single PT415 pulsetube cooler. The first level of IR filtering therefore attempts to reject the majority of IR loading before it reaches the 50 K stage, and consists of a stack of 10 metal-mesh IR-reflective filters mounted just within the receiver window at room temperature. These filters are made up of \(\mathcal{O}\)(10 \\(\\)m) capacitive squares created by laser ablation of aluminized thin plastic films (3.5 \\(\\)m Mylar or 6 \\(\\)m polypropylene (PP)/polyethylene (PE) blend), as introduced in Ref. . In all, through both reflection and absorption/reemission, the metal-mesh filter stack allows \(\sim\)`<!-- -->`{=html}19 W to reach the 50 K stage. One additional metal-mesh filter is mounted just behind the 50 K alumina filter. The remaining power is absorbed by a series of three filters (and thereby removed by the pulsetube) in order to shield the sub-Kelvin detector stage: a 10 mm thick alumina filter on the 50 K stage, between the metal-mesh filter stack and the objective lens; a 5 mm thick nylon filter on the 4 K stage, between the two lenses; and a second 4 K nylon filter, 9.5 mm thick, between the field lens and the focal plane. These IR filters and lenses are labeled in Fig. [\[fig:b3_rx_cutaway\]](#fig:b3_rx_cutaway){reference-type="ref" reference="fig:b3_rx_cutaway"}. Finally, low-pass metal-mesh edge filters with a 4 cm\\(^-1^\\) (120 GHz) cutoff are mounted onto each detector module to prevent any remaining out-of-band sub-millimeter blue leaks from coupling to the detector bolometers. # Observing Site and Strategy {#sec:obs} The receiver is integrated with its telescope mount and various subsystems in the Dark Sector Laboratory (DSL) building at the South Pole, from where it observes the sky during the austral winter. Support and logistics are managed through the US National Science Foundation's Amundsen-Scott South Pole Station, located 1 km away, including satellite-based daily transfer of the full science data stream. This cold and remote observing site sits at an elevation of over 2800 m on the Antarctic Plateau and provides extremely dry and stable atmospheric conditions, optimal for 's millimeter-wave observations. The warmed indoor environment of DSL is extended beyond the roof level by a flexible insulating boot so that only the receiver window is exposed to the outdoor environment (Fig. [\[fig:b3_roof\]](#fig:b3_roof){reference-type="ref" reference="fig:b3_roof"}). A dry nitrogen gas system fills a protective membrane above the receiver window to prevent snow accumulation, and a cylindrical comoving microwave-absorptive forebaffle extends beyond the receiver window to intercept stray light. A reflective, stationary ground shield is fixed to the roof of DSL and acts as a second barrier against signal contamination from nearby ground sources. continues to use the original three-axis mount in DSL that also supported . By removing some non-essential structures, the mount was modified to fit 's larger size while retaining the same functions: ranges of \>360° in azimuth, 50°-- 90° (zenith) in elevation, and boresight rotation about the optical axis (Fig. [\[fig:b3_axes\]](#fig:b3_axes){reference-type="ref" reference="fig:b3_axes"}). Due to the larger diameter of and the protruding vacuum shell extension that houses the pulsetube cooler, boresight rotation is limited by mechanical interference to a range of 255°. All electrical cabling, high-pressure helium gas hoses for the pulsetube, and other necessary lines are routed through cable carriers for the three axes of motion, with the exception of a 4-channel high-pressure rotary joint[^3] used to bypass the two helium hoses at the azimuth stage. By pressurizing two outer channels to act as buffers to atmosphere, the 4-channel rotary joint achieves much better leak performance than the 2-channel unit used during 's 2015 season; the optimal pulsetube system pressure is maintained by small, non-disruptive refills to the compressor less than once a month. focuses its CMB observations on the same sky patch as the *Keck* *Array* and previous B[icep]{.smallcaps} experiments, spanning right ascension \(\ang{-60}<\textrm{RA}<\ang{60}\) and declination \(\ang{-70}<\delta<\ang{-40}\). Its larger optical FOV results in an effective sky area of \(\sim\) 600 deg\\(^2^\\), larger than the \(\sim\) 400 deg\\(^2^\\) area covered by /*Keck*. The fundamental observing block is a constant-elevation 'scanset' of 50 back-and-forth 2.8 deg/s fixed-center azimuth scans spanning 64.4°. Each scanset is bookended by a 26 s elevation nod of 1.3° peak to peak for airmass-based detector gain calibration and \(\sim\) 1 minute of partial TES loadcurves (non-superconducting) to monitor detector parameters, for a total duration of \(\sim\) 50 minutes. The azimuth center is adjusted every other scanset to track the changing RA of the sky patch, using that drift in combination with the field of view to fill in the full range in RA. Elevation is staggered by twenty 0.25° offsets amongst the scansets to fill in coverage of the detector beam pattern. Observations are on a two-sidereal day cadence, including 40 CMB scansets and 7 scansets aimed at the galactic plane. Each two-day cycle uses one of four boresight angles within the allowed boresight envelope to help constrain polarization systematics: two sets of 180° opposing orientations, offset from each other by 45°, and clocked to optimize coverage symmetry of both the Stokes \(Q\) and \(U\) polarizations on the sky. The first 6.5 hrs of each cycle is spent re-cycling the helium sorption fridge to keep the detectors cold for the rest of the two days. # Current Performance {#sec:perf} 's first season performance, during the 2015 austral winter, is summarized in Ref. . During that initial engineering season, 9 of the possible 20 detector modules were installed, for a total of 540 polarization-sensitive detector pairs centered at 95 GHz. Yield after accounting for known hardware issues (e.g. suspected broken wire bonds, SQUID MUX chip yield) was 436 good polarization-sensitive pairs, or 81%. 51 of the 104 lost pairs were attributed to a single anomalously low-yield detector module, potentially due to wire bond or other damage to the detector tile during installation and receiver cool-down. Median per-detector CMB noise-equivalent temperature (NET) for the 2015 season is estimated at 449 \\(\\)K\\(\\), based on polarization pair-differenced timestream data as shown in Fig. [\[fig:netspec\]](#fig:netspec){reference-type="ref" reference="fig:netspec"} and explained later in this section. underwent a number of improvements during the 2015/16 austral summer in preparation for its second season in the 2016 austral winter, which is the first dedicated science season. As originally planned, the focal plane was populated with the full complement of 20 detector modules, including 12 new detector modules and the removal of the anomalously low-yield module mentioned above. Additionally, several important changes were made in response to data and operations knowledge collected during the 2015 season. First, cryogenic improvements have added over 20% to the helium sorption fridge duty cycle and allow the current two-day observing schedule cadence. During the 2015 season, continuous observing time was limited by the fridge to 12 hrs. This cryogenic fix came in two parts, both vetted by parallel lab testing in North America during 's 2015 season observations. Lowering of the helium sorption fridge bath temperature by 0.6 K, to 3.5 K, was achieved by increasing thermal conduction through the copper heatstrap to the pulsetube's 2nd stage and thereby allowed increased helium condensation during the fridge re-cycling. Furthermore, a foil shielding included on a small number of electrical cables was measured to give an excess 200 \\(\\)W parasitic conductive heat load to the nominal 350 mK stage, at least tripling the total expected load. Removal of this incorrectly specified shielding resulted in a large 280 mK reduction on that stage, allowing it to reach 360 mK, which in turn allowed lowering the focal plane operating temperature from 313 mK to 274 mK for the 2016 season. Second, two faulty metal-mesh IR-reflective filters were replaced for 2016 and have significantly reduced broad and polarization-dependent scattering of the 95 GHz signal. This has been measured both as a per-detector 26% median reduction of loading from the comoving forebaffle, and as an improvement in detector loading median mismatch from \(\sim\) 10% to \(<3\)% between co-located orthogonally polarized detectors when observing broad sources (forebaffle and atmosphere). Lab inspection has shown these faulty filters to be plagued by incomplete and inconsistent laser ablation of the aluminization on micron scales, now easily remedied with replacements and improved screening. Finally, 2015 season data showed to be sensitive to 450 MHz interference from the main South Pole Station radio system, owing to its larger receiver aperture than /*Keck* *Array* (Sec. [2.3](#subsec:op){reference-type="ref" reference="subsec:op"}) and correspondingly lower high-pass waveguide cutoff frequency. The interference caused a mostly ground-fixed signal, but also contributed to transients in the feedback-based detector readout. For the 2016 season, internal RF shielding improvements at the focal plane have given a \(\sim\) 10 dB reduction in RF signal sensitivity, and changes at the main station (reduced output power and a directional antenna) have reduced the signal strength at DSL by \(\sim\) 35 dB, for a total \(\sim\) 45 dB reduction. As is routine for the B[icep]{.smallcaps}/*Keck* experiments, was subject to several calibration measurements during the 2015/16 austral summer in order to characterize the instrument before the 2016 observation season. Per-detector spectra were measured for all 20 modules with a custom Martin-Puplett Fourier Transform Spectrometer (FTS) mounted to the window of while in the mount. Optical efficiency through the receiver was measured with an aperture-filling liquid nitrogen load mounted above the receiver window, which also acted as an ambient temperature source when unfilled. Results from both measurements can be found in the concurrently presented Ref. . Far-field beam characteristics were measured over several weeks using a chopped thermal source mounted high above the nearby Martin A. Pomerantz Observatory building (the location of *Keck* *Array*) and are covered in detail in the concurrently presented Ref. . As of writing in early June 2016, has observed for \(\sim2.5\) months in its second season and accumulated roughly 1224 receiver-hours of new CMB data. Overall detector yield due to known hardware issues is 951 good polarization-sensitive pairs, or 79%. 180 of the 249 lost pairs are due to electrical open circuits that developed on 7 SQUID MUX row-select lines shared by a group of 5 detector modules. The cause of the bad connections is yet to be determined and only became present after cryogenic cooling of the receiver. Because these row-select lines are run in series, it is possible that poor wire bond adhesion within a single detector module could be the culprit. The number of lost detector pairs is magnified due to polarization pairing across adjacent MUX rows (pairs are wired along the same MUX column, rather than row, sharing common parameters such as SQUID and TES biases). The median noise spectrum and per-detector, per-scanset noise distribution is shown in Fig. [\[fig:netspec\]](#fig:netspec){reference-type="ref" reference="fig:netspec"}, based on the timestreams for all CMB data so far during the 2016 season. The timestream data has been differenced between polarization pairs, with basic cuts applied so that only good pairs remain (based on both the known hardware yield and preliminary data-quality cuts). Pair-summed spectra are also shown to demonstrate the \(1/f\) noise rejection of the pair-difference polarization data. Conversion to CMB temperature is done by calibration of a CMB temperature map to 100 GHz Planck data. The histogram shows science-band average noise after filtering the timestreams with a 3rd-order polynomial during every unidirectional azimuth scanning movement (100 moves per scanset) to mimic final map making. 's 2015 season data is also shown for comparison. An improvement in per-detector sensitivity is clear, summarized by the improved per-detector median band-average NET of 347 \\(\\)K\\(\\) for the 2016 season versus 449 \\(\\)K\\(\\) from the 2015 season data (right panel of Fig. [\[fig:netspec\]](#fig:netspec){reference-type="ref" reference="fig:netspec"}). However, *Keck* *Array* 95 GHz data shows a median 280 \\(\\)K\\(\\) with the same treatment. Reduction of scattering and the resulting detector photon load from warm surfaces, along with the reduction of RF interference, both detailed earlier in this section, can explain 's improvement in the 2016 season, but a detailed quantification and an understanding of the performance relative to *Keck* *Array* detectors is ongoing. Combining the band-average NET values across all detectors for each scanset gives a 2016 median array NET of 9.91 \\(\\)K\\(\\). This value accounts for hardware yield (79%) and for preliminary data-quality cuts (currently a \(\sim\)`<!-- -->`{=html}45% loss). 2015 data gives a much higher median array NET of 24.6 \\(\\)K\\(\\) due to the higher per-detector NET, the only partially populated focal plane, and a higher (\(\sim\)`<!-- -->`{=html}58%) loss to data-quality cuts. All above quoted NET values in this section are based on timestream data. This is an important distinction and the values are not meant to be directly comparable to map-based NET estimates. Specifically, Ref.  and Ref.  quote map-based per-detector NET estimates of 395 \\(\\)K\\(\\) (2015 season) and 333 \\(\\)K\\(\\) (2016 season). The difference arises from details of the two NET estimations. The map-based value is an estimate based on the standard deviation across all map pixels of a polarization noise map, accounting for the weighted integration time of each pixel, and is therefore insensitive to details of individual detector spectra and assumes a Gaussian noise distribution that will not necessarily reproduce a median of the distributions shown in Fig. [\[fig:netspec\]](#fig:netspec){reference-type="ref" reference="fig:netspec"}. Additionally, the timestream-based value averages across 0.1--1 Hz, not the full range of scales considered in the map-based method described. The two methods converge more closely for the total array NET, for which the map-based method gives 9.72 \\(\\)K\\(\\). # Future Multi-Receiver Array {#sec:array} Plans have begun to upgrade the *Keck* *Array* to a 3rd-generation, or "Stage 3", experiment with -class receivers. The *Keck* *Array*, the culmination of the /*Keck* 2nd-generation, demonstrates the scalability of -class receiver technology and the efficiency of frequency expansion through an array model in which the optics and detectors of each receiver can be tuned to a single frequency band. For the 2014 observing season, two of the five 150 GHz *Keck* *Array* receivers were replaced with 95 GHz detectors and matching optics. That additional data allowed a more extensive model of the polarized galactic foreground, and for the first time showed B-mode constraints to be the most powerful on the IGW signal (beyond constraints derived from CMB temperature and other evidence). In the following 2015 observing season, two more 150 GHz *Keck* *Array* receivers were refitted, this time with 220 GHz detectors and optics, further broadening the spectral coverage of the combined data set by reaching to the higher-frequency polarized dust signal. Most recently, now that is fully populated at 95 GHz for the current 2016 observing season, the two 95 GHz *Keck* *Array* receivers were also refitted to 220 GHz, for a total of four at 220 GHz. With its 5\(\times\) higher per-receiver detector count, the design acts as a building block for this 3rd-generation multi-frequency CMB polarimeter that we have chosen to name B[icep]{.smallcaps} Array. In order to continue to further constrain the IGW signal and aim for primordial B-mode detection, B[icep]{.smallcaps} Array will focus its increased total detector count (\(\sim\) 12\(\times\) greater than *Keck* *Array*) on the same sky patch from the South Pole. This shared sky patch with all other B[icep]{.smallcaps}/*Keck* experiments covers \(\sim\)`<!-- -->`{=html}1% of the sky, sufficient area to search for the degree-scale (multipole \(l=80\)) recombination bump B-mode signal. With the large amount of accumulated observations, it remains the preferred patch to continue with deeper, multi-frequency observations for targeted degree-scale IGW signal detection and improved galactic foreground separation. B[icep]{.smallcaps} Array will be 4 new -class receivers with frequency band centers spanning 35 to 270 GHz. The design of a -class receiver was detailed in Sec. [2](#sec:inst){reference-type="ref" reference="sec:inst"}, showing how expanded to higher detector count and larger aperture while maintaining the advantages of the compact, on-axis, cold refractor design shared by the B[icep]{.smallcaps}/*Keck* program. A number of changes are expected to improve on the design. Testing of the metal-mesh IR-reflective filter design has been ongoing since its first development for , and potential changes include double-sided aluminization and laser patterning so that fewer filters are required to achieve the same IR rejection. Heatsinking the metal-mesh filter stack to 50 K, rather than the room temperature vacuum shell as in , could also reduce IR loading from the emissive substrate films. Some changes are also expected for the detector modules used in B[icep]{.smallcaps} Array. The four receivers will separately have focal planes centered on 35, 95, and 150 GHz, and lastly a 220/270 GHz hybrid. Each focal plane will be made up of -style detector module packages but will use 6" rather than 4" silicon wafer processes and require many fewer than 20 modules to fill the focal plane, thereby reclaiming focal plane area lost to the module borders. Table [1](#tab:numbers){reference-type="ref" reference="tab:numbers"} shows the nominal number of detectors that will fit into each focal plane of B[icep]{.smallcaps} Array. The 220/270 GHz dust-control receiver will have detectors at the two frequencies in a checkerboard array on the focal plane, based on the current 220 GHz *Keck* *Array* detectors and a lab-tested 270 GHz design that will be introduced into *Keck* *Array* for the 2017 season. The extremely high 220/270 GHz detector count motivates frequency domain multiplexing readout schemes as a possible alternative to 's SQUID-based TDM readout (which will continue to be used for the lower frequency receivers). The 35 GHz receiver will help constrain galactic synchrotron emission at the level required for B[icep]{.smallcaps} Array's sensitivity; prototyping and testing of the 35 GHz detectors is currently underway. Fig. [\[fig:bkhistory\]](#fig:bkhistory){reference-type="ref" reference="fig:bkhistory"} shows a comparison of the proposed B[icep]{.smallcaps} Array design to its predecessors, including the focal plane size and frequency mix. B[icep]{.smallcaps} Array will use a new mount at the South Pole to hold all 4 -class receivers, each arranged around a common boresight, as the increased receiver size cannot be accommodated by the existing *Keck* *Array* mount (originally built for the D[asi]{.smallcaps} experiment). The use of high-pressure helium rotary joints, operating successfully with , will help reduce the total size of the new mount by avoiding a significant volume of helium hose in the azimuth and theta stages. B[icep]{.smallcaps} Array will replace the *Keck* *Array* and observe simultaneously with the ongoing program. The 35 GHz receiver is under development and is planned to be the first of the B[icep]{.smallcaps} Array, beginning as early as the 2018 observing season. Three *Keck* *Array* receivers outfitted at 270 GHz will fill the remaining spots of the new mount. Before the second observing season, the 95 and 150 GHz receivers will be installed, and finally the 220/270 GHz receiver will complete the array for the third and subsequent observing seasons. A useful metric to quantify the CMB polarization sensitivity and the statistical sensitivity to \(r\) that will be achieved with B[icep]{.smallcaps} Array is the survey weight \(W=2A/N^2\) (units of \(\mu\)K\(^{-2}\)), where \(A\) is the map area and \(N\) is the polarization map depth (e.g. units of \(\mu\)K\(\cdot\)degrees). Conveniently, survey weight scales linearly with detector count and integration time, and Table [1](#tab:numbers){reference-type="ref" reference="tab:numbers"} shows a comparison of survey weights per year for *Keck* *Array* and B[icep]{.smallcaps} Array receivers. The *Keck* *Array* 95 GHz survey weight is based on the 2014 season published result, while the 150 GHz survey weight is based on both performance and the latest published 3 receiver-years of *Keck* *Array*. The *Keck* *Array* 220 GHz survey weight is based on preliminary results from the first season (2015) of receivers at that frequency. These achieved values are then simply scaled by the squared ratio of nominal per-detector NET values (taking into account e.g. detector parameters, loading from optical elements and the atmosphere) to 35 and 270 GHz, and the B[icep]{.smallcaps} Array values are further scaled by the ratio of detector counts. In this way, all values take into account all real world observing inefficiencies of the /*Keck* experiments. From this treatment it is clear that the B[icep]{.smallcaps} Array represents an order of magnitude upgrade in total raw experimental sensitivity versus *Keck* *Array*. # Conclusion These proceedings have summarized the instrument design and observation strategy of the polarimeter, and reported on changes and current performance for its ongoing second season. Important changes included full detector population of the focal plane, cryogenic improvements that led to increased observation duty cycle, replacement of under-performing IR filters, and a significant reduction in RF interference. Based on the first \(\sim1224\) hours of new CMB data this season (2016), has a timestream-based median per-detector NET of 347 \\(\\)K\\(\\) and an array NET of 9.91 \\(\\)K\\(\\) for the whole instrument, an improvement over the 2015 per-detector median of 449 \\(\\)K\\(\\). Plans were also outlined for an upgrade to the multi-frequency *Keck* *Array*, named B[icep]{.smallcaps} Array. B[icep]{.smallcaps} Array will be four -class receivers spanning 35 to 270 GHz and begin a staggered deployment as soon as 2018. It will use a new mount at the existing South Pole *Keck* *Array* site, and have more than an order of magnitude greater survey weight per receiver versus *Keck* *Array*. # MISCELLANEOUS FORMATTING DETAILS This work is supported by the National Science Foundation (grant nos. 1313158, 1313010, 1313062, 1313287, 1056465, 0960243), the SLAC Laboratory Directed Research and Development Fund, the Canada Foundation for Innovation, Science and Technology Facilities Council Consolidated Grant (ST/K000926/1), and the British Columbia Development Fund. The development of detector technology was supported by the JPL Research and Technology Development Fund and grants 06-ARPA206-0040, 10-SAT10-0017, and 12-SAT12-0031 from the NASA APRA and SAT programs. The development and testing of detector modules was supported by the Gordon and Betty Moore Foundation. [^1]: Cryomech Inc., Syracuse, NY 13211, USA (`www.cryomech.com`) [^2]: Chase Research Cryogenics Ltd., Sheffield, S10 5DL, UK (`www.chasecryogenics.com`) [^3]: Dynamic Sealing Technologies, Inc., Andover, MN 55304, USA (`www.dsti.com`)

{'timestamp': '2016-07-19T02:00:45', 'yymm': '1607', 'arxiv_id': '1607.04668', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04668'}

# Introduction {#intro} Deep neural networks (DNNs) have been successful in modeling many aspects of natural language including word meanings, machine translation, syntactic parsing, language modeling and image captioning. Given sufficient training data, DNNs are highly accurate and can be trained end-to-end without the need for intermediate knowledge representations or explicit feature extraction. With recent interest in conversational user interfaces such as virtual assistants and chatbots, the application of DNNs to facilitate meaningful conversations is an area where more progress is needed. While sequence to sequence models based on recurrent neural networks (RNNs) have shown initial promise in creating intelligible conversations, it has been noted that more work is needed for these models to fully capture larger aspects of human communication including conversational goals, personas, consistency, context, and word knowledge. Since discourse analysis considers language at the conversation-level, including its social and psychological context, it is a useful framework for guiding the extension of end-to-end neural conversational models. Drawing on concepts from discourse analysis such as *coherence* and *cohesion*, we can codify what makes conversations more intelligent in order to design more powerful neural models that reach beyond the sentence and utterance level. For example, by looking for features that indicate deixis, anaphora, and logical consequence in the machine-generated utterances we can benchmark the level of coherence and cohesion with the rest of the conversation, and then make improvements to models accordingly. In the long run, if neural models can encode the long-range structure of conversations, they may be able to express conversational discourse similar to the way the human brain does, without the need for explicitly building formal representations of discourse theory into the model. To that end, we explore RNN-based sequence to sequence architectures that can capture long-range relationships between multiple utterances in conversations and look at their ability to exhibit discourse relationships. Specifically, we look at 1) a baseline RNN encoder-decoder with attention mechanism and 2) a model with an additional discourse RNN that encodes a sequence of multiple utterances. Our contributions are as follows: - We examine two RNN models with attention mechanisms to model discourse relationships across different utterances that differ somewhat compared to what has been done before - We carefully construct controlled experiments to study the relative merits of different models on multi-turn conversations - We perform a sensitivity analysis on how the amount of context provided by previous utterances affects model performance - We quantify how neural conversational models display coherence by measuring the prevalence of specific syntactical features indicative of deixis, anaphora, and logical consequence. # Related Work {#relatedwork} Building on work done in machine translation, sequence to sequence models based on RNN encoder-decoders were initially applied to generate conversational outputs given a single previous message utterance as input. In several models were presented that included a "context" vector (for example representing another previous utterance) that was combined with the message utterance via various encoding strategies to initialize or bias a single decoder RNN. Some models have also included an additional RNN tier to capture the context of conversations. For example, includes a hierarchical "context RNN" layer to summarize the state of a dialog, while includes an RNN "intension network" to model conversation intension for dialogs involving two participants speaking in turn. Modeling the "persona" of the participants in a conversation by embedding each speaker into a \(K\)-dimensional embedding was shown to increase the consistency of conversations in. Formal representations such as Rhetorical Structure Theory (RST) have been developed to identify discourse structures in written text. Discourse parsing of cue phrases and coherence modeling based on co-reference resolution of named-entities have been applied to tasks such as summarization and text generation. Lexical chains and narrative event chains provide directed graph models of text coherence by looking at thesaurus relationships and subject-verb-temporal relationships, respectively. Recurrent convolutional neural networks have been used to classify utterances into discourse speech-act labels and hierarchical LSTM models have been evaluated for generating coherent paragraphs in text documents . Our aim is to develop end-to-end neural conversational models that exhibit awareness of discourse without needing a formal representation of discourse relationships. ## Models Since conversations are sequences of utterances and utterances are sequences of words, it is natural to use models based on an RNN encoder-decoder to predict the next utterance in the conversation given \(N\) previous utterances as source input. We compare two types of models: **seq2seq+A**, which applies an attention mechanism directly to the encoder hidden states, and **Nseq2seq+A**, which adds an additional RNN tier with its own attention mechanism to model discourse relationships between \(N\) input utterances. In both cases the RNN decoder predicts the output utterance and the RNN encoder reads the sequence of words in each input utterance. The encoder and decoder each have their own vocabulary embeddings. As in we compute the attention vector at each decoder output time step \(t\) given an input sequence \((1...,T_{A})\) using: \[\begin{aligned} u_{i}^{t} & = & v^{T}tanh(W_{1} h_{i} + W_{2} d^{t}) \nonumber \\ a_{i}^{t} & = & softmax(u_{i}^{t}) \nonumber \\ c^{t} & = & \sum_{i=1}^{T_{A}} a_{i}^{t} h_{i} \nonumber \end{aligned}\] where the vector \(v\) and matrices \(W_{1}\), and \(W_{2}\) are learned parameters. \(d^{t}\) is the decoder state at time \(t\) and is concatenated with \(c^{t}\) to make predictions and inform the next time step. In **seq2seq+A** the \(h_{i}\) are the hidden states of the encoder \(e_{i}\), and for **Nseq2seq+A** they are the \(N\) hidden states of the discourse RNN (see Fig. [\[schematic\]](#schematic){reference-type="ref" reference="schematic"}.) Therefore, in **seq2seq+A** the attention mechanism is applied at the word-level, while in **Nseq2seq+A** attention is applied at the utterance-level. ## seq2seq+A {#seq2seq} As a baseline starting point we use an attention mechanism to help model the discourse by a straightforward adaptation of the RNN encoder-decoder conversational model discussed in. We join multiple source utterances using the *EOS* symbol as a delimiter, and feed them into the encoder RNN as a single input sequence. As in , we reversed the order of the tokens in each of the individual utterances but preserved the order of the conversation turns. The attention mechanism is able to make connections to any of the words used in earlier utterances as the decoder generates each word in the output response. ## Nseq2seq+A {#Nseq2seq} Since conversational threads are ordered sequences of utterances, it makes sense to extend an RNN encoder-decoder by adding another RNN tier to model the discourse as the turns of the conversation progress. Given \(N\) input utterances, the RNN encoder is applied to each utterance one at a time as shown in Fig. [\[schematic\]](#schematic){reference-type="ref" reference="schematic"} (with tokens fed in reverse order.) The output of the encoder from each of the input utterances forms \(N\) time step inputs for the discourse RNN. The attention mechanism is then applied to the \(N\) hidden states of the discourse RNN and fed into the decoder RNN. We also considered a model where the output of the encoder is also combined with the output of the discourse RNN and fed into the attention decoder, but found the purely hierarchical architecture performed better. ## Learning For each model we chose identical optimizers, hyperparameters, etc. in our experiments in order to isolate the impact of specific differences in the network architecture, also taking computation times and available GPU resources into account. It would be straightforward to perform a grid search to tune hyperparameters, try LSTM cells, increase layers per RNN, etc. to further improve performance individually for each model beyond what we report here. For each RNN we use one layer of Gated Recurrent Units (GRUs) with 512 hidden cells. Separate embeddings for the encoder and decoder, each with dimension 512 and vocabulary size of 40,000, are trained on-the-fly without using predefined word vectors. We use a stochastic gradient descent (SGD) optimizer with \(L2\) norms clipped at \(5.0\), an initial learning rate of \(0.5\), and a learning rate decay factor of \(0.99\) is applied when needed. We trained with mini-batches of 64 randomly selected examples, and ran training for approximately 10 epochs until validation set loss converged. # Experiments We first present results comparing our neural discourse models trained on a large set of conversation threads based on the OpenSubtitles dataset. We then examine how our models are able to produce outputs that indicate enhanced coherence by searching for discourse markers. ## OpenSubtitles dataset {#opensubtitles} A large-scale dataset is important if we want to model all the variations and nuances of human language. From the OpenSubtitles corpus we created a training set and validation set with 3,642,856 and 911,128 conversation fragments, respectively[^1]. Each conversation fragment consists of 10 utterances from the previous lines of the movie dialog leading up to a target utterance. The main limitation of the OpenSubtitles dataset is that it is derived from closed caption style subtitles, which can be noisy, do not include labels for which actors are speaking in turn, and do not show conversation boundaries from different scenes. We considered cleaner datasets such as the Ubuntu dialog corpus, Movie-DiC dialog corpus, and SubTle corpus but found they all contained orders of magnitude fewer conversations and/or many fewer turns per conversation on average. Therefore, we found the size of the OpenSubtitles dataset outweighed the benefits of cleaner smaller datasets. This echoes a trend in neural networks where large noisy datasets tend to perform better than small clean datasets. The lack of a large-scale clean dataset of conversations is an open problem in the field. ## Results We compared models and performed a sensitivity analysis by varying the number of previous conversation turns fed into the encoder during training and evaluation. ::: # Conclusions We studied neural discourse models that can capture long distance relationships between features found in different utterances of a conversation. We found that a model with an additional discourse RNN outperforms the baseline RNN encoder-decoder with an attention mechanism. Our results indicate that providing more context from previous utterances improves model performance up to a point. Qualitative examples illustrate how the discourse RNN produces increased coherence and cohesion with the rest of the conversation, while quantitative results based on text mining of discourse markers show that the amount of deixis, anaphora, and logical consequence found in the decoder output can be sensitive to the size of the context window. In future work, it will be interesting to train discourse models on even larger corpora and compare conversations in different domains. By examining the attention weights it should be possible to study what discourse markers the models are "paying attention to" and possibly provide a powerful new tool for analyzing discourse relationships. By applying multi-task sequence to sequence learning techniques as in we may be able to combine the conversational modeling task with other tasks such as discourse parsing and/or world knowledge modeling achieve better overall model performance. Not just for conversations, neural discourse modeling could also be applied to written text documents in domains with strong patterns of discourse such as news, legal, healthcare. [^1]: The training and validation sets consisted of \(320M\) and \(80M\) tokens, respectively [^2]: We use perplexity as our performance metric, because it is simple to compute and correlates with human judgements, though it has well-known limitations. [^3]: here, there, then, now, later, this, that [^4]: she, her, hers, he, him, his, they, them, their, theirs [^5]: so, after all, in addition, furthermore, therefore, thus, also, but, however, otherwise, although, if, then

{'timestamp': '2016-07-18T02:08:34', 'yymm': '1607', 'arxiv_id': '1607.04576', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04576'}

# Introduction It is well known that a degree \(n\) Blaschke product is an \(n\)-to-\(1\) (counting multiplicity) analytic self-map of the unit disk \(\mathbb{D}\), and moreover that every such analytic \(n\)-to-\(1\) self-map of the unit disk is a degree \(n\) Blaschke product. We can restate this in a slightly more general setting. Throughout this paper we let \(G\subset\mathbb{C}\) denote a simply connected domain, and we let \(\mathcal{S}\subset G\) be a smooth Jordan curve, with bounded face \(D\). The purpose of this note is to generalize the result of Theorem [\[thm: Blaschke prod. version.\]](#thm: Blaschke prod. version.){reference-type="ref" reference="thm: Blaschke prod. version."} to the context of meromorphic functions and pseudo-lemniscates. In order to state the theorem, we begin with several definitions. For the remainder of the paper we let \(\Gamma\) denote some smooth Jordan curve in \(\mathbb{C}\), with bounded face \(\Omega_-\) and unbounded face \(\Omega_+\) (where \(\Omega_+\) includes the point \(\infty\)). In the special case that \(\Gamma=\mathbb{T}\), the above definition reduces to the classical definition of a lemniscate. The definition of a pseudo-lemniscate was introduced in  wherein conformal equivalence of functions on domains bounded by pseudo-lemniscates was studied. We now proceed to the promised generalization of Theorem [\[thm: Blaschke prod. version.\]](#thm: Blaschke prod. version.){reference-type="ref" reference="thm: Blaschke prod. version."}. Using the fact that any meromorphic function is conformal away from its critical points, we may use Theorem [\[thm: Main.\]](#thm: Main.){reference-type="ref" reference="thm: Main."} to supply a test for when the image of \(\mathcal{S}\) under \(f\) will not be a Jordan curve. In Section [2](#sect: Lemmas and proof.){reference-type="ref" reference="sect: Lemmas and proof."} we will provide proofs of Theorem [\[thm: Main.\]](#thm: Main.){reference-type="ref" reference="thm: Main."} and Corollary [\[cor: Non-Jordan curve or crit. point.\]](#cor: Non-Jordan curve or crit. point.){reference-type="ref" reference="cor: Non-Jordan curve or crit. point."}. In Section [3](#sect: Example.){reference-type="ref" reference="sect: Example."} we apply Corollary [\[cor: Non-Jordan curve or crit. point.\]](#cor: Non-Jordan curve or crit. point.){reference-type="ref" reference="cor: Non-Jordan curve or crit. point."} to show that the image of a certain smooth Jordan curve \(\mathcal{S}\) under a certain meromorphic function \(f\) is not a Jordan curve, without explicitly finding a point of intersection in \(f(\mathcal{S})\). # Lemmas and Proofs {#sect: Lemmas and proof.} We begin with a topological observation on the convergence of preimages under continuous maps. First we introduce a bit of standard notation. Our second lemma shows that, under the conditions of Item 2 from Theorem [\[thm: Main.\]](#thm: Main.){reference-type="ref" reference="thm: Main."}, if \(E\) is a component of \(f^{-1}(\Omega_-)\cap D\), and \(z\in\partial E\), then \(f(z)\in\partial\Omega_-\). Since \(h\) is continuous and \(\Omega\) and \(D\) are open, there is a neighborhood of \(z_0\) contained in \(h^{-1}(\Omega)\cap D\), contradicting our choice of \(z_0\). Set \(m=\mathcal{N}_h(w_0)\), and let \(z_1,z_2,\ldots,z_k\) be the distinct preimages of \(w_0\) in \(D\), having multiplicities \(m_1,m_2,\ldots,m_k\) (so that \(m=m_1+m_2+\cdots+m_k\)). Choose some \(\delta>0\) such that \(\delta<\min(|z_i-z_0|:1\leq i\leq k)\). Choose disjoint neighborhoods \(D_1,D_2,\ldots,D_k\subset D\) such that for each \(1\leq i\leq k\), the following holds. - \(z_i\in D_i\). - \(D_i\cap B(z_0;\delta)=\emptyset\). - \(h\) is exactly \(m_i\)-to-\(1\) on \(D_i\setminus\{z_i\}\). Choose some point \(x\in B(z_0;\delta)\cap E\) sufficiently close to \(z_0\) so that \[h(x)\in\displaystyle\bigcap_{i=1}^kh(D_i).\] Then \(h^{-1}(h(x))\cap D\) contains \(x\), along with \(m_i\) distinct points in \(D_i\), for each \(1\leq i\leq k\). Thus since \(x\notin D_i\) for any \(1\leq i\leq k\), we have that \(\mathcal{N}_h(h(x))\) is at least \[m_1+m_2+\ldots+m_k+1>m.\] Since \(x\in E\), \(h(x)\in\Omega\), so we have a contradiction of the choice of \(m\). We therefore conclude that \(w_0\notin\Omega\). Since \(h\) is continuous we must have that \(w_0\in\overline{\Omega}\), so we conclude that \(w_0\in\partial\Omega\). ◻ # Application of Corollary [\[cor: Non-Jordan curve or crit. point.\]](#cor: Non-Jordan curve or crit. point.){reference-type="ref" reference="cor: Non-Jordan curve or crit. point."} {#sect: Example.} **Example.** Let \(\mathcal{S}\) denote some smooth curve in \(\mathbb{C}\) which contains the quadrilateral with vertices \(0\), \(1\), \(2i\), and \(1+4i\), and which approximates the boundary of this shape closely. Let \(D\) denote the bounded face of \(\mathcal{S}\). Then setting \(f(z)=e^z\), we claim that \(f(\mathcal{S})\) is not a Jordan curve. To see this, observe that for \(w_1=e^0i\), \(w_2=e^1i\), and \(w_3=e^2i\), we have \(\mathcal{N}_f(w_1)=1\), \(\mathcal{N}_f(w_2)=2\), and \(\mathcal{N}_f(w_3)=0\). Therefore Corollary [\[cor: Non-Jordan curve or crit. point.\]](#cor: Non-Jordan curve or crit. point.){reference-type="ref" reference="cor: Non-Jordan curve or crit. point."} tells us that either \(f\) has a critical point on \(\mathcal{S}\) or \(f(\mathcal{S})\) is not a Jordan curve. Since \(f(z)=e^z\) has no critical points, we conclude that \(f(\mathcal{S})\) is not a Jordan curve.

{'timestamp': '2016-07-18T02:09:10', 'yymm': '1607', 'arxiv_id': '1607.04607', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04607'}

# Introduction {#introduction .unnumbered} M. F. Atiyah introduced the notion of the \(\Gamma\)-index \(\mathrm{index}_{\Gamma}(\widetilde{D})\) for a lifted elliptic differential operator \(\widetilde{D}\) on a Galois \(\Gamma\)-covering over a closed manifold and proved that the \(\Gamma\)-index \(\mathrm{index}_{\Gamma}(\widetilde{D})\) of a lifted operator \(\widetilde{D}\) equals the Fredholm index \(\mathrm{index} (D)\) of the elliptic differential operator \(D\) on the base manifold. On the other hand, Atiyah also investigate properties of a \(\Gamma\)-trace \(\mathrm{tr}_{\Gamma}\) at the same time. The \(\Gamma\)-trace is a trace of the \(\Gamma\)-trace operators, so it induces a homomorphism \((\mathrm{tr}_{\Gamma})_{\ast}\) from \(K_{0}\)-group of the \(\Gamma\)-compact operators to the real numbers. Out of a lifted elliptic differential operator \(\widetilde{D}\), we can define the \(\Gamma\)-index class \(\mathrm{Ind}_{\Gamma}(\widetilde{D})\) by using the Connes-Skandalis idempotent and send it by the induced homomorphism \((\mathrm{tr}_{\Gamma})_{\ast}\), then the image \((\mathrm{tr}_{\Gamma})_{\ast}(\mathrm{Ind}_{\Gamma}(\widetilde{D}))\) equals the \(\Gamma\)-index \(\mathrm{index}_{\Gamma}(\widetilde{D})\) and thus the Fredholm index \(\mathrm{index} (D)\) of the elliptic differential operator \(D\) on the base manifold. On the other hand, there is another pseudo-differential calculus on Heisenberg manifolds which is called the Heisenberg calculus; see, for instance. Roughly speaking, Heisenberg calculus is "weighted" calculus and the product of the "Heisenberg principal symbols" are defined by convolution product. When the Heisenberg principal symbol of \(P\) is invertible, we call \(P\) a Heisenberg elliptic operator. Note that any Heisenberg elliptic operator is not elliptic. For a Heisenberg elliptic operator \(P\), we can construct a parametrix by using its inverse, so \(P\) is a Fredholm operator if the base manifold is closed. Thus the Fredholm index of \(P\) on a closed manifold is well defined, but a solution of an index problem of \(P\) does not obtained in general. However, index problems for Heisenberg elliptic operators on contact manifolds or foliated manifolds are solved by E. van Erp and P. F. Baum; see, , ,. In this note, we study that we can define the \(\Gamma\)-index \(\mathrm{index}_{\Gamma}(\widetilde{P})\) and the \(\Gamma\)-index class \(\mathrm{Ind}_{\Gamma}(\widetilde{P})\) for a lifted Heisenberg elliptic differential operator \(\widetilde{P}\) by using a parametrix. Once these ingredients are defined, the proof of the matching of three ingredients \(\mathrm{index}_{\Gamma}(\widetilde{P})\), \((\mathrm{tr}_{\Gamma})_{\ast}(\mathrm{Ind}_{\Gamma}(\widetilde{P}))\) and \(\mathrm{index} (P)\), is straightforward; see subsection [2.1](#sub:main){reference-type="ref" reference="sub:main"}. We also investigate an example of Heisenberg differential operators on a contact manifold with non-trivial \(\Gamma\)-index by using the index formula in; see subsection [2.2](#sub:exm){reference-type="ref" reference="sub:exm"}. # Short review of Atiyah's \(\Gamma\)-index theorem {#sec:atiyah} In this section, we recall Atiyah's \(\Gamma\)-index theorem in ordinary pseudo-differential calculus. The main reference of this section is Atiyah's paper. Let \(\widetilde{M} \to M\) be a Galois covering with a deck transformation group \(\Gamma\) over a closed manifold \(M\) with a smooth measure \(\mu\) and \(D: C^{\infty}(E) \to C^{\infty}(F)\) an elliptic differential operator on Hermitian vector bundles \(E,F \to M\). We lift these ingredients on \(\widetilde{M}\) and denote by \(\widetilde{D}: C^{\infty}(\widetilde{E}) \to C^{\infty}(\widetilde{F})\) and \(\tilde{\mu}\). Let \(\mathrm{Ker}_{L^{2}}(\widetilde{D})\) (resp. \(\mathrm{Ker}_{L^{2}}(\widetilde{D}^{\ast})\)) be the \(L^{2}\)-solutions of \(\widetilde{D}u = 0\) (resp. \(\widetilde{D}^{\ast}u = 0\)) and denote by \(\Pi_{0}\) (resp. \(\Pi_{1}\)) the orthogonal projection on a closed subspace \(\mathrm{Ker}_{L^{2}}(\widetilde{D})\) (resp. \(\mathrm{Ker}_{L^{2}}(\widetilde{D}^{\ast})\)) of the \(L^{2}\)-sections. A \(\Gamma\)-invariant bounded operator \(T\) on the \(L^{2}\)-sections \(L^{2}(\widetilde{E})\) of \(\widetilde{E}\) is of *\(\Gamma\)-trace class* if \(\phi T \psi \in L^{2}(\widetilde{E})\) is of trace class for any compactly suppoted smooth functions \(\phi, \psi\) on \(\widetilde{M}\). Denote by \(\mathcal{L}^{1}_{\Gamma}\) the set of \(\Gamma\)-trace class operators and \(\mathrm{tr}_{\Gamma}(T)\) the *\(\Gamma\)-trace* of a \(\Gamma\)-trace class operator \(T\) defined by \[\mathrm{tr}_{\Gamma}(T) = \textrm{Tr}(\phi T \psi) \in \mathbb{C}.\] Here, the right hand side is the trace of a trace class operator \(\phi T \psi\) and this quantity does not depend on the choice of functions \(\phi, \psi\). By using ellipticity of \(\widetilde{D}\), operators \(\phi \Pi_{0} \psi\) and \(\phi \Pi_{1} \psi\) are smoothing operators on compact sets. Thus \(\Pi_{0}\) and \(\Pi_{1}\) are of \(\Gamma\)-trace class and thus one obtains the \(\Gamma\)-index of \(\widetilde{D}\): \[\mathrm{index}_{\Gamma}(\widetilde{D}) = \mathrm{tr}_{\Gamma}(\Pi_{0})-\mathrm{tr}_{\Gamma}(\Pi_{0}) \in \mathbb{R}.\] In the context of the \(\Gamma\)-index theorem, the most important class of \(\Gamma\)-trace class operators is the lifts of almost local smoothing operators on \(M\). Let \(S\) be an almost local smoothing operator with a smooth kernel \(k_{S}\) and \(\widetilde{S}\) a lift of \(S\). Then \(\widetilde{S}\) is of \(\Gamma\)-trace class and its \(\Gamma\)-trace is calculated by the following: \[\label{eq:gammatr} \mathrm{tr}_{\Gamma}(\widetilde{S}) = \int_{M}\mathrm{tr}(k_{S}(x,x)) d\mu = \mathrm{Tr}(S). \tag{\(\ast\)}\] Denote by \(\mathcal{K}_{\Gamma}\) the \(C^{\ast}\)-closure of \(\mathcal{L}^{1}_{\Gamma}\) and \(K_{0}(\mathcal{K}_{\Gamma})\) the analytic \(K_{0}\)-group. Then \(\mathrm{tr}_{\Gamma}\) induces a homomorphism of abelian groups by substitution: \[(\mathrm{tr}_{\Gamma})_{\ast} : K_{0}(\mathcal{K}_{\Gamma}) \to \mathbb{R}.\] On the other hand, since \(D\) is an elliptic differential operator, there exist an almost local parametrix \(Q\) and almost local smoothing operators \(S_{0},S_{1}\) such that one has \(QD = 1-S_{0}\) and \(DQ = 1-S_{1}\). Denote by \(\widetilde{Q}\), \(\widetilde{S_{0}}\) and \(\widetilde{S_{1}}\) lifts of these operators and then one has same relations \(\widetilde{Q}\widetilde{D} = 1-\widetilde{S_{0}}\) and \(\widetilde{D}\widetilde{Q} = 1-\widetilde{S_{1}}\). Set \[e_{\widetilde{D}} = \begin{bmatrix} \widetilde{S_{0}}^{2} & \widetilde{S_{0}}(1 + \widetilde{S_{0}}) \widetilde{Q} \\ \widetilde{S_{1}}\widetilde{D} & 1-\widetilde{S_{1}}^{2} \end{bmatrix} \quad \text{and} \quad e_{1} = \begin{bmatrix} 0 & 0 \\ 0 & 1 \end{bmatrix}.\] By \(\widetilde{Q}\widetilde{S_{1}} = \widetilde{S_{0}}\widetilde{Q}\) and \(\widetilde{S_{1}}\widetilde{D} = \widetilde{D}\widetilde{S_{0}}\), one has \(e_{\widetilde{D}}^{2} = e_{\widetilde{D}}\), that is, \(e_{\widetilde{D}}\) is an idempotent. Note that this idempotent \(e_{\widetilde{D}}\) is called the Connes-Skandalis idempotent; see, for instance. Moreover, a difference \(e_{\widetilde{D}}-e_{1}\) is of \(\Gamma\)-trace class. Hence we can define a \(\Gamma\)-index class \[\mathrm{Ind}_{\Gamma}(\widetilde{D}) = [e_{\widetilde{D}}]-[e_{1}] \in K_{0}(\mathcal{K}_{\Gamma}).\] By the definition of a map \((\mathrm{tr}_{\Gamma})_{\ast}\) and Atiyah's paper, one has the following: As described in subsection [2.1](#sub:main){reference-type="ref" reference="sub:main"}, Atiyah's proof of matching of these ingredients in the above equality does not essentially use ellipticity. Note that ellipticity of \(D\) and \(\widetilde{D}\) is only used in the definition of these ingredients. # Atiya's \(\Gamma\)-index theorem in Heisenberg calculus Let \((M,H)\) be a closed Heisenberg manifold, that is, \(M\) is a closed manifold and \(H \subset TM\) is a hyperplane bundle. Let \(P: C^{\infty}(E) \to C^{\infty}(F)\) be a Heisenberg elliptic differential operator on Hermitian vector bundles \(E,F \to (M,H)\), that is, the Heisenberg principal symbol \(\sigma_{H}(P)\) of \(P\) is an invertible element. In this section, we prove the \(\Gamma\)-index theorem for \(P\), which is analogous to Atiyah's \(\Gamma\)-index theorem. Note that \(P\) is not an elliptic operator in the sense of ordinary pseudo-differential calculus. ## Statement and proof {#sub:main} By, there exist parametrix \(Q\) and smoothing operators \(S_{0},S_{1}\) such that one has \(QP = 1-S_{0}\) and \(PQ = 1-S_{1}\). Thus \(P\) is a Fredholm operator and one has the Fredholm index \(\mathrm{index} (P) \in \mathbb{Z}\) of \(P\) by compactness of \(M\). Moreover, since a integral kernel of \(Q\) is smooth off the diagonal, we can choose \(Q\) as an almost local operator and then \(S_{0}\) and \(S_{1}\) are also almost local operators. Let \(\widetilde{M} \to M\) be a Galois covering with a deck transformation group \(\Gamma\) over a closed manifold \(M\) with a smooth measure \(\mu\). We lift all structures on \(M\) to \(\widetilde{M}\). Then \((\widetilde{M}, \widetilde{H})\) is a Heisenberg manifold, \(\widetilde{P}: C^{\infty}(\widetilde{E}) \to C^{\infty}(\widetilde{F})\) is a Heisenberg elliptic differential operator and one has \(\widetilde{Q}\widetilde{P} = 1-\widetilde{S}_{0}\) and \(\widetilde{P}\widetilde{Q} = 1-\widetilde{S}_{1}\). Since \(P\) is a differential operator (in particular, \(P\) is local), there exists a constant \(C = C(\widetilde{P}, \phi ) > 0\) such that we have an inequality \[\| \widetilde{P}(\phi f) \|_{L^{2}} \leq C( \|\chi \widetilde{P}f \|_{L^{2}} + \| \chi f \|_{L^{2}} )\] for any \(f \in C^{\infty}(\widetilde{E})\); see. Here, \(\phi, \chi \in C_{c}^{\infty}(\widetilde{M})\) are compactly supported smooth functions and one assumes \(\chi = 1\) on the support of \(\phi\). Thus by using Atiyah's technique of the proof of, we have the following: By Lemma [\[lem:closure\]](#lem:closure){reference-type="ref" reference="lem:closure"}, \(\widetilde{P}\) has the unique closed extension denoted by the same letter \(\widetilde{P}\) and thus the closure of the formal adjoint of \(\widetilde{P}\) (the formal adjoint is also Heisenberg elliptic) equals the Hilbert space adjoint \(\widetilde{P}^{\ast}\). On the other hand, any \(L^{2}\)-solutions of \(\widetilde{P}u = 0\) and \(\widetilde{P}^{\ast}u = 0\) are smooth by existence of a parametrix. Thus the orthogonal projection \(\Pi_{0}\) (resp. \(\Pi_{1}\)) onto a closed subspace \(\mathrm{Ker}_{L^{2}}(\widetilde{P})\) (resp. \(\mathrm{Ker}_{L^{2}}(\widetilde{P}^{\ast})\)) of the \(L^{2}\)-sections is of \(\Gamma\)-trace class since operators \(\phi \Pi_{0} \psi\) and \(\phi \Pi_{1} \psi\) are smoothing operators on compact sets. Thus one obtains the well-defined \(\Gamma\)-index of \(\widetilde{P}\). By using operators \(\widetilde{P}, \widetilde{Q}, \widetilde{S}_{0}\) and \(\widetilde{S}_{1}\), we define \[e_{\widetilde{P}} = \begin{bmatrix} \widetilde{S_{0}}^{2} & \widetilde{S_{0}}(1 + \widetilde{S_{0}}) \widetilde{Q} \\ \widetilde{S_{1}}\widetilde{P} & 1-\widetilde{S_{1}}^{2} \end{bmatrix} \quad \text{and} \quad e_{1} = \begin{bmatrix} 0 & 0 \\ 0 & 1 \end{bmatrix}.\] Since a difference \(e_{\widetilde{P}}-e_{1}\) is of \(\Gamma\)-trace class, one can define a \(\Gamma\)-index class of \(\widetilde{P}\). By using a \(\Gamma\)-trace, we have the \(\Gamma\)-index theorem in Heisenberg calculus. ## Example {#sub:exm} Index problems for Heisenberg elliptic operators on arbitrary closed Heisenberg manifolds are not solved yet. However, van Erp and Baum and van Erp solved the index problem on contact manifolds, which are good examples of Heisenberg manifolds. In this subsection, we investigate an example of Heisenberg elliptic differential operators with non-trivial \(\Gamma\)-index on a Galois covering over a closed contact manifold by using the index formula for subLaplacians. Let \(T^{2} = S^{1} \times S^{1} = \{ (e^{ix},e^{iy}) \}\) be a \(2\)-torus and set \[e(x,y) = \begin{bmatrix} f(x) & g(x) + h(x)e^{iy} \\ g(x) + h(x)e^{-iy} & 1-f(x) \end{bmatrix}.\] Here, let \(f\) be a \([0,1]\)-valued \(2\pi \mathbb{Z}\)-periodic function on \(\mathbb{R}\) such that \(f(0) = 1\) and \(f(\pi) = 0\) and set \(g(x) = \chi_{[0, \pi]}(x)\sqrt{f(x)-f(x)^{2}}\) and \(h(x) = \chi_{[\pi, 2\pi]}(x)\sqrt{f(x)-f(x)^{2}}\); see. Moreover, we assume \(f\), \(g\) and \(h\) are smooth functions. Then \(e\) defines an \(M_{2}(\mathbb{C})\)-valued smooth function on \(T^{2}\). Since \(e\) is an idempotent of rank \(1\), \(e\) defines an complex line bundle \(E\) on \(T^{2}\). As well known, the first Chern class \(c_{1}(E)\) of \(E\) is given by a \(2\)-form \[\frac{-1}{2\pi i}\mathrm{tr}(e(de)^{2}) = \frac{-1}{\pi}(hh' + 2f'h^{2}-2fhh')dx \wedge dy.\] Thus, by using an equality \(h^{2} = f-f^{2}\) on \([\pi, 2\pi]\), we can calculate the first Chern number of \(E\): \[\int_{T^{2}}c_{1}(E) =-\int_{\pi}^{2\pi}f'dx =-1.\] Let \(T^{3} = T^{2} \times S^{1} = \{ (e^{ix}, e^{iy}, e^{iz}) \}\) be a \(3\)-torus and \(q: T^{3} \to T^{2}\) the projection onto \(T^{2}\) of the first component. Set \(\theta_{k} = \cos (kz)dx-\sin (kz)dy\) for a positive integer \(k\), \(H_{k} = \mathrm{ker}(\theta_{k})\), \(f_{l}(x,y,z) = e^{ilz} + 1\) for a integer \(l\) and \(F = q^{\ast}E\). Then \((T^{3},H_{k})\) is a contact manifold and \(H_{k}\) is a flat vector bundle. Denote by \(T_{k}\) the Reeb vector field for \(\theta_{k}\) and \(\Delta_{H_{k}}^{F} =-\nabla_{X_{k}}^{F}\nabla_{X_{k}}^{F} -\nabla_{Y_{k}}^{F}\nabla_{Y_{k}}^{F}\) the sum of squares on \(F\), where \(\{ X_{k}, Y_{k} \}\) is a local frame of \(H_{k}\). Set \[P_{k,l} = \Delta_{H_{k}}^{F} + if_{l}\nabla_{T_{k}}^{F}.\] Since the values of \(f_{l}-n\) contained in \(\mathbb{C}^{\times}\) for any odd integer \(n\), an operator \(P_{k,l}: C^{\infty}(F) \to C^{\infty}(F)\) is a Heisenberg elliptic differential operator of Heisenberg order \(2\). By the index formula for \(P_{k,l}\) in, one has \[\mathrm{index} (P_{k,l}) = \int_{T^{3}}\frac{-1}{2\pi i}e^{-ilz}de^{ilz} \wedge c_{1}(F) = \frac{-1}{2\pi i}\int_{S^{1}}e^{-ilz}de^{ilz} \int_{T^{2}}c_{1}(E) = l.\] Note that a contact structure \(H_{k}\) is a lift of \(H_{1}\) by a \(k\)-fold cover \(p_{k}: T^{3} \to T^{3}\); \((e^{ix},e^{iy},e^{iz}) \mapsto (e^{ix},e^{iy},e^{ikz})\). Since the lift \(\widetilde{P_{1,l}}\) of a subLaplacian \(P_{1,l}\) by \(p_{k}\) equals \(P_{k,kl}\), we have the \(\Gamma (= \mathbb{Z}/k\mathbb{Z})\)-index of \(\widetilde{P_{1,l}}\): \[\mathrm{index}_{\Gamma}(\widetilde{P_{1,l}}) = \frac{1}{k}\mathrm{index} (\widetilde{P_{1,l}}) = \frac{1}{k}\mathrm{index} (P_{k,kl}) = l = \mathrm{index} (P_{1,l}).\] Next, we consider a general Galois covering of \(T^{3}\). Let \(X \to T^{3}\) be a Galois covering with a deck transformation group \(\Gamma\), which is a quotient of \(\pi_{1}(T^{3}) = \mathbb{Z}^{3}\), for example, \(X = \mathbb{R}^{3}\) and \(\Gamma = \mathbb{Z}^{3}\) the universal covering. By Theorem [\[thm:GammaHeisen\]](#thm:GammaHeisen){reference-type="ref" reference="thm:GammaHeisen"}, we have a non-trivial \(\Gamma\)-index as follows: \[\mathrm{index}_{\Gamma} (\widetilde{P_{k,l}}) = \mathrm{index} (P_{k,l}) = l.\]

{'timestamp': '2016-07-28T02:03:45', 'yymm': '1607', 'arxiv_id': '1607.04820', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04820'}

# Introduction {#Sec_Introduction} Modification of hadron properties in extreme environments such as finite temperature and density is one of the important problems in hadron physics, which is useful to clarify the relation between hadrons and the vacuum. Similarly, an external magnetic field comparable to the typical scale of quantum chromodynamics (QCD) may drastically change hadron properties. Such an intense field can be induced in noncentral heavy-ion collisions between two charged nuclei, where the strength is estimated to reach \(|eB| \sim m_\pi^2 \sim 0.02 \, \mathrm{GeV^2}\) at the Relativistic Heavy Ion Collider (RHIC) and \(|eB| \sim 15 m_\pi^2 \sim 0.3 \, \mathrm{GeV^2}\) at the Large Hadron Collider (LHC). Moreover, lattice QCD simulations are powerful tools to numerically confirm hadron properties in strong magnetic fields. For light mesons (with only up, down and strange quarks) in a magnetic field, many studies have been performed using SU(2) and SU(3) lattice gauge theory, AdS/QCD as well as other phenomenological models. Some of them were devoted to discussions on \(\rho\) meson condensation. There have been a few more studies for light baryons. It is known that the spontaneous chiral symmetry breaking (\(\langle \bar{q} q \rangle\) condensate) in the QCD vacuum is enhanced by a magnetic field, which is the so-called magnetic catalysis of \(\langle \bar{q} q \rangle\). This effect has been studied using various QCD-like models as well as lattice QCD simulations (see Refs.  for recent reviews). As temperature or density dependence of the \(\langle \bar{q} q \rangle\) condensate can modify hadron properties, the contribution of the magnetic field dependence of the condensate to hadron properties is also one of the more attractive topics in hadron physics. Properties of hadrons with heavy quarks (charm and bottom) in a magnetic field were also investigated by various approaches. They include potential models, QCD sum rules and AdS/QCD. Among them, the constituent quark model is a suitable tool to investigate properties of heavy hadrons in an external field. In such a model, there are mainly two physical effects which lead to the change of hadronic properties: (i) *mixing between different spin states* by the \(-\bm{\mu}_i \cdot \bm{B}\) term and (ii) *modification of quark kinetic energy* by the \(\bm{B} \times \bm{r}\) term. In addition, for heavy-light meson systems such as \(D\) and \(B\) mesons, we expect two additional effects (iii) *Zeeman splitting* by the \(-\bm{\mu}_i \cdot \bm{B}\) term and (iv) *magnetic field dependence of constituent light quark mass* induced by the magnetic catalysis. Furthermore, (v) *the anisotropy of the linear and Coulomb potential* between a quark and an antiquark manifests itself in a magnetic field. The latter effect was evaluated by recent lattice QCD simulations and its correction in the potential model for quarkonia was investigated in Ref. . In this work, we will discuss the effects of (i)-(iv) by focusing on the neutral heavy meson systems. The main purpose of this paper is a systematic understanding of heavy-meson spectra in a magnetic field, while our previous work is devoted to the findings of the level structure and wave function deformation in charmonia. To this end, we emphasize the importance of the comparison between charmonia and bottomonia, which is characterized by their different masses and electric charge. Furthermore, comparing quarkonia and heavy-light mesons is helpful to clarify the role of a light quark in heavy-meson systems under magnetic field. It is not technically easy to solve the two-body Schrödinger equation with a confining potential under strong magnetic field and to extract the anisotropic wave functions for ground and excited states simultaneously. In Refs. , the finite differential time domain (FDTD) method is applied to obtain the eigenvalues and eigenstates for the ground states of quarkonia. In our previous work, we proposed an alternative approach based on the variational method, which we call the cylindrical Gaussian expansion method (CGEM). This method is an extension from the conventional Gaussian expansion method (GEM) mainly used for nuclear and atomic few-body systems. The CGEM has the following nice properties: (i) it respects the symmetry of the Hamiltonian under constant magnetic field, (ii) it can deal fully with higher excited states and (iii) it reduces the computational cost substantially. This paper is organized as follows. In Sec. [2](#Sec_Formalism){reference-type="ref" reference="Sec_Formalism"}, we construct a two-body Hamiltonian for heavy mesons in a magnetic field and develop the CGEM to solve the Schrödinger equation in the cylindrical coordinate. Numerical results obtained from this approach are shown in Sec. [3](#Sec_Results){reference-type="ref" reference="Sec_Results"} and their physical pictures are discussed. Section [4](#Sec_Conclusion and outlook){reference-type="ref" reference="Sec_Conclusion and outlook"} is devoted to our conclusion and outlook. # Formalism {#Sec_Formalism} ## Quark model in a magnetic field Nonrelativistic two-body Hamiltonian of the constituent quark model in a magnetic field is written as \[H = \sum_{i=1}^2 \left[ \frac{1}{2m_i} \left( \bm{p}_i-q_i \bm{A} \right)^2-\bm{\mu}_i \cdot \bm{B} +m_i \right] +V(r),\] where \(m_i\), \(\bm{p_i}\), \(q_i\) and \(\bm{\mu}_i\) are the constituent quark mass, the momentum of quarks, the quark electric charge and the quark magnetic moment, respectively. For the vector potential, we choose the symmetric gauge \(\bm{A}(\bm{r}_i) = \frac{1}{2} \bm{B} \times \bm{r}_i\). We introduce the center of mass and relative coordinates, \(\bm{R}=(m_1 \bm{r}_1 + m_2 \bm{r}_2)/M\) and \(\bm{r} = \bm{r}_1-\bm{r}_2\), where \(M=m_1+m_2\) is the total mass of the two particles. As a new conserved quantity, we define the pseudomomentum: \[\hat{\bm{K}} = \sum_{i=1}^2 \left[ \bm{p}_i +\frac{1}{2} q_i \bm{B} \times \bm{r}_i \right]\] whose commutation relation is given by \([\hat{K}_i, \hat{K}_j] =-i(q_1+q_2) \epsilon_{ijk} B_k\). When the system is charge neutral, the components \(\hat{K}_i\) commute with each other. By using the pseudomomentum, the total wave function for the Hamiltonian can be factorized as follows: \[\Phi(\bm{R},\bm{r}) = \exp \left[ i (\bm{K}-\frac{1}{2} q\bm{B} \times \bm{r}) \cdot \bm{R} \right] \Psi(\bm{r}).\] For a neutral system with \(q_1=-q_2=q\), the total Hamiltonian can be reduced into \[\begin{aligned} H_\mathrm{rel} &=& \frac{\bm{K}^2}{2M}-\frac{q}{M}(\bm{K} \times \bm{B}) \cdot \bm{r} \nonumber\\ && + \frac{q}{2} \left( \frac{1}{m_1}-\frac{1}{m_2} \right) \bm{B} \cdot( \bm{r} \times \bm{p}) \\ && + \frac{\bm{p}^2}{2\mu} + \frac{q^2}{8\mu}( \bm{B} \times \bm{r})^2 + V(r) + \sum_{i=1}^2 [-\bm{\mu}_i \cdot \bm{B} + m_i], \nonumber \end{aligned}\] where \(\bm{p}=\frac{m_2\bm{p_1}-m_1\bm{p_2}}{m_1+m_2}\), \(\mu = m_1 m_2/M\) are the relative momentum and the reduced mass, respectively. If the orbital angular momentum along \(\bm{B}\) is zero, we have \(\bm{B} \cdot (\bm{r} \times \bm{p}) = 0\). When we focus on a magnetic field along the \(z\)-axis, \(\bm{B} =(0,0,B)\), \(\bm{B} \times \bm{r} = B(-y,x,0)\) and \(\bm{K} =(K_x,K_y,0)\), the Hamiltonian is reduced to, by using \(p=-i \nabla\), \[\begin{aligned} H_\mathrm{rel} &=& \frac{\bm{K}^2}{2M}-\frac{\nabla^2}{2\mu} + \frac{q^2 B^2}{8\mu} \rho^2 + \frac{qB}{4\mu} K_x y-\frac{qB}{4\mu} K_y x \nonumber\\ && + V(r) + \sum_{i=1}^2 [-\bm{\mu}_i \cdot \bm{B} + m_i], \label{H_rel2} \end{aligned}\] where \(\rho=\sqrt{x^2+y^2}\). In this work, we investigate only the case of \(\bm{K}=0\). In this situation, we can calculate the eigenvalues and eigenstates for mesons at rest. Then, the Hamiltonian maintains the rotational symmetry on the transverse \(x\)-\(y\) plane and the reflection symmetry along the \(z\)-axis. For the potential term in the model, we adopt the Cornell potential with a spin-spin interaction: \[\begin{aligned} V(r) &=& \sigma r-\frac{A}{r} + \alpha (\bm{S}_1 \cdot \bm{S}_2) e^{-\Lambda r^2} + C \nonumber\\ &=& \sigma \sqrt{\rho^2 + z^2}-\frac{A}{\sqrt{\rho^2 + z^2}} + \alpha (\bm{S}_1 \cdot \bm{S}_2) e^{-\Lambda (\rho^2 + z^2)} \nonumber\\ && + C, \end{aligned}\] where \(\bm{S}_1 \cdot \bm{S}_2 =-3/4\) and \(1/4\) for the spin singlet and triplet, respectively. Here, choosing the conventional Gaussian form for the spin-spin interaction leads to the analytic expressions of the matrix elements. Let us now consider the contribution of the effects induced by the term \(-\bm{\mu}_i \cdot \bm{B}\), where the quark magnetic moment is given by \(\bm{\mu}_i = g q_i \bm{S}_i/2m_i\) with the Land\(\acute{\mathrm{e}}\) \(g\)-factor. We have assumed \(g=2\). As a result, for a neutral meson system, we obtain \[\begin{aligned} -(\bm{\mu}_1 + \bm{\mu}_2) \cdot \bm{B} &=&-\frac{gq}{2} \left( \frac{\bm{S}_1}{m_1}-\frac{\bm{S}_2}{m_2} \right) \cdot \bm{B} \nonumber\\ &=&-\frac{gq}{4} \left( \frac{\bm{\sigma}^1}{m_1}-\frac{\bm{\sigma}^2}{m_2} \right) \cdot \bm{B}, \end{aligned}\] where \(\bm{\sigma}^i\) is the Pauli matrix. Then the eigenstates with the different spin quantum numbers, the singlet \(| 00 \rangle =(| \uparrow \downarrow \rangle-| \downarrow \uparrow \rangle) / \sqrt{2}\) and "longitudinal\" component \(| 10 \rangle =(| \uparrow \downarrow \rangle + | \downarrow \uparrow \rangle) / \sqrt{2}\) of the triplet, mix with each other through the off-diagonal matrix element: \[\langle 10 |-(\bm{\mu}_1 + \bm{\mu}_2) \cdot \bm{B} |00 \rangle =-\frac{gqB}{4} \left( \frac{1}{m_1} + \frac{1}{m_2} \right). \label{mat_ele_1}\] To take into account the off-diagonal components of the matrix elements, Eq. ([\[mat_ele_1\]](#mat_ele_1){reference-type="ref" reference="mat_ele_1"}), we have to solve a *coupled-channel* Schrödinger equation. On the other hand, "transverse\" components \(| 1 \pm1 \rangle\) (\(| \uparrow \uparrow \rangle\) and \(| \downarrow \downarrow \rangle\)) of the triplet cannot mix with other states and the off-diagonal components are zero. For neutral systems with \(m_1 \neq m_2\) (e.g. heavy-light mesons), the expectation values of these diagonal components are given by \[\langle 1 \pm1 |-(\bm{\mu}_1 + \bm{\mu}_2) \cdot \bm{B} | 1 \pm1 \rangle = \mp \frac{gqB}{4} \left( \frac{1}{m_1}-\frac{1}{m_2} \right). \label{mat_ele_2}\] For systems with \(m_1=m_2\) (quarkonia), these values become zero. Then, the couplings between the quark magnetic moments and the magnetic field are completely canceled between the quark and the antiquark. ## Cylindrical Gaussian expansion method In this section, we review the cylindrical Gaussian expansion method. After that, we introduce a coupled-channel basis to solve the coupled-channel Schrödinger equation for the mixing between the spin singlet \(| 00 \rangle\) and longitudinal component \(| 10 \rangle\) of the spin triplet. ### Cylindrical Gaussian basis We have no analytic solution for the two-body Hamiltonian with the confinement potential and a magnetic field. To solve the two-body Schrödinger equation, we use the GEM based on the Rayleigh-Ritz variational method, which is a powerful tool in nuclear and atomic physics. In a magnetic field, a wave function should be expanded individually by the Gaussian basis on the transverse plane, \(e^{-\beta_n \rho^2}\), and the basis along the longitudinal axis, \(e^{-\gamma_n z^2}\), in the cylindrical coordinate \((\rho,z,\phi)\). Such a cylindrical basis was first applied to the quark model in our previous work and a similar form was also successfully used for atomic systems such as hydrogen and helium in a magnetic field (e.g. Refs. ). The spatial part of the trial wave function for \(l_z=0\) is given as follows: \[\begin{aligned} && \Psi (\rho,z,\phi) = \sum_{n=1}^N C_n \Phi_n (\rho,z,\phi), \label{basis1} \\ && \Phi_n (\rho,z,\phi) = N_n e^{-\beta_n \rho^2} e^{-\gamma_n z^2}, \label{basis2} \end{aligned}\] where \(N\) and \(C_n\) are the number of basis functions and the expansion coefficients. \(N_n\) is the normalization constant defined by \(\langle \Phi_n| \Phi_n \rangle =1\). \(\beta_n\) and \(\gamma_n\) are Gaussian range parameters for the \(\rho\) and \(z\) directions, respectively. It is empirically known that the best set of the range parameters, \(\beta_1 \cdots \beta_N\) and \(\gamma_1 \cdots \gamma_N\), are the geometric progressions: \[\begin{aligned} && \beta_n = \frac{1}{\rho_n^2}, \ \ \ \rho_n = \rho_1 b^{n-1}, \label{geo1} \\ && \gamma_n = \frac{1}{z_n^2}, \ \ \ z_n = z_1 c^{n-1}. \label{geo2} \end{aligned}\] Here, the four parameters, \(\rho_1\), \(\rho_N\), \(z_1\) and \(z_N\), will be optimized as the energy eigenvalue is minimized. The analytic form of the matrix elements from the Hamiltonian ([\[H_rel2\]](#H_rel2){reference-type="ref" reference="H_rel2"}) and the wave function basis ([\[basis1\]](#basis1){reference-type="ref" reference="basis1"}) are summarized in Appendix [\[App_mat\]](#App_mat){reference-type="ref" reference="App_mat"}. We checked the applicability of our numerical code by comparing our variational result with the analytic solution of the three-dimensional anisotropic harmonic oscillator. ### Coupled-channel basis for spin mixing Since the spin singlet \(| 00 \rangle\) and longitudinal component \(| 10 \rangle\) of the spin triplet mix with each other in a magnetic field as seen in Eq. ([\[mat_ele_1\]](#mat_ele_1){reference-type="ref" reference="mat_ele_1"}), the wave function should take into account the coupled channel. Then the basis function is given as a linear combination of the spin-0 and spin-1 components: \[\begin{aligned} && \Psi (\rho,z,\phi) = \sum_{n=1}^N C^{S=0}_n \Phi^{S=0}_n (\rho,z,\phi) \nonumber \\ && \hspace{43pt} + \sum_{n=N+1}^{2N} C^{S=1}_n \Phi^{S=1}_n (\rho,z,\phi), \label{basis3} \\ && \Phi^{S=0}_n (\rho,z,\phi) = N^A_n e^{-\beta^A_n \rho^2} e^{-\gamma^A_n z^2}, \\ && \Phi^{S=1}_n (\rho,z,\phi) = N^B_n e^{-\beta^B_n \rho^2} e^{-\gamma^B_n z^2}, \end{aligned}\] where we have the four range parameters, \(\beta^A_n\), \(\gamma^A_n\), \(\beta^B_n\) and \(\gamma^B_n\). These parameters are also expressed by the geometrical progression, so that we finally optimize the eight parameters. A brief review of the generalized eigenvalue problem for the coupled-channel basis is shown in Appendix [\[App_GEP\]](#App_GEP){reference-type="ref" reference="App_GEP"}. ## Numerical setup {#subsec_parameter} In this section, we set the parameters of our constituent quark model. For parameters in charmonium systems, the charm-quark kinetic mass \(m_c=1.7840 \, \mathrm{GeV}\), the Coulomb parameter \(A=0.713\) and the string tension \(\sqrt{\sigma} = 0.402 \, \mathrm{GeV}\) are determined from the equal-time \(Q\bar{Q}\) Bethe-Salpeter amplitude in lattice QCD. For the spin-dependent potential given as \(\alpha \exp{(-\Lambda r^2)}\), we adopt \(\Lambda=1.020 \, \mathrm{GeV}^2\) obtained from lattice QCD by fitting a charmonium potential. The remaining parameters, \(\alpha=0.4778 \, \rm{GeV}\) and \(C=-0.5693 \, \rm{GeV}\), are chosen to reproduce the experimental values of the masses of the ground states of \(\eta_c\) and \(J/\psi\). We checked that these parameters can reproduce the masses of the ground and excited states of the charmonium well. For bottomonium system, we have no information from lattice QCD. Therefore, we fix \(\sqrt{\sigma}\) on the same value as that of charmonium. Furthermore, we put \(C=0\) and tune the bottom-quark kinetic mass \(m_b\), \(A\) and \(\alpha\) to reproduce the experimental masses of \(\eta_b\), \(\eta_b(2S)\) and \(\Upsilon(1S)\). For \(D_s\) meson systems, we know the values of \(m_s=0.656 \, \mathrm{GeV}\), \(A=1.30 \, \mathrm{GeV}\) and \(\sqrt{\sigma}=0.324 \, \mathrm{GeV}\) from the lattice QCD. Then, we tune only \(\alpha\) and \(C\) by matching our results with the experimental masses of \(D_s\) and \(D^\ast_s\) mesons. For \(D\), \(B\) and \(B_s\) meson systems, we fix \(A\) and \(\sqrt{\sigma}\) on the same values as those of \(D_s\) meson. The light-quark kinetic mass is fixed as \(m_q=0.3 \, \mathrm{GeV}\) and only \(\alpha\) and \(C\) are tuned by the experimental masses of \(D\)--\(D^\ast\), and \(B\)--\(B^\ast\) and \(B_s\)--\(B_s^\ast\) mesons. Under these criteria, the resulting parameters are summarized in Tables [1](#paramter_list_M){reference-type="ref" reference="paramter_list_M"} and [2](#paramter_list_P){reference-type="ref" reference="paramter_list_P"}. A list of the predicted meson masses is shown in Table [3](#paramter_list_com){reference-type="ref" reference="paramter_list_com"}.

{'timestamp': '2016-11-07T02:03:02', 'yymm': '1607', 'arxiv_id': '1607.04935', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04935'}

# Introduction {#sec:introduction} Quantitative investigations in logic, where properties and behaviour of typical objects are studied, form a rich and well-established branch of mathematics on the border of logic, combinatorics and theoretical computer science. From a combinatorial point of view, logical formulae are objects with finite representations and, sometimes several, naturally associated notions of size. In cases when the assumed size notion imposes finitely many objects of any size, we can consider uniformly random formulae within such a family of objects. Analysing the asymptotic behaviour of the probability that a uniformly random object of size \(n\) satisfies a certain property \(P\) as \(n\) tends to infinity, yields the notion of asymptotic density of \(P\) and, in consequence, leads to investigations how certain natural properties, such as satisfiability etc., behave in the case of typical formulae. There is a long history of using this kind of asymptotic approach applied to logic (see, e.g. ) and recently its computational aspects. In , David et al. initiated quantitative investigations in lambda calculus and combinatory logic. Considering the set of closed \(\lambda\)-terms in a canonical representation where variables do not contribute to the term size, David et al. showed that typical \(\lambda\)-terms are strongly normalising whereas in the case of \(\S \mathbf{K}\)-combinators the situation is precisely opposite--asymptotically almost no \(\S \mathbf{K}\)-combinator is strongly normalising. Somewhat contrary to their result, Bendkowski et al. in  considered a different representation of \(\lambda\)-terms with de Bruijn indices, showing that similarly to the case of \(\S \mathbf{K}\)-combinators, asymptotically almost no \(\lambda\)-term is strongly normalising. Despite many efforts, the associated counting problem for closed \(\lambda\)-terms is still one of the remaining major open problems. Throughout the years, different variants of lambda calculus have been considered. In , Bodini et al. studied the enumeration of BCI \(\lambda\)-terms. John Tromp in , as well as Grygiel and Lescanne in , considered the counting problem in the so called binary lambda calculus. Recently, in  Gittenberger and Gołębiewski considered \(\lambda\)-terms in the de Bruijn notation with various size notions, giving tight lower and upper asymptotic bounds on the number of closed \(\lambda\)-terms. Due to the lack of bound variables, combinatory logic circumvents the intrinsic issues present in lambda calculus, serving as a minimalistic formalism capable of expressing all lambda-definable functions. Its simple syntax yields a natural representation using plane binary trees with labelled leaves, which greatly facilitates the analysis of asymptotic properties. The paper is organised as follows. In  we recall preliminary definitions and notation regarding combinatory logic. In the following sections [3](#sec:combinatorial-classes){reference-type="ref" reference="sec:combinatorial-classes"} and [4](#sec:generating-functions-and-analytic-tools){reference-type="ref" reference="sec:generating-functions-and-analytic-tools"}, we state basic notions of combinatorial classes and generating functions, listing main tools used to study asymptotic properties of combinators, in particular, the analytic methods of singularity analysis. In  we study the class of plane binary trees with labelled leaves, deriving basis-independent combinatory logic results as easy corollaries. In  we focus on the \(\S \mathbf{K}\)-basis proving, inter alia, that for each positive \(n\), the set of combinators reducing in \(n\) normal-order reduction steps has positive asymptotic density. Finally, in  we discuss some experimental results. # Combinatory logic {#sec:combinatory-logic} Let \(\mathcal{B}\) be a finite *basis* of primitive combinators. The set \(\class{\mathcal{B}}\) of *\(\mathcal{B}\)-combinators* is defined inductively as follows. If \(\comb{X} \in \mathcal{B}\), then \(\comb{X} \in \class{\mathcal{B}}\). If \(N, M \in \class{\mathcal{B}}\), then \((N M) \in \class{\mathcal{B}}\). In the latter case, we say that \((N M)\) is an *application* of \(N\) to \(M\). If the underlying basis is clear from the context, we simply write *combinators* instead of \(\mathcal{B}\)-combinators. Following standard notational conventions (see, e.g. ), we omit outermost parentheses and drop parentheses from left-associated combinators, e.g. instead of writing \(((M N) (P Q))\) we write \(M N (P Q)\). Each primitive combinator \(\comb{X} \in \mathcal{B}\) contributes a *reduction rule* of the form \[\comb{X} N_1 \ldots N_m \to M\] where \(m \geq 1\), \(M \in \class{\set{N_1, \ldots, N_m}}\) and \(N_1 \ldots N_m\) are arbitrary combinators. In other words, \(\comb{X} N_1 \ldots N_m\) *reduces* to some determined combinator \(M\) built from \(N_1,\ldots,N_m\) and term application. The reduction relation \(\to\) is then extended onto all combinators such that if \(P \to Q\), then \(PR \to QR\) and \(RP \to RQ\) for each combinator \(R\). Let \(P\) be a combinator. If there exists no combinator \(Q\) such that \(P \to Q\), then \(P\) is said to be in *normal form*. If there exists a finite sequence of combinators \(P_0,P_1,\ldots,P_k\) such that \(P = P_0 \to P_1 \to \ldots \to P_k\) and \(P_k\) is in normal form, then \(P\) is *weakly normalising*, or simply *normalising*. If there does not exist an infinite sequence of combinators \(P_0,P_1,\ldots\) such that \(P=P_0 \to P_1 \to \ldots\), then we say that \(P\) is *strongly normalising*. Naturally, strong normalisation implies weak normalisation. A major part of combinatory logic is devoted to its simple type theory corresponding to the implicational fragment of minimal logic (see, e.g. ) and hence also to simply typed \(\lambda\)-calculus (see ). In its most common basis \(\mathcal{B} = \set{\comb{S},\comb{K}}\), the type-assignment deduction system \(TA_{\comb{S} \comb{K}}\) is given by the following two axiom schemes for \(\comb{S}\) and \(\comb{K}\), with a single *modus ponens* inference rule. \[\tag{Axiom \(\comb{S}\)} \begin{mathprooftree} \AxiomC{} \UnaryInf\(\comb{S}: (\alpha \to \beta \to \gamma) \to (\alpha \to \beta) \to \alpha \to \gamma \fCenter\) \end{mathprooftree}\] \[\tag{Axiom \(\comb{K}\)} \begin{mathprooftree} \AxiomC{} \UnaryInf\(\comb{K}: \alpha \to \beta \to \alpha\fCenter\) \end{mathprooftree}\] \[\tag{Modus ponens} \begin{mathprooftree} \AxiomC{\(N: \alpha \to \beta\)} \AxiomC{\(M: \alpha\)} \BinaryInf\((N M): \beta\fCenter\) \end{mathprooftree}\] The primitive combinators \(\comb{S}\) and \(\comb{K}\) can be assigned any types fitting to their axiom schemes. On the other hand, an application \((N M)\) can be assigned a type \(\beta\), denoted \((N M): \beta\), if and only if \(N: \alpha \to \beta\), whereas \(M: \alpha\). If \(P: \alpha\) for some type \(\alpha\), we say that \(P\) is *typeable*. Naturally, not every combinator is typeable, e.g. \(\omega := \comb{S} \comb{I} \comb{I}\), where \(\comb{I} := \comb{S} \comb{K} \comb{K}\). Though not all strongly normalising combinators are typeable, the converse implication holds, i.e. each typeable combinator is strongly normalising. In this paper, we focus mostly on bases which are capable of expressing all computable functions. A sufficient and necessary condition for a basis \(\mathcal{B}\) to be *Turing-complete* is to span \(\mathcal{B}\)-combinators extensionally-equivalent to \(\comb{S}\) and \(\comb{K}\). In such a case, we make the natural assumption that \(\mathcal{B}\) defines a set of axiom schemes, one for each primitive combinator, in such a way that when enriched with the modus ponens inference rule, \(TA_{\mathcal{B}}\) constitutes a sound typing system--if \(N\) is extensionally-equivalent to either \(\comb{S}\) or \(\comb{K}\), then the types of \(N\) in \(TA_{\mathcal{B}}\) are the same as the types of, respectively, \(\comb{S}\) or \(\comb{K}\) in \(TA_{\comb{S} \comb{K}}\). Throughout the paper, we use the over line notation \(\overline{\comb{X}}\) to denote arbitrary terms extensionally-equivalent to \(\comb{X}\). We refer the reader to  or  for a more detailed exposition of combinatory logic. # Combinatorial classes {#sec:combinatorial-classes} Let \(B\) be a countable set of objects with an associated *size* function \(f \colon B \to \mathbb{N}\). If for each \(n \in \mathbb{N}\) the set of \(B\)'s objects of size \(n\) is finite, then \(B\) together with \(f\) forms a *combinatorial class* (see, e.g. ). In such a case, we can associate a *counting sequence* \(\set{b_n}_{n \in \mathbb{N}}\) of natural numbers \(b_n\) capturing the number of objects in \(B\) of size \(n\). Naturally, if \(A \subseteq B\), then \(A\) is a combinatorial class as well with \(a_n \leq b_n\) for each \(n \in \mathbb{N}\). Assuming that \(b_n > 0\) for each \(n\in \mathbb{N}\), we can then define the *asymptotic density* \(\density{A}{B}\) of \(A\) in \(B\) as \[\density{A}{B} = \lim_{n\to\infty} \frac{a_n}{b_n}.\] Note that if it exists, we can interpret \(\density{A}{B}\) as the asymptotic probability of finding an object of \(A\) in the class of objects \(B\). In other words, the likelihood that \(A\) represents 'typical' objects in \(B\). Unfortunately, sometimes we do not know whether the asymptotic density of \(A\) in \(B\) exists, however we can use the *lower* and *upper limits* defined as \[\densityminus{A}{B} = \liminf_{n\to\infty} \frac{a_n}{b_n} \quad \text{and} \quad \densityplus{A}{B} = \limsup_{n\to\infty} \frac{a_n}{b_n}.\] As \(0 \leq \frac{a_n}{b_n} \leq 1\), these two numbers are well defined for any set \(A\), even when the limiting ratio \(\density{A}{B}\) is not known to exist. Henceforth, given a combinatorial class \(A\), we use \(a_n\) to denote the number of objects in \(A\) of size \(n\). # Generating functions and analytic tools {#sec:generating-functions-and-analytic-tools} Let \(A\) be a combinatorial class. The formal power series \(A(z) = \sum_{n \geq 0} a_n z^n\) with \(A\)'s counting sequence \(\set{a_n}_{n\in \mathbb{N}}\) as coefficients, is called the *ordinary generating function* of \(A\). Using the powerful theory of *Analytic Combinatorics* developed by Flajolet and Sedgewick , many questions concerning the asymptotic behaviour of \(\set{a_n}_{n \in \mathbb{N}}\) can be efficiently resolved by analysing the behaviour of \(A(z)\) viewed as an analytic function in some neighbourhood around the complex plane origin. This is the approach we take to study the asymptotic fractions of interesting combinatory logic terms. Throughout the paper we use \(A(z)\) to denote the ordinary generating function associated with the combinatorial class \(A\). We write \([z^n]A(z)\) to denote the coefficient standing by \(z^n\) in the Taylor series expansion of \(A(z)\) around \(z = 0\). We say that two sequences \(\set{a_n}_{n \in \mathbb{N}}\) and \(\set{b_n}_{n \in \mathbb{N}}\) are *asymptotically equivalent* if \(\lim_{n \to \infty} \frac{a_n}{b_n} = 1\). In such a case we write \(a_n \sim b_n\). ## Main tools {#subsec:main-tools} In our endeavour to study the asymptotic behaviour of 'typical' classes of combinatory logic terms, we use the method of *singularity analysis* . Starting with a particular class \(A\) of combinators, we find its corresponding generating function \(A(z)\). The location of \(A(z)\)'s dominant singularities determines the exponential growth rate of \(\set{a_n}_{n \in \mathbb{N}}\) as dictated by the following theorem. In the case of analytic functions derived from combinatorial classes, the location of dominant singularities is significantly simplified as it suffices to look for singularities on the real line. The sub-exponential factors determining the asymptotic growth rate of \(\set{a_n}_{n \in \mathbb{N}}\) can be then further established using, in our case, the standard function scale for algebraic singularities of square-root type. Finally, combining the scaling rule for Taylor expansions and the Newton-Puiseux expansion of algebraic functions with unique dominating singularities, we obtain the following corollary theorem. In order to simplify the reasoning about the type and location of singularities of generating functions given without explicit closed-form solutions, we use the following technical lemma guaranteeing certain natural closure properties of analytic functions with a single square-root type dominating singularity. ## Removable singularities {#subsec:removable-singularities} In order to apply  to the analysis of a generating function \(A(z)\), we have to guarantee that \(A(z)\) is analytic at \(z = 0\). If it is not the case, yet \(A(z)\) has a *removable pole singularity* at \(z = 0\), we can consider its analytic extension \(\widetilde{A}(z)\), instead of \(A(z)\). The following theorem due to Bernhard Riemann provides a sufficient and necessary condition to determine whether \(A(z)\)'s pole singularity at \(z = 0\) can be removed. As a direct consequence, we obtain the following technical lemma. # Basis-independent results {#sec:basis-independent-results} In this section we are interested in universal basis-independent asymptotic properties of combinatory logic. We prove certain general results about labelled plane binary trees, deriving the combinatory logic results as immediate corollaries. Let us notice that the asymptotic growth rate of \(L\)-trees greatly depends on the asymptotic approximation of Catalan numbers \(\catalan{n}\) counting the number of plane binary trees with \(n\) inner nodes. It is well known that \[\catalan{n} = \frac{1}{n+1}{2n \choose n} \qquad \text{and} \qquad \catalan{n} = 4^n \frac{n^{-\nicefrac{3}{2}}}{\sqrt{\pi}}.\] ## Counting \(L\)-trees containing fixed \(L\)-trees as subtrees {#subsec:counting-l-trees} Suppose that \(t \in \T{L}\). Let \(\overline{\T{L}}(z)\) denote the generating function counting the cardinalities of \(L\)-trees containing \(t\) as a subtree. In the following series of propositions, we derive the closed-form solution for \(\overline{\T{L}}(z)\) and check the conditions of  used subsequently to show that in fact \([z^n]\overline{\T{L}}(z) \sim [z^n]\T{L}(z)\), independently of \(L\). The generating functions \(\T{L}(z)\) and \(\overline{\T{L}}(z)\) are not defined at \(z = 0\), however due to , both have analytic extensions to functions analytic in the origin and we can consider them instead of \(\T{L}(z)\) and \(\overline{\T{L}}(z)\) in the subsequent theorem. Immediately, we obtain the following corollary theorem. The above theorem provides a general result showing that 'local' properties of \(L\)-trees propagating to supertrees, span asymptotically almost the whole set of \(L\)-trees. Using the natural bijection between \(\mathcal{B}\)-combinators and \(\mathcal{B}\)-trees, we can reinterpret this observation in the language of combinatory logic and state that each 'local' property of \(\mathcal{B}\)-combinators closed under taking superterms is typical, i.e. has asymptotic probability \(1\) in the set of all \(\mathcal{B}\)-combinators. In particular, we obtain the following corollaries generalising the results in . ## Counting normalising combinators {#subsec:counting-normalizing-combinators} Let \(\mathcal{B}\) be a Turing-complete set of primitive combinators. Let us start with the classical observation is that the set of normalising \(\mathcal{B}\)-combinators is undecidable. It follows that its corresponding generating function has no computable closed-form solution. For that reason we take the following approach. We find feasible subclasses of normalising and non-normalising \(\mathcal{B}\)-combinators and use them to bound the density of normalising combinators \(\mathcal{WN}_\mathcal{B}\). Let us start with recalling the famous standardisation theorem. Although usually stated in the \(\S \mathbf{K}\)-basis, the standardisation theorem easily generalises to every Turing-complete basis of combinators, e.g. through the classical translation to \(\lambda\)-calculus (see, e.g. ). Hence, we obtain the following result. Using the fact that asymptotically no \(\mathcal{B}\)-combinator is strongly normalising (see ), we obtain the following corollary generalising the result in . # SK-combinators {#sec:sk-combinators} In this section we address the problem of estimating the asymptotic density of normalising \(\comb{S}\comb{K}\)-combinators in the set of all \(\comb{S}\comb{K}\)-combinators. In  authors provided the following bounds: \[\frac{1}{32} \leq \densityminus{\mathcal{WN}_{\comb{S}\comb{K}}}{\class{\comb{S}\comb{K}}} \quad \text{and} \quad \densityplus{\mathcal{WN}_{\comb{S}\comb{K}}}{\class{\comb{S}\comb{K}}} \leq 1-\frac{1}{2^{18}}.\] Here, we prove that \(\comb{S}\comb{K}\)-combinators reducing in exactly \(n>0\) normal-order reduction steps (leftmost outermost redex first, see ) have positive asymptotic density in the set of all \(\comb{S}\comb{K}\)-combinators. We provide a constructive method of finding their asymptotic approximations and related densities, yielding a systematic approach to improving the above lower bound. For simplicity, we use \(\class{\comb{S}\comb{K}}\) and \(\rg{0}\) to denote the set of \(\comb{S}\comb{K}\)-combinators and the set of normal forms, respectively. Let us start with a few technical propositions regarding the generating functions \(C(z)\) and \(\rgf{0}\). Our method relies on the effective computation and asymptotic properties of *normal-order reduction grammars* \(\set{\rg{n}}_{n\in\mathbb{N}}\). In  the author provided a recursive algorithm, which for given \(n \geq 1\), constructs the \(n\)th normal-order reduction grammar \(\rg{n}\) defining the set of \(\S \mathbf{K}\)-combinators reducing in exactly \(n\) normal-order reduction steps. Applying the Symbolic Method of Flajolet and Sedgewick , \(\rg{n}\) is then translated into a functional equation \[\label{eq:rgf-gen-fun-eq-psi} \rgf{n} = \Psi(\rgf{n},z)\] involving the generating function \(\rgf{n}\) and its formal parameter \(z\). Due to the specific structure of \(\Phi(\rg{n})\)--the set of productions \(\alpha \in \rg{n}\) not referencing \(\rg{n}\)-- [\[eq:rgf-gen-fun-eq-psi\]](#eq:rgf-gen-fun-eq-psi){reference-type="eqref" reference="eq:rgf-gen-fun-eq-psi"} turns out to be linear in \(\rgf{n}\), yielding a unique closed-form solution. In the reminder of this section, we exploit the structure of the normal-order reduction grammars, showing the following main result. Before we provide a proof, let us present two propositions preparing the background for . In the following proposition we use the notions of major and minor productions, showing that major productions contribute to the asymptotic growth rate of \(\rgf{n}\)'s counting sequence, whereas minor ones are asymptotically negligible. Now we are in a position to prove . As \(\domsing{m} = \nicefrac{1}{8}\) for every \(m \geq 1\), we can easily compute the coefficients \(\overline{C}_m\) in the asymptotic approximation of \([z^n]\rgf{m}\) using available computer algebra systems, e.g. Mathematica  . The quotient \(\nicefrac{\overline{C}_m}{-4}\) (see ) yields the desired asymptotic density of \(\S \mathbf{K}\)-combinators normalising in \(m\) normal-order reduction steps in the set of all \(\S \mathbf{K}\)-combinators. Hence, we obtain the following corollary. Using an implementation of the normal-order reduction grammar algorithm together with Mathematica  we were able to compute the densities of combinators reducing in \(m\) normal-order reduction steps \(\rg{m}\) in \(C_{\S \mathbf{K}}\) for \(m = 1,\ldots,7\). The results are summarised in the following figure. And so, exploiting the finite additivity of asymptotic density we obtain the following improved lower bound: \[0.34010402598726164 \leq \densityminus{\mathcal{WN}_{\comb{S}\comb{K}}}{\class{\comb{S}\comb{K}}}.\] Clearly, the above lower bound can be further improved if we compute the next asymptotic densities \(\density{\rg{8}}{\class{\S \mathbf{K}}},\density{\rg{9}}{\class{\S \mathbf{K}}},\density{\rg{10}}{\class{\S \mathbf{K}}}\), etc. Unfortunately, due to the sheer amount of major productions, this process is immensely time and memory consuming, quickly requiring resources exceeding current desktop computer capabilities. Nevertheless, \(\density{\rg{m}}{\class{\S \mathbf{K}}} \leq 1\) for each \(m\) and hence \(\sum_{m \geq 0} \density{\rg{m}}{\class{\S \mathbf{K}}}\) is necessarily convergent to some value \(0 < \zeta < 1\). Moreover, if \(\density{\mathcal{WN}_{\comb{S}\comb{K}}}{\class{\S \mathbf{K}}}\) exists, then we have \(\zeta \leq \density{\mathcal{WN}_{\comb{S}\comb{K}}}{\class{\S \mathbf{K}}}\). The intriguing problem of determining whether the inequality can be replaced by the equality still remains open. # Experimental results {#sec:experimental-results} Our method developed in  allows us to improve the lower bound on \(\densityminus{\mathcal{WN}_{\comb{S}\comb{K}}}{\class{\comb{S}\comb{K}}}\), provided we have enough computational resources to find and manipulate generating functions \(\rgf{m}\) for higher \(m\). Unfortunately, the current gap between the lower and upper bound on the density of normalising combinators is still quite significant. In this section we present some experimental results regarding the aforementioned density as well as numerical evaluations of the obtained approximation error. ## Super-computer results {#subsec:super-computer-results} Consider the following experiment scheme \(G(s,n,r)\) with three positive integer parameters \(s,n,r\). We draw \(s\) uniformly random \(\S \mathbf{K}\)-combinators of size \(n\) using, e.g. Rémy's algorithm (see )--drawing a uniformly random plane binary tree with \(n\) inner nodes--combined with a random \(\S \mathbf{K}\)-labelling. Then we reduce each of the \(s\) samples using up to \(r\) normal-order reduction steps. We record then the number of normalised samples, with their corresponding reduction lengths. For samples which did not normalise in \(r\) reduction steps, we artificially record their reduction lengths as \(-1\). Afterwards, we collect the reduction lengths, plotting the obtained function mapping reduction lengths to the number of samples attaining the given reduction length. We preformed our experiments on the *Prometheus*  super-computer cluster granted by ACC Cyfronet AGH in Kraków, Poland ( out of 500 world's most powerful supercomputer in 2016, with theoretical computational power of 2.4 Pflops) . The following figure summarises the experiment result for \(G(s = 1200,n = 50000000,r = 1000)\). Even though the number \(r = 1000\) bounding the number of performed reductions was significantly smaller than the size of considered samples, the experiment revealed that normalising combinators have short reduction lengths. In fact, the mean reduction length \(E(X)\) over all normalising samples is approximately \(E(X) \approx 31.5810\), whereas \(\log_2 n \approx 25.5754\). Out of \(1200\) samples, only \(176\) did not normalise in \(r\) steps, yielding approximately \(0.146\) percent of all considered samples. Similar results were obtained with different experiment parameters, suggesting that the ratio of normalising \(\S \mathbf{K}\)-combinators should be approximately equal to \(85\%\), whereas the mean reduction length of normalising terms is \(\Theta(\log_2 n)\). Our Haskell implementation of the program, as well as all the obtained data sets are available at . ## Approximation error {#subsec:approximation-error} In  we showed that \([z^n]\rgf{m} = \Theta([z^n]C(z))\) for \(m \geq 1\). Nonetheless, techniques used to obtain this result do not yield the convergence rate of sequences in question. Using Mathematica  we compared \([z^n]\rgf{m}\) with its approximation \(8^n \widetilde{C}_m n^{-\nicefrac{3}{2}}\). The following figures summarise values for \([z^n]\rgf{1}\) and their relative error \(\delta([z^n]\rgf{1})\). Since \([z^n]\rgf{1} \sim 8^n \widetilde{C}_1 n^{-\nicefrac{3}{2}}\), the relative error \(\delta([z^n]\rgf{1})\) is inevitably tending to \(0\) as \(n \to \infty\). Remarkably, the error \(\delta([z^n]\rgf{1})\) converges much slower than one would expect. With \(n = 300\) the error is just of order \(10^{-2}\). We observed similar results in the relative errors for higher \(n\), where the convergence rate is even slower than in the case of \([z^n]\rgf{1}\). Our Mathematica  scripts and an implementation of the algorithm computing \(\rgf{m}\) are available at . # Conclusion {#sec:conclusion} We presented several basis-independent results regarding asymptotic properties of combinatory logic. In particular, we generalised previously known results for \(\S \mathbf{K}\)-combinators from , showing that they span to arbitrary Turing-complete combinator bases. Exploiting the results of , we gave a systematic approach to finding better lower bounds on the density of normalising \(\S \mathbf{K}\)-combinators, improving the previously best known lower bound from about \(3\%\) to approximately \(34\%\). Performed super-computer experiments suggest that the searched density, if it exists, should be approximately \(85\%\), conjecturing that both lower and upper bounds are still quite far from the actual density. Naturally, the following interesting question emerges--is our approach converging to the actual density? As the asymptotic density is always bounded, the series \(\sum_{m\geq 0} \density{R_m}{\class{\comb{S}\comb{K}}}\) converges to some value \(\zeta \in (0,1)\). Is \(\zeta\) the desired asymptotic density of normalising \(\S \mathbf{K}\)-combinators? If not, how far is it from the actual density? Due to the immense computational resources required in the computations and the sheer amount of major normal-order reduction grammar productions, our approach renders brute-force methods of closing the density gap virtually impossible. We expect that more sophisticated techniques are required in order to address this intriguing open problem.

{'timestamp': '2016-07-19T02:08:36', 'yymm': '1607', 'arxiv_id': '1607.04908', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04908'}

# Introduction Let \(d\geq 3\) and let \(G(n,d)\) denote the random \(d\)-regular graph on \(n\) vertices (see e.g. the monograph ). Equivalently, we can think of \(G(n,d)\) as a random model of symmetric \(0-1\) matrices in which the row sums are conditioned to be \(d\). It is expected that the spectral properties of \(G(n,d)\) are closely related to random matrix theory; however, many questions in the area are still open. It is well known that the spectral measure of \(G(n,d)\) converges to the so called Kesten--McKay measure in the weak topology as \(n\) goes to infinity. This gives an approximate semicircle law if \(d\) is large. A famous result by J. Friedman solves Alon's second eigenvalue conjecture showing that \(G(n,d)\) is almost Ramanujan. Much less is known about the scaled eigenvalue spacing and about the structure of the eigenvectors. Recent results in the area include on the second eigenvalue; on eigenvalue spacing, local semicircle law and functional limit theorems; on the delocalization of the eigenvectors. In the present paper we study approximate eigenvectors or shortly: *almost eigenvectors* of \(G(n,d)\) i.e. vectors that satisfy the eigenvector equation \((A-\lambda I)v=0\) with some small error. Almost eigenvectors are not necessarily close to proper eigenvectors. They are much more general objects. For example any linear combination of eigenvectors with eigenvalues in the interval \([\lambda-\varepsilon,\lambda+\varepsilon]\) is an almost eigenvector with error depending on \(\varepsilon\). In general a vector is an almost eigenvector if and only if its spectral measure is close to a Dirac measure in the weak topology. We show that despite of this generality, almost eigenvectors of \(G(n,d)\) have a quite rigid structure if \(n\) is big. Our main result implies that every almost eigenvector (with entry sum 0 and normalized to have length \(\sqrt{n}\)) has an entry distribution close to some Gaussian distribution \(N(0,\sigma)\) in the weak topology where \(0\leq\sigma\leq 1\). Note that if \(\sigma=0\), then the \(\ell^2\)-weight of the vector is concentrated on a small fraction of the vertices. Such vectors are called localized. Our main result holds even in a stronger sense where joint distributions are considered using the local structure of the graph. In some sense our result is best possible since there are examples for both localized and delocalized almost eigenvectors (see chapter [3](#prelim){reference-type="ref" reference="prelim"}). Note that proper eigenvectors are conjectured to be delocalized. The issue of eigenvector Gaussianity goes back to random matrix theory. It is not hard to show that in the GUE (Gaussian Unitary Ensemble) random matrix model every eigenvector has a near Gaussian entry distribution. It is much harder to analyze the random model when the elements of the matrix are chosen from a non Gaussian distribution. Nevertheless Gaussianity of the eigenvectors is proved under various conditions for generalized Wigner matrices and also for various other models (see e.g. and for recent surveys). Sparser models are harder to analyze. This paper deals with the sparsest case, where the matrix is the adjacency matrix of a random \(d\)-regular graph with some fixed \(d\). In this case there is a stronger meaning of eigenvector Gaussianity. For example it is natural to ask about the Gaussianity of the joint distribution of the entries at the two endpoints of a randomly chosen edge in the graph. More generally one can look at the joint distribution of the entries in random balls of radius \(r\). Our Gaussianity results for almost eigenvectors are established in this strong sense. Furthermore, it makes sense to study the joint distribution of the entries of several almost eigenvectors. More precisely, a \(k\)-tuple of almost eigenvectors can be interpreted as a function from the vertices to \(\mathbb R^k\). This function evaluated at a randomly chosen vertex gives a probability distribution on \(\mathbb R^k\) that we define as the joint distribution of the entries. It is an interesting question whether such joint distributions are Gaussian. We prove this if \(k\) is fixed, \(n\) is large and the eigenvalues corresponding to the almost eigenvectors are close to each other. Moreover, we also prove this result when the joint distribution of many eigenvectors is considered in random neighborhoods. The proof of our main theorem is based on the so-called local-global graph limits. However, to keep the paper self-contained, we use a slightly simplified framework (see section [5](#typpro){reference-type="ref" reference="typpro"}) optimized for this particular problem. We relate the properties of random regular graphs to random processes on the infinite \(d\)-regular tree \(T_d\). Most of the work is done in this convenient limiting framework. An invariant process on \(T_d\) is a joint distribution of random variables \(\{X_v\}_{v\in T_d}\) labeled by the vertices of \(T_d\) such that it is invariant under the automorphism group of \(T_d\). A special class of invariant processes, called *typical processes*, was introduced in. Roughly speaking, an invariant process is typical if it can be obtained as the Benjamini--Schramm limit of colored random regular graphs. There is a correspondence between the properties of typical processes on \(T_d\) and the properties of large random \(d\)-regular graphs. Our main theorem is equivalent to the statement that if a typical process satisfies the eigenvector equation at every vertex of \(T_d\) and has finite variance (at every vertex), then the process is jointly Gaussian. Note that such Gaussian eigenvector processes on \(T_d\) are completely characterized and there is a unique one for each possible eigenvalue. A key ingredient in our proof is a general entropy inequality for typical processes. It implies that typical eigenvector processes obey another inequality involving differential entropy. From here we finish the proof using heat propagation on the space of typical eigenvector processes combined with DeBrujin's identity for Fisher information. Gaussianity will follow from the fact that heat propagation converges to a Gaussian distribution. A well studied subclass of typical processes is the class of factor of i.i.d. processes. These processes appeared first in ergodic theory but they are also relevant in probability theory, combinatorics, statistical physics and in computer science. Not every typical process is factor of i.i.d.; this follows from the results of Gamarnik and Sudan; see also Rahman and Virág. Despite of recent progress in the area, a satisfying understanding of factor of i.i.d. processes is only available in the case \(d=2\), which is basically equivalent to the framework of classical ergodic theory of \(\mathbb{Z}\) actions. Our results imply that if an invariant process (with finite variance) is in the weak closure of factor of i.i.d. processes and satisfies the eigenvector equation then the process is Gaussian. This answers a question of B. Virág. **Outline of the paper.** In Chapter [2](#chap:main){reference-type="ref" reference="chap:main"}, we formulate the main results for finite random \(d\)-regular graphs. Chapter [3](#prelim){reference-type="ref" reference="prelim"} and [4](#chap:eigen){reference-type="ref" reference="chap:eigen"} contain general statements about invariant processes, eigenvector processes, entropy and almost eigenvectors. Chapter [5](#typpro){reference-type="ref" reference="typpro"} provides the translation of our main result to the infinite setting using typical eigenvector processes on the infinite \(d\)-regular tree. In Chapter [6](#chap:entr){reference-type="ref" reference="chap:entr"}, we prove a necessary condition for a process to be typical in the form of an entropy inequality. In Chapter [7](#smooth){reference-type="ref" reference="smooth"} we reduce the limiting form of the main theorem to a special family of eigenvector processes called smooth eigenvector processes. Chapter [8](#chap:entrtyp){reference-type="ref" reference="chap:entrtyp"} gives a differential entropy inequality for smooth eigenvector processes. In Chapter [9](#chap:cov){reference-type="ref" reference="chap:cov"}, we calculate the eigenvalues of special submatrices of the covariance matrices of eigenvector processes corresponding to balls around vertices and edges on the tree. In Chapter [10](#chap:impr){reference-type="ref" reference="chap:impr"}, we use the results from Chapter [9](#chap:cov){reference-type="ref" reference="chap:cov"} to prove that among smooth typical eigenvector processes the Gaussian minimizes the differential entropy formula from Chapter [8](#chap:entrtyp){reference-type="ref" reference="chap:entrtyp"}. On the other hand, in Chapter [11](#chap:heat){reference-type="ref" reference="chap:heat"}, we show that the Gaussian eigenvector process maximizes the same formula. Moreover, we finish the proof of the main result in Chapter [11](#chap:heat){reference-type="ref" reference="chap:heat"}. # The main theorem {#chap:main} In this chapter we state the main theorem first in a simpler but weaker form and later in the strong form. An \(\varepsilon\)-almost eigenvector of a matrix \(A\in\mathbb{R}^{n\times n}\) with eigenvalue \(\lambda\) is a vector \(v\in\mathbb{R}^n\) such that \(\|v\|_2=1\) and \(\|Av-\lambda v\|_2\leq\varepsilon\). To every vector \(v\) in \(\mathbb{R}^n\) we associate a probability distribution \({\rm distr}(v)\) on \(\mathbb{R}\) obtained by choosing a uniform random entry from \(v\). If \(\|v\|_2=1\), then the second moment of \({\rm distr}(v)\) is \(1/n\). Thus, to avoid degeneracy in this case, it is more natural to consider \({\rm distr}(\sqrt{n}v)\) whose second moment is \(1\). We will compare probability distributions on \(\mathbb{R}\) using an arbitrary but fixed metrization of the weak convergence of probability measures. Our theorem in a weak form says the following Note that if \({\rm distr}(\sqrt{n}v)\) is close to the degenerate distribution \(N(0,0)\) then most of the \(\ell^2\) weight of \(v\) is concentrated on \(o(n)\) points. Such vectors are called localized. In general, if \(\sigma\) is smaller than, \(1\) then some of the \(\ell^2\) weight is concentrated on \(o(n)\) vertices and the rest is Gaussian. To formulate our main theorem in the strong form we need some more notation. Recall that \(T_d\) denotes the infinite \(d\)-regular tree and \(o\) is a distinguished vertex called \({\it root}\) in \(T_d\). We will denote the vertex set \(V(T_d)\) of \(T_d\) by \(V_d\). For two vertices in a graph we write \(v\sim w\) if they are connected to each other. Let \([n]\) denote the set \(\{1,2,\dots,n\}\) and let \(G\) be a \(d\)-regular graph on the vertex set \([n]\). We denote by \({\rm Hom}^*(T_d,G)\) the set of all covering maps from \(T_d\) to \(G\). In other words \({\rm Hom}^*(T_d,G)\) is the set of maps \(\phi:V_d\rightarrow V(G)\) such that for every vertex \(v\in V_d\) the neighbors of \(v\) are mapped bijectively to the neighbors of \(\phi(v)\). The set \({\rm Hom}^*(T_d,G)\) has a natural probability measure. We first choose the image of \(o\) uniformly at random in \(V(G)\). Then we recursively extend the map \(\phi\) to larger and larger neighborhoods of \(o\) in a random (conditionally independent) way preserving the local bijectivity. It is easy to see that this probability distribution is independent from the choice of \(o\). Let \(X\) be a topological space and \(f:[n]\rightarrow X\) be a function. We define the probability distribution \({\rm distr}^*(f,G)\) on \(X^{V_d}\) as the distribution \(f\circ \phi\) where \(\phi\) is a random covering in \({\rm Hom}^*(T_d,G)\). In other words \({\rm distr}^*(f,G)\) is a random lift of \(f\) to \(T_d\) using a random covering of \(G\) with \(T_d\). By regarding vectors \(v\in\mathbb{R}^n\) as functions form \([n]\) to \(\mathbb{R}\) it makes sense to use \({\rm distr}^*(v,G)\). To formulate our main theorem we need the concepts of eigenvector processes and Gaussian waves on \(T_d\) (see ). An *eigenvector process* with eigenvalue \(\lambda\) is a joint distribution \(\{X_v\}_{v\in V_d}\) of real valued random variables with variance 1 such that it is \({\rm Aut}(T_d)\) invariant and satisfies the eigenvector equation \[\label{eq:sv}\sum_{v\sim o} X_v=\lambda X_o\] with probability \(1\). Note that the group invariance implies that the eigenvector equation is satisfied at every vertex on \(T_d\). We call an eigenvector process *trivial* if \(\mathbb E(X_o)\neq 0\). Notice that the triviality of an eigenvector process implies that \(\lambda=d\). This follows by taking expectation in [\[eq:sv\]](#eq:sv){reference-type="eqref" reference="eq:sv"} and using the invariance of the process. Furthermore, trivial eigenvector processes are constants in the sense that \(X_v=X_w\) holds with probability one for every pair of vertices \(v,w\) in \(V_d\). A *Gaussian wave* is an eigenvector process whose joint distribution is Gaussian. It is proved in that for every \(-d\leq\lambda\leq d\) there is a unique Gaussian wave \(\Psi_\lambda\) with eigenvalue \(\lambda\). Let us choose a fix metrization of the weak topology on \(\mathbb{R}^{V_d}\). Our main theorem on random \(d\)-regular graphs is the following. # Preliminaries {#prelim} ### Invariant processes {#invariant-processes .unnumbered} For a separable metric space \(Y\) we denote by \(I_d(Y)\) the set of Borel probability measures on \(Y^{V_d}\) that are invariant under the automorphisms of the tree. More precisely, for every \(\tau\in \mathrm{Aut}(T_d)\) (not necessarily fixing the root), the probability measure on \(Y^{V_d}\) is required to be invariant under the natural \(Y^{V_d}\rightarrow Y^{V_d}\) map induced by \(\tau\). Note that eigenvector processes are in \(I_d(\mathbb{R})\). If \(\mu\in I_d(Y)\) and \(F\subseteq V_d\), then we denote by \(\mu_F\) the marginal distribution of \(\mu\) at \(F\). We can equivalently think of \(\mu\in I_d(Y)\) as a joint distribution \(\{X_v\}_{v\in V_d}\) of \(Y\)-valued random variables that is invariant under the automorphism group of \(T_d\). In this language \(\mu_F\) is the same as the joint distribution \(\{X_v\}_{v\in F}\). If both \(F\) and \(Y\) are finite then \(\mu_F\) is a probability distribution on the finite set \(Y^F\). In this case we denote the entropy of \(\mu_F\) by \(\mathbb{H}(F)\). By invariance of \(\mu\) the quantity \(\mathbb{H}(F)\) depends only on the isomorphism class of \(F\). We will consider convergence in \(I_d(Y)\) with respect to the weak topology. Since the weak topology is metrizable we can always choose a fixed metrization of it in advance. ### Almost eigenvectors {#almost-eigenvectors .unnumbered} Let \(A\in \mathbb R^{n\times n}\) be a matrix. An \(\varepsilon\)-almost eigenvector of \(A\) (with eigenvalue \(\lambda\)) is a vector \(v\in \mathbb R^n\) such that \(\|v\|_2=1\) and \(\|Av-\lambda v\|_2\leq \varepsilon\). As we mentioned in the introduction, we will give examples for both localized and delocalized almost eigenvectors on essentially large girth \(d\)-regular graphs. (Note that a graph is called essentially large girth if most vertices are not contained in short cycles.) The purpose of these examples is to show that our results on the almost eigenvectors of random regular graphs are best possible in the sense that all \(0\leq \sigma \leq 1\) can indeed occur in the statement. We will need some preparation. For \(k\geq 1\) and \(x\in [-1,1]\) let \[\label{eq:fkx}f(k,x)=\frac{1}{\sqrt{d(d-1)^{k-1}}}q_k(x),\] where \[\label{eq:qkx}q_k(x)=\sqrt{\frac{d-1}{d}}U_k(x)-\frac{1}{\sqrt{d(d-1)}}U_{k-2}(x); \qquad U_k(\cos \vartheta)=\frac{\sin((k+1)\vartheta)}{\sin \vartheta}.\] (\(U_k(x)\) is the Chebyshev polynomial of the second kind.) Let \(g_{\lambda}: V_d\rightarrow \mathbb R\) be the function defined by \[g_{\lambda}(v)=f(|v|, \lambda/(2\sqrt{d-1})),\] where \(|v|\) denotes the distance of \(v\) and the root \(o\). It is easy to see (and well known) that \(g_{\lambda}\) satisfies the eigenvector equation with eigenvalue \(\lambda\) at every vertex \(v\) (for \(\lambda\in [-2\sqrt{d-1}, 2\sqrt{d-1}]\)). Now let \(G\) be a \(d\)-regular graph such that there is a vertex \(w\in V(G)\) with the property that the shortest cycle containing \(w\) has length at least \(2n\). We define the function \(g'_{\lambda}: V(G)\rightarrow \mathbb R\) by \(g'_{\lambda}(v) =f(d(w,v), \lambda/{2\sqrt{d-1}})\) if \(d(v,w)<n\) and \(0\) otherwise. It is easy to see that \(u=g'_{\lambda}/\|g'_{\lambda}\|_2\) is an almost eigenvector with eigenvalue \(\lambda\) with error tending to \(0\) as \(n\rightarrow \infty\). Furthermore, if \(|V(G)|\) is much larger than \(n\), then \(u\) is close to the constant \(0\) distribution in the weak topology. Thus we obtain examples for completely localized almost eigenvectors for \(d\)-regular graphs (for all \(\lambda\in [-2\sqrt{d-1}, 2\sqrt{d-1}]\)), which corresponds to the case \(\sigma=0\). We switch to the delocalized example. In, the authors construct eigenvector processes on \(T_d\) for every \(\lambda\in [-2\sqrt{d-1}, 2\sqrt{d-1}]\) that are weak limits of factor of i.i.d. processes. These processes have the property that they can be arbitrarily well approximated on any essential large girth \(d\)-regular graph. It is easy to see that these approximations are almost eigenvectors that are completely delocalized. This corresponds to the case \(\sigma=1\). Finally, every \(\sigma\) occurs by mixing completely localized and completely delocalized almost eigenvectors corresponding to the same \(\lambda\). # Eigenvectors and eigenvector processes on the tree {#chap:eigen} For a vertex set \(F\subseteq V_d\) we denote by \(B_k(F)\) the neighborhood of radius \(k\) around \(F\). Let \(F\subseteq V_d\) be a subset of the vertices of the tree, and let \(f\in \mathbb R^{F}\). We say that \(v\) satisfies the eigenvector equation with eigenvalue \(\lambda\) if for every \(v\in F\) with \(B_1(v)\subseteq F\) we have that \(\lambda f(v)=\sum_{w\sim v} f(w)\). It is clear that for a fixed \(\lambda\) these vectors form a linear subspace of \(\mathbb R^F\) that we denote by \(W_{\lambda}(F)\). We will need a formula for \({\rm dim}\, W_{\lambda}(F)\) for a family of special finite sets \(F\). Given \(F\), we say that \(F_0\subseteq F\) is a basis if for all \(f\in \mathbb R^{F_0}\) and \(\lambda \in \mathbb R\) the subspace \(W_{\lambda}(F)\) contains exactly one extension of \(f\) to \(F\). It is clear that if \(F_0\) is a basis in \(F\), then \({\rm dim}\, W_{\lambda}(F)=|F_0|\) for all \(\lambda\in \mathbb R\). \(\square\) We will use \(C\) to denote the star \(B_1(o)\) and we will use \(e\) to denote a distinguished edge in \(T_d\). It is clear that if \(v\) is a neighbor of \(o\), then \(C\setminus \{v\}\) is a basis of \(C\). Similarly \(e\) is a basis of itself. Using Lemma [\[lem:comb\]](#lem:comb){reference-type="ref" reference="lem:comb"} and induction, we obtain that \[\label{eq:dim} {\rm dim}\, W_{\lambda}(B_k(C))=|\partial B_k(C)|=d(d-1)^{k}; \qquad {\rm dim}\, W_{\lambda}(B_k(e))=|\partial B_k(e)|=2(d-1)^k\] holds for every \(\lambda \in \mathbb R\) and \(k \in \mathbb N\), where \(\partial F\) denotes the boundary of a set \(F\). Recall that an invariant process \(\{X_v\}_{v\in V_d}\) is an eigenvector process with eigenvalue \(\lambda\) if it satisfies the eigenvector equation [\[eq:sv\]](#eq:sv){reference-type="eqref" reference="eq:sv"} with probability \(1\) and \(\mathrm{Var}(X_v)=1\). Notice that if \(F\) is any vertex set in \(T_d\), then \(\{X_v\}_{v\in F}\) is supported on \(W_{\lambda}(F)\). In the rest of this chapter we investigate the joint distribution \(\{X_v\}_{v\in S}\) in a non-trivial eigenvector process where \(S\) is one of \(B_k(C)\) or \(B_k(e)\) for some \(k\in\mathbb{N}\). Our goal is to find uncorrelated linear combinations of the variables \(\{X_v\}_{v\in S}\) with the property that they linearly generate every random variable in \(\{X_v\}_{v\in S}\). Observe that since the covariance matrix of \(\{X_v\}_{v\in S}\) has rank at most \({\rm dim}(W_\lambda(S))=|\partial S|\), it is enough to find that many uncorrelated linear combinations with nontrivial variance to guarantee that they generate everything. Let \(Y=(Y_1, Y_2,\dots,Y_n)\) be an \(n\)-dimensional distribution with \(0\) mean such that the covariance matrix is of full rank. We can always find a linear transformation \(T:\mathbb{R}^n\rightarrow\mathbb{R}^n\) with the property that the covariance matrix of \(TY\) is the identity matrix. The probability distribution \(TY\) is unique up to an orthogonal transformation on \(\mathbb{R}^n\). We call it the standardized version of \(Y\). Let \(\mu\in I_d(\mathbb{R})\) represented by an invariant joint distribution \(\{X_v\}_{v\in V_d}\) of random variables. Assume that \(\mathbb{E}(X_o)=0\). With every directed edge \((v,w)\) of \(T_d\) we associate a probability distribution on \(\mathbb{R}^{d-2}\) defined up to an orthogonal transformation. Let \(\{v_i\}_{i=1}^{d-1}\) be the set of neighbors of \(w\) different from \(v\). We denote by \(A_{v,w}\) the standardized version of \(\{X_{v_i}-X_{v_{d-1}}\}_{i=1}^{d-2}\). It is easy to see that (the orthogonal equivalence class of) \(A_{v,w}\) does not depend on the labeling of the vectors \(v_i\). We introduce the symmetric relation \(r\) on directed edges of \(T_d\) such that \(((x,y),(v,w))\in r\) if and only if the unique shortest path connecting \(y\) and \(w\) contains at least one of \(x\) and \(v\). The next lemma implies that if \(((x,y),(v,w))\in r\) then \(A_{x,y}\) and \(A_{v,w}\) are uncorrelated and so if \(\mu\) is Gaussian then \(A_{x,y}\) and \(A_{v,w}\) are independent. Let \(S\subset V_d\) be either \(B_k(C)\) or \(B_k(e)\) for some \(k\in\mathbb{N}\). Assume that \(\mu\) is an eigenvector process. Let \(p\in S\) be such that it has distance at least one from the boundary \(\partial S\). Let \(D\) denote the set of directed edges \((v,w)\) inside \(S\) with the following three properties: \((a)\) the unique shortest path connecting \(p\) and \(w\) contains \(v\), \((b)\) \(p\notin\{v,w\}\), \((c)\) \(B_1(w)\subset S\). Let \(B_p\) denote the joint distribution \((X_v-X_p)_{v\sim p}\). We denote by \(\mathcal{Q}(S,p,\mu)\) the joint distribution of the random variables \(\{A_{v,w}\}_{(v,w)\in D}\) and \(B_p\). Lemma [\[arcuncorr\]](#arcuncorr){reference-type="ref" reference="arcuncorr"} implies that the components of \(\mathcal{Q}(S,p,\mu)\) are uncorrelated multidimensional random variables. It is easy to see that if \(|\lambda|<d\), then the correlation matrix of each such multidimensional random variable is of full rank. This implies that if \(|\lambda|<d\), then \(\mathcal{Q}(S, p, \mu)\) provides a linear basis for \(\{X_v\}_{v\in S}\). By counting dimensions, this yields the following corollary. # Typical processes and the limiting form of the main theorem {#typpro} Now we describe a limiting form of our main theorem using typical processes on \(T_d\). Typical processes on \(T_d\) were first introduced in to study the properties of random regular graphs via ergodic theory. Here we use a slightly different definition which extends the original notion to processes that take values in a separable, metrizable space \(Y\). Our main theorem in the limit setting is the following. Using the fact that weak limits of factor of i.i.d processes are typical (see ) we obtain the next corollary which answers a question of B. Virág. Note that Corollary [\[mainlimcorr\]](#mainlimcorr){reference-type="ref" reference="mainlimcorr"} implies that if many eigenvector processes corresponding to the same eigenvalue are coupled in a way that the coupling is a weak limit of factor of i.i.d. processes, then its distribution is jointly Gaussian. We emphasize that the first part of the statement in Theorem [\[mainlim\]](#mainlim){reference-type="ref" reference="mainlim"}, namely that \(|\lambda|\leq 2\sqrt{d-1}\) is a consequence of Friedman's theorem. We prove this implication in Lemma [\[lem:equiv\]](#lem:equiv){reference-type="ref" reference="lem:equiv"}. In addition, the main goal of this chapter is to show that Theorem [\[mainlim\]](#mainlim){reference-type="ref" reference="mainlim"} implies Theorem [\[main\]](#main){reference-type="ref" reference="main"}. We need a few notions and lemmas. Let \(\mathcal{P}\) denote the set of Borel probability distributions \(\mu\) on \(\mathbb{R}^{V_d}\) which have a second moment bounded from above by \(1\) at each coordinate. By tightness of \(\mathcal P\), we have that \(\mathcal{P}\) is compact with respect to the weak topology of measures. Let \(m\) be a fixed metrization of the weak topology on \(\mathcal{P}\). Let us denote by \(\mathcal{T}\) the set of closed subsets in \(\mathcal{P}\), and by \(d_H\) the the Hausdorff metric on \(\mathcal{T}\). We have that \(d_H\) induces a compact topology on \(\mathcal{T}\). Let us define the distance \(m^*\) for \(d\)-regular graphs in the following way. If \(G_1\) and \(G_2\) are \(d\)-regular graphs then \(m^*(G_1,G_2)\) is the infimum of the numbers \(\delta\) with the property the if \(f_1:V(G_1)\rightarrow \mathbb{R},f_2:V(G_2)\rightarrow \mathbb{R}\) are arbitrary functions with \(\mathbb{E}({\rm distr}(f_1^2)),\mathbb{E}({\rm distr}(f_2^2))\leq 1\) then there are functions \(f_1':V(G_2)\rightarrow\mathbb{R}, f_2':V(G_1)\rightarrow \mathbb{R}\) with \(\mathbb{E}({\rm distr}(f_1'^2)),\mathbb{E}({\rm distr}(f_2'^2))\leq 1\) such that \[m({\rm distr}^*(f_1,G_1),{\rm distr}^*(f_1',G_2))\leq\delta ~~~,~~~m({\rm distr}^*(f_2,G_2),{\rm distr}^*(f_2',G_1))\leq\delta.\] We can describe the metric \(m^*\) in terms of the metric \(d_H\) as follows. For a graph \(G\), let \[S(G)=\{\mathrm{distr}^*(f, G)|f\in \mathbb{R}^{V(G)},\mathbb{E}(\mathbb{\rm distr}(f^2))\leq 1\}.\] We have that \(m^*(G_1, G_2)=d_H(S(G_1), S(G_2))\). \(\square\) \(\square\) Now we enter the proof of Lemma [\[lem:equiv\]](#lem:equiv){reference-type="ref" reference="lem:equiv"}. \(\square\) # Entropy inequality for typical processes {#chap:entr} Let \(X\) be a separable metric space; let \(F\) be a finite set and let \(\mathcal{P}(F)\) denote the set of probability distributions on \(F\) equipped with the topology generated by total variation distance. A continuous discretization of \(X\) is a continuous function \(\phi:X\rightarrow\mathcal{P}(F)\). If \(\alpha\in X^V\) is an \(X\)-coloring of a finite or countable set \(V\), then we denote by \(\phi*\alpha\) the probability distribution on \(F^V\) obtained by independently choosing an element from \(F\) for each \(v\in V\) with distribution \(\phi(\alpha(v))\). If \(\mu\) is a probability distribution on \(X^V\), then we denote by \(\phi*\mu\) the probability distribution obtained by first taking a \(\mu\) random element \(\alpha:V\rightarrow X\) and then in a second round of randomization we take \(\phi*\alpha\). The main result of this chapter is the next entropy inequality for typical processes. Before proving Theorem [\[discmain\]](#discmain){reference-type="ref" reference="discmain"} we need some preparation. Let us fix a metrization of the weak topology on \(I_d(X)\). For a finite \(d\)-regular graph \(G\) let \(\omega(G)\) denote the infimum of the numbers \(\varepsilon>0\) for which it is true that at least \(1-\varepsilon\) fraction of the vertices of \(G\) are not contained in a cycle of length at most \(\lfloor 1/\varepsilon\rfloor\). The quantity \(\omega(G)\) measures how similar the graph \(G\) is to the tree \(T_d\) in the Benjamini--Schramm metric. Throughout this chapter, \(G\) is always assumed to be finite. The operator \({\rm distr}^*\) maps \(X\)-colored \(d\)-regular graphs \((\alpha\in X^{V(G)},G)\) to invariant processes in \(I_d(X)\). Using this correspondence and the metric on \(I_d(X)\) we define the distance of an \(X\)-colored graph \((\alpha,G)\) and a process \(\mu\in I_d(X)\) as \(\omega(G)\) plus the distance of \({\rm distr}^*(\alpha,G)\) and \(\mu\) in \(I_d\). Note that if \((\alpha\in X^{V(G)},G)\) is an \(X\)-colored \(d\)-regular graph, then \((\phi*\alpha,G)\) is a probability distribution on \(F\)-valued colorings of the vertices of \(G\). In this case \({\rm distr}^*(\phi*\alpha,G)\) is a probability distribution on \(I_d(X)\), while \(\phi*{\rm distr}^*(\alpha,G)\) is a single element in \(I_d(X)\). The proof of Proposition [\[discapprox\]](#discapprox){reference-type="ref" reference="discapprox"} relies on the next technical lemma. We continue with the proof of Proposition [\[discapprox\]](#discapprox){reference-type="ref" reference="discapprox"}. \(\square\) For the next lemma let \(R(d,n)\) denote the number of \(d\)-regular graphs on the vertex set \([n]\). In case \(d\) is odd we will always assume that the number of vertices is even. Now we arrived to the proof of Theorem [\[discmain\]](#discmain){reference-type="ref" reference="discmain"}. # Smooth eigenvector processes {#smooth} Let \(\mu\) be an eigenvector process. If \(F\subseteq V_d\) is a finite set then the distribution of \(\mu\), when restricted to \(F\), is concentrated on the subspace \(W_{\lambda}(F)\) (recall Chapter [4](#chap:eigen){reference-type="ref" reference="chap:eigen"}). We denote by \(\mathbb D_{\rm sp}(F,\mu)\) the differential entropy of \(\mu_F\) measured inside this subspace using the Euclidean structure inherited from \(\mathbb{R}^F\). We say that \(\mu\) is *smooth* if \(\mathbb{D}_{\rm sp}(B_k(C))\) and \(\mathbb{D}_{\rm sp}(B_k(e))\) are finite for every \(k\). In this chapter we reduce Theorem [\[mainlim\]](#mainlim){reference-type="ref" reference="mainlim"} to smooth eigenvector processes. The reduction will rely on the following statement. Assume that Theorem [\[mainlim\]](#mainlim){reference-type="ref" reference="mainlim"} holds for smooth typical eigenvector processes. Using Proposition [\[prop:masodik\]](#prop:masodik){reference-type="ref" reference="prop:masodik"} we obtain that if \(\{X_v\}_{v\in V_d}\) is an arbitrary typical eigenvector process then \(\{X_v+aY_v\}_{v\in V_d}\) is smooth for all \(a>0\). In addition, by Proposition [\[prop:elso\]](#prop:elso){reference-type="ref" reference="prop:elso"} and Proposition [\[prop:hv\]](#prop:hv){reference-type="ref" reference="prop:hv"}, \(\{X_v+aY_v\}_{v\in V_d}\) is typical. By our assumption, we obtain the Gaussianity of \(\{X_v+aY_v\}_{v\in V_d}\) for all \(a>0\). This implies the Gaussianity of \(\{X_v\}_{v\in V_d}\) by going to \(0\) with \(a\). # Entropy Inequality for typical eigenvector processes {#chap:entrtyp} In this chapter we prove the next theorem. To prove the above theorem we will need some preparation. For \(a\in\mathbb{N}\) let us define the continuous discretization \(t_{0,a}\) of \(\mathbb{R}\) in the following way. If \(x>a\) (resp. \(x<-a\)), then \(t_{0,a}(x)=a\) (resp. \(t_{0,a}(x)=-a\)) with probability \(1\). Otherwise let \(t_{0,a}(x)\) denote the probability distribution that takes \(\lfloor x*a\rfloor/a\) with probability \(1+\lfloor x*a\rfloor-x*a\) and takes \(1/a+\lfloor x*a\rfloor/a\) with probability \(x*a-\lfloor x*a\rfloor\). For \(\sigma>0\) we define the discretization \(t_{\sigma,a}\) by \(t_{\sigma,a}(x)=t_{0,a}(x+\sigma N)\) where \(N\) is a random variable with standard normal distribution. We denote by \(t_{\sigma,a}^n\) the continuous discretization of \(\mathbb{R}^n\) obtained by the coordinatewise independent application of \(t_{\sigma,a}\). The main difficulty of the proof of Lemma [\[finombecsles\]](#finombecsles){reference-type="ref" reference="finombecsles"} comes from the fact that the support of \(X\) is not necessarily compact. We have to treat a situation where we refine and increase the interval of discretization simultaneously. In the following lemma, first we calculate the entropy of the random variable \(t_{\sigma, a}(x)\) for every fixed \(x\), and then we take the expectation of this quantity with respect to the distribution of a random variable \(X\). We finish this chapter with the proof of Theorem [\[mainineq\]](#mainineq){reference-type="ref" reference="mainineq"}. Let \(\mu_{\sigma,a}\) denote the process in which we pointwise discretize \(\mu\) using \(t_{\sigma,a}\) (using the notation of Chapter [5](#typpro){reference-type="ref" reference="typpro"}, we have \(\mu_{\sigma, a}=t_{\sigma, a}*\mu\)) and let \(\mu_\sigma\) denote the process obtained from \(\mu\) by adding \(\sigma\) times the i.i.d normal distribution. By Lemma [\[finombecsles\]](#finombecsles){reference-type="ref" reference="finombecsles"} and \(|B_k(C)|-(d/2)|B_k(e)|=1\) we obtain that \[\mathbb{H}(B_k(C),\mu_{\sigma,a})-\frac{d}{2}\mathbb{H}(B_k(e),\mu_{\sigma,a})=\log a + \mathbb D(B_k(C),\mu_\sigma)-\frac{d}{2}\mathbb D(B_k(e),\mu_\sigma)+o(1).\] By Theorem [\[discmain\]](#discmain){reference-type="ref" reference="discmain"} and the typicality of \(\mu\) we get that \[\mathbb{H}(B_k(C),\mu_{\sigma,a})-\frac{d}{2}\mathbb{H}(B_k(e),\mu_{\sigma,a})\geq \mathbb{E}_{\mu_o}(\mathbb{H}(t_{\sigma,a}(x))).\] Using the previous formulas, Lemma [\[varhato\]](#varhato){reference-type="ref" reference="varhato"} and the limit \(a\rightarrow\infty\) we obtain that \[\label{eq:eht}\mathbb D(B_k(C),\mu_\sigma)-\frac{d}{2} \mathbb D(B_k(e),\mu_\sigma)\geq D(N(0,\sigma)\] for every \(\sigma>0\). Let \(S\subset V_d\) be either \(B_k(C)\) or \(B_k(e)\). We denote by \(\mu_{S, \sigma}\) the probability measure obtained by convolving the measure \(\mu_S\) with the standard normal distribution on \(W_{\lambda}(S)\) (for the definition of \(W_\lambda(S)\) see chapter [4](#chap:eigen){reference-type="ref" reference="chap:eigen"}), where \(\lambda\) is the eigenvalue corresponding to \(\mu\). Observe that the standard normal distribution on \(\mathbb R^S\) is the independent sum of the standard normal distribution on \(W_{\lambda}(S)\) and on \(W_{\lambda}(S)^{\perp}\). Then by using that \(\mathrm{dim}\, W_{\lambda}(S)^{\perp}=|S|-|\partial S|\), we have \[\mathbb D(S, \mu_{\sigma})=\mathbb D_{\rm sp}(\mu_{S, \sigma})+(|S|-|\partial S|)\mathbb D(N(0,\sigma)).\] Using this formula for \(S=B_k(C)\) and \(S=B_k(e)\) in [\[eq:eht\]](#eq:eht){reference-type="eqref" reference="eq:eht"} together with \(|\partial B_k(C)|=(d/2)|\partial B_k(e)|\), we obtain that \[\mathbb D_{\rm sp}(\mu_{B_k(C), \sigma})-\frac{d}{2}\mathbb D_{\rm sp}(\mu_{B_k(e), \sigma})\geq 0.\] If \(\sigma\rightarrow 0\), then we obtain the statement of Theorem [\[mainineq\]](#mainineq){reference-type="ref" reference="mainineq"}. \(\square\) # Eigenvalues of the covariance operator {#chap:cov} The main goal of this section is to prove the following statement on the differential entropies of Gaussian waves. Recall the definition of \(\mathbb D_{\rm sp}\) from Chapter [7](#smooth){reference-type="ref" reference="smooth"}. In addition, by \(\mathrm{det}_{\rm sp}\) we mean the product of the non-zero eigenvalues of a matrix. Let \(\Sigma_k\) be the covariance matrix of the joint distribution of \(\Psi_{\lambda}\) restricted to the ball \(B_k(C)\), and \(\Sigma'_k\) be the similar covariance matrix on \(B_k(e)\). The differential entropy of a multivariate normal random variable with covariance matrix \(\Sigma\) of rank \(m\) is given by \(\frac 12 \log \big((2\pi e)^m \det_{\rm sp} \Sigma)\big)\) if we measure differential entropy inside the support of the variable. Equation [\[eq:dim\]](#eq:dim){reference-type="eqref" reference="eq:dim"} and Corollary [\[cor:qsp\]](#cor:qsp){reference-type="ref" reference="cor:qsp"} imply that the rank of \(\Sigma_k\) is \(d/2\) times the rank of \(\Sigma_k'\). Hence we need to prove that \[\label{eq:logdet}\log \det\!_{\rm sp}\, \Sigma_k-\frac{d}{2}\log \det\!_{\rm sp} \, \Sigma_k'\rightarrow 0 \qquad (k\rightarrow\infty).\] Notice that if \(s\) is an eigenvalue of both \(\Sigma_k\) and \(\Sigma_k'\), and its multiplicity in the first case is \(d/2\) times its multiplicity in the second case, then it is canceled out in the difference. In order to find the eigenvalues that do not cancel out, we decompose both \(\mathbb R^{|B_k(C)|}\) and \(\mathbb R^{|B_k(e)|}\) as a union of orthogonal subspaces that are invariant under the corresponding covariance operators. First we need some notation. We will use the genealogical labeling of the vertices in \(B_k(C)\) and in \(B_k(e)\) (in this section, we will not distinguish vertices and labels). In \(B_k(C)\), the root gets label \(\emptyset\), and we put the labels on the vertices such that the labels of neighbors differ only in the last coordinate (i.e. \(1, 2, \ldots, d\) are the neighbors of the root; \(11, 12, \ldots, 1(d-1)\) are the further neighbors of \(1\), and so on). For a vertex \(v\) the length of its label is denoted by \(|v|\), which is its distance from \(\emptyset\). For a vertex \(v\) and a sequence \(y\), by \(vy\) we mean the vertex with the label obtained by concatenating \(v\) and \(y\). We say that \(v\) is an ancestor of \(w\) (denoted by \(v\rightarrow w\)), if \(w=vy\) for some \(y\neq \emptyset\). As for \(B_k(e)\), we use a similar notation, but keeping track of symmetry with respect to the central edge \(e\). The endpoints of \(e\) have labels \(\emptyset\) and \(\emptyset'\). The descendants of \(\emptyset\) have labels \(1, \ldots, d-1\), their descendants have labels \(11, 12, \ldots, 1(d-1), 21, \ldots\) and so on. Similarly, the descendants of \(\emptyset'\) have labels \(1', \ldots, (d-1)'\), their descendants have labels \(11', 12', \ldots, 1(d-1)', 21', \ldots\) and so on. Note that \(|v|\) still denotes the length of the label. We assign a linear subspace to each vertex in \(B_k(C)\setminus \partial B_k(C)\) and \(B_k(e)\setminus \partial B_k(e)\). Fix \(v\in B_k(C)\setminus \partial B_k(C)\). Let \(\mathcal E_v\) be the elements \(\alpha\in\mathbb R^{B_k(C)}\) for which the following hold. \((i)\) \(\alpha_w=0\) if \(v\) is not an ancestor of \(w\). \((ii)\) Suppose that \(1\leq j\leq (d-1)\) and \(y, z\) are labels with \(|y|=|z|\). Then \(\alpha_{vjy}=\alpha_{vjz}\). (To put it in another way, for descendants of \(v\), the value of \(\alpha\) depends only on the first coordinate after \(v\) and the distance from \(v\).) \((iii)\) We have \(\sum_{y: |y|=r} \alpha_{vy}=0\) for \(r\geq 1\). In addition, we introduce the following subspace: \[\mathcal G=\{\alpha\in\mathbb R^{B_k(C)}: \alpha_v=\alpha_w \text{ if } |v|=|w|\}.\] We will also refer to \(\mathcal E_v'\) (when \(v\in B_k(e)\setminus \partial B_k(e)\)), which are linear subspaces of \(\mathbb R^{B_k(e)}\) defined similarly. The definition of the complement subspace is somewhat different: \[\mathcal G_1'=\{\alpha\in\mathbb R^{B_k(e)}: \alpha_v=\alpha_w \text{ if } |v|=|w|\}.\] \[\mathcal G_2'=\{\alpha\in\mathbb R^{B_k(e)}: \alpha_v=\alpha_w \text{ if } \emptyset \rightarrow v, \emptyset \rightarrow w, |v|=|w|;\ \alpha_{v'}=-\alpha_v \text{ if } v=\emptyset \text{ or } \emptyset \rightarrow v\}.\] \((b)\) Fix \(v_1, v_2\in B_k(C)\). If none of them is ancestor of the other one, then the support of any vector of \(\mathcal E_{v_1}\) is disjoint from the support of any vector in \(\mathcal E_{v_2}\), which implies orthogonality. If \(v_1\rightarrow v_2\), then the value of a vector in \(\mathcal E_{v_1}\) is the same at all vertices of type \(v_2y\) with \(|y|\) fixed. Multiplying this by the values of a vector in \(\mathcal E_{v_2}\) and summing this up for different \(y\)s (of fixed length) we get \(0\), because of property \((iii)\). This implies the orthogonality. The other cases are similar; we omit the details. \((c)\) The dimensions of these subspaces are as follows. \[\mathrm{dim}\ \mathcal E_{v}=(k-|v|+1)(d-2) \qquad \text{for } \emptyset \neq v\in B_k(C)\setminus \partial B_k(C);\] \[\mathrm{dim}\ \mathcal E_{\emptyset}=(k+1)(d-1); \qquad {\rm dim}\ \mathcal G=k+2;\] \[\mathrm{dim}\ \mathcal E'_{v}=(k-|v|+1)(d-2) \qquad \text{for } \emptyset \neq v\in B_k(e)\setminus \partial B_k(e);\] \[\mathrm{dim}\ \mathcal G'_{1}=k+1; \qquad {\rm dim}\ \mathcal G_2'=k+1.\] The following equalities are easy to check by induction on \(k\): \[1+(k+1)d+\sum_{j=1}^{k} d(d-1)^{j-1} (k+1-j)(d-2)=|B_k(C)|;\] \[2(k+1)+2\sum_{j=1}^{k-1} (d-1)^{j-1} (k-j)(d-2)=|B_k(e)|.\] Hence the sum of the dimension of the linear subspaces \(\mathcal E_v\) and \(\mathcal G\) is equal to the dimension of the space \(\mathbb R^{B_k(C)}\). Since the subspaces are pairwise orthogonal by part \((b)\) of the lemma, this implies that the sum must be equal to \(\mathbb R^{B_k(C)}\). A similar argument works for \(B_k(e)\). \(\square\) For the following lemma, recall the definition of \(f(k, x)\) from equation [\[eq:fkx\]](#eq:fkx){reference-type="eqref" reference="eq:fkx"}. Furthermore, the calculation about this recurrence relation in imply that if we take \(\lambda=2\sqrt{d-1}x\), then the covariance of the values at distance \(k\) in the Gaussian wave \(\Psi_{\lambda}\) is equal to \(f(k,x)\). \(\square\) As we have discussed, it is sufficient to show that [\[eq:logdet\]](#eq:logdet){reference-type="eqref" reference="eq:logdet"} holds. First fix \(k\), and recall Lemma [\[lem:ev\]](#lem:ev){reference-type="ref" reference="lem:ev"}. Notice that for every \(1\leq r\leq k-1\), if we take two vertices \(v_1, v_2\) in \(B_k(C)\) such that \(|v_1|=|v_2|=r\), then the linear transformation \(\Sigma_k\) restricted to \(\mathcal E_{v_1}\) is isomorphic to the linear transformation \(\Sigma_k\) restricted to \(\mathcal E_{v_2}\) (we use again that the entries of the covariance matrix depend only on the distance of the vertices). Hence the set of eigenvalues of \(\Sigma_k\) corresponding to the invariant subspaces \(\mathcal E_{v_1}\) and \(\mathcal E_{v_2}\) are the same. Furthermore, this linear transformation is isomorphic to the linear transformation \(\Sigma_k'\) restricted to \(\mathcal E_{v}'\), if \(|v|=r\) holds. For every \(1\leq r\leq k-1\) we have \[\big|\{v: |v|=r,\, v\in B_k(C)\}\big|=\frac d2 \big|\{v: |v|=r,\, v\in B_k(e)\}\big|.\] This means that all eigenvalues belonging to the spaces \(\mathcal E_v\) and \(\mathcal E'_v\) cancel out in the expression in [\[eq:logdet\]](#eq:logdet){reference-type="eqref" reference="eq:logdet"} for \(|v|>0\). Therefore only the eigenvalues corresponding to \(\mathcal G, \mathcal E_{\emptyset}', \mathcal G_1'\) and \(\mathcal G_2'\) are left. These are calculated with multiplicities in Lemma [\[lem:s1\]](#lem:s1){reference-type="ref" reference="lem:s1"}, and hence Lemma [\[lem:kul\]](#lem:kul){reference-type="ref" reference="lem:kul"} finishes the proof by showing that the difference goes to zero as \(k\rightarrow \infty\). \(\square\) # Improved differential entropy inequality {#chap:impr} In this chapter we use a combination of Theorem [\[mainineq\]](#mainineq){reference-type="ref" reference="mainineq"} and Theorem [\[thm:nulla\]](#thm:nulla){reference-type="ref" reference="thm:nulla"} to prove an improved version of Theorem [\[mainineq\]](#mainineq){reference-type="ref" reference="mainineq"} for the case \(k=0\). We will use the notation from Chapter [4](#chap:eigen){reference-type="ref" reference="chap:eigen"}. If \(S\) is either \(B_k(C)\) or \(B_k(e)\) for some \(k\geq 0\) and \(p\) is in \(S\setminus\partial S\), then we have for the Gaussian wave \(\Psi_\lambda\) with distribution \(\mu\) that \[\label{dfor} \mathbb{D}(\mathcal{Q}(S,p,\mu))=\mathbb{D}(B_p)+\sum_{(v,w)\in D}\mathbb{D}(A_{v,w}).\] We obtain from ([\[dfor\]](#dfor){reference-type="ref" reference="dfor"}) that the differential entropy \(\mathbb{D}(\mathcal{Q}(S,p,\mu))\) does not depend on \(p\) for a Gaussian wave with distribution \(\mu\). Observe that changing the vertex \(p\) to \(p'\) results in a linear transformation \(T_{p,p'}\) in the system \(\mathcal{Q}(S,p,\mu)\). The invariance of the differential entropy shows by Lemma [\[difftrdet\]](#difftrdet){reference-type="ref" reference="difftrdet"} that \(T_{p,p'}\) has determinant \(1\). Now applying Lemma [\[difftrdet\]](#difftrdet){reference-type="ref" reference="difftrdet"} for an arbitrary smooth eigenvector process \(\nu\) with eigenvalue \(\lambda\) we obtain that the value of \(\mathcal{D}(\mathcal{Q}(S,p,\nu))\) is independent of \(p\) since the transformations \(T_{p,p'}\) depend only on the quadruple \(p,p',\lambda,S\) and can be calculated from the eigenvector equation. For a general smooth eigenvector process \(\nu\) we define \(\mathbb{D}(S,\nu)\) as this unique differential entropy of \(\mathbb{D}(\mathcal{Q}(S,p,\nu))\). We will need the next definition. Note that the \(2\)-Markov property implies that the marginal distribution \(Y_C=\{Y_v\}_{v\in C}\) determines the whole process because we can build up the distribution \(\{Y_v\}_{v\in V_d}\) using iterated conditionally independent couplings of \(Y_C\) along edges. More precisely, if for some connected subgraph \(K\) of \(T_d\) the distribution \(\{Y_v\}_{v\in V(K)}\) is already constructed and \(w\in V_d\) is a vertex such that the star \(B_1(w)\) intersects \(K\) in a single edge \(e\), then the joint distribution \(\{Y_v\}_{v\in V(K)\cup B_1(w)}\) is the conditionally independent coupling of \(\{Y_v\}_{v\in V(K)}\) and \(\{Y_v\}_{v\in B_1(w)}\) with respect to \(\{Y_v\}_{v\in e}\). The invariance of the process implies that \(\{Y_v\}_{v\in B_1(w)}\) has the same distribution as \(Y_C\). By iterating this procedure we can build up the marginal distribution on any finite connected subgraph of \(T_d\) and thus the whole process in uniquely determined. Let \(\alpha(S,\nu)\) denote the difference \(\mathbb{D}(\mathcal{Q}(S,p,\nu))-\mathbb{D}(\mathcal{Q}(S,p,\mu))\) where \(\mu\) is the Gaussian eigenvector process with the same eigenvalue as \(\nu\). If we apply the same change of basis to both \(\mathbb{D}(\mathcal{Q}(S,p,\nu))\) and \(\mathbb{D}(\mathcal{Q}(S,p,\mu))\), they change with the same additive constant by lemma [\[difftrdet\]](#difftrdet){reference-type="ref" reference="difftrdet"} and thus \(\alpha(S,\nu)\) remains unchanged. This shows the basis independence of \(\alpha(S,\nu)\). In particular we have that if \(\nu\) is a smooth eigenvector process then \(\alpha(S,\nu)=\mathbb{D}_{\rm sp}(S,\nu)-\mathbb{D}_{\rm sp}(S,\mu)\) and so \[\alpha(B_k(C),\nu)-(d/2)\alpha(B_k(e),\nu)=\] \[\label{imprlem3eq} \Bigl(\mathbb{D}_{\rm sp}(B_k(C),\nu)-(d/2)\mathbb{D}_{\rm sp}(B_k(e),\nu)\Bigr)-\Bigl(\mathbb{D}_{\rm sp}(B_k(C),\mu)-(d/2)\mathbb{D}_{\rm sp}(B_k(e),\mu)\Bigr).\] Using Theorem [\[mainineq\]](#mainineq){reference-type="ref" reference="mainineq"} and Theorem [\[thm:nulla\]](#thm:nulla){reference-type="ref" reference="thm:nulla"} we get the following consequence. The main theorem of this chapter is the following. # Heat equation and the proof of the main theorem {#chap:heat} Notice that the covariance conditions of definition [\[stardist\]](#stardist){reference-type="ref" reference="stardist"} guarantee that \(\mathbb{E}((X_1+X_2+\dots+X_d-\lambda Z)^2)=0\) and thus \(X_1+X_2+\dots+X_d=\lambda Z\) holds with probability \(1\). This implies that the joint distribution \(F\) is concentrated on the \(1\) co-dimensional (\(d\)-dimensional) subspace \(W_\lambda(C)\) in \(\mathbb{R}^{d+1}\). The subspace differential entropy in definition [\[stardist\]](#stardist){reference-type="ref" reference="stardist"} is measured in this subspace. In this chapter we think of \(\lambda\in[-2\sqrt{d-1},2\sqrt{d-1}]\) and \(d\) as fixed values and most of the times our notation will not indicate the dependence on these values even if there is such a dependence. Our goal is to solve the extremal problem of maximizing \(\mathcal{D}\) inside \(\mathcal{F}_{d,\lambda}\) (see Theorem [\[gausscond\]](#gausscond){reference-type="ref" reference="gausscond"}). This will provide the last step in the proof of our main theorems (see Theorem [\[main\]](#main){reference-type="ref" reference="main"} and Theorem [\[mainlim\]](#mainlim){reference-type="ref" reference="mainlim"}) as we will explain at the end of this chapter. It will be important that there is a unique element \(F^*=(X_1^*,X_2^*,\dots,X_d^*,Z^*)\) in \(\mathcal{F}_{d,\lambda}\) such that \(F^*\) is Gaussian. (The covariances define the Gaussian system uniquely, which clearly satisfies the symmetry conditions.) Note that \(F^*\) depends on both \(d\) and \(\lambda\). Most of this chapter deals with the proof of the the next theorem which says that the entropy formula \(\mathcal{D}\) in \(\mathcal{F}_{d,\lambda}\) is maximized by the Gaussian distribution \(F^*\). To get rid of the subspace differential entropy, we apply a change of basis to the systems in \(\mathcal{F}_{d,\lambda}\). We can choose a fix linear transformation \(T:\mathbb{R}^{d+1}\rightarrow \mathbb{R}^d\) (depending on \(d\) and \(\lambda\)) such that for every \(F\in\mathcal{F}_{d,\lambda}\) the system \(T(F)=(B_1,B_2,\dots,B_d)\) satisfies \(\mathbb{E}(B_iB_j)=\delta_{i,j}\) for \(1\leq i,j\leq d\). Using that \(M=T(F^*)\) is Gaussian we obtain that \(M\) is the standard normal distribution on \(\mathbb{R}^d\). Using that linear transformations change differential entropy with a fix constant (depending on the transformation; see Lemma [\[difftrdet\]](#difftrdet){reference-type="ref" reference="difftrdet"}), the statement of the theorem is equivalent to the fact that \[\mathbb D(T(F))-(d/2)\mathbb D(X_1,Z)\leq \mathbb D(M)-(d/2) \mathbb D(X_1^*,Z^*)\] and equality holds if and only if \(T(F)=M\). The proof of Theorem [\[gausscond\]](#gausscond){reference-type="ref" reference="gausscond"} relies on the following proposition. Note that the choice of \(F_t\) comes from the heat equation in \(\mathbb{R}^d\) (see chapter [12](#heatedrandom){reference-type="ref" reference="heatedrandom"}). We first show that Proposition [\[gaussprop\]](#gaussprop){reference-type="ref" reference="gaussprop"} implies Theorem [\[gausscond\]](#gausscond){reference-type="ref" reference="gausscond"}. The joint distribution \(F_t\) does not satisfy the covariance conditions of definition [\[stardist\]](#stardist){reference-type="ref" reference="stardist"} but it is clear that the scaled version \(F_t(1+2t)^{-1/2}\) is in \(\mathcal{F}_{d,\lambda}\). Notice that scaling does not change the differential entropy formula because the extra additive constants coming from scaling exactly cancel each other. By using the claim for \(G=F_t(1+2t)^{-1/2}\) we obtain from \(\Lambda_G'(0)\geq 0\) that \(\Lambda_F'(t)\geq 0\) holds for \(F\) with every \(t\geq 0\). Since \(F_t(1+2t)^{-1/2}=F(1+2t)^{-1/2}+F^*(2t/(1+2t))^{1/2}\) converges to \(F^*\) as \(t\rightarrow \infty\), we obtain that \(F^*\) maximizes \(\mathcal{D}\). To see that only the Gaussian system \(F^*\) attains the maximum assume that \(F\) attains the maximum. In this case \(\Lambda_F'(t)\geq 0\) is only possible if \(\Lambda_F'(t)=0\) holds for every \(t\geq 0\). This implies that \(F\) is Gaussian by the second part of Proposition [\[gaussprop\]](#gaussprop){reference-type="ref" reference="gaussprop"}. It remains to prove Proposition [\[gaussprop\]](#gaussprop){reference-type="ref" reference="gaussprop"}. We start with some notation and lemmas. Assume that the measure \(\mu\) is the distribution and \(f\) is the density function of \(B=T(F)\). We can choose a matrix \(Q\in \mathbb R^{(d+1)\times d}\) such that \(X_i=\sum_k Q_{i,k}B_k\) holds for \(i=1, \ldots, d\), and \(Z=\sum_k Q_{d+1, k} B_k\) is also satisfied. We define \(v_i=(Q_{i, 1}, \ldots, Q_{i, d})\) for \(i=1, \ldots, d\) and \(w=(Q_{d+1, 1}, \ldots, Q_{d+1, d})\). Notice that the covariance matrix of \(F\) is \(QQ^T\). Hence the vectors \(\{v_i\}_{i=1}^d\) and \(w\) are unit vectors such that \((v_i,v_j)=(\lambda^2-d)d^{-1}(d-1)^{-1}\) and \((v_i,w)=\lambda/d\) for every \(1\leq i,j\leq d\). The joint distribution \(((v_1,B),(v_2,B),\dots,(v_d,B),(w,B))\) is the same as \(F\). We have that \(\mu\) is invariant with respect to every orthogonal transformation permuting the system \(\{v_i\}_{i=1}^d\) and fixing \(w\). Furthermore we have that \(\mu\) projected to the space spanned by \(v_i\) and \(w\) is invariant with respect to the reflection interchanging \(v_i\) and \(w\). We can choose two real numbers \(\alpha\) and \(\beta\) such that the vectors \(a_i=\alpha w+\beta v_i\) and \(b_i=\alpha v_i+\beta w\) satisfy \((a_i,b_i)=0\) and \(\|a_i\|_2=\|b_i\|_2=1\) for every \(1\leq i\leq d\) (in particular, we need that \((\alpha^2+\beta^2)\lambda/d+2\alpha \beta=0\), and \((\alpha^2+\beta^2)+2\alpha\beta\lambda/d=1\)). The choice of \(\alpha\) and \(\beta\) is unique up to multiplying both by \(-1\) or switching them. Their values can be determined using elementary geometry. Note that construction of the vector system \(\{v_i\}_{i=1}^d, w, \{a_i\}_{i=1}^d, \{b_i\}_{i=1}^d\) in \(\mathbb{R}^d\) is purely linear algebraic. Such system, with the scalar products given above, can be constructed for an arbitrary \(|\lambda|\leq d\) however in the case of \(|\lambda|\leq2\sqrt{d-1}\) they satisfy a useful geometric property expressed in the following lemma. \(\square\) Now we return to the proof of Proposition [\[gaussprop\]](#gaussprop){reference-type="ref" reference="gaussprop"}. For \(1\leq i\leq d\) let \(f_i\) denote the orthogonal projection of \(f\) to the two dimensional space \(V_i=\langle w,v_i\rangle_{\mathbb{R}}\). This means that \[f_i(x)=\int_{z\in V_i^{\perp}}f(x+z)\] for \(x\in V_i\). Let \(T_2:\mathbb{R}^2\rightarrow\mathbb{R}^2\) denote the linear transformation \(T_2(x,y)=(\alpha x+\beta y, \alpha y+\beta x)\) with \(\alpha\) and \(\beta\) defined above. We have that \(f_i\) (when written in the orthonormal basis \(a_i,b_i\)) is the density function of \(T_2(X_i,Z)\). We can write \(\Lambda_F(t)\) as \[k+\mathbb D(T(F_t))-(d/2)\mathbb D(T_2(X_1+\sqrt{2t}X_1^*,Z+\sqrt{2t}Z^*))\] where the constant \(k\) comes from the change of basis \(T_2\). Then by the de Bruijn identity (see equation [\[diffheat\]](#diffheat){reference-type="eqref" reference="diffheat"} and Lemma [\[korlat\]](#korlat){reference-type="ref" reference="korlat"}) we get \[\Lambda_F'(0)=\int_{\mathbb{R}^d} \|\triangledown f\|_2^2/f-(d/2)\int_{V_1} \|\triangledown f_1\|_2^2/f_1.\] From Lemma [\[wr\]](#wr){reference-type="ref" reference="wr"} we have that \[\|\triangledown f\|_2^2=\sum_{i=1}^d (t_1(\partial_{a_i}f)^2+t_2(\partial_{b_i}f)^2)\] holds for some \(t_1, t_2\geq 0\) such that \(t_1+t_2=1\). Using \[\|\triangledown f_1\|_2^2=(\partial_{a_1}f_1)^2+(\partial_{b_1}f_1)^2\] and the above equations it follows that \[\Lambda_F'(0)=\sum_{i=1}^d \int_{\mathbb{R}^d} (t_1(\partial_{a_i}f)^2+t_2(\partial_{b_i}f)^2)/f-(d/2)\int_{V_1} ((\partial_{a_1}f_1)^2+(\partial_{b_1}f_1)^2)/f_1.\] By the symmetries of \(f\) we have that the terms in the above sum are all the same and thus \[\label{eq:aa11} \Lambda_F'(0)=d\int_{\mathbb{R}^d} (t_1(\partial_{a_1}f)^2+t_2(\partial_{b_1}f)^2)/f-(d/2)\int_{V_1} ((\partial_{a_1}f_1)^2+(\partial_{b_1}f_1)^2)/f_1.\] We apply Lemma [\[csch\]](#csch){reference-type="ref" reference="csch"} for \(a_1\) and \(b_1\) and ([\[eq:aa11\]](#eq:aa11){reference-type="ref" reference="eq:aa11"}) to obtain that \[\label{eq:ab1}\Lambda_F'(0)\geq (dt_1-d/2)\int_{V_1} (\partial_{a_1}f_1)^2/f_1+(dt_2-d/2)\int_{V_1} (\partial_{b_1}f_1)^2/f_1.\] Using the symmetry of \((X_1,Z)\) we obtain that \[\label{eq:ab2}\int_{V_1} (\partial_{a_1}f_1)^2/f_1=\int_{V_1} (\partial_{b_1}f_1)^2/f_1.\] It follows that \(\Lambda_F'(0)\geq 0\). The proof of the first part of Proposition [\[gaussprop\]](#gaussprop){reference-type="ref" reference="gaussprop"} is now complete. We arrived to the second part of Proposition [\[gaussprop\]](#gaussprop){reference-type="ref" reference="gaussprop"}. Assume that \(F\) satisfies \(\Lambda_F'(t)=0\) for every \(t\geq 0\). Using the notation from Lemma [\[wr\]](#wr){reference-type="ref" reference="wr"} we have by \(t_1+t_2=1\) that at least one of \(t_1>0\) and \(t_2>0\) holds. Without loss of generality we assume that \(t_1>0\). Let \(g_t:\mathbb{R}^d\rightarrow\mathbb{R}\) denote the logarithm of the density function of \(B_t=T(F_t)\). This implies by Lemma [\[csch\]](#csch){reference-type="ref" reference="csch"} that \(\partial_{a_1}g_t\) satisfies the property that \(\partial_{a_1}g_t(x)=\partial_{a_1}g_t(x+z)\) holds whenever \(x\in \mathbb R^d\) and \(z\in V_1^{\perp}\). For convenience, we will write every element \(x\in \mathbb R^d\) as a triple \((\alpha(x), \beta(x), \gamma(x))\), where \(\alpha(x)=(x, a_1)\), \(\beta(x)=(x, b_1)\) and \(\gamma(x)\) is the projection of \(x\) to \(V_1^{\perp}\). Using this notation, we have that \(\partial_{a_1}g_t(x)=h_t(\alpha(x), \beta(x))\). This means that there exists a function \(\widehat{h_t}: \mathbb R^2\rightarrow \mathbb R\) such that \(\partial_{a_1} \widehat{ h_t}(\alpha(x), \beta(x))=\partial_{a_1}g_t(x)\). We have by \(\partial_{a_1}(g_t-\widehat{h_t})=0\) that \(g_t(x)-\widehat{h_t}(\alpha(x), \beta(x))=s_t(\beta(x), \gamma(x))\) for some function \(s_t\). We obtain that the density function of \(B_t\) can be written in the following form. \[\exp(\widehat{h_t}(\alpha(x), \beta(x)))\cdot \exp(s_t(\beta(x), \gamma(x)).\] In other words, this means that the random variables of \(\alpha(B_t)\) and \(\gamma(B_t)\) are conditionally independent with respect to \(\beta(B_t)\). This implies by Lemma [\[twoposs\]](#twoposs){reference-type="ref" reference="twoposs"} that one of the following two possibilities holds: either \(\alpha(B)\) is independent of \((\beta(B), \gamma(B))\) or \(\gamma(B)\) is independent of \((\alpha(B), \beta(B))\). In the first case we obtain (using the terminology of Lemma [\[indepgauss\]](#indepgauss){reference-type="ref" reference="indepgauss"}) that \(a_1\) is an independent direction for \(B\). By symmetries of \(B\) we obtain that \(\{a_i\}_{i=1}^d\) are all independent directions for \(B\). If \(t_1<1\) then \((a_i,a_j)\neq 0\) for every pair \(1\leq i<j\leq d\) and Lemma [\[indepgauss\]](#indepgauss){reference-type="ref" reference="indepgauss"} finishes the proof. If \(t_1=1\), then \(\{a_i\}_{i=1}^d\) is an orthonormal basis in \(\mathbb{R}^d\) and \(b_1=\sum_{i=2}^{d-1}a_i(d-1)^{-1/2}\). We have that \((B,a_i)\) are identically distributed independent random variables and that \((B,b_1)=\sum_{i=2}^{d-1}(B,a_i)(d-1)^{-1/2}\) has the same distribution. This is only possible if this \((B,a_i)\) is normal for every \(i\). We obtain that \(B\) is Gaussian. In the case when \(\gamma(B)\) is independent of \((\alpha(B), \beta(B))\) we have that \(\partial_u g(x)=\partial_ug(x+z)\) holds for every triple \(u\in V_1, x\in\mathbb{R}^d,z\in V_1^\perp\). The symmetries of \(f\) imply that \(\partial_u g(x)=\partial_ug(x+z)\) holds for every triple \(u\in V_i, x\in\mathbb{R}^d,z\in V_i^\perp\). Let \(r_i\) denote the orthogonal projection of \(a_i\) to \(w^\perp\) for \(1\leq i\leq d\). Note that the vector system \(\{r_i\}_{i=1}^d\) is completely symmetric in the sense that the origin is the center of a regular simplex whose vertices are given by these vectors. We have for every \(1\leq i\leq d\) that \(\partial_{r_i} g(x)=h_i((x,r_i),(x,w))\) for some two variable function \(h_i:\mathbb{R}^2\rightarrow\mathbb{R}\). The symmetries of \(f\) imply that \(h_i\) does not depend on \(i\) and thus \(h_i=h\) for some \(h\) for every \(i\). The next step is to prove that \(h(x,y)=xh^*(y)\) for some one variable function \(h^*\). We have by \(\sum_{i=1}^d r_i=0\) that \(\sum_{i=1}^d\partial_{r_i}g(x)=0\) and thus \(\sum_{i=1}^d h_i((x,r_i),(x,w))=0\) holds for every \(x\in\mathbb{R}^d\). For arbitrary numbers \(x_1,x_2,\dots,x_d,y\in\mathbb{R}\) with \(\sum_{i=1}^d x_i=0\) we can choose \(x\in\mathbb{R}^d\) such that \((x,r_i)=x_i\) and \((x,w)=y\). It follows that \(\sum_{i=1}^d h(x_i,y)=0\) holds for arbitrary numbers with \(\sum_{i=1}^d x_i=0\). Assume first that all \(x_i\) is \(0\) then we have that \(dh(0,w)=0\) and thus \(h(0,w)=0\) for every \(w\). Then assume that \(x_1=a,x_2=-a\) and \(x_i=0\) if \(i\geq 3\). We obtain that \(h(a,w)+h(-a,w)=0\) and thus \(h(-a,w)=-h(a,w)\) holds for every \(a\) and \(w\). Finally let \(x_1=a,x_2=b,x_3=-a-b\) and \(x_i=0\) if \(i\geq 4\). We obtain that \(h(a,w)+h(b,w)=-h(-a-b,w)=h(a+b,w)\). Since \(h\) is additive and continuous in the first coordinate a well known fact implies that \(h\) is a linear function in the first coordinate and thus we obtain \(h(x,y)=xh^*(y)\). Now we have \(\partial_{r_i} g(x)=(x,r_i)h^*((x,w))\) for every \(1\leq i\leq d\). It is easy to see that this implies that \(g(x)=\|x-(x,w)w\|_2^2h^*((x,w))+c^*((x,w))\) where \(c^*:\mathbb{R}\rightarrow\mathbb{R}\) is some function. We have that \(\partial_w g(x)=\|x-w(x,w)\|_2^2(h^*)'((x,w))+(c^*)'((x,w))\). On the other hand we have that \(\partial_w g(x)=\partial_w g(x+z)\) holds whenever \(z\in V_i^\perp\) for every \(1\leq i\leq d\). It follows that \((h^*)'=0\) everywhere and so \(g(x)=c\|x-(x,w)w\|_2^2+c^*((x,w))\) with some constant \(c\). Now from \(f(x)=\exp(c\|x-(x,w)w\|_2^2+c^*((x,w)))\) we have that \(B-(B,w)w\) and \((B,w)\) are independent random variables. Furthermore \(B-(B,w)w\) is a Gaussian distribution concentrated on the orthogonal space of \(w\). This means that the pair \((B,r_1),(B,w)\) of random variables is independent and \((B,r_1)\) is Gaussian. We know that \((B,v_1)\) is a linear combination of \((B,r_1)\) and \((B,w)\) (with a non-zero coefficient for \((B,r_1)\)) and its distribution is the same as the distribution of \((B,w)\) (here we use the symmetries of \(B\)). It follows that \((B,w)\) is also Gaussian and thus \(c^*((x,w))=c_2(x,w)^2+c_3\) for some constants \(c_2,c_3\). Thus we have that \(B\) is a Gaussian joint distribution implying that \(F\) is also joint Gaussian, as it is a linear function of \(B\). \(\square\)   From Proposition [\[lem:equiv\]](#lem:equiv){reference-type="ref" reference="lem:equiv"} we have that theorem [\[mainlim\]](#mainlim){reference-type="ref" reference="mainlim"} implies theorem [\[main\]](#main){reference-type="ref" reference="main"}. Let \(\mu\) be a smooth typical eigenvector process corresponding to eigenvalue \(\lambda\) represented by a system of random variables \(\{X_v\}_{v\in V_d}\). According to the results in chapter [7](#smooth){reference-type="ref" reference="smooth"} it is enough to show that \(\mu\) is a Gaussian wave. We have by Lemma [\[lem:friedman\]](#lem:friedman){reference-type="ref" reference="lem:friedman"} that \(\lambda\in [-2\sqrt{d-1},2\sqrt{d-1}]\). Let \(F=\{X_v\}_{v\in C}\). It is clear that \(F\in\mathcal{F}_{d,\lambda}\). We have by Theorem [\[thm:imprmain\]](#thm:imprmain){reference-type="ref" reference="thm:imprmain"} that \(\mathcal{D}(F)\geq\mathcal{D}(F^*)\) and thus by Theorem [\[gausscond\]](#gausscond){reference-type="ref" reference="gausscond"} we obtain that \(F=F^*\). Again by Theorem [\[thm:imprmain\]](#thm:imprmain){reference-type="ref" reference="thm:imprmain"} we have that \(\mu\) is \(2\)-Markov and so \(\{X_v\}_{v\in V_d}\) can be obtained by iterating conditionally independent couplings of \(C\) along edges (see chapter [10](#chap:impr){reference-type="ref" reference="chap:impr"}). This shows the Gaussianity of the whole system \(\{X_v\}_{v\in V}\). # Appendix A: On heated random variables {#heatedrandom} Let \(X\) be a random variable with values in \(\mathbb{R}^n\) and let \(M\) be the standard normal distribution on \(\mathbb{R}^n\). Let \(f_t\) denote the density function of \(X+\sqrt{2t}M\) and let \(\mu_t\) denote the corresponding measure on \(\mathbb{R}^n\). The standard heat equation says that \(\partial_tf_t=\triangle f_t\) holds for every \(t>0\). It is useful to compute the variation of the differential entropy \(\mathbb D(f_t)\). The de Bruijn identity (see e.g. ) says that \[\label{diffheat} \partial_t(\mathbb D(f_t))=\partial_t \int_{\mathbb{R}^n}-f_t\log f_t=\int_{\mathbb{R}^n}-\triangle f_t (1+\log f_t)=\int_{\mathbb{R}^n} \|\triangledown f_t\|_2^2/f_t.\] However the validity of ([\[diffheat\]](#diffheat){reference-type="ref" reference="diffheat"}) relies on the fact that both \(\partial_i f_t\) and \(\partial_i f_t\log f_t\) vanish at infinity for every \(1\leq i\leq n\). This fact is proved in Lemma [\[korlat\]](#korlat){reference-type="ref" reference="korlat"}. Notice that if \(X\) has finite variance then also \(X+\sqrt{2t}M\) has finite variance and if \(t>0\) then \(\mathbb D(f_t)\) is a finite quantity (Lemma [\[lem:veges\]](#lem:veges){reference-type="ref" reference="lem:veges"}). In general if \(\sigma>0\) then the density function \(f\) of \(X+\sigma M\) is smooth, non-vanishing and analytic restricted to every line in \(\mathbb{R}^n\). More precisely, if \(p,q\in\mathbb{R}^n\) then the real function \(\lambda\mapsto f(p+\lambda(q-p))\) extends to an entire analytic function on \(\mathbb{C}\). Furthermore every partial derivative of \(f\) has this property. In the rest of this appendix we prove several other facts about heated random variables. \(\square\) \(\square\) Let \(X\) be a random variable with values in \(\mathbb{R}^d\). We say that \(v\in\mathbb{R}^d\) is an independent direction for \(X\) if the \(\mathbb{R}\)-valued random variable \((X,v)\) is independent from the projection of \(X\) to the \(d-1\) dimensional space \(v^{\perp}\). Note that every direction is independent for the standard normal distribution. # Appendix B: Differential entropy Differential entropy is defined as follows for absolutely continuous random vectors. Some properties of the discrete entropy are preserved (e.g. it is additive if we put together independent random variables), others do not hold any more; an essential difference is that differential entropy does not have to be nonnegative. To see the connection between entropy in the discrete case and differential entropy, recall Theorem 9.3.1 from. This says that if we divide the range of \(X\) into bins of length \(\delta\), and \(X^{\delta}\) denotes the quantized version of \(X\) with respect to this grid, then \[\mathbb H(X^{\delta})+\log \delta\rightarrow \mathbb H(X)\] as \(\delta\rightarrow 0\), assuming that the density of \(X\) is Riemann integrable. The following well-known lemma shows how the differential entropy is modified when we apply a linear transformation to the random vector (see e.g. corollary to Theorem 9.6.4 in. The following lemma is equivalent to the fact that the nonnegativity of conditional mutual information holds for differential entropy as well. We include a proof for completeness. \(\square\) # Appendix C: Factor of i.i.d. processes Let \(f:[0,1]^{V_d}\rightarrow Y\) be a measurable function such that it is invariant under root preserving automorphisms. We can use \(f\) to construct an invariant process in the following way. First we put independent uniformly random elements from \([0,1]\) on the vertices of \(T_d\). Then, at each vertex \(v\), we evaluate \(f\) for this random labeling such that the root is placed to \(v\). If \(f\) depends only on finitely many coordinates, then the corresponding process is called a block factor of i.i.d. process. The next proposition was proved by Harangi and Virág in.

{'timestamp': '2016-07-19T02:05:12', 'yymm': '1607', 'arxiv_id': '1607.04785', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04785'}

# Introduction ## Introduction The Fisher--Rao metric on the space \(\on{Prob}(M)\) of probability densities is invariant under the action of the diffeomorphism group \(\on{Diff}(M)\). Restricted to finite-dimensional submanifolds of \(\on{Prob}(M)\), so-called statistical manifolds, it is called Fisher's information metric. A uniqueness result was established for Fisher's information metric on finite sample spaces and extended it to infinite sample spaces. The Fisher--Rao metric on the infinite-dimensional manifold of all positive probability densities was studied in, including the computation of its curvature. In it was proved that any \(\on{Diff}(M)\)-invariant Riemannian metric on the space \(\on{Dens}_+(M)\) of smooth positive densities on a compact manifold \(M\) without boundary is of the form \[\label{Fisher-Rao-metric} G_\mu(\alpha,\beta)=C_1(\mu(M)) \int_M \frac{\alpha}{\mu}\frac{\beta}{\mu}\,\mu + C_2(\mu(M)) \int_M\alpha \cdot \int_M\beta\] for some smooth functions \(C_1,C_2\) of the total volume \(\mu(M)\). This implies that the Fisher--Rao metric on \(\on{Prob}(M)\) is, up to a multiplicative constant, the unique \(\on{Diff}(M)\)-invariant metric. By Cauchy--Schwarz the metric [\[Fisher-Rao-metric\]](#Fisher-Rao-metric){reference-type="eqref" reference="Fisher-Rao-metric"} is positive definite if and only if \(C_2(m)>-\frac1m C_1(m)\) for all \(m>0\). ## The setting Let \(M^m\) be a smooth compact manifold. It may have boundary or it may even be a manifold with corners; i.e., modelled on open subsets of quadrants in \(\mathbb R^m\). For a detailed description of the line bundle of smooth densities we refer to or. We let \(\on{Dens}_+(M)\) denote the space of smooth positive densities on \(M\), i.e., \(\on{Dens}_+(M) = \{ \mu \in \Gamma(\on{Vol}(M)) \,:\, \mu(x) > 0\; \forall x \in M\}\). Let \(\on{Prob}(M)\) be the subspace of positive densities with integral 1 on \(M\). Both spaces are smooth Fréchet manifolds; in particular they are open subsets of the affine spaces of all densities and densities of integral 1 respectively. For \(\mu \in \on{Dens}_+(M)\) we have \(T_\mu \on{Dens}_+(M) = \Gamma(\on{Vol}(M))\) and for \(\mu\in \on{Prob}(M)\) we have \[T_\mu\on{Prob}(M)=\{\alpha\in \Gamma(\on{Vol}(M)): \int_M\alpha =0\}.\] The Fisher--Rao metric, given by \(G^{\operatorname{FR}}_\mu(\alpha,\beta) = \int_M \frac{\alpha}{\mu}\frac{\beta}{\mu}\mu\) is a Riemannian metric on \(\on{Prob}(M)\); it is invariant under the natural action of the group \(\on{Diff}(M)\) of all diffeomorphisms of \(M\). If \(M\) is compact without boundary of dimension \(\ge 2\), the Fisher-Rao metric is the unique \(\on{Diff}(M)\)-invariant metric up to a multiplicative constant. This follows, since any \(\on{Diff}(M)\)-invariant Riemannian metric on \(\on{Dens}_+(M)\) is of the form [\[Fisher-Rao-metric\]](#Fisher-Rao-metric){reference-type="eqref" reference="Fisher-Rao-metric"} as proved in. ## Overview We will study four different representations of the metric \(G\) in [\[Fisher-Rao-metric\]](#Fisher-Rao-metric){reference-type="eqref" reference="Fisher-Rao-metric"}. The first representation is \(G\) itself on the space \(\on{Dens}_{+}(M)\). Next we fix a density \(\mu_0 \in \on{Prob}(M)\) and consider the mapping \[R:\on{Dens}_+(M)\to C^\infty(M,\mathbb R_{>0})\,, \qquad R(\mu) = f = \sqrt{\frac{\mu}{\mu_{0}}}\.\] This map is a diffeomorphism with inverse \(R^{-1}(f)= f^2\mu_0\), and we will denote the induced metric by \(\tilde G = \left(R^{-1}\right)^\ast G\); it is given by the formula \[\tilde G_{f}(h,k) = 4 C_1(\| f\|^2) \langle h, k \rangle + 4 C_2(\| f \|^2) \langle f, h \rangle \langle f, k \rangle\,,\] with \(\| f\|^2 = \int_M f^2 \mu_0\) denoting the \(L^2(\mu_0)\)-norm, and this formula makes sense for \(f \in C^{\infty}(M,\mathbb R)\). See Sect. [1.5](#geodesics){reference-type="ref" reference="geodesics"} for calculations. Next we take the pre-Hilbert space \((C^{\infty}(M,\mathbb R), \langle\;,\;\rangle_{L^2(\mu_0)})\) and pass to polar coordinates. Let \(S = \{ \varphi \in L^2(M,{\mathbb R}) \,:\, \int_M \varphi^2 \mu_0 = 1\}\) denote the \(L^2\)-sphere. Then \[\Phi: C^\infty(M, {\mathbb R}_{>0}) \to {\mathbb R}_{>0} \times (S \cap C^\infty_{>0})\,,\qquad \Phi(f) = (r,\varphi) = \left(\| f \|, \frac f{\|f\|}\right)\,,\] is a diffeomorphism, where \(C^\infty_{>0} = C^\infty(M,{\mathbb R}_{>0})\); its inverse is \(\Phi^{-1}(r,\varphi)= r.\varphi\). We set \(\bar G = \left(\Phi^{-1}\right)^\ast \tilde G\); the metric \(\bar G\) has the expression \[\bar G_{r,\varphi} = g_1(r) \langle d\varphi, d\varphi \rangle + g_2(r) dr^2\,,\] with \(g_1(r) = 4C_1(r^2) r^2\) and \(g_2(r) = 4\left(C_1(r^2) + C_2(r^2) r^2\right)\). Finally we change the coordinate \(r\) diffeomorphically to \[s= W(r) = \int_1^r \sqrt{g_2(\rho)} \,d\rho\.\] Then, defining \(a(s) = 4 C_1(r(s)^2) r(s)^2\), we have \[\bar G_{s,\varphi} = a(s) \langle d\varphi, d\varphi \rangle + ds^2\.\] We will use \(\bar G\) to denote the metric in both \((r,\varphi)\) and \((s,\varphi)\) coordinates. Let \(W_-= \lim_{r\to 0+} W(r)\) and \(W_+ = \lim_{r\to \infty} W(r)\). Then \(W: {\mathbb R}_{>0} \to (W_-, W_+)\) is a diffeomorphism. This completes the first row in Fig. [\[fig:diagram\]](#fig:diagram){reference-type="ref" reference="fig:diagram"}. The geodesic equation of \(G\) in the various representations will be derived in Sect. [1.5](#geodesics){reference-type="ref" reference="geodesics"}. The formulas for the geodesic equation and later for curvature are infinite-dimensional analoga of the corresponding formulas for warped products; see or. The four representations are summarized in the following diagram. \[\on{Dens}_+(M) \ar[r]^-R & C^\infty(M,{\mathbb R}_{>0}) \ar[r]^-\Phi & {\mathbb R}_{>0} \times (S \cap C^\infty_{>0}) \ar[r]^-{W \times \on{Id}} & (W_-, W_+) \times (S \cap C^\infty_{>0})\. }\] Since \(\bar G\) induces the canonical metric on \((W_-, W_+)\), a necessary condition for \(\bar G\) to be geodesically complete is \((W_-, W_+) = (-\infty, +\infty)\). Rewritten in terms of the functions \(C_1\) and \(C_2\) this becomes \[W_+ = \infty \Leftrightarrow\left( \int_1^\infty r^{-1/2} \sqrt{C_1(r)} \,dr = \infty \text{ or } \int_1^\infty \sqrt{C_2(r)} \,dr = \infty \right)\,,\] and similarly for \(W_-=-\infty\), with the limits of integration being 0 and 1. If \(\bar G\) is incomplete, i.e., \(W_->-\infty\) or \(W_+<\infty\), there are sometimes geodesic completions. See Sect. [1.8](#completions){reference-type="ref" reference="completions"} for details. We now assume that \((W_-, W_+) = (-\infty, +\infty)\). The metrics \(\bar G\) and \(\tilde G\) can be extended to the spaces \({\mathbb R}\times S \cap C^\infty\) and \(C^\infty(M,{\mathbb R}) \setminus \{0\}\) and the last two maps in the diagram \[\Gamma_{C^1}(\on{Vol}(M)) \!\setminus\! \{0\} & C^\infty(M,{\mathbb R}) \!\setminus\! \{0\} \ar[l]_{\qquad R^{-1}} \ar[r]^-\Phi & {\mathbb R}_{>0} \times S \cap C^\infty \ar[r]^-{W \times \on{Id}}& {\mathbb R} \times S \cap C^\infty }\] are bijections. The extension of \(R^{-1}\) is given by \(R^{-1}(f) = f |f| \mu_0\); it does not map into smooth densities any more, but only into \(C^1\)-sections of the volume bundle; however, \(R^{-1}\) is not surjective into \(C^1\)-sections, because the loss of regularity for \(R^{-1}(f)\) occurs only at point where \(f\) is \(0\). The last two maps, \(\Phi\) and \(W \times \on{Id}\), are diffeomorphisms. The following will be shown in Sect. [1.7](#completeness){reference-type="ref" reference="completeness"}: \((W_-, W_+) = (-\infty, +\infty)\) implies that \(({\mathbb R} \times S \cap C^\infty, \bar G)\) is geodesically complete and hence so are \(({\mathbb R}_{>0} \times S \cap C^\infty, \bar G)\) and \((C^\infty(M,{\mathbb R}) \setminus \{0\}, \tilde G)\). Finally we consider the metric completions, still assuming that \((W_-, W_+) = (-\infty, +\infty)\). For \(\bar G\) this is \({\mathbb R} \times S\) or \({\mathbb R}_{>0} \times S\) in \((s,\varphi)\) or \((r,\varphi)\)-coordinates, respectively, as shown in Sect. [1.7](#completeness){reference-type="ref" reference="completeness"}. The metrics and maps can be extended to \[\Gamma_{L^1}(\on{Vol}(M)) \setminus \{0\} \ar[r]^-R & L^2(M,{\mathbb R}) \setminus \{0\} \ar[r]^-\Phi & {\mathbb R}_{>0} \times S \ar[r]^-{W \times \on{Id}} & {\mathbb R} \times S }\.\] Here \(\Gamma_{L^1}\) denotes the space of \(L^1\)-sections. The extension of \(R\) is given by \(R(\mu) = \on{sgn}(\mu) \sqrt{|\mu|/\mu_0}\) and its inverse is \(R^{-1}(f) = f |f| \mu_0\) as before. The last two maps are diffeomorphisms and hence \((L^2(M,{\mathbb R}) \setminus \{0\}, \tilde G)\) is metrically complete. The extension of \(R\) is bijective, but not a diffeomorphism. It is continuous, but not \(C^1\), and its inverse is \(C^1\), but not \(C^2\); furthermore \(DR^{-1}(f)\) is not surjective if \(f=0\) on a set of positive measure. However we can use \(R\) to pull back the geodesic distance function from \(L^2(M,{\mathbb R}) \setminus \{0\}\) to \(\Gamma_{L^1}(\on{Vol}(M)) \setminus \{0\}\) to obtain a complete metric on the latter space, that is compatible with the standard topology. ## The inverse \(R^{-1}\) and geodesic completeness There is more than one choice for the extension of \(R^{-1}(f) = f^2 \mu_0\) from \(C^\infty(M,{\mathbb R}_{>0})\) to \(C^\infty(M,{\mathbb R})\). The choice \(R^{-1}(f) = f |f| \mu_0\) remains injective and can be further extended to a bijection on the metric completion \(L^2(M,{\mathbb R}) \setminus \{0\}\). We can consider the equally natural extension \(Q\) and its factorizarion given by \[\begin{gathered} C^\infty(M, \mathbb R) \ar[rr]^{Q} \ar[dr]_{Q_1} && \Gamma_{\ge 0}(\on{Vol}(M))\\ & \{|f|: f\in C^\infty(M,\mathbb R)\} \ar[ur]_{Q_2}& } \\ Q(f) = f^2\mu_0\,,\quad Q_1(f) = |f|\,, \quad Q_2(|f|) = |f|^2\mu_0\. \end{gathered}\] into the space of smooth, nonnegative sections. The map \(Q\) is not surjective; see for a discussion of smooth non-negative functions admitting smooth square roots. The image \(\{|f|: f\in C^\infty(M,\mathbb R)\}\) of \(Q_1\) looks somewhat like the orbit space of a discrete reflection group: An example of a codimension 1 wall of the image could be \(\{|f|: f\in C^\infty(M,\mathbb R), f(x)=0\}\) for one fixed point \(x\in M\). Since this is dense in the \(L^2\)-completion of \(T_f C^\infty(M,\mathbb R)\) with respect to \(\tilde G_f\), we do not have a reflection at this wall. Fixing \(\varphi_0\in S\cap C^\infty\) and considering \(\{(r,\varphi) \in \mathbb R_{>0}\times S\cap C^\infty: \langle \varphi_0, \varphi\rangle = 0 \}\) we can write the orthogonal reflection \((r, t_1 \varphi_0 + t_2 \varphi)\mapsto (r,-t_1\varphi_0 + t_2\varphi)\). Geodesics in \((C^\infty(M,\mathbb R),\tilde G)\) are mapped by \(Q_1\) to curves that are geodesics in the interior \(C^\infty(M,\mathbb R_{>0})\), and that are reflected following Snell's law at any hyperplanes in the boundary for which the angle makes sense. The mapping \(Q_2\) then smoothes out the reflection to a 'quadratic glancing of the boundary' if one can describe the smooth structure of the boundary. It is tempting to paraphrase this as: *The image of \(Q\) is geodesically complete.* But note that: The metric \(G\) becomes ill-defined on the boundary. The boundary is very complicated; each closed subset of \(M\) is the zeroset of a smooth non-negative function and thus corresponds to a 'boundary component'. Some of them 'look like reflection walls'. One could try to set up a theory of infinite dimensional stratified Riemannian manifolds and geodesics on them to capture this notion of geodesic completeness, similarly to. But the situation is quite clear geometrically, and we prefer to consider the geodesic completion described by the inverse \(R^{-1}\) used in this paper, which is perhaps more natural. ## Geodesics of the Fisher-Rao metric on \(\on{Dens}_+(M)\) {#geodesics} In it was shown that \(\on{Prob}(M)\) has constant sectional curvature for the Fisher-Rao metric. For fixed \(\mu_{0}\in \on{Prob}(M)\) we consider the mapping \[R:\on{Dens}_+(M)\to C^\infty(M,\mathbb R_{>0}), \qquad R(\mu) = \sqrt{\frac{\mu}{\mu_{0}}}\.\] The inverse \(R^{-1}: C^{\infty}(M,\mathbb R_{>0})\to \on{Dens}_{+}(M)\) is given by \(R^{-1}(f)=f^{2}\mu_{0}\); its tangent mapping is \(T_{f}R^{-1}.h = 2fh\mu_{0}\). The Fisher--Rao metric induces the following metric on the open convex cone \(C^{\infty}(M,\mathbb R_{>0})\subset C^{\infty}(M,\mathbb R)\): \[\begin{gathered} \tag{a} \left((R^{-1})^{*}G\right)_{f}(h,k) = G_{R^{-1}(f)}(T_{f}R^{-1}.h, T_{f}R^{-1}.k) = G_{f^2\mu_0}(2fh\mu_0,2fk\mu_0) \\ = C_1(\|f\|^2_{L^2(\mu_0)}) \int \frac{2fh\mu_0}{f^2\mu_0}\frac{2fk\mu_0}{f^2\mu_0} f^2\mu_0 + C_2(\|f\|^2_{L^2(\mu_0)})\int 2fh\mu_0\cdot \int 2fk\mu_0 \\ = 4C_1(\|f\|^2) \int hk\mu_0 + 4C_2(\|f\|^2)\int fh\mu_0\cdot \int fk\mu_0 \\= 4C_1(\|f\|^2) \langle h,k\rangle + 4C_2(\|f\|^2)\langle f,h\rangle \langle f,k\rangle \\ \shoveleft{= 4C_1(\|f\|^2) \Big\langle h-\frac{\langle f,h\rangle}{\|f\|^2} f,k-\frac{\langle f,k\rangle}{\|f\|^2} f\Big\rangle\; +} \\ + 4\big(C_2(\|f\|^2).\|f\|^2 + C_1(\|f\|^2)\big)\Big\langle \frac{f}{\|f\|},h\Big\rangle \Big\langle \frac{f}{\|f\|},k\Big\rangle\,, \end{gathered}\] where in the last expression we split \(h\) and \(k\) into the parts perpendicular to \(f\) and multiples of \(f\). We now switch to *polar coordinates on the pre-Hilbert space*: Let \(S = \{\varphi\in L^2(M,{\mathbb R}): \int \varphi^2\mu_0 =1 \}\) denote the sphere, and let \(S\cap C^{\infty}_{>0}\) be the intersection with the positive cone. Then \(C^{\infty}(M,\mathbb R) \setminus \{0\}\cong \mathbb R_{>0}\times S\cap C^{\infty}\) via \[\Phi: C^\infty(M, {\mathbb R}) \setminus \{0\} \to {\mathbb R}_{>0} \times S\,,\qquad \Phi(f) = (r,\varphi) = \left(\| f \|, \frac f{\|f\|}\right)\.\] Note that \(\Phi(C^{\infty}(M,\mathbb R_{>0}))=\mathbb R_{>0}\times S\cap C^{\infty}_{>0}\). We have \(f = \Phi^{-1}(r,\varphi) = r.\varphi\) thus \(df = r\,d\varphi +\varphi\,dr\), where \(r\,d\varphi(h)= h-\langle\varphi,h\rangle\varphi\) is the orthogonal projection onto the tangent space of \(S\) at \(\varphi\) and \(dr(h)=\langle\varphi,h\rangle\). The Euclidean (pre-Hilbert) metric in polar coordinates is given by \[\begin{aligned} \langle df,df \rangle &= \langle \varphi.dr + r.d\varphi,\varphi.dr + r.d\varphi\rangle = \langle\varphi,\varphi\rangle dr^2 + 2r.\langle\varphi,d\varphi\rangle.dr + r^2\langle d\varphi,d\varphi\rangle \\& = dr^2 + r^2\langle d\varphi,d\varphi\rangle\. \end{aligned}\] The pullback metric is then \[\begin{aligned} \tag{b} \bar G=\left((\Phi^{-1})^\ast \tilde G\right) &= 4C_1(r^2) r^2 \langle d\varphi,d\varphi\rangle + 4\left( C_2(r^2) r^2 + C_1(r^2)\right) dr^2 \\ &= g_1(r) \langle d\varphi,d\varphi\rangle + g_2(r) dr^2 \\ &= a(s) \langle d\varphi,d\varphi\rangle + ds^2 \,, \end{aligned}\] where we introduced the functions \[g_1(r) = 4C_1(r^2) r^2 \quad \text{ and }\quad g_2(r) = 4\left( C_2(r^2) r^2 + C_1(r^2)\right)\,,\] and where in the last expression we changed the coordinate \(r\) diffeomorphically to \[s(r) = 2 \int_1^r \sqrt{ C_2(\rho^2) \rho^2 + C_1(\rho^2)}\,d\rho \quad\text{ and let }a(s) = 4C_1(r(s)^2) r(s)^2.\] The resulting metric is a radius dependent scaling of the metric on the sphere times a different radius dependent scaling of the metric on \({\mathbb R}_{>0}\). Note that the metric (as well as the metric in the last expression of ) is actually well-defined on \(C^{\infty}(M,\mathbb R)\setminus \{0\} \cong \mathbb R_{>0}\times S \cap C^\infty\); this leads to a (partial) geodesic completion of \((\on{Dens}_+(M), G)\). Geodesics for the metric follow great circles on the sphere with some time dependent stretching, since reflection at any hyperplane containing this great circle is an isometry. We derive the geodesic equation. Let \([0,1]\times(-\varepsilon,\varepsilon)\ni (t,s)\mapsto (r(t,s),\varphi(t,s))\) be a smooth variation with fixed ends of a curve \((r(t,0),\varphi(t,0))\). The energy of the curve and its derivative with respect to the variation parameter \(s\) are as follows, where \(\nabla^S\) is the covariant derivative on the sphere \(S\). \[\begin{aligned} E(r,\varphi) &= \int_0^1 \left( \frac 12 g_1(r) \langle \varphi_t, \varphi_t \rangle + \frac 12 g_2(r). r_t^2 \right) dt \\ \partial_s E(r,\varphi) &= \int_0^1 \Big( \frac 12 g_1'(r).r_s \langle \varphi_t, \varphi_t \rangle + g_1(r) \langle \nabla^S_{\partial_s} \varphi_t, \varphi_t \rangle + \\ &\qquad\qquad +\frac 12 g_2'(r).r_s.r_t^2 + g_2(r).r_t.r_{ts} \Big) dt \\ &= \int_0^1 \Big( \frac 12 g_1'(r).r_s \langle \varphi_t, \varphi_t \rangle-g_1'(r).r_t \langle \varphi_s, \varphi_t \rangle-g_1(r) \langle \varphi_s, \nabla^S_{\partial_t} \varphi_t \rangle + \\ &\qquad\qquad +\frac 12 g_2'(r).r_s.r_t^2-g_2'(r).r_t^2.r_s-g_2(r).r_{tt}.r_s \Big) dt \\ &= \int_0^1 \Big( \frac 12 g_1'(r) \langle \varphi_t, \varphi_t \rangle-\frac 12 g_2'(r).r_t^2-g_2(r).r_{tt} \Big) r_s \\ &\qquad\qquad-\Big( g_1'(r).r_t \langle \varphi_s, \varphi_t \rangle + g_1(r) \langle \varphi_s, \nabla^S_{\partial_t} \varphi_t \rangle \Big) dt\. \end{aligned}\] Thus the geodesic equation is \[\tag{c}\boxed{\; \begin{aligned} \nabla^S_{\partial_t}\varphi_{t} &=-\partial_t\left( \log g_1(r) \right) \varphi_t \\ r_{tt} &= \frac 12 \frac{ g_1'(r)}{g_2(r)} \langle \varphi_t, \varphi_t \rangle-\frac 12 \partial_t \left( \log g_2(r) \right) r_t \end{aligned} \;}\] Using the first equation we get: \[\begin{aligned} \partial_t\langle \varphi_t,\varphi_t\rangle &= 2\langle\nabla_{\partial_t}\varphi_t,\varphi_t\rangle =-2\, \partial_t\left( \log g_1(r) \right) \langle \varphi_t, \varphi_t \rangle \\ \partial_t \left( \log \langle \varphi_t, \varphi_t \rangle \right) &=-2\, \partial_t\left( \log g_1(r) \right) \\ \log(\|\varphi_t\|^2) &=-2 \log g_1(r) + 2\log A_0 \quad\text{ with }\quad A_0 = g_1(r)\, \| \varphi_t \| \,, \end{aligned}\] which describes the speed of \(\varphi(t)\) along the great circle in terms of \(r(t)\); note that the quantity \(g_1(r) \| \varphi_t\|\) is constant in \(t\). The geodesic equation simplifies to \[\tag{d}\boxed{\; \begin{aligned} \nabla^S_{\partial_t}\varphi_{t} &= -\partial_t\left( \log g_1(r) \right) \varphi_t \\ r_{tt} &= \frac {A_0^2}2 \frac{g_1'(r)}{g_1(r)^2 g_2(r)}-\frac 12 \partial_t \left( \log g_2(r) \right) r_t \end{aligned} \;}\] with \(g_1(r) = 4C_1(r^2) r^2\) and \(g_2(r) = 4\left( C_2(r^2) r^2 + C_1(r^2)\right)\). We can solve equation for \(\varphi\) explicitely. Given initial conditions \(\varphi_0, \psi_0\), the geodesic \(\tilde \varphi(t)\) on the sphere with radius \(1\) satisfying \(\tilde \varphi(0) = \varphi_0\), \(\tilde \varphi_t(0) = \psi_0\) is \[\tilde\varphi(t) = \cos(\| \psi_0\| t) \varphi_0 + \sin(\|\psi_0\| t ) \frac{\psi_0}{\|\psi_0\|}\.\] We are looking for a reparametrization \(\varphi(t) = \tilde \varphi(\alpha(t))\). Inserting this into the geodesic equation we obtain \[\begin{aligned} \partial_t^2 \left(\tilde \varphi(\alpha) \right)-\left\langle \partial_t^2 \left(\tilde \varphi(\alpha)\right), \frac{\tilde \varphi(\alpha)}{\| \tilde \varphi(\alpha) \|} \right\rangle \tilde \varphi(\alpha) &=-\partial_t \left(\log g_1(r) \right) \partial_t \left(\tilde \varphi(\alpha)\right) \\ \left(\nabla^S_{\partial_t}\tilde \varphi_t\right)(\alpha) \alpha_t^2 + \tilde\varphi_t(\alpha) \alpha_{tt}-\left\langle \tilde\varphi_t(\alpha) \alpha_{tt}, \frac{\tilde \varphi(\alpha)}{\| \tilde \varphi(\alpha) \|} \right\rangle \tilde \varphi(\alpha) &=-\partial_t \left(\log g_1(r) \right) \tilde\varphi_t(\alpha) \alpha_t \\ \alpha_{tt} = \partial_t \left(\log g_1(r) \right) \alpha_t\. \end{aligned}\] With intial conditions \(\alpha(0) = 0\) and \(\alpha_t(0) = 1\) this equation has the solution \[\alpha(t) ={g_1(r_0)} \int_0^{t} \frac{1}{g_1(r(\tau))} \,d\tau\,,\] where \(r_0 = r(0)\) is the initial condition for the \(r\)-component of the geodesic. If the metric is written in the form \(\bar G = ds^2 + a(s)\langle d\varphi,d\varphi\rangle\), equation becomes \[s_{tt} = \frac {A_0^2}2 \frac{a'(s)}{a(s)^2}\,,\quad \text{ for } A_0 = a(s) \|\varphi_t\|\,,\] where \(\varphi(t)\) is given explicitly as above. This can be integrated into the form \[\tag{e}\boxed{\; s_{t}^2 =-\frac {A_0^2}{a(s)} + A_1\,,\quad A_1 \text{ a constant.} \;}\] ## Relation to hypersurfaces of revolution {#revolution} We consider the metric \(\bar G\) on \((W_-,W_+)\times S\cap C^\infty\) where \(\bar G_{s,\varphi}=a(s) \langle d\varphi, d\varphi \rangle + ds^2\) and \(a(s) = 4C_1(r(s)^2) r(s)^2\). Then the map \(\Psi\) is an isometric embedding (remember \(\langle \varphi,d\varphi\rangle=0\) on \(S\cap C^\infty\)), \[\begin{aligned} &\Psi: ((W_-,W_+)\times S\cap C^\infty, \bar G) \to \big(\mathbb R\times C^{\infty}(M,\mathbb R), du^2 + \langle df,df\rangle\big)\,, \\& \Psi(s,\varphi) = \Big(\int_0^s \sqrt{1-\frac{ a'(\sigma)^2}{4 a(\sigma)}}\,d\sigma\;,\; \sqrt{ a(s)}\varphi\Big)\,, \end{aligned}\] In fact it is defined and smooth only on the open subset \[\left\{(s,\varphi)\in (W_-,W_+)\times S\cap C^{\infty}: a'(s)^2 < 4 a(s)\right\}\.\] We will see in Sect. [1.9](#curvature){reference-type="ref" reference="curvature"} that the condition \(a'(s)^2 < 4 a(s)\) is equivalent to a sign condition on the sectional curvature; to be precise \[a'(s)^2 < 4 a(s) \Leftrightarrow \on{Sec}_{(s,\varphi)}(\on{span}(X,Y)) > 0\,,\] where \(X,Y \in T_\varphi S\) is any \(\bar G\)-orthonormal pair of tangent vectors. Fix some \(\varphi_0\in S\cap C^\infty\) and consider the generating curve \[\gamma(s) = \Big(\int_0^s \sqrt{1-\frac{ a'(\sigma)^2}{4 a(\sigma)}}\,d\sigma\;,\; \sqrt{ a(s)} \varphi_0\Big)\in {\mathbb R} \times C^\infty(M,{\mathbb R})\,;\] then \(\gamma(s)\) is already arc-length parametrized! Any arc-length parameterized curve \(I\ni s\mapsto (c_1(s),c_2(s))\) in \(\mathbb R^2\) generates a hypersurface of revolution \[\{(c_1(s), c_2(s)\varphi): s\in I, \varphi\in S\cap C^\infty\}\subset \mathbb R\times C^\infty(M,\mathbb R)\,,\] and the induced metric in the \((s,\varphi)\)-parameterization is \(c_2(s)^2\langle d\varphi,d\varphi\rangle + ds^2\). This suggests that the moduli space of hypersurfaces of revolution is naturally embedded in the moduli space of all metrics of the form \(a(s) \langle d\varphi, d\varphi \rangle + ds^2\). Let us make this more precise in an example: In the case of \(S=S^1\) and the tractrix \((c_1,c_2)\), the surface of revolution is the pseudosphere (curvature \(-1\)) whose universal cover is only part of the hyperbolic plane. But in polar coordinates we get a space whose universal cover is the whole hyperbolic plane. In detail: the arc-length parametrization of the tractrix and the induced metric are \[\begin{aligned} &c_1(s) = \int_0^s\sqrt{1-e^{-2\sigma}}\,d\sigma = \on{Arcosh}\big({e^{s}}\big)-\sqrt{1-e^{-2s}} ,\quad c_2(s) = e^{-s},\quad s> 0 \\ &a(s)\,d\varphi^2 + ds^2 = e^{-2s}d\varphi^2 + ds^2, \qquad s\in \mathbb R\. \end{aligned}\] ## Completeness In this section we assume that \((W_-, W_+) = (-\infty, +\infty)\), which is a necessary and sufficient condition for completeness. First we have the following estimate for the geodesic distance \(\on{dist}\) of the metric \(\bar G\), which is valid on bounded metric balls. Let \(\on{dist}_{S}\) denote the geodesic distance on \(S\) with respect to the standard metric. The map \(W \times \on{Id} \circ \Phi: L^2(M,{\mathbb R}) \setminus \{ 0 \} \to {\mathbb R} \times S\) is a diffeomorphism and an isometry with respect to the metrics \(\tilde G\) and \(\bar G\). It remains to consider the existence of minimal geodesics. ## Some geodesic completions {#completions} The relation to hypersurfaces of revolution in Sect. [1.6](#revolution){reference-type="ref" reference="revolution"} suggests that there are functions \(C_1\) and \(C_2\) such that geodesic incompleteness of the metric \(G\) is due to a 'coordinate singularity' at \(W_-\) or at \(W_+\). Let us write \(I=(W_-,W_+)\). We work in polar coordinates on the infinite-dimensional manifold \(I \times (S\cap C^\infty)\) with the metric \(\bar G = ds^2 + a(s)\langle d\varphi,d\varphi\rangle\). *Example.* For \(I=(0,\infty)\) the metric \(ds^2 + s^2\langle d\varphi,d\varphi\rangle\) describes the flat space \(C^\infty(M,\mathbb R)\setminus \{0\}\) with the \(L^2\)-metric in polar coordinates. Putting \(0\) back in geodesically completes the space. Moreover, for \(\beta\in (0,\pi/4]\) the metric \(ds^2 + \sin^2(\beta) s^2 \langle d\varphi,d\varphi\rangle\) describes the cone with radial opening angle \(\beta\). Putting in 0 generates a tip; sectional curvature is a delta distribution at the tip of size \(2(1-\sin(\beta))\pi\). This is an orbifold with symmetry group \(\mathbb Z/k\mathbb Z\) at the tip if \(\sin(\beta)=1/k\). More generally, \(ds^2 + K^2 s^2 \langle d\varphi,d\varphi\rangle\) describes the generalized cone whose 'angle defect' at the tip is \(2\pi (1-K)\); there is negative curvature at the tip if \(K>1\) in which case we cannot describe it as a surface of revolution. *Example.* For \(I= (0,\pi)\), the metric \(ds^2 + \sin^2(s)\langle d\varphi,d\varphi\rangle\) describes the infinite-dimensional round sphere 'of 1 dimension higher' with equator \(S\cap C^\infty\) and with north-and south-pole omitted. This can be seen from the formula for sectional curvature from Sect. [1.9](#curvature){reference-type="ref" reference="curvature"} below, or by transforming it to the hypersurface of revolution according to Sect. [1.6](#revolution){reference-type="ref" reference="revolution"}. Putting back the two poles gives the geodesic completion. To realize this on the space of densities, we may choose a smooth and positive function \(g_2(r)\) freely, and then put \[\begin{aligned} g_1(r) &= \sin^2\Big(\int_1^r g_2(\rho)^{1/2}\,d\rho\Big)\,, \quad C_1(m) = \frac{g_1(\sqrt{m})}{4m}\,, \\ C_2(m) &= \frac1{4m} g_2(\sqrt{m})-\frac1{4m^2} g_1(\sqrt{m})\. \end{aligned}\] Choosing \(g_2(r)= 4r^2\) we get \(g_1(r)=\sin^2(r^2-1)\) so that \(C_1(m)=\frac1{4m}\sin^2(m-1)\) and \(C_2(m)= 1-\frac1{4m^2}\sin^2(m-1)\). The general situation can be summarized in the following result: ## Covariant derivative and curvature {#curvature} In this section we will write \(I = (W_-,W_+)\). In order to calculate the covariant derivative we consider the infinite-dimensional manifold \(I \times S\) with the metric \(\bar G = ds^2 + a(s)\langle d\varphi,d\varphi\rangle\) and smooth vector fields \(f(s,\varphi)\partial_s + X(s,\varphi)\) where \(X(s,\;)\in \mathfrak X(S)\) is a smooth vector field on the Hilbert sphere \(S\). We denote by \(\nabla^S\) the covariant derivative on \(S\) and get \[\begin{aligned} \partial_s &\bar G\big(g\partial_s+ Y,h\partial_s + Z\big) = \partial_s\big(gh + a\langle Y,Z\rangle\big) = \\& = g_s h + g h_s + a_s\langle Y,Z\rangle + a\langle Y_s,Z\rangle + a\langle Y,Z_s\rangle \\& = \bar G\big( g_s\partial_s + \frac{a_s}{2a}Y + Y_s, h\partial_s + Z\big) + \bar G\big(g\partial_s + Y, h_s\partial_s + \frac{a_s}{2a}Z + Z_s\big) \\ X &\bar G\big(g\partial_s+ Y,h\partial_s + Z\big) = X\big(gh + a\langle Y,Z\rangle\big) \\& = dg(X).h + g.dh(X) + a\langle \nabla^S_XY,Z\rangle + a \langle Y,\nabla^S_XZ,\rangle \\& = \bar G\big(dg(X)\partial_s + \nabla^S_XY, h\partial_s+Z\big) + \bar G\big(g\partial_s+Y,dh(X)\partial_s + \nabla^S_XZ\big)\. \end{aligned}\] Thus the following covariant derivative on \(I \times S\), which is not the Levi-Civita covariant derivative, \[\begin{aligned} \bar\nabla_{f\partial_s+X}(g\partial_s+Y) = f.g_s\partial_s + f\frac{a_s}{2a}Y + fY_s + dg(X)\partial_s + \nabla^S_XY\,, \end{aligned}\] respects the metric \(ds^2 + a\langle d\varphi,d\varphi\rangle\). But it has torsion which is given by \[\begin{aligned} \on{Tor}&(f\partial_s+X,g\partial_s+Y)= \\& =\bar\nabla_{f\partial_s+X}(g\partial_s+Y)-\bar\nabla_{g\partial_s+Y}(f\partial_s+X)-[f\partial_s + X, g\partial_s + Y] = \\& = \frac{a_s}{2a}(fY-gX)\. \end{aligned}\] To remove the torsion we consider the endomorphisms \[\begin{gathered} \on{Tor}_{f\partial_s +X}, \on{Tor}_{f\partial_s +X}^{\top}: T(I \times S)\to T(I \times S)\,, \\ \on{Tor}_{f\partial_s +X}(g\partial_s+Y) = \on{Tor}(f\partial_s+X,g\partial_s + Y)\,, \\ \bar G\big(\on{Tor}_{f\partial_s +X}^\top(g\partial_s + Y), h\partial_s + Z) = \bar G\big(g\partial_s + Y,\on{Tor}_{f\partial_s +X}(h\partial_s+ Z)\big) \end{gathered}\] The endomorphism \[\begin{aligned} &A_{f\partial_s+ X}(g\partial_s+Y) := \\& = \tfrac12\big(\on{Tor}(f\partial_s+ X,g\partial_s+Y)-\on{Tor}_{f\partial_s+ X}^\top(g\partial_s+Y)-\on{Tor}_{g\partial_s+ Y}^\top(f\partial_s+X)\big) \end{aligned}\] is then \(\bar G\)-skew, so that \[\nabla_{f\partial_s + X}(g\partial_s + Y) = \bar\nabla_{f\partial_s + X}(g\partial_s + Y)-A_{f\partial_s+ X}(g\partial_s+Y)\] still respects \(\bar G\) and is now torsion free. In detail we get \[\begin{aligned} \on{Tor}_{f\partial_s+ X}^\top(g\partial_s+Y) &=-\frac{a_s}2\langle X,Y\rangle \partial_s + \frac{a_s}{2a}fY \\ A_{f\partial_s+ X}(g\partial_s +Y) &= \frac{a_s}{2}\langle X,Y\rangle\partial_s-\frac{a_s}{2a}gX\,, \end{aligned}\] so that \(\nabla\) is the Levi-Civita connection of \(\bar G\): \[\boxed{\; \begin{aligned} \nabla_{f\partial_s + X}(g\partial_s + Y) &= \big(f.g_s+ dg(X)-\frac{a_s}{2}\langle X,Y\rangle\big)\partial_s \\&\quad + \frac{a_s}{2a}(fY + gX)+ fY_s + \nabla^S_XY\. \end{aligned} \;}\] For the curvature computation we assume from now on that all vector fields of the form \(f\partial_s+X\) have \(f\) constant and \(X=X(\varphi)\) so that in this case \[\begin{aligned} \nabla_{f\partial_s + X}(g\partial_s + Y) &= -\frac{a_s}{2}\langle X,Y\rangle \partial_s + \frac{a_s}{2a}(fY + gX) + \nabla^S_XY\,, \\ [f\partial_s + X,g\partial_s + Y] &= [X,Y]^S\,, \end{aligned}\] in order to obtain \[\begin{aligned} &\nabla_{f\partial_s + X}\nabla_{g\partial_s + Y}(h\partial_s+Z) = \nabla_{f\partial_s + X} \big(-\frac{a_s}{2}\langle Y,Z\rangle \partial_s + \frac{a_s}{2a}(gZ + hY) + \nabla^S_YZ\big) \\& =\Big(-f\frac{a_{ss}}{2}\langle Y,Z\rangle-\frac{a_s}{2}\langle \nabla^S_XY,Z\rangle-\frac{a_s}{2}\langle Y,\nabla^S_XZ\rangle \\&\qquad\qquad-\frac{a_s^2}{4a}g\langle X, Z\rangle -\frac{a_s^2}{4a}h\langle X,Y\rangle -\frac{a_s}2\langle X, \nabla^S_YZ\rangle \Big)\partial_s \\&\quad + \frac{a_s^2}{4a^2}fgZ + \frac{a_s^2}{4a^2}fhY + \frac{a_s}{2a}f\nabla^S_YZ-\frac{a_s^2}{4a}\langle Y,Z\rangle X \\&\quad +\big(\frac{a_s}{2a}\big)_sfgZ + \big(\frac{a_s}{2a}\big)_sfhY + \frac{a_s}{2a}g\nabla^S_XZ + \frac{a_s}{2a}h\nabla^S_XY + \nabla^S_X\nabla^S_YZ \\-&\nabla_{[f\partial_s + X,g\partial_s + Y]}(h\partial_s+Z) =-\nabla_{[X,Y]^S}(h\partial_s+Z) \\& = +\frac{a_s}{2}\langle [X,Y]^S,Z\rangle\partial_s-\frac{a_s}{2a}h[X,Y]^S-\nabla^S_{[X,Y]^S}Z \end{aligned}\] Summing up we obtain the curvature (for general vector fields, since curvature is of tensorial character) \[\begin{aligned} &\mathcal R(f\partial_s + X,g\partial_s + Y)(h\partial_s + Z) = \\& = \big(\frac{a_{ss}}{2}-\frac{a_s^2}{4a}\big)\langle gX-fY,Z\rangle\partial_s + \mathcal R^S(X,Y)Z \\&\quad-\big(\big(\frac{a_s}{2a}\big)_s + \frac{a_s^2}{4a^2}\big)h(gX-fY) +\frac{a_s^2}{4a}\big(\langle X,Z\rangle Y-\langle Y,Z\rangle X\big)\. \end{aligned}\] and the numerator for sectional curvature \[\begin{aligned} &\bar G\big(\mathcal R(f\partial_s + X,g\partial_s + Y)(g\partial_s + Y),f\partial_s +X\big)= a\langle\mathcal R^S(X,Y)Y,X\rangle \\&\quad-\big(\frac{a_{ss}}{2}-\frac{a_s^2}{4a}\big)\big\langle gX-fY ,gX-fY\big\rangle +\frac{a_s^2}{4}\big(\langle X,Y\rangle^2-\langle Y,Y\rangle \langle X,X\rangle\big)\. \end{aligned}\] Let us take \(X,Y\in T_\varphi S\) with \(\langle X,Y\rangle=0\) and \(\langle X,X\rangle=\langle Y,Y\rangle = 1/a(s)\), then \[\begin{aligned} \on{Sec}_{(s,\varphi)}(\on{span}(X,Y)) &= \frac1a-\frac{a_s^2}{4a^2}\,, & \on{Sec}_{(s,\varphi)}(\on{span}(\partial_s,Y)) &=-\frac{a_{ss}}{2a} + \frac{a_s^2}{4a^2}\,, \end{aligned}\] are all the possible sectional curvatures. Compare this with the formulae for the principal curvatures of a hypersurface of revolution in and with the formulas for rotationally symmetric Riemannian metrics in. ## Example The simplest case is the choice \(C_1(\lambda) = \frac 1\lambda\) and \(C_2(\lambda) = 0\). The Riemannian metric is \[G_\mu(\alpha, \beta) = \frac{1}{\mu(M)} \int_M \frac \alpha\mu \frac \beta\mu \mu\.\] Then \(g_1(r) = 4\) and \(g_2(r) = \frac 4{r^2}\). This metric is geodesically complete on \(C^\infty(M,\mathbb R)\setminus \{0\}\). The geodesic equation simplifies to \[r_{tt} = \frac{r_t^2}{r}\.\] This ODE can be solved explicitely and the solution is given by \[r(t) = r(0) \exp \left( \frac{r_t(0)}{r(0)} t \right)\.\] The reparamterization map is \(\alpha(t) = t\) and thus the geodesic \[\varphi(t) = \cos\left(\| \varphi_t(0)\| t\right) + \sin\left(\| \varphi_t(0)\| t\right) \frac{\varphi_t(0)}{\| \varphi_t(0)\|}\,,\] describes a great circle on the sphere with the standard parametrization. Note that geodesics with \(r_t(0) = 0\) are closed with period \(2\pi / \| \varphi_t(0)\|\). The spiraling behaviour of the geodesics can be seen in Fig. [\[fig:sim\]](#fig:sim){reference-type="ref" reference="fig:sim"}. ## Example. By setting \(C_1(\lambda) = 1\) and \(C_2(\lambda) = 0\) we obtain the Fisher--Rao metric on the space of all densities. The Riemannian metric is \[G_\mu(\alpha, \beta) = \int_M \frac \alpha\mu \frac \beta\mu \mu\.\] In this case \(g_1(r) = 4 r^2\) and \(g_2(r) = 4\). The metric is incomplete towards 0 on \(C^\infty(M,\mathbb R)\setminus \{0\}\). The pullback metric is \[\tilde G = 4 r^2 \langle d\varphi, d\varphi \rangle + 4 dr^2\,,\] and hence geodesics are straight lines in \(C^\infty(M,{\mathbb R}) \setminus \{0\}\). In terms of the variables \((r,\varphi)\), the geodesic equation for \(r\) is \[r_{tt} = \frac{A_0^2}{16} \frac 1{r^3}\,,\] with \(A_0 = 4 r(0)^2 \| \varphi_t(0)\|\). ## Example Setting \(C_1(\lambda) = 1\) and \(C_2(\lambda) = 1\) we obtain the extended metric \[G_\mu(\alpha, \beta) = \int_M \frac \alpha\mu \frac \beta\mu \mu + \int_M \alpha \int_M \beta\.\] In this case \(g_1(r) = 4r^2\) and \(g_2(r) = 4r^2 + 4\). The geodesic equation is \[r_{tt} = \frac{A_0^2-16 r^4 r_t^2}{16 r^3 \left( 1+ r^2 \right)}\.\] The metric on \(C^\infty(M,\mathbb R)\setminus \{0\}\) is incomplete towards 0. Geodesics for the metric can be seen in Fig. [\[fig:ext_inf\]](#fig:ext_inf){reference-type="ref" reference="fig:ext_inf"}. Note that only the geodesic going straight into the origin seems to be incomplete. ## Example Setting \(C_1(\lambda) = \frac{1}{\lambda^2}\) and \(C_2(\lambda) = 0\) we obtain the metric \[G_\mu(\alpha, \beta) = \frac{1}{\mu(M)^2} \int_M \frac \alpha\mu \frac \beta\mu \mu\,,\] which is complete towards 0, but incomplete towards infinity on \(C^\infty(M,\mathbb R)\setminus \{0\}\). We have \(g_1(r) = 4/r^2\) and \(g_2(r) = 4 / r^2\). The geodesic equation is \[r_{tt} = \frac{2 r_t^2-A_0^2 r^6}{16 r}\.\] Examples of geodesics can be seen in Fig. [\[fig:ext_inf\]](#fig:ext_inf){reference-type="ref" reference="fig:ext_inf"}. Note that the geodesic ball extends more towards infinity than towards the origin.

{'timestamp': '2018-05-10T02:05:43', 'yymm': '1607', 'arxiv_id': '1607.04550', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04550'}

# Introduction Galaxy clusters are the largest virialized structures in the Universe and form via aggregation of less massive systems. During merger events, the intra-cluster medium (ICM) is heated by shocks and is believed to become turbulent. Part of the energy involved in these processes is converted into non-thermal phenomena that exhibit themselves in the radio band as halo and relic emissions. Both radio sources are diffuse cluster-scale sources with steep spectra[^1] (\(\alpha\gtrsim1\)). Radio halos are generally morphologically connected with the X-ray emission of the hosting cluster, whereas radio relics are elongated, polarized and found in cluster peripheries [e.g. @feretti12rev for an observational overview]. In particular, radio relics are believed to form at the gigantic shocks that are generated in major mergers, where cosmic ray electrons (CRe) are (re)accelerated [see @bruggen12rev; @brunetti14rev for reviews]. This scenario is supported by the arc-shaped morphologies of relics, their high level of polarization and by the fact that an increasing number of shocks have been detected at the location of radio relics [e.g. @akamatsu13systematic; @bourdin13; @shimwell15; @eckert16; @botteon16]. The main difficulty in the understanding of the origin of radio relics resides in the low Mach number (\(\mathcal{M}\lesssim 3-4\)) associated with merger shocks. The acceleration efficiency at these weak shocks is indeed expected to be small and in several cases it is in tension with the observational requirements [e.g. @markevitch05; @macario11; @kang12; @pinzke13; @kang14; @vanweeren16toothbrush; @botteon16].\ ACT-CL J0102--4915 is the most massive cluster detected in the far Universe, at a redshift of \(z=0.87\). For its extraordinary mass of \(M_{500}\sim8.8\times10^{14}\) \(\rm{M_\odot}\), it is also known with the nickname of 'El Gordo'. The cluster was firstly discovered by its strong Sunyaev-Zel'dovich (SZ) signal and later confirmed through optical and X-ray observations. The system is in a complex merger state, as revealed by the double peaked galaxy distribution and the elongated morphology of its hot (\(kT\sim15\) \(\rm{keV}\)) ICM . In the radio band, a tenuous halo and a double relic system at the cluster NW and SE X-ray boundaries were discovered .\ In this paper we report the discovery of a strong shock associated with a radio relic in 'El Gordo' cluster. In particular, our joint  and  (GMRT) analysis provides interesting insights about the origin of the relic. Throughout the paper, we assume a concordance \(\Lambda\)CDM cosmology with \(H_0 = 70\) \(\rm{km\,s^{-1}\,Mpc^{-1}}\), \(\Omega_{m}=0.3\) and \(\Omega_{\Lambda}=0.7\), in which \(1'' = 7.713\) \(\rm{kpc}\) at the cluster redshift (\(z=0.87\)). Reported uncertainties are 68%, unless stated otherwise. # Observations and data reduction ## X-ray data reduction 'El Gordo' was observed three times (ObsID: 12258, 14022, 14023) with  in ACIS-I configuration and `VFAINT` mode for a total exposure time of 360 \(\rm{ks}\). We carried out the standard data reduction by using `CIAO 4.7` and  `CALDB 4.6.9`. In particular, soft proton flares were inspected analyzing the light curves extracted from the S3 chip in the 0.5-2 \(\rm{keV}\) band for each ObsID and removed using the `lc_clean` routine. We then used the `merge_obs` task to make the final 0.5-2 \(\rm{keV}\) cleaned image (340 \(\rm{ks}\)) shown in Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}a.\ We created a single exposure-corrected point spread function (PSF) map with minimum size for the merged image by combining the PSF and exposure maps of each ObsID. Once the PSF of the instrument is known, the `wavdetect` task allows to identify discrete sources in the surface brightness (SB) image of the cluster. These were detected using wavelet radii of 1, 2, 4, and 8 pixels, confirmed by eye and excluded in the SB profile analysis. In order to create a single background image, the `reproject_event` task was used to match the background templates to the corresponding event files for every ObsID. This single background image was normalized by counts in the band 9.5-12 \(\rm{keV}\) and subtracted during the SB analysis.\ Dealing with spectral analysis of low SB sources as in the case of cluster outskirts requires a detailed treatment of the astrophysical and instrumental background emission. In this respect, we modeled the sky component due to the Galactic emission with two thermal plasmas with \(kT_1 = 0.14\) \(\rm{keV}\) and \(kT_2 = 0.25\) \(\rm{keV}\), the cosmic X-ray background with an absorbed power-law with photon index \(\Gamma=1.4\) and the ACIS-I particle background by using the analytical approach prosed by. Spectra were extracted in the same region for every ObsID and simultaneously fitted in the 0.5-11 \(\rm{keV}\) energy band with the package `XSPEC v12.9.0o`. Since the low X-ray count rate, we kept the metal abundance of the `APEC` model, which accounts for the ICM thermal emission, fixed at the value of 0.3 \(\rm{Z_\odot}\) (solar abundance table by @anders89) and used Cash statistics during the fits. ## Radio data reduction Archival GMRT 610 \(\rm{MHz}\) observations of 'El Gordo' (project code \(22\_001\), PI: R. R. Lindner) taken on 26 August 2012 were analyzed using the Astronomical Image Processing System (`AIPS`). The GMRT Software Backend was used to record the parallel polarization products RR and LL with a bandwidth of 33.3 \(\rm{MHz}\) divided into 256 channels. The source 3C48 was used for flux and bandpass calibration and the calibrator \(0024-420\) was used for phase calibration towards the target. The total on-target observing time was 170 minutes. Standard steps of flagging (excision of bad data) and calibration were carried out. The resulting calibrated visibilities towards the target were split and used for imaging. A few rounds of phase-only self-calibration and a round of amplitude and phase self-calibration were carried out to improve the sensitivity of the image. The final image with visibilities weighted according to `ROBUST 0` in the task `IMAGR` and resolution \(11'' \times 4.8''\) (position angle \(4.8\[^\circ\)) is presented in Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}b. The image was corrected for the GMRT primary beam using the task `PBCOR`. The off-source noise level is 50 \(\rm{\mu Jy\,b^{-1}}\) and a 10% error on the absolute flux calibration scale was assumed.\ As a preliminary result of new GMRT data, we also used observations taken at 327 \(\rm{MHz}\) (project code \(25\_023\), PI: R. Kale) to perform spectral analysis. The complete analysis of the new radio dataset will be presented in the forthcoming paper (Kale et al., in preparation). # Results ## X-ray/radio analysis {#ch:x-radio} 'El Gordo' X-ray emission remarkably recalls the famous 'Bullet' cluster : a dense cool core (\(kT\sim 6\) \(\rm{keV}\)) is moving in the SE-NW direction producing a prominent cold front which is expected to follow a shock wave. The cluster is elongated along the merger direction and presents a couple of X-ray tails that give to the system a comet-like morphology (Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}a).\ Our 610 \(\rm{MHz}\) radio image of 'El Gordo' recovers extended emission better than previously done by as we considered baselines down to 0.2 \({\rm k\lambda}\) (instead of 0.48 \({\rm k\lambda}\)). This allows to study the morphology of the diffuse sources in more detail. In our image shown in Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}b, the prominent and elongated radio halo connects a pair of radio relics, located in opposite directions at the NW and SE edges of the cluster X-ray emission (Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}c). The strongest part of the halo coincides with the disrupted cluster core, whereas a radio tail appears to remarkably follow the northern tail visible in the X-rays (Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}c).\ Our work is focused on the NW radio relic, whose flux densities at 610 and 327 \(\rm{MHz}\) are \(F_{610} = 27.5\pm2.8\) \(\rm{mJy}\) and \(F_{327} = 64.6\pm6.6\) \(\rm{mJy}\), respectively. These result in a radio spectral index \(\alpha=1.37\pm0.20\). The flux density of the relic at 2.1 \(\rm{GHz}\) measured by with the  implies \(\alpha\sim1.5\) from 2100 to 327 \(\rm{MHz}\), which is consistent with what we estimated in the narrower frequency range. Nevertheless, we will use the spectral index from 610 to 327 \(\rm{MHz}\) since it is taken from two high sensitivity images obtained from GMRT observations with matched inner-*uv* coverage (*uv*\(_{min}\) = 0.2 \({\rm k\lambda}\)). ## Relics and shocks Double relics have been observed in different systems and are believed to form in mergers between two clusters with similar mass where diametrically opposite shocks move outwards along the merger axis (re)accelerating particles . pointed out the possible presence of a couple of shocks by analyzing a \(60\:\rm{ks}\)  unsharp-masked image of 'El Gordo'. For these reasons we created the unsharp-masked images shown in Fig. [\[fig:un-mask\]](#fig:un-mask){reference-type="ref" reference="fig:un-mask"} and searched for sharp edges in the X-ray SB image, identifying at least one discontinuity in the cluster. We used `PROFFIT v1.3.1` to extract the SB profiles in the red sectors shown in Fig. [\[fig:sectors\]](#fig:sectors){reference-type="ref" reference="fig:sectors"}, where the NW relic stands out. An underlying broken power-law density profile is usually adopted to describe the SB in presence of a discontinuity. In the case of spherical symmetry, the downstream density \(\rho_d\) is higher by a factor of \(\mathcal{C} \equiv \rho_d/\rho_u\), with \(\rho_u\) upstream density, at the shock putative distance \(r_{sh}\). In formula \]\label{eq:break-pl} \begin{array}{ll} \rho_d (r) = \mathcal{C} \rho_0 \left( \frac{r}{r_{sh}} \right)^{a_1}, & \mbox{if} \quad r \leq r_{sh} \\ \\ \rho_u (r) = \rho_0 \left( \frac{r}{r_{sh}} \right)^{a_2}, & \mbox{if} \quad r > r_{sh} \end{array}\[ where \(\rho_0\) is the density normalization, \(a_1\) and \(a_2\) are the power-law indices and \(r\) is the radius from the center of the sector (in Fig. [\[fig:sectors\]](#fig:sectors){reference-type="ref" reference="fig:sectors"} RA: \(+15^\circ.7275\), DEC: \(-49^\circ.2724\), J2000). We used this density shape to fit the X-ray SB keeping all parameters of the model free to vary.\ We firstly report results concerning sector 1+2 (opening angle (OA): \(30^\circ-98^\circ\)) because it covers the whole extension of the feature shown in Fig. [\[fig:un-mask\]](#fig:un-mask){reference-type="ref" reference="fig:un-mask"} and it gives the maximum SB drop with the best statistics (a discussion on the sector choice is presented in Section [3.2.1](#ch:syste){reference-type="ref" reference="ch:syste"}). In Fig. [\[fig:sb_profile\]](#fig:sb_profile){reference-type="ref" reference="fig:sb_profile"} we report the best broken power-law model fit, which is in excellent agreement with data. We detect a large SB drop, corresponding to a density compression factor \(\mathcal{C}=3.4^{+0.4}_{-0.3}\), co-spatially located with the relic. For a shock, the Rankine-Hugoniot (RH) jump conditions for a monatomic gas \]\label{eq:mach-from-dens} \mathcal{C} \equiv \frac{\rho_d}{\rho_u} = \frac{4\mathcal{M}_{\rm{SB}}^2}{\mathcal{M}_{\rm{SB}}^2 + 3}\[ would lead to a Mach number \(\mathcal{M}_{\rm{SB}}=4.1^{+3.4}_{-0.9}\). A \(\mathcal{M}>3\) shock is quite a rarity in galaxy clusters and so far only two of them have been detected (\(\mathcal{M}=3.0\pm0.4\) in the 'Bullet' cluster, @markevitch06; \(\mathcal{M}=3.0\pm0.6\) in A665, @dasadia16a665). Shocks heat the downstream gas and this allows to distinguish them from cold fronts, other kinds of SB discontinuities found in galaxy clusters with inverted temperature jumps. For this reason we performed spectral analysis in the yellow sectors shown in Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}c. Spectra for sector 1+2 are reported in Fig. [\[fig:spectra\]](#fig:spectra){reference-type="ref" reference="fig:spectra"}. We found evidence of a very high downstream temperature, \(kT_d=17.9^{+3.3}_{-2.8}\) \(\rm{keV}\), while only a lower limit to the upstream one was obtained, \(kT_u>6.1\) \(\rm{keV}\). In principle, this is not enough to confirm the shock nature of the discontinuity but, similarly to the E shock in the 'Bullet' cluster, the presence of a cold front is very unlikely because it would imply a too high temperature (\(kT_u > 20\) \(\rm{keV}\)) at such a large cluster distance.\ Although current data do not allow to measure a temperature jump at the position of the shock, we can use the lower limit to such a jump to provide independent constraints on the shock Mach number. According to RH conditions, the upstream and downstream temperature are related by \]\label{eq:mach-from-temp} \frac{T_d}{T_u} = \frac{5\mathcal{M}_{\rm kT}^4 + 14\mathcal{M}_{\rm kT}^2 -3}{16\mathcal{M}_{\rm kT}^2}\[ which implies \(\mathcal{M}_{\rm{kT}}<2.9\) if we insert the upper \(1\sigma\) limit of \(T_d\) and the lower limit on \(T_u\). We anticipate that this value is consistent with the Mach number inferred from SB jump once systematic errors are taken into account (see the following Section)\ A visual inspection of Fig. [\[fig:gordo_cluster\]](#fig:gordo_cluster){reference-type="ref" reference="fig:gordo_cluster"}a suggests the presence of a drop in SB also at the position of the SE relic. A shock in this region is expected due to the presence of the radio relic and in analogy with the 'Bullet' cluster. However, current data do not allow us to characterize statistically the SB drop because of the low X-ray counts in this region. Nevertheless we found evidence for high temperature in the putative downstream gas, \(kT_d=30.1^{+10.5}_{-6.2}\) \(\rm{keV}\) in the yellow sector in the SE (Fig. [\[fig:sectors\]](#fig:sectors){reference-type="ref" reference="fig:sectors"}), somewhat supporting this possibility. Since typical temperatures in cluster outskirts are of a few keV, such an high \(kT_d\) would likely imply a \(\mathcal{M}\gtrsim3.5-4\) shock. ### Systematic errors on X-ray analysis {#ch:syste} Results in the previous Section are based on measurements obtained for a particular sector (1+2). This entirely covers the feature found in the unshap-mask images (Fig. [\[fig:un-mask\]](#fig:un-mask){reference-type="ref" reference="fig:un-mask"}) and allows the best characterization of the SB jump due to the statistics of the fit.\ We checked the impact due to the choice of the SB extracting region in the determination of the NW X-ray discontinuity and the resulting Mach number. Firstly, we re-performed SB and spectral analysis by splitting the red and yellow sectors of Fig. [\[fig:sectors\]](#fig:sectors){reference-type="ref" reference="fig:sectors"} in two sub-regions; the dashed line distinguishes between sector 1 (OA: \(60^\circ.5-98^\circ\)), which is oriented in the N direction, to sector 2 (OA: \(30^\circ-60^\circ.5\)), which is in the NW direction and better overlaps the relic. In both regions, the SB profile is well described by a compression factor \(\mathcal{C} \gtrsim 3\), implying \(\mathcal{M}_{\rm{SB}} \gtrsim 3\).\ We then repeated the SB analysis by excluding data at \(r<2'\) and keeping the discontinuity distance frozen at the values found in the \(1.2'-4.9'\) radial range. Although with a large error, spectral analysis allowed to constrain the upstream temperature in sector 1, implying a 68% confidence interval estimate for the Mach number \(\mathcal{M}_{\rm{kT}}=1.9-3.4\) (taking into account the asymmetric errors on the two temperatures), whereas only lower limits to \(kT_u\) can be obtained in sectors 1+2 and 2. The results of the fits obtained for the three regions are summarized in Tab. [\[tab:sectors-sb\]](#tab:sectors-sb){reference-type="ref" reference="tab:sectors-sb"}.\ Finally, we checked the variation on \(\mathcal{M}_{\rm{SB}}\) in sector 1+2 due to different shock curvature radii from the best fit value found in Fig. [\[fig:sb_profile\]](#fig:sb_profile){reference-type="ref" reference="fig:sb_profile"} (\(r_{curv} \sim 1\) \(\rm{Mpc}\)). Results are reported in Tab. [1](#tab:curvature){reference-type="ref" reference="tab:curvature"} and the impact of \(r_{curv}\) on the shock compression factor is presented in Fig. [\[fig:r_curv\]](#fig:r_curv){reference-type="ref" reference="fig:r_curv"}.\ Spectral analysis requires a careful determination of the background sources and its systematic uncertainties. In this respect, we varied background normalization levels within \(\pm1\sigma\) and re-performed spectral fits. We achieved results consistent with the reported cluster parameters within \(1\sigma\). Nonetheless, we highlight that the measurement of high temperatures is critical with  given its low effective area at energies higher than \(5\) \(\rm{keV}\), in particular the estimated confidence range may not reflect entirely the true statistical and systematic error range.\ As a final test, a more complex model of a two-temperature thermal plasma was adopted to fit the downstream spectra. In this case, the high-\(T\) component is not constrained while the low-\(T\) component gives unreasonably low temperatures (e.g. \(kT_{high}>21.2\) \(\rm{keV}\) and \(kT_{low}=2.2^{+4.0}_{-1.2}\) \(\rm{keV}\), for sector 1+2). As pointed out in the case of the Coma cluster, the low-\(T\) component mitigates the fit residuals at low energy rather than describing a physical condition. The high-\(T\) spectral component instead supports the presence of a high temperature plasma in the downstream region. ## Constraints on the downstream magnetic field {#ch:ic} Relativistic electrons scattering with the cosmic microwave background (CMB) photons is expected to produce inverse Compton (IC) emission. From the ratio between radio and X-ray emission it is possible to constraints the magnetic field in the source region . For this reason we preformed spectral analysis in a region enclosing the NW relic and introduced, in addition to the canonical thermal model for the ICM, a power-law in the spectral fit.\ We assume that the IC spectrum is a power-law with photon index related to the synchrotron spectral index via \(\Gamma=\alpha+1\). Initially we set \(\Gamma=2.37\) (see Section [3.1](#ch:x-radio){reference-type="ref" reference="ch:x-radio"}) and kept it frozen in the fit while thermal parameters were free to vary. In this case we obtain 0.5-2 \(\rm{keV}\) upper limit to the non-thermal component \(F_{[0.5-2\:\rm{keV}]} < 6.76 \times 10^{-15}\) \(\rm{erg\,s^{-1}\,cm^{-2}}\).\ The IC measurement is a very complicated issue and can be influenced by several factors. We investigated the impact on IC flux estimation by: using the IC power-law slope in the range \(2.17-2.57\) (consistently with the values reported in Section [3.1](#ch:x-radio){reference-type="ref" reference="ch:x-radio"}), keeping the temperature frozen at 13.5 and 17 \(\rm{keV}\) (which covers a range of \(kT\) obtained for different sector choices in the relic region), varying background normalization levels within \(\pm1\sigma\) and re-performing the fits in the 0.7-11 and 0.9-11 \(\rm{keV}\) energy bands. In summary, we found upper limits in the range \((2.95-8.51) \times 10^{-15}\) \(\rm{erg\,s^{-1}\,cm^{-2}}\) for the IC flux, representing \(\lesssim50\)% of the thermal model flux in the same energy band (0.5-2 \(\rm{keV}\)). However, in the case of temperature frozen at 17 \(\rm{keV}\), the fits formally result in a IC detection both for \(\Gamma=2.17\) and 2.57; we do not consider these detections solid enough due to the systematics related to the presence of multi-temperature components and background characterization. We further explore the possibility of IC detection in Section [4.2](#ch:ic-bump){reference-type="ref" reference="ch:ic-bump"}.\ By using our upper limits to IC flux, we conservatively obtained[^2] the following lower limits to the downstream magnetic field strength: \(B\geq3.1\) \(\rm{\mu G}\) for \(\Gamma= 2.17\), \(B\geq4.9\) \(\rm{\mu G}\) for \(\Gamma= 2.37\) and \(B\geq7.6\) \(\rm{\mu G}\) for \(\Gamma= 2.57\). These values are in line with other estimates for radio relics [e.g. @bonafede09double; @finoguenov10; @vanweeren10; @vanweeren11zwcl0008]. ## Acceleration efficiency The relic--shock connection is nowadays supported by many observational studies. Nevertheless, theoretical models of relic formation are challenged by the low Mach numbers associated with cluster shocks. In fact, the commonly adopted diffusive shock acceleration (DSA) model is severely challenged for weak shocks if CRe are accelerated by the thermal pool [e.g. @brunetti14rev for review]. A connected problem is the ratio of cosmic ray protons and electrons that would be generated at these shocks and that current \(\gamma\)-ray  limits constrain at values that are significantly smaller than that in supernova remnants.\ If the downstream synchrotron luminosity emitted at frequency \(\geq \nu_0\) originates from electrons in steady state conditions, the bolometric (\(\geq \nu_0\)) synchrotron luminosity that is generated via shock acceleration from a shock with speed \(V_{sh}\) and surface \(S\) can be estimated as \]\label{eq:luminosity} \int_{\nu_0} L(\nu)\,d\nu \simeq {1 \over 2 } \eta_e \Psi \rho_u V_{sh}^3 \left( 1-{1 \over{\mathcal{C}^2}} \right) {{ B^2 }\over{B_{cmb}^2 + B^2}} S\[ where \(\eta_e\) is the efficiency of electron acceleration, \]\label{eq:psi} \Psi = {{\int_{p_{min}} N_{inj}(p) E\,dp}\over{\int_{p_0} N_{inj}(p) E\,dp}}\[ accounts for the ratio of the energy flux injected in "all" electrons and those visible in the radio band (\(\nu \geq \nu_0\)), \(p_0\) is the momentum of the relativistic electrons emitting the synchrotron frequency \(\nu_0\) in a magnetic field \(B\) and \(B_{cmb} = 3.25(1 + z)^2\) \(\rm{\mu G}\) accounts for IC scattering of CMB photons. The injection spectrum of accelerated CRe is given by \]\label{eq:n_inj} N_{inj}(p) = (\delta_{inj} +2 ) p^{-\delta_{inj}} \int_{p_{min}}^{p} x^{\delta_{inj}-1} N_u(x)\,dx\(\) where \(N_u\) is the spectrum of seed particles upstream. In the case of CRe acceleration from the thermal pool this is \(N_u \propto p^{-\delta_{inj}}\), where \(\delta_{inj} = 2 (\mathcal{M}^2+1)/(\mathcal{M}^2-1)\), implying a synchrotron spectral index \(\alpha=\delta_{inj}/2\) in the case of steady state conditions.\ In Fig. [\[fig:eff\]](#fig:eff){reference-type="ref" reference="fig:eff"} we report the acceleration efficiency that is necessary to explain the radio luminosity observed in the NW relic assuming DSA of thermal electrons upstream.\ We assumed shock Mach numbers \(\mathcal{M}=3.5\) and 2.5, in line with the values derived from the X-ray analysis. These Mach numbers would imply a synchrotron spectrum of the relic \(\alpha = 1.18\) and \(1.38\), respectively, that are in line with radio measurements (Section [3.1](#ch:x-radio){reference-type="ref" reference="ch:x-radio"}). Calculations were obtained assuming upstream number density and temperature \(2.4 \times 10^{-4}\) \(\rm{cm^{-3}}\) and 6.1 \(\rm{keV}\), respectively, and a surface of the relic \(S=\pi \times 350^2\) \(\rm{kpc^2}\).\ In Fig. [\[fig:eff\]](#fig:eff){reference-type="ref" reference="fig:eff"} we also show the lower limits to the magnetic field in the relic that are derived from the upper limits to the IC flux assuming the two values of the spectral index (Section [3.3](#ch:ic){reference-type="ref" reference="ch:ic"}).\ Despite we are dealing with a high-velocity shock, \(V_{sh} \sim 4000\) \(\rm{km\,s^{-1}}\), we note that the efficiency of CRe acceleration that is requested to explain the radio relic is large. This is due to the fact that the NW relic in the 'El Gordo' is one of the most luminous radio relics known so far and because, for few \(\rm{\mu G}\) magnetic fields, most of the CRe energy is radiated via IC emission (due to the high redshift of the cluster). Still, contrary to the case of weaker shocks (see e.g. A754, @macario11; 1RXS J0603.3+4214, @vanweeren16toothbrush; A115, @botteon16), we conclude that in this case DSA of thermal electrons is still a viable option. Indeed for Mach number \(\geq 3.5\) the electron acceleration efficiency appears energetically viable \(\eta_e \leq 0.01\), whereas for \(\mathcal{M}\sim 3-3.5\) additional mechanisms of pre-acceleration of thermal electrons downstream (see @guo14a [@guo14b]) may be required.\ The other possibility is that the NW relic is due to shock re-acceleration of seeds (relativistic or supra-thermal) electrons. In this case the efficiency necessary to explain the radio emission is much smaller simply because the bulk of the energy is channelled directly into highly relativistic particles (Eq. [\[eq:n_inj\]](#eq:n_inj){reference-type="ref" reference="eq:n_inj"}, e.g. @markevitch05 [@kang12]). Seeds can be broadly distributed in the cluster outskirts where the life time of 100 \(\rm{MeV}\) electrons is very long (e.g. @pinzke13 [@donnert16arx]) or they can be in radio ghost/clouds generated by past AGN activity . The two possibilities have different predictions on the upstream synchrotron emission that in principle can be tested with very deep radio observations, but that are well beyond the aim of our paper. # Discussion ## Overall considerations 'El Gordo' is a high redshift (\(z=0.87\), @menanteau12) and high mass (\(M_{500}\sim8.8\times10^{14}\) \(\rm{M_\odot}\), @planck14xxix) galaxy cluster. It is the most distant massive cluster with the brightest X-ray and SZ emission and the farthest hosting diffuse radio sources (halo and relics). Our study makes it is also the most distant cluster where a shock (with one of the highest Mach number) has been detected.\ Optical and X-ray observations revealed that 'El Gordo' is in a merging state . Recent numerical simulations were able to reproduce the overall system properties assuming a nearly head on major merger .\ Double relic systems are expected to trace shocks moving outwards in cluster outskirts. So far, studies on 'El Gordo' were mainly focused on the SE relic located in front of the dense cool core, which is expected to follow a shock in analogy with the well know 'Bullet' cluster case. However, current X-ray data do not allow to characterize the jump in this region because of the low statistics. We instead discovered a \(\mathcal{M}\gtrsim3\) shock spatially coincident with the NW relic. Our detection is based on the  SB jump. Although with large uncertainties, spectral analysis is also consistent with the presence of a strong shock in the ICM. Further indications of the shock are given by the unsharp-masked images of Fig. [\[fig:un-mask\]](#fig:un-mask){reference-type="ref" reference="fig:un-mask"}. We also mention the striking similarity between 'El Gordo' shock/reverse-shock and X-ray morphology with A2146 [@russell10; @russell12] even though the latter is a less massive system (\(M_{500}\sim3.8\times10^{14}\) \(\rm{M_\odot}\), @planck14xxix) and does not host any radio relics at the sensitivity level of current observations . ## Hints of inverse Compton emission? {#ch:ic-bump} The search for IC emission from galaxy clusters has been undertaken for many years with many instruments. However, no confirmed detection has been obtained so far. The most famous dispute in this field regards the case of the nearby Coma cluster [@rephaeli99; @rephaeli02; @fusco99; @fusco04; @fusco07; @rossetti04; @wik09; @wik11; @gastaldello15]. The excellent spatial resolution and good spectral capabilities of the  satellite allow to minimize the contamination from the thermal X-ray emission and open to the possibility to search for non-thermal emission also in the soft X-ray band .\ 'El Gordo' is a perfect target to search for IC emission. It hosts a bright radio relic in an external region with low thermal SB and it is at high redshift (\(z=0.87\)), where the equivalent CMB magnetic field strength is large, \(B_{cmb}=3.25(1+z)^2=11.4\) \(\rm{\mu G}\).\ We used the deep  observations to look for IC signatures in the NW relic (Section [3.3](#ch:ic){reference-type="ref" reference="ch:ic"}). In the case the temperature is frozen at 17 \(\rm{keV}\), the spectral analysis provides a formal detection and a significant fraction (up to \(40-50\)%) of the X-ray brightness across a relic would be contributed by IC emission from the relic itself. This should be visible in the X-ray images and profiles across the relic. For this reason we extracted a SB profile in the 0.5-2.0 \(\rm{keV}\) across a narrow sector (OA: \(37^\circ.2-67^\circ.5\)) containing the NW radio relic. The fit of a broken power-law model (Eq. [\[eq:break-pl\]](#eq:break-pl){reference-type="ref" reference="eq:break-pl"}) in this restricted region provides a good description of the SB jump, as shown in Fig. [\[fig:ic-bump\]](#fig:ic-bump){reference-type="ref" reference="fig:ic-bump"} (dashed red line), leading to \(\chi^2 /\rm{d.o.f.} = 48.5/41\). However, a SB excess is present in the region of the relic. For this reason we also attempt to fit the SB by adding a Gaussian component to the downstream power-law. The addition of this Gaussian improves the fit (Fig. [\[fig:ic-bump\]](#fig:ic-bump){reference-type="ref" reference="fig:ic-bump"}, solid blue line) with \(\chi^2 /\rm{d.o.f.} = 40.3/38\). Slight different sector centers and apertures do not influence this excess. The Gaussian component coincident with the relic could represent an excess due to IC emission from electrons in the relic. The excess flux associated with the Gaussian component is in line with that expected from the spectral analysis. The combination of this excess with the formal detection of IC emission obtained from the spectral analysis (with \(kT\) frozen) is tantalizing, however deeper observations (i.e. \(>500\) \(\rm{ks}\)) are required to firmly conclude about this possible detection. # Conclusions We presented an X-ray/radio study of the famous 'El Gordo' cluster located at \(z=0.87\) focusing on the non-thermal activity in the cluster.\ Our GMRT radio observations at 610 and 327 \(\rm{MHz}\) confirmed the presence of a halo and a system of double relics. These represent the most distant diffuse radio sources detected in a galaxy cluster so far. The halo is quite elongated in the NW-SE, i.e. in the merger direction, and remarkably follows the ICM emission of the northern X-ray tail. The two relics are found at the boundaries of the X-ray emission. We focused on the NW relic which has a synchrotron spectral index \(\alpha=1.37\pm0.20\) between 610 and 327 \(\rm{MHz}\).\ The deep  observations (340 \(\rm{ks}\)) allowed us to discover a shock at the position of the NW relic. The SB profile taken is this region abruptly drops at the relic location. The density compression factor \(\mathcal{C} \gtrsim 3\) and the high downstream temperature provide the indication of a strong shock (\(\mathcal{M} \gtrsim 3\)) in the ICM. This is one of the three strongest shocks detected in galaxy clusters and the most distant (\(z=0.87\)) observed so far.\ The detection of a shock co-spatially located with a relic strongly supports the relic--shock connection. The NW shock in 'El Gordo' cluster allows to study particle acceleration in a rare regime of strong shock. We found that DSA of thermal electrons is consistent with measured synchrotron spectrum. Nonetheless, only shocks with \(\mathcal{M} > 3.5\) appear energetically viable while for weaker shocks re-acceleration models would be preferred.\ The presence of relativistic particles emitting a bright synchrotron relic at \(z=0.87\) makes 'El Gordo' a suitable cluster candidate to search for IC emission from the relic. From the X-ray spectral analysis we obtained possible hints for IC emission from the relic, however we could not firmly conclude the presence of IC excess and conservatively we derived only lower limits to the downstream magnetic field that have been used to improve constraints on particle acceleration. However, we also found hints of an excess in the 0.5-2 \(\rm{keV}\) SB profile across the relic region. The combination of a possible IC excess in the spectral analysis with the hints of excess in the SB is tantalizing and certainly deserves deeper  observations.

{'timestamp': '2016-08-19T02:07:22', 'yymm': '1607', 'arxiv_id': '1607.04641', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04641'}

# Introduction Adaptation is often viewed as a cumulative succession of fixations of beneficial mutations, and the waiting time between two successive beneficial mutations is viewed as one of the main factor limiting adaptation rate. This is however not true in general since several beneficial mutations can compete in a single population, which generates different interference phenomena, such as the Muller's ratchet or more generally Hill-Robertson interfences. Competition between several beneficial mutations is especially expected in large asexual populations with high mutation rates and adaptation is then expected to be slowed down, a phenomenon called "clonal interference". Clonal interference and its consequences on adaptation rates seems general as they have been observed in bacteria, viruses, yeasts or cancer tumors.\ Despite the ubiquity and importance of clonal interference, experiments showed that adaptation is not blocked and keep on even after thousands generations. These observations led more recently to the development of an alternative view of clonal interference, considering that beneficial mutations can occur in a single lineage, thus increasing fitness gradually. In large populations with high mutation rates, it is expected that adaptation follows a wave bounded by the most and the least fit lineages. Adaptation rate is expected to be constant in this model and depends only on the width of the wave, in other words on the genetic variability present in the population.\ Models of clonal interference have been useful in understanding the dynamics of adaptation in asexual populations, and have known many empirical successes. However, many experiments have shown that the dynamics of beneficial mutations can be much more complex. Different experiments showed that dynamics can be non-linear: lineages can show multiple frequency peaks during the course of adaptation and different lineages can coexist for a long time in a single population. This suggests that frequency-dependent fitness can occur. Moreover, cooperation between lineages and niche construction have been observed multiple times in tumoral cancers and bacteria, which suggests that taking into account different ecological interactions might be important for a global understanding of the evolution of asexual populations with large population size and high mutation rates.\ Clonal interference has been mostly theoretically studied using population genetics models with strong assumptions: beneficial mutations are assumed to have simple epistatic interactions, the effect of mutations on fitness is assumed to be transitive, selection is assumed not to depend on the mutants frequency, and the environment is assumed constant. Non-transitive fitness and non-linear dynamics are consequently considered as special cases, of less important interest. There is however a large literature dealing with Lotka-Volterra models with more than two species, especially with non-transitive competitive interactions. These studies try to define the conditions for the coexistence of multiple competitive species in a community. Such studies are generally deterministic and investigate the conditions of stability of dynamical systems. When dealing with stochastic dynamics, simulations are generally performed in spatialized context again in order to determine the conditions for the coexistence of multiple competing species. Those studies do not however investigate the probability and time of fixation of mutations and are thus of limited interest regarding the effect of competitive interactions between several clones in the course of adaptation, especially on the adaptation rate.\ Here we propose a stochastic model with three different lineages under competition, where the competitive interactions are not necessarily transitive, thus relaxing one important assumption of the models studying clonal interference so far. Our model embraces a large variety of phenomena observed in the course of adaptation of asexual species. We recover classical results from clonal interference models and we also show that unexpected behaviours are expected. For instance, we show that in some cases competitive interactions between three clones can lead to a higher rate of adaptation. Our results generally show that non-linear dynamics are likely in large clonal populations, which challenges the interpretation of experimental results.\ # Model and methods ## Definitions and assumptions We denote \(i \in [0,1,2]\) a type of individuals (types can be phenotypes, alleles, strains, clonal species, mutants, etc.). For the sake of simplicity, we will use in the rest of the paper the term *mutant* \(i\) when referring to type \(i\) individuals. \(N(t)=(N_i(t))\) is a vector whose elements are the number of mutants \(i\) in the population at time \(t\), with \(N_i(t)\) a random variable. We assume that the environment has a fixed quantity of available resources: the carrying capacity \(K\) modulates the intensity of competition between individuals. We investigate the population dynamics of three clonal types as a birth-death process with competition in continuous time. Each mutant \(i\) is characterized by its individual ecological parameters: \(\beta_i\) and \(\delta_i\) are respectively the individual birth and natural death rates, and \(C_{ij}\) is the effect of competition of a single mutant \(j\) on a single mutant \(i\), assuming \(C_{ij} \geq 0\) and \(C_{ii}>0\). For simplicity, we assume that competition between individuals affects mortality. The individual death rate of a mutant \(i\) thus depends on both an intrinsic component (\(\delta_i\)) and a component due to competition: \(d_i(N(t))=\delta_i+C_{i0} N_0(t)/K +C_{i1}N_1(t)/K+C_{i2}N_2(t)/K\).\ We suppose that the resident population is only composed of mutants \(0\), at its ecological steady-state equilibrium, say \(N_0(0)\). A single mutant \(1\) is introduced in the population (\(N_1(0)=1\)). The population of mutants \(1\) follows a stochastic dynamics that depends on the ecological parameters and the competitive interactions between mutants \(0\) and mutants \(1\). The time taken for mutants \(1\) to invade the resident population, and eventually get fixed, is of order \(\log K\). Since the dynamics is stochastic, mutants \(1\) can either spread or be lost. Since we are interested in the dynamics of three competing clones, we will focus on cases where neither mutants \(0\) nor mutants \(1\) are lost when a single mutant \(2\) enters the population by mutation (or migration). We assume that the time at which this event occurs is \(\alpha \log K\), \(\log K\) being the time scale of the whole stochastic dynamics. There are two general cases. Either \(\alpha\) is low enough that mutants \(1\) are still in too few numbers to affect the invasion of mutants \(2\), or \(\alpha\) is large enough that mutants \(1\) have invaded the population and thus affect the invasion of mutants \(2\). We will investigate both situations and show that the time \(\alpha \log K\) at which mutant \(2\) enters the population is crucial and largely affect the final state of the population.\ ## The stochastic dynamics is a succession of several phases When the population is large, the dynamics followed by the population can be divided into a succession of two kinds of phases. First, when a mutation enters a population, say a mutant \(j\) enters in an single copy in a \(i\) resident population, the dynamics of mutants \(j\) is well approximated by a branching birth-death process without interactions until its population size \(N(j)\) is large enough, i.e. when it is of order \(K\). In the case of the joint dynamics of three interacting clones, the dynamics of two mutants \(j\) and \(k\) in a resident population \(i\) is also well approximated by a branching birth-death process without interactions (see the proofs in the companion paper ). Second, when two or three mutants have a population size of order \(K\), then the stochastic dynamics of these populations is well approximated by a competitive Lotka-Volterra deterministic system. During this phase, if the population size of one of the mutants is of order lower than \(K\), then its population size essentially does not change until the deterministic equilibrium is reached. Even though the whole dynamics is stochastic, we will respectively call these two kinds of phases "stochastic" and "deterministic", for the sake of simplicity (see Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}). The dynamics of the population can finally be described as a succession of "stochastic" and "deterministic" phases. Note that we only give in this paper the relevant biological results and a summarized version of the model (mathematical proofs and detailed computations are given in a companion paper ).\ **Deterministic phases.** We denote by \(n_i\) the size of the population of mutants \(i\) when dealing with the deterministic dynamics, while we will keep the notation \(N_i\) when dealing with stochastic dynamics (\(n_i\) is the population size rescaled by the carrying capacity \(N_i / K\) when \(N_i\) is of order \(K\)). When the population size of the three mutants are of order \(K\), the dynamics of the rescaled process can be well approximated by the following system of ordinary differential equations, \[\label{EDO} \left\{\begin{array}{ll} \dot{n}_0=(\beta_0-\delta_0-C_{0,0}n_0-C_{0,1}n_1-C_{0,2}n_2)n_0,\\ \dot{n}_1=(\beta_1-\delta_1-C_{1,0}n_0-C_{1,1}n_1-C_{1,2}n_2)n_1,\\ \dot{n}_2=(\beta_2-\delta_2-C_{2,0}n_0-C_{2,1}n_1-C_{2,2}n_2)n_2. \end{array} \right.\] Under the assumptions that competitive parameters are \(C_{ij} \geq 0\) and \(C_{ii}>0\) for all \(\{i,j\}\), this system of deterministic equations is a three-dimensions Lotka-Volterra competitive model. Such a three species population shows different possible dynamics: different equilibrium states (either monomorphic or polymorphic, with two or three coexistent mutants), or stable limit cycles.\ The fate of a mutant \(i\) entering in a single copy a resident \(j\) population is associated to the so-called "invasion fitness", denoted \(S_{ij}=\beta_i-\delta_i-C_{ij} \bar{n}^j\), where \(\bar{n}^j=\frac{\beta_j-\delta_j}{C_{jj}}\) is the population size of mutant \(j\) at equilibrium when there are only mutants \(j\) in the population. The invasion fitness corresponds to the initial growth rate of the mutant when it is rare. If the resident population is composed of both mutants \(i\) and \(j\) at equilibrium, then the fate of a mutant \(k\) entering in a single copy is associated with the invasion fitness denoted \(S_{kij}=\beta_k-\delta_k-C_{ki} \bar{n}^{i}_{ij}-C_{kj} \bar{n}^{j}_{ij}\), where \[\label{eqpop} \begin{array}{ll}\\ \bar{n}^i_{ij}=\frac{C_{jj}(\beta_i-\delta_i)-C_{ij}(\beta_j-\delta_j)} {C_{ii}C_{jj}-C_{ij}C_{ji}}, \bar{n}^j_{ij}=\frac{C_{ii}(\beta_j-\delta_j)-C_{ji}(\beta_i-\delta_i)}{C_{ii}C_{jj}-C_{ij}C_{ji}}, \end{array}\] is the equilibrium of Eq. [\[EDO\]](#EDO){reference-type="ref" reference="EDO"} when there are only mutants \(i\) and \(j\) in the population. If \(S_{kij}>0\), mutation \(k\) is favorable when rare in the polymorphic resident population \((i,j)\) and can invade.\ **The stochastic phase.** When the population size of at least one mutant, say \(i\), is low (*i.e.* its population size is of order lower than \(K\)), while the other mutants are at their deterministic steady state (their population size is of order \(K\)), the dynamics of mutants \(i\) is close to a pure birth-death process with birth and death rates respectively \(\beta_i\) and \(\delta_i+\sum_{j \neq i}C_{ij}n_j\), thus neglicting the effect of the competition between individuals \(i\). When a mutant \(i\) enters a resident \(j\) population in a single copy, the probability of invasion of the mutant \(i\), defined as the probability that the population of mutants \(i\) reaches a size of order \(K\), is \(S_{ij}/\beta_i\) when \(S_{ij}>0\), and a probability 0 if \(S_{ij}\leq 0\). The time taken by a mutant \(i\) which enters a \(j\) resident population to reach the threshold population size is of order \(\log K / S_{ij}\). We can similarly define the probability of invasion of a mutant \(k\) in a resident population with both mutants \(i\) and \(j\) as \(S_{kij}/\beta_k\) when \(S_{kij}>0\), and a probability 0 if \(S_{kij}\leq 0\). The time taken by a mutant \(k\) which enters a \(i\) and \(j\) resident population to reach the threshold population size is of order \(\log K / S_{kij}\). In both cases, the stochastic phase ends when the mutant is either lost or reaches a threshold population size of order \(K\). ## Stochastic dynamics and final states with three competitive clones When a single favorable mutant enters a resident population, the stochastic dynamics can be decomposed into three successive phases (see Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}a and [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}b): First, a stochastic phase corresponding to the beginning of the mutant invasion and where the mutant has a population size of order lower than \(K\); Second, a deterministic phase when the population size of the new mutant is large enough (of order \(K\)); Third, a new stochastic phase until the resident mutant is lost (if both mutants do not stably coexist at deterministic equilibrium). In the case of competition between three clones with two mutations entering a resident population, the succession of stochastic and deterministic phases must be decomposed into a higher and not limited number of phases.\ **What determines the different dynamics and final states when there is competition between three clones?** The different dynamics and final states depend on the ecological parameters and also on the following conditions (see details of computation in Appendix A1 and illustrations in Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"} and [\[fig:clonalreinforcement\]](#fig:clonalreinforcement){reference-type="ref" reference="fig:clonalreinforcement"}, simulation algorithm given in Appendix A3):\ i) Does mutant \(2\) enter the population during the first (Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}a) or second (Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}b) stochastic phase (depending on the time \(\alpha \log K\))? If mutant 2 enters the population during the first stochastic phase then mutant 1 has a population size of order lower than \(K\). Consequently, mutant 2 suffers the competitive effect of mutants 0 only. If mutant 2 enters the population during the second stochastic phase then, assuming mutant 1 is favorable in the resident population, there are two possibilities: either mutants 0 and 1 stably coexist, in which case mutant 2 suffers the competitive effects of both mutants, either mutant 0 has a population size of order lower than \(K\), in which case mutant 2 suffers the competitive effects of mutant 1 only;\ ii) When mutant 2 enters the population during the first stochastic phase (Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}a), does mutant 1 or 2 reaches first a population size of order \(K\)? Since mutants 1 and 2 have a population size lower than order \(K\), the speed at which they invade the resident population only depends on their competitive interactions with the resident mutants 0. Hence, which mutants reaches first a population size of order \(K\) depends on their invasion fitness \(S_{10}\) and \(S_{20}\) and on the time when mutant 2 enters the population \(\alpha \log K\). The first mutant which reaches a population size of order \(K\) determines the initial state of the succeeding deterministic phase;\ iii) What is the equilibrium of the first deterministic phase: stable coexistence of two mutants, i.e. two mutants have a population size of order \(K\), or a single mutant has a population size of order \(K\)? This only depends on the sign of the invasion fitnesses (Eq. [\[EDO\]](#EDO){reference-type="ref" reference="EDO"}). For instance, mutants 0 and 1 stably coexist if \(S_{01}>0\) and \(S_{10}>0\).\ iv) What is the population size of all mutants when the second stochastic phase begins? It depends on whether two mutants stably coexist or not at the end of the deterministic phase (step iii), and on the population size of the mutant which did not invade and still has a size lower than order \(K\);\ v) Does a mutant go extinct before the start of the next deterministic phase? When a mutant has a population size of order lower than \(K\) and is deleterious in a given context, it is expected to go extinct. However, its time to extinction can take longer than the time for another rare mutation to reach a population size of order \(K\). In this case, a new deterministic phase begins. The ecological context of the deleterious mutation can change before it goes extinct, which changes its fate.\ vi) Steps ii-v are again applied for the further succeeding phases (when applicable) as often as necessary until a final steady state is reached.\ **The different possible final states and their hitting times.** Our goal is to investigate how clonal interference might affect the dynamics of mutant populations and thus adaptation. We will thus especially focus on cases where mutation \(1\) and \(2\) have a positive invasion fitness when 2 enters the population during the first stochastic phase ( \(S_{10}>0\)and \(S_{20}>0\)), and on cases where mutation \(1\) has a positive invasion fitness( \(S_{10}>0\)) if mutation 2 enters the population during the second stochastic phase (in the latter case, the invasion fitness of mutant 2 depends on the currect state of the population). All the possible final states and their hitting times are compiled in Tables [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"} and [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}. We do not give all detailed calculations for all cases here, we only give one detailed example in Appendix A2 as an illustration (for complete and detailed computations see the companion paper ). Roughly, two classes of final states are possible: either one mutant goes to fixation (it can be either 0, 1 or 2), or two or three mutants coexist (in all possible combinations). ## Likelihood of the final states We want to estimate the likelihood of the different possible final states assuming ecological parameters are drawn in given prior distributions. The complexity of the model can be reduced to: \(\rho_i=\beta_i-\delta_i\), the net individuals reproductive rate of mutants \(i\), and \(\widetilde{C}_{ij}=\frac{C_{ij}}{C_{jj}}\) the ratio of the between and within competitive interactions. We drew \(10^6\) different sets of parameters in prior distributions. For a given set of parameters and a given \(\alpha\), the final state is given by Tab. [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"} and [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}, which allows to estimate the posterior distribution of the final states among the \(10^6\) random sets of parameters.\ The time \(\alpha \log K\) at which the second mutation enters the population has a large impact on the final states. We can determine the final states for a given parameter set for any \(\alpha\) using the results of our model (Tab. [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"} and [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}). Assuming mutation 2 enters the population during the first stochastic phase, we know that mutant 2 necessarily appears before mutant 1 spreads out, i.e. \(\alpha<1/S_{10}\). If mutation 2 enters the population during the second stochastic phase, mutant 2 necessarily appears after mutant 1 spreads out and before mutant 0 goes extinct, i.e. \(1/S_{10}<\alpha<1/S_{10}+1/|S_{01}|\). We also know that there are two threshold values \(\alpha < 1/S_{10}+1/S_{20}(S_{21}/|S_{01}|-1)\) and \(\alpha < S_{02} S_{21}/(S_{10} |S_{12}| |S_{01}|)-1/S_{01}\) (Tab. [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"}) which determine which dynamics is followed by the population when mutation 2 enters the population during the first stochastic phase. Similarly, there are two threshold values \(1/S_{10}<\alpha < 1/S_{10}+1/|S_{01}|-1/S_{21}\) and \(1/S_{10}<\alpha <1/S_{10}+S_{02}/(|S_{12}||S_{01}|)-1/S_{21}\) (Tab. [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}) which determine which dynamics is followed by the population when mutation 2 enters the population during the second stochastic phase (see above). We can thus finally compute the probability of any final states given an ecological parameter set and assuming \(\alpha\) is uniformly drawn in the interval \(\left[ 0, 1/S_{10}\right]\) or \(\left[ 1/S_{10}, 1/S_{10}+1/|S_{01}|\right]\) respectively when mutation 2 enters the population during the first or second stochastic phase.\ **Effect of mutations on the reproductive rates.** In bacteria, yeasts or some eukaryotes, fitness is generally estimated as the initial growh rate (at low density) of mutants (see Table 2 in and the Appendix in ). We will thus assume that the effect of mutations on the growth rate of mutant \(i\) follows a Fisher's geometric model. Given the net reproduction rate of mutants 0 is \(\rho_0\), we assumed that the reproductive rate of mutant \(i\) is \(\rho_i=\rho_0 + x_i\) with \(x_i\), the effect of mutation \(i\), being drawn in a shifted negative Gamma distribution which is an approximation of a Fisher's geometric model for adaptation. Note that when mutation 2 enters the population during the second stochastic phase, mutation 2 is assumed to occur in the most frequent mutation at equilibrium: \(\rho_2=\rho_1 + x_2\) when mutant 1 is more frequent than mutant 0, \(\rho_2=\rho_0 + x_2\) otherwise.\ **Effect of mutations on competition.** There is, to our knowledge, no theoretical or empirical consensus on the distribution of mutation effects on the competitive abilities \(\widetilde{C}_{ij}\). Without any knowledge about the distribution of competitive abilities, we simply assumed that the ratio of competitive interaction \(\widetilde{C}_{ij}\) follows arbitrary chosen distributions. First, we assumed that it follows a uniform distribution in the interval \(\left[1-\mu, 1+\mu \right]\), with \(0 \leq \mu \leq 1\). Second we assumed it follows an exponential distribution with parameter \(\mu>0\). When \(\mu=0\), all \(\widetilde{C}_{ij}=1\), fitnesses are necessarily transitive, while if \(\mu>0\), frequency-dependent fitnesses can occur. As \(\mu\) increases, the variance of the competitive ratio \(\widetilde{C}_{ij}\) also increases, i.e. the more different can the competitive interactions be between mutants.\ # Results Given that two favorable mutations successively enter in a single copy a resident population 0, Tables [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"} and [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"} show all possible dynamics and final states, depending on the sign of the invasion fitnesses \(S_{ij} \text{ and } S_{ijk}, \{i,j,k\} \in \{0,1,2 \}\), and the time of appearance of the second mutation \(\alpha \log K\). When the second mutation enters the population during the first stochastic phase, Tab. [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"} shows that six final states are possible: fixation of either mutation 1 or 2, stable polymorphic equilibrium with two (mutants 0 and 1, 1 and 2 or 0 and 2) or three mutants (mutants 0, 1 and 2). Tab. [\[tableass1\]](#tableass1){reference-type="ref" reference="tableass1"} also shows that a given possible final state can be reached under different conditions, and consequently after different durations. For instance, mutant 2 can fix under three different cases (a subcase of B, and cases E and I), with potentially different fixation times. When mutation 2 enters the population during the second stochastic phase, Tab. [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"} shows that a seventh final state is possible: the fixation of mutant 0, even if mutants 1 and 2 are advantageous when they enter the population. Interestingly, Rock-Paper-Scissors cyclical dynamics can only occur if the second mutation enters the population during the second stochastic phase. Assuming three co-occuring competing clones, our model can thus capture a large diversity of dynamics and allows to determine their duration, final states and likeliness. In the following, we first show that despite the complexity and variety of the possible stochastic dynamics with three competing clones, six possible general dynamics can be defined. Second, we focus on several special cases of particular interest. We especially argue that our model, despite its simple assumptions, captures a large range of dynamics diversity observed in experiments. ## Beyond clonal interference: Six possible dynamics Two categories of dynamics have been proposed in the empirical literature to explain observations in experiments with several interacting clones: clonal interference, when adaptation is slowed down because of the interaction between advantageous mutations, or clonal reinforcement, also called niche construction or frequency-dependent selection elsewhere, when several clones stably coexist. Using our model, focusing on the second mutation (mutant 2), we can embrace the two previous proposed categories, and propose an alternative categorization, with more accurate definitions.\ Competition between three clones can affect adaptation for three reasons: 1) It can promote or hinder polymorphism maintenance; 2) The invasion probability of mutation 2 can be increased or decreased (relatively to the case where mutant 2 enters alone the resident population, i.e. there are only two interacting clones); 3) In cases where mutation 2 goes to fixation, its fixation time can be be shorter or longer. Hence, dynamics with three competing clones can be classified in six general cases: When clonal interaction promotes polymorphism maintenance, we call the dynamics "clonal coexistence". When the mutant 2 goes to fixation, we call "clonal assistance" when the duration of the sweep is shorter, and "clonal interference" (following ) when it is larger than with only two competing clones. Finally, we call "soft" vs. "hard" the dynamics depending on wheter the invasion probability of the mutant 2 is lower vs. higher than with only two competing clones. This gives 6 possible general dynamics, summarized in Table [\[tab:cat\]](#tab:cat){reference-type="ref" reference="tab:cat"}.\ **Rate of adaptation: fixation time vs. probability of invasion of beneficial mutants.** Clonal interference is viewed in the literature as the phenomenon of the increase in fixation time of co-occuring beneficial mutants in a population, which consequently decreases the rate of adaptation of clonal species. However, the rate of adaptation can be affected also by the rate at which new mutants invade a population, i.e. by their probability of invasion. To our knowledge, in all models dealing with clonal interference so far, derived from population genetics models, the probability of invasion of a new mutant only depends on its own features: in general, a selection coefficient \(s\) is arbitrarily assigned to a mutant independently of the composition and state of the resident population when this mutant occurs. Its invasion probability and its time of fixation are approximately \(2s\) and \(1/s\): increasing \(s\) necessarily both increases the probability of fixation and decreases the time of fixation. In our approach, we have a more general point of view: both the fixation time and probability of invasion depend on the state of the population when a mutant enters the population. We thus argue that fixation times and probabilities of invasion should be considered independently in order to evaluate to which extent interaction between various clones can affect rate of adaptation. Interestingly, depending on the state of the population when mutation occurs and on their own ecological specificities, the rate of adaptation can increase: the probability of invasion can be higher or the time to fixation can be shorter, what we propose to call "hard clonal assistance".\ **Clonal coexistence: cooperative interactions are not necessary.** Several experiments of competition between clones have shown stable persistence of different strains in a single well-mixed population, which has been explained by frequency-dependent selection, niche construction or cooperative interactions. Especially, introduced the concept of "clonal reinforcement" when "the emergence of one genotype favors the emergence and persistence of other genotypes via cooperative interactions". Here we show that clones can favor either the emergence ("hard" vs. "slow" dynamics), or the persistence ("clonal coexistence" vs. "clonal assistance" or "clonal interference") or both ("hard clonal coexistence"), without any cooperative interactions between clones, but only competition. We do not argue that persistence observed in are not effectively due to cooperative interactions, rather we propose an alternative hypothesis: both facilitated emergence and stable persistence of clones can be due to non-transitive competitive interactions, or frequency-dependent selection. It would need specific experimental work to show whether or not clones effectively cooperate. ## Two specific dynamics when the second mutant lately enters the population: Rock-Paper-Scissors or annihilation of adaptation Two related specific dynamics are encountered only when mutant 2 enters the resident population during the second stochastic phase (\(\alpha>1/S_{10}\), Tab. [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}): Rock-Paper-Scissor dynamics (final state J in Tab. [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}), or a return to the initial state, i.e. a population fixed for mutant 0 (final state G in Tab. [\[tableass2\]](#tableass2){reference-type="ref" reference="tableass2"}). These two different dynamics are illustrated in Fig. [\[fig:phases\]](#fig:phases){reference-type="ref" reference="fig:phases"}b and Fig. [\[fig:clonalreinforcement\]](#fig:clonalreinforcement){reference-type="ref" reference="fig:clonalreinforcement"}b. Both dynamics have the same parameters, except the time at which mutant 2 enters the population \(\alpha \log K\). This illustrates the importance of considering stochastic dynamics: if mutant 2 enters the population early enough that mutant 1 is not extinct when mutant 2 invades, then Rock-Paper-Scissors cyclical dynamics take place, otherwise mutant 0 goes to fixation and adaptation is annihilated despite the occurrence of two beneficial mutations. In a deterministic model, for the same parameters, a mutant can not go extinct and only Rock-Paper-Scissor dynamics is possible.\ Our results also show that Rock-Paper-Scissors dynamics can be obtained in a narrow set of parameters. In addition to specific parameters values regarding the competitive interactions between the three clones, the second mutant must occur in the population in a narrow time frame. First, it must occur after mutant 1 invaded, since mutant 2 is deleterious in a mutant 0 resident population. Second, if mutant 2 occurs too late during the second stochastic phase, then mutant 0 can be extinct before mutant 2 invades, in which case mutant 1 goes to fixation. These results have important consequences regarding our understanding of empirical Rock-Paper-Scissor dynamics observed in natural populations: either the three types of individuals involved in such stable cycles have effectively entered the population by mutation or migration in a single individual, in which case the third type of individuals has necessarily entered the population in a narrow time frame. Otherwise, the alternative explanation is that the three types of individuals went together in a single population with a sufficiently large enough population size such that the dynamics initially followed an almost-deterministic dynamics, which certainly occurred by a massive migration and mixing of three different and complementary types of individuals ## Likehood of the final states assuming prior distributions of the ecological parameters Figures [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"} and [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"} show the posterior probability of the dynamics and final states when mutation 2 enters the population during the first and second stochastic phases, assuming that the competition abilities are respectively drawn in an uniform or exponential distribution. When the variance of the distribution of the \(\widetilde{C}_{ij}\) is low, all clones have similar competitive abilities (\(\widetilde{C}_{ij} \simeq 1\)), *i.e.* invasion fitness are mostly transitive. We naturally recover predictions from population genetics models: The likeliest scenari are the fixation either of mutant 1 or 2 (Fig. [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"}c, [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"}d, [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"}c, [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"}d). Rapidly, when the variance of the uniform and exponential distributions increases, polymorphic final states become the likeliest. When the effect of mutation on competitive abilities become large, the likeliness of all dynamics rapidly reaches a plateau when competitive abilities (\(\widetilde{C}_{ij}\)) are drawn in an uniform distribution (Fig. [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"}). When competitive abilities are drawn in an exponential distribution, the likeliest state is the one with three coexisting clones. Our results suggest that non-transitive fitness are mostly expected to occur when several clones are interacting as soon as mutations affect their competitive abilities. This further supports that clonal coexistence is likely to occur even when considering only competitive interactions: cooperative interactions are not necessary to explain the stable coexistence of several clones. Finally, our results show that Rock-Paper-Scissors dynamics and annihilation of adaptation are weakly probable.\ Comparing left and right columns in Fig. [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"} and [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"} shows that the time at which mutation 2 enters the population only marginally affects the dynamics and the final states. Interestingly, comparing the final states between cases with two or three interacting clones (Fig.[\[fig:Distrib2clones\]](#fig:Distrib2clones){reference-type="ref" reference="fig:Distrib2clones"}) shows that more polymorphic final states are expected when three clones are interacting, even though the difference is small. Whether a further increase in the number of interacting clones could even more promote the maintenance of polymorphism because of non-transitive fitness. Fig. [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"}, [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"} and [\[fig:Distrib2clones\]](#fig:Distrib2clones){reference-type="ref" reference="fig:Distrib2clones"} also show that the prior distribution of the competitive abilities has important effect: polymorhic final states are more expected when the distribution is exponential.\ Finally, Fig. [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"}e-f and [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"}e-f show the likeliness of clonal interference *vs.* clonal assistance. When competitive abilities follow an uniform distribution, clonal interference, *i.e.* the slowing down of fixation of beneficial mutation, is the most probable, especially when the competitive abilities are similar between clones (small \(\mu\)). However, rapidly when the difference between competitive abilities increases (large \(\mu\)) the likeliness of clonal assistance increases, and reaches a plateau. When mutation 2 enters the population during the second stochastic phase, clonal assistance is even likelier than clonal interference. The situation is different and more complex when competitive abilities follow an exponential distribution, especially when the second mutation enters the population during the second stochastic phase. Contrarily to the uniform distribution, the exponential distribution is asymetrical around the mean, which generates asymetrical invasion fitness distributions and explains the difference observed in our results. Indeed, when the mean \(\mu\) is low, the distribution of the invasion fitness is skewed towards negative values, while it is skewed towards positive values when \(\mu\) is large (not shown), which explains why polymorphic states are the likeliest for large \(\mu\). Globally, our results thus suggest that clonal interference might indeed be an important factor affecting adaptation rate, but clonal assistance can be as important given non-transitive fitnesses are possible. # Discussion Both theory and experimental observations tend to agree regarding the rate of adaptation in evolving large asexual populations: the speed at which new beneficial mutations go to fixation should decrease during the course of adaptation and reach a plateau, what is effectively found in experimental evolution of bacteria. Clonal interference is believed to be the most important factor causing this limit to adaptation: different beneficial mutations occurring simultaneously in competing lineages tend to mutually decrease their probability of fixation and increase their time to fixation, what has been effectively observed in viruses, bacteria or cancer. This congruence between models and data can be challenged regarding our results. Indeed, by explicitly taking into account competitive interactions between different clones, we showed that the evolutionary dynamics can be much more complex: competitive interactions between clones can either impede or foster adaptation by modifying fixation times and probabilities, and can promote coexistence of clones in the long-term. Our results are different from previous works because in our model non-transitive fitnesses are allowed, giving rise to frequency-dependent selection, non-linear and cyclical dynamics. Our model is in fact more general than previous works since they are a subcase where all competitive interaction abilities are equal (\(\widetilde{C}_{ij}=1\)). Most importantly, we showed that frequency-dependent selection is very likely when mutations have an effect on competitive interaction abilities (Fig. [\[fig:Distrib1\]](#fig:Distrib1){reference-type="ref" reference="fig:Distrib1"} and [\[fig:Distrib2\]](#fig:Distrib2){reference-type="ref" reference="fig:Distrib2"}), giving rise to almost equiprobable dynamics with either clonal interference or clonal assistance, i.e. respectively a decrease or an increase in the rate of adaptation.\ We might thus wonder why a decreasing rate of adaptation is generally observed in experimental evolution. In other words, why does adaptation in clonal species mostly follow a single subcase of our model where all competitive interaction abilities are almost equal (\(\widetilde{C}_{ij} \simeq 1\)). Several hypotheses can be made. First, as Lenski and colleagues claimed regarding their long-term experimental evolution, the experiments were designed to avoid frequency-dependent selection: no horizontal gene transfer, a well-mixed environment with no spatial structure and a low concentration of the density-limiting resource. One can thus imagine that these experimental conditions were effectively fulfilled and their goal is achieved. However, several experiments showed that even in ideal conditions, frequency-dependent selection occurs. Hence, even if the experimental design limits the occurrence of mutations with non-transitive effects on fitness, it does not obliterate it. A question thus remains: why is frequency-dependent selection so rare in experimental evolution? Second, we only considered three interacting clones in our model, while much more can be competing in a large clonal population. To our knowledge, it is not known whether an increasing number of competing clones decreases the plausibility of non-transitive interactions. However, we showed that non-transitive interactions are more probable with three than two interacting clones (Fig. [\[fig:Distrib2clones\]](#fig:Distrib2clones){reference-type="ref" reference="fig:Distrib2clones"}). Deterministic models with four interacting clones have also shown that even more complex dynamics can be expected, especially chaotic ones. This suggests that a large number of interacting clones could not be the main cause for the rarity of non-transitive interactions. Third, our estimation of the likeliness of frequency-dependent selection are based on two strong hypotheses: the effect of mutations of competitive interaction abilities follows either an uniform or exponential distribution, and a mutation shows no trade-offs between its effect on the growth rate and the competitive abilities. To our knowledge there is neither empirical nor theoretical treatments about the distribution of the competitive interactions and the trade-offs between species or clones. estimated strength of frequency dependent selection in a limited range of frequencies in a well-mixed environment. They especially showed that non-transitive interactions can have similar extent than transitive interactions. detected no intransitive interactions between clones three pairs of bacterial clones in a simple experimental sets, suggesting that competitive interactions are weak. Yet, a general treatment is lacking in order to evaluate the potential importance of non-transitive interactions in adaptive populations. Whether non-transitive interactions can be expected to be frequent or not is an important question since it can challenge the generality of the evolutionary dynamics observed in controlled experimental evolution of clonal species.\ Non-linear dynamics have been observed in various experiments and different explanations and interpretations have been proposed. observed different dynamics in different replicates. They especially observed cases where a lineage showed two successive frequency peaks. They explained this observation by the occurrence of a third cryptic mutation affecting a preexisting lineage, what they called "multiple mutations" dynamics. We have shown in our model that such a dynamics can be explained with three interacting clones only: such cyclical dynamics can be observed in cases of Rock-Paper-Scissor. Importantly, we show that such cyclical dynamics can only be observed if the second mutation occurs during the second stochastic phase, i.e. when the first mutation is near fixation. It would be interesting to look thoroughly into the data to challenge this prediction., and also showed the possible coexistence of clones in the long-term, which can effectively be explained by negative frequency-dependent selection, as proposed here. and showed the long-term coexistence of three lineages derived by mutation from a single initial *Escherichia coli*clone, in a long-run experimental evolution in a chemostat. They interpreted this coexistence as due to cooperative rather than competitive interactions, yet we showed that non-transitive competitive interactions can explain the emergence by mutation and stable coexistence of three clones. showed that the three clones can stably coexist by pairs. Under our framework, it means that \(S_{ij}>0\) for all \(\{i,j\} \in \{ 0,1,2\}\), and we predict that the three clones should indeed coexist because of non-transitive competitive interactions. The three clones consume the same resources (glucose, acetate, glycerol,, ) supporting the existence of competitive interactions. However, looking thoroughly at the data, it is not obvious that one of the clones is maintained in coexistence when in competition with another: its frequency is less than 1% and sometimes even not detected in the experiments (see Fig. 1d in ). Taking an opposite point of view than the authors of the experiments, we can conclude that mutant 2 is not maintained when in competition with mutant 1 (*i.e.* \(S_{12}<0\)). In this case, our model predicts that the three clones can not stably coexist, and consequently cooperation between the three clones can be an exclusive explanation for the observed coexistence.

{'timestamp': '2016-07-19T02:00:32', 'yymm': '1607', 'arxiv_id': '1607.04656', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04656'}

# Introduction ## Overview Distributing points on \(\mathbb{S}^2:=\left\{x\in\mathbb{R}^3: \ |x|=1\right\}\) is a classical problem arising in many settings in numerical analysis and approximation theory, in particular the study of radial basis functions, quadrature and polynomial interpolation, Quasi-Monte Carlo methods for graphics applications, finite element methods for PDE's, cosmic microwave background radiation modeling, crystallography and viral morphology, to name a few. The goal of this paper is to survey some widely used algorithms for the generation of spherical node sets. We will restrict our descriptions to "popular\" point sets, most of which can be generated "reasonably fast.\" Namely, we study - Fibonacci and generalized spiral nodes - projections of low discrepancy nodes from the unit square - zonal equal area nodes and HEALPix nodes - polygonal nodes such as icosahedral, cubed sphere, and octahedral nodes - minimal energy nodes - maximal determinant nodes - random nodes and - "mesh icosahedral equal area nodes.\" The last is a new point set devised to have many desirable properties. For each of the above configurations, we provide illustrations, and analyze several of their properties. We focus our attention primarily on - equidistribution - separation - covering - quasi-uniformity and - Riesz potential energy. For each property we provide numerical calculations, tables, and comparisons, and in some cases we prove theoretical bounds on the mesh ratio. Section 3 is devoted to asymptotic comparisons of various potential energies. We do not consider quadrature of the point sets; however, such a comparison for several of the configurations we describe here can be found in. We now formally introduce the properties we will be studying and, in Section 2, describe the point sets themselves. We leave the technical proofs to Section 4, and at the end of the document, we provide resources for Matlab source codes to generate the point sets. ## Definitions and Properties For low error numerical integration with respect to uniform surface area measure ( and ) as well as in digitizing \(\mathbb{S}^2\) for computer graphics purposes ( and ), it is important for any spherical configuration to have an approximately uniform distribution. A sequence \(\left\{\omega_N\right\}_{N=1}^{\infty}\) of spherical point sets with \(\omega_N\) having cardinality \(N\) is called *equidistributed* if the sequence of normalized counting measures, \[\nu_N(A):=\frac{1}{N}|A\cap\omega_N|,\ \ \ \ \ \ A\ \textup{Borel set},\] associated with the \(\omega_N\)'s converges in the weak-star sense to \(\sigma\), the normalized surface area measure on \(\mathbb{S}^2\), as \(N\rightarrow \infty\). That is, for all continuous functions \(f\) on \(\mathbb{S}^2\), \[\lim_{N\to\infty}\int_{\mathbb{S}^2}f\,\text{d}\nu_N = \int_{\mathbb{S}^2}f\,\text{d}\sigma.\] An equivalent definition is that the \(L_\infty\)-spherical cap discrepancy \[D_C(\omega_N):=\sup_{V\subset \mathbb{S}^2}\bigg|\frac{|V\cap \omega_N|}{N}-\sigma(V)\bigg|\to 0 ,\ \ \ \ \ \ \ N\to\infty,\] where the supremum is taken over all spherical caps \(V\subset \mathbb{S}^2\). For the study of local statistics, separation and covering properties play an important role. The *separation* of a configuration \(\omega_N\subset \mathbb{S}^2\) is \[\delta(\omega_N):=\min_{\substack{x,y\in\omega_N\\ x\neq y}} |x-y|,\] and a sequence of spherical \(N\)-point configurations is said to be *well-separated* if for some \(c>0\) and all \(N\geq 2\), \[\delta(\omega_N)\geq cN^{-1/2}.\] The *covering radius* of \(\omega_N\) with respect to \(\mathbb{S}^2\) is defined to be \[\eta(\omega_N):= \max_{y\in \mathbb{S}^2} \min_{x\in \omega_N} |x-y|,\] and a sequence of spherical \(N\)-point configurations is a *good-covering* if for some \(C>0\) and all \(N\geq 2\), \[\eta(\omega_N)\leq CN^{-1/2}.\] A sequence of configurations \(\left \{\omega_N\right \}_{N=2}^\infty\) is said to be *quasi-uniform* if the sequence \[\bigg\{\gamma (\omega_N) := \frac{\eta (\omega_N)}{\delta (\omega_N)}\bigg\}_{N\geq 2}\] is bounded as \(N\rightarrow \infty.\) The quantity \(\gamma(\omega_N)\) is called the *mesh ratio* of \(\omega_N\). Note that some authors define the mesh ratio as \(2\gamma(\omega_N)\). A sequence of \(N\)-point configurations is quasi-uniform if it is well-separated and a good-covering. We remark that equidistribution does not imply quasi-uniformity or vice versa. In applications involving radial basis functions, "1-bit\" sensing, and finite element methods (,, , and ), there is interest in precise bounds on \(D_C(\omega_N)\), \(\delta(\omega_N),\eta(\omega_N)\), and \(\gamma(\omega_N)\). A trivial lower bound is \(\gamma(\omega_N)\geq 1/2\) for any configuration. Asymptotically, as proved in, \[\gamma(\omega_N) \geq \frac{1}{2\cos\pi/5} + o(1) = \frac{\sqrt{5}-1}{2} +o(1), \ \ \ \ N \rightarrow \infty,\] for any sequence of configurations \(\left \{\omega_N\right \}_{N=2}^\infty \subset \mathbb{S}^2\). We also evaluate the potential energy of our point sets. The problem of minimizing point energies on the sphere dates to at least the beginning of the 20th century when Thomson put forth a model of the ground state configurations of electrons in. Given a lower-semicontinuous, symmetric kernel \(K:\mathbb{S}^2\times\mathbb{S}^2\rightarrow(-\infty,\infty]\), and a spherical configuration \(\omega_N\subset\mathbb{S}^2\), the \(K\)-energy of \(\omega_N\) is defined to be \[\text{E}_K(\omega_N):=\sum_{\substack{x,y\in\omega_N\\ x\neq y}} K(x,y). \label{energy_def}\] The infimum of \(\text{E}_K(\omega_N)\) over all \(N\)-point configurations on \(\mathbb{S}^2\) is attained and is denoted by \(\mathcal{E}_K(N)\). We will restrict our attention to the class of Riesz kernels defined by \[\ \ \ \ \ \ \ \ \ \ \ K_s(x,y) = \frac{1}{|x-y|^s},\ \ \ \ \ \ \ \ s > 0\] \[K_{\text{log}}(x,y) = \text{log}\frac{1}{|x-y|},\] \[\ \ \ \ \ \ \ \ \ \ \ K_s(x,y) =-|x-y|^{-s},\ \ \ \ \ \ \ \ s < 0.\] For brevity, the energy and minimal energy quantities for the Riesz \(s\)-kernel and log kernel will be denoted by \(\text{E}_s(\omega_N)\), \(\text{E}_{\text{log}}(\omega_N)\), \(\mathcal{E}_s(N)\), and \(\mathcal{E}_{\text{log}}(N)\) respectively. Determining an exact minimal configuration for a fixed \(N\) and \(s\) is a highly nonlinear optimization problem. In practice, gradient descent and Newton methods are used to arrive at approximate global minima; however, there is substantial interest in generating nearly optimal points more quickly . The *Voronoi cell* of a point \(x\in\omega_N\subset\mathbb{S}^2\) is the spherical polygon \[V_x(\omega_N):= \left\{y\in\mathbb{S}^2:\ |y-x| \leq |y-z|, \forall \ z\in\omega_N\setminus \left\{x\right\}\right\}.\] The *Voronoi decomposition* of a configuration is \[V(\omega_N):=\left\{V_x(\omega_N)\right\}_{x\in\omega_N}.\] It has been observed that the Voronoi cell decomposition of nearly optimal energy configurations appears to consist primarily of nearly regular spherical hexagons mixed with "scars\" of spherical heptagons and pentagons as shown in Figure [\[MinEnVoronoi\]](#MinEnVoronoi){reference-type="ref" reference="MinEnVoronoi"}. It has been conjectured that for \(s>2\), the dominant term in the asymptotic expansion of \(\mathcal{E}_s(N)\) is related to the Epstein-Zeta function for the hexagonal lattice (see Section 3). We conclude this section by introducing some secondary properties of spherical configurations that we will use: A partition \(P_N := \left\{W_i\right\}_{i=1}^N\) of \(\mathbb{S}^2\) into \(N\) cells whose pairwise intersections have \(\sigma\)-measure \(0\) is *equal area* if \(\sigma(W_i) = 1/N\) for all \(1\leq i\leq N\). A sequence of partitions \(\left \{P_N\right \}_{N=2}^\infty\) of \(\mathbb{S}^2\) such that each \(P_N\) has \(N\) cells is *diameter bounded* if there are constants \(c,C > 0\) such that for all \(N\in \mathbb{N}\) and for every cell \(W^N_i\in P_N\), \[cN^{-1/2}\leq \text{diam }W^N_i \leq CN^{-1/2}, \label{diambdd}\] where diam\((A):=\sup_{x,y\in A}|x-y|\). We will call a sequence of partitions of \(\mathbb{S}^2\) *asymptotically equal area* if \[\lim_{N\rightarrow\infty}N\max_{1\leq i\leq N} \sigma(W_i^{N}) = \lim_{N\rightarrow\infty}N\min_{1\leq i\leq N}\sigma(W_i^{N}) = 1, \label{asympeq}\] and a sequence of spherical configurations \(\left\{\omega_N\right\}_{N=1}^{\infty}\) will be said to be asymptotically equal area if its sequence of Voronoi decompositions is asymptotically equal area. # Point Sets on \(\mathbb{S}^2\) **Generalized Spiral Points**  A spherical spiral on \(\mathbb{S}^2\) is a path in spherical coordinates of the form \[r = 1,\ \ \ \ \ \theta = L\phi, \ \ \ \ \ 0 \leq \phi \leq \pi,\]where \(\phi\) denotes the polar angle and \(\theta\) the azimuth. Modifying a construction by Rakhmanov, Saff, and Zhou, Bauer defines a sequence of \(N\) points lying on a generating spherical spiral, \(S_N\): \[L = \sqrt{N\pi}, \ \ \ h_k = 1-\frac{2k-1}{N}, \ \ \ \phi_k = \cos^{-1}(h_k), \ \ \ \theta_k = L\phi_k, \ \ \ k=1...,N.\] The slope \(L\) is chosen such that for large \(N\), the distance between adjacent points on the same level of the curve is similar to the distance between adjacent levels which differ by \(2\pi\) in \(\theta\). Indeed, the geodesic spacing between turns of the spiral is given by \(2\pi/L = \sqrt{4\pi/N}\). Meanwhile, the total arc length is \[T = \int_{S_N} \sqrt{d\phi^2+d\theta^2\sin^2\phi} = \int_{0}^{\pi} \sqrt{1+L^2\sin^2\phi}\ d\phi = 2\sqrt{1+L^2}\textup{E}\big(L/\sqrt{1+L^2}\big),\] where E\((\cdot)\) is the complete elliptic integral of the second kind. For large \(N\),\(\ \ T\approx~2L\), and the spiral is divided into nearly equal length segments of approximately \(2L/N =\sqrt{4\pi/N}\). We refer to these points as the *generalized spiral points*.   As shown in the proof of Theorem [\[spiral distribution\]](#spiral distribution){reference-type="ref" reference="spiral distribution"}, the Voronoi cells of \(\omega_n\) are asymptotically equal area, but do not approach regular hexagons. Indeed, a typical decomposition is shown in Figure [\[spiral\]](#spiral){reference-type="ref" reference="spiral"}. A comparison of the mesh ratios for several values of \(N\) is shown in Table [1](#tab:spiral){reference-type="ref" reference="tab:spiral"}. Numerically, the mesh ratio appears to converge to 0.8099\.... [\[tab:spiral\]]{#tab:spiral label="tab:spiral"} Numerically, the minimal separation appears to occur at the pole with value \(|x_1-x_4| = |x_{2N+1}-x_{2N-2}|\) and the largest hole appears to occur in the triangles covering the poles, \(\triangle x_2,x_3,x_5\) and \(\triangle x_{2N},x_{2N-1},x_{2N-3}\). In a straightforward computation, it can be shown that \[\lim_{N\to\infty} \sqrt{2N+1}|x_1-x_4| = \sqrt{16-\sqrt{112}\cos(6\pi\varphi^{-1})} = 3.09207...\] and the circumradius \(r\) of the polar triangles satisfies \[\lim_{N\to\infty}\sqrt{2N+1}\,r = 2.72812....\] As shown in Table [2](#tab:Fib){reference-type="ref" reference="tab:Fib"}, the mesh ratios for Fibonacci nodes appear to converge quickly to this ratio \(\approx 0.882298\). [\[tab:Fib\]]{#tab:Fib label="tab:Fib"} The above construction was modified by Bondarenko et al to create a partition with geodesic boundaries for the creation of well-separated spherical designs. More details can be found in. Table [3](#tab:Zonal){reference-type="ref" reference="tab:Zonal"} gives a comparison of the mesh ratios of the zonal points. [\[tab:Zonal\]]{#tab:Zonal label="tab:Zonal"} Numerically, the mesh ratio appears to be bounded by 1, as shown in Table [4](#tab:HEALPix){reference-type="ref" reference="tab:HEALPix"}. [\[tab:HEALPix\]]{#tab:HEALPix label="tab:HEALPix"} Numerically, the mesh ratio appears to be bounded by \(0.86\), as shown in Table [5](#tab:RadIcos){reference-type="ref" reference="tab:RadIcos"}. [\[tab:RadIcos\]]{#tab:RadIcos label="tab:RadIcos"} [\[tab:Cube\]]{#tab:Cube label="tab:Cube"} This bound seems to be near optimal. As shown in Table [7](#tab:Oct){reference-type="ref" reference="tab:Oct"}, the mesh ratio grows to at least \(0.9235\).    [\[tab:Oct\]]{#tab:Oct label="tab:Oct"} \[\tan\lambda = \frac{\sin(\frac{hw}{2})}{\cos(\frac{hw}{2})-\frac{2\cos\psi}{\sqrt{3}}}.\] 4. Transform \((\psi,\lambda)\) into spherical coordinates.  The map \(\Phi\) is extended to \(\mathcal{I}\) by rotations and reflections. This defines the unique azimuthal equal area projection from \(\mathcal{I}\) onto \(\mathbb{S}^2\). The spherical configurations are \(\omega_N :=\Phi(\widetilde{\omega_{N}})\). A typical point set is shown in Figure [\[meshicos\]](#meshicos){reference-type="ref" reference="meshicos"}. The Voronoi cells are almost regular hexagons with 12 pentagonal cells at the vertices of \(\mathcal{I}\), and the Voronoi decomposition forms a spherical Goldberg Polyhedron. To implement the points in Matlab, we derive explicit formulas on a triangular face. More details are provided in the companion paper. Unlike the radial icosahedral points, the sequence of equal area configurations is equidistributed. Regarding quasi-uniformity, the following is proved in . As shown in Table [\[tab:MeshIcos\]](#tab:MeshIcos){reference-type="ref" reference="tab:MeshIcos"}, the mesh ratios appears to stay below \(0.736\). These are the lowest mesh ratios of all point sets discussed.  **Coulomb Points and Log Energy Points (Elliptic Fekete Points)**  For \(s=1\), Riesz \(s\)-energy minimization is the classic Thomson problem for the Coulomb potential. The sequence of minimal Coulomb energy configurations is known to be equidistributed, well-separated, and quasi-uniform. However, no explicit bound is known for the mesh ratio. The Voronoi decomposition of these cells, as shown in Figure [\[Coulomb\]](#Coulomb){reference-type="ref" reference="Coulomb"}, primarily consists of close to regular hexagons with heptagons and pentagons forming scars along the sphere. For relatively small \(N\), the scars grow out from the 12 vertices of the icosahedron like dislocations in a crystal due to displacement deformities. For \(N >~5,000\), the scars become less fixed, spreading across the sphere. For an in-depth discussion of the scarring behavior, see Bowick et al and. The log energy points are minimizers of the Riesz logarithmic potential. The sequence of log energy configurations is known to be equidistributed and well-separated, but covering and quasi-uniformity is an open problem. As shown in Table [8](#tab:CoulLog){reference-type="ref" reference="tab:CoulLog"} below, numerically the log energy points appear to be quasi-uniform. The best known lower bound on separation is due to Dragnev: \[\delta(\omega^{\log}_N)\geq\frac{2}{\sqrt{N}},\ \ \ N\geq 2.\] Their geometric structure is very similar to the Coulomb points as shown in Figure [\[Coulomb\]](#Coulomb){reference-type="ref" reference="Coulomb"}. The energies of log and Coulomb points have the same asymptotic behavior in the dominant and second order term for many Riesz potentials. See Section 3 and Conjecture [\[newconj\]](#newconj){reference-type="ref" reference="newconj"}. Generating these points is a highly nonlinear optimization problem. Unlike the configurations we have described up to now, they are not so quickly obtained. Table [8](#tab:CoulLog){reference-type="ref" reference="tab:CoulLog"} displays the mesh ratios of near minimal Coulomb and log energy configurations. We remark that the sequence appears to have outliers at several values of \(N\), such as \(N = 20, 300,\) and \(4096\). Points for \(N<500\) and \(N = k^2\), \(k \leq 150\), were provided by Rob Womersley. [\[tab:CoulLog\]]{#tab:CoulLog label="tab:CoulLog"} [\[tab:MaxDet\]]{#tab:MaxDet label="tab:MaxDet"} [\[tab:comp1\]]{#tab:comp1 label="tab:comp1"} [\[tab:comp2\]]{#tab:comp2 label="tab:comp2"} We define the energy of a probability measure \(\mu\) on \(\mathbb{S}^2\) with respect to the logarithmic or Riesz \(s\)-potential as \[\ \ \ \ \ \ \ \ \ \ \ \mathcal{I}_{s} [\mu]:=\ \int\int \frac{1}{|\textbf{x}-\textbf{y}|^s} \text{d}\mu(\textbf{x})\text{d}\mu(\textbf{y}),\ \ \ \ \ \ s\neq 0\] \[\mathcal{I}_{\log} [\mu]:=\ \int\int \log\frac{1}{|\textbf{x}-\textbf{y}|} \text{d}\mu(\textbf{x})\text{d}\mu(\textbf{y}).\] The normalized surface area measure \(\sigma\) is the unique minimizer of \(\mathcal{I}_{\log} [\mu]\) and \(\mathcal{I}_{s} [\mu]\) for \(-2<s<2\), \(s\neq 0\), and \[\mathcal{I}_{\log} [\sigma] = \frac{1}{2}-\log2,\] \[\mathcal{I}_s[\sigma] = \frac{2^{1-s}}{2-s}, \ \ \ \ -2<s<2.\] However, for \(s\geq 2\), \(\mathcal{I}_{s} [\mu] = \infty\) for all \(\mu\) supported on \(\mathbb{S}^2\) (see for example ). For random points, the expected value of the \(s\)-energy is easily computed: \[\mathbb{E}[E_{s}(\omega_N^{\text{rand}})] = \mathcal{I}_{s}[\sigma](N(N-1)),\ \ \ \ \-2<s<2. \label{RandomEnergy}\] For \(s\geq 2\), \(\mathbb{E}[E_{s}(\omega_N^{\text{rand}})] = \infty\). The Epstein-Zeta function for a lattice \(\Lambda\) in \(\mathbb{R}^2\) is given by \[\zeta_{\Lambda}(s):= \sum\limits_{0\neq\textbf{x}\in\Lambda} |\textbf{x}|^{-s},\ \ \ \ \ \ \textup{Re}\ s>2.\] Let \(\Lambda_2\) be the regular triangular lattice in \(\mathbb{R}^2\) generated by basis vectors \((1,0)\) and \((1/2,\sqrt{3}/2)\). It is known from number theory that \(\zeta_{\Lambda_2}(s)\) admits the factorization \[\zeta_{\Lambda_2}(s) = 6\zeta(s/2)\text{L}_{-3}(s/2),\ \ \ \ \ \ \text{Re}\ s>2, \label{EpsZeta}\] where \[\text{L}_{-3}(s) := 1-\frac{1}{2^s}+\frac{1}{4^s}-\frac{1}{5^s}+\frac{1}{7^s}-\cdots,\ \ \ \ \ \ \text{Re}\ s>1,\] is a Dirichlet L-series. The right-hand side of ([\[EpsZeta\]](#EpsZeta){reference-type="ref" reference="EpsZeta"}) can be used to extend \(\zeta_{\Lambda_2}(s)\) to \(s\) with Re \(s<2\).   ## Logarithmic Potential The following asymptotic expansion is proven by Betermin and Sandier in: The following extension of Theorem [\[logexpansion\]](#logexpansion){reference-type="ref" reference="logexpansion"} is conjectured by Brauchart et al in: Beltran provides a partial converse to Theorem [\[optimpliesequi\]](#optimpliesequi){reference-type="ref" reference="optimpliesequi"}: As a corollary, Coulomb points and Fibonacci spiral points are asymptotically log optimal. We make the following natural conjecture: Figures [\[logasymptotics1\]](#logasymptotics1){reference-type="ref" reference="logasymptotics1"}, [\[logasymptotics2\]](#logasymptotics2){reference-type="ref" reference="logasymptotics2"}, and [\[logasymptotics3\]](#logasymptotics3){reference-type="ref" reference="logasymptotics3"} display three comparisons of the logarithmic energies of the point sets with cardinality up to 50,000. Some configurations are sampled along subsequences to avoid overcrowding the picture. Due to the computational cost of generating approximate log energy and Coulomb points, these points are only available for \(N<22,500\). The energies of random configurations are plotted by expected value. We do not include the maximal determinant nodes because there is no known algorithm to generate them in polynomial time. Figure [\[logasymptotics1\]](#logasymptotics1){reference-type="ref" reference="logasymptotics1"} shows \(E_{\log}(\omega_N)/N^2\) for the point sets. For the equidistributed and well-separated point sets as well as for the Hammersley nodes, this ratio converges to \(1/2-\log2\) supporting Conjecture [\[logequiconj\]](#logequiconj){reference-type="ref" reference="logequiconj"}. Though the radial icosahedral and cubed sphere points are not equidistributed and thus not asymptotically optimal, their log energy appears to be of leading order \(N^2\). Figure [\[logasymptotics2\]](#logasymptotics2){reference-type="ref" reference="logasymptotics2"} displays \([E_{\log}(\omega_n)-\mathcal{I}_{\log} [\sigma] N^2]/N\log N\) for each configuration. The curves for classes of point sets begin separating. The energies of the octahedral, HEALPix, Fibonacci, generalized spiral, zonal equal area, and Coulomb points all appear to converge to the correct second order term. The energy of the Hammersley nodes appears to have \(N\log N\) second term, though incorrect coefficient. While the equal area icosahedral nodes perform better than the radial icosahedral nodes, their asymptotic energy appears to have a second term different to \(N\log N\). Numerical second order behavior of log energy points is also studied in. Figure [\[logasymptotics3\]](#logasymptotics3){reference-type="ref" reference="logasymptotics3"} compares the energy of the configurations to the conjectured minimal log energy third order term, i.e. \([E_{\log}(\omega_n)-\mathcal{I}_{\log} [\sigma]N^2 + 1/2N\log N]/N\). Conjecture [\[logconjecture\]](#logconjecture){reference-type="ref" reference="logconjecture"} is supported by the behavior of the Coulomb and log points. The octahedral, HEALPix, Fibonacci, zonal equal area, and generalized spiral configurations appear to have the third term of their energy of order \(N\) but the wrong coefficient. Of the algorithmically generated point sets, the generalized spiral and zonal equal area points perform the best with respect to the logarithmic energy.   ## Riesz Potential,-2 \(<\textbf{\textit{s}}<\) 2, \(\textbf{\textit{s}}\neq\) 0  In this range, the best known bounds on \(\mathcal{E}_s(N)\) are due to Wagner (cf. and ). Kuiljaars and Saff conjecture that \(\lim_{N\rightarrow\infty} [\mathcal{E}_s(N)-\mathcal{I}_s[{\sigma}]N^2]/N^{1+s/2}\) exists. Heuristically, the second order coefficient corresponds to the Voronoi decomposition of minimal energy points being composed primarily of close to regular hexagons. A characterization of asymptotically optimal point sets is due to Leopardi. A similar result for \(s<0\) is also known (see for example ). Theorems [\[s1char\]](#s1char){reference-type="ref" reference="s1char"} and [\[sneg1char\]](#sneg1char){reference-type="ref" reference="sneg1char"} are the analogs of Conjecture [\[logequiconj\]](#logequiconj){reference-type="ref" reference="logequiconj"} for \(-2<s<2\), \(s\neq 0\). In the particular case of \(s=-1\), the problem of minimizing energy becomes that of maximizing sums of distances. A characterization of minimal configurations for this case is due to Stolarsky (see also ) and is generalized by Brauchart and Dick in. Let \(V_t(\textbf{z}) = \left\{\textbf{x}\ |\ \textbf{x}\cdot\textbf{z}\geq t\right\}\) denote the spherical cap of height \(1-t\) centered at \(\textbf{z}\in\mathbb{S}^2\). The *\(L_2\) discrepancy* of a configuration \(\omega_N\) is defined to be \[D_{L_2}(\omega_N):= \bigg(\int_{-1}^{1}\int_{\mathbb{S}^2}\bigg(\frac{|\omega_N\cap V_t(\textbf{z})|}{N}-\sigma(V_t(\textbf{z}))\bigg)^2 d\sigma(\textbf{z})dt \bigg)^{1/2}.\] This gives the immediate corollary (see also ): Figure [\[s1asymptotics1\]](#s1asymptotics1){reference-type="ref" reference="s1asymptotics1"} graphs the Coulomb energy of the point sets to first order. From Theorem [\[s1char\]](#s1char){reference-type="ref" reference="s1char"}, \(\lim_{N\rightarrow\infty} E_1(\omega_N)/N^2 = \mathcal{I}_s[\sigma] = 1\) for all the point sets except the radial icosahedral, cubed sphere, and Hammersley nodes. As in the logarithmic case, the energies of the radial icosahedral and cubed sphere nodes appear to have \(N^2\) leading order term but incorrect coefficient. The plot is inconclusive on whether the behavior of the Hammersley nodes is asymptotically optimal. Figure [\[s1asymptotics2\]](#s1asymptotics2){reference-type="ref" reference="s1asymptotics2"} shows the second order \(N^{3/2}\) term from Conjecture [\[s1conj\]](#s1conj){reference-type="ref" reference="s1conj"}. The energies of the generalized spiral, Fibonacci spiral, zonal equal area, HEALPix, and octahedral nodes all appear to converge to an \(N^{3/2}\) term that is different from the conjectured value. None of these point sets have a regular hexagonal lattice structure as \(N\) gets large. Again the energies of equal area icosahedral points have a different second order term. The Hammersley nodes don't have low enough energy to appear on the plot. The asymptotic energy of the log nodes seems to have the same second order coefficient as the Coulomb nodes and both point sets support Conjecture [\[s1conj\]](#s1conj){reference-type="ref" reference="s1conj"}. As in the logarithmic energy case, the generalized spiral and zonal equal area points perform the best of the algorithmically generated points. Figure [\[sneg1asymptotics1\]](#sneg1asymptotics1){reference-type="ref" reference="sneg1asymptotics1"} displays \(E_{-1}(\omega_N)/N^2\) for the point sets. Again the radial icosahedral and cubed sphere points have the correct order though they are not asymptotically optimal. For \(s=-1\) The second term in the minimal energy expansion is of known order \(\sqrt{N}\) with conjectured coefficient \(\approx 0.7985\). As shown in Figure [\[sneg1asymptotics2\]](#sneg1asymptotics2){reference-type="ref" reference="sneg1asymptotics2"}, the behavior of the second order of the asymptotic energy of the point sets resembles that of their log and Coulomb energies. The minimal Coulomb and log energy points perform the best and support Conjecture [\[s1conj\]](#s1conj){reference-type="ref" reference="s1conj"}. The asymptotic energies of the octahedral, HEALPix, Fibonacci, zonal equal area, and generalized spiral appear to converge to the right second order and incorrect coefficient, and the latter two configurations have the lowest energy. ## Riesz Potential, \(\textbf{\textit{s=2}}\)  Less is known about the case \(s=2\). The first order term was proved in. It is not known whether the \(N^2\log N\) term corresponds to any geometric property of the points. There is also no result analogous to Theorems [\[Beltranthm\]](#Beltranthm){reference-type="ref" reference="Beltranthm"} and [\[s1char\]](#s1char){reference-type="ref" reference="s1char"} giving a sufficient condition for a sequence of configurations to be asymptotically optimal. Figure [\[s2asymptotics1\]](#s2asymptotics1){reference-type="ref" reference="s2asymptotics1"} shows \(E_2(\omega_N)/N^2\log N\). The Hammersley points do not appear to have asymptotically minimal energy. The behavior of the octahedral, HEALPix, Fibonacci, generalized spiral, zonal equal area, and equal area icosahedral points is not clear. Their energies remain well below the known coefficient but have larger energy than the Coulomb and log energy points. Calculations beyond \(N=50,000\) are needed. The generalized spiral, zonal equal area, and equal area icosahedral points perform the best.   ## Riesz Potential, \(\textbf{\textit{s}}>2\)  In the hypersingular case \(s>2\) the continuous energy integral is infinite for any probability measure \(\mu\). Instead, the energy is known to be dominated by the nearest neighbor interactions of points as shown in a much more general result by Hardin and Saff. One can consider the analytic continuation of \(\mathcal{I}_s[\sigma]\), \(0<s<2\), \[V_s: = \frac{2^{1-s}}{2-s},\ \ \ \ s\in\mathbb{C}\setminus \left\{2\right\}.\] The following is analogous to Conjecture [\[s1conj\]](#s1conj){reference-type="ref" reference="s1conj"} (see ). Figure [\[s3asymptotics1\]](#s3asymptotics1){reference-type="ref" reference="s3asymptotics1"} plots \(E_3(\omega_N)/N^{5/2}\). The energies of most configurations seem to be going to the correct order but incorrect coefficient. The equal area icosahedral points outperform the spiral and zonal equal area points of the algorithmically generated configurations. This is expected because their Voronoi decomposition is closest to the regular hexagonal lattice. The log energy and Coulomb points again seem to be close to minimal and may converge to the conjectured \(C_3\approx 0.199522\). The Hammersley points are not seen on the plot because their asymptotic energy does not appear to have first order \(N^{5/2}\). We conclude this section with the following observation: The apparent first and second order asymptotically minimal behavior of the energy of the Coulomb and log points for the potentials we have considered is striking and we suggest a conjecture. # Proofs We begin with two auxiliary results.  *Proof of Theorem [\[spiral distribution\]](#spiral distribution){reference-type="ref" reference="spiral distribution"}.* For a fixed \(N\) denote in spherical coordinates \(x_i:=(\phi_i,\theta_i)\in\omega_N\). We first prove the separation bound. Let \(\epsilon_N = 2\sqrt{4\pi/N}\). For large \(N\), by the pigeonhole principle, at least one set of adjacent points \(x_k,x_{k+1}\in\omega_N\) is separated along \(S_N\) by a distance less than \(\sqrt{4\pi/N}\). So we can restrict our attention to \(k\leq N/2\) and \(\epsilon_N\)-balls \(B(x_k,\epsilon_N)\). For large N, if \(k<8\pi+1/2\), then \[\cos\phi_k = 1-\frac{2k-1}{N}\geq 1-\frac{16\pi}{N} \approx \cos(2\sqrt{\frac{4\pi}{N}}),\] and \(x_k\) is within the first two full longitudinal turns of \(S_N\) starting from a pole. Otherwise, \(B(x_k,\epsilon_N)\) contains disjoint levels of \(S_N\). In this case, the minimal distance between levels in \(B(x_k,\epsilon_N)\) is \(\sqrt{4\pi/N}\). We compute the nearest neighbor distance between points in the same level as follows. Let \[f_N(k):=\sqrt{N}|x_k-x_{k+1}|, \ \ \ \ x_k\in \omega_N,\ \ \ \ k<N/2.\] Using the distance formula for spherical coordinates \[|x_j-x_k|^2 = 2-2(\cos\phi_j\cos\phi_k+\sin\phi_j\sin\phi_k\cos(\theta_j-\theta_k)) \label{geodesic}\] and expanding \(\cos^{-1}x\) around \(x=1\) we have \[\lim_{N\to\infty}f_N(k) = \sqrt{8k-4\sqrt{4k^2-1}\cos(\sqrt{2\pi}(\sqrt{2k-1}-\sqrt{2k+1}))}, \label{spirallimit}\] and \(f_N(k)<f_{N+1}(k)\) for all \(k<N/2-1\) and large N. Thus we have the correct order for the minimal separation between adjacent points in \(\omega_N\). Furthermore, ([\[spirallimit\]](#spirallimit){reference-type="ref" reference="spirallimit"}) is increasing as a function of \(k\) and thus \[\lim_{N\to\infty} \min_k f_N(k) = \sqrt{8-4\sqrt{3}\cos(\sqrt{2\pi}(1-\sqrt{3}))}. \label{spiralmin}\] Lastly, around the north pole, \(k_1,k_2<8\pi+1/2\), we can again use ([\[geodesic\]](#geodesic){reference-type="ref" reference="geodesic"}) to show that \[\lim_{N\to\infty} \sqrt{N}|x_{k_1}-x_{k_2}|\] exists and can be computed case by case for pairs \((k_1,k_2)\). The southern hemisphere can be similarly computed. Comparing to ([\[spiralmin\]](#spiralmin){reference-type="ref" reference="spiralmin"}) gives the separation constant ([\[spiralsep\]](#spiralsep){reference-type="ref" reference="spiralsep"}). For covering, given \(y\in\mathbb{S}^2\), \[\textup{dist}(S_N)\leq \sqrt{\frac{\pi}{N}}.\] From ([\[spirallimit\]](#spirallimit){reference-type="ref" reference="spirallimit"}), the maximal distance from any point on \(S_N\) to a point of \(\omega_N\) is \(O(1/\sqrt{N})\) and thus the covering radius of \(\omega_N\) is also \(O(1/\sqrt{N})\). We have two additional observations. First, the Voronoi decompositions of the spiral points are diameter bounded. Secondly, the Voronoi cells are asymptotically equal area on \(K_h := \left\{(x,y,z)\in \mathbb{S}^2:-h\leq z\leq h \right\}\) for any fixed \(0<h<1\). By this we mean \[\lim_{N\rightarrow\infty}N\max_{V_x(\omega_N)\subset K_h} \sigma(V_x(\omega_N)) = \lim_{N\rightarrow\infty}N\min_{V_x(\omega_N)\subset K_h} \sigma(V_x(\omega_N)) = 1.\] Indeed, fixing \(h\), for any \(\epsilon>0\), we can take \(N\) large enough such that given \(x_i=(\phi_i,\theta_i)\in K_h\cap \omega_N\), \(x_{i-1}\), and \(x_{i+1}\) are almost iso-latitudinal with \(x_i\) with separation \(\sqrt{4\pi/N}\). I.e., \[|x_{i\pm 1}-(\phi_i,\theta_i\pm\csc\phi_i\sqrt{4\pi/N})|<\epsilon.\] There exists shifts \(0\leq\lambda_{i+},\lambda_{i-}\leq\sqrt{4\pi/N}\) such that the nearest points in the adjacent spiral levels are within \(\epsilon\) of the points \((\phi_i\pm\sqrt{4\pi/N},\theta_i+\csc\phi_i\lambda_{i\pm})\) and \((\phi_i\pm\sqrt{4\pi/N},\theta_i+\csc\phi_i(\lambda_{i\pm}-\sqrt{4\pi/N}))\). Thus as \(N\to\infty\), the Voronoi cell \(V_i(\omega_N)\) approaches the form of \(V\) in Figure [\[spiralvoronoicell\]](#spiralvoronoicell){reference-type="ref" reference="spiralvoronoicell"} and \[\sigma(V_i(\omega_N)) = \sigma(V) + O(\epsilon^2).\] Furthermore, we can treat \(V\) as a planar polygon in Figure [\[spiralvoronoicell\]](#spiralvoronoicell){reference-type="ref" reference="spiralvoronoicell"}, and \[\sigma(V)=\frac{1}{N}\] independent of the shifts \(\lambda_{i\pm}\). This we show by direct computation. Letting \(a=\sqrt{4\pi/N}\) and centering \(x_i\) at \((0,0)\), the points \(y_1 = (\lambda_{i+}/2,a/2)\) and \(y_2 = ((\lambda_{i+}-a)/2,a/2)\) are the midpoints of the lines connecting \(x_i\) to its nearest neighbors in the adjacent level which are shifted by \(\lambda_{i+}\). The corresponding lines \[L_1: y =-\frac{\lambda_{i+}}{a}(x-\frac{\lambda_{i+}}{2})+\frac{a}{2} \ \ \ \ \ \ \ L_2: y = \frac{a-\lambda_{i+}}{a}(x-\frac{\lambda_{i+}-a}{2})+\frac{a}{2}\] form the top boundary of V and have intersection point \[w:= (\lambda_{i+}-a/2,(\lambda_{i+}a+a^2-\lambda_{i+}^2)/2a).\] From this we calculate the area of the top half of V to be \[\sigma(V_{top}) = \frac{1}{4\pi}\bigg[a\bigg(\frac{-a\lambda_{i+}+\lambda_{i+}^2+a^2}{2a}\bigg)+\frac{a}{2}\bigg(\frac{a\lambda_{i+}-\lambda_{i+}^2}{a}\bigg)\bigg]=\frac{1}{2N}.\] The same calculation holds for the bottom half of \(V\) and thus ([\[asympeq\]](#asympeq){reference-type="ref" reference="asympeq"}) holds. We now consider equidistribution. Because the height steps between points in \(\omega_N\) are uniform, for a spherical cap \(A\) centered at a pole, \[\lim_{N\to\infty}\frac{|\omega_N\cap A|}{N}=\sigma(A). \label{equicap}\] If \(A\) does not contain one of the poles, then \(A\subset K_h\) for some \(h\) and ([\[equicap\]](#equicap){reference-type="ref" reference="equicap"}) holds by Proposition [\[lemmaprop\]](#lemmaprop){reference-type="ref" reference="lemmaprop"}. Finally if \(A\) is a cap containing but not centered at one of the poles, \(A\) can be partitioned into an open cap of height \(h\) centered at the pole and \(A\cap K_h\). Because ([\[equicap\]](#equicap){reference-type="ref" reference="equicap"}) holds on each disjoint subset, it also holds on \(A\). 0◻ *Proof of Proposition [\[Fib mesh ratio\]](#Fib mesh ratio){reference-type="ref" reference="Fib mesh ratio"}.* When \(z=k,\) the basis vector \(\textbf{c}_{k,i}\) has minimum length of \(\sqrt{2}d\). At this latitude, \(\textbf{c}_{k+1,i}\) and \(\textbf{c}_{k-1,i}\) form the next most dominant spirals and have length \(\sqrt{3}d\). For \(z = k + 1/2\), \(\textbf{c}_{k,i}\) and \(\textbf{c}_{k+1,i}\) are equally dominant and have grid length \(\sqrt[4]{5}d\). For a fixed latitude \(\phi\neq \pm\pi/2\), \(z\) increases with \(N\). Points around \(\phi\) will form a locally rectangular grid. Thus the separation approaches \(\sqrt{2}d\) which occurs when \(z=k\) and the largest hole in the triangulation around \(x_i\) will be at most \(\sqrt[4]{5}d/\sqrt{2}\) which occurs when \(z=k\pm 1/2\). Thus off the poles, \(\delta(\omega_{2N+1})\geq \sqrt{8\pi\sqrt{5}}/\sqrt{N}\) and \(\eta(\omega_{2N+1})\leq \sqrt{2\pi/N}.\) Indeed, these inequalities hold for large zone numbers \(z\) where \(F_k\approx \varphi^k/\sqrt{5}\) holds. However, on the polar points, \(x_1\) and \(x_N\), the zone number \[z=\frac{\log((2N+1)\pi\sqrt{5}(1-4N^2/(2N+1)^2))}{\log\varphi^2}\to\frac{\log(2\pi\sqrt{5})}{\log\varphi^2} = 2.75...,\ \ \ \ N\to\infty\] and writing \(\textbf{c}_{k,i}\) in the form of equation ([\[Fib basis2\]](#Fib basis2){reference-type="ref" reference="Fib basis2"}) for small \(k\) overestimates the length of the vector. Using equation ([\[Fib basis\]](#Fib basis){reference-type="ref" reference="Fib basis"}) and noticing that \(k-1/2\leq z \leq k+1/2\) implies \[\frac{\varphi^{2k-1}}{(2N+1)\pi\sqrt{5}}\leq\cos^2\phi\leq\frac{\varphi^{2k+1}}{(2N+1)\pi\sqrt{5}},\] we have \[|\textbf{c}_{k,i}|^2\geq \frac{4\pi^2\varphi+20\pi F_k^2\varphi^{-2k+1}}{(2N+1)\pi\sqrt{5}}=O\bigg(\frac{1}{2N+1}\bigg), \ \ \ \ \ k-1/2\leq z \leq k+1/2.\] Since \(|\textbf{c}_{k,i}|\) is the minimal separation distance for \(k-1/2\leq z\leq k+1/2\), we have the correct order of separation on \(\mathbb{S}^2\). By a similar computation, we have the upper bound \[|\textbf{c}_{k,i}|\leq O\bigg(\frac{1}{\sqrt{2N+1}}\bigg), \ \ \ \ \ \ \ \ k-3/2\leq z \leq k+3/2.\] Since the triangulation of \(\omega_{2N+1}\) in each zone consists of \(\textbf{c}_{k,i}\), \(\textbf{c}_{k-1,i}\), and \(\textbf{c}_{k+1,i}\), the covering of \(\omega_{2N+1}\) is of the correct order on all of \(\mathbb{S}^2\). 0◻ *Proof of Proposition [\[zonal quasi\]](#zonal quasi){reference-type="ref" reference="zonal quasi"}.* The diameter boundedness of the partition and Proposition [\[lemmaprop\]](#lemmaprop){reference-type="ref" reference="lemmaprop"} gives equidistribution and covering, \[\eta(\omega_N) \leq \frac{3.5}{\sqrt{N}}.\] In and, it is established that there exists \(c_1,c_2>0\) such that for all partitions \(P_N\), with \(\phi_j\), \(n\) as defined above, \[\frac{c_1}{\sqrt{N}}\leq\phi_{j+1}-\phi_j\leq \frac{c_2}{\sqrt{N}},\ \ \ \ \ \ \ \ \ 0\leq j \leq n. \label{zonal collar order}\] This gives the correct order of separation between collars. For neighbors \(x_1,x_2 \in \omega_N\) within collar \(j\), wlog suppose \(\phi_j<\pi/2\). Using the fact that the normalized area of each cell can be expressed as \[\frac{(\cos\phi_j-\cos\phi_{j+1})}{2y_i} = \frac{1}{N},\] we have \[|x_1-x_2|\geq \frac{2\pi\sin\phi_j}{y_i} = \frac{4\pi\sin\phi_j}{N(\cos\phi_j-\cos\phi_{j+1})}.\] So it suffices to show there exists \(c_3>0\) such that \[\frac{\sin\phi_j}{\cos\phi_j-\cos\phi_{j+1}}\geq c_3\sqrt{N}\ \ \ \ \ \ \forall N, 0\leq j\leq n. \label{zonalsep}\] For a fixed \(h>0\) and \(\phi_j\geq h\), this follows from ([\[zonal collar order\]](#zonal collar order){reference-type="ref" reference="zonal collar order"}) and the fact that \(\cos\) is Lipschitz. On the other hand, for sufficiently small \(\phi_j\), there exists \(c_4>0\) such that \[\frac{\sin\phi_j}{\cos\phi_j-\cos\phi_{j+1}}\geq c_4\frac{\phi_j}{\phi_j^2-\phi_{j+1}^2}.\] Again applying ([\[zonal collar order\]](#zonal collar order){reference-type="ref" reference="zonal collar order"}) twice, \[\textup{RHS}\ (\ref{zonalsep})\geq \frac{c_4}{c_1}\frac{\phi_j}{\phi_j+\phi_{j+1}}\sqrt{N}\geq \frac{c_4}{c_1}\frac{\phi_j}{2\phi_j+\frac{c_2}{\sqrt{N}}}\sqrt{N}\geq c_3\sqrt{N}.\] In the last step we used the fact that for some \(c_5>0\) and all \(j\) \[\phi_j\geq\cos^{-1}(1-\frac{2}{N})\geq \frac{c_5}{\sqrt{N}}.\] 0◻ *Proof of Proposition [\[HEALPix quasi\]](#HEALPix quasi){reference-type="ref" reference="HEALPix quasi"}.* The pixels are diameter bounded and thus by Proposition [\[lemmaprop\]](#lemmaprop){reference-type="ref" reference="lemmaprop"} the nodes are equidistributed. To establish separation, we examine the five cases of nearest neighbor points: The points lie in 1) the polar region or 2) the equatorial region, and the points lie in a) the same ring or b) adjacent rings, and 3) the points lie in adjacent rings at the boundary of the polar region and the equatorial region. Case 1a: If nearest neighbor points lie in the polar region along the same ring \(1\leq i\leq k\) that has radius \[r_i = \sin\phi_i= \sqrt{\frac{2i^2}{3k^2}-\frac{i^4}{9k^4}}\] and \(4i\) equally spaced points, then the separation \(\delta\) satisfies \[\delta = 2r_i\sin\frac{\pi}{4i} = 2\sqrt{\frac{2i^2}{3k^2}-\frac{i^4}{9k^4}}\sin\frac{\pi}{4i}\geq\frac{\sqrt{2}}{i}\sqrt{\frac{2i^2}{3k^2}-\frac{i^4}{9k^4}} = O\bigg(\frac{1}{k}\bigg) = O\bigg(\frac{1}{\sqrt{N}}\bigg).\] In the middle inequality we use the fact that \[\sin x \geq \frac{\sqrt{2}/2}{\pi/4}x\ \ \ \ \ \ \ \ \ 0\leq x \leq \pi/4. \label{sineineq}\] Case 2a: Suppose the nearest neighbor points lie in the equatorial region along the same ring. Since each ring has \(4k\) points, the smallest separation occurs at the ring farthest from the equator and closest to \(z = 2/3\). Using ([\[sineineq\]](#sineineq){reference-type="ref" reference="sineineq"}) again, we have \[\delta = 2r_i\sin\frac{\pi}{4k}\geq \frac{2\sqrt{5}}{3}\sin\frac{\pi}{4k}\geq\frac{\sqrt{10}}{3k} = O\bigg(\frac{1}{\sqrt{N}}\bigg).\] Case 1b: We split up the rings in the polar region into the outer half, \(1\leq i \leq k/2\) and the inner half, \(k/2\leq i\leq k\). On the outer half, the separation between rings is \[\begin{aligned} \delta \geq r_{i+1}-r_i =& \frac{(i+1)\sqrt{6k^2-(i+1)^2}-i\sqrt{6k^2-i^2}}{3k^2}\\ \geq& \frac{(k/2+1)\sqrt{6k^2-(k/2+1)^2}-(k/2)\sqrt{6k^2-(k/2)^2}}{3k^2} = O\bigg(\frac{1}{\sqrt{N}}\bigg).\\ \end{aligned}\] On the inner polar rings, the separation between rings is \[\delta \geq |\cos\phi_{i+1}-\cos\phi_i| = \frac{2i+1}{3k^2}\geq \frac{k+1}{3k^2} =O\bigg(\frac{1}{\sqrt{N}}\bigg).\] Case 2b: In the equatorial region, the ring height \(z\) increases linearly with respect to the index \(i\) giving \[\delta \geq \frac{2}{3k} =O\bigg(\frac{1}{\sqrt{N}}\bigg).\] Case 3 follows from Case 1b and 2b. The covering of the points follows by similar geometric arguments. 0◻ *Proof of Proposition [\[rad icos quasi\]](#rad icos quasi){reference-type="ref" reference="rad icos quasi"}.* It suffices to show \(\Pi\) is locally bi-Lipschitz on each face \(\mathcal{F}\) of the icosahedron. If for some \(\delta\), \(L_1\), \(L_2 > 0\) \[L_1|x-y|\leq|\Pi(x)-\Pi(y)|\leq L_2|x-y|, \ \ \ \ \ x,y\in \mathcal{F},\ \ |x-y|<\delta,\] then \[\gamma(\omega_N)\leq\frac{L_2}{L_1}\gamma(\widetilde{\omega_{N_k}}) = \frac{L_2}{L_1\sqrt{3}}.\] Let \(c := \min_{x \in \text{Icos}} |x| = \sqrt{(1/3+2\sqrt{5}/15)}\). For \(x,y\in \mathcal{F}\) with angle \(\theta\), we have \[|x-y|\geq 2c\sin\frac{\theta}{2}.\] Using the fact that \(\sin^{-1}x\leq \frac{\pi}{2}x\) for \(x\leq 1\) and \(|x-y|< 2c\), \[|\Pi(x)-\Pi(y)|\leq \theta\leq 2 \sin^{-1}\bigg(\frac{|x-y|}{2c}\bigg)\leq\pi c |x-y|.\] For the other inequality, wlog suppose \(c\leq|x|\leq|y|\leq 1\), and consider the line \(P \subset \mathcal{F}\) connecting \(x\) and \(y\). Denote \(z\) as the projection of \(0\) onto \(P\). Defining \(\phi := \cos^{-1}(|z|/|x|)\), we have \[|y|-|x| = |z|\sec(\theta+\phi)-|z|\sec\phi\leq \sec(\theta+\phi)-\sec\phi.\] Since \(\phi \leq \cos^{-1}c-\theta\) and secant is convex on \((0,\pi/2)\), \[|y|-|x|\leq\frac{1}{c}-\sec(\cos^{-1}c-\theta)=:g(\theta).\] Thus, \[\begin{aligned} |x-y|^2 &= |x|^2+|y|^2-2|x||y|\cos \theta\leq\max_{c\leq|y|\leq 1} (|y|+g)^2+|y|^2-2|y|(|y|+g)\cos\theta\\ & = (1+g)^2+1-2(1+g)\cos\theta.\\ \end{aligned}\] Since \[f(\theta):=\frac{(1+g)^2+1-2(1+g)\cos\theta}{2-2\cos\theta}\] is continuous for \(\theta\in (0,\pi/2+\cos^{-1}c)\) and \(\lim_{\theta\rightarrow 0} f(\theta)\) exists, there exists \(L>0\) such that \[|x-y|\leq L|\Pi(x)-\Pi(y)|, \ \ \ \ \ \ \ \ \ \theta\in (0,\pi/2+\cos^{-1}c).\] 0◻ *Proof of Theorem [\[oct quasi\]](#oct quasi){reference-type="ref" reference="oct quasi"}.* Restricting ourselves to the face of \(\mathbb{K}\) with all positive coordinates, label the vertices of the partition \(\left \{A_{i,j}\right \}_{0\leq i+j\leq k}\) by \[A_{i,j} = \bigg( \frac{iL}{k\sqrt{2}}, \frac{jL}{k\sqrt{2}}, \frac{L}{\sqrt{2}}\bigg(1-\frac{i+j}{k}\bigg)\bigg).\] Let \(\mathcal{A}_{i,j} = \mathcal{U}(A_{i,j})\). Then \[\mathcal{A}_{i,j} = \bigg( \frac{i+j}{k}\sqrt{2-\frac{(i+j)^2}{k^2}}\cos\frac{\pi j}{2(i+j)}, \frac{i+j}{k}\sqrt{2-\frac{(i+j)^2}{k^2}}\sin\frac{\pi j}{2(i+j)}, 1-\frac{(i+j)^2}{k^2}\bigg).\] Then \[\delta (\omega_N) = \min_{i,j}\{ \|\mathcal{A}_{i+1,j}-\mathcal{A}_{i,j}\|, \|\mathcal{A}_{i,j+1}-\mathcal{A}_{i,j}\|, \|\mathcal{A}_{i+1,j}-\mathcal{A}_{i,j+1}\|\}. \label{OctNeighbors}\] Adapting, we have \[\|\mathcal{A}_{i+1,j}-\mathcal{A}_{i,j}\|^2 = 2\frac{(i+j+1)^2}{k^2}+2\frac{(i+j)^2}{k^2}-2\frac{(i+j+1)^2(i+j)^2}{k^4} \ \ \ \ \ \ \ \ \ \ \] \[\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ -2\frac{(i+j)(i+j+1)}{k^2}\sqrt{2-\frac{(i+j+1)^2}{k^2}}\sqrt{2-\frac{(i+j)^2}{k^2}}\cos\frac{\pi j}{2(i+j)(i+j+1)}.\] Along the line \(i+j = c\), the minimum is obtained when the cosine term is maximized, i.e. at \(j=0\). Thus \[\min_{i,j}\|\mathcal{A}_{i+1,j}-\mathcal{A}_{i,j}\|^2 = \min_{0\leq i\leq k} \|\mathcal{A}_{i+1,0}-\mathcal{A}_{i,0}\|^2\] \[= \frac{2}{k^2}\min_{0\leq i \leq k}\bigg((i+1)^2 +i^2+\frac{(i+1)^2i^2}{k^2}-i(i+1)\sqrt{2-\frac{(i+1)^2}{k^2}}\sqrt{2-\frac{i^2}{k^2}}\bigg)= \frac{2}{k^2}.\] By symmetry of the above expressions in \(i\) and \(j\), \[\min_{i,j}\|\mathcal{A}_{i,j+1}-\mathcal{A}_{i,j}\|^2 = \min_{0\leq j \leq k}\|\mathcal{A}_{0,j+1}-\mathcal{A}_{0,j}\|^2 = \frac{2}{k^2}.\] Lastly, \[\|\mathcal{A}_{i+1,j}-\mathcal{A}_{i,j+1}\|^2 = 4\frac{(i+j+1)^2}{k^2}\bigg(2-\frac{(i+j+1)^2}{k^2}\bigg)\sin^2\frac{\pi}{4(i+j+1)}\] which again depends on only \(i+j\). Using ([\[sineineq\]](#sineineq){reference-type="ref" reference="sineineq"}) we have \[\begin{aligned} \min_{i,j}\|\mathcal{A}_{i+1,j}-\mathcal{A}_{i,j+1}\|^2 &= \min_{0\leq i \leq k} \|\mathcal{A}_{i+1,0}-\mathcal{A}_{i,1}\|^2\\ &= 4 \min_{0\leq i \leq k} \frac{(i+1)^2}{k^2}\bigg(2-\frac{(i+1)^2}{k^2}\bigg)\sin^2\frac{\pi}{4(i+1)}\\ &\geq 4\min_{0\leq i\leq k} \frac{(i+1)^2}{k^2}\bigg(2-\frac{(i+1)^2}{k^2}\bigg)\frac{1}{2(i+1)^2}\\ &= 4\min_{0\leq i \leq k} \frac{1}{2k^2}\bigg(2-\frac{(i+1)^2}{k^2}\bigg) = \frac{2}{k^2}\bigg(2-\frac{(k+1)^2}{k^2}\bigg).\\ \end{aligned}\] Thus from ([\[OctNeighbors\]](#OctNeighbors){reference-type="ref" reference="OctNeighbors"}) \[\delta(\omega_{4k^2+2})^2 \geq \frac{2}{k^2}\bigg(2-\frac{(k+1)^2}{k^2}\bigg).\] Taking the square root and substituting \(N = 4k^2 +2\) gives \[\liminf_{N\rightarrow\infty}\delta(\omega_N)\sqrt{N} \geq \sqrt{8}.\] Finally, the diameter bound in gives an immediate upper bound for the covering radius from which ([\[Octmesh\]](#Octmesh){reference-type="ref" reference="Octmesh"}) follows: \[\eta(\omega_{4k^2+2},\mathbb{S}^2) \leq \sqrt{\frac{4+\pi^2}{8k^2}}.\] 0◻ # Matlab Code Many thanks to Grady Wright for his help in implementing some of these point sets in Matlab. Code for the Fibonacci, Hammersley, HEALPix, and cubed sphere nodes authored by Wright is available at\ ***https://github.com/gradywright/spherepts***. The approximate Coulomb, log energy, and maximum determinant points were provided from computations by Rob Womersley. Code for the zonal equal area points is available in Paul Leopardi's Recursive Zonal Equal Area Toolbox from ***eqsp.sourceforge.net.*** Code for the generalized spiral, octahedral, radial icosahedral, and equal area icosahedral nodes was created by the authors and is available upon request.

{'timestamp': '2016-10-25T02:07:57', 'yymm': '1607', 'arxiv_id': '1607.04590', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04590'}

# Introduction The problem of finding the best packings of lines, one-dimensional subspaces of a real or complex Euclidean space, is easy to state. Despite its simple geometric formulation, it has given rise to a surprisingly diverse literature over many years, ranging from relatively elementary, low dimensional examples to more sophisticated constructions, some involving combinatorial or group-theoretic aspects and results on bounds on the relationship between the number of lines and achievable angles. Maximal sets of equiangular lines are known to be optimal packings, but the number of lines that can be realized is hard to determine. Special regard has been given to the construction of complex examples, motivated by applications in quantum information theory. Numerical searches indicate that they exist in many cases, but a rigorous proof of their existence is restricted to low dimensions, see and references therein. Next to lines, packings of higher-dimensional subspaces have also been investigated. In this case, even less seems to be known about general construction principles that realize tight bounds . More recently, these packing problems have been studied in the context of frame theory. Apart from geometric optimality criteria, frame design aims at tightness, which implies that the projections onto the subspaces sum to a multiple of the identity. The case of higher-dimensional subspaces corresponds to fusion frames. If the number of subspaces is not too large, then in close similarity to line packings, equi-distant fusion frames present optimal solutions. Examples of such constructions follow similar strategies as in the frame case. For a larger number of subspaces, such equiangular arrangements cannot be realized and one needs to find an alternative bound for the characterization of optimal packings, for example the orthoplex bound for lines or subspaces. In an earlier paper, we constructed optimal line packings when the number of lines goes slightly beyond the threshold beyond which equiangular lines are impossible to realize. In this paper, we study the orthoplex bound for subspace packings and investigate cases in which the bound is achieved while the number of subspaces is maximal. The main results are as follows. In order to maximize the number of subspaces while achieving the orthoplex bound, the dimension of the subspaces is necessarily half of the dimension of the ambient space and the chordal distance between subspaces assumes only two values. Because of the relation with the orthoplex bound, we call these subspace packings maximal orthoplectic fusion frames. The family of examples we describe here has already appeared in the literature, either as optimal real subspace packings, whose discovery is ascribed to a "remarkable coincidence", or among the more general family of quantum 2-designs in complex Hilbert spaces of prime power dimensions. In the complex case, it was observed that the projections are affine, Grassmannian designs, see also, where the construction is attributed to Rötteler. In the present paper, we examine rigidity properties in the construction of maximal orthoplectic fusion frames obtained from the theory of packings and designs. We treat the real and complex case on the same footing, involving a new construction principle. To this end, we leverage earlier constructions of orthoplex-bound achieving, optimal line packings associated with mutually unbiased bases introduced by Schwinger. Maximal sets of mutually unbiased bases are known to exist in the complex case in prime power dimensions and in the real case if the dimension is a power of four, see also. We obtain maximal orthoplectic fusion frames by augmenting these maximal mutually unbiased bases with block designs, subsets of the index set that satisfy certain combinatorial conditions. The designs we construct for our purposes are known as balanced incomplete block designs and at the same time associated with optimal constant-weight binary codes, see also. The resulting families of subspaces are constructed in any real Hilbert space whose dimension is a power of four or in any complex Hilbert space whose dimension is a power of two. This paper is organized as follows. After the introduction, we fix notation and recall known distance and cardinality bounds on fusion frames in Section [2](#sec:bounds){reference-type="ref" reference="sec:bounds"}. We relate the orthoplex bound for fusion frames with the notion of mutual unbiasedness in Section [\[sec:muff\]](#sec:muff){reference-type="ref" reference="sec:muff"} and study implications for the structure of maximal orthoplectic fusion frames as packings and as Grassmannian designs. Finally, Section [\[sec:moff\]](#sec:moff){reference-type="ref" reference="sec:moff"} presents the construction of a family of maximal orthoplectic fusion frames. # Distance bounds and Grassmannian fusion frames {#sec:bounds} By the polarization identity, the tightness property is equivalent to the fusion frame resolving the identity \(I_m\) on \(\BB{F}^m\) according to \[\frac 1 A \sum_{j=1}^n P_j = I_m \,.\] More general types of fusion frames are obtained by relaxing the condition that all subspaces have the same dimension and by scaling the projections with non-negative weight factors. For any two projections \(P\) and \(P'\) onto \(l\)-dimensional subspaces of \(\BB{F}^m\), the *chordal distance* is defined by \(d_c(P,P') = \frac{1}{\sqrt 2} \| P-P'\| = (l-\mathop{\mathrm{tr}}( P P' ) )^{1/2} \,.\) In order to characterize optimal packings with respect to \(d_c\), we use an embedding that maps the projections to vectors in a higher dimensional Hilbert space. We denote the dimension of this space as \[d_\BB{F} (m) = \left\{ \begin{array}{cc} \frac{(m+2)(m-1)}{2}, & \BB{F}= \BB{R} \\ m^2-1, & \BB{F} =\BB{C} \end{array} \right..\] We use Rankin's distance bound for vectors on the sphere in the formulation used by Conway, Hardin and Sloane. In terms of the inner products between \(n\) unit vectors in \(\BB{R}^d\), the Rankin bound is \[\max_{j, k \in [\![ n ]\!], j \ne k} \langle v_j, v_k \rangle \ge-\frac{1}{n-1}, \,\] and if \(n>d+1\), then it improves to \[\max_{j, k \in [\![ n ]\!], j \ne k} \langle v_j, v_k \rangle \ge 0 \,.\] Using the embedding from Theorem [\[thm:embed\]](#thm:embed){reference-type="ref" reference="thm:embed"}, we reformulate the Rankin bound for the Hilbert-Schmidt inner products of the projections of a fusion frame. This results in a bound that has already been derived in an alternative way before and in an improved bound for a larger number of subspaces, as noted in. We wish to construct maximal orthoplectic fusion frames, which means the projections must embed exhaustively into the vertices of an orthoplex in \(\BB{R}^{d_{\BB{F}}(m)}\). The feasability of this depends on whether the embedding admits antipodal points in the higher dimensional Euclidean sphere, which in turn depends on the relationship between \(l\) and \(m\). Partitioning a maximal orthoplectic fusion frame into orthogonal pairs of projections shows, together with the preceding proposition, that \(m=2l\). Maximal orthoplectic fusion frames enjoy another property that has been studied in the literature: they are part of a family of *Grassmannian 2-designs*, as shown by Zauner. For our purposes, this is important because it imposes more rigidity on their construction. We follow Zauner's convention for the definition of these designs. Equivalently, the right-hand side of the defining identity can be averaged with respect to the Haar measure \(\mu\) on the group \(\CAL U\) of orthogonal or unitary \(m \times m\) matrices. This formulation implies a simple characterization of the design property based on the *\(t\)-coherence tensor* \[K_{t,l,m} = \int_{\CAL{U}} (U P U^*)^{\otimes t} d\mu(U) \,,\] where \(P\) is any rank-\(l\) orthogonal projection matrix. Because of the analogy with bounds for constant-weight codes, Zauner calls the following estimate a generalized Sidelnikov inequality. Because it is of independent interest, we study the \(t\)-coherence tensor and derive an alternative proof of the Grassmannian 2-design property of maximal orthoplectic fusion frames.

{'timestamp': '2016-07-18T02:07:51', 'yymm': '1607', 'arxiv_id': '1607.04546', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04546'}

# Error Probability Bound {#appendix:APPENDIX A} In this section, we give a direct proof for . Specifically, we bound the error probability of the maximum likelihood decoder, and show that under the condition on \(T\) in the lemma, the error probability is indeed approaching \(0\) as \(N \to \infty\). Under the assumption that all items are equality likely to be defective, we consider the error probability given that the first \(K\) items are defective, that is, is the defective set. Denote this probability by \(P_{e|1}\). We have \[P_{e|1} \leq \sum_{i=1}^{K}P(E_i),\] where \(E_i\) is the event of a decoding error in which the decoder declares a defective set which differs from the true one (\(S_1\)) in exactly \(i\) items. In general, we follow the derivation in. However, there is a key difference. In the code construction suggested in , for each item there are \(M\) possible codewords (a "bin\" of size \(M\)). Only one of these codewords is selected by the mixer to decide in which pool tests the item will participate. Thus, when viewing this problem as a channel coding problem, if an item is defective, one and only one codeword out of its bin is actually transmitted (and summed with the codewords of the other \(K-1\) defective items). Since the decoder does not know which codewords were selected in each bin (the randomness is known only to the mixer), there are multiple error events to consider. E.g., events where the decoder choose the wrong codeword for some items, yet identified parts of the bins correctly, and, of course, events where the codeword selected was from a wrong bin. This complicates the error analysis significantly. Moreover, we wish to employ the correction suggested in, which results in a simpler yet stricter bound. Consider the event \(E_i\). Clearly, \(E_i\) can be broken down into two disjoint events. The first is \(E_i\) *and* the event that the codewords selected for the correct \(K-i\) items are the true transmitted ones, and the second is the event of both \(E_i\) and the event that *at least one of the codewords selected* for the correct items is wrong. Denote the first event as \(E'_i\). Now, consider the case where we have \(E_i\), that is, a correct decision on \(K-i\) items, yet, out of these \(K-i\) items, the decoder identified only \(j\) codewords right, \(0 \leq j < K-i\), and for the rest, it identified a wrong codeword in the right bin. We claim that the probability for this event is exactly \(P(E'_{K-j})\), as the decoder has to mistake \(\bf{X'}\) for \(\bf{X}_{S_1}\), where both are collections of \(K\) codewords, and \(\bf{X'}\) shares exactly \(j\) codewords with \(\bf{X}_{S_1}\). Thus, we have: \[\begin{aligned} \label{E and E'} P_{e|1} &\leq& \sum_{i=1}^{K}P(E_i) \nonumber\\ &=& \sum_{i=1}^{K}\Big( P(E'_i) + \sum_{j=0}^{K-i-1} {K-i-1 \choose j}P(E'_{K-j}) \Big) \nonumber\\ &\leq&\revisedr{2^K} \sum_{i=1}^{K}P(E'_i). \end{aligned}\] The last inequality in [\[E and E\'\]](#E and E'){reference-type="eqref" reference="E and E'"} is loose, yet it suffices for the proof in the regime where \(K=O(1)\). A more delicate bounding technique might be required if \(K\) might grow with \(N\). We now bound \(P(E'_i)\). Particularly, we will establish the following lemma. # Problem Formulation {#formulation} In *SGT*, a legitimate user desires to identify a small unknown subset \(\mathcal{K}\) of defective items from a larger set \(\mathcal{N}\), while the number of measurements \(T\) and keeping the eavesdropper, which is able to observe a subset of the tests results, ignorant regarding the status of the \(\mathcal{N}\) items. \(N=|\mathcal{N}|\), \(K=|\mathcal{K}|\) denote the total number of items, and the number of defective items, respectively. [^1] Throughout the paper, we use boldface to denote matrices, capital letters to denote random variables, lower case letters to denote their realizations, and calligraphic letters to denote the alphabet. Logarithms are in base \(2\) and \(h_b(\cdot)\) denotes the binary entropy function. gives a graphical representation of the model. In general, and regardless of security constraints, GT is defined by a testing matrix \[\textbf{X}=[X_{1}^{T};X_{2}^{T};\ldots; X_{N}^{T}] \in \{0,1\}^{N\times T},\] where each row corresponds to a separate item \(j\in\{1,\ldots,N\}\), and each column corresponds to a separate pool test \(t\in\{1,\ldots,T\}\). For the \(j\)-th item, \[X_{j}^{T}=\{X_{j}(1),\ldots, X_{j}(T)\}\] is a binary row vector, with the \(t\)-th entry \(X_{j}(t)=1\) if and only if item \(j\) participates in the \(t\)-th test. If \(A_j\in \{0,1\}\) an indicator function for the \(j\)-th item, determining whether it belongs to the defective set, i.e., \(A_j=1\) if \(j \in \mathcal{K}\) and \(A_j = 0\) otherwise, \[\revised{Y(t)=\bigvee_{j=1}^{N}X_{j}(t)A_j=\bigvee_{d\in \mathcal{K}}X_{d}(t)},\] where \(\bigvee\) is used to denote the boolean OR operation. In SGT, we assume an eavesdropper observes a noisy vector \(Z^{T}=\{Z(1),\ldots, Z(T)\}\), generated from the outcome vector \(Y^{T}\). In the erasure case considered in the work, the probability of erasure is \(1-\delta\), i.i.d. for each test. That is, on average, \(T\delta\) outcomes are not erased and are accessible to the eavesdropper via \(Z^{T}\). Therefore, in the erasure case, if \(B_t\in \{1,?\}\) is an erasure indicator function for the \(t\)-th pool test, i.e., \(B_t=1\) with probability \(\delta\), and \(B_t=?\) with probability \(1-\delta\), the eavesdropper observes \[Z(t)=Y(t)B_t= \revised{\left(\bigvee_{j=1}^{N}X_{j}(t)A_j\right)B_t},\text{ }\text{ } t=1,\ldots,T.\] Denote by the index of the *subset of defective items*. We assume \(W\) is uniformly distributed, that is, there is no *a priori* bias any specific subset.[^2] Further, denote by \(\hat{W}(Y^T)\) the index recovered by the legitimate decoder, after observing \(Y^T\). In this work, we assume that the mixer may use a *randomized* testing matrix. In this case, the random bits used are know only to the mixer, and are not assumed to be shared with the decoder. In other words, the "codebook\" which consists of all possible testing matrices is known to all parties, Alice, Bob and Eve. \(\textbf{X}\), the random value is not shared with Bob or Eve. We refer to the codebook consisting of all possible matrices, together with the decoder at Bob's side as SGT algorithm. As we are interested in the asymptotic behavior, In , we give an algorithm which applies to any \(K\).[^3] The following definition lays out the goals of SGT algorithms. # Conclusions {#conc} In this paper, we proposed a novel non-adaptive SGT algorithm, which with parameters \(N,K\) and \(T\) is asymptotically *reliable* and *secure*. Specifically, when the fraction of tests observed by Eve is \(0 \leq \delta <1\), we prove that the number of tests required for both correct reconstruction at the legitimate user (with high probability) and negligible mutual information at Eve's side is \(\frac{1}{1-\delta}\) times the number of tests required with no secrecy constraint. We further provide sufficiency and necessity bounds on the number of tests required in the SGT model to obtains both, *reliability* and *secrecy* constraints. Moreover, we analyze in the proposed secure model, computationally efficient algorithms at the legitimate decoder, previously considered for the non-secure GT in the literature which identify the definitely non-defective items. # Efficient Algorithms {#efficient_algorithms} The achievability result given in uses a random codebook and ML decoding at the legitimate party. The complexity burden in ML, however, prohibits the use of this result for large \(N\). In this section, we derive and analyze an efficient decoding algorithm, which maintains the reliability result using a much simpler decoding rule, at the price of only slightly more tests. The secrecy constraint, as will be clear, is maintained by construction, as the codebook and mixing process do not change compared to the achievability result given before. Moreover, the result in this section will hold for any \(K\), including even the case were \(K\) grows linearly with \(N\). Specifically, we assume the same codebook generation and the testing procedure given in , and analyze the *Definite Non-Defective* (DND) algorithm, previously considered for the non-secure GT in the literature. The decoding algorithm at the legitimate user is as follows. Bob attempts to match the rows of **X** with the outcome vector \(Y^T\). If a particular row \(j\) of **X** has the property that all locations \(t\) where it has \(1\), also corresponds to a \(1\) in \(Y(t)\), then that row *can correspond to a defective item*. If, however, the row has \(1\) at a location \(t\) where the output has \(0\), then it is not possible that the row corresponds to a defective item. The problem, however, when considering the code construction in this paper for SGT, is that the decoder does not know which row from each bin was selected for any given item. Thus, it takes a conservative approach, and declares an item as defective if at least one of the rows in its bin signals it may be so. An item is not defective only if *all the rows in its bin prevent it from being so*. It is clear that this decoding procedure has no false negatives, as a defective item will always be detected. It may have, though, false positives. A false positive may occur if all locations with ones in a row corresponding to a non-defective item are hidden by the ones of other rows corresponding to defective items and selected by the mixer. To calculate the error probability, fix a row of \(\textbf{X}\) corresponding to a non-defective item (a row in its bin). Let \(j_1;\ldots;j_k\) index the rows of \(\textbf{X}\) corresponding to the \(K\) defective items, and selected by the mixer for these items (that is, the rows which were actually added by the Boolean channel). An error event associated with the fixed row occurs if at any test where that row has a \(1\), at least one of the entries \(X_{j_1}(t), \ldots, X_{j_k}(t)\) also has a \(1\). The probability for this to happen, per column, is \(p(1-(1-p)^K)\). Hence, the probability that a test result in a fixed row is hidden from the decoder, in the sense that it cannot be declared as non defective due to a specific column, is \[p(1-(1-p)^K)+(1-p) = 1-p(1-p)^K.\] Since this should happen for all \(T\) columns, the error probability for a fixed row is \(\left(1-p(1-p)^K\right)^T\). Now, to compute the error probability for the entire procedure we take a union bound over all \(M(N-K)\) rows corresponding to non-defective items. As a result, we have \[\begin{aligned} \label{eq:PeComp} P_e &\leq& M(N-K)\left(1-p(1-p)^{K}\right)^T \nonumber\\ & \stackrel{(a)}{\leq}& MN\left(1-\frac{y}{K}\left(1-\frac{y}{K}\right)^{K}\right)^{\beta K\log N} \nonumber\\ &=& MN\left(1-\frac{y}{K}\left(1-\frac{y}{K}\right)^{K-1}\left(1-\frac{y}{K}\right)\right)^{\beta K\log N} \nonumber\\ & \stackrel{(b)}{\leq}& MN\left(\left(1-\frac{y\left(1-\frac{y}{K}\right)}{Ke^{y}}\right)^{K}\right)^{\beta\log N} \nonumber\\ & \stackrel{(c)}{\leq}& MNe^{-y\left(1-\frac{y}{K}\right)e^{-y}\beta\log N}\nonumber\\ &\leq& MN^{1-y\left(1-\frac{y}{K}\right)e^{-y}\beta\frac{1}{\ln 2}} \nonumber\\ & \stackrel{(d)}{=}& MN^{1-\frac{1}{2}\beta (1-\frac{\ln2}{K})} \nonumber\\ &\stackrel{(e)}{=}& 2^{\beta (\delta-\epsilon) \log N}N^{1-\frac{1}{2}\beta (1-\frac{\ln2}{K})} \nonumber\\ &=& N^{\beta (\delta-\epsilon)}N^{1-\frac{1}{2}\beta (1-\frac{\ln2}{K})} \nonumber\\ &\leq& N^{1-\beta\left(\frac{1}{2}(1-\frac{\ln 2}{K})-\delta \right)}. \end{aligned}\] \[\begin{aligned} \label{eq:PeComp} P_e &\leq& M(N-K)\left(1-p(1-p)^{K}\right)^T \nonumber\\ & \stackrel{(a)}{\leq}& MN\left(1-\frac{y}{K}\left(1-\frac{y}{K}\right)^{K}\right)^{\beta K\log N} \nonumber\\ &=& MN\left(1-\frac{y}{K}\left(1-\frac{y}{K}\right)^{K-1}\left(1-\frac{y}{K}\right)\right)^{\beta K\log N} \nonumber\\ & \stackrel{(b)}{\leq}& MN\left(\left(1-\frac{y\left(1-\frac{y}{K}\right)}{Ke^{y}}\right)^{K}\right)^{\beta\log N} \nonumber\\ & \stackrel{(c)}{\leq}& MNe^{-y\left(1-\frac{y}{K}\right)e^{-y}\beta\log N}\nonumber\\ &\leq& MN^{1-y\left(1-\frac{y}{K}\right)e^{-y}\beta\frac{1}{\ln 2}} \nonumber\\ & \stackrel{(d)}{=}& MN^{1-\frac{1}{2}\beta (1-\frac{\ln2}{K})} \nonumber \end{aligned}\] \[\begin{aligned} &\hspace{-3.4cm} \stackrel{(e)}{=}& 2^{\beta (\delta-\epsilon) \log N}N^{1-\frac{1}{2}\beta (1-\frac{\ln2}{K})} \nonumber\\ &\hspace{-3.4cm}=&N^{\beta (\delta-\epsilon)}N^{1-\frac{1}{2}\beta (1-\frac{\ln2}{K})} \nonumber\\ &\hspace{-3.4cm}\leq&N^{1-\beta\left(\frac{1}{2}(1-\frac{\ln 2}{K})-\delta \right)}. \end{aligned}\] In the above, (a) follows by taking \(p=y/K\) and setting \(T\) as \(\beta K\log N\), for some positive \(y\) and \(\beta\), to be defined. (b) follows since \(e^{-y} \leq (1-y/n)^{n-1}\) for small \(y>0\) and any integer \(n >0\). In the sequence below, we will use it with \(y=\ln 2\), for which it is true. (c) follows since \(e^{-x} \ge (1-x/n)^{n}\) for \(x>0\) and any integer \(n >0\). (d) follows by choosing \(y=\ln 2\). (e) is by setting \(M=2^{T\frac{\delta-\epsilon}{K}}\) and substituting the value for \(T\). The result in [\[eq:PeComp\]](#eq:PeComp){reference-type="eqref" reference="eq:PeComp"} can be interpreted as follows. As long as \(\delta\), the leakage probability at the eavesdropper, is smaller than \(\frac{1}{2}(1-\frac{\ln 2}{K})\), choosing \(T=\beta K \log N\) with a large enough \(\beta\) results in an exponentially small error probability. For example, for large enough \(K\) and \(\delta=0.25\), one needs \(\beta > 4\), that is, about \(4K\log N\) tests to have an exponentially small (with \(N\)) error probability while using an efficient decoding algorithm. To see the dependence of the error probability on the number of tests, denote \[\epsilon = \beta\left(\frac{1}{2}\left(1-\frac{\ln 2}{K}\right)-\delta\right)-1.\] Then, if the number of tests satisfies \[T \ge \frac{1+\epsilon}{\frac{1}{2}(1-\frac{\ln 2}{K})-\delta} K \log N\] one has \[P_e \leq N^{-\epsilon}.\] Thus, while the results under ML decoding () show that any value of \(\delta <1\) is possible (with a \(\frac{1}{1-\delta}\) toll on \(T\) compared to non-secure GT), the analysis herein suggests that using the efficient algorithm, one can have a small error probability only for \(\delta < 1/2\), and the toll on \(T\) is greater than \(\frac{1}{\frac{1}{2}-\delta}\). This is consistent with the fact that this algorithm is known to achieve only half of the capacity for non-secure GT. However, both these results may be due to coarse analysis, and not necessarily due to an inherent deficiency in the algorithm. includes simulation results of the secure DND GT algorithm proposed, compared with ML decoding and the upper and lower bounds on the performance of ML. # Introduction {#intro} The classical version of Group Testing (GT) was suggested during World War II in order to identify syphilisinfected draftees while dramatically reducing the number of required tests. Specifically, when the number of infected draftees, \(K\), is much smaller than the population size, \(N\), instead of examining each blood sample individually, one can conduct a small number of of *pooled samples*. Each pool outcome is negative if it contains no infected sample, and positive if it contains at least one infected sample. The problem is thus to identify the infected draftees via as few pooled tests as possible. (a)-(c) depicts a small example. Since its exploitation in WWII, GT utilized in numerous fields, including biology and chemistry, communications, sensor networks, pattern matching and web services. GT also found applications in the emerging field of Cyber Security, e.g., detection of significant changes in network traffic, Denial of Service attacks and indexing information for data forensics. Many which utilize GT sensitive information which should not be revealed if some of the tests leak if it is positive, at least one of the items in the pool is defective (see (d) for a short example). Accordingly, that a leakage of a fraction of the pool-tests to undesirable or malicious eavesdroppers does not give them any useful information on the status of the items. It is very important to note that *protecting GT is different from protecting the communication between the parties*. However, in GT, we do not want to assume one entity has access to all pool-tests, and can apply some encoding function before they are exposed. We also do not want to assume a mixer can add a certain substance that will prevent a third party from testing the sample. To protect GT, one should make sure that without altering mixed samples, if a fraction of them leaks, either already tested or not, information is not revealed. While the current literature includes several works on the privacy in GT algorithms for digital objects, these works are based on cryptographic schemes, assume the testing matrix is not known to all parties, a high computational burden, and, last but not least, assume the computational power of the eavesdropper is limited. *Information theoretic security* , on the other hand, if applied appropriately to GT, can offer privacy at the price of *additional tests*, without keys, obfuscation or assumptions on limited power. ## Main Contribution {#main-contribution .unnumbered} In this work, we formally define Secure Group Testing (SGT), suggest SGT algorithms based on informationtheoretic principles and analyse their performance. In the considered model, there is an eavesdropper observe part of the . The goal of the test designer is to design the tests such that a legitimate decoder can decode the status of the items ( defective or not) with an arbitrarily small error probability. long as the eavesdropper gains only part of the output vector (a fraction \(\delta\) cannot (asymptotically, ) gain any significant information on the status of the items. We propose a SGT code and algorithms which , as well as conditions We do so by proposing a model, which is, in a sense, analogous to a , as depicted in . The decoding algorithm is analogous to a channel decoding process, and the eavesdropped signal is the output of *an erasure channel*, namely, having only any part of the transmitted signal from the legitimate source to the legitimate receiver. Namely, we use *stochastic encoding*, and each vector pool-tests . For each item the lab picks one of the vectors in its set at random, and the item participates in the pool-tests according to this randomly chosen vector. A schematic description of our procedure is depicted in . Accordingly, by obtaining a pool-test result, without knowing the specific vectors chosen by the lab for each item, the eavesdropper may gain only negligible information regarding the items themselves. Specifically, we show that even though the pool-tests in which each item participated chosen randomly and even though the legitimate user does not know *a-priori* in which pool-tests each item has participated, the legitimate user while the eavesdropper, observing only a subset of the pool-test results, will have no significant information regarding the status of the items. The structure of this work is as follows. In , we present an extensive survey and summarize the related work. In , a SGT model is formally described. includes our main results, with the direct proved in and converse proved in . describes efficient algorithm, and proves an upper bound on its error probability. concludes the paper. # Code Construction and a Proof for {#LowerBound} In order to keep the eavesdropper, which obtains only a fraction \(\delta\) of the outcomes, ignorant regarding the status of the items, we *randomly* map the items to the tests. Specifically, as depicted in , for each item we generate a bin, containing several rows. The number of such rows *corresponds* to the number of tests that the eavesdropper can obtain, yet, unlike wiretap channels, *it is not identical* to Eve's capacity, and should be normalized by the number of defective items. Then, for the \(j\)-th item, we randomly select a row from the \(j\)-th bin. This row will determine in which tests the item will participate. In order to rigorously describe the construction of the matrices and bins, determine the exact values of the parameters (e.g., bin size), and analyze the reliability and secrecy, we first briefly review the representation of the GT problem as a channel coding problem, together with the components required for SGT. A SGT code consists of an index set , its \(w\)-th item corresponding to the \(w\)-th subset \(\mathcal{K}\subset \{1,\ldots,N\}\); A discrete memoryless source of randomness \((\mathcal{R},p_R)\), with known alphabet \(\mathcal{R}\) and known statistics \(p_R\); An encoder, \[f: \revisedSr{\mathcal{W}} \times \mathcal{R} \rightarrow \mathcal{X}_{S_w}\in\{0,1\}^{K\times T}\] which maps the index \(W\) of the defective items to a matrix \(\textbf{X}^{T}_{S_{w}}\) of codewords, each of its rows corresponding to a different item in the index set \(S_w\). The need for a *stochastic encoder* is similar to most encoders ensuring information theoretic security, as randomness is required to confuse the eavesdropper about the actual information. Hence, we define by \(R_K\) the random variable encompassing the randomness required *for the \(K\) defective items*, and by \(M\) the number of rows in each bin. Clearly, \(M^K=H(R_K)\). At this point, important clarification is in order. The lab, of course, does not know which items are defective. Thus, operationally, it needs to select a row for each item. However, in *the analysis*, since only the defective items affect the output (that is, only their rows are to give \(Y^T\)), we refer to the "message\" as the index of the defective set \(w\) and refer only to the random variable \(R_K\) required to choose the rows in their bins. In other words, unlike the analogous communication problem, in GT, *nature* performs the actual mapping from \(W\) to \(\textbf{X}^{T}_{S_{w}}\). The mixer only mixes the blood samples according to the (random in this case) testing matrix it has. A decoder at the legitimate user is a map \[\hat{W}: \revisedSr{\mathcal{Y}}^{T} \rightarrow \revisedSr{\mathcal{W}}.\] The probability of error is \(P(\hat{W}(Y^{T})\neq W)\). The probability that an outcome test leaks to the eavesdropper is \(\delta\). We assume a memoryless model, i.e., each outcome \(Y(t)\) depends only on the corresponding input \(X_{S_w}(t)\), and the eavesdropper observes \(Z(t)\), generated from \(Y(t)\) according to \[p(Y^T,Z^T|X_{S_w})=\prod_{t=1}^{T}p(Y(t)|X_{S_w}(t))p(Z(t)|Y(t)).\] We may now turn to the detailed construction and analysis. ### Codebook Generation Using a distribution \(P(X^{T})=\prod^{T}_{i=1}P(x_i)\), for each item generate \(M\) independent and identically distributed codewords \(x^{T}(m)\), \(1 \leq m \leq M\). The codebook is depicted in the left hand side of . Reveal the codebook to Alice and Bob. We assume Eve may have the codebook as well. ### Testing For each item \(j\), the mixer/lab selects uniformly at random one codeword \(x^{T}(m)\) from the \(j\)-th bin. Therefore, the SGT matrix contains \(N\) randomly selected codewords of length \(T\), one for each item, defective or not. Amongst is an unknown subset \(X^{T}_{S_{w}}\), with the index \(w\) representing the true . An entry of the \(j\)-th random codeword is \(1\) if the \(j\)-item is a member of the designated pool test and \(0\) otherwise. ### Decoding at the Legitimate Receiver The decoder looks for a collection of \(K\) codewords \(X_{S_{\hat{w}}}^{T}\), *one from each bin*, for which \(Y^T\) is most likely. Namely, \[P(Y^{T}|X_{S_{\hat{w}}}^{T})>P(Y^{T}|X_{S_{w}}^{T}), \forall w \neq \hat{w}.\] Then, the legitimate user (Bob) declares \(\hat{W}(Y^T)\) as the set of bins in which the rows \(\hat{w}\) reside. ## Reliability Let \((\mathcal{S}^{1},\mathcal{S}^{2})\) denote a partition of the defective set \(S\) into disjoint sets \(\mathcal{S}^{1}\) and \(\mathcal{S}^{2}\), with cardinalities \(i\) and \(K-i\), respectively.[^4] Let \(I(X_{\mathcal{S}^1};X_{\mathcal{S}^2},Y)\) denote the mutual information between \(X_{\mathcal{S}^1}\) and \(X_{\mathcal{S}^2},Y\), under the i.i.d. distribution with which the codebook was generated and remembering that \(Y\) is the output of a Boolean channel. The following lemma is a key step in proving the reliability of the decoding algorithm. Next, we , which extends the results in to the codebook required for SGT. Specifically, to obtain a bound on the required number of tests as given in , we first state Lemma 2, which bounds the error probability of the ML decoder using a Gallager-type bound. Applying to the expression in , we have \[\frac{\log\binom{N-K}{i}M^i}{I(X_{\mathcal{S}^1};X_{\mathcal{S}^2},Y)} \leq \frac{\log\binom{N-K}{i}M^i}{\frac{i}{K}}.\] Hence, substituting \(M=2^{T\frac{\delta-\revised{\epsilon_K}}{K}}\), a sufficient condition for reliability is \[\begin{aligned} T &\ge& \max_{1 \leq i \leq K}\frac{1+\varepsilon}{\frac{i}{K}}\left[ \log\binom{N-K}{i} + \frac{i}{K}T(\delta-\revised{\epsilon_K}) \right] \end{aligned}\] Rearranging terms results in \[\begin{aligned} \label{eq:reduce_h} T & \ge & \max_{1 \leq i \leq K} \frac{1}{1-\revised{\delta +\epsilon_K-\varepsilon\delta+\varepsilon\epsilon_K}}\frac{1+\varepsilon}{i/K}\log\binom{N-K}{i},\nonumber \end{aligned}\] Noting that this is for large \(K\) and \(N\), and that \(\varepsilon\) is independent of them, achieves the bound on \(T\) provided in and reliability is established. ## Information Leakage at the Eavesdropper {#LowerBoundLeakage} We now prove the security constraint is met. Hence, we wish to show that \(I(W;Z^{T})/T\rightarrow 0\), as \(T\rightarrow \infty\). Denote by \(\mathcal{C}_T\) the random codebook and by \(X_{\mathcal{S}}^T\) the set of codewords corresponding to the true defective items. We have, \[\begin{aligned} && \hspace{-0.8cm}\textstyle\textstyle \frac{1}{T}I(W;Z^T|\mathcal{C}_{T}) = \frac{1}{T}\left(I(W,R_K;Z^T|\mathcal{C}_{T})-I(R_K;Z^T|W,\mathcal{C}_{T})\right)\nonumber\\ &\hspace{-0.6cm} \stackrel{(a)}{=} &\hspace{-0.5cm}\textstyle \frac{1}{T}\left(I(X_{\mathcal{S}}^T;Z^T|\mathcal{C}_{T})-I(R_K;Z^T|W,\mathcal{C}_{T})\right)\nonumber\\ &\hspace{-0.6cm} = &\hspace{-0.5cm}\textstyle \frac{1}{T}(I(X_{\mathcal{S}}^T;Z^T|\mathcal{C}_{T})-H(R_K|W,\mathcal{C}_{T})+H(R_K|Z^T,W,\mathcal{C}_{T}))\nonumber\\ &\hspace{-0.6cm} \stackrel{(b)}{=} & \hspace{-0.5cm}\textstyle \frac{1}{T}\left(I(X_{\mathcal{S}}^T;Z^T|\mathcal{C}_{T})-H(R_K)+H(R_K|Z^T,W,\mathcal{C}_{T})\right)\nonumber\\ &\hspace{-0.6cm} \stackrel{(c)}{=}&\hspace{-0.5cm} I(X_{\mathcal{S}};Z|\mathcal{C}_{T})-\frac{1}{T}K\log M + \frac{1}{T}H(R_K|Z^T,W,\mathcal{C}_{T})\nonumber\\ &\hspace{-0.6cm} \stackrel{(d)}{\leq}&\hspace{-0.5cm} \delta-\frac{1}{T}K\left(T \frac{\delta-\epsilon^{\prime}}{K}\right) + \frac{1}{T}H(R_K|Z^T,W,\mathcal{C}_{T})\stackrel{(e)}{\leq} \epsilon_T + \revisedSr{\epsilon^{\prime}},\nonumber \end{aligned}\] where \(\epsilon_T \to 0\) as \(T \to \infty\). (a) is since there is a \(1:1\) correspondence between \((W,R_K)\) and \(X_{\mathcal{S}}^T\); (b) is since \(R_K\) is independent of \(W\) and \(\mathcal{C}_T\); (c) is since in this *direct result* the codebook is defined by the construction and is memoryless, as well as the channel; (d) is since by choosing an i.i.d. distribution for the codebook one easily observes that \(I(X_{\mathcal{S}};Z| \mathcal{C}_T) \leq \delta\). Finally, (e) is for the following reason: Given \(W\) and the codebook, Eve has a perfect knowledge regarding the *bins from which the codewords were selected*. It requires to see whether she can indeed estimate \(R_K\). Note that the channel Eve sees in this case is the following *multiple access channel*: each of the defective items can be considered as a "user\" with \(M\) messages to transmit. Eve's goal is to decode the messages from all users. This is possible if the rates at which the users transmit are within the capacity region of this MAC. Indeed, this is a (binary) Boolean MAC channel, followed by a simple erasure channel. The sum capacity cannot be larger than \(\delta\), and this sum capacity is easily achieved by letting one user transmit at a time, or, in our case, # Main Results {#main results} Under the model definition given in , our main results are the following sufficiency (direct) and necessity (converse) conditions, characterizing the maximal number of tests required to guarantee both reliability and security. The proofs are deferred to and . ## Direct (Sufficiency) The sufficiency part is given by the following theorem. The construction of the SGT algorithm, together with the proofs of reliability and secrecy are deferred to Section [6](#LowerBound){reference-type="ref" reference="LowerBound"}. However, a few important remarks are in order now. First, rearranging terms in [\[main_result_eq\]](#main_result_eq){reference-type="ref" reference="main_result_eq"}, we have \[\label{main_result} T \geq \frac{1}{1-\delta} \max_{i=1,\ldots ,K}\frac{(1+\varepsilon)K}{i}\log\binom{N-K}{i}.\] That is, compared to only a reliability constraint, the number of tests required for *both reliability and secrecy* is increased by the multiplicative factor \(\frac{1}{1-\delta}\), where, again, \(\delta\) is the leakage probability at the eavesdropper. ## Converse (Necessity) The necessity part is given by the following theorem. ## Secrecy capacity in SGT Returning to the analogy in between channel capacity and group testing, one might define by \(C_s\) the (asymptotic) minimal threshold value for \(\log\binom{N}{K}/T\), above which no reliable and secure scheme is possible. Under this definition, the result in this paper show that \(C_s = (1-\delta)C\), where \(C\) is the capacity without the security constraint. Clearly, this can be written as \[C_s = C-\delta C,\] raising the usual interpretation as the *difference* between the capacity to the legitimate decoder and that to the eavesdropper. Note that as the effective number of tests Eve sees is \(T_e = \delta T\), her GT capacity is \(\delta C\). ## Efficient Algorithms {#efficient-algorithms} Under the SGT model definition given in , we further consider a computationally efficient algorithm at the legitimate decoder. Specifically, we analyze the *Definite Non-Defective* (DND) algorithm (originally called *Combinatorial Orthogonal Matching Pursuit* (COMP)), considered for the non-secure GT model in the literature. The theorem below states that indeed efficient decoding (with arbitrarily small error probability) and secrecy are possible, at the price of a higher \(T\). Interestingly, the theorem applies to any \(K\), and not necessarily only to \(K = O(1)\). This is, on top of the reduced complexity, an important benefit of the suggested algorithm. The construction of the DND GT algorithm, together with the proofs of reliability and secrecy are deferred to . Clearly, the benefits of the algorithm above come at the price of additional tests and a smaller range of \(\delta\) it can handle. # Background and Related Work {#BooleanCompressed} ## Group-testing ### Performance Bounds GT can be *non-adaptive*, where the testing matrix is designed beforehand, , where each new test can be designed while taking into account previous test *results*, or a combination of the two, where testing is adaptive, yet with batches of non-adaptive tests. It is also important to distinguish between *exact recovery* and a *vanishing probability of error*. To date, the best known lower bound on the number of tests required (non-adaptive, exact recovery) is \(\Omega(\frac{K^2}{\log K}\log N)\). The best known explicit constructions were given in, resulting in \(O(K^2 \log N)\). However, focusing on exact recovery requires more tests, and forces a combinatorial nature on the problem. Settling for high probability reconstructions allows one to reduce the number of tests to the order of \(K\log N\).[^5] For example, see the channel-coding analogy given in. A similar analogy to wiretap channels will be at the basis of this work as well. In fact, probabilistic methods with an error probability guarantee appeared in, without explicitly mentioning GT, yet showed the \(O(K \log N)\) bound. Additional probabilistic methods can be found in for *support recovery*, or in, when an interesting *phase transition* phenomenon was observed, yielding tight results on the threshold (in terms of the number of tests) between the error probability approaching one or vanishing. ### A Channel Coding Interpretation As mentioned, the analogy to channel coding proved useful. defined the notion of *group testing capacity*, that is, the value of \(\lim_{N \to \infty} \frac{\log{{N}\choose{K}}}{T}\) under which reliable algorithms exist, yet, over which, no reliable reconstruction is possible. A converse result for the Bernoulli, non-adaptive case was given in. Strong converse results were given in, again, building on the channel coding analogy, as well as converses for noisy GT. In, adaptive GT was analyzed as a channel coding *with feedback* problem. ### Efficient Algorithms {#efficient-algorithms-1} A wide variety of techniques were used to design efficient GT decoders. Results and surveys for early non-adaptive decoding algorithms were given in. Moreover, although most of the works described above mainly targeted fundamental limits, some give efficient algorithms as well. In the context of this work, it is important to mention the recent COMP, DD and SCOMP algorithms, concepts from which we will use herein. [^6] - - - ### - - - . # Information Leakage at the Eavesdropper (Strong Secrecy) {#strong_secrecy} We wish to show that \(I(W;Z^{T})\rightarrow 0\). Denote by \(\mathcal{C}_T\) the random codebook and by \(\textbf{X}_{\mathcal{S}_w}^T\) the set of codewords corresponding to the true defective items. We assumed \(W \in \{1,\ldots, {N \choose K}\}\) is uniformly distributed, that is, there is no *a-priori* bias to any specific subset. Further, the codebook includes independent and identically distributed codewords. The eavesdropper, observing \(Z^T\), wish to decode the true \(K\) independent and identically distributed codewords, which correspond to the defective items, one of \({N \choose K}\) subsets. To analyze the information leakage at the eavesdropper, we note that the channel Eve sees in this case is the following *multiple access channel*. Each of the items can be considered as a user\" with \(M\) specific codewords. Eve's goal is to identify the active users. Note the Eve's channel can be viewed as (binary) Boolean MAC, followed by a BEC\((1-\delta)\). The sum capacity to Eve cannot be larger than \(\delta\). . Consequently, Eve may obtain a sub-matrix \(\tilde{\textbf{Z}}\) of possibly transmitted codewords, where from each codeword Eve sees at most a capacity of \(\delta/K\). However, in this analysis, we help Eve even further to get the true sub-matrix \(\tilde{\textbf{Z}}\), with the maximum rate \(\delta/K\). That is, we assume Eve gets the full rate, identifies the codewords of the users except for the erasures. We give Eve the power not to be confused from the existence of other, \(N-K\) users which did not transmit eventually. Providing this information to Eve only makes her stronger and thus a coding scheme that succeeds to keep Eve ignorant, will also succeed against the original Eve. Now, once Eve obtains the true sub-matrix \(\tilde{\textbf{Z}}\), since the codebook and the subsets of the items was generated independently and identically, we will analyze the information that leaks at the eavesdropper from each possibly transmitted codeword in \(\tilde{\textbf{Z}}\) separately. Namely, per original codeword \(X_j^T\), on average, \(\frac{\delta}{K}T\) from the \(T\) outcomes are not erased and are accessible to the eavesdropper via \(\tilde{Z}_{\tilde{j}}^{T}\). Thus, out of the \(M\) independent and identically distributed codewords, there is an exponential number of codewords, that from eavesdropper perspective, may have participated in \(\textbf{X}_{\mathcal{S}_w}^T\) and could have resulted in the same \(\tilde{Z}_{\tilde{j}}^{T}\). These consistent codewords are exactly those that lie in a ball around \(\tilde{Z}_{\tilde{j}}^{T}\) of radius \(d \approx (1-\frac{\delta}{K}) T\) as depicted in . The eavesdropper does not know what is the codeword \(X^T_j\) in \(\textbf{X}_{\mathcal{S}_w}^T\), which was selected by the mixer and she even does not know what \(d_H = \|X^T_j-\tilde{Z}_{\tilde{j}}^{T}\|\) is exactly (where \(\|\cdot\|\) is the hamming distance), other than the fact that \(d_H \approx (1-\frac{\delta}{K})T\). However, we help Eve by providing \(d\) and by choosing a small yet exponential set of codewords (of size \(2^{T\frac{\epsilon}{K}}\), for an appropriately small \(\epsilon\)) from the codebook \(\mathcal{C}\), chosen from the set of all codewords at distance \(d\) from \(\tilde{Z}_{\tilde{j}}^{T}\), with the additional guarantee that it also contains the true \(X^T_j\). We refer to this set as the oracle-given set \(\textbf{X}_{\mathcal{O}racel}^T\). Again, providing this information to Eve only makes her stronger and thus a coding scheme that succeeds to keep Eve ignorant, will also succeed against the original Eve. Conditioned on Eve's view \(\tilde{Z}_{\tilde{j}}^{T}\) and \(\textbf{X}_{\mathcal{O}racel}^T\), each of the codewords in \(\textbf{X}_{\mathcal{O}racel}^T\) is equally likely to have been participated in the pool tests. Nonetheless, Eve still has a reasonable amount of uncertainty about which of the \(2^{T\frac{\epsilon}{K}}\) codewords was actually participated indeed, this is the uncertainty that we leverage in our analysis. We define, for any \(\tilde{Z}_{\tilde{j}}^{T}\) and \(d\), a ball and a shell as \[\mathcal{V}ol(\tilde{Z}_{\tilde{j}}^{T},d)=\{X^T_j:d_H(X^T_j,\tilde{Z}_{\tilde{j}}^{T})\leq d\},\] \[\mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)=\{X^T_j:d_H(X^T_j,\tilde{Z}_{\tilde{j}}^{T})= d\}.\] Hence, we define the probability for a codeword to fall in shell as \[\begin{gathered} Pr(X^T_j \in\mathcal{C} \cap X^T_j\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d))\\ = \frac{\mathcal{V}ol(\mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d))}{2^T} = \frac{2^{\left(1-\frac{\delta}{K}\right)T}}{2^{T}}. \end{gathered}\] \[Pr(X^T_j \in\mathcal{C} \cap X^T_j\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)) = \frac{\mathcal{V}ol(\mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d))}{2^T} = \frac{2^{\left(1-\frac{\delta}{K}\right)T}}{2^{T}}.\] For each item we have \(M =2^{\left(\frac{\delta+\revised{\revisedSr{\epsilon^{\prime}}}}{K}\right)T}\) codewords. Thus, *on average*, the number of codewords Eve sees on a shell, per defective item is \[\begin{gathered} |\{w:X^T_j(w)\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)\}|\\ = \frac{2^{\left(\frac{\delta+\revised{\revisedSr{\epsilon^{\prime}}}}{K}\right) T}\cdot 2^{\left(1-\frac{\delta}{K}\right)T}}{2^T} = 2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}. \end{gathered}\] \[|\{w:X^T_j(w)\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)\}| = \frac{2^{\left(\frac{\delta+\revised{\revisedSr{\epsilon^{\prime}}}}{K}\right) T}*2^{\left(1-\frac{\delta}{K}\right)T}}{2^T} = 2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}.\] Hence, we can conclude that, on average, for every item, Eve has quite a few options in a particular shell. Now that we established that the average number of codewords per item \(|\{w:X^T_j(w)\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)\}|\) is \(2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}\), we wish to calculate the probability that the *actual number of options* deviates from the average by more than \(\varepsilon\). We define \[\begin{gathered} \mathcal{E}_{C_1}(\tilde{Z}_{\tilde{j}}^{T},d):= Pr\{ (1-\varepsilon)2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}\leq \\ |\{w:X^T_j(w)\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)\}| \leq (1+\varepsilon)2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}\} \end{gathered}\] \[\mathcal{E}_{C_1}(\tilde{Z}_{\tilde{j}}^{T},d):= Pr\{ (1-\varepsilon)2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}\leq |\{w:X^T_j(w)\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)\}| \leq (1+\varepsilon)2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}\}.\] Let us fix a pair \(\tilde{Z}_{\tilde{j}}^{T}\) and \(d\) for large enough \(T\). By the Chernoff bound, and taking union bound over \(\tilde{Z}_{\tilde{j}}^{T}\) and \(d\), we have \[Pr(\mathcal{E}_{C_1}(\tilde{Z}_{\tilde{j}}^{T},d)) \geq 1-2^{-\varepsilon^{\prime}2^{T\frac{\revised{\revisedSr{\epsilon^{\prime}}}}{K}}}.\] Due to the super exponential decay in \(T\), even when we take a union bound over all the codewords and all shells, we get that the probability that a codeword will have \(|\{w:X^T_j(w)\in \mathcal{S}h(\tilde{Z}_{\tilde{j}}^{T},d)\}|\) options significantly different than \(2^{T\frac{\revised{\epsilon^{\prime}}}{K}}\) is very small. Actually, with high probability, Eve has almost the same number of options per item. Hence, for Eve all the codewords are almost equiprobable. In other words, Eve distribution on the items converges super-exponentially fast to a uniform one, hence \(H(W|Z^T)\rightarrow H(W)\) and we have \(I(W;Z^{T})\rightarrow 0\). # Converse (Necessity) {#converse} In this section, we derive the necessity bound on the required number of tests. Let \(\bar{Z}\) denote the random variable corresponding to the tests which are not available to the eavesdropper. Hence, \(Y = (Z,\bar{Z})\). By Fano's inequality, if \(P_e\rightarrow 0\), we have \[H(W|Y) \leq T\epsilon'_T,\] where \(\epsilon'_T \rightarrow 0\) as \(T\rightarrow\infty\). Moreover, the secrecy constraint implies \[\label{eq:Necessity19} I(W;Z)\leq T\epsilon''_T,\] where \(\epsilon''_T\rightarrow 0\) as \(T\rightarrow \infty\). Consequently, \[\begin{aligned} \label{eq:R_s} \log \binom{N}{K} &=& H(W) \nonumber\\ &=& I(W;Y) + H(W|Y) \nonumber\\ &\stackrel{(a)}{\leq}& I(W;Z,\bar{Z}) + T\epsilon'_T \nonumber\\ &=& I(W;Z) + I(W;\bar{Z}|Z) + T\epsilon'_T \nonumber\\ &\stackrel{(b)}{\leq} & I(W;\bar{Z}|Z) + T\epsilon'_T + T\epsilon''_T \nonumber\\ &\stackrel{(c)}{\leq} & I(\textbf{X}_{\mathcal{S}_w};\bar{Z}|Z) + T(\epsilon'_T + \epsilon''_T) \nonumber\\ &=& H(\bar{Z}|Z)-H(\bar{Z}|\textbf{X}_{\mathcal{S}_w},Z) + T(\epsilon'_T + \epsilon''_T) \nonumber\\ &\stackrel{(d)}{\leq} & H(\bar{Z}) + T(\epsilon'_T + \epsilon''_T) \nonumber \end{aligned}\] where (a) follows from Fano's inequality and since \(Y = (Z,\bar{Z})\), (b) follows from [\[eq:Necessity19\]](#eq:Necessity19){reference-type="eqref" reference="eq:Necessity19"}, (c) follows from the Markov chain \(W\rightarrow X_{\mathcal{S}_w}^{T} \rightarrow Y \rightarrow Z\) and (d) is since conditioning reduces entropy. We now evaluate \(H(\bar{Z})\). Denote by \(\mathcal{\bar{E}}\) the set of tests which are not available to Eve and by \(\bar{E}_\gamma\) the event \(\{|\mathcal{\bar{E}}| \leq T(1-\delta)(1+\gamma)\}\) for some \(\gamma >0\). We have \[\begin{aligned} H(\bar{Z}) &=& P(\bar{E}_\gamma)H(\bar{Z}| \bar{E}_\gamma) + P(\bar{E}^c_\gamma)H(\bar{Z}| \bar{E}^c_\gamma) \\ &\leq& T(1-\delta)(1+\gamma) + TP(\bar{E}^c_\gamma) \\ & \leq& T(1-\delta)(1+\gamma) + T2^{-T(1-\delta)f(\gamma)}, \end{aligned}\] where the last inequality follows from the Chernoff bound for i.i.d. Bernoulli random variables with parameter \((1-\delta)\) and is true for some \(f(\gamma)\) such that \(f(\gamma)>0\) for any \(\gamma > 0\). Thus, we have \[\log \binom{N}{K} \leq T(1-\delta)(1+\gamma) + T2^{-T(1-\delta)f(\gamma)} + T(\epsilon'_T + \epsilon''_T).\] That is, \[T \ge \frac{1-\epsilon_T}{1-\delta}\log \binom{N}{K},\] for some \(\epsilon_T\) such that \(\epsilon_T \to 0\) as \(T \to \infty\). This completes the converse proof. [^1]: [^2]: [^3]: [^4]: [^5]: A simple information theoretic argument explains a lower bound. There are \(K\) defectives out of \(N\) items, hence \(N \choose K\) possibilities to cover: \(\log {N \choose K}\) bits of information. Since each test carries at most one bit, this is the amount of tests required. Stirling's approximation easily shows that for \(K \ll N\), the leading factor of that is . [^6]:

{'timestamp': '2018-01-17T02:06:19', 'yymm': '1607', 'arxiv_id': '1607.04849', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04849'}

# Introduction Let \(T\) be a bounded operator on a complex Hilbert space \(\mathcal{H}\). Then \(T\) is called complex symmetric if there exists a conjugation \(C\) such that \(T=CT^*C\). Here a conjugation is a conjugate-linear, isometric involution on \(\mathcal{H}\). The operator \(T\) may also be called \(C\)-symmetric if \(T\) is complex symmetric with respect to a specifical conjugation \(C\). For more details about complex symmetric operators one may turn to and. In this paper, we are particularly interested in the complex symmetry of composition operators induced by analytic self-maps of \(D\). This subject was started by Garcia and Hammond in. Recall that for each analytic self-map \(\varphi\) of the unit disk \(D\), the composition operator given by \[C_\varphi f=f\comp\varphi\] is always bounded on \(H^2(D)\). Here, the Hardy-Hilbert space \(H^2(D)\) is the set of analytic functions on \(D\) such that \[||f||_{H^2}^2=\sup_{0<r<1}\int_{0}^{2\pi}|f(re^{i\theta})|^2\frac{d\theta}{2\pi}<\infty.\] Several simple examples of complex symmetric composition operators on \(H^2(D)\) arise immediately. For example, every normal operator is complex symmetric (see ), so when \(\varphi(z)=sz\) with \(|s|\leqslant 1\), \(C_\varphi\) is normal hence complex symmetric on \(H^2(D)\). Also, Theorem 2 in states that each operator satisfying a polynomial equation of order \(2\) is complex symmetric. So when \(\varphi\) is an elliptic automorphism of order \(2\), then \(C_\varphi^2=I\), thus \(C_\varphi\) is complex symmetric on \(H^2(D)\). In Waleed Noor find the conjugation \(C\) such that \(C_\varphi\) is \(C\)-symmetric when \(\varphi\) is an elliptic automorphism of order \(2\). However, we are still far away from our final destination: finding out all composition operators which are complex symmetric on \(H^2(D)\). The first step is, of course, to determine the complex symmetric composition operators induced by automorphisms. Bourdon and Waleed Noor considered this problem in. The next Proposition is Proposition 2.1 in Thanks to this proposition, we only need to investigate the elliptic automorphisms of the unit disk \(D\). It turns out that things depend much on the orders of the automorphisms. The next proposition is one of the main result in. However, the order 3 elliptic case remains as an open question, which is posed by Bourdon and Waleed Noor in : The aim of this paper is to solve this problem. By dong this we complete the project of finding out all invertible composition operators which are complex symmetric \(H^2(D)\). In our main result Theorem [\[main\]](#main){reference-type="ref" reference="main"}, we prove that if \(\varphi\) is an elliptic automorphism of order \(3\) and not a rotation, then the composition operator \(C_\varphi\) cannot be complex symmetric on \(H^2(D)\). Then we can come to our final result as follows. It will be given as Corollary [\[last\]](#last){reference-type="ref" reference="last"} in the third section. # Preliminary The Hardy space \(H^2(D)\) is naturally a Hilbert space, with the inner product \[\langle f,g\rangle=\sup_{0<r<1}\int_{0}^{2\pi}f(re^{i\theta})\overline{g(re^{i\theta})}\frac{d\theta}{2\pi}.\] For each \(w\in D\), let \[K_w(z)=\frac{1}{1-\overline{w}z}.\] Then \(K_w\in H^2(D)\) is the reproducing kernel at the point \(w\), i.e., \[\langle f,K_w\rangle=f(w)\] for all \(f\in H^2(D)\). It is well known that the automorphisms of the unit disk \(D\) fall into three categories: parabolic and hyperbolic automorphisms have not fixed point in \(D\), and besides them are the elliptic automorphisms who have a unique fixed point in \(D\). Note that if the order of an automorphism \(\varphi\) is one, then \(\varphi\) is identity on \(D\). If \(\varphi\) has order two, then \(\varphi\) is of the form \[\varphi(z)=\varphi_a(z)=\frac{a-z}{1-\overline{a}z}\] for some \(a\in D\). By Propositions [\[propa\]](#propa){reference-type="ref" reference="propa"} and [\[propb\]](#propb){reference-type="ref" reference="propb"}, we only need to be concerned with the elliptic automorphisms that have order \(3\). Moreover, if the fixed point of a automorphism is \(0\), then it is a rotation and of course complex symmetric. So throughout this paper we will always assume that \(\varphi\) is an elliptic automorphism of order \(3\) with fixed point \(a\in D\backslash\{0\}\).\ For a complex symmetric operator \(T\), one should keep the following simple result in mind. The next lemma gives the eigenvectors of \(C_\varphi\) on \(H^2(D)\) when \(\varphi\) is an elliptic automorphism of order \(3\). The main calculation of this lemma is done in, with the help of Theorem 9.2 in Cowen and MacCluer's book. # Proof of the main result In this section, we will focus on the proof of our main result, i.e., Theorem [\[main\]](#main){reference-type="ref" reference="main"}, which assert that no elliptic automorphism of order \(3\) except for rotations can induce a complex symmetric composition operator on \(H^2(D)\). We would like to point out here that throughout the rest of this paper, each notation will always represent the same thing as it did initially. For example, \(\varphi\) is always a elliptic automorphism of order \(3\) in what follows, \(a\) is always the fixed point of \(\varphi\) in \(D\backslash\{0\}\), and \(\rho\) always represents the same constant \(-\frac{\overline{a}^2}{a}\cdot\frac{1-|a|^2}{1-|a|^4}\). We will assume that \(C_\varphi\) is \(C\)-symmetric with respect to some conjugation \(C\), and finally we will show this assumption leads to a contradiction. Now we start by determining the image of a certain vector under the conjugation \(C\). The notation \(\{e_j\}_{j=0}^\infty\) in Lemma [\[le\]](#le){reference-type="ref" reference="le"} is still valid in this section. Now we can prove our final result as follows. At last we can get our main result as a corollary. It gives a complete description of the automorphisms which can induce complex symmetric operators on \(H^2(D)\).

{'timestamp': '2016-08-02T02:09:19', 'yymm': '1607', 'arxiv_id': '1607.04801', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04801'}

# Introduction This paper explores Turán-type problems for a class of relational structures that can each be thought of as generalized directed hypergraphs. This class includes the standard undirected \(r\)-uniform hypergraphs that have been extensively studied in combinatorics as well as totally directed \(r\)-uniform hypergraphs where each edge is a set of \(r\) vertices under a linear ordering. Instances of this latter structure have been studied in the extremal setting by Erdős, Brown, Simonovits, Harary, and others. Other instances of this class are uniform versions of the model used to represent definite Horn formulas in the study of propositional logic and knowledge representation. The combinatorial properties of this model have been recently studied by Langlois, Mubayi, Sloan, and Gy. Turán in and by this author in and. Other structures in this class are slight variations on the \(d\)-simplex structures studied by Leader and Tan in. Turán-type extremal problems for uniform hypergraphs make up a large and well-known area of research in combinatorics that ask the following: "Given a family of forbidden \(r\)-uniform hypergraphs \(\mathcal{F}\) what is the maximum number of edges an \(r\)-uniform hypergraph on \(n\) vertices can have without containing any member of \(\mathcal{F}\) as a (not necessarily induced) subgraph?\" Such problems were named after Paul Turán due to his important early results and conjectures concerning forbidden complete \(r\)-graphs. A related question for undirected hypergraphs was proposed by Erdős known as the jumping constant conjecture. A real number \(\alpha \in [0,1)\) is called a jump for an integer \(r \geq 2\) if there exists some positive constant \(c\) which depends only on \(\alpha\) such that for any \(\epsilon > 0\) and positive integer \(l\) there exists a positive integer \(N\) for which any \(r\)-uniform hypergraph on \(n \geq N\) vertices which has edge density at least \(\alpha + \epsilon\) contains a subgraph on \(l\) vertices with edge density at least \(\alpha + c\). It is well-known that when \(r=2\), every \(\alpha \in [0,1)\) is a jump. Moreover, every \(\alpha \in \left[0, \frac{r!}{r^r}\right)\) is a jump for \(r \geq 3\). In 1984, Frankl and Rödl disproved the jumping constant conjecture when they found the first instance of a nonjump for each \(r \geq 3\). Since then many infinite sequences of nonjumps have been found, but the smallest known nonjump to date is \(\frac{5r!}{2r^r}\) for each \(r \geq 3\) determined by Frankl, Peng, Rödl, and Talbot in. The only additional jumps that have been found are all \(\alpha \in [0.2299,0.2316), \left[0.2871,\frac{8}{27} \right)\) for \(r=3\) found by Baber and Talbot in using Razborov's flag algebra method. Extremal issues of these kinds have also been considered for digraphs and multigraphs (with bounded multiplicity) in and and for the more general directed multi-hypergraphs in. In, Brown and Harary determined the extremal numbers for several types of specific directed graphs including all tournaments-that is, a digraph with one edge in some orientation between every pair of vertices. In, Brown, Erdős, and Simonovits determined the general structure of extremal sequences for every forbidden family of digraphs analogous to the Turán graphs for simple graphs. In, Brown and Simonovits proved several general extremal results about \(r\)-uniform directed \(q\)-hypergraphs. In this model the edges are ordered \(r\)-tuples of vertices with multiplicity up to \(q\) for some fixed positive integer \(q\). Among their results on this model are three that will be reproduced in this paper in a more general setting: Supersaturation, Continuity, and Approximation. Roughly speaking supersaturation implies that a large graph with an edge-density more than the Turán density for a particular forbidden family must contain many copies of members of that family. Continuity shows that given an infinite forbidden family, we can get arbitrarily close to its extremal number with a finite subfamily. Approximation is a structural result that shows that given a forbidden family, we can approximate an extremal sequence to an arbitrarily small difference by taking some sequence of graphs that all exclude this family and which all fall into some "nice\" form. All of these notions will be made rigorous in the paper. In and, Langlois, Mubayi, Sloan, and Gy. Turán studied extremal properties of certain small configurations in a directed hypergraph model. This model can be thought of as a \(2 \rightarrow 1\) directed hypergraph where each edge has three verticies, two of which are "tails\" and the third is a "head.\" They determined the extremal number for one such subgraph with two edges, and found the extremal number of a second configuration with two edges up to asymptotic equivalence. In and, this author followed up this work and found the exact extremal numbers for every \(2 \rightarrow 1\) directed hypergraph with exactly two edges. This \(2 \rightarrow 1\) model is one instance of the class of models discussed in this paper. The graph theoretic properties of a more general definition of a directed hypergraph were studied by Gallo, Longo, Pallottino, and Nguyen in. There a directed hyperedge was defined to be some subset of vertices with a partition into head vertices and tail vertices. This is a nonuniform version of models considered in this paper. The totally directed hypergraph model considered in and the \(r \rightarrow 1\) directed hypergraph model resulting from the study of Horn clauses both lead to the natural question of all possible ways to define a directed hypergraph. The definition in this paper of the class of general directed hypergraph models attempts to unify all of the possible "natural\" ways one could define a directed hypergraph so that certain extremal questions can be answered about all of them at once. Adding to the motivation of considering more general structures is the recent interest in Razborov's flag algebra method which applies to all relational theories and not just undirected hypergraphs. The fact that the \(d\)-simplex model studied by Leader as well as many other somewhat geometric models come out of the class defined in this paper was a very interesting accident. This paper is organized as follows. In Section 2, we define the class of generalized directed hypergraphs and extend the concepts of Turán density, blowups, and supersaturation to this setting. In Section 3, we define the idea of a jump for a given model of directed hypergraphs and prove several results about these jumps and how the jumps from one instance of the class relate to jumps in another. In Section 4, we adapt a couple of results proved in for totally directed hypergraphs with multiplicity to any GDH. In Section 5, we ask some questions that arose from studying these structures and discuss alternate definitions that would generalize the concept further. # Basic Definitions and Results The following definition for a generalized directed hypergraph is intended to include most uniform models that could reasonably be called uniform directed hypergraphs. This includes models where the edges are \(r\)-sets each under some partition into \(k\) parts of fixed sizes \(r_1,\ldots,r_k\) with some linear ordering on the \(k\) parts. The definition only includes structures where an \(r\)-set could include multiple edges up to the number of possible orientations allowed. That is, we do not consider the "oriented\" versions of the models where only one edge is allowed per \(r\)-set. The definition is given in terms of logic and model theory for convenience only. No deep results from those subfields are used. The use of this notation also makes further generalizations like nonuniform directed hypergraphs or oriented directed hypergraphs easy. Note that this definition includes graphs, hypergraphs, and \(r \rightarrow 1\) directed hypergraphs. For example, the theory for a \(2 \rightarrow 1\) directed hypergraph is \[T = \{\forall xyz E(x,y,z) \implies x \neq y \land x \neq z \land y \neq z \land E(y,x,z)\}.\] It is easy to see that when \(r=2\) we have only two GDH theories. The theory associated with the group \(S_2\) is the theory of graphs, and the theory associated with the trivial group is the theory of directed graphs. When \(r=3\) there are six subgroups of \(S_3\). Three of these are all isomorphic to \(\mathbb{Z}_2\) with each generated by a permutation that swaps two elements. The corresponding GDH theory for any of these can be thought of as having pointed \(3\)-sets for edges or as being (\(2 \rightarrow 1\))-graphs. Of the other subgroups, \(S_3\) itself gives the theory of undirected \(3\)-uniform hypergraphs, the trivial group gives totally directed \(3\)-edges, and the subgroup generated by a three-cycle isomorphic to \(\mathbb{Z}_3\) yields a GDH theory where the edges can be thought of as \(3\)-sets that have some kind of cyclic orientation-either clockwise or counter-clockwise. Figure [\[lattice\]](#lattice){reference-type="ref" reference="lattice"} summarizes the models of GDHs when \(r=3\). Note that in general, \(S_r\) always corresponds to the normal undirected \(r\)-graph model and the trivial group always corresponds to totally directed hypergraphs. A fun thought experiment is to consider the kinds of edges that arise when \(r=4\). Many of them are geometric in nature. For instance, the alternating group \(A_4\) gives a theory where edges can be thought of tetrahedrons (at least in an abstract sense). In fact, in Leader and Tan study the "oriented\" versions of the models that come from the alternating groups for any \(r \geq 3\). In this paper when the theory is not specified we are simply discussing GDHs that are all models of the same fixed theory. When discussing multiple theories we will often refer to \(T\)-graphs to mean models of a GDH theory \(T\). Throughout the paper, \(J_T\) will always stand for the subgroup \(J_T \subseteq S_r\) that determines the GDH theory \(T\) and \(m_T\) will always be the order of this subgroup, \(m_T = |J_T|\). Also, \(V_G\) and \(E_G\) will be used to denote the underlying set of elements of a model \(G\) and its relation set respectively. The following basic propositions are given without proof. The first is a simple consequence thatwe are working in a relational language, and the second results from the fact that \(J_T\) is a group. We can now use these propositions to extend the concepts of extremal graph theory to GDHs in a natural way. Note that since \[e_T(G) = \frac{|E_G|}{m_T},\] then the density is \[d_T(G) = \frac{(n-r)!|E_G|}{n!}\] and could have been defined this way while mostly avoiding talk of edges as equivalence classes of \(E_G\). However, the above definition makes the following extremal concepts reduce to their standard definitions in the undirected case. Our first main result is to show that these Turán densities exist for any GDH theory. The proof is the standard averaging argument used to show that these limiting densities exist for families of undirected hypergraphs. ## Blowups and Blowup Density We'll now extend the concept of the blowup of uniform hypergraphs to the more general setting of GDHs and define the corresponding notion of the blowup density. As with hypergraphs, the blowup of a GDH can be thought of as the replacement of each vertex with many copies and taking all of the resulting edges. Formally, Next, we consider the edge density of a given blowup by defining the edge polynomial for a GDH. This polynomial is a simple generalization of the standard edge polynomial for undirected hypergraphs. To see this more easily note that for a given GDH \(G\), the edges of \(G\) are in bijection with the monomials the sum \(p_G\) were we to write the sum out with no coefficients greater than one. From this we see that the edge density of the \((t_1,\ldots,t_n)\)-blowup of \(G\) is \[\frac{p_G(t_1,\ldots,t_n)}{\frac{r!}{m_T}{t \choose r}} = m_T\frac{p_G(t_1,\ldots,t_n)}{t(t-1)\cdots(t-r+1)}\] where \(t= \sum t_i\). Let \(t\) increase to infinity and for each \(t\) pick a vector \((t_1,\ldots,t_n)\) that maximizes this edge density. Then this sequence of densities is asymptotically equivalent to the sequence of numbers \[m_T p_G\left(\frac{t_1}{t},\ldots,\frac{t_n}{t}\right).\] This motivates the following definition. Since any \(x \in S^n\) is the limit of some sequence \(\left\{\left(\frac{t_1}{t},\ldots,\frac{t_n}{t}\right)\right\}\) with positive \(t_i\) as \(t \rightarrow \infty\), then the blowup density of a GDH \(G\) is the best limiting density of any sequence of blowups of \(G\). The remaining definition and basic result about blowups given in this subsection will be useful when extending results about jumps and nonjumps from undirected hypergraphs to GDHs generally in Section 3. ## Supersaturation and Related Results Supersaturation holds for GDHs as it does for undirected hypergraphs, and the proof of this result is the same as the one for hypergraphs found in with only minor differences. Similarly, the following theorem is an extension from the same result for undirected hypergraphs, and the proof is an adaptation of the one found in. The fact that the Turán density of a blowup equals the Turán density of the original GDH leads to the following nice characterization of degenerate families of GDH-those families with Turán density zero. # Jumps Now we turn to the issue of finding jumps and nonjumps for GDH theories. The definition of a jump for undirected hypergraphs extends naturally to this setting as does the important connection between jumps and blowup densities. Note that by Theorem [\[degenerate\]](#degenerate){reference-type="ref" reference="degenerate"} every \(\alpha \in \left[0,\frac{m_T}{r^r}\right)\) is a jump for any \(r\)-ary GDH theory \(T\). This generalizes the well-known result of Erdős that every \(\alpha \in [0,\frac{r!}{r^r})\) is a jump for \(r\)-graphs. The following important theorem on jumps for GDH theories was originally shown by Frankl and Rödl for undirected hypergraphs. Their proof works equally well in this setting so the differences here are in name only. The following proposition is needed to compare jumps between different GDH theories. We will now look at how jumps are related between two different GDH theories for some fixed edge size \(r\). We will see that in general jumps always "pass up\" the subgroup lattice. That is, if \(J_{T'} \subseteq J_T\) for GDH theories \(T'\) and \(T\), then a jump for \(T'\) is a jump for \(T\). The converse is not true in general. In fact, for any GDH theories \(T'\) and \(T\) with \(J_{T'} \subseteq J_T\) such that the order of \(J_T\) is at least three times that of \(J_{T'}\) we will show that the set of jumps for \(T'\) is not equal to the set of jumps for \(T\). The case where \(m_T = 2m_{T'}\) is open. ## Jumps pass up the lattice First, we will show that for GDH theories \(T\) and \(T'\) with \(J_{T'} \subseteq J_T\) the set of Turán densities of forbidden families of \(T\)-graphs is a subset of the set of Turán densities for \(T'\). The converse of Theorem [\[thmC\]](#thmC){reference-type="ref" reference="thmC"} is false in general. For example, the permutation subgroup for the theory \(T'\) of (\(2 \rightarrow 1\))-uniform directed hypergraphs is a subgroup of the permutation group for the theory \(T\) of undirected \(3\)-graphs, \(S_3\). The extremal number for the directed hypergraph is \(F = \{ab \rightarrow c, cd \rightarrow e\}\) (see Figure [\[F\]](#F){reference-type="ref" reference="F"}) is \[ex_{T'}(n,F) = \left\lfloor \frac{n}{3} \right\rfloor {\left\lceil \frac{2n}{3} \right\rceil \choose 2}\] as shown in. Therefore, the Turán density is \(\pi_{T'}(F) = \frac{4}{27}\). However, it is well-known that no Turán densities exist for \(3\)-graphs in the interval \(\left(0,\frac{6}{27}\right)\). Corollary [\[jumpsgoup\]](#jumpsgoup){reference-type="ref" reference="jumpsgoup"} immediately implies that all nonjumps found for \(r\)-uniform undirected hypergraphs must also be non-jumps for any GDH with an \(r\)-ary relation. However, the converse is not true in general. ## Jumps do not pass down the lattice Roughly speaking, the current best method of demonstrating that a particular \(\alpha\) is not a jump for \(r\)-uniform hypergraphs is to construct a sequence of hypergraphs each with blowup densities that are strictly larger than \(\alpha\) but for which any relatively small subgraph has blowup density at most \(\alpha\). This method originated in and generalizes to GDHs as the following definition and lemma demonstrate. We can now show that a demonstrated nonjump for a GDH theory \(T\) yields multiple nonjumps of equal and lesser values down the lattice to GDH theories \(T'\) for which \(J_{T'} \subseteq J_T\). Constructions of sequences of undirected \(r\)-graphs which show that \(\frac{5r!}{2r^r}\) is a demonstrated nonjump for each \(r \geq 3\) were given in. This gives the following corollary. This in turn shows that the set of jumps for a theory \(T'\) is a proper subset of the set of jumps for \(T\) for any \(T\) such that \(J_{T'} \subseteq J_T\) and \(m_T \geq 3m_{T'}\). # Continuity and Approximation The following two results are direct adaptations of two theorems from. They are both general extremal results related to everything discussed in this paper but did not fit nicely into the other sections. The first result, Continuity, relates extremal numbers of any infinite family of GDHs to the extremal numbers of its finite subfamilies. The second, Approximation, discusses structural aspects of (nearly) extremal sequences for any forbidden family. Theorem 6 in is the Approximation Theorem for totally directed \(r\)-uniform hypergraphs with bounded multiplicity. We will use the following equivalent statement (in the case of multiplicity one) written in terms of Turán densities as a lemma to prove that this approximation result holds for all GDHs. # Conclusion Some questions naturally come up in studying GDHs. Most notably it would be nice to show that the set of jumps for some GDH theory \(T'\) is a proper subset of the set of jumps of any theory \(T'\) up the lattice including those for which \(m_T = 2m_{T'}\). Or on the other hand it would be very interesting to learn that this is not true in certain cases for \(r \geq 3\)! It is known by a result in that every \(\alpha \in [0,1)\) is a jump for digraphs. Therefore, the conjecture is not true when \(r=2\). On a related note, is it always true that when \(J_{T'} \subset J_T\), there always exists a family \(\mathcal{F}'\) of \(T'\)-graphs such that \(\pi_{T'}(\mathcal{F}')\) is not contained in the set of Turán densities for \(T\)? Finally, it would be nice to generalize the definition of a GDH to include other combinatorial structures. For instance we could easily change the current formulation to include multiple relations in order to capture nonuniform GDHs and those with edges that have bounded multiplicity like the structures studied in. We could even allow these theories to contain general statements that relate the different relations. An example of this might be the theory of some kind of GDH with an edge-coloring that behaves in a certain way (at least locally). In another direction we could take away the requirement that all vertices of an edge be distinct to allow for kinds of generalized loops or add a condition that the existence of certain edges preclude the existence of others such as in the oriented cases studied in, , and.

{'timestamp': '2016-07-19T02:09:36', 'yymm': '1607', 'arxiv_id': '1607.04927', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04927'}

# Motivation Understanding of the analytic structure of the contribution of a graph to a Feynman amplitude, a time-honored problem, is related to an analysis of its reduced graphs and the graphs in which internal edges are on the mass-shell. The former case relates to graphs in which internal edges shrink. The latter case relates to graphs with cut edges. The set of cut edges is uniquely determined by the choice of a spanning forest for the graph: such a spanning forest defines a unique set of edges connecting distinct components of the forest. It is those edges we will put on the mass-shell. Pairs of graphs and a chosen ordered spanning tree or forest deliver the cubical chain complex. A given ordering of the edges of the spanning tree \(T\) defines a sequence of spanning forests \(F\), and to any pair \((\Gamma,F)\) for fixed \(\Gamma\) we can associate:\ -a reduced graph \(\Gamma_F\) obtained by shrinking all edges of \(\Gamma\) to length zero which do not connect different components of the spanning forest\ -a cut graph \(\Gamma^F\) where all those edges connecting different components are put on-shell, so are marked by a Cutkosky cut,\ -the set of graphs \(G^F=\Gamma-E_{\Gamma_F}\) obtained from \(\Gamma\) by removing the edges which connect distinct components of the spanning forest. Such data define a cell-complex. With it they define a set of lower triangular matrices, one for each ordering of the edges in \(T\), which allow to analyse a graph amplitude from its reduced graphs and the variations obtained by putting internal edges on-shell. ## Results A sequence of cuts (edge sets \(\epsilon_i\) determines from \(i\)-component forests, \(i\geq 2\)) \[\epsilon_2\to \epsilon_3\to\cdots\to \epsilon_{v_\Gamma}\] will shift the normal threshold \(s_0(\epsilon_2)\) associated with a chosen cut \(\epsilon_2\) to anomalous thresholds \[s_0(\epsilon_2)\to s_1(\epsilon_3)\to\cdots\to s_{v_\Gamma-2}(\epsilon_{v_\Gamma}).\] The resulting sequence of anomalous thresholds \(s_i(\epsilon_{i+2})\), \(i>0\) is a sequence of values for a channel variable \(s\) defined by \(\epsilon_2\). They are computed from the divisors associated to \(\epsilon_{i+2}\). The latter are functions of all kinematical variables. For example, for the one-loop triangle the divisor in \(\mathbb{C}^3\) associated to \(\epsilon_3\) is a simple function of \[\lambda(p_1^2,p_2^2,p_3^2)=p_1.p_2^2-p_1^2p_2^2=p_2.p_3^2-p_2^2p_3^2=p_3.p_1^2-p_3^2p_1^2,\,p_1+p_2+p_3=0.\] The three representations of \(\lambda(p_1^2,p_2^2,p_3^2)\) allow to compute \(s_1(\epsilon_3)\) for the three choices of a channel variable \(s=p_3^2\) or \(s=p_1^2\) or \(s=p_2^2\) respectively. As a result, to a graph \(\Gamma\) we can assign a collection of lower triangular matrices \(M_i^\Gamma\) with the following properties:\ i) All entries in the matrix correspond to well-defined integrable forms under on-shell renormalization conditions.\ ii) Anomalous thresholds \(s_i\) are determined from properties of graph polynomials. They provide lower boundaries for dispersion integrals associated to these integrable forms.\ iii) Along the diagonal in the matrices \(M_i^\Gamma\) we find leading threshold entries: all quadrics for all edges in a graph are on the mass-shell.\ iv) The variation of a column in \(M_i^\Gamma\) wrt to a given channel is given by the column to the right.\ v) The subdiagonal entries \((M_i^\Gamma)_{k,k-1}\) are determined from the diagonal entries \((M_i^\Gamma)_{k-1,k-1}\) and \((M_i^\Gamma)_{k,k}\) via a dispersion integral. This gives \((k-1)\) two-by-two matrices each of which has an interpretation via the optical theorem. This hence determines the first subdiagonal.\ vi) Continuing, all subdiagonals and hence the whole matrix \((M_i^\Gamma)_{r,s}\) is determined via iterated dispersion. This answers the question how to continue the optical theorem beyond two-point functions. # The cubical chain complex We follow . Consider a pair \((\Gamma,T)\) of a bridge free graph \(\Gamma\) and a chosen spanning tree \(T\) for it. Assume \(T\) has \(k\) edges. Consider the \(k\)-dimensional unit cube. It has origin \((0,\cdots,0)\) and \(k\) unit vectors \((1,0,\cdots,0),\ldots\), \((0,\cdots,0,1)\) form its edges regarded as \(1\)-cells. A change of ordering of the edges of \(T\) permutes those edges. The origin is decorated by a rose on \(|\Gamma|\) petals, and the corner \((1,1,\cdots,1)\) decorated by \((\Gamma,V_\Gamma)\), with \(k=|V_\Gamma|-1\), and we regard \(V_\Gamma\) as a spanning forest consisting of \(k+1\) distinct vertices. The complex is best explained by assigning graphs as in the following example. \[{\;\raisebox{10mm}{\epsfysize=80mm\epsfbox{cubical.eps}}\;}\] The cell is two-dimensional as each of the five spanning trees of the graph \(\Gamma\), the dunce's cap graph, in the middle of the cell has length two. We have chosen a spanning tree \(T\) provided by the edges \(e_1\) and \(e_3\), indicated in red. The boundary of our two-dimensional cell has four one-dimensional edges, bounded by two of the four 0-dimensional corners each. To these lower dimensional cells we assign graphs as well as indicated. The spanning tree has length two and so there are \(2=2!\) orderings of its edges, and hence two lower triangular \(3\times 3\) matrices \(M^\Gamma_i\) which we can assign to this cell. They look as follows: \[{\;\raisebox{10mm}{\epsfxsize=140mm\epsfbox{cubicalM.eps}}\;}\] These square matrices are lower triangular. Note that we have cuts which separate the graph into two components determining a normal threshold which appears already in a reduced graph on the diagonal, and in the lower right corner a cut into three components, which determines an anomalous threshold. All these cuts determine variations, as stated in Cutkosky's theorem. # Cutkosky's theorem We quote from where you find details. For a graph \(\Gamma\) and a choosen spanning forest \(F\) we let the quotient graph \(\Gamma''\)-the reduced graph-be the graph obtained by shrinking all edges \(e\in E'\) of \(\Gamma\) which do not connect distinct components of \(F\), so \(E'=E_\Gamma-E''\), and \(E''\) all edges of \(\Gamma\) which do connect distinct components of \(F\). Assume the reducedgraph \(\Gamma''\) has a physical singularity at an external momentum point \(p''\), i.e. the intersection \(\bigcap_{e\in E''}Q_e\) of the propagator quadrics associated to edges in \(E''\) has such a singularity at a point lying over \(p''\). Let \(p\) be an external momentum point for \(\Gamma\) lying over \(p''\). Then the variation of the amplitude \(I(\Gamma)\) around \(p\) is given by Cutkosky's formula \[\text{var}(I(\Gamma)) = (-2\pi i)^{\# E''}\int\frac{\prod_{e\in E''} \delta^+(\ell_e)}{\prod_{e\in E'} \ell_e}.\] # Anomalous thresholds Let us come back to a generic graph \(\Gamma\). We want to determine anomalous thresholds. With their help, dispersion relations can be established when real analycity in kinematical variables can be established. We analyse the Landau singularities of \(\Gamma\) in terms of \(\Gamma/e\), where \(e\) is such an edge. To completely analyse the graph, we have to consider all possibilities to shrink it edge after edge (the generalization to multiple edges is in ). We have for the second Symanzik polynomial \(\Phi\) \[\Phi(\Gamma)=\overbrace{\Phi(\Gamma/e)}^{=:X}+A_e\overbrace{\left\{ \Phi(\Gamma-e)-m_e^2\psi(\Gamma/e) \right\} }^{=:Y}-A_e^2\overbrace{m_e^2\psi(\Gamma-e)}^{Z}.\] Solving \(\Phi(\Gamma)=0\) for \(A_e\) is a quadratic equation with coefficients \(X,Y,Z\). Note that \(Z>0\) is independent of kinematical variables, while \(X,Y\) depend on momenta and masses. In particular, for a chosen channel variable \(s\) we can write \(X=sX_s+N\), with \(X_s\) independent of kinematics and \(N\) a constant in the channel variabel \(s\). It depends on other kinematic variables though. In terms of parametric variables, \(X,Y\) are functions of the parametric variables of the reduced graph. The above quadratic equation has a discriminant \(D=Y^2+4XZ\), and we find a physical Landau singularity for positive \(Y\) and vanishing discriminant \(D=0\). Define \(Y_0:=Y(p_A^{\Gamma/e},\{Q,M\})\) to be the evaluation of \(Y\) by evaluating parametric variables at the point of the Landau singularity for the reduced graph. The condition \(D=0\) allows to determine the anomalous threshold from \[s(\{A\},\{Q,M\})=\frac{Y^2-4ZN}{4ZX_s},\] minimizing over parametric variables \(A_e\geq 0\). Let \(\mathit{T}_s^\Gamma\) be the set of all ordered spanning trees \(T\) of a fixed graph \(\Gamma\) which allow for the same associated channel variable \(s\). We have the following result.\ i) A necessary and sufficient condition for a physical Landau singularity is \(Y_0>0\) with \(D=0\).\ ii) The corresponding anomalous threshold \(s_F\) for fixed masses and momenta \(\{M,Q\}\) is given as the minimum of \(s(\{a,b\},\{Q,M\})\) varied over edge variables \(\{a,b\}\). It is finite (\(s_F>-\infty\)) if the minimum is a point inside \(p\in\mathbb{P}^{e_\Gamma-1}\) in the interior of the integration domain \(A_i>0\). If it is on the boundary of that simplex, \(s_F=-\infty\).\ iii) If for all \(T\in \mathit{T}_s^\Gamma\) and for all their forests \((\Gamma,F)\) we have \(s_F>-\infty\), the Feynman integral \(\Phi_R(\Gamma)(s)\) is real analytic as a function of \(s\) for \(s<\min_F\{s_F\}\).\ iv) For \(Y>0\) and \(X<0\), both zeroes of \(\Phi(\Gamma)=0\) appear for \(A_e>0\). For \(Y>0\) and \(X>0\), only one zero is inside the domain of integration. As a result for \(X=0\) corresponding to the threshold provided by the reduced \(\Gamma/e\) we have a discontinuity. # Example We consider the triangle graph \(\Delta\). In fact, we augment it with one of its three possible spanning trees, say on edges \(e_2,e_3\), so \(E_T=\{e_2,e_3\}\). The corresponding cell in the cubical chain complex is \[\label{trianglecell} {\;\raisebox{-40mm}{\epsfxsize=80mm\epsfbox{trianglecubical.eps}}\;}\] For the Cutkosky cut we choose two of the three edges, say \(\epsilon_2=\{e_1,e_2\}\). This defines the channel \(s=p_a^2\) and the matrix \(M^\Delta_1\). The other cut in that matrix is the full cut separating all three vertices. We give \(M^\Delta_1\) in the following figure: \[M^\Delta_1={\;\raisebox{-40mm}{\epsfxsize=80mm\epsfbox{trianglematrix.eps}}\;}\] We now calculate: \[\begin{aligned} \Phi_\Delta & = & \overbrace{p_a^2A_1A_2-(m_1^2A_1+m_2^2A_1)(A_1+A_2)}^{=\Phi_{\Gamma/e_3}}\\ & & +A_3((p_b^2-m_3^2-m_1^2)A_1+(p_c^2-m_1^2-m_3^2)A_2)-A_3^2m_3^2, \end{aligned}\] so \[\begin{aligned} \Phi_\Delta & = & \Phi_{\Delta/e_3} +A_3Y-A_3^2 m_3^2\overbrace{\psi_{\Delta-e_1}}^{=1}, \end{aligned}\] as announced: \[X=\Phi_{\Delta/e_3},\,Y=\overbrace{(p_b^2-m_3^2-m_1^2)}^{=:l_1}A_1+\overbrace{(p_c^2-m_1^2-m_3^2)}^{=:l_2}A_2,\,Z=m_3^2.\] We have \(Y_0=m_2 l_1+m_1 l_2\), and need \(Y_0>0\) for a Landau singularity. Solving \(\Phi(\Delta/e_3)=0\) for a Landau singularity determines the familiar physical threshold in the \(s=p_a^2\) channel, leading for the reduced graph to \[p_Q: s_0=(m_2+m_3)^2,\, p_A: A_1m_1=A_2m_2.\] We let \(D=Y^2+4XZ\) be the discriminant. For a Landau singularity we need \[D=0.\] We have \[\Phi_\Delta=-m_3^2\left(A_3-\frac{Y+\sqrt{D}}{2m_3^2}\right)\left(A_3-\frac{Y-\sqrt{D}}{2m_3^2}\right),\] where \(Y,D\) are functions of \(A_1,A_2\) and \(m_1^2,m_2^2,m_3^2,s,p_b^2,p_c^2\). Note that at \(D=0\) we have \[2m_3^2A_3=A_1l_1+A_2l_2,\] which determines a co-dimension one (a line) hypersurface of \(\mathbb{P}^2\). Finding the anomalous thresold determines a point on this line (it fixes the ratio \(A_1/A_2\)), and hence the anomalous threshold determines a point in \(\mathbb{P}^2\). We can write \[0=D=Y^2+4Z(sA_1A_2-N),\] with \(N=(A_1m_1^2+A_2m_2^2)(A_1+A_2)\) \(s\)-independent. This gives \[s(A_1,A_2)=\frac{4ZN-(A_1l_1+A_2l_2)^2}{4ZA_1A_2}=:\frac{A_1}{A_2}\rho_1+\rho_0+\frac{A_2}{A_1}\rho_2.\] Define two Kallen functions \(\rho_1=-\lambda_1=-\lambda(p_b^2,m_1^2,m_3^2)\) and \(\rho_2=-\lambda_2=-\lambda(p_c^2,m_2^2,m_3^2)\). Both are real and non-zero off their threshold or pseudo-threshold. Then, for \[\rho_1>0,\,\rho_2>0,\] we find the threshold \(s_1\) at \[s_1=(m_1+m_2)^2+\frac{4m_3^2(\sqrt{\lambda_2}m_1-\sqrt{\lambda_1}m_2)^2-(\sqrt{\lambda_1}l_2+\sqrt{\lambda_2}l_1)^2}{4m_3^2\sqrt{\lambda_1}\sqrt{\lambda_2}}.\] On the other hand for the coefficients of \(\rho_1<0\) and/or \(\rho_2<0\) we find a minimum \[s_1=-\infty,\] along the boundaries \(A_1=0\) or \(A_2=0\). The domains \(Y>0,X<0\) and \(Y>0,X>0\) determine the domains of parametric integration for the variation prescribed by Cutkosky's theorem, whilst the normal and anomalous threshold (when finite) determine the lower boundaries of the dispersion integrals needed to reconstruct the function from its variation. Let us now discuss the triangle in more detail. It allows three spanning trees on two edges each, so we get six matrices \(M^\Delta_i\), \(i=1,\ldots,6\) altogether, by having two possibilities to order the two edges for each spanning tree. The six matrices \(M^\Delta_i\) come in groups of two for each spanning tree. For each of the three spanning trees we get a cell as in ([\[trianglecell\]](#trianglecell){reference-type="ref" reference="trianglecell"}). The boundary operator for such a cell in the cubical cell complex of is the obvious one stemming from co-dimension one hypersurfaces at \(0\) or \(1\) with suitable signs. So the square populated by the triangle \(\Delta\) in ([\[trianglecell\]](#trianglecell){reference-type="ref" reference="trianglecell"}) has four boundary components, the edges populated by the four graphs as indicated. Those four edges are the obvious boundary of the square. If we now consider all graphs in ([\[trianglecell\]](#trianglecell){reference-type="ref" reference="trianglecell"}) as evaluated by the Feynman rules, we can consider for a given cell a boundary operator which replaces evaluation at the \(x_e=0\)-hypersurface by shrinking edge \(e\), and evaluation at the \(x_e=1\)-hypersurface by setting edge \(e\) on the mass-shell. Then, to check that this is a boundary operator for the amplitudes defined by the graphs in ([\[trianglecell\]](#trianglecell){reference-type="ref" reference="trianglecell"}) we need to check that the amplitudes for the four graphs at the four corners are uniquely defined from the amplitudes of the graphs at the adjacent edges: for example, the imaginary part of the amplitude of the graph on the left vertical edge is related to the amplitude of the graph at the upper left corner: This imaginary part must be also obtained from shinking edge \(e_3\) in the graph on the upper horizontal edge by setting \(A_3\) to zero in the integrand and integrating over the hypersurface \(A_3=0\) of the integration simplex \(\sigma_\Delta\). This is indeed the case, and similar checks work for all other corners. In summary, the analytic structure of Feynman amplitudes realized the structure of the cubical chain complex. The latter is highly non-trivial. Its further study in the conext needed for physics will inform our understanding of amplitudes considerably. Future work will be dedicated in understanding the relation between the monodromy of physical singularities and the fundamental group underlying Outer Space as used in.

{'timestamp': '2016-07-19T02:07:15', 'yymm': '1607', 'arxiv_id': '1607.04861', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04861'}

# Introduction Let \(\lambda_1,\ldots,\lambda_n\) be \(n\) distinct values. The Lagrange interpolation formula says that a polynomial \(P(x)\) of degree not greater than \(n-1\) in one variable can be written in the following: \[P(x) = \sum_{i=1}^nP(\lambda_i)L_i(x),\] where \[L_i(x) = \prod_{j\neq i}\frac{x-\lambda_j}{\lambda_i-\lambda_j}.\] This formula not only provides a means of polynomial approximation, but also plays a significant role in establishing combinatorial identities in many variables including identities on symmetric polynomials. For instance, given \(n\) distinct values \(\lambda_1,\lambda_2,\ldots,\lambda_n\) \((n \geq 2)\) and an integer \(m \geq 0\), we have the following identity: \[\label{id1} \sum_{i=1}^n\frac{\lambda_i^m}{\displaystyle\prod_{j\neq i}(\lambda_i-\lambda_j)}=h_{m-n+1}(\lambda_1,\ldots,\lambda_n),\] where \(h_k\) is the \(k\)-th homogeneous symmetric polynomial in \(n\) variables, which is defined to be zero for \(k<0\). The identity can be easily derived from the Lagrange interpolation formula (see for more details). In particular, for \(m = 0, 1, \ldots,n-1\), the left hand side of ([\[id1\]](#id1){reference-type="ref" reference="id1"}) is independent to \(\lambda_1,\ldots,\lambda_n\). It is equal to zero if \(m < n-1\) and one if \(m=n-1\). This implies the following result. The proof of Proposition [\[pro1\]](#pro1){reference-type="ref" reference="pro1"} is elementary. It is a direct corollary of the Lagrange interpolation formula. The first goal of this paper is to give a generalization of Proposition [\[pro1\]](#pro1){reference-type="ref" reference="pro1"} for multivariate symmetric polynomials. Throughout we always assume that all polynomials are over the field of rational numbers or any field of characteristic zero. For convenience, we shall write \([n]\) for the set \(\{1,2,\ldots,n\}\). Let \(\lambda_1,\ldots,\lambda_n\) be \(n\) distinct values. For each subset \(I = \{i_1,\ldots,i_k\} \subset [n]\), we denote by \(\lambda_I = (\lambda_{i_1},\ldots,\lambda_{i_k})\) and \(I^c = [n]\setminus I\). Recall that a polynomial \(P(x_1,\ldots,x_k)\) is said to be symmetric if it is invariant under permutations of \(x_1,\ldots,x_k\). We obtain the following result. More generally, we also obtain an identity involving doubly symmetric polynomials. Recall that a polynomial \(P(x_1,\ldots,x_k,y_1,\ldots,y_{n-k})\) is said to be doubly symmetric if it is invariant under permutations of \(x_1,\ldots,x_k\) and permutations of \(y_1,\ldots,y_{n-k}\) respectively. We obtain the following result. The second goal of this paper is to give a way of dealing with integrals over Grassmannians. The idea is as follows. Localization in equivariant cohomology allows us to express integrals in terms of some data attached to the fixed points of a torus action. In particular, for Grassmannians, we obtain interesting formulas with nontrivial relations involving rational functions. Let \(G(k,n)\) be the Grassmannian of \(k\)-dimensional linear spaces in \(\mathbb C^n\). Consider the following integrals: \[\int_{G(k,n)}\Phi(\mathcal S)\quad ,\quad \int_{G(k,n)}\Psi(\mathcal Q)\quad,\quad \int_{G(k,n)}\Delta(\mathcal S,\mathcal Q),\] where \(\Phi(\mathcal S), \Psi(\mathcal Q)\) are respectively characteristic classes of the tautological sub-bundle \(\mathcal S\) and quotient bundle \(\mathcal Q\) on the Grassmannian \(G(k,n)\), and \(\Delta(\mathcal S,\mathcal Q)\) is a characteristic class of both \(\mathcal S\) and \(\mathcal Q\). Using localization in equivariant cohomology, Weber and Zielenkiewicz presented a way of expressing the integrals as iterated residues at infinity of holomorphic functions. However, there is no guarantee expressions appearing in the Atiyah-Bott-Berline-Vergne formula for Grassmannians are in fact constants rather than rational functions. Our identities provide a way of handling such expressions. As a special case, if \(\Delta(\mathcal S,\mathcal Q) = \Phi(\mathcal S)\Psi(\mathcal Q)\) is represented by the product of two symmetric polynomials \(P(x_1,\ldots,x_k)\) and \(Q(y_1,\ldots,y_{n-k})\), then we have the following corollary. The rest of the paper is organized as follows: The identities are proved in Section 2. Section 3 is to give a review of equivariant cohomology and the proof of Theorem [\[integral1\]](#integral1){reference-type="ref" reference="integral1"}. # Proof of the identities We shall use \(X\) to denote the set \(\{x_1,\ldots,x_k\}\) or the \(n\)-tuple \((x_1,\ldots,x_k)\), and \(Y\) to denote the set \(\{y_1,\ldots,y_{n-k}\}\) or the \((n-k)\)-tuple \((y_1,\ldots,y_{n-k})\). We denote by \[Y-X = \prod_{i=1}^{n-k}\prod_{j=1}^k(y_i-x_j).\] For \(I=\{i_1,\ldots,i_k\}\) and \(J=\{j_1,\ldots,j_{n-k}\}\), we denote by \(\lambda_I = (\lambda_{i_1},\ldots,\lambda_{i_k})\), \(\lambda_J = (\lambda_{j_1},\ldots,\lambda_{j_{n-k}})\), and \[\lambda_I-\lambda_J = \prod_{i\in I}\prod_{j\in J}(\lambda_i-\lambda_j).\] # Localization in equivariant cohomology In this section, we recall some basic definitions and results in the theory of equivariant cohomology. For more details on this theory, we refer to. Throughout we consider all cohomologies with coefficients in the complex field \(\mathbb C\). Let \(T = (\mathbb C^*)^n\) be an algebraic torus of dimension \(n\), classified by the principal \(T\)-bundle \(ET \to BT\), whose total space \(ET\) is contractible. Let \(X\) be a compact space endowed with a \(T\)-action. Put \(X_T = X\times_T ET\), which is itself a bundle over \(BT\) with fiber \(X\). Recall that the \(T\)-equivariant cohomology of \(X\) is defined to be \(H^*_{T}(X) = H^*(X_T)\), where \(H^*(X_T)\) is the ordinary cohomology of \(X_T\). Note that \(H^*_T(\pt) = H^*(BT)\). By pullback via the map \(X\to \pt\), we see that \(H^*_T(X)\) is an \(H^*(BT)\)-module. Thus we may consider \(H^*(BT)\) as the coefficient ring for equivariant cohomology. A \(T\)-equivariant vector bundle is a vector bundle \(E\) on \(X\) together with a lifting of the action on \(X\) to an action on \(E\) which is linear on fibers. Note that \(E_T\) is a vector bundle over \(X_T\). The \(T\)-equivariant Chern classes \(c_i^T(E) \in H^*_T(X)\) are defined to be the Chern classes \(c_i(E_T)\). If \(E\) has rank \(r\), then the top Chern class \(c_r^T(E)\) is called the \(T\)-equivariant Euler class of \(E\) and is denoted \(e^T(E) \in H^*_T(X)\). More generally, the \(T\)-equivariant characteristic class \(c^T(E) \in H^*_T(X)\) is defined to be the characteristic class \(c(E_T)\). Let \(\chi(T)\) be the character group of the torus \(T\). For each \(\rho \in \chi(T)\), let \(\mathbb C_{\rho}\) denote the one-dimensional representation of \(T\) determined by \(\rho\). Then \(L_{\rho} = (\mathbb C_{\rho})_T\) is a line bundle over \(BT\), and the assignment \(\rho \mapsto-c_1(L_{\rho})\) defines an group isomorphism \(f: \chi(T) \simeq H^2(BT)\), which induces a ring isomorphism \(\Sym(\chi(T)) \simeq H^*(BT)\). We call \(f(\rho)\) the weight of \(\rho\). In particular, we denote by \(\lambda_i\) the weight of \(\rho_i\) defined by \(\rho_i(x_1,\ldots,x_n) = x_i\). We thus obtain an isomorphism \[H^*_T(\pt) = H^*(BT) \simeq \mathbb C[\lambda_1,\ldots,\lambda_n].\] Let \(\mathcal R_T \simeq \mathbb C(\lambda_1,\ldots,\lambda_n)\) be the field of fractions of \(\mathbb C[\lambda_1,\ldots,\lambda_n]\). An important result in equivariant cohomology is the localization theorem. Historically, localization in equivariant cohomology was studied by Borel and then further investigated by Quillen, Atiyah-Bott, and Berline-Vergne. Among many versions of the formulation of the localization theorem, we choose the one by Atiyah and Bott. Moreover, Atiyah and Bott also gave an explicit formula for the inverse isomorphism. If \(X\) is a compact manifold and \(X^T\) is finite, then the localization theorem can be rephrased as follows: For many applications, the Atiyah-Bott-Berline-Vergne formula can be formulated in more down-to-earth terms. We are mainly interested in the computation of integrals over Grassmannians.

{'timestamp': '2016-09-12T02:04:24', 'yymm': '1607', 'arxiv_id': '1607.04850', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04850'}

# Introduction The Aubry-André-Harper (AAH) model has been extensively used as a quasiperiodic model to theoretically study the phase transition between extended, critical and localized phases. With the realization of the AAH model in photonic crystals and ultracold atoms, this model has gained attention in recent years. The abundant phenomena revealed by the AAH model make it an ideal test field for topological phases and the transitions between them both in the incommensurate and commensurate cases. When the lattice is incommensurate, the system will go through a transition from an extended phase to a localized phase due to the disordered on-site potential. If the lattice is commensurate, the AAH model can be used to explore emerging topological states of matter. In addition, quantum many-body localization phenomena are also widely explored in this model. Recently, the connection between the AAH model, the one-dimensional Kitaev chain, and topological superconductivity are investigated. In Ref., a generalized AAH model which has modulations both in the on-site potentials and the hopping amplitudes between the neighboring lattice sites is discussed, and the topological states are attributed to the topological properties of the one-dimensional Majorana chain. This model is also connected to the Su-Schrieffer-Heeger (SSH) model when the lattice is commensurate. In Ref., the phase diagram of a non-Abelian AAH model with p-wave superfluidity is discussed and the phase transition from a metallic phase to an insulator phase is checked. However, most AAH models explored before have only modulations in the on-site potential or do not include the superconducting pairing, thus limiting the usage of these models. If we introduce superconducting pairing into the generalized AAH models discussed in Ref., more interesting phenomena can be expected. With p-wave superconducting pairing and modulations in the model, there would be both SSH-like and Kitaev-like topologically nontrivial phases, and thus so-called electron fractionalization to the Majorana fermions can be realized, as discussed in Ref.. In this paper, we consider a generalized AAH model with a superconducting pairing potential. Both the hopping amplitudes between the nearest neighboring lattice sites and the on-site potentials are modulated by a cosine function which has a periodicity of \(1/\alpha\). When \(\alpha\) is irrational, the lattice is incommensurate, and we find that the system will go through a phase transition from the extended states to localized states when the disordered on-site potential increases beyond some critical value. This critical value as well as the critical region of the system will be significantly reduced due to the modulation in the hopping amplitude and thus makes the system much easier to be localized. On the other hand, in the commensurate lattice when \(\alpha\) is rational, this model shows different topological phases as we tune the system parameters. We mainly discuss the case with \(\alpha=1/2\) and find that there are three phases presented in the system: a Su-Schrieffer-Heeger (SSH)-like trivial phase, a SSH-like topological phase, and a Kitaev-like topological phase. By calculating the topological numbers of the system, we present the phase diagrams and phase boundaries for these phases. This generalized AAH model provides us with a useful test field for the study of phase transitions between extended states and Anderson localized states, or the transitions between different topological phases. The rest of the paper is organized as follows. In Sec. [2](#sec2){reference-type="ref" reference="sec2"}, we introduce the generalized AAH model and write down the Hamiltonian. Then we consider the incommensurate case in Sec. [3](#sec3){reference-type="ref" reference="sec3"}, where we will discuss the phase transition from the extended phase to the localized phase, and especially the influences of the modulations in the hopping amplitude on this transition. In Sec. [4](#sec4){reference-type="ref" reference="sec4"}, we check the commensurate lattice with \(\alpha=1/2\) and investigate those different phases presented in this system. The last section (Sec. [5](#sec5){reference-type="ref" reference="sec5"}) is dedicated to a brief summary. # Model Hamiltonian {#sec2} The Hamiltonian of the generalized one-dimensional (1D) Aubry-André-Harper model we consider in this paper is described by the following Hamiltonian \[\label{Eq1} H= \sum_{j=1}^{N} V_j c_j^\dagger c_j + \sum_{j=1}^{N-1} [-t_j c_{j+1}^\dagger c_j +\Delta c_{j+1}^\dagger c_j^\dagger + H.c. ]\] where \(c_j^\dagger\) (\(c_j\)) is the creation (annihilation) operator at site \(j\), \(V_{j} = V \cos(2\pi \alpha j + \varphi_V)\) is the on-site potential, \(t_j = t [1+\lambda \cos (2\pi \alpha j + \varphi _\lambda)]\) is the hopping amplitude between the nearest neighboring lattice sites, and \(\Delta\) is the superconducting pairing gap which is taken to be real. This one-dimensional chain has \(N\) sites and both the on-site potential and the hopping amplitude are modulated by a cosine function with periodicity \(1/\alpha\) but with phase factor \(\varphi_v\) and \(\varphi_\lambda\), respectively. If \(\Delta=0\), this model will reduce to the generalized AAH model introduced in Ref., which covers both the diagonal and off-diagonal AAH model. If \(\lambda=0\), then the Hamiltonian is the same as the 1D Kitaev chain with on-site potential modulations which are investigated in detail in Refs.. It has been predicted that if \(\alpha\) is irrational, the Kitaev model \((\lambda=0, V \neq 0)\) or the diagonal AAH model \((\lambda=0, \Delta=0, V \neq 0)\) will go through a localization transition where all extended states become localized as \(V\) is increased beyond some critical value. However, if \(\alpha\) is rational, it is known that there is no such phase transition in these two models. For simplicity, we will set \(\varphi_V = \varphi_\lambda = \varphi\) in this paper but we emphasize that all the results obtained in this paper can also be extended to the \(\varphi_V \neq \varphi_\lambda\) case. By introducing Majorana operators \(c_j = (\gamma_{j,1} + i \gamma_{j,2})/2\), the Hamiltonian could be rewritten as \[\label{Eq2} \begin{split} H &= \frac{i}{2} \sum_{j} \{ V_{2j-1} \gamma_{2j-1,1} \gamma_{2j-1,2} + V_{2j} \gamma_{2j,1} \gamma_{2j,2} \\ &+ (-t_{2j-1}-\Delta ) \gamma_{2j,1} \gamma_{2j-1,2} + (-t_{2j} + \Delta ) \gamma_{2j,1} \gamma_{2j+1,2} \\ &+ ( t_{2j-1}-\Delta ) \gamma_{2j,2} \gamma_{2j-1,1} + ( t_{2j} + \Delta ) \gamma_{2j,2} \gamma_{2j+1,1} \}. \end{split}\] Thus the system can be taken as two 1D Majorana chains coupled by potential \(V_j\) and we may expect that there will be different topological phases depending on whether or not the Majorana fermions at the ends of the Majorana chains are paired. The phase transition from the extended state to the localized state could also be expected according to previous works on the quasiperiodic 1D Kitaev chain. In the following, we will discuss the properties of this generalized AAH model both in the incommensurate case and the commensurate case. # Incommensurate case {#sec3} When \(\alpha\) is irrational, the model system becomes quasiperiodic. Since the model is similar to a 1D Kitaev model with modulated on-site potentials which shows an Anderson localization transition when \(V\) becomes large enough \[\(V > 2(t+\Delta)\)\], we may expect that the same transition would also show up in our system. We take \(\alpha = (\sqrt{5}-1)/2\) as an example, but the results can also be generalized to other incommensurate situations. The Hamiltonian can be diagonalized by using the Bogoliubov-de Gennes (BdG) transformation, \[\label{} \eta_n^\dagger = \sum_{j=1}^{N} [u_{n,j} c_{n,j}^\dagger + v_{n,j} c_{n,j}],\] where \(u_{n,j}\) and \(v_{n,j}\) are chosen to be real and \(n\) is the energy band index. Then the wave function of the Hamiltonian is \[\label{} | \Psi_n \rangle = \eta_n^\dagger | 0 \rangle = \sum_{j=1}^{N} [u_{n,j} c_{n,j}^\dagger + v_{n,j} c_{n,j}] | 0 \rangle.\] From the Schrödinger equation \(\mathcal{H} | \Psi_n \rangle = \epsilon_n | \Psi_n \rangle\), we have where \[\label{} A_j = \begin{pmatrix} V_j & 0 \\ 0 &-V_j \end{pmatrix},\] \[\label{} B=\begin{pmatrix} -t_j &-\Delta \\ \Delta & t_j \end{pmatrix},\] and \[\label{} C=\begin{pmatrix} -t_{j+1} & \Delta \\ -\Delta & t_{j+1}. \end{pmatrix}\] Here we have assumed a periodic boundary condition which implies that \(c_{j+N}=c_j\). This is legitimate when \(N\) is large enough since the irrational number \(\alpha\) can be approximated by rational numbers. From this Hamiltonian, we can get \(u_{n,j}\) and \(v_{n,j}\) and thus the wave function of the system. In order to investigate the phase transition of this incommensurate 1D chain, we need to calculate the inverse participation ratio (IPR) which for a normalized wave function \(\Psi_n\) is defined as \(IPR=\sum_j (u_{n,j}^4 + v_{n,j}^4)\). The IPR measures the inverse of the number of occupied lattice sites and is a very useful quantity in characterizing the localization transitions of quasiperiodic systems. We further define the mean inverse participation ratio (MIPR) as \(MIPR = \frac{1}{2N} \sum_{n=1}^{2N} \sum_{j=1}^{N}(u_{n,j}^4 + v_{n,j}^4)\). MIPR is close to zero for extended states of the system, however, it will tend to a finite value of \(O(1)\) for localized states. Figure [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} shows the variation of the MIPR as the parameter of the system changes. Throughout this paper, we take \(t=1\) as the energy unit. Other parameters in Fig. [\[fig1\]](#fig1){reference-type="ref" reference="fig1"} are \(\Delta=0.2t\), \(\varphi=0\), and the number of lattice sites \(N=377\). The solid blue dots in Fig.[\[fig1a\]](#fig1a){reference-type="ref" reference="fig1a"} corresponds to the usual 1D Kitaev chain with a quasiperiodic disordered on-site potential (\(\lambda=0, V \neq 0\)). The phase transition takes place when \(V_c>2(t+\Delta)\), as was reported in earlier work. There is a plateau in the MIPR before it increases sharply when \(\lambda=0\), which corresponds to the critical phase before the system enters the localized phase, which is discussed in Ref.. If \(\lambda\) becomes nonzero and gets stronger, as is shown in Fig. [\[fig1a\]](#fig1a){reference-type="ref" reference="fig1a"}, the critical value of the phase transition \(V_c\) is reduced and the critical region will disappear gradually. It becomes more significant when \(\lambda\) is close to the value of \(t\), as shown by the triangular green dots and the open purple dots in Fig. [\[fig1a\]](#fig1a){reference-type="ref" reference="fig1a"}, where the MIPR increases rapidly at a much smaller value of \(V\). This is understandable since as \(\lambda\) increases, especially when it reaches the strength of \(t\), the modulation in the hopping amplitude will become increasingly stronger, the coupling strength between some lattice sites then will be weakened and the system thus can be localized more easily and more quickly. In Fig. [\[fig1b\]](#fig1b){reference-type="ref" reference="fig1b"}, we present the evolution of the MIPR as a function of \(\lambda\) with different \(V\). From this figure, we can see that when \(V=0\), the MIPR will increase from zero to a finite though small value, which means that even without the disorders coming from the on-site potentials, the modulation in the hopping amplitude alone can also influence the extended state of the system. When \(V \neq 0\), it is clear that the MIPR increases sharply when \(\lambda\) is close to \(1\) and reaches to a maximum, then drops again when \(\lambda\) becomes stronger. This signifies that the system is more likely to be localized when the hopping amplitudes are also modulated, which is consistent with the results we get in Fig. [\[fig1a\]](#fig1a){reference-type="ref" reference="fig1a"}. So due to the modulation in the hopping amplitude between the nearest neighboring lattice sites, we find that the incommensurate generalized AAH model with p-wave superconducting pairing is easier to be localized when tuning the disordered on-site potential. The critical value of the potential beyond which the phase transition happens is significantly reduced especially when the modulation in the hopping amplitude is strong. These results can also be generalized to situations with other irrational \(\alpha\) values. # Commensurate case {#sec4} When \(\alpha\) is rational, the lattice is commensurate. It is well known that the system will not undergo a localization transition as that in the incommensurate case. In this section, we mainly discuss the \(\alpha=1/2\) case. The generalized AAH model with p-wave superconducting pairing in this case can be tuned between different topological phases including the Su-Schrieffer-Heeger-like (SSH-like) topological phase and the Kitaev-like topological phase. When \(\alpha=1/2\), the Hamiltonian in Eq. ([\[Eq1\]](#Eq1){reference-type="ref" reference="Eq1"}) becomes This Hamiltonian is very similar to the one discussed in Ref., except that here the modulations are added to the on-site potential and the hopping amplitude by the function \(\lambda \cos \varphi\). It is reduced to the SSH model when \(\mu=\Delta=0\) and to the 1D Kitaev model when \(\lambda=0\) except that the chemical potential is positive or negative alternatively. Since both SSH-like and Kitaev-like models are included in this generalized system, we may expect that with appropriate system parameters, different topological phases would show up. To verify this we can check the energy spectrum of the system, which can be reached by diagonalizing the Hamiltonian rewritten in the Majorana fermions representation \[see Eq. ([\[Eq2\]](#Eq2){reference-type="ref" reference="Eq2"}))\]. In Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}, we present the spectra of the system with an even site number (\(N=100\)) and an odd site number (\(N=101\)). From Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(a), we can see that there are gapped topologically trivial regimes and nontrivial regimes with zero energy modes. However, comparing with Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(b), it is clear that when \(V\) becomes nonzero, parts of the zero modes in the topologically nontrivial regime split into two non zero energy modes, but the rest of the zero modes remain unchanged. More interestingly, when the site number is odd, there are also zero modes in the trivial regime when the site number is even and those zero modes also split when \(V\) becomes nonzero \[see Fig. [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(c) and [\[fig2\]](#fig2){reference-type="ref" reference="fig2"}(d)\]. This even-odd effect in the energy spectrum is a feature of the SSH model. The regime with zero energy modes which will not split when \(V\) is finite is the Kitaev topological regime. So there are three phases in this generalized system: a SSH-like trivial phase, a SSH-like topological phase, and a Kitaev-like topological phase. In the SSH-like topological phase, there will be one Dirac fermion at each end of the chain, while in the Kitaev-like topological phase, there will be one Majorana fermion at each end of the chain. The zero energy modes originate from the particle-hole symmetry which has two sources: In the SSH-like phase, it is the sublattice symmetry (when \(V=0\)) while in the Kitaev-like phase, it is the superconductivity. As the superconducting pairing strength becomes stronger, the superconductivity will dominate and there will be no SSH-like phase in the system. Figure [\[fig3\]](#fig3){reference-type="ref" reference="fig3"} shows the energy spectra of the system with \(\Delta =0.4t\) and \(0.5t\) both in the even and odd site number situations. When \(\Delta=0.4t\), there is only a very small region which shows the even-odd effect, as shown in Figs. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(a) and [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(b). When \(\Delta\) becomes larger, there are always zero modes as long as the system is in the topological nontrivial phase \[Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(c) and [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}(d)\]. So the Kitaev-like phase dominates and the SSH-like phase disappears. As the parameters change, there would be a phenomenon of electron fractionization to the Majorana fermions. To give a more intuitive picture of the different topological phases in this commensurate situation, we illustrate the schematic diagrams of the AAH model in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"} by setting \(\alpha=1/2\) in Eq. ([\[Eq2\]](#Eq2){reference-type="ref" reference="Eq2"}). As shown by Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(a), the AAH model can be taken as two Majorana fermion chains which are coupled by the on-site chemical potentials. The different bonding strengths between the Majorana fermions will lead to different topological phases. When the system is in the SSH-like topological phase, there is a fermion at each end of the one-dimensional chain, \[see Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(b)\]. If the system is in the Kitaev-like topological phase, there will be an unpaired Majorana fermion at each end of the chain, which is shown in Fig. [\[fig4\]](#fig4){reference-type="ref" reference="fig4"}(c). In order to characterize these different topological phases and determine the phase boundaries, we need to calculate the topological numbers. First, we transform the Hamiltonian into the momentum representation. By defining \[\Psi = [c_{qA}^\dagger \ c_{qB}^\dagger \ c_{-qA} \ c_{-qB}]^T,\] the Hamiltonian can be written as \(\mathcal{H}=\frac{1}{2} \sum_q \Psi^\dagger \mathcal{H} (q) \Psi\), where \[\label{} \mathcal{H} (q)= \begin{pmatrix} -V\cos \varphi & g(q) & 0 & h(q) \\ g^*(q) & V\cos \varphi &-h^*(q) & 0 \\ 0 &-h(q) & V\cos \varphi &-g(q) \\ h^*(q) & 0 &-g^*(q) &-V\cos \varphi \end{pmatrix}\] with \(g(q) =-t [ (1-\lambda \cos \varphi) + (1+ \lambda \cos \varphi) e^{-iq} ]\) and \(h(q) =-\Delta (1-e^{-iq})\). After diagonalizing \(\mathcal{H} (q)\) the eigenvalues can be obtained as \[\label{} \begin{split} E^2&(q) = (V \cos \varphi)^2 + |g(q)|^2 + |h(q)|^2 \\ & \pm 2\sqrt{(V\cos \varphi)^2 |h(q)|^2 + 4\Delta^2 t^2 (\lambda \cos \varphi)^2 (\cos q-1)^2}. \end{split}\] Thus we have \[\label{eigenvalues} \begin{split} E(0)=&\pm \sqrt{(V\cos \varphi)^2 + 4t^2}, \\ E(\frac{\pi}{2}) =& ( \sqrt{(V\cos \varphi)^2 + 2t^2 (\lambda \cos \varphi)^2} \pm \sqrt{2} \Delta)^2 + 2t^2, \\ E(\pi) =& \pm \sqrt{(V\cos \varphi)^2 + 4t^2 (\lambda \cos \varphi)^2} \pm 2\Delta. \end{split}\] The gap will close at \(q=\pi\) when \((V\cos \varphi)^2 = 4[\Delta^2-t^2 (\lambda \cos \varphi)^2]\). For \(\Delta=0\), the spectrum is reduced to \[\label{} \begin{split} &E(q)=\\ &\pm \sqrt{V^2 \cos^2 \varphi + 2t^2 [ 1+\lambda^2 \cos^2 \varphi + (1-\lambda^2 \cos^2 \varphi) \cos q ]} \end{split}\] which is similar to the SSH model spectrum. If \(\lambda=0\), then the spectrum is reduced to \[\label{} E(q) = \pm \sqrt{(V\cos \varphi \pm 2\Delta \sin \frac{q}{2})^2 + 4t^2 \cos^2 \frac{q}{2}},\] which is similar to the Kitaev model spectrum. The difference here is that the system is topological for \(|V \cos \varphi| <2|\Delta|\) and trivial for \(|V \cos \varphi|>2|\Delta|\). However, the spectrum for a usual Kitaev model is \(E(q)=\pm \sqrt{(2t\cos \frac{q}{2}-\mu)^2 + 4 \Delta^2 \sin^2 \frac{q}{2}}\), where the system is topological for \(|\mu|<2t\) and trivial for \(|\mu|>2t\). In order to make further investigations about the topological phases of the commensurate lattice, we need to check the topological quantum numbers of the system with different parameters. Following the discussions in Ref., we find that the topological class of our model is also BDI and we can characterize our system by the \(\mathbb{Z}\) index. Now we calculate the topological numbers of the system. Firstly, we consider the sublattice symmetric case (\(V=0\)) whose topological number can be calculated as (check the Appendix for details) \[\label{} N_1 = \Theta (t \lambda \cos \varphi-\Delta) + \Theta (t\lambda \cos \varphi +\Delta)\] with \(\Theta(x)\) being the step function. Figure [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}(a) illustrates the phase diagram of the system according the value of topological number \(N_1\). There are three phases: (i) When \(|\Delta|< t \lambda |\cos \varphi|\), \(\cos \varphi < 0\), \(N_1=0\), the system is in the SSH-like trivial phase; (ii) when \(|\Delta| < t\lambda |\cos \varphi|\), \(\cos \varphi >0\), \(N_1 =2\), the system is in the SSH-like topological phase; (iii) when \(|\Delta|>t\lambda |\cos \varphi|\), \(N_1=1\), the system is in the Kitaev-like topological phase. As shown in this figure, when \(|\Delta| > \lambda\) (here \(\lambda=0.4t\)), the system will be totally in the Kitaev-like topological phase, which is consistent with the numerical results for the energy spectra we presented in Fig. [\[fig3\]](#fig3){reference-type="ref" reference="fig3"}. The topological number \(N_1\) can be interpreted as the number of Majorana fermion pairs at the ends of the chain. So when \(N_1=2\), there will be two Majorana fermions at each end of the chain which will pair up with each other and become a Dirac fermion, which is the feature of the SSH model. If \(N_1=1\), there will be one single unpaired Majorana fermion at each end of the chain, so the system is in the Kitaev-like topological phase. Next we check the \(V\neq 0\) case. Using similar techniques, the topological number \(N_2\) for this sublattice asymmetric case is (see the Appendix) \[\label{} N_2 =-\int_{-\pi}^{\pi} \frac{dq}{2\pi i} \partial_q \ln Z(q),\] where \[\label{} \begin{split} Z(q) =& Det M_2(q) = (V \cos \varphi)^2 + (g-h)(g^* + h^*) \\ =& (V \cos \varphi)^2 + 2[ t^2(1+\lambda^2 \cos^2 \varphi)-\Delta^2 ] \\ & + 2[ t^2(1-\lambda^2 \cos^2 \varphi) + \Delta^2 ]\cos q-4i t\Delta \sin q. \end{split}\] Here, \(N_2\) is the winding number of \(Z(q)\), and is determined by the cross points of the real axis at \(q=0\) and \(\pi\). When \(\Delta > 0\), we have \[\label{} \begin{split} Z(0)Z(\pi) < 0 &\Rightarrow N_2=1, \\ Z(0)Z(\pi) > 0 &\Rightarrow N_2=0, \end{split}\] with \[\label{} \begin{split} Z(0) =& (V \cos \varphi)^2 + 4 t^2, \\ Z(\pi) =& (V\cos \varphi)^2 + 4(t^2 \lambda^2 \cos^2 \varphi-\Delta^2). \end{split}\] When \(\Delta<0\), we have \(N_2=-1\) in the topological regime but it is equivalent to the \(N_2=1\) phase. In Figs. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}(b)-Fig. [\[fig5\]](#fig5){reference-type="ref" reference="fig5"}(d), we present the phase diagrams of the system according to the topological number \(N_2\). There are two phases when \(V \neq 0\): \(N_2 = \pm1\) is the Kitaev-like topological phase, and \(N_2=0\) is the trivial phase. The phase boundary between these two phases is \((V\cos \varphi)^2 = 4[\Delta^2-t^2 (\lambda \cos \varphi)^2]\), which is the gap closing condition we got earlier \[see Eq. ([\[eigenvalues\]](#eigenvalues){reference-type="ref" reference="eigenvalues"})\]. So, by checking the energy spectra and the topological numbers of the commensurate AAH model, we can completely describe the phases of the system. In addition, we can also tune the system into different topological phases by changing the system parameters such as the superconducting pairing amplitude and the modulation of the hopping amplitude, etc. # Summary {#sec5} In this paper, we have investigated a generalized Aubry-André-Harper model with superconducting pairing both in the incommensurate and commensurate case. Due to the modulations in the hopping amplitude and the on-site potential, this model shows aboundant physical phenomena as we tune the system parameters. When the lattice is incommensurate \[\(\alpha=(\sqrt{5}-1)/2\)\], the system shows a phase transition from extended states to localized states when the disordered potential becomes larger than some critical value. This critical value as well as the critical region will be reduced due to the modulation in the hopping amplitude especially when this modulation become strong. On the other hand, in the commensurate lattice case (\(\alpha=1/2\)), the system presents three different phases in different parameter regimes: a SSH-like trivial phase, a SSH-like topological phase, and a Kitaev-like topological phase. The topological numbers and the phase diagram are illustrated in this work. This generalized AAH model combined with superconducting pairing unifies those different phases and can provide an ideal test field for different kinds of phenomena. All these results can be generalized to discuss other cases with different \(\alpha\) values, both rational and irrational. As proposed in Ref. (see the Supplemental Material there), a generalized AAH model with modulations in the hopping amplitude could be realized in cold atom systems. The generalized Aubry-André-Harper model we propose here, though more complicated, might also be realized in cold atom optical lattices, or semiconductor structures in the future. This generalized AAH model may also have potential applications in understanding the topological properties of other similar systems, as discussed in Refs.. # Appendix {#appendix .unnumbered} In this Appendix, we investigate the topological class and calculate the topological numbers of the system. This part is mainly following the discussions in Ref.. Our model is time-reversal symmetric, \(T \mathcal{H} (q) T^{-1} = \mathcal{H} (-k)\), since the coefficients \(V\), \(t\), \(\lambda\) and \(\Delta\) are all real. The time-reversal operator is defined as \(T=K\), which takes the complex conjugate. When \(V=0\), the system has sublattice symmetry. The operator for sublattice symmetry is defined by \[\label{} C_1 = \sigma_z = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 &-1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 &-1 \end{pmatrix},\] where \(\sigma_z\) is the Pauli matrix acting on the sublattice degree of freedom. It can be shown that \(C_1 \mathcal{H} (q) C_1^{-1} =-\mathcal{H}(q)\). Since we have \(T^2=1\) and \(C_1^2 = 1\), the topological class of the system is BDI. If \(V \neq 0\), the system will no longer have sublattice symmetry anymore. However, due to the superconducting pairing, particle-hole symmetry can still be guaranteed. The particle-hole operator is defined by \(P=\tau_x K\), with \(\tau_x\) being the Pauli matrix acting on the particle-hole space. The Hamiltonian satisfies \(P \mathcal{H} (q) P^{-1} =-\mathcal{H} (-q)\). We can introduce the chiral operator as \[\label{} C_2 = TP = \tau_x = \begin{pmatrix} 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \\ 1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \end{pmatrix}.\] Then we have \(C_2 \mathcal{H}(q) C_2^{-1} =-\mathcal{H} (q)\) and \(C_2^2 = 1\), so the topological class is still BDI. The 1D system in the BDI class is characterized by the \(\mathbb{Z}\) index. Now we calculate the topological numbers of the system. Firstly, we consider the sublattice symmetric case (\(V=0\)) whose topological number is defined as \[\label{} N_1 = Tr \int_{-\pi}^{\pi} \frac{dq}{4\pi i} C_1 g^{-1} \partial_q g,\] where \(g(q)=-\mathcal{H}^{-1} (q)\) is the Green's function at zero energy. \(N_1\) is equivalent to the chiral index. After introducing a unitary transformation \[\label{} U_1 = \begin{pmatrix} 1 & 0 & 0 & 0 \\ 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 0 & 1 \end{pmatrix},\] we have \[\label{} U_1 C_1 U_1^\dagger = \tau_z, \hspace{1cm} U_1 \mathcal{H} U_1^\dagger = \begin{pmatrix} 0 & M_1 \\ M_1^\dagger & 0 \end{pmatrix},\] with \[\label{} M_1= \begin{pmatrix} g & h \\ -h &-g \end{pmatrix}.\] where \(g(q)\) and \(h(q)\) are defined earlier. Then the chiral index is given by \[\label{} \begin{split} N_1= &-Tr \int_{-\pi}^{\pi} \frac{dq}{2\pi i} M_1^{-1} \partial_q M_1 \\ =&-\int_{-\pi}^{\pi} \frac{dq}{2\pi i} \partial_q \ln Det[M_1]\\ =&-\int_{-\pi}^{\pi} \frac{dq}{2\pi i} \partial_q \ln [(h-g)(h+g)] \end{split}\] with \[\label{} \begin{split} h-g =& [t (1-\lambda \cos \varphi)-\Delta] + [ t(1+\lambda \cos \varphi) + \Delta ]e^{-iq}, \\ h+g =& [-t(1-\lambda \cos \varphi)-\Delta]-[t(1+\lambda \cos \varphi)-\Delta] e^{-iq}. \end{split}\] So the topological number \(N_1\) is \[\label{} N_1 = \Theta ( t \lambda \cos \varphi-\Delta) + \Theta (t\lambda \cos \varphi + \Delta )\] with \(\Theta(x)\) being the step function. Next we check the \(V\neq 0\) case. The topological number \(N_2\) associated with the chiral operator \(C_2\) is \[\label{} N_2 = Tr \int_{-\pi}^{\pi} \frac{dq}{4\pi i} C_2 g^{-1} \partial_q g.\] Considering a unitary transformation \[\label{} U_2= \frac{1}{\sqrt{2}} \begin{pmatrix} 1 & 0 & 1 & 0 \\ 0 & 1 & 0 & 1 \\ -i & 0 & i & 0 \\ 0 &-i & 0 & i \end{pmatrix},\] which corresponds to the Majorana fermion representation \[\label{} c_j = \frac{1}{2}(\gamma_{j,1} + i \gamma_{j,2}), \hspace{1cm} c_j^\dagger = \frac{1}{2}(\gamma_{j,1}-i \gamma_{j,2}).\] It is easy to check that \[\label{} U_2 C_2 U_2^\dagger = \tau_z, \hspace{1cm} U_2 \mathcal{H}U_2^\dagger = \begin{pmatrix} 0 & M_2 \\ M_2^\dagger & 0 \end{pmatrix},\] with \[\label{} M_2= \begin{pmatrix} -i V\cos \varphi & i(g-h) \\ i(g^* + h^*) & i V\cos \varphi \end{pmatrix}.\] Thus we have \[\label{} N_2 =-Tr \int_{-\pi}^{\pi} \frac{dq}{2\pi i} M_2^{-1} \partial_q M_2 =-\int_{-\pi}^{\pi} \frac{dq}{2\pi i} \partial_q \ln Z(q),\] where \[\label{} \begin{split} Z(q) =& Det M_2(q) = (V \cos \varphi)^2 + (g-h)(g^* + h^*) \\ =& (V \cos \varphi)^2 + 2[ t^2(1+\lambda^2 \cos^2 \varphi)-\Delta^2 ] \\ & + 2[ t^2(1-\lambda^2 \cos^2 \varphi) + \Delta^2 ]\cos q-4i t\Delta \sin q. \end{split}\] \(N_2\) is the winding number of \(Z(q)\), and is determined by the cross points of the real axis at \(q=0\) and \(\pi\). For \(\Delta > 0\), we find \[\label{} \begin{split} Z(0)Z(\pi) < 0 &\Rightarrow N_2=1, \\ Z(0)Z(\pi) > 0 &\Rightarrow N_2=0, \end{split}\] with \[\label{} \begin{split} Z(0) =& (V \cos \varphi)^2 + 4 t^2, \\ Z(\pi) =& (V\cos \varphi)^2 + 4(t^2 \lambda^2 \cos^2 \varphi-\Delta^2). \end{split}\] For \(\Delta<0\), we have \(N_2=-1\) in the topological regime but it is equivalent to the \(N_2=1\) phase.

{'timestamp': '2016-09-26T02:01:24', 'yymm': '1607', 'arxiv_id': '1607.04835', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04835'}

# Results {#results .unnumbered} ## Estimating point-wise uncertainty {#estimating-point-wise-uncertainty .unnumbered} Estimating uncertainty is closely related to estimating positions, where the maximum likelihood estimate (MLE) is generally considered the optimal method. A maximum likelihood method starts with a likelihood function, i.e., the probability density function of a probabilistic model for generating images of spots (pixel counts in a small region around a spot) with the emitter position among the adjustable parameters. The MLE are the parameters that maximize the likelihood function for a particular spot image. Following common practice, we model EMCCD camera noise with the high-gain approximation and Gaussian readout noise, and the spot shape by a symmetric Gaussian intensity profile plus a constant background intensity. The fit parameters are thus spot position \((\mu_x,\mu_y)\), background intensity \(b\), PSF width \(\sigma\), and spot amplitude \(N\) (see Methods, ), while the camera noise is assumed known from calibration. What is localization uncertainty? The localization error is the difference \(\mu_\text{est.}-\mu_\text{true}\) between estimated and true positions. By localization uncertainty, we seek the distribution of the error, either in a Bayesian posterior sense, or in the sense of repeated localizations of spots with the same uncertainty. The uncertainty of the localization is related to the shape of the likelihood maximum: a sharply peaked maximum means that only a narrow set of parameters are likely, while a shallow maximum means greater uncertainty. The Cramér-Rao lower bound (CRLB) is the smallest possible variance of an unbiased estimator for a given set of model parameters, and is related to the expected sharpness of the likelihood maximum (see Methods, ). While this is strictly speaking not a statement about a single image, but rather about the average information content of data generated by a model, it is often used to estimate localization uncertainty. A Bayesian alternative is the Laplace approximation, which derives an approximate Gaussian posterior distribution for the fit parameters in terms of the sharpness of the likelihood maximum for each particular image (see Methods, ). Both estimators supply an estimated variance \(\varepsilon^2\), but none of them are well characterized as estimators of localization uncertainty. To test these estimators, we analyzed a large set of simulated movies of a fluorescent particle diffusing at D=1 μ^2^m/s in an *E. coli*-like geometry. The movies cover a broad range of experimentally relevant imaging conditions (see Methods) and include realistic EMCCD noise, background fluorescence, a non-Gaussian and space-dependent PSF, and motion blur. A basic consistency check is that the average estimated error variance, \(\mean{\varepsilon^2}\), agrees with the variance of the actual errors, \(\mean{(\mu_\text{est.}-\mu_\text{true})^2}\). In Fig. [\[fig:simErr\]](#fig:simErr){reference-type="ref" reference="fig:simErr"}a, we compare the square root of these quantities for different imaging conditions, based on MLE localizations combined with either CRLB or Laplace uncertainty estimators. We obtain consistency under good imaging conditions, where the spots are bright and the average errors low. However, as conditions worsen and the errors increase, the uncertainty is underestimated, especially by the CRLB. There are several possible reasons for this behavior. The Laplace approximation is based on a truncated Taylor expansion of the log likelihood (see Methods). The CRLB is strictly not a point-wise error estimator at all, and is further defined in terms of the true parameter values, although by necessity evaluated with the fitted ones. The simplified Gaussian PSF model performs well for localization, but this does not guarantee good uncertainty estimates. Any of these approximations might fail in noisy or low light conditions. Another possibility is that localizations under poor imaging conditions are corrupted by a sub-population of fits that converge to a local minimum that does not reflect the underlying spot shape, e.g., fitting a single bright pixel as a very narrow spot, or misinterpreting the PSF shoulders as background or vice versa. To address the latter complication, we constrained the localizations using prior distributions on selected parameters, thus replacing MLE with maximum a aposteriori estimation (MAP). Fixing parameter values is not suitable here, since both the size and shape of spots fluctuate due to fluorophore motion and varying imaging conditions. Furthermore, it is experimentally easier to obtain independent information about the background and PSF width than about the spot intensity. We therefore limit our attention to background and PSF width, and found the following priors to perform well: a log-normal prior centered on the true value for the background intensity, and a skewed log-normal prior for the PSF width to penalize fits with unphysical widths below that of an in-focus spot. The PSF width prior together with the distribution of fitted PSF widths are shown in Fig. [\[fig:simErr\]](#fig:simErr){reference-type="ref" reference="fig:simErr"}b. PSF widths below 1 pixel (80 nm) are virtually eliminated in the MAP fits. Background and spot amplitudes (see Fig. [\[SI:fig:parPrior\]](#SI:fig:parPrior){reference-type="ref" reference="SI:fig:parPrior"}) are shifted somewhat downwards and upwards, respectively. As seen in Fig. [\[fig:simErr\]](#fig:simErr){reference-type="ref" reference="fig:simErr"}c, this substantially improves the agreement between true and estimated errors under all imaging conditions. The Laplace estimator still outperforms the CRLB by a small margin, and numerical experiments on a wider range of priors (see SI Fig. [\[SI:fig:priors_SI\]](#SI:fig:priors_SI){reference-type="ref" reference="SI:fig:priors_SI"}) further confirm that the Laplace estimator is more robust to non-optimal priors than the CRLB. How does the prior help? A direct comparison of the true errors for the MLE and MAP fits (SI Fig. [\[SI:fig:MLE_vs_MAPerrors\]](#SI:fig:MLE_vs_MAPerrors){reference-type="ref" reference="SI:fig:MLE_vs_MAPerrors"}) reveals very small differences, indicating that the improvement is mainly due to improved uncertainty estimates. This is consistent with the theoretical observation that the Fisher information matrix for the localization problem is nearly block-diagonal with very weak coupling between two groups of fitting parameters: positions (\(\mu_x,\mu_y\)) and shape parameters (\(N,b,\sigma\)). Thus, additional information about one of these groups does not reduce the errors of the other significantly. For uncertainty estimation, this is not the case: information about the background and PSF shape does help in estimation position uncertainty. While mean-square errors are useful, we are ultimately interested in the full distribution of errors. In particular, most (but not all ) statistical models of SPT data assume Gaussian errors, but this assumption has not been tested. The Laplace approximation () is a Gaussian approximation. If it was exact, the errors normalized by the estimated standard deviation would be Gaussian with unit variance, and produce a straight line in the Gaussian probability plot of Fig. [\[fig:simErr\]](#fig:simErr){reference-type="ref" reference="fig:simErr"}d. The actual normalized errors based on MAP localizations is more Gaussian than the MLE ones, and follow the straight line for almost four standard deviations. This shows that the improved consistency of the MAP estimates also translates to more Gaussian error distributions. Overall, these results show that point-wise uncertainty estimation using the Laplace approximation works well in a wide range of experimentally relevant conditions, but that experimentally accessible additional information about background and PSF shape is necessary to get consistent results in low light conditions. Moreover, we see good support for the common assumption of Gaussian errors. ## Validation on real data {#validation-on-real-data .unnumbered} To test the above conclusions on real data, we imaged immobilized fluorescent beads, alternating strong and weak excitation as shown in Fig. [\[fig:realData\]](#fig:realData){reference-type="ref" reference="fig:realData"}a. We used images under strong excitation conditions to extract a high accuracy ground truth for testing the uncertainty estimates in the dim images. We estimated the position and uncertainty of spots using the MAP estimation described above, with a background prior centered around the mean background (0.8 photons/pixel) seen in dim frames. A drift-corrected ground truth was estimated by cubic spline interpolation between the mean positions obtained from each block of 10 consecutive bright images. Since the intensity differs by about a factor 10 between bright and dim frames, the RMS errors of the ground truth should be approximately 10-fold lower than that of a single dim spot. Fig. [\[fig:realData\]](#fig:realData){reference-type="ref" reference="fig:realData"}c shows the resulting error-uncertainty comparison, with every point corresponding to a single bead. It reproduces the behavior on simulated images in Fig. [\[fig:simErr\]](#fig:simErr){reference-type="ref" reference="fig:simErr"}b, thus confirming our conclusion that the Laplace approximation is preferable to CRLB for uncertainty estimators, and that the good performance of our proposed PSF width prior is not limited to that particular simulated data set. ## Diffusion constants {#diffusion-constants .unnumbered} Next, we consider how estimated localization uncertainties leads to better estimates of diffusion constants, arguably the most common analysis of single particle tracking data. We divided the synthetic data shown in Fig. [\[fig:frameExamples\]](#fig:frameExamples){reference-type="ref" reference="fig:frameExamples"} into trajectories of length 10, estimated positions and uncertainties with the MAP and Laplace estimators described above, and estimated diffusion constants using the covariance-based estimators of Ref.  with and without the use of uncertainty estimates. Fig. [\[fig:Dest\]](#fig:Dest){reference-type="ref" reference="fig:Dest"} shows the mean value and 1% quantiles of the two estimators applied to every imaging condition, plotted against the signal-to-noise ratio. As expected, the use of estimated uncertainties improves the variability of the diffusion estimates substantially. The covariance-based estimators only use the average uncertainty in each trajectory (see Methods). We also implemented a maximum likelihood estimator for the diffusion constant which makes explicit use of the point-wise uncertainties (see SI text [\[SI:sec:diffusionMLE\]](#SI:sec:diffusionMLE){reference-type="ref" reference="SI:sec:diffusionMLE"}), but found no further improvement. ## Analysis of multi-state data {#analysis-of-multi-state-data .unnumbered} We now turn to a more challenging problem where point-wise errors do matter: data where both the diffusion constant and localization error change significantly on similar time scales. In single particle tracking, changes in diffusion constant can be used as a non-invasive reporter on intracellular binding and unbinding events. However, diffusive motion and localization errors contribute additively to the observed step length statistics (see Eq. [\[MT:eq:dxcov\]](#MT:eq:dxcov){reference-type="ref" reference="MT:eq:dxcov"}), and thus changes in diffusion constants and localization errors cannot be reliable distinguished. As an example, we consider a protein that alternates between free diffusion (\(D=1\) μ^2^m/s) and a bound state simulated by slow diffusion (\(D=0.1\) μ^2^m/s). We focus on a single trajectory with 4 binding/unbinding events, two of which occur about 400 nm out of focus, and thus are accompanied by substantial broadening of the PSF and accompanying increases in localization errors. This defocus matches roughly the radius of an *E. coli* cell, and the scenario could model tracking experiments with cytoplasmic proteins that can bind to the inner cell membrane. Using SMeagol, we simulated 12 000 replicas of the above set of events, at a camera frame rate of 200 Hz, continuous illumination, and 300 photons/spot on average. Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}a shows the \(z\) coordinates in the input trajectory, and the frame-wise RMS errors produced by the MAP localization algorithm described above. The input points are sparse (tens of ms apart), which allows SMeagol to produce simulated movies that contain the same binding events and defocus trends, but vary in the detailed diffusion trajectory as well as in noise realizations. Examples of simulated spots along a trajectory are shown in Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}b. To analyze this challenging data set, we extended the maximum likelihood diffusion constant estimator with explicit point-wise errors to include multiple diffusion states governed by an hidden Markov model (HMM), for which we derived a variational EM algorithm (see SI text [\[SI:sec:HMM\]](#SI:sec:HMM){reference-type="ref" reference="SI:sec:HMM"}). We then analyzed each simulated trajectory with three different 2-state HMMs: (i) vbSPT, which models the observed positions as pure diffusion and neglects blur and localization errors, (ii) the above-mention HMM that explicitly models these effects, and (iii) the Kalman filter limit of the HMM in (ii), which models localization errors but not blur effects. Since multi-state Kalman-type algorithms have been studied previously, it is interesting to compare models with and without blur. Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}c shows the inferred average state from the three different models. As expected, the HMMs that include localization errors outperform vbSPT at detecting the strongly defocused first and third binding events. The full HMM does not give the best classification of the two short binding events, but it does give the lowest overall misclassification rate, 9% versus 9.6% and 17% for the Kalman and vbSPT models, respectively, so the apparent worse time resolution might reflect an overall tendency to invent spurious transitions by the Kalman and vbSPT models. Next, we look at estimated diffusion constants. Here, the Kalman and vbSPT models make systematic errors as seen in the bare parameters in Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}d. However, by comparing the step length statistics between the full and simplified models, one can derive heuristic correction factors for these models (see Methods, ). As shown in Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}e, this reduces the bias substantially, especially for the high diffusion constant. To finally compare the different HMMs on more well-behaved data, we reran the same experiment but with all \(z\) coordinates rescaled by a factor 1/5 in the PSF model, which essentially removes the \(z\)-dependent defocus effects. On this less challenging data set, event detection is much improved and the differences between the three HMMs are much less pronounced. However, with misclassification rates of 4.8%, 6.6%, and 7.6% for the HMM, Kalman model, and vbSPT, respectively, the full HMM still does overall better. ## Position refinement {#position-refinement .unnumbered} Since the new HMM includes the true trajectory as a hidden variable and performs a global analysis, it can be used to refine individual localized positions, and in principle beat the Cramer-Rao lower bound for single image localizations. Fig. [\[fig:pos\]](#fig:pos){reference-type="ref" reference="fig:pos"}a shows the true, measured, and refined positions for part of a two-state trajectory, and Fig. [\[fig:pos\]](#fig:pos){reference-type="ref" reference="fig:pos"}b the relative change of the RMS error for each frame in Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}a after refinement. The refinement improves the RMS errors with up to 50%. Large localization errors and small diffusion constant leads to larger relative improvement, which is intuitive since those factors both make a single point less informative relative its neighbors. A few points show a small error increase. On closer inspection, these turn out to be situated near hard-to-detect state-changes, and therefore tend to be refined using an incorrect diffusion constant. # Discussion {#discussion .unnumbered} Fluorophore positions are not the only useful kind of information in super-resolution microscopy images. Here, we have shown that point-wise position uncertainty can also be extracted and used to improve quantitative data analysis. This is particularly important for live cell data where dynamic phenomena can be studied, and one may expect more heterogeneous imaging conditions where many of the theoretical estimates of localization errors, that may apply for cells with immobilized internal structure ("fixed"), will have little relevance. In general, our results indicate that estimation of position uncertainty is more sensitive to the fit model than estimation of position. For parctical use, we find that an estimate based on the Laplace approximation combined with external information about the fluorescent background and PSF shape performs well in a wide range of experimentally relevant conditions. An intuitive reason why the CRLB performs worse may be that it makes explicit use of the fit model twice, for both fitting and predicting the model-dependent uncertainty, while the Laplace approximation only requires the first step. Theoretically, the posterior density formalism of the Laplace approximation also fits more naturally with model-based time-series analysis where the particle trajectory is treated as a latent variable to be integrated out. While we have focused on 2D localization using conventional optics, the extension to e.g., three dimensions using dual plane imaging or engineered PSFs present no principal difficulties, as long as appropriate PSF models for localization can be formulated. Note however that even a perfect characterization of an experimental PSF does not constitute a perfect localization model, since it does not describe random PSF shape fluctuations due to motion blur. Current techniques for 3D localization are inherently asymmetric and yield different in-plane and axial accuracy, which further underscores the need for downstream analysis methods to incorporate heterogeneous localization uncertainty. Most super-resolution microscopy applications are however not aimed at particle tracking, but imaging. For PALM/STORM type imaging of fixed samples, the ability to estimate the uncertainty of individual spots does not improve the localizations themselves, but it may still improve the final resolution since uncertain points can be omitted on a quantitative basis. However, the consequences for live cell imaging are more interesting, since the same fluorophore may be detected in different positions over different frames if the target is moving. For this case, we have shown that the combination of estimating uncertainty and modeling the fluorophore motion can produce refined position estimates, in principle pushing the localization errors below the single-spot Cramer Rao lower bound, by merging information from consecutive frames in an optimal way. ## Synthetic data {#synthetic-data .unnumbered} We generated synthetic microscopy data using SMeagol, a software for accurate simulations of dynamic fluorescence microscopy at the single molecule level. We modeled the optics using a Gibson-Lanni PSF model with 584 nm wavelength and NA=1.4, and an EMCCD camera with the high gain approximation plus Gaussian readout noise. For localization and diffusion estimation tests, we simulated simple diffusion (\(D=1\)μ^2^m/s) in a cylinder of length 14.4 μm and diameter 2 μm, similar to long *E. coli* cells, in order to avoid confinement artifacts in the longitudinal direction. We generated 48 data sets spanning a wide range of imaging conditions by varying frame duration (8 ms or 10 ms, exposure time 2 ms, 4 ms, 6 ms, and 3 ms, 5 ms, 8 ms respectively, background intensity (1 or 2 photons per pixel), average spot brightness (75-300 photons/spot), EM gain 20 or 30, and readout noise level 4 or 8. Each data set contained \(10^4\) individual spots. For the simulated multi-state data, we hand-modified a single SMeagol input trajectory from a simulated 2-state model to contain 4 binding events with different durations and z-coordinates as seen Fig. [\[fig:HMM\]](#fig:HMM){reference-type="ref" reference="fig:HMM"}a, and also thinned out the input trajectory to create more variability in the particle paths between different realizations. We then simulated many realizations from this input trajectory, using the same PSF model as above, continuous illumination with a sample time of 5 ms, EMCCD gain 90, an average spot intensity of 300 photons/spot, and a time-dependent background that decays exponentially from 0.95 to 0.75 background photons per pixel with a time-constant of 0.75 s. ## Real data {#real-data .unnumbered} For estimating localization errors in the real imaging conditions, we use immobilized fluorescent beads with the diameter of 0.1 μm (TetraSpeck Fluorescent Microspheres, ThermoFischer T7284). The beads where diluted in ethanol and then placed on a coverslip where we let them dry in before adding water as a mounting medium. Imaging was done with a Nikon Ti-E microscope which was configured for EPI-illumination with a 514 nm excitation laser (Coherent Genesis MX STM) together with matching filters (Semrock dichroic mirror Di02-R514 with emission filter Chroma HQ545/50M-2P 70351). Intensity modulation was made possible by an acousto-optic tunable filter (AOTF) (AA Opto Electronics, AOTFnC) which was triggered by a waveform generator (Tektronix, AFG3021B). The waveform used was a sequence of square pulses, high for 200 ms and low for 1800 ms. The two illumination intensities, high and low, corresponds to 10.7 kW/c^2^m and 0.63 kW/c^2^m, respectively. Fluorescent beads where viewed through a 100x (CFI Apo TIRF 100x oil, Na=1.49) objective with a 2X (Diagnostic instruments DD20NLT) extension in front an Andor Ultra 897 EMCCD camera (Andor Technology Ltd.). This configuration puts the pixel size to 80 nm which is the same pixel size set in the simulated data. The data set constituted of 1000 frames (Fig. [\[fig:realData\]](#fig:realData){reference-type="ref" reference="fig:realData"}a) with an exposure time of 30 ms. EMCCD noise characteristics (gain, offset, readout noise) were determined by analyzing a "dark" movie obtained with the shutter closed. ## Localization {#localization .unnumbered} We perform maximum likelihood (MLE) localization using a high-gain EMCCD noise model, which relates the probability \(q(c_i|E_i)\) of the offset-subtracted pixel count \(c_i\) for a given pixel intensity \(E_i\) (expected number of photons/frame) in pixel \(i\). For the intensity \(E(x,y)\), we model the PSF with a symmetric Gaussian, \[\label{eq:gaussPSF} E(x,y)=\frac{b}{a^2}+\frac{N}{\sqrt{2\pi\sigma^2}} \exp\Big(-\frac{(x-\mu_x)^2+(y-\mu_y)^2}{2\sigma^2}\Big),\] with pixel size \(a\), background \(b\) (expected number of photons/pixel), spot width \(\sigma\), amplitude \(N\) (expected number of photons/spot), and spot position \((\mu_x,\mu_y)\), and approximate the pixel intensity \[E_i=\int_\text{px. \(i\)} E(x,y)dxdy\] by numerical quadrature. The log likelihood of an image containing a single spot is then given by \[\ln L(\theta) = \ln q_0(\theta)+\sum_{i\in ROI} \ln q(c_i|E_i(\theta)),\] where \(\theta=(\mu_x,\mu_y,b,N,\sigma)\) are fit parameters, and \(q_0\) is a prior distribution (we set \(\ln q_0=0\) for MLE fitting). To avoid the complications of spot identification, we use known positions to determine the ROI and initial guess for \((\mu_x,\mu_y)\). All localizations are performed using a \(9\times 9\) ROI. Fits that failed to converge or estimated uncertainties larger than 2 pixels where discarded. ## Cramer-Rao lower bound {#cramer-rao-lower-bound .unnumbered} The CRLB is a lower bound on the variance of an unbiased estimator. We use an accurate approximation to the CRLB for a symmetric Gaussian PSF from Ref. , \[\label{eq:CRLB} \varepsilon^2_\text{CRLB}=2\frac{\sigma_a^2}{N}\left( 1+4\tau+\sqrt{\frac{2\tau}{1+4\tau}} \right),\] with \(\tau=2\pi\sigma_a^2 b/(Na^2)\), \(\sigma_a^2=\sigma^2+a^2/12\), and the prefactor 2 accounts for EMCCD excess noise. ## Laplace approximation {#laplace-approximation .unnumbered} An alternative way to approximate the uncertainty of the fit parameters is to Taylor expand the likelihood around the maximum-likelihood parameters \(\theta^*\) to second order, that is \[\label{eq:laplace} \ln L(\theta)\approx \ln L(\theta^*)+ \frac{\partial \ln L}{\partial\theta}\Big|_{\theta^*}(\theta-\theta^*) +\frac12(\theta-\theta^*)^T \frac{\partial^2\ln L}{\partial \theta^2}\Big|_{\theta^*} (\theta-\theta^*).\] The first-order term is zero, since \(\theta^*\) is a local maximum. This approximates the likelihood by a Gaussian with covariance matrix given by the inverse Hessian, i.e., \(\Sigma=[\partial^2\ln L/\partial \theta^2]^{-1}\). In a Bayesian setting, this expresses the (approximate) posterior uncertainty about the fit parameters. The estimated uncertainties (posterior variances) are given by the diagonal entries of the covariance matrix, e.g., \(\varepsilon^2_\text{Lap.}(\mu_x)=\Sigma_{\mu_x,\mu_x}\). We compute the Hessian numerically using Matlab's built-in optimization routines, and use the log of the scale parameters \(b,N,\sigma\) for fitting, since they are likely to have a more Gaussian-like posterior. ## Prior distributions {#prior-distributions .unnumbered} The prior distributions for localization used in the main text are: normal priors with mean value \(\ln b_\text{true}\) and std. 0.2 for the log background intensity, resulting in the log-normal prior \(b\in\ln N(\ln b_\text{true},0.2^2)\). For the PSF width \(\sigma\), we use a skewed normal prior for \(\ln \sigma\) to penalize fits with \(\sigma\) below the minimum in-focus width of the PSF. The skew normal density is given by \[\label{eq:skewed_prior} f(x,x_0,w,\alpha) = \frac{2}{w} \phi\left(\frac{x-x_0}{w}\right)\Phi\left(\alpha\frac{x-x_0}{w}\right),\] where \(\phi\) is the probability distribution function and \(\Phi\) is the cumulative distribution function for the unit normal distribution \(N(0,1)\). We used \(x_0=\log(1.5)\), \(w=1\), and \(\alpha = 5\), as sketched in Fig. [\[fig:simErr\]](#fig:simErr){reference-type="ref" reference="fig:simErr"}b, with the pixel width \(a=80\) nm as the length unit. These priors are shown in SI Fig. [\[SI:fig:parPrior\]](#SI:fig:parPrior){reference-type="ref" reference="SI:fig:parPrior"}. ## Covariance-based diffusion estimator {#covariance-based-diffusion-estimator .unnumbered} If \(x_k\) (\(k=0,1,\ldots\)) is the measured trajectory of a freely diffusing particle with diffusion constant \(D\), the widely used model for camera-based tracking by Berglund predicts that the measured step lengths \(\Delta_k=x_{k+1}-x_k\) are zero-mean Gaussian variables with covariances given by \[\label{eq:dxcov} \mean{\Delta_k^2}=2D\Delta t(1-2R)+2\varepsilon^2,\quad \mean{\Delta_k\Delta_{k\pm1}}=2D\Delta tR-\varepsilon^2,\] and uncorrelated otherwise. Here, \(0\le R\le 1/4\) is a blur coefficient that depends on how the images are acquired (e.g., \(R=1/6\) for continuous illumination), \(\Delta t\) is the measurement time-step, and \(\varepsilon^2\) is the variance of the localization errors. Substituting sample averages for \(\mean{\Delta_k^2}\) and \(\mean{\Delta_k\Delta_{k+1}}\) and solving for \(D\) yields a covariance-based estimator (CVE) with good performance. If \(\varepsilon^2\) is known or can be estimated independently, the first relation in alone yields a further improved estimate of \(D\). As we argue in Sec. [\[SI:sec:cve\]](#SI:sec:cve){reference-type="ref" reference="SI:sec:cve"}, these estimators apply also for variable localization errors if \(\varepsilon^2\) is replaced by the average \(\mean{\varepsilon^2}\). ## Maximum likelihood and multi-state diffusion {#maximum-likelihood-and-multi-state-diffusion .unnumbered} The Berglund model can also be used directly for maximum likelihood inference, which has the potential advantage that point-wise errors can be modeled. The basic assumption is to model the observed positions \(x_k\) as averages of the true diffusive particle path \(y(t)\) during the camera exposure, plus a Gaussian localization error, i.e., \[\label{eq:blurdef} x_k=\int_{0}^{\dt}y(k\dt+t)f(t)dt+\varepsilon_k\] where \(f(t)\) is the normalized shutter function, \(\varepsilon_k\) is the localization uncertainty (standard deviation) at time \(k\), and \(unit variance. Continuous illumination is described by a constant shutter function,\)f(t)=1/\(. The limit where blur effects are neglected can be described by setting\)f(t)\(to a delta function, which corresponds to instantaneous position measurement. This reduces \Eq{eq:blurdef} to a standard Kalman filter, and leads to\)R=0\(in \Eq{eq:dxcov}. In SI text \ref{SI:sec:diffusionMLE}, we derive a maximum likelihood estimator that learns both\)D\(and\)y(t)\(. In SI Sec.~\ref{SI:sec:HMM}, we extend the model to multi-state diffusion, by letting the diffusion constant switch randomly between different values corresponding to different hidden states in an HMM, and derive a variational EM algorithm for maximum likelihood inference of model parameters, hidden states, and refined estimates of the measured positions. To interpret estimated diffusion constants from simplified models, one may ``derive'' corrected diffusion estimates\)D\^\*\(by equating expressions for the step length variance\[from \Eq{eq:dxcov} with and without those effects present. For the Kalman (\)R=0\() and vbSPT (\)R=\() models, we get \begin{equation}\label{eq:Dcorr} D^*=\frac{D_\text{Kalman}}{1-2R},\quad \text{and } D^*=\frac{D_\text{vbSPT}-\mean{\varepsilon^2}/\dt}{1-2R}, \end{equation} respectively, which is what we use in Fig.~\ref{fig:HMM}e. \subsection*{Software} Matlab open source code for the EM and localization algorithms are available at \href{https://github.com/bmelinden/uncertainSPT}{github.com/bmelinden/uncertainSPT}. \section*{Methods} \input{methods.tex} \subsection*{Acknowledgements} \input{acknowledgements.tex} \subsection*{Author contributions} \input{contributions.tex} \input{osak_arxiv.bbl} \clearpage \onecolumngrid \section{Localization error estimates with different priors}\label{sec:locpriors} In this section we present how different prior functions assigned to the PSF width\]and the background\)b\(influence the accuracy of the computed RMS errors and RMS uncertainties for the synthetic data set considered in this paper. We considered log-normal priors for the background intensity, either centered around the true (simulated) background in each data set, or around the average background intensity in all data sets. For the PSF width we tested two types of priors: (i) a log-normal prior around the expected value for the PSF width; (ii) a skew-normal prior on\)()\(, which assigns very low probability to PSF widths below\[pixels, as well as low probabilities for very wide PSFs. We do not impose any prior on the spot amplitude\)N\(. The probability density function for the log-normal prior, denoted as\)N(,\^2)\(, is defined by \begin{equation} \label{eq:logNormal_prior} f_{ln}(x,\sigma,\mu) = \frac{1}{x\sqrt{2\pi\sigma^2}}e^{-\frac{(ln x-\mu)^2}{2\sigma^2}}, \end{equation} where\]and\[are the mean and standard deviation parameters. The skew-normal distribution has a density function given by \begin{equation} f_{sn}(x,x_0,\omega,\alpha) = \frac{2}{\omega}\phi\left(\frac{x-x_0}{\omega}\right)\Phi\left(\alpha\frac{x-x_0}{\omega}\right), \end{equation} where\]is the probability density function and\[is the cumulative distribution function for the unit normal distribution\)N(0,1)\(, and the parameter\]determines the degree of asymmetry; with\[positive (negative) increasing the weight above (below) the mean value parameter\)x_0\(. For our investigations on the effects of different prior distributions, we used two background priors: \begin{itemize} \item Log-normal ``correct'' prior on\)b\(:\)N(b\_,0.2\^2)\(, where\)b\_\(is either 1 or 2 photons/pixel, which is the background level used for simulating that particular data set. \item Log-normal ``average'' prior on\)b\(:\)N(,0.2\^2)\(, centered around the average background intensity in all data sets, that is, 1.5 photons/pixel. \end{itemize} For the PSF width, we considered the two following priors: \begin{itemize} \item Log-normal prior on\]:\)N(,0.2\^2)\(, since the half-width of the focused PSF width is around 1.5 pixels wide. \item Skew-normal prior on\)()\(:\)f\_sn(x,x_0,,)\(, where\)x_0=(1.5)\(,\)= 5\(and\[. \end{itemize} Error-uncertainty comparisons for various combinations of the above priors are shown in Fig.~\ref{fig:priors_SI}. As seen in Fig.~\ref{fig:priors_SI}a (which reproduces Fig.~\ref{MT:fig:simErr}a in the main text), imposing a flat prior (no prior) leads to estimated uncertainties that substantially underestimate the true errors, especially for points with low photon count. In comparison, all other prior combinations present an average improvement. Overall, the Laplace estimator seems more consistent (RMS uncertainty mostly closer to RMS error), and also less sensitive to the choice of priors. The latter trend is especially clear in the examples using the ``average'' background prior, which is always either a systematic over-or underestimate, which results in substantially larger variability or the CRLB results. \section{Diffusion with time-varying localization errors}\label{sec:blurmodel} Here, we give a detailed derivation of the various estimators and models mentioned in the main text methods section, starting from \Eq{MT:eq:blurdef}, the assumption that localizing a moving object amounts to detecting the time-averaged position with some (independent) localization error, which we will assume to be Gaussian. Parts of these derivations have been given elsewhere, but are restated here in a different form that facilitates a generalization to multi-state models. \subsection{Diffusive camera-based tracking with blur and localization errors} To streamline the presentation, we will use units where\], use the subscript\)t=0,1,2,...\(to denote discrete time dependence, and use the step length and localization variances,\)\_t=2D_t\(and\)v_t=\_t\^2\(, respectively. We allow both of them to be time-dependent, but assume them to be statistically independent and restrict the diffusion constant to be constant throughout each frame. Then, \Eq{MT:eq:blurdef} reads \begin{equation}\label{eq:localization} x_t=\int_0^1f(t') y(t+t')dt'+\sqrt{v_t} \end{equation} where\) variables, and \(y(t)\) is the true trajectory of the particle being localized. The shutter distribution \(f(t)\) is a probability density on \([0,1]\), which describes the image acquisition process (e.g., \(f(t)=1\) for continuous acquisition), but neglects stochastic elements such as fluorophore blinking. It has the distribution function \[F(t)=\int_0^t f(t')dt'.\] We divide \(y(t)\) in two parts, the true positions \(y_t\) at the beginning of each frame, which evolve according to \[\label{eq:ydiffusion} y_{t+1}=y_t+\sqrt{\lambda_t}\eta_t,\] where \(\eta_t\) are again iid N(0,1), and a conditional interpolating process between them, described by Brownian bridges. Thus, for \(0\le t'\le 1\), we write \[\label{eq:interp} y(t+t')=y_t + t'(y_{t+1}-y_t)+\sqrt{\lambda_t}B_t(t'),\] where \(B_t\) are a set of iid standard Brownian bridges. These are Gaussian processes on the interval \[0,1\], defined by \[\label{eq:bridges} B_t(0)=B_t(1)=0,\quad \mean{B_t(t')}=0,\quad \mean{B_t(t')B_t(t'')}=t'(1-t'')\text{, for \(t'\le t''\)},\] and also independent on different intervals, so that \(\mean{B_r(t')B_v(t'')}=0\) if \(r\ne v\). Substituting the interpolation formula in the localization model , we get \[\label{eq:interp2} x_t=y_t(1-\tau)+ y_{t+1}\tau+\sqrt{v_t} +\sqrt{\lambda_t}\int_0^1f(t')B_t(t')dt',\] where we have introduced the shutter average, given by \[\tau=\int_0^1tf(t)dt.\] Using the properties of Brownian bridges, , one can show that the last integral in is a Gaussian random variable with mean zero and variance \[\beta\equiv\Var\left[\int_0^1f(t')B_t(t')dt'\right]=\tau(1-\tau)-R,\] where \(R\) is the blur coefficient of Ref. , given by \[R=\int_0^1F(t)\big(1-F(t)\big)dt.\] By assumption, \(v_t\) and \(B_t\) are statistically independent, and thus one can add up the noise in the measurement model and arrive at \[\label{eq:interp3} x_t=y_t(1-\tau)+ y_{t+1}\tau+\sqrt{v_t+\beta\lambda_t}\zeta_t,\] where \(\zeta_t\) are again iid N(0,1). ## Constant exposure An important class of shutter distribution are those that are constant during some fraction \(t_E\) of the each frame, and then zero, i.e., \[f(t)=\left\{\begin{array}{ll} \frac{1}{t_E},&t\le t_E,\\ 0,&t> t_E,\\ \end{array}\right.\] which leads to \[\tau=\frac{t_E}{2},\quad R=\frac{t_E}{6},\qua \beta=\frac14 t_E(\frac43-t_E)\] We see that \(R\le \frac16\) and \(\tau\le\frac12\), with maxima at continuous exposure (\(t_E=1\)). On the other hand, \(\beta\) has a maximum of \(\frac19\) at \(t_E=\frac23\), and the value of \(\frac{1}{12}\) at continuous exposure can only be further lowered when \(t_E<\frac13\). It is unclear if this is significant, since with non-constant exposure there is some freedom in how the shutter distribution is defined, and one could also place it symmetrically in the interval and get \(\tau=0.5\) for all exposure times. We defer further investigations of this issue to future work. ## Covariance relations {#sec:cve} The covariance matrix for the steps \(\Delta_t=x_{t+1}-x_t\) can be found from Eqs. ([\[eq:ydiffusion\]](#eq:ydiffusion){reference-type="ref" reference="eq:ydiffusion"},[\[eq:interp3\]](#eq:interp3){reference-type="ref" reference="eq:interp3"}). With some manipulations, we get \[\begin{aligned} \mean{\Delta_t^2}= &(1-\tau)\lambda_t+\tau\lambda_{t+1} -(\lambda_{t+1}+\lambda_t)R+v_t+v_{t+1},\\ \mean{\Delta_t\Delta_{t+1}}= &\lambda_{t+1}R-v_{t+1},\\ \mean{\Delta_t\Delta_{t+t'}}= &0\text{, if \(|t'|>1\)}, \end{aligned}\] where the expectations \(\mean{\cdot}\) are understood to be over the noise distributions only. If we further assume simple diffusion, \(\lambda_t=2D\dt=const.\), and average over time as well, we recover covariance relations of the same form as , \[\mean{\Delta_t^2}=2D\dt(1-2R)+2\mean{v_t},\quad \mean{\Delta_t\Delta_{t\pm 1}}=2D\dt R-\mean{v_t},\] where the averages are now over time as well, and \(\varepsilon^2\) is identified as the time-average \(\mean{v_t}\). Thus, the covariance-based estimators of Ref.  should apply also to non-constant localization errors. # Maximum likelihood estimator {#sec:diffusionMLE} For the case of a single trajectory of simple diffusion in one dimension, the likelihood for the diffusion constant follows from Eqs. [\[eq:ydiffusion\]](#eq:ydiffusion){reference-type="ref" reference="eq:ydiffusion"} and [\[eq:interp3\]](#eq:interp3){reference-type="ref" reference="eq:interp3"}, \[\begin{aligned} p(x|\lambda) &=& \int dy p(x|y,\lambda)p(y|\lambda),\label{eq:lamlik}\\ p(y|\lambda) &=&\prod_{t=1}^T \big(2\pi\lambda\big)^{-\frac12} \exp\Big[-\frac{(y_{t+1}-y_t)^2}{2\lambda}\Big],\\ p(x|y,\lambda)&=& \prod_{t=1}^T \big(2\pi(v_t+\beta\lambda)\big)^{-\frac12} \exp\Big[-\frac{1}{2}(v_t+\beta\lambda)^{-1} \big(x_t-(1-\tau)y_t-\tau y_{t+1}\big)^2\Big], \end{aligned}\] where we neglected to supply a starting density for \(y_1\), since the problem is translation invariant. The integral over the hidden path in is a multivariate Gaussian and can be solved exactly in several ways. Defining \[\vec{y}=[y_1,\;y_2,\ldots,y_{T+1}]^\dagger,\quad \vec{x}=[x_1,\;x_2,\ldots,x_{T}]^\dagger,\] where \(^\dagger\) denotes matrix transpose, we get \[\begin{gathered} \label{eq:pxlam1} p(x|\lambda)=\int d\vec{y} \exp\bigg[ -T\ln(2\pi)-\frac{T}{2}\ln(\lambda) -\frac12\sum_{t=1}^T\ln(v_t+\beta\lambda) -\frac12\Big(\vec{y}^\dagger\Lambda\vec{y}-2\vec{y}^\dagger W\vec{x} +\vec{x}^\dagger V\vec{x} \Big)\bigg]\\ =\exp\bigg[ -T\ln(2\pi)-\frac{T}{2}\ln(\lambda) -\frac12\sum_{t=1}^T\ln(v_t+\beta\lambda)\bigg]\\ \times\int d\vec{y}\exp\bigg[ -\frac12\Big( (\vec{y}-\Lambda^{-1}W\vec{x})^\dagger\Lambda(\vec{y}-\Lambda^{-1}W\vec{x}) +\vec{x}^\dagger (V-W^\dagger\Lambda^{-1} W)\vec{x} \Big)\bigg], \end{gathered}\] where \(\Lambda,W,V\) are matrices whose are found by comparing terms. This is a multivariate Gaussian in \(\vec{y}\), with mean value \(\mu=\Lambda^{-1}W\vec{x}\) and covariance matrix \(\Sigma=\Lambda^{-1}\), and the marginalized likelihood is therefore given by \[\label{eq:pxlam2} p(x|\lambda)= \exp\bigg[ -\frac{T-1}{2}\ln(2\pi)-\frac{T}{2}\ln(\lambda) -\frac12\sum_{t=1}^T\ln(v_t+\beta\lambda) -\frac12\vec{x}^\dagger (V-W^\dagger\Lambda^{-1}W)\vec{x} -\frac12\ln|\Lambda|\bigg].\] By comparing terms, we see that \(\Lambda\) is symmetric tridiagonal and positive definite, \(V\) is diagonal, and \(W\) has only non-zero elements on the diagonal and first upper diagonal, and so it is possible to compute the above matrix expression in linear time. We minimize \(\ln p(x|\lambda)\) by standard optimization routines in Matlab. Repeating the analysis of Fig. [\[MT:fig:Dest\]](#MT:fig:Dest){reference-type="ref" reference="MT:fig:Dest"} in the main text for this MLE estimator, see Fig. [\[fig:DestMLE\]](#fig:DestMLE){reference-type="ref" reference="fig:DestMLE"}, we find almost no improvement over the CVE that uses average estimated uncertainties. However, the fact that the MLE estimator is limited to return positive estimated diffusion constants, and handles missing data points (by assigning them infinite or very large uncertainties), may be useful in applications. # Variational EM algorithm for diffusive HMM {#sec:HMM} Here, we extend the above diffusion model to include multiple diffusion constants. We start by writing down a diffusive HMM, which includes both the hidden path of the above diffusion model, but also a set of hidden states with different diffusion constants, that evolve as a discrete Markov process. Similar models (that however did not include explicit motion blur effects), have previously been solved by stochastic EM algorithms. Here, we instead describe a deterministic variational approach inspired by algorithms for factorial HMMs. In the rest of this section, we proceed as follows: We start by specify the diffusive HMM model for a single 1-dimensional trajectory. We then outline the variational EM approach, derive high level update equations, and describe the procedure for re-estimating localized positions. However, do not give a detailed derivation of all steps in the algorithm, as large parts of it closely resembles previously published derivations of variational algorithms for HMMs. ## Model In addition to the measured (\(\vec{x}\)) and true (\(\vec{y}\)) positions, we include a hidden state trajectory \(\vec{s}=[s_1,s_2,\ldots,s_T]\), such that \(s_t\) determines the diffusion constant on the interval \([t,t+1]\). The hidden states are numbered from 1 to \(N\), and evolve according to a Markov process with transition matrix \(A\) and initial state probability \(\vpp\). For a single 1-dimensional trajectory, this leads to a complete data likelihood of the form \[p(\] with factors \[\begin{aligned} p(\s|A,\vpp)&=& \prod_{m=1}^N\pp_m^{\delta_{m,s_1}} \prod_{t=2}^T\prod_{i,j=1}^N A_{ij}^{\delta_{is_t}\delta_{js_{t+1}}},\\ p(\y|\s,\lambda)&=& \prod_{t=1}^T\prod_{j=1}^N (2\pi \lambda_j)^{-\frac12\delta_{js_t}} \exp\left[-\delta_{js_t}\frac{(y_{t+1}-y_t)^2}{2\lambda_j}\right],\\ p( \prod_{t=1}^T\prod_{j=1}^N \big(2\pi(v_t+\beta\lambda_j)\big)^{-\frac12\delta_{js_t}} \exp\left[-\delta_{js_t} \frac{(x_t-(t-\tau)y_t-\tau y_{t+1})^2}{2(v_t+\beta\lambda_j)} \right]. \end{aligned}\] ## Variational EM approach We would like to perform maximum-likelihood inference of the model parameters, which means maximizing the likelihood with latent variables \(\s,\y\) integrated out, \[L(A,\vpp,\lambda)=\int d\y\sum_{\s} p(\] Since this problem is intractable, we make a variational approximation, meaning we approximate \(\ln L\) with a lower bound \[\label{eq:lowerbound} \ln L =\ln \int d\y\sum_{\s} p( \ge \int d\y\sum_{\s} q(\s)q(\y)\ln \frac{p(\] where the inequality follows from Jensen's inequality. Here, \(q(\s)\), \(q(\y)\) are arbitrary variational distributions that need to be optimized together with the model parameters to achieve the tightest lower bound, and can be used for approximate inference about the latent variables. In particular, it turns out that the optimal variational distributions approximate the posterior distribution of \(\y,\s\) in the sense of minimizing a Kullback-Leibler divergence. We use gradient descent for the optimization, i.e., we iteratively optimize each variational distribution and the parameters with the others fixed, which leads to an EM-type algorithm. To optimize \(F\) w.r.t. the variational distributions, we set the functional derivatives of \(F\) to zero and use Lagrange multipliers to enforce normalization. After some work, one arrives at \[\begin{aligned} \ln q(\s)&=&-\ln Z_s +\mean{\ln p( +\mean{\ln p(\y|\s,\lambda}_{q(\y)} +\ln p(\s|A,\vpp),\label{eq:lnqs}\\ \ln q(\y)&=&-\ln Z_y +\mean{\ln p( +\mean{\ln p(\y|\s,\lambda}_{q(\s)} +\underbrace{\mean{\ln p(\s|A,\vpp)}_{q(\s)}}_\text{independent of \(\y\)}, \end{aligned}\] where \(Z_{s,y}\) are normalization constants originating from the Lagrange multipliers, and \(\mean{\cdot}_{f}\) denotes an expectation value computed with respect to the distribution \(f\). As it turns out, these equations are individually tractable, and results in \(q(\y)\) being a multivariate Gaussian, and \(q(\s)\) adopting the standard HMM form amenable to efficient forward-backward iterations. For the initial state and transition probability parameters, the only dependence in \(F\) comes from \(p(\s|A,\lambda)\), and arrives at the classical Baum-Welch reestimation formulas. However, optimizing \(F\) w.r.t. step length variances does not lead to a tractable update equation, and we are instead forced to optimize the \(\lambda_j\)-dependent parts of \(F\) numerically, i.e., \[\lambda_j=\mathrm{argmax}_{\lambda_j} \mean{\ln p( \mean{\ln p(\y|\s,\lambda)}_{q(\y)q(\s)},\] ## The lower bound The lower bound can be computed as well, and gets a particularly simple form just after the update of \(q(\s)\). Substituting the update equation into the expression for \(F\), , we get \[\begin{gathered} F=\int d\y\sum_{\s} q(\s)q(\y)\ln\Big[ \ln p( +\ln Z_s-\mean{\ln p(-\ln q(\y)\Big] =\ln Z_s-\mean{\ln q(\y)}_{q(\y)}. \end{gathered}\] Here, \(\ln Z_s\) is the normalization constant of \(q(\s)\) that can be computed as part of the forward-backward iteration, and since (as we noted above) \(q(\y)\) is a multivariate Gaussian with dimension \(T+1\), we get \[-\mean{\ln q(\y)}_{q(\y)}= \frac{d}{2}(T+1)\big(1+\ln(2\pi)\big)+\frac12\ln|\Sigma|,\] where \(\Sigma\) is the covariance matrix of \(q(\y)\). The inverse of \(\Sigma\) is analogous to the matrix \(\Lambda\) appearing in the single-state diffusion estimator, Eqs. ([\[eq:pxlam1\]](#eq:pxlam1){reference-type="ref" reference="eq:pxlam1"},[\[eq:pxlam2\]](#eq:pxlam2){reference-type="ref" reference="eq:pxlam2"}), and in particular \(\Sigma^{-1}\) is also symmetric, tridiagonal, and positive definite, and thus the determinant \(\ln|\Sigma|=-\ln|\Sigma^{-1}|\) can be robustly computed in linear time. Since the EM algorithm approximates the parameter likelihood, the lower bound cannot be used for model selection as in the case of variational maximum evidence calculations. However, it is still useful for numerical convergence control, and possibly for model selection together with some complexity penalty such as the Bayesian or Akaike information criterion. ## Refined positions {#sec:positions} An additional use for the HMM analysis is to use it to refine the localizations. Since the HMM pools information about many spots, it is in principle possible to beat the Cramér-Rao lower bound for single image localizations in this way. To set up the refinment problem, we refer back to and the different contributions to the observed position \(x_t=z_t+\sqrt{v_t} \begin{equation}\label{eq:zrefine1} z_t=y_t(1-\tau)+ y_{t+1}\tau+\sqrt{\lambda_{s_t}}\int_0^1f(t')B_t(t')dt' \end{equation} is the motion-averaged position that the localization algorithm tries to estimate (according to this model). To refine the localization, we need to compute the posterior density of\)z_t\(, i.e.,\)p(z_t\| We start by recalling that the Brownian bridge integral in is Gaussian with mean zero and variance \(\beta\lambda_{s_t}\), and using the compact notation \(\overline{y_t}=(1-\tau)y_t+\tau y_{t+1}\) and \(\theta\) to denote all the model parameters, we therefore have \[p(z_t|\y,\s,\theta)=p(z_t|\overline y_t,s_t,\theta) =N(\overline y_t,\beta\lambda_{s_t}),\] where \(N(a,b)\) denotes a Gaussian density with mean \(a\) and variance \(b\). Furthermore, the localization uncertainty is also assumed Gaussian and independent of the underlying kinetics: \[p(x_t|\z,\y,\s,\theta)=p(x_t|z_t)=N(z_t,v_t).\] Applying Bayes theorem to these relations, we get \[p(z_t|x_t,\y,\s,\theta) = \frac{p(x_t|z_t)p(z_t|\y,\s,\theta)}{p(x_t|\y,\s,\theta))} =\frac{N(z_t,v_t)N(\overline y_t,\beta\lambda_{s_t})}{ N(\overline y_t,\beta\lambda_{s_t}+v_t)} = N\Big(\frac{v_t\overline y_t+\beta\lambda_{s_t}x_t}{v_t+\beta\lambda_{s_t}}, \frac{\beta\lambda_{s_t}v_t}{v_t+\beta\lambda_{s_t}} \Big).\] The predictive distribution is finally given by marginalizing \(p(z_t|x_t,\y,\s,\theta)\) over the posterior for \(\y,\s\). Using the variational distribution, this means \[p(z_t| \approx\mean{ N\Big(\frac{v_t\overline y_t+\beta\lambda_{s_t}x_t}{v_t+\beta\lambda_{s_t}}, \frac{\beta\lambda_{s_t}v_t}{v_t+\beta\lambda_{s_t}}\Big)}_{q(\y)q(\s)}.\] In particular, the posterior mean of \(z_t\) is then given by \[\mean{z_t| \frac{v_t\overline y_t+\beta\lambda_{s_t}x_t}{v_t+\beta\lambda_{s_t}} }_{q(\y)q(\s)} =\mean{ \frac{(1-\tau)\mu_t+\tau\mu_{t+1} +\frac{\beta\lambda_{s_t}}{v_t}x_t}{1+\frac{\beta\lambda_{s_t}}{v_t}} }_{q(\s)},\] which we will use as our estimator for refining the localizations. Here, \(\mu_t=\mean{y_t}_{q(\y)}\) is the variational mean value. The variational average over hidden states is done numerically.

{'timestamp': '2016-07-19T02:01:52', 'yymm': '1607', 'arxiv_id': '1607.04675', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04675'}

# Introduction Structure-from-Motion targets at recovering 3D structure and camera motion from monocular 2D feature tracks. Conventional SFM primarily concerns with the 3D reconstruction of a single rigidly moving object seen by a static camera, or a static and rigid scene observed by a moving camera--in both cases there are only one relative rigid motion involved. SFM has been extended to the areas of *multi-body* SFM, as well as *single body* non-rigid SFM (see Table.[\[tab:summary\]](#tab:summary){reference-type="ref" reference="tab:summary"} for a classification of different SFM problems). Since, NRSFM is central to many computer vision applications such as motion capture, activity recognition, human-computer interaction, and dynamic scene reconstruction, etc. Therefore, it has been extensively researched (e.g. in ) in the past. So far, all existing methods for NRSFM have implicitly assumed that there is only one deformable shape or object in the view. However, the real world scenarios can be much more complicated involving multiple, independently deforming objects in the scene. Multiple nonrigid objects are commonly encountered in our daily lives, for example, in motion capture, multiple persons perform different activities with possible interactions (see Fig. [\[fig:multiple_nonrigid\]](#fig:multiple_nonrigid){reference-type="ref" reference="fig:multiple_nonrigid"} for an example); in human-computer interaction, different users may conduct different gesture commands; and in traffic scene, multiple vehicles and walking pedestrians create **multi-body non-rigid deformations**. To handle such multiple non-rigid deformations in 3D reconstruction, a natural idea would be to simply treat the multiple non-rigid deformations as a single (though more complex) non-rigid deformation (with higher order or higher rank), and then apply any state-of-the-art non-rigid structure-from-motion methods such as. However, by this idea, the inherent structure in the problem has not been exploited, which may hinder the success of 3D reconstruction. Even if the method succeeds in obtaining 3D reconstruction, it cannot tell meaningful segmentation of the multiple non-rigid objects. Another choice would be conducting the tasks of non-rigid motion segmentation and non-rigid 3D reconstruction successively. However, in this way, the solution of each sub-task does not benefit from the solution of the other sub-task. For example, non-rigid motion segmentation provides critical information to correct non-rigid 3D reconstruction while non-rigid 3D reconstruction constrains the corresponding multi-body motion segmentation. Therefore, we would like to emphasize that since non-rigid deformation originally occurs in 3D space, it's more intuitive to perform non-rigid motion segmentation and reconstruction simultaneously in 3D space rather than solving this problem using two step process. To demonstrate the advantage of the concurrent procedure as shown in Fig. [\[fig:flow_diagram\]](#fig:flow_diagram){reference-type="ref" reference="fig:flow_diagram"} and Table [\[tab:summary\]](#tab:summary){reference-type="ref" reference="tab:summary"}, this paper introduces an approach to perform non-rigid 3D reconstruction and its motion segmentation simultaneously. Specifically, we represent the multiple non-rigid motion as union of 3D trajectory subspaces [^1]. By using the self-expressiveness model in representing multiple linear/affine subspace, where each 3D trajectory can be expressed with other trajectories in the same subspace only, enables us in compact representation of trajectories. In this way, we are able to exploit the inherent grouping structure in *3D trajectory space*. For dense non-rigid reconstruction, we could further enforce the spatial coherence constraint. By contrast to existing methods, this endows us the following benefits: (i) A compact representation for component non-rigid deformation in 3D trajectory space; (ii) Joint multiple deformable objects motion segmentation and 3D non-rigid structure reconstruction; (iii) Improved spatial regularity in 3D non-rigid dense reconstruction (in contrast, a hard segmentation of the 2D tracks may result in discontinuity at the segmentation boundary). The readers are invited to have a preview of results obtained by our method, as shown in Fig. [\[fig:multiple_nonrigid\]](#fig:multiple_nonrigid){reference-type="ref" reference="fig:multiple_nonrigid"}, where we show a daily life motion capture of two person performing dancing and yoga individually as in Fig. [\[fig:dance_yoga_2D_1\]](#fig:dance_yoga_2D_1){reference-type="ref" reference="fig:dance_yoga_2D_1"} and Fig. [\[fig:dance_yoga_2D_2\]](#fig:dance_yoga_2D_2){reference-type="ref" reference="fig:dance_yoga_2D_2"}. Similarly, in Fig. [\[fig:result_3object\]](#fig:result_3object){reference-type="ref" reference="fig:result_3object"} three different person are performing stretch, dance and yoga. By applying our proposed method, we are able to achieve both non-rigid 3D reconstruction and non-rigid motion segmentation as shown in Fig. [\[fig:dance_yoga_3D_1\]](#fig:dance_yoga_3D_1){reference-type="ref" reference="fig:dance_yoga_3D_1"}, Fig. [\[fig:dance_yoga_3D_2\]](#fig:dance_yoga_3D_2){reference-type="ref" reference="fig:dance_yoga_3D_2"} and Fig. [\[fig:multi_object3\]](#fig:multi_object3){reference-type="ref" reference="fig:multi_object3"}. # Related Work Ever since the seminal work by Bregler modeling a non-rigid shape as lying in a "*shape space*" (a linear combination of basis shapes), considerable progress has been made in the area of non-rigid 3D reconstruction. In 2004, Xiao showed the inherent ambiguity in modeling non-rigid shape and proposed a remedy of "basis constraints" to derive a closed-form solution. In 2008, Akhter presented a dual approach by modeling 3D trajectories, "*trajectory space*". In 2009, Akhter proved that even there is an ambiguity in shape bases or trajectory bases, non-rigid shapes can still be solved uniquely without any ambiguity. Based on this, in 2012, Dai proposed a "prior-free" method to recover camera motion and 3D non-rigid deformation by exploiting the low rank constraint only. Besides shape basis model and trajectory basis model, the shape-trajectory approach combines two models and formulates the problems as revealing trajectory of the shape basis coefficients. Besides linear combination model, Lee proposed a Procrustean Normal Distribution (PND) model, where 3D shapes are aligned and fit into a normal distribution. Simon proposed to exploit the Kronecker pattern in the shape-trajectory (spatial-temporal) priors. Zhu and Lucey applied the convolutional sparse coding technique to NRSFM using point trajectories. However, the method requires to learn an over-complete basis of 3D trajectories, prior to performing 3D reconstruction. Despite of the above success, NRSFM is still far behind its rigid counterpart. This is mainly due to difficulty in modeling real world non-rigid deformation. Real world non-rigid reconstruction generally requires the ability to handle long-term, complex and dense non-rigid shape variations. Such complex and dense non-rigid motion not only increases the computational complexity but also adds difficulty in modeling various kinds of different motions. Zhu proposed to represent long-term complex non-rigid motion as lying in a union of shape sub-spaces rather than sum of sub-spaces. Cho represented complex shape variations probabilistically by a mixture of primitive shape variations. By contrast to the above methods dealing with sparse NRSFM, dense NRSFM methods such as aim at achieving 3D reconstruction for each pixel in a video sequence, where spatial constraint has been widely exploited to regularize the problem. Garg presented a variational formulation to dense non-rigid reconstruction by exploiting the spatial smoothness in 3D shapes, which in principle deals with single non-rigid deformation in contrast to our multiple non-rigid deformations. Fragkiadaki solved the problem in sequel, namely, video segmentation by multi-scale trajectory clustering, 2D trajectory completion, rotation estimation and 3D reconstruction. Recently, Yu bridges template based method and feature track based method by proposing a dense template based direct approach to deformable shape reconstruction from monocular sequences. Russell proposed to simultaneously segment a complex dynamic scene containing mixture of multiple objects into constituent objects and reconstruct a 3D model of the scene by formulating the problem as hierarchical graph-cut based segmentation, where the entire scene is decomposed into background and foreground objects and the complex motion of non-rigid or articulated objects are modeled as a set of overlapping rigid parts. Our proposed method differs from this method in the following aspects: 1) We provide a compact representation to multiple non-rigid deformation problem; 2) We propose an efficient and elegant optimization based on ADMM; 3) Our method could deal with both sparse and dense scenarios elegantly. # Formulation We seek to reconstruct the 3D trajectories such that they satisfy the union of affine subspace constraint (the 3D trajectories lie in a union of affine subspaces) and non-rigid shape constraints (low rank and spatial coherent). Let us consider a monocular camera observing multiple non-rigid objects. In this paper, we use the \(orthographic\) \(camera\) model and eliminate the translation component in camera motion as suggested in. The image measurement \(\m w_{ij} = [u_{ij}, v_{ij}]^T\) and 3D point \(\m S_{ij}\) on the non-rigid shape are related by the camera motion \(\m R_i\) as: \(\mathbf{w}_{ij} = \m R_i \m{S}_{ij}\), where \(\m R_i \in \mathbb{R}^{2\times 3}\) denotes the first two rows of the \(i\)-th camera rotation. Under this representation, stacking all the \(\m F\) frames of measurements and all the \(\m P\) points in a matrix form will give us: \[\label{eq:motion_shape} \m W = \m R \m S,\] where \(\m R = \mathrm{blkdiag}(\m R_1,\cdots,\m R_F) \in \mathbb{R}^{2\m F\times 3\m F}\) denotes the camera motion matrix. NRSFM aims at recovering the *camera motion* \(\m R\) and 3D non-rigid reconstruction \(\m S \in \mathbb{R}^{3\m F\times \m P}\) from the 2D *measurement matrix* \(\m W \in \mathbb{R}^{2\m F \times \m P}\) such that \(\m W=\m R\m S\). ## Representing multi-body non-rigid structure as a union of affine subspace We propose that multiple non-rigid structures that corresponds to distinct motion lies in a union of affine subspace. Here, the underlying assumption is that the trajectories belong to different non-rigid objects spans a distinct affine subspace. Figure [\[fig:A_algo\]](#fig:A_algo){reference-type="ref" reference="fig:A_algo"} clearly validates such assumption as there are only connection within clusters and no connection between clusters or block diagonal structure. Now, consider each trajectory \(\m S_i\) that corresponds to a \(3 \m F\) dimensional vector formed by stacking the 3D feature tracks of point \(i\) across all frames. \[\m S_i = [\m S_{1i}^T, \m S_{2i}^T,..., \m S_{\m Fi}^T]^T \in \mathbb{R}^{3\m F},\] where \(\m S_{fi}\) \(\in\) \(\mathbb{R}^3\). Under the orthographic camera model, feature trajectory associated with non-rigid motion lie in an affine subspace of dimension \(\mathbb{R}^{3 \m F}\). After taking \(\m P\) such \(3\m F\) dimensional trajectory and stacking into the column of a matrix we form \(\m S\) matrix \(\in\) \(\mathbb{R}^{3\m F \times \m P}\). Mathematically, it implies that \(\m S\) is a real valued matrix whose columns are drawn from a union of \(n\) subspace of \(\mathbb{R}^{3\m F}\) of dimension less than min(\(3\m F\), \(\m P\)). Therefore, each trajectory in a union of affine subspace could be faithfully reconstructed by a combination of other trajectories in the same subspace. This leads to \(self\)-\(expressiveness\) of 3D trajectories. Concretely, \[\m S_i = \m S \m C_i, \m C_{ii} = 0.\] Here, \(\m C_i\) is a \(\m P \times 1\) coefficient vector and \(\m C_{ii}\) = \(0\) takes care of the trivial solution \(\m S_i = \m S_i\). Stacking all such coefficient vectors, we form a matrix \(\m C \in \mathbb{R}^{\m P \times \m P}\) that captures the similarity between different trajectories. Using the fact that any trajectory of \(\m S\) in an affine subspace of dimension \(3\m F\) can be written as affine combination of \(3 \m F + 1\) other points from \(\m S\) and to cluster trajectories that lies near to union of affine subspaces, we arrive at the following equation. \[\label{eq:SSC_constraints} \m S = \m S \m C, \m 1^T \m C = 1^T, \mathrm{diag}(\m C) = 0.\] Figure [\[fig:Union of Subspace representation\]](#fig:Union of Subspace representation){reference-type="ref" reference="fig:Union of Subspace representation"} shows the affinity matrix \(\m A = |\m C| + |\m C^T|\) obtained for two objects that undergo non-rigid deformation. The obtained sparse solution clearly shows that multi-body non-rigid structure can be represented as union of affine subspace. ## Representing multiple non-rigid deformations in case of sparse feature tracks To solve the problem of multi-body NRSFM in case of sparse feature tracks, we propose the following optimization framework for simultaneous reconstruction and segmentation of objects that are undergoing non-rigid deformation, \[\begin{aligned} & \displaystyle \underset{\m C, \m S, \m S^{\sharp}} {\text{minimize}} \frac{1}{2}\|\m W-\m R \m S \|_F^2 + \lambda_1\|\m C\|_1 + \lambda_2\|\m S^{\sharp}\|_{*} \\ & \text{{subject to: }} \\ & \displaystyle \m S^{\sharp} = g(\m S), \displaystyle \m S = \m S \m C, \displaystyle \m 1^{T} \m C = \m 1^{T}, \displaystyle \mathrm{diag}(\m C) = \m 0. \end{aligned} \label{eq:sparse_version}\] The first term in the above optimization is meant for penalizing reprojection error under \(orthographic\) projection. Under single-body NRSFM configuration, 3D shape \(\m S\) can be well characterized as lying in a single low dimensional linear subspace. However, when there are multiple non-rigid objects, each non-rigid object could be characterized as lying in an affine subspace. One can argue that affine subspace of dimension \(n\) can be considered as a subset of \((n+1)\) dimension that includes origin. Nonetheless, such representation may result in ambiguous solution while clustering different subspace. The fact that the 3D trajectories lie in a union of affine subspaces as argued previously we put the Eq. [\[eq:SSC_constraints\]](#eq:SSC_constraints){reference-type="ref" reference="eq:SSC_constraints"} as our optimization constraints. In addition to this, to reveal the intrinsic structure of multi-body non-rigid structure-from-motion (NRSFM), we seek for sparsest solution of \(\m C\) (). So, the second term in **Eq.** [\[eq:sparse_version\]](#eq:sparse_version){reference-type="eqref" reference="eq:sparse_version"} enforces \(l_1\) norm minimization of \(\m C\) matrix. Lastly, we enforce a global shape constraint for compact representation of multi-body non-rigid objects by penalizing the rank of entire non-rigid shape. Similar to and, we penalize the nuclear norm of reshuffled shape matrix \(\m S^{\sharp} \in \mathbb{R}^{\m F\times 3 \m P}\), this is because nuclear norm is known as convex envelope of the rank function. Here, \(\m g(\m S)\) denotes mapping from \(\m S\) \(\in\) \(\mathbb{R}^{3\m F\times \m P}\) to \(\m S^\sharp\) \(\in\) \(\mathbb{R}^{\m F\times 3 \m P}\). ## Representing multi-body non-rigid deformations in case of dense feature tracks When per pixel feature tracks are available, we can enforce spatial regularization (\(Markovian\) \(assumption\)) i.e there will be a high correlation between neighbouring features. To exploit this property, the spatial smoothness can be used as a regularization term to further constrain the non-rigid reconstruction. Garg proposed to use the total variation of the 3D shape \(\|\m S\|_{TV}\). By contrast, we propose to enforce the spatial smoothness constraint on the coefficient matrix \(\m C\) directly by using the \(l_1\) norm, \[\label{eq:C_TV} \sum_{(i,j) \in \mathcal{N}} \|\m C_i-\m C_j\|_1,\] , the total variation of \(\m C\). This definition gives us the benefit in solving the problem as proved later. In essence, the total variations of \(\m C\) and \(\m S\) are correlated. However, it is desirable that \(\m C\) matrix must cater the self-expressiveness of the non-rigid shape deformation as compact as possible. So, we incorporate spatial smoothness constraint on coefficient matrix than on shape matrix \(\m S\). By introducing an appropriately defined matrix \(\m D\) encoding the neighbouring relation, Eq. [\[eq:C_TV\]](#eq:C_TV){reference-type="eqref" reference="eq:C_TV"} can be equivalently expressed as: \[\|\m C \m D \|_1 = \sum_{(i,j) \in \mathcal{N}} \|\m C_i-\m C_j\|_1. \label{eq:C_D}\] In Fig. [\[fig:DMatrix_exp\]](#fig:DMatrix_exp){reference-type="ref" reference="fig:DMatrix_exp"}, we illustrate the process of how to obtain the matrix \(\m D\). Adding the spatial constraint to the optimization equation [\[eq:sparse_version\]](#eq:sparse_version){reference-type="eqref" reference="eq:sparse_version"}, that facilitates this neighboring constraint, we procure the following optimization for dense tracks \[\begin{aligned} & \displaystyle \underset{\m C, \m S, \m S^{\sharp} }{\text{minimize}}~ \frac{1}{2}\|\m W-\m R \m S \|_F^2 + \lambda_1\|\m C\|_1 + \lambda_2 \| \m C\m D \|_1 + \lambda_3\|\m S^{\sharp}\|_{*} \\ & \text{{subject to:}} \\ & \displaystyle \m S^{\sharp} = g(\m S), \displaystyle \m S = \m S \m C, \displaystyle \m 1^{T} \m C = \m 1^{T}, \displaystyle \mathrm{diag}(\m C) = \m 0. \end{aligned} \label{eq:Optimization_function}\] # Solution Due to the bilinear term \(\m S = \m S \m C\), the overall optimization of Eq.-[\[eq:Optimization_function\]](#eq:Optimization_function){reference-type="eqref" reference="eq:Optimization_function"} is non-convex. We solve it via the ADMM, which has a proven effectiveness for many non-convex problems and is widely used in computer vision. The ADMM works by decomposing the original optimization problem into several sub-problems, where each sub-problem can be solved efficiently. To this end, we seek to decompose Eq.-[\[eq:Optimization_function\]](#eq:Optimization_function){reference-type="eqref" reference="eq:Optimization_function"} into several sub-problems. First note that the two \(l_1\) terms \(\|\m C\|_1\) and \(\|\m C \m D\|_1\) can be put together as \(\|\m C [\m I ~\m D]\|_1\). Without loss of generality, we still denote the new term as \(\m C \m D\), the only difference is, the new dimension of \(\m D\) will be \(\m P \times 5\m P\), thus the cost function becomes: \(\frac{1}{2} \|\m W-\m R \m S \|_F^2 + \lambda_1\|\m C \m D \|_1 + \lambda_2\|\m S^{\sharp}\|_{*}\). To further decouple the constraint, we introduce an auxiliary variable \(\m E = \m C \m D\). With these operations, the optimization problem Eq.-[\[eq:Optimization_function\]](#eq:Optimization_function){reference-type="eqref" reference="eq:Optimization_function"} can be reformulated as: \[\begin{aligned} & \displaystyle \underset{\m E, \m S, \m S^{\sharp}, \m C} {\text{minimize}} ~\frac{1}{2}\|\m W-\m R \m S \|_F^2 + \lambda_1\|\m E \|_1 + \lambda_2\|\m S^{\sharp}\|_{*} \\ & \text{{subject to:}} \\ & \displaystyle \m S^{\sharp} = g(\m S), \displaystyle \m S = \m S \m C, \displaystyle \m C \m D = \m E, \displaystyle \m 1^{T} \m C = \m 1^{T}, \displaystyle \mathrm{diag}(\m C) = \m 0. \end{aligned} \label{eq:main_optimization}\] The Augmented Lagrangian formulation for Eq.-[\[eq:main_optimization\]](#eq:main_optimization){reference-type="eqref" reference="eq:main_optimization"} is: \[\begin{aligned} \mathcal{L}(\m S, \m S^{\sharp}, \m C, \m E, \{\m Y_i\}_{i=1}^4) = & \frac{1}{2}\|\m W-\m R \m S \|_F^2 + \lambda_1\|\m E \|_1 + \lambda_2\|\m S^{\sharp}\|_{*} + <\m Y_1, \m S^{\sharp}-g(\m S)> \nonumber \\ & + \frac{\beta}{2}\|\m S^{\sharp}-g(\m S) \|_F^2 + <\m Y_2, \m S-\m S \m C> + \frac{\beta}{2}\|\m S-\m S \m C \|_F^2 \\ \nonumber & + <\m Y_3, \m C \m D-\m E> + \frac{\beta}{2}\|\m C \m D-\m E \|_F^2 \nonumber \\ \nonumber & + <\m Y_4, 1^T \m C-\m 1^T > + \frac{\beta}{2}\|\m 1^T \m C-\m 1^T \|_F^2. \label{eq:Lagrangian} \end{aligned}\] where \(\{\m Y_i\}_{i=1}^4\) are the matrices of Lagrange multipliers corresponding to the four equality constraints, and \(\beta\) is the penalty parameter. We do not need to introduce a Lagrange multiplier for the diagonal constraint of \(\mathrm{diag}(\m C) = \m 0\) as we will enforce this constraint exactly in the solution of \(\m C\). The ADMM iteratively updates the individual variable so as to minimize \(\mathcal{L}\) while the other variables are fixed. In the following, we derive the update for each of the variables. ### The solution of \(\m S\): \[\begin{array}{l} \m {S} = \arg\min_{\m S} \frac{1}{2}\|\m W-\m R \m S \|_F^2 + <\m Y_1, \m S^{\sharp}-g(\m S)> + \frac{\beta}{2}\|\m S^{\sharp}-g(\m S) \|_F^2 \\ \qquad \qquad \qquad + <\m Y_2, \m S-\m S \m C> + \frac{\beta}{2}\|\m S-\m S \m C \|_F^2. \end{array}\] The sub-problem for \(\m S\) reaches a least squares problem. The closed-form solution of \(\m S\) can be derived as: \[\label{eq:Update_S} \frac{1}{\beta}(\m R^T \m R + \beta \m I) \m S + \m S(\m I-\m C)(\m I-\m C^T) = \frac{1}{\beta}\m R^T \m W + (g^{-1}(\m S^{\sharp}) + \frac{g^{-1}(\m Y_1)}{\beta }-\frac{\m Y_2}{\beta}(\m I-\m C^T)),\] which is a Sylvester equation. ### The solution of \(\m S^{\sharp}\): \[\label{eq:update_S_Sharp} \m {S}^{\sharp} = \arg\min_{\m {S}^{\sharp}} ~ \lambda_2\|\m S^{\sharp}\|_{*} + <\m Y_1, \m S^{\sharp}-g(\m S)> + \frac{\beta}{2}\|\m S^{\sharp}-g(\m S) \|_F^2\] A close-form solution exists for this sub-problem. Let's define the soft-thresholding operation as \(\mathcal{S}_{\tau}[x] = \mathrm{sign}(x)\max(|x|-\tau, 0)\). The optimal solution to Eq.-[\[eq:update_S\_Sharp\]](#eq:update_S_Sharp){reference-type="eqref" reference="eq:update_S_Sharp"} can be obtained as: \[\label{eq:Update_S_Sharp} \m S^{\sharp} = \m U \mathcal{S}_{\lambda_2/\beta}(\Sigma) \m V,\] where \([\m U, \Sigma, \m V] = \mathrm{svd}(g(\m S)-\m Y_1/\beta)\). ### The solution of \(\m E\): \[\m E = \arg\min_{\m E} \lambda_1\|\m E \|_1 + <\m Y_3, \m C \m D-\m E> + \frac{\beta}{2}\| \m C \m D-\m E \|_F^2,\] A close-form solution exists for this sub-problem by using element-wise shrinkage. \[\label{eq:Update_E} \m E = \mathcal{S}_{\lambda_1/\beta}(\m C \m D + \frac{\m Y_3}{\beta}).\] ### The solution of \(\m C\): \[\begin{array}{l} \m C = \arg\min_{\m C} ~ <\m Y_2, \m S-\m S \m C> + \frac{\beta}{2}\|\m S-\m S \m C\|_F^2 + <\m Y_3, \m C \m D-\m E> + \frac{\beta}{2}\|\m C \m D-\m E\|_F^2 + \\ \qquad \qquad \qquad <\m Y_4, \m 1^T \m C-\m 1^T> + \frac{\beta}{2}\|\m 1^T \m C-\m 1^T \|_F^2 \end{array}\] The closed-form solution of \(\m C\) is derived as: \[\label{eq:update_C} (\m S^T \m S + \m 1 \m 1^T)\m C + \m C(\m D \m D^T) = \m S^T \m S + \m S^T \frac{\m Y_2}{\beta} + \m E \m D^T-\m Y_3 \frac{\m D^T}{\beta} + \m 1 \m 1^T-\m 1\frac{\m Y_4}{\beta}\] \[\label{eq:Update_C} \m C = \m C-\mathrm{diag}(\m C),\] Finally, the Lagrange multipliers \(\{\m Y_i\}_{i=1}^4\) and \(\beta\) are updated as: \[\label{eq:update_Y1} \m Y_1 = \m Y_1 + \beta(\m S^{\sharp}-g(\m S)), \m Y_2 = \m Y_2 + \beta(\m S-\m S \m C),\] \[\label{eq:update_Y3} \m Y_3 = \m Y_3 + \beta(\m C \m D-\m E), \m Y_4 = \m Y_4 + \beta(\m 1^T \m C-\m 1^T),\] \[\label{eq:update_beta} \beta = \min(\beta_m, \rho\beta),\] ### Initialization: As our method tries to solve a non-convex optimization problem [\[eq:main_optimization\]](#eq:main_optimization){reference-type="eqref" reference="eq:main_optimization"}, a proper initialization is needed. In this paper, we initialize our implementation by using the single-body non-rigid structure-from-motion method. Other methods can also be used. Note that these single-body non-rigid structure-from-motion was only used for a proper camera motion. In our current implementation, we have fixed the camera motion while updating the 3D non-rigid reconstruction and segmentation. In future, we will put the update of camera rotation in the loop. ### Sparse feature tracks: In the above section, we provide the derivation for solving the dense feature tracks case. Sparse feature tracks case could be viewed as a simplified form of the dense case by removing the spatial constraint term \(\|\m C \m D \|\) directly. [\[tab:comparisons_baseline\]]{#tab:comparisons_baseline label="tab:comparisons_baseline"} [\[tab:result_umpm\]]{#tab:result_umpm label="tab:result_umpm"} ## Experiment 4: Analysis on Dense Dataset To expeditiously test the effectiveness of our implementation over available dense sequences, we tested our method on the uniformly sampled version of the original sequences. We performed experimentation on benchmark NRSFM synthetic and real data-set sequence introduced by Grag *et al.* Table [4](#tab:dense_3d_synthetic){reference-type="ref" reference="tab:dense_3d_synthetic"} lists the obtained 3D reconstruction error on these freely available standard synthetic data-sequence. Figure [\[fig:Dense_result_synthetic\]](#fig:Dense_result_synthetic){reference-type="ref" reference="fig:Dense_result_synthetic"} and [\[fig:real_face_back\_\]](#fig:real_face_back_){reference-type="ref" reference="fig:real_face_back_"} provides visual insight of the procured 3D shapes over synthetic and real sequence respectively. To test the segmentation of different structure on real dataset, we performed experiments by combining two real dataset sequence. Figure [\[fig:real_face_heart\_\]](#fig:real_face_heart_){reference-type="ref" reference="fig:real_face_heart_"} shows the segmentation of non-rigid feature tracks to their corresponding classes. ::: # Conclusions This paper has filled in a missing gap in the Structure-from-Motion family by proposing a new framework for "Multi-body Non-rigid-Structure-from-Motion". It achieves a joint non-rigid reconstruction and non-rigid shape segmentation of multiple deformable structures observed in a single image sequence. Under our new multi-body NRSFM framework, the solutions for motion segmentation can better constrain the solutions for 3D reconstruction. This way, we achieved superior performance in both 3D non-rigid reconstruction and non-rigid motion segmentation, compared with the alternative, two stage method (first segment, then reconstruct). In future, we plan to investigate the scalability issue observed in our current implementation; we aim to apply the new method to denser feature tracks in longer video sequences. In addition to this, we plan to improve over state-of-the-art 3D reconstruction method to get much robust reconstruction for the objects undergoing non-rigid deformation.

{'timestamp': '2016-07-18T02:07:21', 'yymm': '1607', 'arxiv_id': '1607.04515', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04515'}

# Numerical simulations of the quantum dynamics and the associated transient-absorption spectra {#Equations of motion and transient-absorption spectra} We model Rb atoms as three-or five-level systems, as shown in Fig. 2(b) in the article. We consider a quantum system in the state \(|\psi(t)\rangle = \hat{\mathcal{U}}(t,t_0)|\psi(t_0)\rangle\), where the evolution operator \(\hat{\mathcal{U}}(t,t_0)\), describing the dynamics from \(t_0\) to \(t\), is solution of the Schrödinger equation \[\frac{\mathop{}\!\mathrm{d} \hat{\mathcal{U}}(t,t_0)}{ \mathop{}\!\mathrm{d} t} =-\mathrm{i}[\hat{H}_0 + \hat{H}_{\mathrm{int}}]\,\hat{\mathcal{U}}(t,t_0)\ \ \ \hat{\mathcal{U}}(t_0,t_0) = \hat{I}. \label{eq:evolution-operator}\] The total Hamiltonian consists of the atomic-structure Hamiltonian \(\hat{H}_0 = \sum_i{(\omega_{i}-\mathrm{i}\gamma_i/2 ) |i \rangle\langle i|}\), complex-valued and including energies \(\omega_{i}\) and linewidths \(\gamma_{i}\) of each atomic state \(|i\rangle\), and of the electric-dipole (\(E1\)) light-matter interaction \(\hat{H}_{\mathrm{int}} =-[\hat{\boldsymbol{D}}^{-}\cdot\,\boldsymbol{\mathcal{E}}^{+}(t) + \mathrm{H.\,c.}]\), in the rotating-wave approximation. Here, \(\hat{I} = \sum_i{|i \rangle\langle i|}\) is the identity operator, while \(\hat{\boldsymbol{D}}^{-} = \sum_{i>j} \boldsymbol{D}_{ij} |i \rangle\langle j|\) is the negative-frequency part of the dipole-moment operator, of matrix elements \(\boldsymbol{D}_{ij}\), and \(\boldsymbol{\mathcal{E}}^{+}(t)\) is the complex electric field. Pump and probe pulses are modeled by functions of the form \(\boldsymbol{\mathcal{E}}(t-t_{\mathrm{c}},\phi) = \boldsymbol{\mathcal{E}}^{+}(t-t_{\mathrm{c}},\phi ) + \mathrm{c.c.} = \mathcal{E}_0\,f\left(t-t_{\mathrm{c}}\right)\,\cos\left[\omega_{\mathrm{L}} \left(t-t_{\mathrm{c}}\right)+\phi\right]\hat{\boldsymbol{e}}_z\), centered around \(t_{\mathrm{c}}\), with carrier-envelope phase (CEP) \(\phi\) and envelope function of the form \(f(t) = \cos^2{(\pi t/T)}\,R(t/T)\). Here, \(R(x) = \theta(x+1/2)-\theta(x-1/2)\) is defined in terms of the Heaviside step function \(\theta(x)\), such that \(T = \pi T_{\mathrm{FWHM}}/[2\arccos{(\sqrt[4]{1/2})}]\), with \(T_{\mathrm{FWHM}}\) being the full width at half maximum of \(f^2(t)\). The evolution operator \(\hat{\mathcal{U}}_{0}(t, t_0)\) introduced in the article is associated with a pulse centered on \(t_{\mathrm{c}} = 0\) and with vanishing CEP \(\phi = 0\). The generic evolution operator reads \(\hat{\mathcal{U}}(t, t_0) = \hat{\varPhi}^{\dagger}(\phi)\,\hat{\mathcal{U}}_0\left(t-t_{\mathrm{c}},t_0-t_{\mathrm{c}}\right)\,\hat{\varPhi}(\phi)\), with the diagonal operator \(\hat{\varPhi}(\phi)\) accounting for the pulse CEP. In order to better present our key idea, we employ a Schrödinger-equation-based formalism throughout the manuscript, with decay rates \(\gamma_2 = \gamma_3 = 1/(1500\,\mathrm{fs})\) modeling experimental linewidth due to, e.g., Doppler or collision-induced broadening. Radiative decay times of \(27\,\mathrm{ns}\) and \(25.7\,\mathrm{ns}\), from the excited states 2 and 3, respectively, are more than 4 orders of magnitude larger than the duration of the femtosecond pulses used for our control scheme. Therefore, spontaneous decay can be neglected for our system. In this time scale, also the employed decay rates \(\gamma_2 = \gamma_3 = 1/(1500\,\mathrm{fs})\) are small and the difference with a density-matrix treatment is therefore negligible. In Sec. [3](#Density-matrix formulation){reference-type="ref" reference="Density-matrix formulation"} we display how our results can be generalized to a more complex case requiring a density-matrix formulation. We solve the Schrödinger equation ([\[eq:evolution-operator\]](#eq:evolution-operator){reference-type="ref" reference="eq:evolution-operator"}) to calculate the time evolution of the system interacting with delayed pump and probe pulses, and use the obtained solution to simulate transient-absorption spectra. The same procedure is used to calculate the dynamics of the controlled quantum system, displayed in Figs. 3(c) and 4 in the article, for an optimal sequence of two intense pulses. To avoid repetitions, in the following we immediately refer to the more general case of a five-level scheme. The state of the system \(|\psi^{(5)}(t,\tau)\rangle = \sum_{i=1}^5c^{(5)}_i(t,\tau)|i\rangle\) has a vector representation given by \(\boldsymbol{c}^{(5)} = (c^{(5)}_1,\,c^{(5)}_2,\,c^{(5)}_3,\,c^{(5)}_4,\,c^{(5)}_5)^{\mathrm{T}}\). The free-evolution operator is given by Photoionization of states \(|4\rangle\) and \(|5\rangle\) in the presence of an optical pulse has also been included as an effective loss of amplitude at the rate \(\gamma_{\mathrm{P},l}(t) = \sigma_l(\omega_{L}) \mathcal{I}_{\mathrm{L}}(t)\), with photoionization cross sections \(\sigma_l(\omega_{L})\) calculated with, and optical flux \(\mathcal{I}_{\mathrm{L}}(t) = I_{\mathrm{L}}(t)/\omega_{L}\) defined in terms of the time-dependent pulse intensity \(I_{L}(t) = [\mathcal{E}_0\,f(t)]^2/(8\pi\alpha)\). The dynamics of the five-level system are associated with the \(5\times 5\) evolution operator \(\hat{\mathcal{U}}^{(5)}(t, t_0)\) and the related interaction operator \[\begin{split} \hat{U}^{(5)} &= \hat{V}^{(5)}(-T/2)\,\hat{\mathcal{U}}^{(5)}_{0}(T/2,-T/2)\,[\hat{V}^{(5)}(T/2)]^{-1} \\ &= \begin{pmatrix} \hat{U}^{(5)}_{3\times3} & \hat{U}^{(5)}_{3\times2}\\ \hat{U}^{(5)}_{2\times3} & \hat{U}^{(5)}_{2\times2} \end{pmatrix}. \end{split}\] The time evolution of the five-level system, calculated from Eq. ([\[eq:timeevolution5\]](#eq:timeevolution5){reference-type="ref" reference="eq:timeevolution5"}), is used to numerically simulate transient-absorption spectra \(\mathcal{S}^{(5)}(\omega,\tau)\). Fast oscillations as a function of the time delay \(\tau\) are averaged out in a noncollinear geometry and cannot be discerned. We numerically calculate the averaged time-delay-dependent spectrum by performing a convolution of the fast oscillating spectrum with a normalized Gaussian function \(G(\tau, \Delta\tau)\) of width \(\Delta\tau = 5\times2\pi/\omega_{\mathrm{L}}\). For a three-level model associated with the state \(\left|\psi^{(3)}(t,\tau)\right\rangle =\sum_{i=1}^3 c^{(3)}_i(t,\tau)|i\rangle\) and the density-matrix elements \(\rho^{(3)}_{ij}(t,\tau)= c^{(3)}_i(t,\tau)\,[c^{(3)}_j(t,\tau)]^*\), \(i,\,j\in\{1,\,2,\,3\}\), then the absorption spectrum results from \[\mathcal{S}^{(3)}(\omega,\tau)\propto-\omega\mathop{}\!\mathbf{Im}\left[\sum_{k=2}^3 D_{1k}^* \int_{-\infty}^{\infty}\rho^{(3)}_{1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i} \omega (t-\tau)}\mathop{}\!\mathrm{d} t\right]. \label{eq:spectrum3}\] Similarly, the dynamics of the three-level system is described in terms of the operators \(\hat{V}^{(3)}(t) = \mathrm{diag} (1,\,\mathrm{e}^{-(\frac{\gamma_2}{2} + \mathrm{i} \omega_{21})t},\,\mathrm{e}^{-(\frac{\gamma_3}{2} + \mathrm{i} \omega_{31})t})\), \(\hat{\varPhi}^{(3)}(\phi) = \mathrm{diag}(1,\,\mathrm{e}^{\mathrm{i}\phi},\,\mathrm{e}^{\mathrm{i}\phi})\), \(\hat{\varLambda}^{(3)}(t) = \mathrm{diag}(1,\,\mathrm{e}^{\mathrm{i}\omega_{\mathrm{L}}t},\,\mathrm{e}^{\mathrm{i}\omega_{\mathrm{L}}t})\), while \[\hat{M}^{(3)}(t) = \begin{pmatrix} 0 & \mathrm{i} \frac{\varOmega_{12}(t)}{2} & \mathrm{i} \frac{\varOmega_{13}(t)}{2} \\ \mathrm{i} \frac{\varOmega_{12}^*(t)}{2} &-\frac{\gamma_2}{2}-\mathrm{i} \omega_{21} & 0\\ \mathrm{i} \frac{\varOmega_{13}^*(t)}{2} & 0 &-\frac{\gamma_3}{2}-\mathrm{i} \omega_{31} \end{pmatrix}, \label{eq:M3}\] resulting in a \(3\times 3\) evolution operator \(\hat{\mathcal{U}}^{(3)}(t, t_0)\) and a related interaction operator \(\hat{U}^{(3)} = \hat{V}^{(3)}(-T/2)\,\hat{\mathcal{U}}^{(3)}_{0}(T/2,-T/2)\,[\hat{V}^{(3)}(T/2)]^{-1}\). It is important to stress that the \(3\times 3\) matrix \(\hat{U}^{(5)}_{3\times3}\) is different from \(\hat{U}^{(3)}\), since \(\hat{U}^{(5)}_{3\times3}\) is influenced by the presence of states \(|4\rangle\) and \(|5\rangle\), which could not be taken into account if one solved the Schrödinger equation for a three-level system exclusively. By using TAS to extract strong-field interaction (SFI) operators by means of the analytical fitting model presented in Sec. [2](#Fitting analytical model of transient-absorption spectra in terms of intensity-dependent control operators){reference-type="ref" reference="Fitting analytical model of transient-absorption spectra in terms of intensity-dependent control operators"}, one has thus access to \(\hat{U}^{(5)}_{3\times3}\), in contrast to methods based exclusively on a theory which could only provide \(\hat{U}^{(3)}\). # Analytical fitting model of transient-absorption spectra in terms of interaction operators {#Fitting analytical model of transient-absorption spectra in terms of intensity-dependent control operators} To interpret transient-absorption spectra from Rb atoms, modeled by the few-level scheme depicted in Fig. 2(b) in the article, we focus on a \(V\)-type three-level model, accounting for the two transitions \(|1\rangle\rightarrow|k\rangle\), \(k\in\{2,\,3\}\). Starting from these assumptions and the associated definition of the spectrum in Eq. ([\[eq:spectrum3\]](#eq:spectrum3){reference-type="ref" reference="eq:spectrum3"}), we derive an analytical fitting model to relate experimental transient-absorption spectra to the operators \(\hat{U}_{\mathrm{pu}}(I)\) and \(\hat{U}_{\mathrm{pr}}\) ## Interaction operator for weak short probe pulses {#Evolution operator for weak short probe pulses} The operator \(\hat{U}_{\mathrm{pr}}\) models a weak probe pulse, with FWHM of \(15\,\mathrm{fs}\) and intensity of \(1\times 10^{8}\,\mathrm{W/cm^2}\). In order to interpret results from TAS, we approximate the envelope function \(f_{\mathrm{pr}}(t)\) of such a short (broadband) weak pulse with a Dirac-\(\delta\)-like peak \[\tilde{f}_{\mathrm{pr}}(t) = \delta(t)\,\int_{-T_{\mathrm{pr}}/2}^{T_{\mathrm{pr}}/2} f_{\mathrm{pr}}(t')\,\mathop{}\!\mathrm{d} t',\] where \(T_{\mathrm{pr}}\) is the duration of the pulse. An explicit solution of Eq. ([\[eq:evolution-operator\]](#eq:evolution-operator){reference-type="ref" reference="eq:evolution-operator"}) provides \[\hat{\mathcal{U}}_{\mathrm{pr}}(t,t_0) = \hat{V}(t)\,\mathrm{e}^{\hat{B}\,[\theta(t)-\theta(t_0 )]}\,\hat{V}(-t_0),\] where \(\theta(x)\) is the Heaviside step function, \(\hat{V}(t)\) is the free-evolution operator, and \[\hat{B} = \begin{pmatrix} 0 & \mathrm{i} \frac{\vartheta_2}{2} & \mathrm{i} \frac{\vartheta_3}{2}\\ \mathrm{i} \frac{\vartheta_2^*}{2} &0 & 0\\ \mathrm{i} \frac{\vartheta_3^*}{2} & 0 &0 \end{pmatrix}\] is defined in terms of the pulse areas \(\vartheta_k =\int_{-T_{\mathrm{pr}}/2}^{T_{\mathrm{pr}}/2} \varOmega_{\mathrm{pr},1k}(t)\,\mathop{}\!\mathrm{d} t\) of the probe-pulse Rabi frequencies \(\varOmega_{\mathrm{pr},1k}(t)\), for \(k\in\{2,\,3\}\). Taking advantage of the weak probe-pulse intensity, the probe operator \(\hat{U}_{\mathrm{pr}} = \hat{V}(-T_{\mathrm{pr}}/2)\,\hat{\mathcal{U}}_{\mathrm{pr}}(T_{\mathrm{pr}}/2,-T_{\mathrm{pr}}/2)\,\hat{V}^{-1}(T_{\mathrm{pr}}/2)\) reads \[\hat{U}_{\mathrm{pr}} = \mathrm{e}^{\hat{B}} \approx \hat{I}_{3} + \hat{B} = \begin{pmatrix} 1 & \mathrm{i} \frac{\vartheta_2}{2} & \mathrm{i} \frac{\vartheta_3}{2}\\ \mathrm{i} \frac{\vartheta_2^*}{2} &1 & 0\\ \mathrm{i} \frac{\vartheta_3^*}{2} & 0 &1 \end{pmatrix}. \label{eq:Upr}\] ## Interpretation of transient-absorption spectra and reconstruction of strong-field interaction operators {#Transient-absorption spectra} While first-order perturbation theory is used to write \(\hat{U}_{\mathrm{pr}}\) explicitly \[Eq. ([\[eq:Upr\]](#eq:Upr){reference-type="ref" reference="eq:Upr"})\], no assumption is required for the operator \(\hat{U}_{\mathrm{pu}}\), describing the interaction with strong pump pulses. We show how its 9 matrix elements are linked to measurable transient-absorption spectra, deriving an analytical fitting model which can be used to extract \(\hat{U}_{\mathrm{pu}}\) directly from observable data. ### Probe-pump scheme For a probe-pump scheme (\(\tau<0\)), the weak probe pulse generates the initial excited state, which is nonperturbatively modified by the action of the second-arriving intense pump pulse. In terms of the operators \(\hat{U}_{\mathrm{pr}}\) and \(\hat{U}_{\mathrm{pu}}(I)\), modeling light-matter interactions as effectively instantaneous transformations, the effective evolution of the time-delay-dependent state \(\left|\psi_{\mathrm{fit}}(t,\tau)\right\rangle = \sum_{i=1}^3 c_i(t,\tau)|i\rangle\) from the effective initial state \(|\psi_0\rangle = |1\rangle\) is given by \[|\psi_{\mathrm{fit}}(t,\tau)\rangle = \left\{ \begin{aligned} &|\psi_0\rangle, &\text{if \(t<\tau\)},\\ &\hat{V}(t-\tau)\hat{U}_{\mathrm{pr}}|\psi_0\rangle, &\text{if \(\tau<t<0\)},\\ &\hat{V}(t)\hat{U}_{\mathrm{pu}}(I)\,\hat{V}(-\tau)\,\hat{U}_{\mathrm{pr}}|\psi_0\rangle, &\text{if \(t>0\)}. \end{aligned} \right. \label{eq:evolutionprobepump}\] We can analogously define the effective density-matrix elements \(\rho_{\mathrm{fit},ij}(t,\tau)= c_i(t,\tau)\,[c_j(t,\tau)]^*\), in terms of the components \(c_{i}(t,\tau)\) of \(\left|\psi_{\mathrm{fit}}(t,\tau)\right\rangle\). By inserting \(\rho_{\mathrm{fit},ij}(t,\tau)\) into Eq. ([\[eq:spectrum3\]](#eq:spectrum3){reference-type="ref" reference="eq:spectrum3"}), an analytical interpretation model for the probe-pump spectrum is obtained as \[\begin{split} &\mathcal{S}_{\mathrm{fit}}(\omega,\tau,U_{\mathrm{pu},ij}) \\ \propto\,&-\omega\mathop{}\!\mathbf{Im}\biggl\{\sum_{k=2}^3 D_{1k}^* \biggr[\int_{\tau}^{0}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ &\ \ \ \ \ \ \ \ \ \ \ \ + \int_{0}^{\infty}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t\biggr]\biggr\}\. \end{split} \label{eq:spprpu}\] The first integral in Eq. ([\[eq:spprpu\]](#eq:spprpu){reference-type="ref" reference="eq:spprpu"}) is equal to \[\begin{split} &\,\int_{\tau}^{0}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ =&\,U_{\mathrm{pr,}11}\, U^*_{\mathrm{pr,}k1} \int_{\tau}^{0} V^*_{kk}(t-\tau) \,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ = &\,-\mathrm{i}\frac{\vartheta_k}{2}\,\frac{1-\mathrm{e}^{\mathrm{i}(\omega-\omega_{k1}) \tau}\,\mathrm{e}^{\frac{\gamma_k}{2}\tau}}{\mathrm{i}(\omega-\omega_{k1}) + \frac{\gamma_k}{2}}, \end{split} \label{eq:spprpu1stcomponent}\] which, for \(\omega \approx \omega_{k1}\), does not feature fast oscillations as a function of the time delay \(\tau\). The second integral in Eq. ([\[eq:spprpu\]](#eq:spprpu){reference-type="ref" reference="eq:spprpu"}) is given by \[\begin{split} &\,\int_{0}^{\infty}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ =&\,\sum_{j,j' =1}^3U_{\mathrm{pu},1j}\,U_{\mathrm{pu},kj'}^*\,V_{jj}(-\tau)\,V_{j'j'}^*(-\tau)\,U_{\mathrm{pr},j1}\,U_{\mathrm{pr},j'1}^*\\ &\times\,\int_{0}^{\infty} V_{kk}^*(t)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)} \,\mathop{}\!\mathrm{d} t \\ = &\ \frac{1}{\,\mathrm{i}(\omega-\omega_{k1}) + \frac{\gamma_k}{2}}\sum_{j,j' =1}^3U_{\mathrm{pu},1j}\,U_{\mathrm{pu},kj'}^*\,\\ &\ \ \times\,\left[V_{jj}(-\tau)\,V_{j'j'}^*(-\tau)\,\mathrm{e}^{\mathrm{i}\omega\tau}\right]\,U_{\mathrm{pr},j1}\,U_{\mathrm{pr},j'1}^*. \end{split} \label{eq:spprpu2ndcomponent}\] Firstly, all terms in the above sum depend upon nonvanishing elements of the probe-pulse operator \(\hat{U}_{\mathrm{pr}}\) \[Eq. ([\[eq:Upr\]](#eq:Upr){reference-type="ref" reference="eq:Upr"})\]. In a collinear geometry all terms in the above sum contribute to the resulting transient-absorption spectrum, which can be used to extract SFI matrix-element products such as \(U_{\mathrm{pu}, 1j}\,U_{\mathrm{pu}, kj'}\). In a noncollinear geometry, however, such as the one utilized in Ref. , fast oscillating terms due to \([V_{jj}(-\tau)\,V_{j'j'}^*(-\tau)\,\mathrm{e}^{\mathrm{i}\omega\tau}]\) are averaged out and do not contribute to the resulting spectra. Equation ([\[eq:spprpu2ndcomponent\]](#eq:spprpu2ndcomponent){reference-type="ref" reference="eq:spprpu2ndcomponent"}) can be employed to recognize and eliminate all terms which, at \(\omega\approx \omega_{k1}\), feature fast oscillations in \(\tau\). As a result, in terms of the two functions \[\begin{split} A_2(\tau) &= U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}22} + U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}23}\, \frac{\vartheta_{3}}{\vartheta_{2}}\, \mathrm{e}^{-\mathrm{i}\omega_{32}\tau}\,\mathrm{e}^{\frac{\gamma_3-\gamma_2}{2}\tau} ,\\ A_3(\tau) &= U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}33}+ U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}32}\,\frac{\vartheta_{2}}{\vartheta_{3}}\, \mathrm{e}^{\mathrm{i}\omega_{32}\tau}\,\mathrm{e}^{\frac{\gamma_2-\gamma_3}{2}\tau}, \end{split}\] average experimental spectra can be modeled by \[\begin{split} \langle \mathcal{S}_{\mathrm{fit}}(\omega,\tau,U_{\mathrm{pu},ij})\rangle_{\tau} & \propto-\omega\mathop{}\!\mathbf{Im}\biggl\{\sum_{k=2}^3\frac{-\mathrm{i}\,D_{1k}^*\frac{\vartheta_k}{2}}{\mathrm{i}(\omega-\omega_{k1}) + \frac{\gamma_k}{2}}\\ &\ \times \bigl[1+\mathrm{e}^{\mathrm{i}(\omega-\omega_{k1})\tau}\,\mathrm{e}^{\frac{\gamma_k}{2}\tau}\, (A_k(\tau)-1)\bigr]\biggr\}. \end{split} \label{eq:Sexpprpu}\] The proportionality symbol \(\propto\) stresses that these spectra depend on a density-dependent multiplication factor \(K\), which we treat here as a fitting parameter, along with the SFI matrix-element products \(U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}22}\), \(U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}33}\), \(U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}32}\), and \(U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}23}\). ### Pump-probe scheme For a pump-probe scheme (\(\tau>0\)), in terms of the operators \(\hat{U}_{\mathrm{pr}}\) and \(\hat{U}_{\mathrm{pu}}(I)\), the effective evolution of the time-delay-dependent state \(\left|\psi_{\mathrm{fit}}(t,\tau)\right\rangle\) from the effective initial state \(|\psi_0\rangle = |1\rangle\) can be modeled as \[|\psi_{\mathrm{fit}}(t,\tau)\rangle = \left\{ \begin{aligned} &|\psi_0\rangle, &\text{if \(t<0\)},\\ &\hat{V}(t)\hat{U}_{\mathrm{pu}}(I)|\psi_0\rangle, &\text{if \(0<t<\tau\)},\\ &\hat{V}(t-\tau)\hat{U}_{\mathrm{pr}}\,\hat{V}(\tau)\,\hat{U}_{\mathrm{pu}}(I)|\psi_0\rangle, &\text{if \(t>\tau\)}. \end{aligned} \right. \label{eq:evolutionpumpprobe}\] By including in Eq. ([\[eq:spectrum3\]](#eq:spectrum3){reference-type="ref" reference="eq:spectrum3"}) the effective evolution of the matrix elements \(\rho_{\mathrm{fit},ij}(t,\tau)\) from Eq. ([\[eq:evolutionpumpprobe\]](#eq:evolutionpumpprobe){reference-type="ref" reference="eq:evolutionpumpprobe"}), an analytical interpretation model for the pump-probe spectrum is derived, which can be split into the following sum: \[\begin{split} &\mathcal{S}_{\mathrm{fit}}(\omega,\tau, U_{\mathrm{pu},ij}) \\ \propto\,&-\omega\mathop{}\!\mathbf{Im}\biggl\{\sum_{k=2}^3 D_{1k}^* \biggr[\int_{0}^{\tau}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ &\ \ \ \ \ \ \ \ \ + \int_{\tau}^{\infty}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t\biggr]\biggr\}\. \end{split} \label{eq:sppupr}\] The first integral in Eq. ([\[eq:sppupr\]](#eq:sppupr){reference-type="ref" reference="eq:sppupr"}) is equal to \[\begin{split} &\,\int_{0}^{\tau}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ =&\,U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}k1}\,\int_{0}^{\tau}V_{kk}^*(t)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ = &\, U_{\mathrm{pu,}11}\,U^*_{\mathrm{pu,}k1}\,\mathrm{e}^{\mathrm{i}\omega \tau}\,\frac{1-\mathrm{e}^{-\mathrm{i}(\omega-\omega_{k1}) \tau}\,\mathrm{e}^{-\frac{\gamma_k}{2}\tau}}{\mathrm{i}(\omega-\omega_{k1}) + \frac{\gamma_k}{2}}. \end{split} \label{eq:sppupr1stcomponent}\] As described in the case of the probe-pump scheme, the fast oscillations at frequencies \(\omega \approx \omega_{k1}\) are averaged out in a noncollinear geometry, and this first integral does not contribute to the associated average absorption spectrum. The second integral in Eq. ([\[eq:sppupr\]](#eq:sppupr){reference-type="ref" reference="eq:sppupr"}) is given by \[\begin{split} &\,\int_{\tau}^{\infty}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ = &\,\sum_{j,j' =1}^3U_{\mathrm{pr},1j}\,U_{\mathrm{pr},kj'}^*\,V_{jj}(\tau)\,V_{j'j'}^*(\tau)\,U_{\mathrm{pu},j1}\,U_{\mathrm{pu},j'1}^*\\ &\times\,\int_{\tau}^{\infty} V_{kk}^*(t-\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)} \,\mathop{}\!\mathrm{d} t \\ = &\,\frac{1}{\,\mathrm{i}(\omega-\omega_{k1}) + \frac{\gamma_k}{2}} \sum_{j,j' =1}^3U_{\mathrm{pr},1j}\,U_{\mathrm{pr},kj'}^*\,\\ &\times\,\left[V_{jj}(\tau)\,V_{j'j'}^*(\tau)\right]\,U_{\mathrm{pu},j1}\,U_{\mathrm{pu},j'1}^*. \end{split} \label{eq:sppupr2ndcomponent}\] By using Eq. ([\[eq:Upr\]](#eq:Upr){reference-type="ref" reference="eq:Upr"}) and removing fast oscillating terms due to \(V_{jj}(\tau)\,V_{j'j'}^*(\tau)\), which would not appear in a noncollinear geometry, we conclude that \[\begin{split} &\langle \mathcal{S}_{\mathrm{fit}}(\omega,\tau, U_{\mathrm{pu},ij})\rangle_{\tau} \\ \propto &-\omega\mathop{}\!\mathbf{Im}\biggl\{\frac{D_{12}^*}{\mathrm{i}(\omega-\omega_{21}) + \frac{\gamma_2}{2}}\biggl[ \Bigl(-\mathrm{i}\frac{\vartheta_{2}}{2}\, |U_{\mathrm{pu,}11}|^2\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{2}}{2}\, |U_{\mathrm{pu,}21}|^2 \,\mathrm{e}^{-\gamma_2\tau} \Bigr)\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{3}}{2}\,U_{\mathrm{pu,}31} \,U^*_{\mathrm{pu,}21} \,\mathrm{e}^{-\mathrm{i}\omega_{32}\tau}\,\mathrm{e}^{-\frac{\gamma_2+\gamma_3}{2}\tau} \biggr]\\ &\ \ \ \ \ + \frac{D_{13}^*}{\mathrm{i}(\omega-\omega_{31}) + \frac{\gamma_3}{2}}\biggl[ \Bigl(-\mathrm{i}\frac{\vartheta_{3}}{2}\, |U_{\mathrm{pu,}11}|^2\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{3}}{2}\, |U_{\mathrm{pu,}31}|^2 \,\mathrm{e}^{-\gamma_3\tau}\Bigr)\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{2}}{2}\, U_{\mathrm{pu,}21}\,U^*_{\mathrm{pu,}31}\,\mathrm{e}^{\mathrm{i}\omega_{32}\tau}\,\mathrm{e}^{-\frac{\gamma_2+\gamma_3}{2}\tau}\biggr]\biggr\}, \end{split} \label{eq:Sexppupr}\] which explicitly depends upon \(|U_{\mathrm{pu,}11}|^2\), \(|U_{\mathrm{pu,}22}|^2\), \(|U_{\mathrm{pu,}33}|^2\), and \((U_{\mathrm{pu,}31} \,U^*_{\mathrm{pu,}21})\). The same multiplication factor \(K\) should be used which was extracted from the probe-pump spectrum \[see Eq. ([\[eq:Sexpprpu\]](#eq:Sexpprpu){reference-type="ref" reference="eq:Sexpprpu"})\]. ### Additional remarks The matrix elements of the SFI operator \(\hat{U}_{\mathrm{pu}}\) are reconstructed by using the above formulas to fit experimentally measurable transient-absorption spectra. In particular, one can fit probe-pump spectra to quantify \(U_{\mathrm{pu},11}\,U_{\mathrm{pu},22}^*\), \(U_{\mathrm{pu},11}\,U_{\mathrm{pu},33}^*\), \(U_{\mathrm{pu},11}\,U_{\mathrm{pu},32}^*\) and \(U_{\mathrm{pu},11}\,U_{\mathrm{pu},23}^*\), and the common multiplication factor \(K\). Once \(K\) is known, it can be employed to fit pump-probe spectra, and thereby extract \(|U_{\mathrm{pu},11}|^2\), \(|U_{\mathrm{pu},21}|^2\), \(|U_{\mathrm{pu},31}|^2\), and \((U_{\mathrm{pu,}31} \,U^*_{\mathrm{pu,}21})\). All elements of the interaction operator \(\hat{U}_{\mathrm{pu}}\) which can be retrieved from probe-pump spectra are inferred from product terms \(U_{\mathrm{pu},11} U_{\mathrm{pu}, ij}^*\), with \(|U_{\mathrm{pu},11}|\) coming from pump-probe spectra. This has two consequences. Firstly, the phase \(\beta\) of \(U_{\mathrm{pu},11} = |U_{\mathrm{pu},11}|\mathrm{e}^{\mathrm{i}\beta}\) cannot be accessed, resulting in SFI matrix elements \[U_{\mathrm{pu}, ij} = \frac{(U_{\mathrm{pu},11} U_{\mathrm{pu}, ij}^*)^*}{|U_{\mathrm{pu},11}|}\mathrm{e}^{\mathrm{i}\beta}, \label{eq:ratio}\] known up to this common phase \(\beta\). However, this is not a limitation, since it only implies that the final state can be measured and controlled up to a nonrelevant phase term, with access to the information about relevant relative phases. Secondly, when \(|U_{\mathrm{pu}, 11}|\) is very close to \(0\), small uncertainties in its reconstructed value are amplified when the division in Eq. ([\[eq:ratio\]](#eq:ratio){reference-type="ref" reference="eq:ratio"}) is performed to retrieve the remaining matrix elements. For a certain range of pump-pulse intensities, we verified that this is indeed the main source of uncertainty in the extraction of \(\hat{U}_{\mathrm{pu}}\), yet always below the relative level of \(8\%\). In a noncollinear geometry, the remaining pump-probe fitting parameter \((U_{\mathrm{pu,}31} \,U^*_{\mathrm{pu,}21})\) can be used to quantify the relative phase \(\arg{(U_{\mathrm{pu,}31})}-\arg{(U_{\mathrm{pu,}21})}\), but not the absolute phases of \(U_{\mathrm{pu,}21}\) and \(U_{\mathrm{pu,}31}\)---at least up to the same common phase \(\beta\) we introduced before. However, this can be easily circumvented by directly observing the absorption spectrum of a single intense pump pulse. The associated spectral lines are given by \[\mathcal{S}_{\mathrm{fit}}(\omega) \propto\,-\omega\mathop{}\!\mathbf{Im}\biggl\{\mathcal{M} + \sum_{k=2}^3 D_{1k}^* \int_{0}^{\infty}\rho_{\mathrm{fit},1k}(t)\,\mathrm{e}^{-\mathrm{i}\omega t}\,\mathop{}\!\mathrm{d} t \biggr\}\,, \label{eq:Sexpnopr}\] where \(\mathcal{M}\) is a fitting parameter modeling the broadband (and hence constant for small frequency intervals) Fourier transform of \(\rho^{(3)}_{1k}(t)\) in the interval \([-T/2,\,T/2]\), i.e., in the presence of the pump pulse. The remaining integral \[\begin{split} &\,\int_{0}^{\infty}\rho_{\mathrm{fit},1k}(t)\,\mathrm{e}^{-\mathrm{i}\omega t }\,\mathop{}\!\mathrm{d} t \\ = &\,U_{\mathrm{pu},11}\,U_{\mathrm{pu},k1}^*\int_{0}^{\infty} \,\mathrm{e}^{-\mathrm{i}(\omega-\omega_{k1})t} \,\mathrm{e}^{-\frac{\gamma_k}{2}t}\,\mathop{}\!\mathrm{d} t \\ = &\,U_{\mathrm{pu},11}\,U_{\mathrm{pu},k1}^*\,\frac{1}{\mathrm{i}(\omega-\omega_{k1}) + \frac{\gamma_k}{2}} \end{split} \label{eq:integral1pulse}\] exploits the effective evolution of the system, given by \[|\psi_{\mathrm{fit}}(t)\rangle = \left\{ \begin{aligned} &|\psi_0\rangle, &\text{if \(t<0\)},\\ &\hat{V}(t)\hat{U}_{\mathrm{pu}}(I)|\psi_0\rangle, &\text{if \(t>0\)}. \end{aligned} \right. \label{eq:evolutionnoprobe}\] Up to the same common phase \(\beta\) already mentioned above, this allows one to retrieve the phases of \(U_{\mathrm{pu},21}\) and \(U_{\mathrm{pu},31}\), complementing the amplitude information accessible by fitting pump-probe spectra. In order to independently extract \(U_{\mathrm{pu},12}\) and \(U_{\mathrm{pu},13}\) in a noncollinear geometry, we show that, when the envelope function \(f(t) = f(-t)\) is a symmetric function of time, then the associated interaction operator \(\hat{U} = \hat{V}(-T/2)\,\hat{\mathcal{U}}_0(T/2,-T/2)\,\hat{V}^{-1}(T/2)\) is a symmetric matrix. We recall that the time evolution of \(\hat{\mathcal{U}}_0(t,t_0)\) is calculated for a pulse with \(\phi = 0\) and \(t_{\mathrm{c}} = 0\), such that \[\begin{split} \frac{\mathop{}\!\mathrm{d} \hat{\mathcal{U}}_0(t,t_0)}{\mathop{}\!\mathrm{d} t} &= [\hat{\varLambda}^{(3)}(t)]^{\dagger}\,\hat{M}^{(3)}(t)\,\hat{\varLambda}^{(3)}(t)\,\hat{\mathcal{U}}(t,t_0),\\ \hat{\mathcal{U}}(t_0,t_0) &= \hat{I}, \end{split} \label{eq:timeevolution3U}\] compare this with Eq. ([\[eq:timeevolution5\]](#eq:timeevolution5){reference-type="ref" reference="eq:timeevolution5"}). The assumption of a symmetric envelope, \(f(t) = f(-t)\), implies that also \(\hat{M}^{(3)}(t)\) is symmetric in time. We thus introduce the operator \(\hat{\mathcal{Z}}(t,t_0) = \hat{\mathcal{U}}_0(-t,-t_0)\), which is solution of the differential equation \[\begin{split} \frac{\mathop{}\!\mathrm{d} \hat{\mathcal{Z}}(t,t_0)}{\mathop{}\!\mathrm{d} t} &=-\hat{\varLambda}^{(3)}(t)\,\hat{M}^{(3)}(t)\,[\hat{\varLambda}^{(3)}(t)]^{-1}\,\hat{\mathcal{Z}}(t,t_0),\\ \hat{\mathcal{Z}}(t_0,t_0) &= \hat{I}, \end{split} \label{eq:evolutionZ}\] where \(\hat{\varLambda}^{(3)}(-t) = [\hat{\varLambda}^{(3)}(t)]^{-1}\) and \(\hat{M}^{(3)}(-t) = \hat{M}^{(3)}(t)\). Both \(\hat{\varLambda}^{(3)}(t) = [\hat{\varLambda}^{(3)}(t)]^{\mathrm{T}}\) and \(\hat{M}^{(3)}(t) = [\hat{M}^{(3)}(t)]^{\mathrm{T}}\) are symmetric matrices. As a result, \[\begin{split} \frac{\mathop{}\!\mathrm{d} \hat{\mathcal{Z}}^{-1}(t,t_0)}{\mathop{}\!\mathrm{d} t} &= \hat{\mathcal{Z}}^{-1}(t,t_0)\hat{\varLambda}^{(3)}(t)\,\hat{M}^{(3)}(t)\,[\hat{\varLambda}^{(3)}(t)]^{-1},\\ \hat{\mathcal{Z}}^{-1}(t_0,t_0) & = \hat{I}, \end{split} \label{eq:evolutionZ-1}\] and \[\begin{split} \frac{\mathop{}\!\mathrm{d} (\hat{\mathcal{Z}}^{-1})^{\mathrm{T}}(t,t_0)}{\mathop{}\!\mathrm{d} t} &\,=[\hat{\varLambda}^{(3)}(t)]^{-1}\,\hat{M}^{(3)}(t)\,\hat{\varLambda}^{(3)}(t)\,(\hat{\mathcal{Z}}^{-1})^{\mathrm{T}}(t,t_0),\\ (\hat{\mathcal{Z}}^{-1})^{\mathrm{T}}(t_0,t_0) &\,= \hat{I}. \end{split} \label{eq:evolutionZ-1T}\] Since this equation corresponds to Eq. ([\[eq:timeevolution3U\]](#eq:timeevolution3U){reference-type="ref" reference="eq:timeevolution3U"}), the evolution operators \((\hat{\mathcal{Z}}^{-1})^{\mathrm{T}}(t,t_0)\) and \(\hat{\mathcal{U}}_0(t, t_0)\) are solution of the same differential equations and are therefore identical, and hence \(\hat{\mathcal{U}}_0(T/2,0) = \hat{\mathcal{U}}_0^{-1}(0, T/2) = \hat{\mathcal{Z}}^{-1}(0,-T/2) = \hat{\mathcal{U}}_0^{\mathrm{T}}(0,-T/2)\). As a result, \[\begin{split} \hat{\mathcal{U}}_0(T/2,-T/2) &= \hat{\mathcal{U}}_0(T/2,0)\,\hat{\mathcal{U}}_0(0,-T/2)\\ &=\hat{\mathcal{U}}_0^{\mathrm{T}}(0,-T/2) \,\hat{\mathcal{U}}_0(0,-T/2), \end{split}\] and \[\begin{split} \hat{\mathcal{U}}_0^{\mathrm{T}}(T/2,-T/2) &= \hat{\mathcal{U}}_0^{\mathrm{T}}(0,-T/2) \,\hat{\mathcal{U}}_0(0,-T/2) \\ &= \hat{\mathcal{U}}_0(T/2,-T/2), \end{split}\] i.e., \(\hat{\mathcal{U}}_0(T/2,-T/2)\) is a symmetric matrix. Since \(\hat{V}(t)\) is a diagonal (and hence symmetric) matrix, it follows that \(\hat{U} = \hat{V}(-T/2)\,\hat{\mathcal{U}}_0(T/2,-T/2)\,\hat{V}^{-1}(T/2)\) is symmetric. We hence use the above result for the operator \(\hat{U}_{\mathrm{pu}}\) reconstructed from strong-field TAS, and fix \(U_{\mathrm{pu},1k} = U_{\mathrm{pu},k1}\). We stress that this is not a disadvantage of the operator-reconstruction scheme proposed here. As shown by Eq. ([\[eq:spprpu2ndcomponent\]](#eq:spprpu2ndcomponent){reference-type="ref" reference="eq:spprpu2ndcomponent"}), an independent extraction of \(U_{\mathrm{pu},12}\) and \(U_{\mathrm{pu},13}\), without additional assumptions on the envelope \(f(t)\), is possible in a collinear geometry. # Density-matrix formulation {#Density-matrix formulation} The equations of motion have been investigated thus far in the framework of the Schrödinger equation, assuming a decay-rate model with a complex-valued atomic structure Hamiltonian. For our implementation with Rb atoms, this was a proper treatment, in the light of the very different time scales in which the strong-field dynamics and the relaxation of the system take place. For more general cases, however, a density-matrix formulation of our operator-reconstruction and quantum-control scheme is possible, as we show in the following. In this section, we will focus on the case of a three-level system, although the equations can be easily generalized to higher numbers of levels if necessary. We describe the system in terms of a \(3\times 3\) density matrix \(\hat{\rho}(t)\), with a diagonal element \(\rho_{ii}(t)\) associated with the population of the \(i\)th level, and off-diagonal matrix elements for the atomic coherences. A density-matrix formulation may be necessary, e.g., to explicitly account for the spontaneous decay of the two excited levels \(2\) and \(3\) into the ground state \(1\), at rates respectively given by \(\gamma_{22}\) and \(\gamma_{33}\). Similarly, it is necessary to account for additional decoherence processes affecting the atomic coherences \(\rho_{12}\), \(\rho_{13}\), and \(\rho_{23}\). In this Section, this is modeled by the decoherence rates \(\gamma_{21}\), \(\gamma_{31}\), and \(\gamma_{23}\), respectively. The evolution of the density matrix \(\hat{\rho}(t)\) is then given by the master equation \[\frac{\mathop{}\!\mathrm{d} \hat{\rho}(t)}{\mathop{}\!\mathrm{d} t} = \mathrm{i} [\hat{\rho}(t),\,\hat{H}_0 + \hat{H}_{\mathrm{int}}] + \mathcal{L}[\hat{\rho}(t)],\] with the atomic structure Hamiltonian now given by \(\hat{H}_ 0 = \sum \omega_i |i\rangle\langle i|\) and the superoperator \(\mathcal{L}[\hat{\rho}(t)]\) describing dissipation. In the absence of external fields, the equations of motion satisfied by the elements of the density matrix can hence be written as: \[\begin{split} \frac{\mathop{}\!\mathrm{d} \rho_{11}}{\mathop{}\!\mathrm{d} t} &= \gamma_{22}\rho_{22} + \gamma_{33}\rho_{33},\\ \frac{\mathop{}\!\mathrm{d} \rho_{kk}}{\mathop{}\!\mathrm{d} t} &=-\gamma_{kk}\rho_{kk},\\ \frac{\mathop{}\!\mathrm{d} \rho_{1k}}{\mathop{}\!\mathrm{d} t} &=-\gamma_{k1}\rho_{1k} + \mathrm{i} \omega_{k1} \rho_{1k},\\ \frac{\mathop{}\!\mathrm{d} \rho_{23}}{\mathop{}\!\mathrm{d} t} &=-\gamma_{23}\rho_{23} + \mathrm{i} \omega_{32} \rho_{23}, \end{split}\] with \(k\in\{2,3\}\) and \(\rho_{ij} = \rho_{ji}^*\). The explicit solution reads \[\begin{split} \rho_{11}(t) &= \rho_{11}(0) + \rho_{22}(0)\left(1-\mathrm{e}^{-\gamma_{22}t} \right) \\ &\ \ \ + \rho_{33}(0)\left(1-\mathrm{e}^{-\gamma_{33}t} \right),\\ \rho_{kk}(t) &= \rho_{kk}(0) \mathrm{e}^{-\gamma_{kk}t},\\ \rho_{1k}(t) &= \rho_{1k}(0) \mathrm{e}^{-\gamma_{k1}t} \mathrm{e}^{\mathrm{i} \omega_{k1} t},\\ \rho_{23}(t) &= \rho_{23}(0) \mathrm{e}^{-\gamma_{23}t} \mathrm{e}^{\mathrm{i} \omega_{32} t}, \end{split}\] summarized by the elements of the operator \(\boldsymbol{X}(t)\) via \[\rho_{ij}(t) = \sum_{i'j'}\boldsymbol{X}_{ij,i'j'}(t)\rho_{i'j'}(0).\] In analogy to Eq. (1) in the article, we can introduce an operator \(\boldsymbol{Y}\) describing the effect of ultrashort pump and probe pulses on the density matrix of the system as an instantaneous interaction, i.e., \[\rho^+ = \boldsymbol{Y}(I)\rho^-.\] For ultrashort pulses, acting on a time scale much shorter than the time scale of the decay processes, we can assume that \(\boldsymbol{Y}(I) = \hat{U}(I)\otimes \hat{U}^*(I)\), i.e., \(\boldsymbol{Y}_{ij,i'j'} = U_{ii'}\,U_{jj'}^*\). In contrast to the previous Sections, however, here we fully account for the relaxation of the system during its free evolution in terms of the operator \(\boldsymbol{X}(t)\). A similar approach can be used to model quantum systems also in the presence of a more structured reservoir, provided that the free evolution of the system is known and can be used to define the operator \(\boldsymbol{X}(t)\). For a TAS experiment in a probe-pump setup, we can describe the effective evolution of the time-delay-dependent density matrix \(\hat{\rho}_{\mathrm{fit}}(t,\tau)\) from the effective initial state \(\rho_0 = |1\rangle\langle 1|\) as \[\hat{\rho}_{\mathrm{fit}}(t,\tau) = \left\{ \begin{aligned} &\hat{\rho}_0, &\text{if \(t<\tau\)},\\ &\boldsymbol{X}(t-\tau)\boldsymbol{Y}_{\mathrm{pr}}\hat{\rho}_0, &\text{if \(\tau<t<0\)},\\ &\boldsymbol{X}(t)\boldsymbol{Y}_{\mathrm{pu}}(I)\,\boldsymbol{X}(-\tau)\,\boldsymbol{Y}_{\mathrm{pr}}\hat{\rho}_0, &\text{if \(t>0\)}. \end{aligned} \right.\] By inserting \(\rho_{\mathrm{fit},ij}(t,\tau)\) into Eq. ([\[eq:spectrum3\]](#eq:spectrum3){reference-type="ref" reference="eq:spectrum3"}), an analytical interpretation model for the probe-pump spectrum is obtained consisting also in this case of two integrals. The first integral reads \[\begin{split} &\,\int_{\tau}^{0}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ =&\int_{\tau}^{0} \sum_{i,j=1}^3\boldsymbol{X}_{1k,ij}(t-\tau)\,\boldsymbol{Y}_{\mathrm{pr};ij,11} \,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ = &\,-\mathrm{i}\frac{\vartheta_k}{2}\,\frac{1-\mathrm{e}^{\mathrm{i}(\omega-\omega_{k1}) \tau}\,\mathrm{e}^{\gamma_{k1}\tau}}{\mathrm{i}(\omega-\omega_{k1}) + \gamma_{k1}}, \end{split}\] where we have used the fact that \(\boldsymbol{X}_{1k,ij}(t) = \mathrm{e}^{-\gamma_{k1}t}\,\mathrm{e}^{\mathrm{i}\omega_{k1}t}\,\delta_{1i}\,\delta_{kj}\). The second integral is similarly given by \[\begin{split} &\,\int_{0}^{\infty}\rho_{\mathrm{fit},1k}(t,\tau)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)}\,\mathop{}\!\mathrm{d} t \\ =&\,\sum_{i,j =1}^3\sum_{i',j' =1}^3\sum_{i'',j'' =1}^3 \boldsymbol{Y}_{\mathrm{pu};ij,i'j'}\,\boldsymbol{X}_{i'j',i''j''}(-\tau)\,\boldsymbol{Y}_{\mathrm{pr},i''j'',11}\\ &\times\,\int_{0}^{\infty} \boldsymbol{X}_{1k,ij}(t)\,\mathrm{e}^{-\mathrm{i}\omega (t-\tau)} \,\mathop{}\!\mathrm{d} t \\ = &\,\frac{1}{\,\mathrm{i}(\omega-\omega_{k1}) + \gamma_{k1}}\sum_{i',j' =1}^3\sum_{i'',j'' =1}^3 U_{\mathrm{pu},1i'}\,U_{\mathrm{pu},kj'}^*\,\\ &\ \ \times\,\left[\boldsymbol{X}_{i'j',i''j''}(-\tau)\,\mathrm{e}^{\mathrm{i}\omega\tau}\right]\, U_{\mathrm{pr},i''1}\,U_{\mathrm{pr},j''1}^*. \end{split}\] By including explicitly the value of \(\boldsymbol{X}_{i'j',i''j''}(-\tau)\), one obtains an equation in which, in contrast to Eq. ([\[eq:spprpu2ndcomponent\]](#eq:spprpu2ndcomponent){reference-type="ref" reference="eq:spprpu2ndcomponent"}), different decay rates for coherences and populations are present and additional terms appear due to the spontaneous decay from the excited states to the ground state. In order to model spectra from a noncollinear geometry, fast oscillating terms due to \([\boldsymbol{X}_{i'j',i''j''}(-\tau)\,\mathrm{e}^{\mathrm{i}\omega\tau}]\) can be eliminated, resulting in the average spectrum \[\begin{split} &\langle \mathcal{S}_{\mathrm{fit}}(\omega,\tau,U_{\mathrm{pu},ij})\rangle_{\tau} \\ \propto &-\omega\mathop{}\!\mathbf{Im}\biggl\{\sum_{k=2}^3 D_{1k}^* \,\frac{1}{\mathrm{i}(\omega-\omega_{k1}) + \gamma_{k1}}\\ &\ \ \ \ \ \ \ \biggl[(-\mathrm{i}{\vartheta_k}/{2})\,\bigl(1-\mathrm{e}^{\mathrm{i}(\omega-\omega_{k1}) \tau}\,\mathrm{e}^{\gamma_{k1}\tau}\bigr)\\ &\ \ \ \ \ \ \ \ + U_{\mathrm{pu},11}\,U_{\mathrm{pu},k2}^*\, (-\mathrm{i}{\vartheta_2}/{2})\,\mathrm{e}^{\mathrm{i}(\omega-\omega_{21})\tau}\,\mathrm{e}^{\gamma_{21}\tau}\\ &\ \ \ \ \ \ \ \ \ +U_{\mathrm{pu},11}\,U_{\mathrm{pu},k3}^*\, (-\mathrm{i}{\vartheta_3}/{2})\, \mathrm{e}^{\mathrm{i}(\omega-\omega_{31})\tau}\,\mathrm{e}^{\gamma_{31}\tau} \biggr]\biggr\}. \end{split}\] In a pump-probe setup, the same formalism can be used, resulting in an effective evolution given by \[\hat{\rho}_{\mathrm{fit}}(t,\tau) = \left\{ \begin{aligned} &\hat{\rho}_0, &\text{if \(t<0\)},\\ &\boldsymbol{X}(t)\,\boldsymbol{Y}_{\mathrm{pu}}(I)\,\hat{\rho}_0, &\text{if \(0<t<\tau\)},\\ &\boldsymbol{X}(t-\tau)\,\boldsymbol{Y}_{\mathrm{pr}}\,\boldsymbol{X}(\tau)\,\boldsymbol{Y}_{\mathrm{pu}}(I)\,\hat{\rho}_0, &\text{if \(t>\tau\)}. \end{aligned} \right.\] By following the same procedure outlined for the probe-pump case, and after removing the fast oscillating terms, one obtains the average spectrum given by \[\begin{split} &\langle \mathcal{S}_{\mathrm{fit}}(\omega,\tau, U_{\mathrm{pu},ij})\rangle_{\tau} \\ \propto &-\omega\mathop{}\!\mathbf{Im}\biggl\{\frac{D_{12}^*}{\mathrm{i}(\omega-\omega_{21}) + \gamma_{21}}\biggl[ \Bigl(-\mathrm{i}\frac{\vartheta_{2}}{2}\, |U_{\mathrm{pu,}11}|^2\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{2}}{2}\, |U_{\mathrm{pu,}21}|^2 \,(2\mathrm{e}^{-\gamma_{22}\tau}-1) \Bigr)\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{3}}{2}\,U_{\mathrm{pu,}31} \,U^*_{\mathrm{pu,}21} \,\mathrm{e}^{-\mathrm{i}\omega_{32}\tau}\,\mathrm{e}^{-\gamma_{23}\tau} \biggr]\\ &\ \ \ \ \ + \frac{D_{13}^*}{\mathrm{i}(\omega-\omega_{31}) + \gamma_{31}}\biggl[ \Bigl(-\mathrm{i}\frac{\vartheta_{3}}{2}\, |U_{\mathrm{pu,}11}|^2\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{3}}{2}\, |U_{\mathrm{pu,}31}|^2 \,(2\mathrm{e}^{-\gamma_{33}\tau}-1)\Bigr)\\ &\ \ \ \ \ \ \ \ \ +\mathrm{i}\frac{\vartheta_{2}}{2}\, U_{\mathrm{pu,}21}\,U^*_{\mathrm{pu,}31}\,\mathrm{e}^{\mathrm{i}\omega_{32}\tau}\,\mathrm{e}^{-\gamma_{23}\tau}\biggr]\biggr\}. \end{split}\] Also in this case, in contrast to Eq. ([\[eq:Sexppupr\]](#eq:Sexppupr){reference-type="ref" reference="eq:Sexppupr"}), we notice different decay rates and the presence of additional terms to account for the decay into the ground state 1. This model can be employed to fit experimental results and extract interaction operators \(\boldsymbol{Y}\) in those cases in which a density-matrix description of the free evolution of the system is required. Once SFI operators are reconstructed, then the sequence of pulses constituting our control method can be modeled as \[\hat{\rho}_{N_{\mathrm{p}}} = \tilde{\boldsymbol{Y}}_{N_{\mathrm{p}}} \ldots \boldsymbol{X}(\bar{\tau}_m)\tilde{\boldsymbol{Y}}_{m}\ldots \boldsymbol{X}(\bar{\tau}_1)\tilde{\boldsymbol{Y}}_1\hat{\rho}_0,\] generalizing Eq. (2) in the article to the case when the system's dynamics have to be described by a density matrix.

{'timestamp': '2016-09-22T02:06:48', 'yymm': '1607', 'arxiv_id': '1607.04570', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04570'}

# INTRODUCTION {#sec:intro} A period of accelerated exponential expansion in the early Universe, while a radical extrapolation from understood physics, naturally solves the flatness, horizon, and monopole problems of standard cosmology. This inflationary period also explains the origin of structure by stretching quantum fluctuations to macroscopic scales. Inflation predicts perturbations to the metric in both scalars (density waves) and tensors (gravitational waves). Both types of perturbations affect the polarization of the cosmic microwave background (CMB) at last scattering: scalars can only generate a gradient-type (\(E\)-mode) pattern, while tensors can also generate a curl-type (\(B\)-mode) pattern. Measurement of \(B\)-mode power in the CMB at degree angular scales in excess of the expectation from gravitational lensing would be direct evidence for an inflationary period, and the amplitude of the signal--parametrized by \(r\), the tensor-to-scalar ratio--would indicate the energy scale of inflation. Since 2006, the Bicep/*Keck Array* CMB polarization experiments have been mapping \(\sim 1\%\) of low-foreground sky from the Amundsen-Scott South Pole Station. The telescopes are optimized to detect fluctuations at degree angular scales, and therefore use a compact, on-axis refracting optical system with the minimum aperture necessary to resolve the \(\ell \sim 100\) peak in the inflationary gravitational wave \(B\)-mode spectrum. In 2014 we reported a detection of \(B\)-mode power at degree angular scales at 150 GHz by Bicep2, which was quickly confirmed by data from the *Keck Array*. Multifrequency maps are required to separate Galactic foreground emission (thermal dust at high frequencies and synchrotron at low frequencies) from the CMB signal. Using all available *Planck* and Wmap maps and the latest *Keck Array* maps at both 95 and 150 GHz, we find that the excess at 150 GHz is consistent with dust emission, and have constrained the tensor-to-scalar ratio to \(r < 0.07\) at \(95\%\) confidence. To push this limit further--or achieve a detection of primordial \(B\)-mode s--we have deployed 220 GHz detectors in the *Keck Array* to measure dust at high precision, and have deployed Bicep3 at 95 GHz to probe deeply at a low-foreground frequency. For the 2016 observing season, Bicep3 was outfitted with a full complement of 20 detector tiles (2400 optically coupled detectors). The Bicep/*Keck Array* telescopes measure polarization by differencing co-located, orthogonally polarized detector pairs. Any mismatch in beam shape between the two polarizations will leak some of the bright CMB temperature signal into polarization. Temperature-to-polarization leakage is the most prominent instrumental systematic effect which may cause a false \(B\)-mode signal and must be precisely characterized. We have developed an analysis technique called deprojection to marginalize over any false polarization signal which would arise from the coupling of the CMB temperature sky to a second-order expansion of the measured differential beam; however, the contribution from higher-order beam mismatch remains. In the Bicep2 final results, using high-fidelity far field beam maps we constrained the undeprojected residual contribution to the final \(B\)-mode spectrum to an equivalent \(r < 3.0 \times 10^{-3}\). In these proceedings, we present an analysis of the far field beam response of the Bicep3 detectors observing in the 2016 season. In February and March of 2016, we obtained far field beam maps of all Bicep3 detectors *in situ* at the South Pole using a chopped thermal source in the far field of the telescope. These measurements serve several purposes: the array-averaged beam profile is used to smooth `Healpix` maps for simulations, the measured beam shapes offer a cross-check that the modes removed by deprojection correspond to actual beam mismatch and also allow for direct subtraction of the deprojection templates, and the beam maps are used in dedicated simulations to explicitly predict the residual \(B\)-mode power after deprojection. In Section [2](#sect:optics){reference-type="ref" reference="sect:optics"} we review the optical design of Bicep3. In Section [3](#sect:ffbm){reference-type="ref" reference="sect:ffbm"} we describe the far field beam measurement setup and the new chopped source constructed in 2016. In Section [4](#sect:beamparam){reference-type="ref" reference="sect:beamparam"} we present two-dimensional Gaussian fits to each detector and characterize beamwidth, ellipticity, differential pointing, differential beamwidth, and differential ellipticity across the entire array. In Section [5](#sect:composites){reference-type="ref" reference="sect:composites"} we coadd all beam measurements to obtain high signal-to-noise composite maps of individual beams, and use these beams to calculate the array-averaged beam response. In Section [6](#sect:deprojection){reference-type="ref" reference="sect:deprojection"} we discuss the measured beam mismatch in the context of deprojection and the residual beam systematics levels achieved by Bicep2 and *Keck Array*. Three companion papers--a Bicep3 status update, Bicep3 detector performance, and measurements of *Keck Array* detector polarization angles with a dielectric sheet calibrator--are also presented at this conference. # OPTICAL DESIGN {#sect:optics} Like Bicep1, Bicep2  and *Keck Array*, Bicep3 is a compact, two-lens on-axis telecentric refracting telescope. However, the aperture is twice as large as that of its predecessors (\(520\) mm vs \(264\) mm), the field of view is larger (\(27^{\circ}\) vs \(15^{\circ}\)), and the optics are faster (\(f/1.7\) vs \(f/2.4\)). For observations at a wavelength of \(\lambda = 3 \mbox{ mm}\), the far field is at roughly \(2D^2 / \lambda = 180 \mbox{ m}\). Figure [\[fig:optics\]](#fig:optics){reference-type="ref" reference="fig:optics"} shows a schematic of the optical system compared to that of Bicep2/*Keck Array*. Mounted directly on the detector modules are metal-mesh low pass filters with a cutoff at \(4 \mbox{ cm}^{-1}\) (120 GHz) to block out-of-band power. Two lenses, made of 99.6% pure alumina ceramic, perform the imaging; a cold Lyot stop skyward of the objective lens defines the aperture. The telescope housing, lenses, and aperture stop are all cooled to 4 K. The alumina lenses are anti-reflection coated with a mix of Stycast 1090 and 2850 epoxy tuned to the correct index of refraction (\(n=1.75\) for alumina at \(n=3.1\)), machined to the correct thickness and laser diced into patches to relieve mechanical stress from differential thermal contraction. The \(670 \mbox{ mm}\) aperture high-density polyethylene vacuum window is \(31.75 \mbox{ mm}\) thick and anti-reflection coated with Teadit 24RGD. The large aperture allows over \(100 \mbox{ W}\) of infrared power to enter the cryostat, and the purely absorptive filtering used in Bicep2 and *Keck Array* would overload the Cryomech PT415 pulse tube; we therefore reject a majority of the loading with a stack of 10 metal-mesh IR-reflective filters mounted just below the receiver window at ambient temperature. Below these are flat absorptive filters: a \(10 \mbox{ mm}\) thick alumina filter at 50 K above the objective lens, a \(5 \mbox{ mm}\) thick nylon filter at 4 K above the field lens, and \(9.5 \mbox{ mm}\) thick nylon filter at 4 K between the field lens and focal plane. There is an additional reflective metal mesh filter below the alumina filter at 50 K. The resulting IR load allows continuous operation of the pulse tube at \(< 4\) K and \(> 48\) hour hold time of the sub-Kelvin refrigerator. To terminate far sidelobes in CMB observations, we install a co-moving absorptive forebaffle coated with a combination of Eccosorb HR-25/AN-75 and weatherproofed with Volara foam. The forebaffle is \(128.9 \mbox{ cm}\) tall, \(141 \mbox{ cm}\) in diameter, and intercepts radiation \(> 14^{\circ}\) from the telescope boresight. The telescope is surrounded by a stationary reflective ground shield which redirects off-axis rays to cold sky; the geometry requires that a ray must diffract twice--around the forebaffle and ground shield--before it hits the ground. # FAR FIELD BEAM MEASUREMENT SETUP {#sect:ffbm} The far field beam mapping setup for Bicep3 is very similar to that used for Bicep2 and the *Keck Array*; see the Beams paper for additional details. ## Flat Mirror and Thermal Chopped Microwave Source {#ss:mirrorchopper} Since the telescope is contained within its ground shield and can only tip down to \(\sim 40^{\circ}\) from zenith, we use a flat mirror to redirect the beams over the ground shield and towards the thermal chopper (Figure [\[fig:mirrorchopper\]](#fig:mirrorchopper){reference-type="ref" reference="fig:mirrorchopper"}, left panel). The 1.7 x 2.5 m aluminum honeycomb mirror is flat to 0.2 mm across its surface and is mounted at a \(45^{\circ}\) angle, such that when the telescope is pointed at zenith, the beams are redirected towards the horizon (the source is \(\sim 2^{\circ}\) above the horizon). The forebaffle is removed to provide room for the mirror. For the 2016 deployment season, we built two identical chopped thermal microwave sources so that Bicep3 and *Keck Array* could take far field beam map data simultaneously. These choppers consist of a rotating blade, an enclosure surrounding the blade, and a motor drive assembly to spin the blade. The blade is a composite material made from carbon fiber, low density foam, and honeycomb Nomex, coated with Eccosorb HR-10. The enclosure is made of a similar carbon fiber composite, coated with aluminum foil so that rays which do not terminate on the source aperture are reflected to cold sky (see the right inset of Figure [\[fig:mirrorchopper\]](#fig:mirrorchopper){reference-type="ref" reference="fig:mirrorchopper"}). The source aperture is 24" in diameter, behind which is a flat mirror directed to zenith, so that the source chops between ambient (\(\sim 260\) K) and sky (\(\sim 10\) K). The entire assembly weighs 55 pounds. The right panel of Figure [\[fig:mirrorchopper\]](#fig:mirrorchopper){reference-type="ref" reference="fig:mirrorchopper"} shows the chopper mounted on a 40 ft high mast on the Martin A. Pomerantz Observatory, 211 meters away from Bicep3. The chopper can spin at a maximum of 9 Hz (18 Hz chop rate due to the two blades); for beam mapping in 2016 we operated at a 14 Hz chop rate. The aperture is sealed with HD-30 Zotefoam[^1] to prevent air turbulence from slowing down the blade. Compared to previous chopped thermal microwave sources deployed at South Pole, the 24" diameter chopper offers higher signal due to the larger aperture and lower noise due to the faster chop rate. ## 2016 Data Set {#ss:dataset} The far field beam maps presented in these proceedings were taken in February and March of 2016. Due to the comparatively high loading of the \(\sim 250\) K source, beam maps are taken with the detectors on the aluminum superconducting transition[^2]. In this higher-loading configuration the sub-Kelvin refrigerator lasts for 12 hours. We scan across the source in \(30^{\circ}\) azimuth rasters at \(0.8^{\circ}\) per second, increasing the elevation by \(0.05^{\circ}\) in between each scan. The scan speed was chosen so that at the fast detector sample rate (150 Hz) and source chop rate (14 Hz), multiple chop cycles would be recorded across the \(0.167^{\circ}\) beam. The full field of view of the instrument is therefore covered in \(\sim 8\) hours. To fill in the gap before the next fridge cycle, we additionally run "half" schedules which only cover the top, middle, or bottom of the focal plane. Schedules are evenly divided among five boresight angles: \(0^{\circ}\), \(45^{\circ}\), \(90^{\circ}\), \(135^{\circ}\), and \(315^{\circ}\). Consistency of beam maps across boresight angles offers a robust check on the repeatability of our beam measurements. Including both full and half focal plane configurations, we took a total of 77 beam mapping schedules. Once the data are taken, we demodulate the timestream data using a reference signal from the chopper and bin into maps with \(0.1^{\circ}\) pixelization. The reflection off the mirror is handled with a pointing model that describes both the mount and the mirror. # GAUSSIAN BEAM PARAMETERS {#sect:beamparam} ## Coordinate System and 2D Gaussian Fits {#ss:coord} To facilitate comparison of beam parameters across measurements taken at multiple telescope orientations, we define a coordinate system similar to that used for Bicep2 and *Keck Array* beams analysis--see the Beams paper for further details. Each pixel \(P\), consisting of two orthogonally polarized detectors, is at a location \((r,\theta)\) from the boresight. \(r\) is the radial distance away from the boresight, and \(\theta\) is the counterclockwise angle looking out from the telescope toward the sky from the \(\theta=0^{\circ}\) ray, defined to be the great circle that runs along tiles 11/10/9 on the focal plane. For each pixel we then define a local \((x',y')\) coordinate system. The positive \(x'\) axis is along the great circle passing through \(P\) which is an angle \(-\theta\) from the \(\hat{r}\) direction of the pixel, while the \(y'\) axis is along the great circle \(+90^{\circ}\) counterclockwise away from the \(x'\) axis. The \((x',y')\) coordinate system is then projected onto a plane at \(P\). This coordinate system is fixed to the instrument and rotates on the sky with the boresight rotation angle \(K\) and the angle with which the receiver is clocked with respect to the \(K=0\) line, or the drum angle \(K'\); for Bicep3, \(K'=1.5^{\circ}\). In a simplified sense, to transform from mount coordinates to detector-centered coordinates, we flip the elevation axis to account for the mirror and rotate by \(K + K'\) to arrive at \((x',y')\). All focal plane and beam map plots in these proceedings are rendered as the instantaneous projection of each beam onto the sky, with the same parity (viewing the projection from inside the celestial sphere) and orientation (with the \(+x'\) axis pointing up and the \(+y'\) axis pointing to the left); this is consistent with the conventions established in our other publications. Figure [\[fig:fpu\]](#fig:fpu){reference-type="ref" reference="fig:fpu"} shows the coordinate axes, the location of each nominally working beam, tile numbers and polarization axes, and the division of tiles into separate Multi-Channel Electronics (MCE) systems used to control the multiplexing and readout of the detectors. Each beam is fit to a two-dimensional elliptical Gaussian model with six free parameters: \[B(\bm{x}) = \frac{1}{\Omega} e^{-\frac{1}{2}(\bm{x}-\bm{\mu})^{T} \Sigma^{-1} (\bm{x}-\bm{\mu})},\] where \(\bm{x} = (x,y)\) is the location of the beam center, \(\bm{\mu} = (0,0)\) is the origin, \(\Omega\) is the normalization, and \(\Sigma\) is the covariance matrix parametrized as \[\Sigma = \begin{pmatrix} \sigma^2 (1 + p) & c \sigma^2 \\ c \sigma^2 & \sigma^2 (1-p) \end{pmatrix}.\] \(\sigma\) is the beamwidth and \(p\) and \(c\) are the ellipticities in the "plus" and "cross" directions respectively, such that an elliptical Gaussian with a major axis oriented along \(x'\) or \(y'\) would have a \(+p\) or \(-p\) ellipticity, and one with a major axis oriented diagonally would have a \(\pm c\) ellipticity (see Figure [\[fig:fpu\]](#fig:fpu){reference-type="ref" reference="fig:fpu"}). Parametrization with (\(\sigma, p, c\)), while equivalent to the more common (\(\sigma_{maj}, \sigma_{min}, \phi\)), is more convenient because differential parameters are simple differences: \(d\sigma = \sigma_A-\sigma_B\), \(dx = x_A-x_B\), \(dy = y_A-y_B\), \(dp = p_A-p_B\), and \(dc = c_A-c_B\), where \(A\) and \(B\) refer to the orthogonally polarized detectors within a pair. These parameters directly scale the deprojection templates used to remove temperature-to-polarization leakage in analysis. Note that we do not use the absolute \((x,y)\) beam locations because of the complicated mirror model, instead obtaining beam centers from cross-correlation with the *Planck* CMB temperature maps. In each of the 77 beam map schedules, we find the best fit \(\Omega\), \(\bm{x}\), \(\sigma\), \(p\), and \(c\) for every beam. Not all beams are covered in each schedule, and occasionally some detectors are not working normally even when they are covered. We then remove maps in which the fit failed, or in which only one of the \(A\) or \(B\) polarizations were working, so that the same number of measurements and detectors are used in per-detector and per-pair parameters. After cuts, the average pair has \(\sim 30\) good measurements. Table [1](#tab:stats){reference-type="ref" reference="tab:stats"} shows summary statatistics derived from these measurements. For each detector/pair, we take the median across the \(\sim 30\) measurements as the best estimate of each parameter, and take half the width of the central 68% of the distribution of those measurements as the measurement uncertainty.[^3] The characteristic uncertainty for an individual measurement--taken to be the median measurement uncertainty for all detectors/pairs across the array--is shown in the "Individual Measurement Uncertainty" column of Table [1](#tab:stats){reference-type="ref" reference="tab:stats"}. To characterize the distribution of parameters, we also find the median across the focal plane ("FPU Median" column) and quantify the variation across the focal plane as half the width of the central 68% of the distribution of *best estimates for each detector/pair* ("FPU Scatter" column). Note that the "FPU Scatter" measures the spread of best-estimate parameters across the array and is not a measurement uncertainty. ::: In this section, we compare Bicep3's differential beam parameters with those from Bicep2 and *Keck Array* (Table [3](#tab:statsvb2){reference-type="ref" reference="tab:statsvb2"}) and comment on the expected level of temperature-to-polarization leakage. In our CMB analysis pipeline, for each detector pair we perform a deprojection procedure in which the timestream data are regressed against templates of the real CMB temperature sky and its derivatives corresponding to the difference modes of two mismatched beams. The template multiplied by the best-fit regression coefficient is removed from the timestream, eliminating all leakage caused by a particular mode without external knowledge of the beams. As an alternative to deprojection, we may also choose to fix the coefficient which scales the deprojection template to the beam-map-derived value--we refer to this procedure as "subtraction." Since differential gain is likely time-variable, we deproject it. Gain is not measured in beam maps because of the different TES used and loading environment between calibration measurements and CMB observations. In our previous results, we have chosen to deproject differential pointing; the measured correlation between beam-map-derived values and deprojection coefficients offers a robust check that our beam maps correspond to reality. Differential beamwidth has been found to be negligible, and so has neither been deprojected nor subtracted. We subtract differential ellipticity, because the real-sky \(TE\) correlation would be filtered out of our \(TE\) and \(EE\) spectra. We first consider differential pointing, which for Bicep2 and *Keck Array* has been the beam mismatch mode contributing the most leakage and is independent of beam size. While Bicep2's \(dx\) and \(dy\) were quite large, significant development effort resulted in subsequent *Keck Array* focal planes at 95, 150, and 220 GHz with differential pointing that was several times smaller. Bicep3 is most similar to *Keck Array* 220 GHz in both the magnitude and the measurement precision. We can expect Bicep3's pre-deprojection leakage from differential pointing to be much less than that of Bicep2; the distribution is also incoherent across the focal plane and will likely average down significantly. Bicep3's differential beamwidth is incoherent across the focal plane, centered close to zero, and is measured with precision similar to that of previous receivers. The level of leakage scales with one power of the beamwidth. Though in previous results we have found differential beamwidth to contribute negligible leakage, if beam map simulations show that this mode is significant for Bicep3, deprojection is a possibility. Differential ellipticity is subtracted in our standard results, and therefore the precision requirements become stronger. We first observe that the measurement uncertainties in Table 2 are lower than those achieved by Bicep2, so we may expect that Bicep3 will be able to subtract leakage from differential ellipticity with higher precision than in previous results. We also note that the pre-subtraction leakage scales as the square of the beam size--Bicep3's beams are \(\sim 75\%\) the size of Bicep2's, so a similar level of inherent differential ellipticity would produce significantly less leakage in Bicep3. Ultimately, the effectiveness of deprojection is simulated explicitly using beam map simulations in which we convolve the *Planck* CMB temperature sky with the individual beam maps and process simulated timestreams just like the real data. We can then deproject these maps which contain only temperature-to-polarization leakage from measured beam mismatch. Based on comparison of the magnitude of the dominant differential pointing in Table [3](#tab:statsvb2){reference-type="ref" reference="tab:statsvb2"}, we expect the inherent leakage from beam mismatch modes to be equal to or lower than that from Bicep2. Though deprojection removes all power corresponding to the coupling of the CMB temperature sky to a second-order expansion of the measured differential beam shape, real beams contain higher-order power which deprojection does not remove. Since the undeprojected beam mismatch residuals are often complicated, vary significantly across the focal plane, and may nontrivially cancel or amplify in the final map, the degree to which they will affect the final \(B\)-mode power spectrum is not immediately obvious. The undeprojected residual level is the final result of the beam map simulations after deprojection. In the Bicep2 final results, we found that after deprojection of relative gain and differential pointing, and subtraction of differential ellipticity, we were able to constrain the residual temperature-to-polarization leakage to an equivalent \(r < 3.0 \times 10^{-3}\). Similar simulations for Bicep3 will be run using the composite beam maps presented above, and we expect the measurement uncertainty on the beam maps to be much lower than that of Bicep2 because of higher signal strength afforded by the new thermal chopper. # CONCLUSIONS {#sect:conclusions} In these proceedings we have presented a preliminary analysis of Bicep3's far field beam performance for the 2016 observing season. Measurements were made during the 2015-2016 deployment season using a new chopped thermal microwave source with a 24" aperture. We find that Bicep3 has a median beamwidth of \(0.167^{\circ}\) (23.6 arcminutes) and present the array-averaged \(B(\ell)\) profile. When detector pairs are fit to 2D elliptical Gaussians, we find that differential pointing, beamwidth, and ellipticity are similar to or lower than those measured for Bicep2 and *Keck Array*. The high-fidelity per-detector composite beam maps we have constructed from 77 beam mapping schedules will allow for explicit simulation of the undeprojected residual temperature-to-polarization leakage expected in the final polarization maps produced by Bicep3. The Bicep3 project has been made possible through support from the National Science Foundation (grant Nos. 1313158, 1313010, 1313062, 1313287, 1056465, and 0960243), the SLAC Laboratory Directed Research and Development Fund, the Canada Foundation for Innovation, and the British Columbia Development Fund. The development of antenna-coupled detector technology was supported by the JPL Research and Technology Development Fund and grants 06-ARPA206-0040 and 10-SAT10-0017 from the NASA ARPA and SAT programs. The development and testing of focal planes were supported by the Gordon and Betty Moore Foundation at Caltech. The computations in these proceedings were run on the Odyssey cluster supported by the FAS Science Division Research Computing Group at Harvard University. Tireless administrative support was provided by Irene Coyle, Kathy Deniston, Donna Hernandez, and Dana Volponi. We are grateful to Samuel Harrison and Hans Boenish as our 2015 and 2016 winterovers, respectively. We thank the staff of the US Antarctic Program and in particular the South Pole Station without whose help this research would not have been possible. We thank our Bicep1, Bicep2, *Keck Array*  and Spider colleagues for useful discussions and shared expertise. [^1]: HD-30 is also used for the *Keck Array* vacuum window and has \(< 2\%\) transmission loss in our band. [^2]: Our transition edge sensor (TES) detectors contain superconductors with different transition temperatures suitable for various loading conditions: titanium for CMB observations and aluminum for the higher loading in calibration measurements. [^3]: This statistic is relatively insensitive to outliers, and would equal \(1 \sigma\) for a Gaussian distribution of measurements.

{'timestamp': '2016-07-18T02:08:19', 'yymm': '1607', 'arxiv_id': '1607.04567', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04567'}

# Introduction {#secIntr} Understanding the nature and structure of the excited states of the nucleon is a key contemporary problem in QCD. The Roper resonance, \(N^*(1440)\), was first deduced from the analysis of \(\pi N\) phase shifts in 1963. However, its structure and nature have aroused interest ever since; it is lighter than the first odd-parity nucleon excitation, \(N^*(1535)\), and has a significant branching ratio into \(N\pi\pi\). Although it is recognized as a well established resonance (four-star ranking in the Review of Particle Physics), the properties of the Roper, such as the mass, width, and decay branching ratios, still suffer large experimental uncertainties. On the theoretical side, there are widely varying models describing the Roper resonance, such as early classical quark models, bag and Skyrme models, dynamically generated by meson-nucleon interactions [@Krehl2000; @Schuetz1998; @Matsuyama2007; @Kamano2010; @Kamano2013; @Hernandez2002a], or a monopole gluonic excitation. However, these descriptions do encounter challenges. For example, predictions of the mass are often too large or predictions for its width are too small. Difficulties are also encountered in explaining its electromagnetic coupling. One expects that lattice QCD simulations will provide unique information concerning the Roper in a finite volume. Current simulation results near the physical quark masses on lattices with spatial length \(L \simeq 3\) fm reveal a \(2S\)-like radial excitation of the nucleon near 1800 MeV, much higher than the infinite volume mass of 1440 MeV. The main task of this paper is to examine the physical phase shifts, inelasticities and pole position associated with the \(N^*(1440)\) Roper resonance and connect these infinite-volume observables to the finite-volume spectrum of lattice QCD. We use the formalism of Hamiltonian effective field theory (HEFT) to achieve this goal and seek an understanding of the observed finite-volume spectrum in the context of empirical scattering observables. The investigations of recent papers have shown how to relate the eigenvalues of a finite-volume Hamiltonian matrix to the spectrum of states observed in lattice QCD. These two papers explored the spectrum of states with the quantum numbers of the \(\Delta(1232)\) resonance and the \(\pi\pi\)-\(K\bar{K}\) system via solutions of the eigen-equation of a finite-volume Hamiltonian matrix. Both papers showed that this Hamiltonian matrix approach is equivalent to the well known Lüscher formulation. Furthermore, Ref.  showed that this method is sufficient for the multi-channel scattering case, where the Lüscher method is more difficult to apply because it needs the phase shifts and inelastic factors in every channel. In contrast, the parameters of the Hamiltonian can be constrained by the empirical phase shifts and inelasticities. As a result, the spectrum is easily obtained in HEFT. This work is a direct application of HEFT in the \(N\frac{1}{2}^+\) multi-channel case including three channels, \(\pi N\), \(\pi\Delta\), and \(\sigma N\). The parameters of the Hamiltonian are fitted to describe the phase shifts and inelasticities of \(\pi N\) scattering up to 1800 MeV. Then, in the finite-volume relevant to lattice QCD, the energy eigenvalues and their associated eigenvectors (describing the wave functions of the eigenstates) are obtained from the Hamiltonian matrix. Both the energy eigenvalues and the eigenstate wave functions are important in understanding the spectrum of the Roper channel obtained in today's lattice QCD simulations. The framework of HEFT is described in Sec. [2](#secFrm){reference-type="ref" reference="secFrm"}. We illustrate how the phase shifts and inelasticities in the infinite volume of Nature are obtained and the manner in which the finite-volume energy eigenstates are calculated. The numerical results and associated discussion are presented in Sec. [3](#secNum){reference-type="ref" reference="secNum"}. Here we present results for three different hypotheses for the internal structure of the Roper. In the first case, the Roper is postulated to have a triquark-like bare or core component with a mass exceeding the resonance mass. This component mixes with attractive virtual meson-baryon contributions, including the \(\pi N\), \(\pi \Delta\), and \(\sigma N\) channels, to reproduce the observed pole position. In the second hypothesis, the Roper resonance is dynamically generated purely from the \(\pi N\), \(\pi \Delta\), and \(\sigma N\) channels. In the third hypothesis, the Roper resonance is coupled to the low-lying bare component associated with the ground state nucleon. Through coupling with the virtual meson-baryon contributions the scattering data and pole position are reproduced. The merit of this third approach lies in its ability to not only describe the observed nucleon energy levels in large-volume lattice QCD simulations but also explain why other low-lying states have been missed in today's lattice QCD results. Finally, a short summary is given in Sec. [4](#secSum){reference-type="ref" reference="secSum"}. # Framework {#secFrm} In this section we provide a short introduction to HEFT and illustrate how it is used in both infinite and finite volumes. The Hamiltonian interactions associated with the \(N\frac{1}{2}^+\) resonance channel are described in Sec. [2.1](#secInt){reference-type="ref" reference="secInt"}. In Sec. [2.2](#secPha){reference-type="ref" reference="secPha"}, the phase shifts and inelasticities are derived from the Hamiltonian model and the pole positions of states are easily obtained via the \(T\)-matrix. The Hamiltonian is then momentum discretised for the finite-volume of the lattice in Sec. [2.3](#secHam){reference-type="ref" reference="secHam"} and the spectrum of energy eigenstates is obtained by solving the Hamiltonian eigen equation. ## Hamiltonian in Channels with \(\mathbf{I(J^P)=\frac12(\frac12^+)}\) {#secInt} The main channels strongly coupled to the Roper are the \(\pi N\), \(\pi \Delta\) and \(\sigma N\) channels. In the rest frame, the Hamiltonian of the system with \(I(J^P)=\frac12(\frac12^+)\) has the following energy-independent form, \[\begin{aligned} H = H_0 + H_I. \label{eq:h} \end{aligned}\] The non-interacting part is \[\begin{aligned} H_0 &=&\sum_{B_0} |B_0\rangle \, m^{0}_{B} \, \langle B_0|+ \sum_{\alpha}\int d^3\vec{k}\nonumber\\ && |\alpha(\vec{k})\rangle \, \left [ \sqrt{m_{\alpha_1}^2 + \vec{k}^2} + \sqrt{m_{\alpha_2}^2 + \vec{k}^2}\, \right ] \langle\alpha(\vec{k})|, \label{eq:h0} \end{aligned}\] where \(B_0\) is the bare baryon (including a bare nucleon \(N_0\) or a bare Roper \(R_0\)) with mass \(m^{0}_{B}\) and \(\alpha\) denotes the included channels \(\pi N\), \(\pi \Delta\), and \(\sigma N\). The masses \(m_{\alpha_1}\) and \(m_{\alpha_2}\) are the masses of the meson and baryon in the channel \(\alpha\), respectively. The interaction Hamiltonian of this system includes two parts \[\begin{aligned} H_I = g + v,\label{eq:hi} \end{aligned}\] where \(g\) describes the vertex interaction between the bare particle and the two-particle channels \(\alpha\) \[\begin{aligned} g &=& \sum_{\alpha\, B_0} \int d^3\vec{k} \left \{ \, |\alpha(\vec{k})\rangle \, G^\dagger_{\alpha, B_0}(k)\, \langle B_0| \right. \nonumber \\ &&\qquad\qquad\quad \left. + |B_0\rangle\, G_{\alpha, B_0}(k)\, \langle \alpha(\vec{k})| \, \right \} \,, \label{eq:int-g} \end{aligned}\] while the direct two-to-two particle interaction is defined by \[\begin{aligned} v = \sum_{\alpha,\beta} \int d^3\vec{k}\; d^3\vec{k}'\, |\alpha(\vec{k})\rangle\, V^{S}_{\alpha,\beta}(k,k')\, \langle \beta(\vec{k}')| \,. \label{eq:int-v} \end{aligned}\] For the vertex interaction between the bare baryon and two-particle channels, the following form is used: \[\begin{aligned} G_{\pi N, B_0}^2(k)&=&\frac{3 g_{B_0\pi N}^2 }{4\pi^2 f^2} \frac{k^2 u_{\pi N}^2(k)}{\omega_\pi(k)},\\ G_{\pi \Delta, B_0}^2(k)&=&\frac{g_{B_0\pi\Delta}^2 }{3\pi^2 f^2} \frac{k^2 u_{\pi \Delta}^2(k)}{\omega_\pi(k)},\\ G_{\sigma N, B_0}^2(k)&=&\frac{g_{B_0\sigma N}^2 }{4\pi^2} \frac{u_{\sigma N}^2(k)}{\omega_\sigma(k)}, \end{aligned}\] where \(f=92.4\) MeV is the pion decay constant, \(\omega_X(k)=\sqrt{k^2+m_X^2}\) is the corresponding energy, and \(u_\alpha(k)\) is the regulator. We consider the exponential form \[u_\alpha(k)=\exp\left (-\frac{k^2}{\Lambda_\alpha^2} \right ) \,,\] where \(\Lambda_\alpha\) is the regularization scale. Although we adopt the exponential form, our main conclusions are not affected if other form factors are used. We have explicitly checked the selection of a dipole form factor \(u_\alpha(k)=\left(1+{k^2}/{\Lambda_\alpha^2}\right)^{-2}\). The phase shifts and inelasticities are fit well and we obtain similar finite-volume results for the three scenarios considered in Sec. [3](#secNum){reference-type="ref" reference="secNum"}. For the two-to-two particle interaction, we introduce the separable potentials for the following five channels \[\begin{aligned} V_{\pi N, \pi N}^S(k,k')&=&g^S_{\pi N}\frac{\bar G_{\pi N}(k)}{\sqrt{\omega_\pi(k)}} \frac{\bar G_{\pi N}(k')}{\sqrt{\omega_\pi(k')}}\, ,\label{eqVspiN}\\ V_{\pi \Delta, \pi \Delta}^S(k,k')&=&g^S_{\pi \Delta}\frac{\bar G_{\pi \Delta}(k)}{\sqrt{\omega_\pi(k)}} \frac{\bar G_{\pi \Delta}(k')}{\sqrt{\omega_\pi(k')}}\,, \label{eqVsPiDelta} \\ V_{\pi N, \pi \Delta}^S(k,k')&=&g^S_{\pi N,\pi \Delta}\frac{\bar G_{\pi N}(k)}{\sqrt{\omega_\pi(k)}} \frac{\bar G_{\pi \Delta}(k')}{\sqrt{\omega_\pi(k')}}\,, \label{eqVPiNPiDel} \\ V_{\sigma N, \sigma N}^S(k,k')&=&g^S_{\sigma N,\sigma N}\frac{\bar G_{\sigma N}(k)}{\sqrt{\omega_\sigma(k)}} \frac{\bar G_{\sigma N}(k')}{\sqrt{\omega_\sigma(k')}}\,, \label{eqVSigmaN}\\ V_{\pi N, \sigma N}^S(k,k')&=&g^S_{\pi N,\sigma N}\frac{\bar G_{\pi N}(k)}{\sqrt{\omega_\pi(k)}} \frac{\bar G_{\sigma N}(k')}{\sqrt{\omega_\sigma(k')}}\,, \label{eqVPiNSigmaN} \end{aligned}\] where \(\bar G_{\alpha}(k)=G_{\alpha, B_0}(k)/g_{B_0 \alpha}\). ## Phase Shift and Inelasticity {#secPha} The T-matrices for two particle scattering can be obtained by solving a three-dimensional reduction of the coupled-channel Bethe-Salpeter equations for each partial wave \[\begin{aligned} T_{\alpha, \beta}(k,k';E)&=&V_{\alpha, \beta}(k,k';E)+\sum_\gamma \int q^2 dq \times \\ &&\hspace{-1.0cm} V_{\alpha, \gamma}(k,q;E)\, \frac{1}{E-\omega_\gamma(q)+i \epsilon}\, T_{\gamma, \beta}(q,k';E) \,, \nonumber \end{aligned}\] where \(\omega_\alpha(k)\) is the center-of-mass energy of channel \(\alpha\) \[\omega_{\alpha}(k)=\sqrt{m_{\alpha_1}^2+k^2}+\sqrt{m_{\alpha_2}^2+k^2},\] and the coupled-channel potential can be calculated from the interaction Hamiltonian \[\begin{aligned} V_{\alpha, \beta}(k,k') &=& \sum_{B_0}\, G^\dag_{\alpha, B_0}(k)\, \frac{1}{E-m_B^0}\, G_{\beta, B_0}(k') \nonumber \\ &&+V^S_{\alpha,\beta}(k,k') \,. \label{eq:lseq-2} \end{aligned}\] With the normalization \(\langle \alpha(\vec{k})\, |\, \beta(\vec{k}^{\,\,'})\rangle = \delta_{\alpha,\beta}\, \delta (\vec{k}-\vec{k}^{\,\,'})\), the S-matrix for \(\pi N \to \pi N\) is related to the T-matrix by \[\begin{aligned} S_{\pi N}(E_{\rm cm})&=&1-2i\pi \frac{\omega_\pi(k_{\rm cm})\, \omega_N(k_{\rm cm})}{E_{\rm cm}}\, k_{\rm cm}\nonumber\\ &&\quad\quad \times T_{\pi N,\pi N}(k_{\rm cm}, k_{\rm cm}; E_{\rm cm}) \,, \end{aligned}\] where \(k_{\rm cm}\) satisfies the on-shell condition \(\sqrt{m_N^2+k_{\rm cm}^2}+\sqrt{m_\pi^2+k_{\rm cm}^2}=E_{\rm cm}\). One can obtain phase shifts \(\delta\) and inelasticities \(\eta\) with \(S_{\pi N}(E_{\rm cm})=\eta \exp(2i\delta)\). In addition to the phase shifts and inelasticities, the pole positions of bound states or resonances can also be obtained by searching for the poles of the T-matrix.\ ## Finite-Volume Matrix Hamiltonian Model {#secHam} We present the formalism of the finite-volume matrix Hamiltonian model by following Refs. . The Hamiltonian \(\mathcal H\) at finite volume is the momentum discretisation of the Hamiltonian \(H\) at infinite volume. It can also be written as a sum of free and interacting Hamiltonians \(\mathcal H=\mathcal H_0+\mathcal H_I\). In the center-of-mass frame, the meson and the baryon in the two-particle states carry the same magnitude of momentum with back-to-back orientation, while the bare baryon is at rest. In the finite periodic volume of the lattice with length \(L\), the momentum of a particle is restricted to \(k_n=2\pi\sqrt{n}/L\), with \(n = n_x^2 + n_y^2 + n_z^2\) such that \(n=0,1,2,\ldots\) In the finite volume, it is convenient to express the Hamiltonian \(\mathcal H\) as a matrix. \(\mathcal H_0\) is a diagonal matrix \[\mathcal H_0={\rm diag}\{m_B^0,\, \omega_{\sigma N}(k_0),~ \omega_{\pi N}(k_1),\, \omega_{\pi\Delta}(k_1),\, \omega_{\sigma N}(k_1),~...\}.\] The corresponding symmetric matrix \(\mathcal H_I\) is where \[\mathcal G_{\alpha,B_0}(k_n)=\sqrt{\frac{C_3(n)}{4\pi}}\, \left (\frac{2\pi}{L}\right )^{3/2}\, G_{\alpha,B_0}(k_n)\,,\] \[\mathcal V^S_{\alpha,\beta}(k_n, k_m)=\frac{\sqrt{C_3(n)\,C_3(m)}}{4\pi}\, \left (\frac{2\pi}{L}\right )^{3}\, V^S_{\alpha,\beta}(k_n, k_m) \,.\] \(C_3(n)\) represents the degeneracy factor for summing the squares of three integers to equal \(n\). One can obtain the eigenstate energy levels on the lattice and analyse the corresponding eigenvector wave functions describing the constituents of the eigenstates with the above Hamiltonian \(\mathcal H\). In addition to the results at physical pion mass, we can also extend the formalism to unphysical pion masses. Using \(m_\pi^2\) as a measure of the light quark masses, we consider the variation of the bare mass and \(\sigma\)-meson mass as \[\begin{aligned} m_B^0(m_\pi^2)&=&m_B^0|_{\rm phy}+\alpha_B^0\, (m_\pi^2-m_\pi^2|_{\rm phy})\,, \\ m_\sigma^2(m_\pi^2)&=&m_\sigma^2|_{\rm phy}+\alpha_\sigma^0\, (m_\pi^2-m_\pi^2|_{\rm phy}), \end{aligned}\] where the slope parameter \(\alpha_B^0\) is constrained by lattice QCD data from the CSSM. In the large quark mass regime where constituent quark degrees of freedom become relevant one expects \(m_\sigma \sim m_\sigma|_{\rm phy} + (2/3)\, \alpha_N^0 \, (m_\pi^2-m_\pi^2|_{\rm phy})\) providing \(\alpha_\sigma^0 \simeq \frac43 \, m_\sigma|_{\rm phy} \, \alpha_N^0\). The nucleon and Delta masses away from the physical point, \(m_N(m_\pi^2)\) and \(m_\Delta(m_\pi^2)\), are obtained via linear interpolation between the corresponding data of lattice QCD. With \(\alpha_N^0 = 1.00~{\rm GeV}^{-1}\), \(\alpha_\sigma^0 \simeq 0.67\). In the following results, we find the \(\sigma N\) channel couples weakly and therefore our conclusions are not sensitive to this value. For the other parameters constrained by experimental data, it is difficult to predict their quark-mass dependence. However, Refs.  show examples where lattice data can be described well without a quark-mass dependence for the couplings. A similar approach has been employed successfully in chiral effective field theory, where one expands in small momenta and masses about the chiral limit. Using fixed couplings, lattice data is described over a wide range of pion masses. In any event, the lightest pion mass considered herein is very close to the physical pion mass. The couplings should not change significantly over this small change in pion mass. # Numerical Results and Discussion {#secNum} ## Fitting the Phase Shift and Inelasticity {#secFitData} Here we examine the phase shifts and inelasticities associated with the \(N^*(1440)\) Roper resonance and connect these infinite-volume observables to the finite-volume spectrum of lattice QCD using HEFT. It is natural to think that the Roper resonance might be dominated by a bare state dressed by meson-baryon states, like the nucleon, but some authors also propose that the Roper may be a dynamically-generated molecular state arising purely from multi-particle meson-baryon interactions. These considerations lead us to explore three hypotheses for the structure of the Roper resonance. In the first case, the Roper is postulated to have a triquark-like bare or core component with a mass exceeding the resonance mass. This component mixes with virtual meson-baryon contributions, including the \(\pi N\), \(\pi \Delta\), and \(\sigma N\) channels, to reproduce the observed pole position. We will refer to this first scenario (Scenario I) as the "bare Roper" scenario. In the second hypothesis, the Roper resonance is dynamically generated purely from the \(\pi N\), \(\pi \Delta\), and \(\sigma N\) channels. We will refer to this second scenario (Scenario II) as the "without bare baryon" scenario. In the third hypothesis, the Roper resonance is composed of the low-lying bare component associated with the ground state nucleon. Through coupling with the virtual meson-baryon contributions the scattering data and pole position are reproduced. We will refer to this third scenario (Scenario III) as the "bare Nucleon" scenario. We fit the experimental data up to centre of mass energies of 1800 MeV. Since the \(\sigma\) meson has a large width and the \(N(\pi\pi)_\texttt{S-wave}\) plays an important role for the inelasticities of the \(\pi N\) channel up to 1450 MeV, we fix the \(\sigma\) mass to a small value of 350 MeV to describe the threshold behavior well. In this way we have included the threshold effects of the \(N(\pi\pi)_\texttt{S-wave}\) channel in an effective manner. The contributions of both \(N(\pi\pi)_\texttt{S-wave}\) and \(N \sigma\) are included in a single effective channel. Similar difficulties are encountered in lattice QCD. Here one needs to include both five-quark momentum-projected \(N \sigma\) and seven-quark momentum-projected \(N \pi \pi\) interpolating fields to separate the \(N(\pi\pi)_\texttt{S-wave}\) and \(N \sigma\) contributions. In the absence of the seven-quark interpolating fields, the \(N \pi \pi\) and \(N \sigma\) contributions will be treated in a similar effective manner, where the combined contributions are treated as a single state. The fitted results for the phase shifts and inelasticities are plotted in Fig. [\[figPSEta\]](#figPSEta){reference-type="ref" reference="figPSEta"}, for the three aforementioned scenarios. The best-fit parameters and the pole positions for each scenario are presented in Table [1](#tabPara){reference-type="ref" reference="tabPara"}. It is interesting to observe that a pole corresponding to the Roper resonance is generated in all three scenarios, whether a bare state is introduced or not. While the imaginary part in Scenarios II and III deviates from the Review of Particle Physics, we note the model is in agreement with the phase shift and inelasticity data.

{'timestamp': '2017-02-16T02:04:50', 'yymm': '1607', 'arxiv_id': '1607.04536', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04536'}

# Introduction and results {#sec:Intro} Consider an \(n\)-variate polynomial of degree at most \(d\): \[p=\sum_{|\alpha| \le d} c_{\alpha} x^{\alpha}\] where \(x:=(x_1,\ldots,x_n)\), \(\alpha \in \ZZ_{\ge 0}^n\), \(|\alpha|:=\sum_i \alpha_i\), \(x^{\alpha}:=\prod_i x_i^{\alpha_i}\), and where each coefficient \(c_\alpha\) is taken from a ground field \(K\). A *determinantal representation* of \(p\) is an \(N \times N\)-matrix \(M\) of the form \[M=A_0 + \sum_{i=1}^n x_i A_i,\] where each \(A_i \in K^{N \times N},\) with \(\det(M)=p\). We call \(N\) the *size* of the determinantal representation. Clearly, since the entries of \(M\) are affine-linear forms in \(x_1,\ldots,x_n\), \(N\) must be at least the degree of \(p\). Determinantal representations of polynomials play a fundamental role in several mathematical areas: from *algebraic geometry* it is known that each plane curve (\(n=2\)) of degree \(d\) over an algebraically closed field \(K\) admits a determinantal representation of size \(d\). Over non-algebraically closed fields, and especially when restricting to symmetric determinantal representations, the situation is much more subtle. For larger \(n\), only certain hypersurfaces have a determinantal representation of size equal to their degree. In *optimisation*, and notably in the theory of *hyperbolic polynomials* , one is particularly interested in the case where \(K=\RR\), \(A_0\) is symmetric positive definite, and the \(A_i\) are symmetric. In this case, the restriction of \(p\) to any line through 0 has only real roots. For \(n=2\) the converse also holds ; for counterexamples to this converse holding for higher \(n\), see. In *complexity theory* a central role is played by Valiant's conjecture that the permanent of an \(m \times m\)-matrix does not admit a determinantal representation of size polynomial in \(m\). Via the *geometric complexity theory* programme this leads to the study of polynomials in the boundary of the orbit of the \(N \times N\)-determinant under the action of the group \(\GL_{N^2}(K)\) permuting matrix entries. Recent developments in this field include the study of this boundary for \(N=3\) and the exciting negative result in that Valiant's conjecture can *not* be proved using occurrence obstructions proposed earlier in. Our motivation comes from *scientific computing*, where determinantal representations of polynomials have recently been proposed for efficiently solving systems of equations. For this application, it is crucial to have determinantal representations not of a *single* polynomial \(p\), but rather of all \(n\)-variate polynomials of degree at most \(d\). Moreover, the representation should be easily computable from the coefficients of \(p\). Specifically, in determinantal representations are constructed for the bivariate case (\(n=2\)) in which the entries of the matrices \(A_0,\ldots,A_n\) themselves *depend affine-linearly* on the coefficients \(c_\alpha\). This is what we call a *uniform determinantal representation* of the generic polynomial \(p\) of degree \(d\) in \(n\) variables; see Section [2](#sec:Problem){reference-type="ref" reference="sec:Problem"} for a precise definition. In applications, the matrix \(M\) is used as input to algorithms in numerical linear algebra that scale unfavourably with \(N\), such as a complexity of \(O(N^6)\). Consequently, we are led to consider the following fundamental question. A construction from shows that for fixed \(n=2\) and \(d \to \infty\) we have \(N^*(2,d) \le \frac{1}{4} \, d^2 + O(d)\); this construction is reviewed in Section [4](#sec:First){reference-type="ref" reference="sec:First"}. We improve the construction from by giving a particularly elegant uniform determinantal representations of bivariate polynomials of size \(2d+1\) in Example [\[ex:2d\]](#ex:2d){reference-type="ref" reference="ex:2d"}, and of size \(2d-1\) in Example [\[ex:2d1\]](#ex:2d1){reference-type="ref" reference="ex:2d1"}. In view of the obvious lower bound of \(d\) this is clearly sharp up to a constant factor for \(d \to \infty\), although we do not know where in the interval \([d,2d-1]\) the true answer lies. We show in Section [8](#sec:Numerics){reference-type="ref" reference="sec:Numerics"} how to use these small determinantal representations of bivariate polynomials for solving systems of equations. Before that, we focus on the asymptotic behaviour of \(N^*(n,d)\) for fixed \(n\) and \(d \to \infty\). In this setting, we derive the following result. We will also compare our results with previous constructions, most notably with those by Quarez, who proves the existence of a symmetric representation of size \({n+\lfloor \frac{d}{2} \rfloor} \choose n\). For fixed \(n\) and \(d \to \infty\), therefore has the asymptotic rate \(\sim d^n\), meaning that the results of this paper represent a clear improvement. For fixed \(d\) and \(n \to \infty\), leads to the asymptotic behavior \(\sim n^{\lfloor d/2 \rfloor}\), which is similar to our bounds; we will discuss more details in Section [9](#sec:Outlook){reference-type="ref" reference="sec:Outlook"}. In Section [2](#sec:Problem){reference-type="ref" reference="sec:Problem"} we formalise the notion of uniform determinantal representations, study their symmetries, and derive some simple properties. In particular, we relate uniform determinantal representations to spaces of singular \(N \times N\)-matrices. In Section [3](#sec:Singular){reference-type="ref" reference="sec:Singular"} we briefly review some of the existing literature on these singular spaces, and we prove that for \(N>4\) there are infinitely many equivalence classes of such objects; this poses an obstruction to a "brute-force" approach towards finding lower bounds on \(N^*(n,d)\). In Section [4](#sec:First){reference-type="ref" reference="sec:First"} we present a first construction, of which however the size is of the order of \(d^n\), rather than \(d^{n/2}\), for \(d \to \infty\). In Section [5](#sec:Second){reference-type="ref" reference="sec:Second"} we give a more efficient construction and prove Theorem [\[thm:Main\]](#thm:Main){reference-type="ref" reference="thm:Main"}. In Section [6](#sec:Small){reference-type="ref" reference="sec:Small"} we give upper bounds on \(N^*(n,d)\) for small \(n\) and \(d\) and determine \(N^*(2,2)\) and \(N^*(3,2)\) exactly. We extend representations from scalar to matrix polynomials in Section [7](#sec:MatPol){reference-type="ref" reference="sec:MatPol"}. In Section [8](#sec:Numerics){reference-type="ref" reference="sec:Numerics"} we give some numerical results that show that for \(n=2\) and small \(d\) we get a competitive method for computing zeros of polynomials systems. Finally, in Section [9](#sec:Outlook){reference-type="ref" reference="sec:Outlook"} we summarise our main conclusions and collect some questions that arise naturally from our work. # Problem formulation and symmetries {#sec:Problem} In this section we give a formal definition of uniform determinantal representations, and introduce a group that acts on such representations. We also show that a uniform determinantal representation gives rise to a vector space consisting entirely of singular matrices; such spaces are the topic of next section. Let \(K\) be a field and fix \(d,n \in \ZZ_{\ge 0}\). Let \(F_d\) denote the polynomials of degree at most \(d\) in the polynomial ring \(K[x_1,\ldots,x_n]\). Furthermore, let \(p_{n,d}\) be the generic polynomial of that degree, i.e., \[\label{eq:polnd} p_{n,d}=\sum_{|\alpha| \le d} c_\alpha x^\alpha,\] where \(x:=(x_1,\ldots,x_n)\), \(\alpha \in \ZZ_{\ge 0}^n\), \(|\alpha|:=\sum_i \alpha_i\), \(x^{\alpha}:=\prod_i x_i^{\alpha_i}\), and where we consider \(c_\alpha\) as a variable for each \(\alpha\). To be explicit, we require each entry of \(M\) to be a \(K\)-linear combination of the monomials \(1,x_i,c_\alpha,c_\alpha x_i,\ (i=1,\ldots,n, |\alpha| \le d)\). This means that we can decompose \(M\) as \(M_0+M_1\), where \(M_0\) contains all terms in \(M\) that do not contain any \(c_\alpha\), and where \(M_1\) contains all terms in \(M\) that do. We will use the notation \(M=M_0+M_1\) throughout the paper. When \(n\) and \(d\) are fixed in the context, we will also speak of a uniform determinantal representation without reference to \(p_{n,d}\). Our ultimate aim is to determine the following quantity. This minimal size could potentially depend on the ground field \(K\), but the bounds that we will prove do not. Note that in the definition of \(N^*(n,d)\) we do not allow terms in \(M\) of degree strictly larger than one in the \(c_\alpha\). Relaxing this condition to polynomial dependence on the \(c_\alpha\) might affect the exact value of \(N^*\), but it will not affect our bounds---see Remark [\[re:Poly\]](#re:Poly){reference-type="ref" reference="re:Poly"}. Given a uniform determinantal representation \(M\) of size \(N\), and given matrices \(g,h\) in \(\SL_N(K)\), the group of determinant-one matrices with entries in \(K\), the matrix \(gMh^{-1}\) is another uniform determinantal representation of \(p_{n,d}\). In this manner, the group \(\lieg{SL}_N(K) \times \lieg{SL}_N(K)\) acts on the set of uniform determinantal representations of \(p_{n,d}\). Moreover, there exist further symmetries, arising from affine transformations of the \(n\)-space. Recall that these transformations form the group \(\AGL_n(K)=\GL_n(K) \ltimes K^n\) generated by invertible linear transformations and translations. The statement of this lemma is empty without making the action explicit, as we do in the proof. The action of the affine group will be used in Section [6](#sec:Small){reference-type="ref" reference="sec:Small"} to determine the exact value of \(N^*(n,2)\) for \(n=2\) and \(3\). We now turn our attention to the component \(M_0\) of a uniform determinantal representation \(M\). Here, as in the rest of this paper, by the product of two spaces of polynomials we mean the \(K\)-linear span of all the products. # Spaces of singular matrices {#sec:Singular} Let \(M=M_0+M_1\) be a uniform determinantal representation of \(p_{n,d}\). Writing \(M_0=B_0+\sum_{i=1}^n x_i B_i\), Lemma [\[lm:SingSpace\]](#lm:SingSpace){reference-type="ref" reference="lm:SingSpace"} implies that the linear span \(\langle B_0,\ldots,B_n \rangle_K \subseteq K^{N \times N}\) consists entirely of singular matrices (and indeed that this remains true when extending scalars from \(K\) to an extension field). There is an extensive literature on such *singular matrix spaces*; see, e.g., and the references therein. The easiest examples are the following. Given any two subspaces \(U,V \subseteq K^N\) with \(\dim V=-1+\dim U\), the space of all matrices which map \(U\) into \(V\) (acting on row vectors) is a compression space with witness \(U\). It is easy to see that these spaces are inclusion-wise maximal among all singular spaces. If \(\mathcal{A}\) is a singular matrix space, then so is \(g \mathcal{A} h^{-1}\) for any pair \((g,h) \in \GL_N(K) \times \GL_N(K)\). We call the latter space *conjugate* to the former. For large \(N\) it seems impossible to classify maximal singular matrix spaces. The construction above already gives an infinite number of conjugacy classes, but there are many other sources of examples. For instance, for infinitely many \(N\) there exists a maximal singular matrix space in \(K^{N \times N}\) of constant dimension \(8\), at least if we assume that \(K\) has characteristic 0. On the other hand, if the singular matrix space \(\mathcal{A}\) has dimension at least \(N^2-N\), then it is a compression space with either a one-dimensional witness or all of \(K^N\) as witness (and hence of dimension exactly \(N^2-N\)). A sharpening of this result is proved in (see also ). It should be noted that in many cases not even the dimension of such singular matrix spaces is known, for fixed values of the size and rank of the matrices. There is a considerable body of work devoted to giving lower and upper bounds for such dimensions, both in the case of bounded and constant rank, but these bounds are rarely sharp, see, among many other references, and the more recent works on skew-symmetric matrices of constant rank. Hence the fact that \(M_0\) represents a singular matrix space of dimension (at most) \(n+1\) does not much narrow down our search for good uniform determinantal representations, except in small cases discussed in Section [6](#sec:Small){reference-type="ref" reference="sec:Small"}. However, for our constructions in Sections [4](#sec:First){reference-type="ref" reference="sec:First"} and [5](#sec:Second){reference-type="ref" reference="sec:Second"} we will only use compression spaces where the witness has dimension 1 or about \(\frac{1}{2} N\), respectively; and our lower bounds on \(N^*(n,d)\) are independent of the literature on singular matrix spaces. # A first construction {#sec:First} In this section we restrict our attention to determinantal representations \(M=M_0+M_1\) where \(M_0\) represents a compression space with a one-dimensional witness (or, dually by transposition, with a full-dimensional witness). Under this assumption we will prove quite tight bounds on the minimal size of a uniform matrix representation. The following fundamental notion will be used throughout below. Note that this implies that \(V_0=\langle 1 \rangle\). We borrow the terminology from the theory of border bases, where a set \(S\) of monomials is called connected to \(1\) if \(1 \in S\) and each nonconstant monomial in \(S\) can be divided by some variable to obtain another monomial in \(S\). The linear span of \(S\) is then connected to \(1\) in our sense. Translating monomials to their exponent vectors, we will call a subset \(S\) of \(\ZZ_{\ge 0}^n\) *connected to \(0\)* if it contains 0 and for each \(\alpha \in S \setminus \{0\}\) there exists an \(i\) such that \(\alpha-e_i \in S\), where \(e_i\) is the \(i\)-th standard basis vector. Let \(V\) be a finite-dimensional subspace of \(K[x_1,\ldots,x_n]\) connected to 1. Choose a \(K\)-basis \(f_1,\ldots,f_m\) of \(V\) whose total degrees increase weakly. For each \(i=2,\ldots,m\) write \[f_i=\sum_{j<i} \ell_{ij} f_j\] for suitable elements \(\ell_{ij} \in F_1\). Let \(M_V\) be the \((m-1) \times m\)-matrix whose \(i\)th row equals \[(-\ell_{i1},-\ell_{i2},\ldots,-\ell_{i,i-1},1,0,\ldots,0).\] Note that \(M_V\) depends on the choice of basis, but we suppress this dependence in the notation, since the property of \(M_V\) in the next lemma does not depend on the choice of basis. We can now formulate our first general construction. This generalises a construction from to the multivariate case. This is the construction of, which shows that there exists a uniform determinantal representation of the generic bivariate polynomial of degree \(d\) of size \(\frac{1}{4} \, d^2 + O(d)\) as \(d \to \infty\). \(\clubsuit\) Now let \(\Delta_d\) be the simplex in \(\RR^n\) with vertices \(0,de_1,\ldots,de_n\), for \(i=1,\ldots,n\) let \(S_i\) be the set of lattice points in \(\Delta_d\) that have \(i\)th coordinate zero, and set \(S_0:=(\ZZ \times \Lambda) \cap \Delta_d\). Define \[S:=S_1 \cup S_2 \cup \ldots \cup S_n \cup S_0,\] a subset of the lattice points in \(\Delta_d\). We claim that \(S\) is connected to 0. Indeed, for each \(i=1,\ldots,n\) the set \(S_i\) is connected to \(0\), and from each point \(\alpha\) in \(S_0\) one can walk within \(S_0\) to \(S_1\) by subtracting \(\alpha_1\) times an \(e_1\). Next, we claim that for each \(\alpha \in \Delta_d \cap \ZZ^n\) there exists a \(\beta \in S\) with \(\alpha-\beta \in \{0,e_2,\ldots,e_n\}\). Indeed, there is a (unique) \(\beta'\) with this property in \(\ZZ \times \Lambda\). If this \(\beta'\) has nonnegative entries, then set \(\beta:=\beta' \in S_0\). Otherwise, \(\alpha\) itself has a zero entry, say on the \(i\)th position, and we set \(\beta:=\alpha \in S_i\). Furthermore, for \(i=1,\ldots,n\) the set \(S_i\) contains \(O(d^{n-1})\) vertices, and \(S_0\) contains \(\frac{1}{n} \cdot \frac{1}{n!} \, d^n + O(d^{n-1})\) vertices. This concludes the construction--note that in the construction of Proposition [\[prop:Cons1\]](#prop:Cons1){reference-type="ref" reference="prop:Cons1"} the matrix \(M_0\) has a zero row, so that it represents a compression space with a one-dimensional witness. For sharpness, assume that \(M=M_0+M_1\) is a uniform determinantal representation of size \(N\) such that the singular matrix space represented by \(M_0\) is a compression space with a one-dimensional witness. After a choice of basis of \(K^n\), we may assume that the first row of \(M_0\) is identically zero; write \(M_0=[0|M_0']^T\) accordingly. Let \(u\) be the first row of \(M_1\) and write \(M_1=[u|M_1']^T\). Then we have \[p=\sum_{|\alpha| \le d} c_\alpha x^\alpha = \det[u|M_0'+M_1'].\] Let \(D_1,\ldots,D_N\) denote the \((N-1) \times (N-1)\) subdeterminants of \(M_0'\). By Lemma [\[lm:Dets\]](#lm:Dets){reference-type="ref" reference="lm:Dets"}, the space \(V\) spanned by these satisfies \(F_1 \cdot V \supseteq F_d\). This already gives a lower bound of \(V\) equal to \(d^n/((n+1)n!) + O(d^{n-1})\). To improve the \(n+1\) in the denominator into an \(n\), we observe that by Cramer's rule, the map \[F_1^N \to K[x_1,\ldots,x_n],\ (\ell_1,\ldots,\ell_N) \mapsto \sum_i (-1)^i \ell_i D_i\] has every column of \(M_0'\) in its kernel. These columns are linearly independent over \(K\) (indeed over \(K(x_1,\ldots,x_n)\); see Lemma [\[lm:SingSpace\]](#lm:SingSpace){reference-type="ref" reference="lm:SingSpace"}). We conclude that \[\label{eq:Ineq} N \cdot \dim F_1-(N-1) \ge \dim F_d,\] so that \[N \ge ((\dim F_d)-1)/n = d^n/(n \cdot n!) + O(d^{n-1}),\] as desired. ◻ Example [\[ex:2d\]](#ex:2d){reference-type="ref" reference="ex:2d"} generalises as follows. We will now prove our main theorem. For the upper bound, we first give a simple construction for even \(n\). For odd \(n\), a trickier analysis is needed (which also applies in the even case); see below. We now give a construction that works for all \(n>2\), for which we thank Aart Blokhuis. The proofs above use the classification of spaces of small singular matrices in an essential manner, as well as the action of \(\AGL_n(K)\) on uniform determinantal representations. We conjecture that \(N^*(4,2)=5\), and that this can still be proved in the same manner, using the classification of \(4 \times 4\)-singular matrix spaces from. But as Theorem [\[thm:Sing5\]](#thm:Sing5){reference-type="ref" reference="thm:Sing5"} shows, fundamentally new ideas will be needed to prove lower bounds in larger situations. For some pairs of small \(n\) and \(d\) we now give the smallest uniform representations that we have been able to find. For the constructions we use Proposition [\[prop:Cons2\]](#prop:Cons2){reference-type="ref" reference="prop:Cons2"} with subspaces \(V,W\subseteq K[x_1,\ldots,x_n]\) spanned by the monomials and connected to \(1\). First, we give in Table [1](#tab:minnd){reference-type="ref" reference="tab:minnd"} the minimal sizes known to us of uniform determinantal representations for some small values of \(n\) and \(d\). # Matrix polynomials {#sec:MatPol} Suppose that we have a uniform representation \(M\) of \(p_{n,d}\) as in [\[eq:polnd\]](#eq:polnd){reference-type="eqref" reference="eq:polnd"}, and write \[\label{eq:linM} M=M_0+M_1=M_0 + \sum_{|\alpha| \le d} c_\alpha M_\alpha,\] where each \(M_\alpha \in F_1^{N \times N}\). Now consider the matrix polynomial (cf. [\[eq:polnd\]](#eq:polnd){reference-type="eqref" reference="eq:polnd"}) \[P_{n,d}=\sum_{|\alpha| \le d} x^\alpha C_\alpha,\] where \(C_\alpha\) is a \(k\times k\) matrix. We will show that under certain assumptions we can construct from \(M\) a matrix \(\widetilde{M}\) that represents \(P_{n,d}\) in the sense that \(\det(\widetilde M)=\det(P_{n,d})\). We obtain \(\widetilde M\) from \(M\) in the following way. Each element of the form \(\alpha +\beta x +\gamma y\) is replaced by the \(k\times k\) matrix \((\alpha+\beta x + \gamma y)I_k\), where \(I_k\) is the \(k \times k\) identity, and each \(c_\alpha\) is replaced by the matrix \(C_\alpha\). Unfortunately, not all uniform determinantal representations induce a determinantal representation of a general matrix polynomial in this manner. As a counterexample, let \(M\) be such a uniform determinantal representation of the polynomial \(p_{n,d}\), \(|\alpha|,|\beta| \le d\), and construct a representation of larger size \[M'=\begin{bmatrix} M & 0\\ 0 & N\\ \end{bmatrix}, \quad\quad \hbox{with}\quad N= \left[ \begin{array}{rrrr} 0 & c_{\alpha} & c_{\beta} & 1\\-c_{\alpha} & 0 & 1 & 0\\-c_{\beta} &-1 & 0 & 0\\-1 & 0 & 0 & 1 \end{array} \right].\] Then \(\det(M')=\det(M)\det(N)= p_{n,d}(1+c_{\alpha}c_{\beta}-c_{\beta}c_{\alpha})=p_{n,d}\), but \(\widetilde{M'}\) is not a representation for the matrix polynomial \(P_{n,d}\) as the coefficient matrices \(C_{\alpha}\) and \(C_{\beta}\) do not commute in general. This motivates the following definition. We speculate that each minimal uniform representation gives rise to a representation for a matrix polynomial. # Numerical experiments {#sec:Numerics} Recently, a new numerical approach for computing roots of systems of bivariate polynomials was proposed in. The main idea is to treat the system as a two-parameter eigenvalue problem using determinantal representations. Suppose that we are looking for roots of a system of bivariate polynomials \[\label{eq:syspol} \begin{matrix}p=\sum_{i+j \le d_1} \alpha_{ij} x^i y^j=0,\\[0.2em] q=\sum_{i+j \le d_2} \beta_{ij} x^i y^j=0,\end{matrix}\] where \(p\) and \(q\) are polynomials of degree \(d_1\) and \(d_2\) over \(\CC\). Let \(P=A_0 + x A_1 + y A_2\) and \(Q=B_0 + x B_1 + y B_2\), where \(A_0,A_1,A_2 \in \CC^{N_1 \times N_1}\) and \(B_0,B_1,B_2 \in \CC^{N_2 \times N_2}\), with \(\det(P)=p\) and \(\det(Q)=q\), be determinantal representations of \(p\) and \(q\), respectively. Then a root \((x,y)\) of [\[eq:syspol\]](#eq:syspol){reference-type="eqref" reference="eq:syspol"} is an eigenvalue of the two-parameter eigenvalue problem \[\label{eq:twopar} \begin{matrix} (A_0+x A_1+y A_2) \, u=0,\\[0.2em] (B_0+x B_1+y B_2) \, v=0, \end{matrix}\] where \(u\in\CC^{N_1}\) and \(v\in\CC^{N_2}\) are nonzero vectors. The standard way to solve [\[eq:twopar\]](#eq:twopar){reference-type="eqref" reference="eq:twopar"} is to consider a joint pair of generalized eigenvalue problems \[\label{eq:twopardelta} \begin{matrix} (\Delta_1-x \Delta_0) \, w=0,\\[0.2em] (\Delta_2-y \Delta_0) \, w=0, \end{matrix}\] where \[\Delta_0=A_1\otimes B_2-A_2\otimes B_1,\quad \Delta_1=A_2\otimes B_0-A_0\otimes B_2,\quad \Delta_2=A_0\otimes B_1-A_1\otimes B_0,\] and \(w=u\otimes v\). In this particular application we can expect that the pencils in [\[eq:twopardelta\]](#eq:twopardelta){reference-type="eqref" reference="eq:twopardelta"} are singular, i.e., \(\det(\Delta_1-x\Delta_0)\equiv 0\) and \(\det(\Delta_1-y\Delta_0)\equiv 0\). Namely, by Bézout's theorem a generic system [\[eq:syspol\]](#eq:syspol){reference-type="eqref" reference="eq:syspol"} has \(d_1d_2\) solutions, while a generic problem [\[eq:twopar\]](#eq:twopar){reference-type="eqref" reference="eq:twopar"} has \(N_1N_2\) eigenvalues. Unless \((d_1,d_2)=(N_1,N_2)\), both pencils in [\[eq:twopardelta\]](#eq:twopardelta){reference-type="eqref" reference="eq:twopardelta"} are singular. In this case we first have to apply the staircase algorithm from to extract the finite regular eigenvalues. The method returns smaller matrices \(\widetilde \Delta_0\), \(\widetilde \Delta_1\), and \(\widetilde \Delta_2\) (of size \(d_1d_2\times d_1d_2\) for a generic [\[eq:syspol\]](#eq:syspol){reference-type="eqref" reference="eq:syspol"}) such that \(\widetilde \Delta_0\) is nonsingular and \({\widetilde \Delta_0}^{-1}\widetilde \Delta_1\) and \({\widetilde \Delta_0}^{-1}\widetilde \Delta_2\) commute. From \[\begin{matrix} (\widetilde\Delta_1-x \widetilde\Delta_0) \, \widetilde w=0,\\[0.2em] (\widetilde\Delta_2-y \widetilde\Delta_0) \, \widetilde w=0, \end{matrix}\] we compute the eigenvalues \((x,y)\) using a variant of the QZ algorithm and thus obtain the roots of [\[eq:syspol\]](#eq:syspol){reference-type="eqref" reference="eq:syspol"}. The above approach is implemented in the Matlab package BiRoots together with the two determinantal representations from. The first one, to which we refer as `Lin1`, is a uniform one from Example [\[ex:lin1\]](#ex:lin1){reference-type="ref" reference="ex:lin1"} of size \(\frac{1}{4} \, d^2+ O(d)\) for a polynomial of degree \(d\). The second one, which we refer to as `Lin2`, is not uniform and involves some computation to obtain a smaller size \(\frac{1}{6} \, d^2+ O(d)\). Although the construction of `Lin2` is more time consuming, this pays off later, when the staircase algorithm is applied to [\[eq:twopardelta\]](#eq:twopardelta){reference-type="eqref" reference="eq:twopardelta"}. Table [3](#tbl:compare1){reference-type="ref" reference="tbl:compare1"} shows the sizes of determinantal representations for polynomials of small degree. As expected, the new uniform determinantal representation of size \(2d-1\), to which we refer as `MinUnif`, returns smaller matrices, which reflects later in faster computational times. It is also important that `Lin1` and `MinUnif` return real matrices for polynomials with real coefficients, which is not true for `Lin2`. It was reported in that the determinantal representation approach for solving systems of bivariate polynomials is competitive for polynomials of degree 9 or less. As we show below, the new uniform representation `MinUnif` extends this to degree 15 and, in addition, performs better than the existing representations for polynomials of degree 6 or more. In the approach was compared numerically to the following state-of-the art numerical methods for polynomial systems: `NSolve` in Mathematica 9 , BertiniLab 1.4  running Bertini 1.5 , NAClab 3.0 , and `PHCLab` 1.04  running PHCpack 2.3.84, which turned out as the fastest of these methods. To show the improved performance of the new determinantal representation, we compare `MinUnif` to `Lin1`, `Lin2`, and `PHCLab` in Table [4](#tbl:two){reference-type="ref" reference="tbl:two"}. For each \(d\) we run the methods on the same set of 50 real and 50 complex random polynomial systems of degree \(d\) and measure the average time. For `Lin1` and `MinUnif`, where determinantal representations have real matrices for real polynomials, we report separate results for polynomials with real and complex coefficients. The timings for `Lin1` and `Lin2` are given only for \(n\le 10\) as for larger \(n\) these two linearisations are no longer competitive. Of course, the computational time is not the only important factor, we also have to consider the accuracy and reliability. In each step of the staircase algorithm a rank of a matrix has to be estimated numerically, which is a delicate task. After several steps it may happen that the gap between the important singular values and the meaningful ones that should be zero in exact computation, virtually disappears. In such case the algorithm fails and does not return any roots. As the number of steps in the staircase algorithm increases with degree of the polynomials, such problems occur more often for polynomials of large degree. A heuristic that usually helps in such cases is to apply the algorithm on a transformed system \[\begin{aligned} \widetilde p &:= \ph{-}c p + s q = 0,\\ \widetilde q &:=-s p + c q = 0 \end{aligned}\] for random \(c\) and \(s\) such that \(c^2+s^2=1\). As this transformation does not change the conditioning of the roots, we can conclude that the difficulties with the staircase algorithm are not directly related to the conditioning. The trick does not work every time, and it seems that for some systems the only way to make the determinantal representation approach to work is to increase the machine precision. We can apply the same approach to systems of polynomials in more than two variables. However, since the size of the corresponding \(\Delta\) matrices is the product of sizes of all representations, this is competitive only for \(n=3\) and \(d\le 3\). For a comparison, if we have a system of three polynomials in three variables of degree \(3\), then the size of the \(\Delta\) matrices is \(343\times 343\). For degree \(4\) the size increases to \(1000\times 1000\) and PHCpack is faster. Finally, for \(n=4\) and the smallest nontrivial \(d=2\) we already get \(\Delta\) matrices of size \(625\times 625\) and the method is not efficient. # Outlook {#sec:Outlook} We have introduced *uniform determinantal representations*, which rather than representing a single polynomial as the determinant of a matrix of affine-linear forms, represent all polynomials of degree at most \(d\) in \(n\) variables as such a determinant. We have seen that in the bivariate case, these determinantal representations are useful for numerically solving bivariate systems of equations; and in the general multivariate case we have determined, up to constants, the asymptotic behaviour of \(N^*(n,d)\), the minimal size of such a representation, for \(n\) fixed and \(d \to \infty\). We now summarise several results that have been shown in the paper. - For fixed \(n\) and \(d \to \infty\), \(N^*(n,d) \sim d^{n/2}\), see Theorem [\[thm:Main\]](#thm:Main){reference-type="ref" reference="thm:Main"}. This is a noticeable improvement on, where an asymptotic rate of \(N^*(n,d) \sim d^n\) is shown, with the remark that the representation in are symmetric. However, symmetry currently cannot be exploited by methods that compute roots of multivariate polynomial systems. - For fixed odd \(d\) and \(n \to \infty\), \(N^*(n,d) \sim n^{(d-1)/2}\), which is the same rate as Quarez, who also manages to get symmetric representations. For fixed even \(d\) and \(n \to \infty\), \(N^*(n,d) \lsim n^{d/2}\), which again is the same rate as in. However, we have a slightly smaller lower bound for the asymptotic rate of \(n^{(d-1)/2}\). - Tables [1](#tab:minnd){reference-type="ref" reference="tab:minnd"} and [2](#tab:wv){reference-type="ref" reference="tab:wv"} give constructions for the smallest representations that we have been able to find for some small values of \(n\) and \(d\). - \(N^*(n,2) \le n+1\); cf. Table [1](#tab:minnd){reference-type="ref" reference="tab:minnd"}. - \(N^*(n,3) \le 2n+1\); cf. Table [1](#tab:minnd){reference-type="ref" reference="tab:minnd"}. - \(N^*(n,4) \le \frac{1}{6} n^2+O(n)\); see Example [\[ex:turan\]](#ex:turan){reference-type="ref" reference="ex:turan"}. - \(N^*(2,d) \le 2d-1\); cf. Table [1](#tab:minnd){reference-type="ref" reference="tab:minnd"}. Note that this result satisfies Dixon  up to an asymptotic factor 2 whereby no computations are necessary for the determinantal representation. In particular, it is a major improvement on the \(\sim \frac{1}{4} d^2\) of. - Due to the smaller sizes of the representations, the numerical approach for bivariate polynomials (\(n=2\)) is competitive to (say) Mathematica for degree \(d\) up to \(d \approx 15\) (see Section [8](#sec:Numerics){reference-type="ref" reference="sec:Numerics"}); this in contrast to \(d \approx 9\) as obtained in. - Under some conditions, the results carry over to the case of matrix coefficients (see Section [7](#sec:MatPol){reference-type="ref" reference="sec:MatPol"}). There are still many interesting open questions, both of intrinsic mathematical interest and of relevance to polynomial system solving. First of all, in a situation where the degree \(d\) is fixed and the number \(n\) of variables grows, what is the asymptotic behaviour of \(N^*(n,d)\)? Although we expect that the above inequalities are equalities, whether this holds is still open. For general fixed \(d\), the proof of Theorem [\[thm:Main\]](#thm:Main){reference-type="ref" reference="thm:Main"} yields a lower bound which is a constant (depending on \(d\)) times \(n^{(d-1)/2}\). For odd \(d\) we obtain a matching upper bound (with a different constant) by using Proposition [\[prop:Cons2\]](#prop:Cons2){reference-type="ref" reference="prop:Cons2"} with \(V=W=F_{(d-1)/2}\). However, for even \(d\) we only know how to obtain \(O(n^{d/2})\). We remark that this latter is (up to a constant) the same bound as obtained in for *symmetric* uniform representations. Second, in the case of fixed \(n\) and varying \(d\) studied in this paper, what are the best constants in Theorem [\[thm:Main\]](#thm:Main){reference-type="ref" reference="thm:Main"}? More specifically, for fixed \(n\), does \(\lim_{d \to \infty} \frac{N^*(n,d)}{d^{n/2}}\) exist, and if so, what is its value? Third, how can our techniques for upper bounds and lower bounds be further sharpened? Can singular matrix spaces *other* than compression spaces be used to obtain tighter upper bounds (constructions) on \(N^*(n,d)\)? Can the action of the affine group be used more systematically to find lower bounds on \(N^*(n,d)\)? Fourth, is it true that each minimal uniform representation gives rise to a representation of the corresponding matrix polynomial (cf. Section [7](#sec:MatPol){reference-type="ref" reference="sec:MatPol"})? Finally, we have restricted our attention to matrices that, apart from being affine-linear in \(x_1,\dots,x_n\), are also affine-linear in the coefficients \(c_\alpha\). Our proofs give the same asymptotic behaviour (with different constants) if we require, in addition, that no quadratic terms \(c_\alpha x_i\) may occur in \(M\). If, instead, we relax the condition that \(M\) be affine-linear in the \(c_\alpha\) to a *polynomial* dependence on the \(c_\alpha\), then the same bounds still apply; see Remark [\[re:Poly\]](#re:Poly){reference-type="ref" reference="re:Poly"}. But what if we relax this to *rational* dependence of \(M\) on the \(c_\alpha\)? Given that \(p_{n,d}\) is only linear in the \(c_\alpha\) it seems unlikely that allowing \(M\) to be rational in the \(c_\alpha\) we would gain anything, but we currently do not know how to formalise this intuition. On the other hand, in cases where a (non-uniform) determinantal representation of size \(d\) is known to exist for every (or sufficiently general) polynomials of degree \(d\) in \(d\) variables (e.g., in the case of plane curves), it follows that this representation can be chosen to have entries *algebraic* in the \(c_\alpha\). This observation rules out approaches aimed at proving lower bounds in a too general setting.

{'timestamp': '2016-07-19T02:07:22', 'yymm': '1607', 'arxiv_id': '1607.04873', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04873'}

# Introduction The dynamics of electrolytic solutions is essential in the description of most most processes in (bio)chemistry and soft matter physics. This includes electrostatic screening via the formation of double layers, electro-osmotic flow, electrophoresis, and self-electrophoresis. Industrial applications range from DNA sequencing and oil recovery to the detection, separation, and characterization of analyte molecules. The coupled occurrence of diffusion, hydrodynamics, and electrostatics, often referred to as electrokinetics or electrohydrodynamics, gives rise to complex physical behavior. Electrokinetic processes involve dynamics on vastly different length scales with double layers typically having a nanometer scale in aqueous solution and system geometries often ranging from several micrometers to millimeters. The small length scales make it difficult to access the nanoscopic details in experiment, while the large discrepancy in length scales poses a challenge to simulation methods. Existing methods that solve the electrokinetic equations numerically are based on a variety of algorithms for hydrodynamics, generally falling into two classes. Firstly, particle-based algorithms, which include dissipative particle dynamics (DPD) and multi-particle collision dynamics (MPCD). Secondly, continuum (grid-based) algorithms, which encompass methods like finite elements (FEM), the boundary element method (BEM), finite differences, finite volumes, electrokinetics based on lattice Boltzmann (LB), and the smooth profile method (SP). For additional details, we refer to the overviews given in Refs. . The particle-based methods solve the full time-dependent problem and are intrinsically able to include mobile charged colloids and macromolecules. The downside of these methods is that solute ions, water, and macromolecules are resolved on the same scale. This sets a low limit for the maximum size of the simulation domain, given current computational abilities. Continuum solvers allow for the study of systems with much larger length scales because the ions are not resolved. In general, introducing moving boundaries in numeric solvers for the continuum electrokinetic equations poses a more challenging problem than for their particle-based counterparts. Tackling this issue, however, is well worth the effort, as it grants access to the wide variety of physical problems beyond those that can be studied via Galilean transformation to the rest frame of the charged object, an approach that is not possible in general. Moving particles will permit the study of interactions of multiple (self-)electrophoretically driven colloids, as well as their interactions with any stationary boundaries in the system. When it comes to resolving large system sizes, the FEM method has proven itself to be a very efficient solver for the stationary electrokinetic equations. FEM has been successfully applied in arbitrary geometries and for a diverse range of physical systems, including: nanopores, self-electrophoresing colloids, and electro-osmotic pumping. The strength of the FEM lies in its ability to adaptively increase the resolution in regions with high gradients, i.e., the electric double layers. The downside is that the treatment of time-dependent boundary positions is computationally exceptionally expensive, due to the constant need for remeshing. However, there are examples of its application to phoretic problems with moving boundaries. An alternative to FEM was proposed by. Their lattice electrokinetics method (EK) combines finite differences and finite volumes to solve the diffusion-advection of solute and the electrostatics problem. For hydrodynamics, they exploit the computational efficiency of the LB method. Unlike earlier methods that aimed to achieve a similar level of description, EK propagates the solute by considering solute fluxes between neighboring lattice cells instead of concentrations in lattice cells, thus ensuring local mass conservation and reducing discretization artifacts. Furthermore, EK, like LB, is a local algorithm, with the exception of the electrostatics solver, making it generally more efficient than FEM and other solvers, which are fully nonlocal. Of specific interest is EK's ability to resolve the particles' surfaces as boundary conditions on the grid. This allows for the study of extended objects and straightforward incorporation of local variation of the surface properties, which is, e.g., useful in studying self-propelled particles. Extending this method further with dynamically-adaptive grid refinement, a common practice in LB, will constitute an algorithm fully capable of directly modeling microfluidic systems. The EK method has previously been used to study problems such as electrophoresis and sedimentation of single charged spherical colloids. Other applications include charge transport in porous media and the translocation of DNA molecules through nanopores, where this method can tremendously reduce computational effort when compared to particle-based methods, and thus enable one to reach experimentally relevant length and time scales. In this manuscript, we present an extension of the lattice electrokinetics method proposed by to include moving boundaries. For the underlying LB method, a moving boundary scheme is already widely employed for moving particles. To adapt the EK method for moving boundaries, we first introduce solute mass and charge conservation by displacing solute from cells that are claimed by a moving particle. We subsequently add a partial volume smoothing to reduce the effect of a cell being claimed in a single time step. We do so by incorporating a term in the solute flux calculation that depends on the amount of volume in a cell that is overlapped by a particle. This allows for a gradual expulsion of solute from a cell as it is increasingly overlapped. We validate our method by considering the problem of electrophoresis of spherical colloids in an external field. We observe that the moving boundary method is capable of reproducing results using the Capuani method with fixed boundaries (in the co-moving frame). Moreover, we observe that the smoothing term is necessary to prevent strong variations in the particle velocity due to the grid discretization---it reduces the oscillations by more than a factor of \(10\). The added computational cost due to the moving boundaries amounts to approximately a \(10\%\) increase per time step, compared to the stationary boundary algorithm of. The method we describe retains the flexibility and efficiency of the LB and EK methods, while granting the EK method access to an entire class of problems where the relative motion of multiple particles is of relevance. The remainder of this paper is laid out as follows: In [2](#sec:equations){reference-type="ref" reference="sec:equations"}, we give the governing equations. In [3](#sec:numerical){reference-type="ref" reference="sec:numerical"}, we introduce the numerical methods, including a novel method for coupling moving particles with EK. In [4](#sec:validation){reference-type="ref" reference="sec:validation"}, we validate the new method by comparing equivalent electrophoretic and electro-osmotic systems. We conclude in [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}. # The Electrokinetic Equations {#sec:equations} In this section, we introduce the electrokinetic equations for bulk electrolytes. These describe the motion of a charged solute in a solvent fluid by a diffusion-advection process. Boundary conditions will be discussed in [3](#sec:numerical){reference-type="ref" reference="sec:numerical"}. For each solute species \(k\), the solute flux is given by \[\begin{aligned} \vec{j}_k(\vec{r},t)=& \underbrace{-D_k\vec{\nabla}\rho_k(\vec{r},t) -\frac{D_k}{k_\text{B} T} z_ke \rho_k(\vec{r},t) \vec{\nabla} \Phi(\vec{r},t)}_{\vec{j}^\text{diff}_k} \nonumber\\ &+ \underbrace{\rho_k(\vec{r},t) \vec{u}(\vec{r},t)}_{\vec{j}^\text{adv}_k} \label{eq:flux} \end{aligned}\] and obeys the continuity equation \[\frac{\partial \rho_k}{\partial t}(\vec{r},t) =-\vec{\nabla} \cdot \vec{j}_k(\vec{r},t) \label{eq:continuity} .\] Here, \(\rho_k(\vec{r},t)\) is the concentration field and \(\vec{j}_k(\vec{r},t)\) is the associated flux. \(\Phi(\vec{r},t)\) refers to the electrostatic potential and \(\vec{u}(\vec{r},t)\) to the fluid velocity. \(\vec{\nabla}\cdot\) is the divergence operator; \(\partial/\partial t\) denotes the time derivative. \(D_k\) is the diffusion coefficient of the \(k\)th solute, \(k_\text{B}\) Boltzmann's constant, \(T\) the absolute temperature, \(z_k\) is the \(k\)th solute's ionic valency, and \(e\) is the unit charge. The electrostatic potential is given by Poisson's equation \[\begin{aligned} \nabla^2 \Phi(\vec{r},t) &=-\frac{1}{\varepsilon} \sum\limits_k z_ke\rho_k(\vec{r},t) \nonumber \\ &=-\frac{4\pi \lambda_\text{B} k_\text{B}T}{e} \sum\limits_k z_k\rho_k(\vec{r},t) , \label{eq:poisson} \end{aligned}\] with the Bjerrum length \[\lambda_\text{B}=\frac{e^2}{4\pi \varepsilon k_\text{B} T} .\] A spatially homogeneous dielectric permittivity \(\varepsilon\) is assumed here. As the relevant length scales in colloidal systems are on the nano-and micrometer scale, we are in the low Reynolds number regime---viscous friction dominates over inertial forces---and the dynamics of the fluid is thus described by the stationary incompressible Stokes equations: \[\begin{aligned} \eta \nabla^2\vec{u}(\vec{r},t) &= \vec{\nabla} p(\vec{r},t)-\vec{f}_\text{ext}(\vec{r},t), \label{eq:stokes} \\ \vec{\nabla}\cdot\vec{u}(\vec{r},t) &= 0 \label{eq:continuity-fluid} , \end{aligned}\] with the kinematic viscosity \(\eta\), the pressure density \(p(\vec{r},t)\), and the applied force density \(\vec{f}_\text{ext}(\vec{r},t)\). The fluid's continuity [\[eq:continuity-fluid\]](#eq:continuity-fluid){reference-type="ref" reference="eq:continuity-fluid"} neglects the contribution of the solute to the total mass, which is reasonable at typical solute concentrations up to several mol/l. The force density couples the motion of the fluid to the dynamics of the solutes via the following equation: \[\vec{f}_\text{ext}(\vec{r},t)=k_BT\sum\limits_k \frac{\vec{j}_k^\text{diff}(\vec{r},t)}{D_k} . \label{eq:fluidcoupling}\] This represents a frictional coupling proportional to the relative velocities of fluid and ions. # Numerical Methods {#sec:numerical} In this section, we introduce the numerical methods employed to solve the electrokinetic equations [\[eq:flux\]](#eq:flux){reference-type="eqref" reference="eq:flux"}-[\[eq:fluidcoupling\]](#eq:fluidcoupling){reference-type="eqref" reference="eq:fluidcoupling"}. cover the LB method and the EK method, respectively, while [3.2](#sec:ladd){reference-type="ref" reference="sec:ladd"} describes the moving boundary method for coupling particles into LB. describes a new method for combining moving boundaries with EK. ## Lattice Boltzmann {#sec:lb} The LB method is based on solving the Boltzmann transport equation, \[\begin{aligned} \frac{\mathrm{d}f}{\mathrm{d}t}= &\frac{\partial f}{\partial t} + \vec{v} \cdot \vec{\nabla}\!_x f + \frac{1}{m} \vec{F} \cdot \vec{\nabla}\!_v f \left. \frac{\partial f}{\partial t} \right| _\text{collision} \equiv \tens{\Omega} f \label{eq:transport} . \end{aligned}\] \(\vec{\nabla}\!_a\) is the gradient operator with respect to \(\vec{a}\). \(f=f(\vec{x},\vec{v}, t)\) is a system's single-particle phase space probability distribution function. It gives the probability density of finding a particle at position \(\vec{x}\) with a velocity \(\vec{v}\) at time \(t\). \(\vec{F}\) is the force acting on the particle. The right-hand side is the collision operator. The Boltzmann transport equation [\[eq:transport\]](#eq:transport){reference-type="eqref" reference="eq:transport"} describes microscopic particle motion, while the Stokes equation [\[eq:stokes\]](#eq:stokes){reference-type="eqref" reference="eq:stokes"} describes flow of a Newtonian continuum fluid. However, both are based on the same conservation laws, those of mass and momentum, and they can therefore be used interchangeably when studying mesoscopic hydrodynamics. The LB method linearizes the collision operator \(\tens{\Omega}\) on the right-hand side of around the equilibrium distribution \(f^\text{eq}\). The total derivative on the left-hand side is discretized using finite differences with a lattice spacing of \(a_\text{grid}\) and a time step of \(\tau\) to obtain \[f_i(\vec{r}+\tau\vec{c}_i, t+\tau) = f_i(\vec{r},t)+\sum\limits_j \Omega_{ij} \underbrace{\left(f_j(\vec{r},t)-f^\text{eq}_j(\vec{r},t)\right)}_{f^\text{neq}_j(\vec{r},t)} , \label{eq:lb}\] with the lattice vectors \(\vec{c}_i\). The collision represents a linear relaxation of the populations towards their equilibrium Maxwell-Boltzmann distribution. At each lattice site, there is a number of fluid populations \(f_i\), e.g., 19 different ones for the widely used D3Q19 model applied in this work, corresponding to a set of velocity vectors \(\vec{c}_i\) connecting the site to neighboring cells. For D3Q19, these are the rest mode, the six face neighbors, and the twelve edge neighbors. In each time step, a collision operation is performed that redistributes fluid between the 19 populations within one cell, i.e., [\[eq:lb\]](#eq:lb){reference-type="ref" reference="eq:lb"} is applied with only \(\tens{\Omega}\) on the right-hand side. Next, a streaming step is performed which transfers the populations to the neighboring cell in the direction of their respective velocity vector, i.e., [\[eq:lb\]](#eq:lb){reference-type="ref" reference="eq:lb"} is applied without \(\tens{\Omega}\) on the right-hand side. From the populations \(f_i\), the macroscopic physical quantities of fluid density and flow velocity are obtained using \[\begin{aligned} \rho(\vec{r},t)&=\sum\limits_i f_i(\vec{r},t), \\ \vec{u}(\vec{r},t)&=\frac{1}{\rho(\vec{r},t)}\sum\limits_i f_i(\vec{r},t)\vec{c}_i. \end{aligned}\] Throughout the manuscript, we choose a two relaxation time (TRT) LB collision operator for its high fidelity of reproducing no-slip (zero velocity) boundaries: \(f_i^{\pm}\) and \(f_i^{\text{eq}\pm}\) are symmetric (\(+\)) and antisymmetric (\(-\)) combinations of populations and equilibrium populations, respectively. The symmetric relaxation parameter \(\lambda_\text{e}\) derives from the fluid viscosity, while the antisymmetric \(\lambda_\text{o}=3/16\) is tuned to minimize boundary slip. The index \(\bar{i}\) is the one whose velocity vector points in the direction opposite to the one with index \(i\), i.e. \(\vec{c}_{\bar{i}}=-\vec{c}_i\). We will have forces acting on the fluid that vary in time and space for moving boundaries, therefore we need to augment the collision operation to be \(\Omega_{ij}f^\text{neq}_j+\Delta_i\) with the appropriate force term \(\Delta_i\). We utilize the one originally constructed by and generalized to multiple relaxation times: where \(\otimes\) denotes the tensor (dyadic) product of two vectors and \(\mathrm{Tr}\) the trace of a matrix. \(w_i\) is the lattice weight factor associated with direction \(i\), \(\vec{f}_\text{ext}\) is the force acting on a lattice cell, and \(\vec{u}\) the flow velocity in that cell. Until we introduce the moving boundaries in [3.2](#sec:ladd){reference-type="ref" reference="sec:ladd"}, we can consider colloidal particles by requiring them to be fixed relative to the lattice. This corresponds to cells inside the colloid being a no-slip boundary for the fluid (corresponding to the co-moving frame, if a Galilean transformation is permitted). In that case, the streaming step is modified so that populations streaming into the boundary are reflected back. A generalization to arbitrary boundary velocities \(\vec{v}_\text{b}\) is possible and leads to \[f_i(\vec{r}_\text{b}+\tau\vec{c}_i,t+\tau) =f_{\bar{i}}(\vec{r}_\text{b},t) + \frac{6 w_i\tau^2}{\rhoa_\text{grid}^2} \vec{c}_i\cdot \vec{v}_\text{b},\] where \(\vec{r}_\text{b}\) is the boundary node position. This is referred to as a velocity bounce-back boundary. We do not include thermal fluctuations of the solvent in this work. Neither do we do so for the solute, as will be discussed in [3.3](#sec:ek){reference-type="ref" reference="sec:ek"}. A consistent thermalization for the pure LB algorithm is available, and such fluctuations may be significant for nanoscale systems. ## Moving LB Boundaries {#sec:ladd} There are multiple methods that incorporate particles into an LB fluid, ranging from point-particle descriptions to particles that are resolved on the grid. In this paper, we are interested in the latter, for which the method by is the obvious choice to achieve coupling. It considers all cells inside a particle as no-slip boundaries in the particle-co-moving frame. In the lab frame, this corresponds to a velocity boundary condition with velocity \(\vec{v}_\text{b}\) equal to the particle's surface velocity \[\vec{v}_\text{b}(\vec{r}_\text{b},t)=\vec{v}(t) + \vec{\omega}(t) \times (\vec{r}_\text{b}-\vec{r}(t)) .\] Here, the particle is located at \(\vec{r}\) and has linear velocity \(\vec{v}\) and angular velocity \(\vec{\omega}\). The linear and angular momentum exchanged during the bounce-back is transferred to the particle by applying an appropriate force and torque, Here, the bounce-back is taking place between the boundary cell \(\vec{r}_\text{b}\) and the fluid cell \(\vec{r}_\text{f}\). After determining these forces, the particle trajectory can be integrated, in our case using a symplectic Euler integrator. These forces, along with any non-hydrodynamic forces applied, cause the particle to move across the lattice. While it does so, the set of cells overlapped by the particle changes. The original method by has some shortcomings, due to the presence of fluid inside the particle. These include forces exerted on the particle by the internal fluid and the requirement that the particle's density is the same or larger than that of the fluid. The modification by removes this unphysical fluid. However, one now needs to destroy fluid in cells that are newly overlapped by a particle and create fluid in cells that are vacated. The former is straightforward to do, while the latter requires choosing a value for each of the \(f_i\) in the cell that becomes 'new' fluid. The initialization of the new fluid cell is problematic, as the D3Q19 LB model has more microscopic degrees of freedom than specified by the hydrodynamic boundary condition. The typical choice is to use LB's equilibrium populations for a velocity equal to the particle's velocity and a density equal to the surrounding fluid cells' average density. To conserve momentum, the created / destroyed fluid's momentum \(\vec{p}_\text{f}\) is removed from / added to the particle as an additional force: \[\vec{F}(t)=\pm\frac{1}{\tau}\vec{p}_\text{f}(\vec{r}_\text{f},t) .\] Fluid mass is never conserved, as conserving it would violate the incompressibility of the fluid. However, the time-averaged fluid mass fluctuates around the correct value while the particle traverses the grid. ## Lattice Electrokinetics {#sec:ek} The method proposed by discretizes [\[eq:flux\]](#eq:flux){reference-type="ref" reference="eq:flux"} with a 19-point stencil on the same lattice that LB is performed on. have shown that the discretization originally proposed in Ref.  introduces errors scaling exponentially with \(E/a_\text{grid}\) where \(E\) is the applied electric field. We instead use \[\begin{aligned} j_{ki}^\text{diff}&(\vec{r}\rightarrow\vec{r}+\vec{c}_i,t) \nonumber\\ =& \frac{D_k}{a_\text{grid}}\left(\rho_k(\vec{r},t)-\rho_k(\vec{r}+\vec{c}_i,t)\right) \nonumber\\ &-\frac{D_k z_ke}{k_\text{B} Ta_\text{grid}(1+2\sqrt{2})} \frac{\rho_k(\vec{r},t) + \rho_k(\vec{r}+\vec{c}_i,t)}{2} \nonumber\\ &\times \left(\Phi(\vec{r},t)-\Phi(\vec{r}+\vec{c}_i,t)\right) \label{eq:flux-disc} \end{aligned}\] for the diffusive term, as in Ref. . For the advective term \(j_{ki}^\text{adv}\), a volume-of-fluid method is used: the solute in a cell at \(\vec{r}\) is virtually displaced by a distance of \(\vec{u}(\vec{r},t)\tau\), where \(\vec{u}\) is the velocity of the LB fluid. This displaced cell is intersected with all 26 neighboring cells to determine the amount of solute that needs to be transferred. Once the fluxes have been calculated, they can be propagated using the discretized form of the continuity equation [\[eq:continuity\]](#eq:continuity){reference-type="eqref" reference="eq:continuity"}, \[\frac{V_0}{\tau}\left(\rho_k(\vec{r},t+\tau)-\rho_k(\vec{r},t)\right) =-A_0 \sum\limits_i j_{ki}(\vec{r},t),\] with \(V_0\) the cell volume and \(A_0\) an effective surface area chosen such that [\[eq:flux-disc\]](#eq:flux-disc){reference-type="ref" reference="eq:flux-disc"} recovers the correct (bulk) diffusive mean-square displacement of \(6Dt\). This results in \(A_0=1+2\sqrt{2}\) for the D3Q19 lattice. The fluid coupling force is a direct discretization of [\[eq:fluidcoupling\]](#eq:fluidcoupling){reference-type="ref" reference="eq:fluidcoupling"} and is applied using [\[eq:guo\]](#eq:guo){reference-type="ref" reference="eq:guo"}: \[\vec{f}_\text{ext}(\vec{r},t)=\frac{k_\text{B}Ta_\text{grid}}{\tau}\sum_k \sum_i \frac{j_{ki}^\text{diff}(\vec{r},t)}{D_k} \vec{c}_i .\] Electrostatics [\[eq:poisson\]](#eq:poisson){reference-type="eqref" reference="eq:poisson"} is treated using any available lattice-based electrostatics solver. For this work, one based on Fast Fourier Transforms (FFT) is used, but iterative solvers for linear equation systems, such as Successive Over-Relaxation (SOR) or Krylov subspace methods, may also be applied. The non-FFT-based solvers are also capable of incorporating inhomogeneous dielectric coefficients; the assumption of homogeneity has only entered into the specific form of Poisson's equation [\[eq:poisson\]](#eq:poisson){reference-type="eqref" reference="eq:poisson"} used here. Besides driving the migrative flux [\[eq:flux-disc\]](#eq:flux-disc){reference-type="eqref" reference="eq:flux-disc"}, electrostatics also results in a force acting on the colloidal particles, which is added to the hydrodynamic drag force [\[eq:force\]](#eq:force){reference-type="eqref" reference="eq:force"} before integrating the particle trajectory. At this point, EK is only capable of handling stationary boundary conditions, such as walls and particles that are stationary with respect to the lattice. They are simply mapped into the LB fluid as a no-slip boundary in all cells inside them and any EK solute fluxes into or out of these cells is set to zero. Using normal fluxes other than zero allows for the incorporation of chemical reactions occurring on the surface that lead to phenomena such as charge regulation and self-electrophoresis. Boundaries may be charged by considering their charge distribution \(\rho_\text{b}(\vec{r},t)\) as an additional summand when solving [\[eq:poisson\]](#eq:poisson){reference-type="ref" reference="eq:poisson"}. As mentioned before, thermal fluctuations are not included. The addition of noise to the concentration fields---while obtaining the proper fluctuation-dissipation relation for the total fluid---is non-trivial and goes beyond the scope of the current manuscript. Thermalization of the EK algorithm based on existing numerical methods for fluctuating hydrodynamics of electrolytes shall be the subject of future study. ## Moving EK Boundaries {#sec:ek-moving} In [3.2](#sec:ladd){reference-type="ref" reference="sec:ladd"}, the well-known moving boundary method for LB was introduced. It is not mass-conserving, so a straightforward adaptation to EK would cause the amount of solute to vary over time and thus violate charge conservation. Furthermore, as the solute charge is typically far less homogeneously distributed than the fluid mass, one can expect that charge would not even be conserved on average over long time scales. A simulation undergoing such a net charge drift will typically not be able to produce physically correct results. ### Simple Charge Conservation Scheme {#sec:simplescheme} To avoid a net charge drift, any solute from cells claimed by a particle needs to be expelled and redistributed into surrounding cells. Such a scheme is illustrated in [\[fig:chargecons\]](#fig:chargecons){reference-type="ref" reference="fig:chargecons"}. Cells vacated behind the particle are refilled by taking the sum of the solute concentration of the surrounding \(N_\text{f}\) non-boundary cells and dividing it by \(N_\text{f}+1\) to account for the fact that this amount of solute will now be shared with one additional cell. This means that the solute concentration \[\rho_k(\vec{r}_\text{new},t+\tau)=\frac{1}{N_\text{f}+1}\sum\limits_{i=1}^{N_\text{f}} \rho_k(\vec{r}_\text{new}+\vec{c}_i,t) \label{eq:refill}\] is put into the vacated cell. To conserve total solute mass, that amount is then removed from the surrounding non-fluid cells in amounts proportional to their current solute concentration: \[\rho_k(\vec{r}+\vec{c}_i,t+\tau) = \rho_k(\vec{r}+\vec{c}_i,t)-\frac{\rho_k(\vec{r}+\vec{c}_i,t)}{N_\text{f}+1} \label{eq:refillneighbors}.\] It should be noted that expulsion and vacation do not necessarily happen at the same time in the front and back of a moving particle, and thus the total number of boundary cells oscillates over time. ### Enhanced Partial Volume Scheme The scheme introduced in [3.4.1](#sec:simplescheme){reference-type="ref" reference="sec:simplescheme"} has the disadvantage that it moves large amounts of solute at a few points in time. Therefore, as we demonstrate in [4](#sec:validation){reference-type="ref" reference="sec:validation"}, the particle's velocity can vary quite strongly during the time steps after a cell has been claimed or vacated. To reduce these effects, we propose a partial volume scheme, which is illustrated in [\[fig:partialvolume\]](#fig:partialvolume){reference-type="ref" reference="fig:partialvolume"}. In the following, \(\Psi(\vec{r},t)\) is a field describing the volume fraction of the cell at \(\vec{r}\) that is overlapped by a particle, with \(\Psi=1\) meaning that the cell is completely inside the particle and \(\Psi=0\) completely outside. In the calculation of the diffusive fluxes [\[eq:flux-disc\]](#eq:flux-disc){reference-type="eqref" reference="eq:flux-disc"}, the concentrations are replaced with ones that take into account that all solute resides in the non-overlapped part of the cells. To prevent the resulting diffusive fluxes from diverging as \(\Psi\rightarrow 1\), we renormalize them by scaling them with the volume. This leads to the following modified expression for the flux: \[\begin{aligned} j_{ki}^\text{diff}&(\vec{r}\rightarrow\vec{r}+\vec{c}_i,t) \nonumber\\ =& \left[ \frac{D_k}{a_\text{grid}}\left(\frac{\rho_k(\vec{r},t)}{1-\Psi(\vec{r},t)}-\frac{\rho_k(\vec{r}+\vec{c}_i,t)}{1-\Psi(\vec{r}+\vec{c}_i,t)}\right) \right. \nonumber\\ &-\frac{D_k z_ke}{2 k_\text{B} Ta_\text{grid}(1+2\sqrt{2})} \left( \frac{\rho_k(\vec{r},t)}{1-\Psi(\vec{r},t)} \right. \nonumber\\ &+ \left.\frac{\rho_k(\vec{r}+\vec{c}_i,t)}{1-\Psi(\vec{r}+\vec{c}_i,t)} \right) \times \left. \vphantom{\frac{1}{1}} \left(\Phi(\vec{r},t)-\Phi(\vec{r}+\vec{c}_i,t)\right) \right] \nonumber\\ & \times \left(1-\Psi(\vec{r},t)\right) \left(1-\Psi(\vec{r}+\vec{c}_i,t)\right) \label{eq:flux-pv} . \end{aligned}\] With this change, refilling vacated cells as per [\[eq:refill,eq:refillneighbors\]](#eq:refill,eq:refillneighbors){reference-type="ref" reference="eq:refill,eq:refillneighbors"} is no longer necessary. They can be set to zero concentration and will be filled up by the diffusive flux again as \(\Psi\) increases. We calculate \(\Psi\) numerically by sub-dividing each cell into 8 equally-sized cells and determining how many of them are completely inside and completely outside the particle. For those cells that are neither, the subdivision is recursively repeated up to a maximum depth of \(4\). Expelling solute from a cell that is claimed by a particle is, however, still necessary---even with the modified expression for the flux---as the cell is not necessarily completely empty by the time it is claimed, due to the discretized motion of the colloid. The expelled amount of solute with [\[eq:flux-pv\]](#eq:flux-pv){reference-type="ref" reference="eq:flux-pv"} is much smaller than with [\[eq:flux-disc\]](#eq:flux-disc){reference-type="ref" reference="eq:flux-disc"} and thus the effect of this sudden change on the simulation is reduced to acceptable levels. One further source of sudden variations in solute fluxes is the change in electrostatic potential when the volume across which a particle's charge is distributed varies due to the fluctuation in the number of boundary cells. Therefore, when calculating the electrostatic potential, each particle's total charge \(Q=Ze\) is distributed among all cells that are at least partially overlapped by that particle: \[\rho_\text{b}(\vec{r},t) = Ze\frac{\Psi(\vec{r},t)}{V_\text{p}} ,\] with \(V_\text{p}\) the particle's (non-discrete) volume. Inhomogeneous charge distributions are also possible as long as the charge in a cell varies smoothly while the cell is slowly claimed or vacated by the colloidal particle. # Validation {#sec:validation} We implement our new algorithm using the [waLBerla]{.sans-serif} framework. It supports several lattice Boltzmann models, including the one introduced in [3.1](#sec:lb){reference-type="ref" reference="sec:lb"}, and correctly handles the moving LB boundaries described in [3.2](#sec:ladd){reference-type="ref" reference="sec:ladd"}. We already added an implementation of the EK model described in [3.3](#sec:ek){reference-type="ref" reference="sec:ek"}. [waLBerla]{.sans-serif} provides excellent scaling on high-performance computing clusters and contains advanced features, such as grid refinement, that may prove useful for EK simulations of complex systems in the future. ## Electrophoresis of a Single Colloid {#sec:singlecolloid} To validate the moving boundary EK method described in [3.4](#sec:ek-moving){reference-type="ref" reference="sec:ek-moving"}, we choose a simple system consisting of an electrophoresing sphere. Specifically, we simulate a single, homogeneously charged, spherical colloid in a cubic box with periodic boundary conditions (PBCs) undergoing electrophoresis in a uniform external electric field. This system has already been studied extensively using the EK method without moving boundaries by considering the equivalent problem of electro-osmotic flow. The parameters used are given in [1](#tab:parameters){reference-type="ref" reference="tab:parameters"}. They were chosen to match one of the data points of Ref. . A salt concentration of \(c=\SI{1e-3}{\mole\per\litre}\) is experimentally relevant and the (effective) colloid charge of \(Z=30\) is realistic for a particle of radius \(R=\SI{4e-9}{\metre}\). The resulting Debye length of \(\lambda_\text{D}=\SI{10}{\nano\metre}\) for that concentration is neither in the thin (\(\lambda_\text{D}\ll R\)) nor thick (\(\lambda_\text{D}\gg R\)) double layer limit and should thus demonstrate the capability of the method to deal with these intermediate regimes. The charge of \(Z=30\) was chosen such that Debye-Hückel theory is not applicable---the requirement being \(e\zeta < k_\text{B}T\) with \(\zeta=Zk_BT\lambda_\text{B}/(R+R^2/\lambda_\text{D})\)---while for \(Z=3\), it is. The salt concentration \(c\) and the grid spacing \(a_\text{grid}\) need to be chosen such that one Debye length is resolved by a minimum of approximately 4 cells, as we will see during validation. This precludes us from using a salt concentration of much higher than \(c=\SI{1e-3}{\mole\per\litre}\) without increasing the grid resolution. However, sufficiently resolving the double layer is a general requirement of the EK method, as well as for FEM and other algorithms, and not specific to the use of moving boundaries. We will compare the results of the different simulation methods via a single number, the reduced electrophoretic mobility: \[\tilde\mu=\frac{6\pi \eta \lambda_\text{B}}{e}\frac{v}{E} ,\] where \(v\) is the speed of the particle's motion relative to the bulk fluid and \(E\) is the applied electric field. compares results obtained for the same parameter set (particle radius \(R\approx\SI{4e-9}{\metre}\), particle charge \(Z=30\), and salt concentration \(c=\SI{1e-3}{\mole\per\litre}\)) for a simulation with a fixed particle, a moving particle without, and a moving particle with the partial volume scheme, respectively. The effective particle radius in the moving scheme is obtained by counting the number of cells that are completely inside the particle, averaging that number over the simulation time, and setting it equal to the sphere volume \((4/3)\pi R^3\). This allows us to choose \(R\) such that the moving and stationary simulations study the same physical system. One can see that both moving particle methods on average agree with the fixed boundary simulation's stationary mobility value to within \(1.4\%\) and \(2.4\%\), respectively. Agreement of the full time evolution is not expected as the electro-osmotic (fixed boundary) and electrophoretic (moving boundary) problems are, in fact, only equivalent in the stationary state. The method without partial volumes results in a mobility oscillating with an amplitude of \(20.3\%\) around the mean and a period corresponding to the time it takes the particle to move forward by one grid cell. The partial volume scheme on the other hand reduces the mobility oscillations to \(1.7\%\), which is generally small enough to consider any instantaneous value to be a good estimate for the true value. The period remains bound to the particle moving forward by one cell. To ensure that the method does not adversely affect the shape of the electric double layer around the colloid, we examine the charge distribution around the colloid in [\[fig:chargedist\]](#fig:chargedist){reference-type="ref" reference="fig:chargedist"}. We see there is a small deviation close to the colloid's surface, but we can attribute this to the slight difference in particle size, as discussed above. We further validated the partial volume method by running additional simulations with various concentrations \(c\), particle charges \(Z\), and particle radii \(R\). The results are presented in [\[fig:results\]](#fig:results){reference-type="ref" reference="fig:results"} and are compared with reference results obtained by solving the electrokinetic equations with the FEM solver [COMSOL]{.sans-serif}. Note that we use a finite but large simulation domain in the FEM model, while the EK model handles PBCs naturally; this may lead to slight differences in the results obtained with both methods. We observe an excellent agreement between fixed and moving boundary simulations, as well as the FEM reference results. The agreement improves as the particle size is increased, and thus the quality of the staircase approximation to its spherical shape. For example, for \(R\approx\SI{7e-9}{m}\), \(Z=30\), and \(c=\SI{1e-3}{\mole\per\litre}\), the oscillation amplitude drops to \(1.3\%\). The remaining difference in the mobility is easily explained by a slight mismatch in volume (and thus effective radius) between the fixed and the moving particle. Finally, we examined the speed of our simulation compared to the method of for the geometries that we considered. The moving boundary algorithm adds only 2% to the simulation time per time step, when compared to the equivalent fixed boundary simulation. Our smoothing results in a further increase of less than 10% in time, when compared to the non-smoothed moving boundaries. Together, this shows that our method does not incur an unreasonable computational cost and can therefore be straightforwardly applied to domain sizes that are currently accessible to the method. ## Electrophoresis of a Colloidal Suspension In this section, we consider the electrophoresis of a system comprised of several spheres. This allows us to verify our moving boundary method in a more physically complex system with interactions that the original method could not access. We pick one parameter set (\(R\approx\SI{4e-9}{\metre}\), \(Z=30\), and \(c=\SI{1e-3}{\mole\per\litre}\)) and repeat the simulation from [4.1](#sec:singlecolloid){reference-type="ref" reference="sec:singlecolloid"}. At these parameters, the volume fraction is about 0.1% and the double layers do not overlap noticeably, therefore only long-ranged hydrodynamic interactions are expected to mediate interactions between the colloids. Instead of having one spherical colloid interact with only its periodic images, we now enlarge the simulation box and add additional identical spheres while keeping the colloidal volume fraction constant. The colloids are either positioned randomly in a cubic box or in a regular lattice in a cuboidal box. This kind of simulation is typically performed to ensure the results do not suffer from artifacts of the PBCs. As opposed to the single colloid and its periodic images, which always maintain their relative positions, the colloids in this simulation are free to move relative to each other. shows that the electrophoretic mobility is indeed mostly independent of the number of colloids, varying \(0.9\%\) from the average and \(1.8\%\) from the value for one colloid in PBCs. This matches the findings of using colloids modeled with the Ahlrichs-Dünweg coupling, where agreement to within a few percent was found. The results of [4.1](#sec:singlecolloid){reference-type="ref" reference="sec:singlecolloid"} are therefore not artifacts of the periodicity of the simulation domain. For the case of a cubic simulation domain, we furthermore observe that the alignment of the colloids on a regular body-centered cubic lattice is a stable configuration. # Conclusion {#sec:conclusion} Summarizing, we have introduced a method to simulate electrokinetic phenomena in colloidal suspensions. Our scheme builds upon the lattice electrokinetics algorithm of, which itself is capable of simulating moving colloids only by considering them as stationary boundary conditions in the Galilean-transformed (co-moving) frame. Motion of the colloids relative to the lattice can be incorporated by employing a method similar to the moving boundary method for LB. A key modification is required, however: We introduce mass conservation for the solute species in order to conserve charge. This is accomplished by redistributing solute from and to cells neighboring the ones that were recently vacated and claimed by the particle, respectively. The above procedure is, in principle, sufficient to enable the simulation of, for example, colloidal electrophoresis. However, in practice, further improvements are desirable to allow for the simulation of colloids with as few LB cells as possible. To reduce the effects of sudden and strong solute fluxes, when a cell is claimed or vacated by a particle, a smoothing scheme by partial volumes is introduced. That is, the electrokinetic equations are solved for an effective concentration that incorporates what fraction of a cell is actually accessible to the solute, i.e., not overlapped by a colloid. This form of smoothing reduces the lattice artifacts by more than a decade in our test simulations. We also showed that having smoothing and moving boundaries only slows down the simulation by about \(10\%\) compared to the original Capuani algorithm. We validated our method for the electrophoresis of spherical colloids. For a single sphere, we find excellent agreement of the reduced steady-state electrophoretic mobility \(\tilde\mu\) with the one obtained for the equivalent problem of electro-osmotic flow around a fixed sphere (co-moving frame). One can also obtain a reliable estimate for the mobility without the smoothing via partial volumes, but in this case it is necessary to average \(\tilde\mu\) over one period of its oscillations. While this is a feasible solution for the system under consideration here, systems where the transient behavior of multiple particles is of interest would require averaging over many periods, thus requiring much longer overall simulation times. Alternatively, the oscillations could be reduced by increasing the grid resolution. This is again not desirable, as it comes with a steep increase in required computation time---performing the same simulation at twice the grid resolution takes eight times as much computing time. It is therefore clear that our smoothing is a prerequisite for the efficient study of electrokinetic moving boundary problems. Our moving boundary EK method was further verified for the electrophoresis of multiple spheres that were free to move and interact. We find the electrophoretic mobility of the spheres to be almost independent of their initial arrangement. Such a simulation can only be performed with moving boundaries and thus demonstrates the power of the method presented. Note that our moving boundary EK method imposes no limitations on simulation parameters beyond those already present in the original non-moving algorithm. These requirements include sufficient discretization of the double layer and limitations on the choice of time step, diffusivity, and grid discretization. Furthermore, it leaves the EK and LB algorithms almost unmodified, which makes it easy to incorporate into existing LB or EK simulation codes. Finally, it also does not influence the scaling behavior of these algorithms as the work required to map a particle onto the lattice is linear in the number of lattice cells and particles. In this investigation, we have primarily studied the case of external electrophoresis of a single colloid. As we have shown, however, our method can also be employed to investigate cases where Galilean transformation to a co-moving frame for the study of an equivalent electro-osmotic is not possible. These types of systems will be the subject of future work and could include a diverse range of systems. For example, our method is readily applicable to study the translocation, characterization, concentration, separation, and transport of DNA, proteins, and other biochemical analytes; interactions between oppositely charged nanoparticles in an oscillatory electric field; collective dynamics of self-electrophoresing colloids, and many others. Our work thus opens the door for the study of a wide range of physical systems that were previously inaccessible to continuum lattice-based methods.

{'timestamp': '2016-11-23T02:05:55', 'yymm': '1607', 'arxiv_id': '1607.04572', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04572'}

# Introduction A major goal of planetary exploration is to obtain a fundamental understanding of planetary environments, both as they are currently and as they were in the past. This knowledge can be used to explore the questions of (a) how conditions for planetary habitability arose and (b) the origins of life. Titan is a unique object of study in this quest. Other than Earth itself, and Pluto [which has also been observed to have photochemically produced haze; @stern15; @gladstone16], Titan is the only solar system body demonstrated to have complex organic chemistry occurring today. Its atmospheric properties---(1) a thick N\(_2\) atmosphere, (2) a reducing atmospheric composition, (3) energy sources for driving disequilibrium chemistry and (4) an aerosol layer for shielding the surface from solar UV radiation---suggest it is a counterpart of the early Earth, before the latter's reducing atmosphere was eradicated by the emergence and evolution of life. A significant number of photochemical models have been developed to investigate the distribution of hydrocarbons in Titan's atmosphere. Recently, more constraints have been placed on the abundance of hydrocarbons and nitriles in the mesosphere of Titan (500--1000 km) from Cassini/UVIS stellar occultations. In combination with the updated version of Cassini/ CIRS limb view, the complete profiles of C\(_2\)H\(_2\), C\(_2\)H\(_4\), C\(_6\)H\(_6\), HCN, HC\(_3\)N are revealed for the first time. C\(_3\) compounds, including C\(_3\)H\(_6\), were modeled by, and the agreement with observations is satisfactory. The chemistry of many of these nitrogen molecules has recently been modeled by. In this paper we introduce our updated Titan chemical model that includes the formation of such potentially astrobiologically important molecules as acrylonitrile. In addition to the usual gas phase chemistry, it also includes a numerically stable treatment of the condensation and sublimation, allowing the formation and destruction of ices in the lower atmosphere to be tracked. Haze formation is also included in a parameterized fashion, allowing for the permanent removal of molecules from the atmosphere. We present here the effects of condensation on the nitrogen chemistry. The interaction of hydrocarbons and nitrile species in the condensed phase is complex and is beyond the scope of this paper [see, for example, Figures 1 and 2 of @anderson16]. We begin with describing our updated model and in particular our treatment of condensation and sublimation (Section [2](#sec:model){reference-type="ref" reference="sec:model"}). We use this updated model to consider the chemistry in Titan's atmosphere from the surface of the moon to an altitude of 1500 km. We explore how condensation processes and haze formation affect the predicted gas phase abundances of observable molecules (Section [3](#sec:results){reference-type="ref" reference="sec:results"}). We also consider where the condensates form within the atmosphere (Section [\[sec:cond\]](#sec:cond){reference-type="ref" reference="sec:cond"}). Section [4](#sec:conc){reference-type="ref" reference="sec:conc"} presents our conclusions. # The Model {#sec:model} We use the Caltech/JPL photochemical model with a recently updated chemical network, and with the addition of condensation and sublimation processes to explore the atmospheric chemistry of Titan. The 1-D model solves the mass continuity equation from the surface of Titan to 1500 km altitude: \[\frac{\partial n_i}{\partial t} + \frac{\partial \psi_i}{\partial z} = P_i-L_i\] where \(n_i\) is the number density of species \(i\), and \(P_i\) and \(L_i\) are its chemical production and loss rates respectively. \(\psi_i\) is the vertical flux of \(i\) calculated from \[\psi_i =-\frac{\partial n_i}{\partial z}(D_i + K_{zz})-n_i \left( \frac{D_i}{H_i} + \frac{K_{zz}}{H_a} \right)-n_i \frac{\partial T}{\partial z} \frac{(1+\alpha_i)D_i + K_{zz}}{T}\] where \(D_i\) and \(H_i\) are the molecular diffusion coefficient and the scale height for species \(i\) respectively, \(H_a\) is the atmospheric scale height, \(\alpha_i\) is the thermal diffusion coefficient of species \(i\), \(T\) is the temperature and \(K_{zz}\) is the eddy diffusion coefficient. The eddy diffusion coefficient used here is taken from and can be summarized as \[log K_{zz}(z) = \left\{ \begin{array}{ll} log(3 \times 10^3), & z < z_1\\ \\ log(3 \times 10^3) \dfrac{z_2-z}{z_2-z_1} + log(2 \times 10^7) \dfrac{z-z_1}{z_2-z_1}, & z_1 \leq z < z_2 \\ \\ log(2 \times 10^7) \dfrac{z_3-z}{z_3-z_1} + log(2 \times 10^6) \dfrac{z-z_2}{z_3-z_2}, & z_2 \leq z < z_3\\ \\ log(2 \times 10^6) \dfrac{z_4-z}{z_4-z_3} + log(4 \times 10^8) \dfrac{z-z_3}{z_4-z_3}, & z_3 \leq z < z_4\\ \\ log(4 \times 10^8), & z \geq z_4 \end{array} \right.\] The atmospheric density and temperature profiles are also taken from, and are based on the T40 Cassini flyby. Aerosols are included in our model, both for the absorption of UV radiation and to provide surfaces onto which molecules can condense. The aerosol properties are from who derived them from a microphysical model validated against Cassini/Descent Imager Spectral Radiometer (DISR) observations. Their results provide the mixing ratio and surface area of aerosol particles as a function of altitude (Figure [\[fig:aerosol\]](#fig:aerosol){reference-type="ref" reference="fig:aerosol"}). To calculate the absorption of UV by dust we assume absorbing aerosols with extinction cross-sections that are independent of wavelength. ## Boundary Conditions The lower boundary of our model is the surface of Titan and the upper boundary is at 1500 km. For H and H\(_2\) the flux at the lower boundary is zero and at the top of the atmosphere these molecules are allowed to escape with velocities of 2.4 \(\times\) 10\(^4\) cms\(^{-1}\) and 6.1 \(\times\) 10\(^{3}\) cms\(^{-1}\) respectively (equivalent to fluxes of 3.78 \(\times\) 10\(^8\) H atoms cm\(^{-2}\) and 6.2 \(\times\) 10\(^9\) H\(_2\) molecules cm\(^{-2}\)). For all other gaseous species the concentration gradient at the lower boundary is assumed to be zero, while they have zero flux at the top boundary. Observations suggest that CH\(_4\) can escape from the top of the atmosphere by sputtering but the same effect can be generated in models by applying a larger eddy diffusivity which is the approach we have taken here. Condensed species have zero flux at both the upper and lower boundaries. Table [1](#tab:molecules){reference-type="ref" reference="tab:molecules"} provides a list of the molecules in our model. The mixing ratio of N\(_2\) is set according to the observational data and held fixed, with values below 50 km taken from the Huygens observations and above 1000 km from Cassini/UVIS data. Between 50 and 1000 km the mixing ratio is assumed to be 0.98. The mixing ratio of CH\(_4\) is fixed to the observed (super-saturated) values below the tropopause and allowed to vary above this. ## Condensation and Sublimation[\[sec:cond\]]{#sec:cond label="sec:cond"} Condensation occurs when the saturation ratio, \(S\), of a molecule is greater than 1. S is defined as n(\(x\))/n\(_{sat}\)(s), where n(\(x\)) is the gas phase mixing ratio of species \(x\) and n\(_{sat}\)(\(x\)) is its saturated density derived from the saturated vapor pressure. For \(S\) \(<\) 1, condensation is switched off and sublimation of any adsorbed molecules can occur. The abrupt change in behavior at \(S\) = 1 can lead to numerical instabilities where the system oscillates between the condensation and sublimation regimes. In previous Titan models various methods have been used to smooth out the transition and prevent such instabilities. For example, parameterized the condensation rate in terms of \(S\): \]\hbox{Loss rate} \propto-\frac{S-1}{S}\[ This results in a relatively constant loss rate as a function of \(S\). A more complicated expression was used by to ensure that the loss rate increases with increasing saturation ratios: \]\hbox{Loss rate} \propto-(S-1) \frac{exp(-0.5/ln(S+1)^2)}{ln(S+1)^2} \hbox{~~~for S \(>\) 1}\[ Other expressions that have been invoked include \]\hbox{Loss rate} \propto-ln S\[. Here we use a numerically stable method to determine the net condensation rate. The rate at which molecules condense on to a pre-existing aerosol particle is given by the collision rate with the particle: \]k_{c} = \alpha_x \sigma v_x n(x) \hbox{~~~molecules s\(^{-1}\)} \label{eq:freeze}\[ where \(\alpha\) is the sticking coefficient of molecule \(x\) (where \(\alpha_x\) \(\leq\) 1), \(\sigma\) is the collisional cross-section of the particles, \(v_x\) is the gas phase velocity of \(x\), and \(n(x)\) is its number density. For a pure ice the saturated vapor pressure is measured when the condensation and sublimation processes are in equilibrium. In this scenario \]k_c n_{sat}(x) = k_s \Theta_x \label{eq:sub}\[ where \(k_s\) is the sublimation rate and \(\Theta_x\) is the surface coverage of molecule \(x\). In the case of a pure ice, \(\Theta_x\) = 1, and hence the sublimation rate, \(k_s\) = \(k_c n_{sat}(x)\). The net condensation rate, \(J_c\) is therefore \]J_c = \alpha_x \sigma v_x [n(x)-n_{sat}(x)\Theta_x] \hbox{~~~molecules s\(^{-1}\)}\[ When sublimation is taking place from a mixture of ices (rather than from pure ice) \(\Theta\) will be less than 1 and the resulting gas phase abundance will be lower than the saturated value. \(\Theta\) is calculated from \]\Theta = n(x^c)/\Sigma_y n(y^c)\[ where \(n(x^c)\) is the number density of \(x\) in the condensed phase, \(\Sigma_y n(y^c)\) is the total number density of all molecules condensed on to the grain surface. We assume that the ices are well-mixed, so that the composition of the surface from which sublimation occurs reflects that of the bulk of the ice. To determine the saturated densities used in this paper we use the expressions for the saturated vapor pressure given in Table [2](#tab:ant){reference-type="ref" reference="tab:ant"}. The values from these fits are extrapolated as necessary to provide saturation vapor pressures over a wider range of temperatures. Figure [\[fig:comp\]](#fig:comp){reference-type="ref" reference="fig:comp"} compares the predicted mixing ratios of HCN and C\(_2\)H\(_2\) with the value predicted directly from the saturated vapor pressure. It can be seen that the model produces good agreement with the saturation vapor pressure in regions where the gas is saturated. ## [\[sec:sed\]]{#sec:sed label="sec:sed"}Sedimentation and Haze Formation We assume that the abundance, size and location of the aerosol particles is fixed. In reality the particles do not remain at the same altitude but rather sediment out towards the surface of Titan, taking any condensates with them. To mimic this effect we have included a loss process for condensed molecules which removes them from the model atmosphere with a rate coefficient of 10\(^{-10}\) s\(^{-1}\). All condensed species are assumed to be lost at the same rate. The assumed size of this reaction rate is somewhat arbitrary and to test the sensitivity of our results to its value we also considered a loss rate of 10\(^{-12}\) molecules s\(^{-1}\). Changing the rate was found to have no effect on the predicted gas phase mixing ratios. In addition to the condensation of ice or liquids onto existing aerosols, molecules can also be incorporated into new or existing aerosols. In this scenario the molecules are then unavailable for return to the gas via sublimation and are permanently removed from the gas. This process is simulated using rates that are proportional to the collision rates between aerosols (assuming mean radii provided by ) and molecules. We simulate this by adding reactions that remove the molecules from the gas with \]\hbox{X} + haze = haze \hbox{~~~k = \(\beta v \sigma n_g\) s\(^{-1}\)} \label{eq:aero}\[ where \(n_g\) is the mixing ratio of aerosol particles, and \(\beta\) is an efficiency factor ranging from 0.01 to 10 depending on the molecule. The value of \(\beta\) was chosen for each molecule to maximize the agreement of the models with the observations. The molecules to be removed in this way are HCN (\(\beta\) = 0.01), C\(_2\)H\(_3\)CN (\(\beta\) = 0.1), HC\(_3\)N and HNC (\(\beta\) = 10) C\(_2\)H\(_5\)CN (\(\beta\) = 1). Other molecules are assumed not to condense in this way--for these molecules agreement of the models with observations is sufficiently good without invoking an additional loss mechanism such as haze formation. # Results {#sec:results} ## The Effect of Condensation Processes We present the results of three models with different assumptions about the condensation and sublimation. Model A is a gas phase only model, with no condensation. Model B includes condensation and sublimation processes as outlined in Section [\[sec:cond\]](#sec:cond){reference-type="ref" reference="sec:cond"}, and the sedimentation of aerosol particles and their condensates. Model C extends Model B to include the removal of molecules from the gas by haze formation. The model parameters are summarized in Table [3](#tab:model){reference-type="ref" reference="tab:model"}. The largest effects are seen for the biggest molecules and in particular for those that contain nitrogen. The addition of sedimentation increases the rate of removal of these species from the gas in the lower atmosphere and improves agreement with the observations. However, some molecules are still found to be over-abundant. Further improvement is achieved between 200 and 600 km for HCN, HNC, HC\(_3\)N and C\(_2\)H\(_5\)CN if these molecules are assumed to be incorporated into haze particles. Below we discuss the chemistry of several species in more detail. ## Distribution of Nitrogen Molecules ### NH\(_3\) In the lower atmosphere upper limits of the NH\(_3\) abundance are provided by Herschel/SPIRE measurements [65--100 km; @teanby13] and from CIRS/Cassini limb observations. In the upper atmosphere the abundance is derived from Cassini/INMS of NH\(_4^+\) at 1100 km. claim a detection of NH\(_3\) in the ionosphere between 950 and 1200km. Their value is an order of magnitude larger than that derived by Vuitton et al. and its origin is a matter of debate. It is possible that this high value is due to spent hydrazine fuel. Our model abundances in the upper atmosphere are a factor of 10 lower than the observations of (Figure [\[fig:nchem1\]](#fig:nchem1){reference-type="ref" reference="fig:nchem1"}). Below 250 km our models are considerably lower (but consistent with) the upper limits derived by and. The main formation processes for NH\(_3\) are \]\begin{aligned} \hbox{C\(_2\)H\(_3\)} + \hbox{NH\(_2\)} & \longrightarrow \hbox{NH\(_3\)} + \hbox{C\(_2\)H\(_2\)} & \hbox{\(<\) 800 km} \end{aligned}\[ with destruction by photodissociation. As discussed by the formation of NH\(_3\) via neutral-neutral reactions depends on the presence of NH\(_2\) which is not efficiently produced in Titan's atmosphere. The inclusion of ion-molecule chemistry may lead to higher abundances of NH\(_3\). ### HCN Observations of HCN have been made from 100 km to 1000 km. The millimeter observations of covered the whole disk and were mainly sensitive to the mid-latitude and equatorial regions. Observations from Cassini/CIRS, UVIS, and INMS provide abundance information between 400 and 1000 km. Abundances in the lower atmosphere are also provided by from Keck observations. used Voyager 1 Ultraviolet Spectrometer measurements to determine abundances between 500 and 900km, although the inferred abundances are much higher than other estimates. The differences between the Voyager 1 HCN abundances and those from Cassini may be due to solar cycle variations. Investigating such differences is beyond the scope of this work. Overall our models are in good agreement with the observational data (Figure [\[fig:nchem1\]](#fig:nchem1){reference-type="ref" reference="fig:nchem1"}). We find that condensation and sublimation are important for HCN below 500 km. The best fit to the observations is obtained with Model C (Figure [\[fig:nchem1\]](#fig:nchem1){reference-type="ref" reference="fig:nchem1"}), where sedimentation and haze formation reduce the abundance of HCN below 500 km. The main formation processes are \]\begin{aligned} \hbox{HNC} + \hbox{H} & \longrightarrow \hbox{HCN} + \hbox{H} & \hbox{300--800 km}\\ \hbox{CN} + \hbox{CH\(_4\)} & \longrightarrow \hbox{HCN} + \hbox{CH\(_3\)} & \hbox{200--600 km}\\ \hbox{N} + \hbox{CH\(_2\)} & \longrightarrow \hbox{HCN} + \hbox{H} & \hbox{600--900 km}\\ \hbox{C\(_2\)H\(_3\)CN} + h\nu & \longrightarrow \hbox{HCN} + \hbox{C\(_2\)H\(_2\)} & \hbox{\(<\) 1000 km}\\ \hbox{H\(_2\)CN} + \hbox{H} & \longrightarrow \hbox{HCN} + \hbox{H}_2 & \hbox{900-1300 km} \end{aligned}\[ Photodissociation plays a role in both the formation of HCN (via photodisssociation of C\(_2\)H\(_3\)CN above 1000 km) and in its destruction (forming CN and H). Below 200 km destruction is by \]\hbox{C\(_2\)H\(_3\)} + \hbox{HCN} \longrightarrow \hbox{C\(_2\)H\(_3\)CN} + \hbox{H}\[ ### HNC The first observations of HNC in Titan were made using Herschel/HIFI by. These measurements do not allow the exact vertical abundance profile to be determined. Several possible profiles can fit the data depending on the mixing ratio and the cut-off altitude assumed. suggest two possible profiles: one where the mixing ratio of HNC is 1.4 \(\times\) 10\(^{-5}\) above 900 km (shown in Figure [\[fig:nchem1\]](#fig:nchem1){reference-type="ref" reference="fig:nchem1"}) and another where the mixing ratio is 6 \(\times\) 10\(^{-5}\) above 1000 km. Our models fall between these two ranges. More recently used ALMA to detect HNC. They found that the emission mainly originates at altitudes above 400 km and that there are two emission peaks that are not symmetrical in longitude. We are able to match their best fit profile reasonably well with model C (green line; Figure [\[fig:nchem1\]](#fig:nchem1){reference-type="ref" reference="fig:nchem1"}), where HNC forms haze providing the best agreement with the data at lower altitudes. The main formation channels of HNC are \]\begin{aligned} \hbox{C\(_2\)H\(_3\)CN} + h\nu & \longrightarrow \hbox{HNC} + \hbox{C\(_2\)H\(_2\)} & \hbox{\(<\) 900 km}\\ \hbox{H\(_2\)CN} + \hbox{H} & \longrightarrow \hbox{HNC} + \hbox{H\(_2\)} & \hbox{\(>\) 900 km}\\ \hbox{N} + \hbox{CH\(_3\)} & \longrightarrow \hbox{HNC} + \hbox{H} & \hbox{\(>\) 900 km} \end{aligned}\[ The main destruction process is by reaction with H atoms forming HCN. This reaction has an activation barrier. In the literature the value for the activation barrier ranges from 800 to 2000 K. Here we are using the rate from the KIDA database which has the highest activation barrier of 2000 K. used the lowest value (800 K), resulting in more efficient HCN production and consequently a lower gas phase abundance of HNC than we see here. We find that reducing the activation barrier does indeed reduce the mixing ratio of HNC but does not result in a good fit to the ALMA observations in this region (Figure [\[fig:HNCcomp\]](#fig:HNCcomp){reference-type="ref" reference="fig:HNCcomp"}). ### HC\(_3\)N HC\(_3\)N has been observed at altitudes from 200 to 1000 km. Below 500 km our models are in excellent agreement with the observations if it is assumed that HC\(_3\)N forms aerosols and thus is permanently removed from the gas (Figure [\[fig:nchem1\]](#fig:nchem1){reference-type="ref" reference="fig:nchem1"}* (bottom left)*). Condensation and sublimation alone result in an over-estimate of the abundance compared to the observations in this region. Good agreement is also seen for all models between 500 km and 700 km. Above this our models tend to under predict the HC\(_3\)N abundance. Below 100 km the mixing ratio follows the saturation level, so that below this altitude the mixing ratio is much reduced compared to the gas only model. Better agreement with the observations below 400 km is obtained in the haze formation model where condensed molecules are assumed to be incorporated into aerosol particles and removed from the gas. The main formation process below 1000 km is \]\hbox{C\(_3\)N} + \hbox{CH\(_4\)} \longrightarrow \hbox{HC\(_3\)N} + \hbox{CH}_3\[ and above 800 km by \]\begin{aligned} \hbox{CN} + \hbox{C\(_2\)H\(_2\)} & \longrightarrow \hbox{HC\(_3\)N} + \hbox{H} \\ \hbox{C\(_2\)H\(_3\)CN} + h\nu & \longrightarrow \hbox{HC\(_3\)N} + \hbox{H}_2 \end{aligned}\[ Destruction is by photodissociation \]\hbox{HC\(_3\)N} + h\nu \longrightarrow \hbox{C\(_3\)N} + \hbox{H}\[ and by reaction with H atoms \]\hbox{H} + \hbox{HC\(_3\)N} + \hbox{M} \longrightarrow \hbox{H\(_2\)C\(_3\)N} + \hbox{M}\[ The observations show a sharp decrease in the abundance of HC\(_3\)N below 400 km. In our models this can be accounted for if HC\(_3\)N is incorporated into haze particles (Model C). An alternative explanation of meridional circulation and condensation in the polar regions has been suggested. ### C\(_2\)N\(_2\) Observations of C\(_2\)N\(_2\) have been made by Cassini/CIRS and by Cassini/INMS. The models with condensation are in very good agreement with both of these datasets (Figure [\[fig:nchem2\]](#fig:nchem2){reference-type="ref" reference="fig:nchem2"}). Without condensation the abundance in the lower atmosphere is over-estimated. The main formation route for C\(_2\)N\(_2\) is by the reaction of CN and HNC: \]\hbox{CN} + \hbox{HNC} \longrightarrow \hbox{C\(_2\)N\(_2\)} + \hbox{H}\[ with destruction via photodissociation forming CN or by \]\hbox{H} + \hbox{C\(_2\)N\(_2\)} + \hbox{M} \longrightarrow \hbox{H\(_2\)CN\(_2\)} + \hbox{M} \hbox{~~~~\(<\) 400 km}\[ ### CH\(_3\)CN Submillimeter observations with the IRAM 30 m telescope detected the CH\(_3\)CN (12-11) rotational line providing a disk average vertical profile up to 500 km, dominated by the equatorial region. Cassini/CIRS and Cassini/INMS provide estimates of the abundance above 1000 km. All models are in good agreement with the observations below 800 km, although all predict slightly lower abundances than observed between 500 and 600 km. The predicted mixing ratio at 1100 km is a factor of 10 lower than the observed value of 3 \(\times\) 10\(^{-5}\). The main formation processes are \]\begin{aligned} \hbox{CN} + \hbox{CH}_4 & \longrightarrow \hbox{CH\(_3\)CN} + \hbox{H} & \hbox{\(<\) 900 km}\\ \hbox{NH} + \hbox{C\(_2\)H\(_4\)} & \longrightarrow \hbox{CH\(_3\)N} + \hbox{H} & \hbox{\(>\) 900 km}\\ \hbox{CH\(_2\)NH} + \hbox{CH} & = \hbox{CH\(_3\)CN} + \hbox{H} & \hbox{\(<\) 900 km} \end{aligned}\[ with destruction by \]\begin{aligned} \hbox{CH\(_3\)CN} + \hbox{H} & \longrightarrow \hbox{CN} + \hbox{CH}_4 & \hbox{\(<\) 1200 km} \end{aligned}\[ ### C\(_2\)H\(_3\)CN Several observations have placed upper limits on the abundance of C\(_2\)H\(_3\)CN. used the IRAM 30m telescope to determine upper limits between 100 and 300 km. Cassini/INMS has provided an upper limit of 4 \(\times\) 10\(^{-7}\) at 1077 km, while Cassini/INMS determined a mixing ratio of 3.5 \(\times\) 10\(^{-7}\) at 1050 km and found 10\(^{-5}\) at 1100 km from observations of ions. Cordiner et al. (private communication) have detected C\(_2\)H\(_3\)CN in the submillimeter and found an average abundance of 1.9 x 10\(^{-9}\) above 300 km.The model abundance of C\(_2\)H\(_3\)CN in the upper atmosphere is within a factor of 2 of the value but 50 times higher than and. None of our models have a constant mixing ratio with altitude between 100 and 300 km as derived from the IRAM observations (Figure [\[fig:nchem2\]](#fig:nchem2){reference-type="ref" reference="fig:nchem2"}). Model B and C (which include condensation) are consistent with the derived mixing ratio at a particular altitude, but neither reproduce the constant value between 100 and 300 km. In the upper atmosphere all models predict mixing ratios within a factor a 3 of the result but are over-abundant compared to the other measurement in this region. The main production mechanism is by reaction of CN with C\(_2\)H\(_4\): \]\begin{aligned} \hbox{C\(_2\)H\(_3\)} + \hbox{HCN} & \longrightarrow \hbox{C\(_2\)H\(_5\)CN} + \hbox{H} & \hbox{\(<\) 200 km}\\ \hbox{CN} + \hbox{C\(_2\)H\(_4\)} & \longrightarrow \hbox{C\(_2\)H\(_3\)CN} + \hbox{H} & \hbox{400--800 km}\\ \hbox{H\(_2\)C\(_3\)N} + \hbox{H} & \longrightarrow \hbox{C\(_2\)H\(_3\)CN} & \hbox{\(>\) 400 km} \end{aligned}\[ Gas phase destruction processes are \]\begin{aligned} \hbox{C\(_2\)H\(_3\)CN} + h\nu & \longrightarrow \left\{ \begin{array}{ll} \hbox{HCN} + \hbox{C\(_2\)H\(_2\)}\\ \hbox{HNC} + \hbox{C\(_2\)H\(_2\)}\\ \hbox{HC\(_3\)N} + \hbox{H}_2 \end{array} \right. \end{aligned}\[ Below 400 km haze formation and sedimentation of aerosol particles play an important role in determining the gas mixing ratio in Model C. ### C\(_2\)H\(_5\)CN Upper limits for the abundance of C\(_2\)H\(_5\)CN have been determined between 100 and 300 km from IRAM 30m observations, with abundances in the upper atmosphere provided by Cassini/INMS data. More recently detected this molecule above 200 km using ALMA. Our models over-estimate the abundance of C\(_2\)H\(_5\)CN in the upper atmosphere (Figure [\[fig:nchem2\]](#fig:nchem2){reference-type="ref" reference="fig:nchem2"}), probably because we do not include ion chemistry [for a discussion of this point see @loison15]. Below 700 km, Model C is in excellent agreement with the ALMA data of Cordiner et al. The main formation process is \]\hbox{CH\(_2\)CN} + \hbox{CH\(_3\)} + \hbox{M} \longrightarrow \hbox{C\(_2\)H\(_5\)CN} + \hbox{M}\[ Destruction is by photodissociation \]\begin{aligned} \hbox{C\(_2\)H\(_5\)CN} + h\nu & \longrightarrow \left\{ \begin{array}{ll} \hbox{C\(_2\)H\(_4\)} + \hbox{HCN} \\ \hbox{CH}_3 + \hbox{CH\(_2\)CN} \\ \hbox{C\(_2\)H\(_3\)CN} + \hbox{H\(_2\)} \end{array} \right. \end{aligned}\[ or by reaction with CH, C\(_2\)H\(_3\) or C\(_2\)H. \]\begin{aligned} \hbox{C\(_2\)H\(_5\)CN} + \hbox{CH} & \longrightarrow \hbox{CH\(_2\)CN} + \hbox{C\(_2\)H\(_4\)}\\ \hbox{C\(_2\)H\(_5\)CN} + \hbox{C\(_2\)H} & \longrightarrow \hbox{HC\(_3\)N} + \hbox{C\(_2\)H\(_5\)} & \hbox{\(<\) 300 km}\\ \hbox{C\(_2\)H\(_3\)CN} + \hbox{C\(_2\)H\(_3\)} & \longrightarrow \hbox{C\(_2\)H\(_3\)CN} + \hbox{C\(_2\)H\(_5\)} & \hbox{\(<\) 300 km} \end{aligned}\(\) ## [\[sec:flux\]]{#sec:flux label="sec:flux"}Condensates in Titan's Atmosphere We find several layers at which condensates are abundant with the location being molecule dependent. The first condensate layer is in the lower atmosphere around the tropopause. Here we find condensates of C\(_2\)H\(_2\), C\(_2\)H\(_4\), C\(_2\)H\(_6\), C\(_3\)H\(_8\) among others. A little further up in the atmosphere around 65--80 km several molecules have peaks in condensation e.g. HCN, C\(_4\)H\(_8\), C\(_4\)H\(_2\), C\(_2\)H\(_3\)CN, C\(_2\)N\(_2\), CH\(_3\)C\(_2\)H. Another layer of C\(_2\)H\(_3\)CN, CH\(_3\)CN, C\(_2\)H\(_5\)CN and CH\(_3\)C\(_2\)H forms around 110--130 km. Several molecules also have high condensation levels between 600 and 900 km e.g. CH\(_3\)C\(_2\)H, HC\(_5\)N, HC\(_3\)N, CH\(_3\)CN, and C\(_2\)H\(_5\)CN. Figure [\[fig:HCN_vap\]](#fig:HCN_vap){reference-type="ref" reference="fig:HCN_vap"} shows the condensation layers for HCN and HC\(_3\)N. Both these molecules have high condensate abundances between 70 and 100 km, but HC\(_3\)N has a further peak around 500 km where the atmospheric temperature dips, and the gas phase abundance of this molecule is high. The net flux of material falling on to the surface of Titan can be calculated from the difference between the atmospheric formation and destruction. Table [4](#tab:flux){reference-type="ref" reference="tab:flux"} presents our predictions of the surface flux of nitrogen molecules. These are in solid form and if evenly distributed across Titan's surface would create a layer 4.4 m deep over a timescale of 1 Gyr. This amount of "fixed nitrogen" could be of biological importance. # Discussion and Conclusions {#sec:conc} The removal of molecules by condensation plays an important role in determining the gas phase composition of Titan's atmosphere, as well as creating new aerosols. Condensates are found throughout the atmosphere. For the majority of molecules, condensation is most efficient below the tropopause. Larger molecules, and in particular nitrogen-bearing molecules have another condensation peak between 200 and 600 km. Relatively high abundances of condensates can also be present above 500 km if the gas phase abundance of a given molecule is high, e.g. HC\(_3\)N, HC\(_5\)N, CH\(_3\)CN and C\(_2\)H\(_5\)CN. These molecules condense in the region where Titan's haze forms. The effect is enhanced if it is assumed that some molecules can be permanently removed from the gas by being incorporated into aerosol particles. This mechanism was able to bring the abundances of HC\(_3\)N, HCN, HNC, CH\(_3\)CN and C\(_2\)H\(_5\)CN into good agreement with the observations below 600 km. Although Titan possesses a rich organic chemistry it is unclear whether this could lead to life. Photochemically produced compounds on Titan, principally acetylene, ethane and organic solids, would release energy when consumed with atmospheric hydrogen, which is also a photochemical product. speculate on the possibility of widespread methanogenic life in liquid methane on Titan. On Earth fixed nitrogen is often a limiting nutrient. Our work shows that an abundant supply of fixed nitrogen, including species of considerable complexity, is available from atmospheric photochemistry. Creating the kinds of lipid membranes that form the basis of lie on Earth depends on the presence of liquid water. Titan's atmosphere contains little oxygen and the surface temperature is well below that at which liquid water can survive. Instead surface liquids are hydrocarbons. Therefore any astrobiological processes, if present, are likely to be quite different to those on Earth. A recent paper by suggests that as alternative to lipids, membranes could be formed from small nitrogen-bearing organic molecules such as acrylonitrile (C\(_2\)H\(_3\)CN). Stevenson et al. calculate that a membrane composed of acrylonitrile molecules would be thermodynamically stable at cryogenic temperatures and would have a high energy barrier to decomposition. All of our models predict abundances of C\(_2\)H\(_3\)CN that are in agreement with observations above 500 km. Below this condensation and incorporation into haze are required to bring the predicted mixing ratios down to the values inferred from observations. If acrylonitrile were to be involved in life formation it needs to reach the surface of Titan. Our predicted flux of this molecule onto Titan's surface is 1.5 \(\times\) 10\(^7\) molecules cm\(^{-2}\) s\(^{-1}\), or \(\sim\) 41.5 gcm\(^{-2}\)/Gyr, a quantity that is potentially of biological importance.

{'timestamp': '2016-09-09T02:00:35', 'yymm': '1607', 'arxiv_id': '1607.04687', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04687'}

# Introduction latest advantages in deep learning technologies has led to explosive growth in machine learning and computer vision for building systems that have shown significant improvements in a huge range of applications such as image classification, and object detection . The fully convolutional neural networks (FCN) permit end-to-end training and testing for image labeling; holistically-nested edge detector (HED) learns hierarchically embedded multi-scale edge fields to account for the low-, mid-, and high-level information for contours and object boundaries; Faster R-CNN is a state-of-the-art object detection method depending on region proposal algorithms to predict object locations. FCN performs image-to-image training and testing, a factor that has become crucial in attaining a powerful modeling and computational capability of complex natural images and scenes. The well-suited solution to image labeling/segmentation in which each pixel is assigned a label from a pre-specified set are FCN family models. However, when it concerns the problem where individual objects need to be identified, they fails to be directly applied to. This is a problem called instance segmentation. In image labeling, two different objects are assigned with the same label as long as they belong to the same class; while in instance segmentation, in addition to obtaining their class labels, it is also demanded that objects belonging to the same class are identified individually. Exited in most organ systems as important structures, glands fulfill the responsibility of secreting proteins and carbohydrates. However, adenocarcinomas, the most prevalent type of cancer, arises form glandular epithelium. The precise instance segmentation of glands in histopathological images is essential for morphology assessment, which has been proven to be not only a valuable tool for clinical diagnosis but also the prerequisite of cancer differentiation. Nonetheless, the task of segmenting gland instances is very challenging due to the striking dissimilarity of glandular morphology in different histologic grades. In computer vision, in spite of the promising results for instance segmentation that a recently developed progress shows, it is suited for segmenting individual objects in natural scenes. With the proposal of fully convolutional network (FCN), the \"end-to-end\" learning strategy has strongly simplified the training and testing process and achieved state-of-the-art results in solving the segmentation problem back at the time. Krahenbuh *et al.* and Zheng *et al.* integrate Conditional Random Fields (CRF) with FCN to achieve finer partitioning result of FCN. However, their inability of distinguishing different objects leads to the failure in instance segmentation problem. The attempt to partition the image into semantically meaningful parts while classifying each part into one of pre-determined classes is called semantic segmentation and has already been well studied in computer vision. One limitation of semantic segmentation is its inability of detecting and delineating different instances of the same class while segmentation at the instance level being an important task in medical image analysis. The quantitative morphology evaluation as well as cancer grading and staging requires the instance segmentation of the gland histopathological slide images. Current semantic segmentation method cannot meet the demand of medical image analysis. The intrinsic properties of medical image pose plenty of challenges in instance segmentation. First of all, objects being in heterogeneous shapes make it difficult to use mathematical shape models to achieve the segmentation. As Fig.1 shows, the cytoplasm being filled with mucinogen granule causes the nucleus being extruded into a flat shape whereas the nucleus appears as a round or oval body after secreting. Second, variability of intra-and extra-cellular matrix is often the culprit leading to anisochromasia. Therefore, the background portion of medical images contains more noise like intensity gradients, compared to natural images. Several problems arose in our exploration of analyzing medical image: 1) some objects lay near the others thus one can only see the tiny gaps between them when zooming in the image on a particular area; or 2) one entity borders another letting their edges adhesive with each other. We call this an issue of *'coalescence'*. If these issues are omitted during the training phase, even there is only one pixel coalescing with another then segmentation would be a total disaster. In this paper, we aim to developing a practical system for instance segmentation in gland histology images. We make use of multichannel learning, region, boundary and location cues using convolutional neural networks with deep supervision, and solve the instance segmentation issue in the gland histology image. Our algorithm is evaluated on the dataset provided by MICCAI 2015 Gland Segmentation Challenge Contest and achieves state-of-the-art performance among all participants and other popular methods of instance segmentation. We conduct a series of experiments in comparison with other algorithms and proves the superiority of the proposed framework. This paper is arranged as follows. In section [2](#related){reference-type="ref" reference="related"}, a review of previous work in relative area is presented. In section [3](#method){reference-type="ref" reference="method"}, the complete methodology of the proposed framework of gland instance segmentation is described. In section [4](#exp){reference-type="ref" reference="exp"}, a detailed evaluation on this method is demonstrated. In section [5](#con){reference-type="ref" reference="con"}, we give our conclusion. # Related Work {#related} In this section, a retrospective introduction about instance segmentation will be delivered. Then, to present related work about our framework as clear as possible, information about channels will be delivered respectively, preceded by an overall review of multi-channel framework. ## Instance segmentation Instance segmentation, a task requires distinguishing contour, location, class and the number of objects in image, is attracting more and more attentions in image processing and computer vision. As a complex problem hardly be solved by traditional algorithms, more deep learning approaches are engaged to solve it. For example, SDS precedes with a proposal generator and then two parallel pathways for region foreground and bounding box are combined as outcome of instance segmentation. Hypercolumn, complishes instance segmentation by utilizing hypercolumn features instead of traditional feature maps. MCNs category predicted pixels via the result of object detection. In DeepMask and SharpMask two branches for segmentation and object score are engaged. Different form DeepMask, InstanceFCN exploits local coherence rather than high-dimensional features to confirm instances. DCAN, the winner of 2015 MICCAI who shares the same dataset with us, combine contour and region for instance segmentation. To sum up, most of the models mentioned above make contributions to instance segmentation by integrating more than one CNN models to provide proposals and do segmentation. ## Multichannel Model Inspired by models mentioned above, since more than one model should engaged to solve instance segmentation problem, in another words, more than one kind of information is required, then building up a multichannel framework is also a plausible method. Multichannel model usually utilized to integrate features of various kinds to achieve more satisfying result. As far as we know, multichannel framework is rarely seen in instance segmentation of medical images. It can be seen in grouping features, face recognizing and image segmentation, which leverage a bag-of-feature pipeline to improve the performance. In our multichannel framework, three channels aim at segmentation, object detection and edge detection are fused together. Related work about them are introduced respectively as follows. ### Image Segmentation Image segmentation aim at producing pixelwise labels to images. In neural network solution, the fully convolutional neural network takes the role of a watershed. Before that, patchwise training is common. Ciresan *et al.* utilize DNN to segment images of electron microscope. Farabet *et al.* segment natural scene and label them. Liu *et al.* extract features of different patches from superpixel. Then CRF is trained to provide ultimate segmentation result. FCN puts forward a more efficient model to train end-to-end. After that, fully convolutional network attracts people's attention. U-net preserve more context information by maintaining more feature channels at up-sampling part compared to FCN. Dai *et al.* improve FCN model to solve instance segmentation problems. We leverage the FCN model to produce the information of probability masks, as the region channel in our framework. ### Object Detection Object detection problem requires a system locate objects of different classes within an image. A common deep convolutional neural network solution is usually running a classier on candidate proposals and many models have been arose and improved on the basis of it. R-CNN is a representative approach that proposals are generated by an unsupervised algorithm and classified by SVMs by features extracted by DNN. To accelerate R-CNN, fast R-CNN and faster R-CNN are put forward one after another. Different from R-CNN, DeepMultiBox generate proposals by using DNN. The network for proposal producing in OverFeat share weights with network designed for classification tasks and reconcile results of classification and proposals as the ultimate result. YOLO is an end-to-end model that predict proposals and class probabilities simultaneously by regard object detection problem as a single regression problem. ### Edge detection Approaches to computational edge detection play a fundamental role in the history of image processing. In recent years, solutions of neural network, at once flourishing and effective, bring about a new access towards solving complicated edge detection problems. HED earns hierarchically embedded multi-scale edge fields to account for the low-, mid-, and high-level information for contours and object boundaries. DeepEdge also utilize multi-scale of image to solve this problem by using deep convolutional neural network. Ganin *et al.* propose an approach to solve edge detection problem by integrate the neural network with the nearest neighbor search. Shen *et al.* propose a new loss function and improve the accuracy of counter detection via a neural network. ## Previous work Earlier conference version of our approach were presented in Xu *et al.*. Here we further illustrate that: (1) we add another channel-detection channel-in this paper, due to the reason that the region channel and the detection channel complement each other; (2) this framework achieves state-of-the-art results; (3) to address the problem of images with rotation invariance, we find a new data augmentation strategy that is proven to be effective. (4) ablation experiments are carried out to corroborate the effectiveness of the proposed framework. # Method In this section, we will introduce details about our framework (as shown in Fig. [\[model1\]](#model1){reference-type="ref" reference="model1"}). By integrating the information generated from different channels, our multichannel framework is capable of instance segmentation. Aiming at solving this problem, we select three channels, foreground segmentation channel for image segmentation, object detection channel for gland detection and edge detection channel. The reason of choosing these three channels is based on the fact that information of region, contour and location contributes receptively and complimentarily to our ultimate purpose and the joint effort of them will perform much better than each of them alone. In our framework, effects of different channels are distinct. The foreground segmentation channel in our framework distinguishes the foreground and background of images. Targeted regions in pathological contains complex morphological features. It is common that glands grow close to one another. This, however, will bring about a negative effect for algorithm that the distance between two adjacent glands are too diminutive to distinguish by computer. Machines tend to conflate two glands as a whole even there do exist a gap between them. Therefore, the object detection channel is in demand of designating the bounding box of each gland to which foreground pixels in that box are belonged. In regard to the overlapping area of bounding boxes, glands boundaries are predicted by the edge detection channel. As for the area that glands are close, edge detection fails to precisely predict boundaries of glands and requires the assistance of object detection channel. Only under the joint effort of various kinds of information, can instance segmentation problem be properly solved. ## Foreground Segmentation Channel The foreground segmentation channel distinguishes glands from the background. With the arising of FCN, image segmentation become more effective thanks to the end-to-end training strategy and dense prediction on attribute-sized images. FCN replace the fully-connected layer with a convolutional layer and upsample the feature map to the same size as the original image through deconvolution thus an end-to-end training and prediction is guaranteed. Compared to the previous prevalent method sliding window in image segmentation, FCN is faster and simpler. FCN family models have achieved great accomplishment in labeling images. Usually, the FCN model can be regarded as the combination of a feature extractor and a pixel-wise predictor. Pixel-wise predictor predict probability mask of segmented images. The feature extractor is able to abstract high-level features by down-sampling and convolution. Though, useful high-level features are extracted, details of images sink in the process of max-pooling and stride convolution. Consequently, when objects are adjacent to each other, FCN may consider them as one. It is natural having FCN to solve image segmentation problems. However, instance segmentation is beyond the ability of FCN. It requires a system to differentiate instances of the same class even they are extremely close to each other. Even so, the probability mask produced by FCN still performs valuable support on solving instance segmentation problems. To compensate the resolution reduction of the feature map due to downsampling, FCN introduce the skip architecture to combine deep, semantic information and shallow, appearance information. Nevertheless, DeepLab proposes the FCN with atrous convolution that empowers the network with the wider receptive field without downsampling. Less downsampling layer means less space-invariance brought by downsampling which is benefit to the enhancement of segmentation precision. Our foreground segmentation channel is a modified version of FCN-32s of which strides of pool4 and pool5 are 1 and subsequent convolution layers enlarge the receptive field by the atrous convolution. Given an input image \(X\) and the parameter of FCN network is denoted as \({w}_{s}\), thus the output of FCN is \[{P}_{s}\left(Y^{*}_{s}=k \mid X;w_{s}\right) = {\mu}_{k}\left(h_{s}\left(X,w_{s}\right)\right),\] where \(\mu(\cdot)\) is the softmax function. \(\mu_{k}(\cdot)\) is the output of the \(k\)th category and \(h_{s}(\cdot)\) outputs the feature map of the hidden layer. ## Object Detection Channel The object detection channel detects glands and their locations in the image. The location of object is helpful on counting number and identifying the range of objects. According to some previous works on instance segmentation, such as MNC, confirmation of the bounding-box is usually the first step towards instance segmentation. After that, segmentation and other options are carried out within bounding boxes. Though this method is highly approved, the loss of context information caused by limited receptive fields and bounding-box may exacerbate the segmentation result. Consequently, we integrate the information of location to the fusion network instead of segmenting instances within bounding boxes. To achieve the location information, Faster-RCNN, a state-of-the-art object detection model, is engaged to solve this problem. In this model, convolutional layers are proposed to extract feature maps from images. After that, Region Proposal Network (RPN) takes arbitrary-sized feature map as input and produces a set of bounding-boxes with probability of objects. Region proposals will be converted into regions of interest and classified to form the final object detection result. Filling is operated in consonance with other two channel and to annotate the overlapping area. Sizes of various channels should be the same before gathering into the fusion network. To guarantee the output size of object detection channel being in accordance with other channels, we reshape it and change the bounding box into another formation. The value of each pixel in region covered by the bounding box equals to the number of bounding boxes it belongs to. For example, if a pixel is in the public area of three bounding boxes, then the value of that pixel will be three. We denote \(w_{d}\) as the parameter of Faster-RCNN and \(\phi\) represents the filling operation of bounding box. The output of this channel is \[P_{d}\left(X,w_{d}\right) = \phi\left(h_{d}\left(X,w_{d}\right)\right).\] \(h_{d}\left(\cdot\right)\) is the predicted coordinate of the bounding box. ## Edge Detection Channel The edge detection channel detects boundaries between glands. The combination of merely the probability mask predicted by FCN and the location of glands tend to fuzzy boundaries of glands, especially between adjacent objects, consequently it is tough to distinguish different objects. To receive precise and clear boundaries, the information of edge is crucial which has also been proved by DCAN. The effectiveness of edge in our framework can be concluded into two aspects. Firstly, edge compensate the information loss caused by max-pooling and other operations in FCN. As a result, the contours become more precise and the morphology become more similar to the ground truth. Secondly, even if the location and the probability mask are confirmed, it is unavoidable that predicted pixel regions of adjacent objects are still connected. Edge, however, is able to differentiate them apart. As expected, the synergies among region, location and edge finally achieve the state-of-the art result. The edge channel in our model is based on Holistically-nested Edge Detector (HED). It is a CNN-based solution towards edge detection. It learns hierarchically embedded multi-scale edge fields to account for the low-, mid-, and high-level information for contours and object boundaries. In edge detection task, pixels of labels are much less than pixels of back ground. The imbalance may decrease the convergence rate or even cause the non-convergence problem. To solve the problem, deep supervision and balancing of the loss between positive and negative classes are engaged. In total, there are five side supervisions which are established before each down-sampling layers. We denote \(w_{e}\) as the parameter of HED, thus the \(m\)th prediction of deep supervision is \[P^{(m)}_{e}(Y^{(m)*}_{e}=1 \mid X;w_{e})=\sigma(h^{(m)}_{e}(X,w_{e}).\] \(\sigma(\cdot)\) denotes sigmoid function-the output layer of HED. \(h^{(m)}_{e}\) represents the output of the hidden layer that relative to \(m\)th deep supervision. The weighted sum of M outputs of deep supervision is the final result of this channel and the weighted coefficient is \(\alpha\). This process is delivered through the convolutional layer. The back propagation enables the network to learn relative importances of edge predictions under different scales. \[P_{e}(Y^{*}_{e}=1 \mid X;w_{e},\alpha) = \sigma(\sum_{m=1}^{M}\alpha^{(m)}\cdot h^{(m)}_{e}(X,w_{e})).\] ## Fusing Multichannel Merely receiving the information of these three channels is not the ultimate purpose of our algorithm. Instance segmentation is. As a result, a fusion system is of great importance to maximize synergies of these three kinds of information above. It is hard for a non-learning algorithm to recognize the pattern of all these information. Naturally, a CNN based solution is the best choice. After obtaining outputs of these three channels, a shallow seven-layer convolutional neural network is used to combine the information and yield the final result. To reduce the information loss and ensure sufficiently large reception field, we once again replace downsampling with the atrous convolution. We denote \(w_{f}\) as the parameter of this network and \(h_{f}\) as the hidden layer. Thus the output of the network is \[P_{f}\left(Y_{f}^{*}=k\mid P_{s},P_{d},P_{e};w_{f}\right)=\mu_{k}\left(h_{f}\left(P_{s},P_{d},P_{e},w_{f}\right)\right).\] # Experiment {#exp} ## Dataset Our method is evaluated on the dataset provided by MICCAI 2015 Gland Segmentation Challenge Contest. The dataset consists of 165 labeled colorectal cancer histological images scanned by Zeiss MIRAX MIDI. The resolution of the image is approximately 0.62μm per pixel. 85 images belong to training set and 80 affiliate to test sets (test A contains 60 images and test B contains 20 images). There are 37 benign sections and 48 malignant ones in training set, 33 benign sections and 27 malignant ones in testing set A and 4 benign sections and 16 malignant ones in testing set B. ## Data augmentation and Processing We first preprocess data by performing per channel zero mean. The next step is to generate edge labels from region labels and perform dilation to edge labels afterward. Whether pixel is edge or not is decided by four nearest pixels (over, below, right and left) in region label. If all four pixels in region channel belongs to foreground or all of them belongs to background, then this pixel is regarded as edge. To enhance performance and combat overfitting, copious training data are needed. Given the circumstance of the absence of large dataset, data augmentation is essential before training. Two strategies of data augmentation has been carried out and the improvement of results is a strong evidence to prove the efficiency of data augmentation. In Strategy , horizontal flipping and rotation operation (\(0^\circ\), \(90^\circ\), \(180^\circ\), \(270^\circ\)) are used in the training images. Besides operations in Strategy , Strategy also includes sinusoidal transformation, pin cushion transformation and shear transformation. Deformation of original images is beneficial to the increasement of robustness and the promotion of final result. After data augmentation, a \(400 \times 400\) region is cropped from the original image as input. ## Evaluation Evaluation method is the same as the competition goes. Three indicators are involved to evaluate performance on test A and test B. Indicators assess detection result respectively, segmentation performance and shape similarity. Final score is the summation of six rankings and the smaller the better. Since image amounts of test A and test B are of great difference, we not only calculate the rank sum as the host of MICCAI 2015 Gland Segmentation Challenge Contest demands, but we also list the weighted rank sum. The weighted rank sum is calculated as: \[Weighted RS=\frac{3}{4}\sum test A Rank+\frac{1}{4}\sum test B Rank.\] The program for evaluation is given by MICCAI 2015 Gland Segmentation Challenge Contest. The first criterion for evaluation reflets the accuracy of gland detection which is called F1score. The segmented glandular object of True Positive (TP) is the object that shares more than 50% areas with the ground truth. Otherwise, the segmented area will be determined as False Positive (FP). Objects of ground truth without corresponding prediction are considered as False Negative (FN). \[F1score = \frac{2\cdot Precision\cdot Recall}{Precision + Recall}\] \[Precision = \frac{TP}{TP+FP}\] \[Recall=\frac{TP}{TP+FN}\] Dice is the second criterion for evaluating the performance of segmentation. Dice index of the whole image is \[D(G,S)=\frac{2(\mid G\cap S\mid)}{\mid G\mid +\mid S\mid},\] of which G represents the ground truth and S is the segmented result. However, it is not able to differentiate instances of same class. As a result, object-level dice score is employed to evaluate the segmentation result. The definition is as follows: \[D_{object}(G,S)=\frac{1}{2}\left[\sum_{i=1}^{n_{S}}w_{i}D(G_{i},S_{i})+\sum_{j=1}^{n_{G}}\widetilde{w}_{j}D(\widetilde{G}_{i},\widetilde{S}_{i})\right],\] \[w_{i}=\frac{\mid S_{i}\mid}{\sum_{j=1}^{n_{S}}\mid S_{j}\mid},\] \[\widetilde{w}_{i}=\frac{\mid \widetilde{G}_{i}\mid}{\sum_{j=1}^{n_{G}}\mid \widetilde{G}_{j} \mid}.\] \(n_{S}\) and \(n_{G}\) are the number of instances in the segmented result and ground truth. Shape similarity reflects the performance on morphology likelihood which plays a significant role in gland segmentation. Hausdorff distance is exploited to evaluate the shape similarity. To assess glands respectively, the index of Hausdorff distance deforms from the original formation: \[H(G,S)=\mathrm{max}\left\{\underset{x\epsilon G}{sup} \underset{y\epsilon S}{inf}\left\|x-y\right\|,\underset{y\epsilon S}{sup} \underset{x\epsilon G}{inf}\left\|x-y\right\|\right\},\] to the object-level formation: \[H_{object}(S,G)=\frac{1}{2}\left[\sum_{i=1}^{n_{s}}w_{i}H(G_{i},S_{i})+\sum_{i=1}^{n_{G}}\widetilde{w}_{i}H(\widetilde{G}_{i},\widetilde{S}_{i})\right],\] where \[w_{i}=\frac{|S_{i}|}{\sum_{j=1}^{n_{S}}|S_{j}|},\] \[\widetilde{w}_{i}=\frac{|\widetilde{G}_{i}|}{\sum_{j=1}^{n_{G}}|\widetilde{G}_{j}|}.\] Similar to object-level dice index \(n_{S}\) and \(n_{G}\) represents instances of segmented objects and ground truth. ## Result and Discussion Our framework performs well on datasets provided by MICCAI 2015 Gland Segmentation Challenge Contest and achieves the state-of-the-art result (as listed in Table ) among all participants. We rearrange the scores and ranks in this table. Our method outranks FCN and other participants based on both rank sum and weighted rank sum. Compared to FCN and FCN with atrous convolution, our framework obtains better score which is a convincing evidence that our work is more effective in solving instance segmentation problem in histological images. Though, FCN with atrous convolution performs better than FCN, for atrous convolution process less poolings and covers larger receptive fields, our framework combines information of region, location and edge to achieve higher score in the dataset. The reason why our framework rank higher, is because most of the adjacent glandular structures have been separated apart, so that more beneficial to meet the evaluation index of instance segmentation, while in FCN and FCN with atrous convolution they are not. Results of comparision are illustrated in Fig. [\[result\]](#result){reference-type="ref" reference="result"}. Ranks of test A are higher than test B in general due to the inconsistency of data distribution. In test A, most images are the normal ones while test B contains a majority of cancerous images which are more complicated in shape and lager in size. Hence, a larger receptive field is required in order to detect cancerous glands. However, before we exploit the atrous convolution algorithm, the downsampling layer not only gives the network larger receptive field but also make the resolution of the feature map decreases thus the segmentation result becomes worse. The atrous convolution algorithm empower the convolutional neural network with larger receptive field with less downsampling layers. Our multichannel framework enhances the performance based on the FCN with atrous convolution by adding two channels-edge detection channel and object detection channel. Since the differences between background and foreground in histopathological image are small (3th row of Fig. [\[result\]](#result){reference-type="ref" reference="result"}), FCN and FCN with atrous convolution sometimes predict the background pixel as gland thus raise the false positive rate. The multichannel framework abates the false positive by adding context of pixel while predicting object location. Compared to CUMedVision1, CUMedVision2 add the edge information thus results of test A improve yet those of test B deteriorate. But our method improves both results of test A and test B after combine the context of edge and location. However, white regions in gland histopathological images are of two kinds: 1) cytoplasm; and 2) there is no cell or tissue (background). The difference between these two kinds is that cytoplasm usually appears surrounded by nuclei or other stained tissue. In the image of the last row in Fig. [\[result\]](#result){reference-type="ref" reference="result"}, glands encircles white regions without cell or tissue causing that the machine mistakes them as cytoplasm. As for images of the 4th and 5th row in Fig. [\[result\]](#result){reference-type="ref" reference="result"}, glands are split when cutting images, which is the reason that cytoplasm is mistaken as the background. **Comparison with instance segmentation methods** Currently, methods suitable for instance segmentation of images of natural scenes predict instances based on detection or proposal, such as SDS, Hypercolumn and MNC. One defect of this logic is its dependence on the precision of detection or proposal. If the object escapes the detection, it would evade the subsequent segmentation as well. Besides, the segmentation being restricted to a certain bounding box would have little access to context information hence impact the result. Under the condition of bounding boxes overlapping one another, which instance does the pixel in the overlapping region belongs to cannot be determined. The overlapping area falls into the category of the nearest gland in our experiment. To further demonstrate the defect of the cascade architecture, we designed a baseline experiment. We first perform gland detection then segment gland instances inside bounding boxes. There is a shallow network (same as the fusion network) combines the information of foreground segmentation and edge detection to generate the final result. Configurations of all experiments are set the same as our method. Results are showed in Table and prove to be less effective than the proposed framework. ## Ablation Experiment ### Data Augmentation Strategy To enhance performance and combat overfitting, data augmentation is essential before training. We observe through experiments that adequate transformation of gland images is beneficial to training. This is because that glands are in various shape naturally and cancerous glands are more different in morphology. Here we evaluate the effect on results of foreground segmentation channel using Strategy and Strategy . We present the results in Table . ### Plausibility of Channels In the convolutional neural network, the main purpose of downsampling is to enlarge the receptive field yet at a cost of decreased resolution and information loss of original data. Feature maps with low resolution would increase the difficulty of upsample layer training. The representational ability of feature maps is reduced after upsampling and further lead to inferior segmentation result. Another drawback of downsampling is the space invariance it introduced while segmentation is space sensitive. The inconsistence between downsampling and image segmentation is obvious. The atrous convolution algorithm empower the convolutional neural network with larger receptive field with less downsampling layers. The comparison between segmentation performances of FCN with and without the atrous convolution shows the effectiveness of it in enhancing the segmentation precision. The foreground segmentation channel with the FCN with atrous convolution improves the performance of the multichannel framework. So does the fusion stage with the atrous convolution. Pixels belonged to edge occupy a extremely small proportion of the whole image. The imbalance between edge and non-edge poses a severe threat to the network training that may lead to non convergence. Edge dilation can alleviate the imbalance in a certain way and improve the edge detection precision. To prove that these three channels truly improve the performance of instance segmentation, we conduct the following two baseline experiments: a) merely launch foreground segmentation channel and edge detection channel; b) merely launch foreground segmentation channel and object detection channel. The results is in favor of the three-channel framework with no surprise. Results of experiments mentioned above are presented in Table . # Conclusion {#con} We propose a new framework called deep multichannel neural networks which achieves state-of-the-art results in MICCAI 2015 Gland Segmentation Challenge. The universal framework extracts features of edge, region and location then concatenate them together to generate the result of instance segmentation. In future work, this algorithm can be expanded in instance segmentation of medical images.

{'timestamp': '2016-07-20T02:11:16', 'yymm': '1607', 'arxiv_id': '1607.04889', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04889'}

# Introduction Multifractal analysis is the study of the concentration of a measure on level sets. In particular, we use measures for a hyperbolic map to study the multifractal spectrum. Multifractal analysts historically modeled an expanding map with a finite state shift. Their main result is the expression of the multifractal spectrum, which is concave and analytic. However, the multifractal spectrum is not analytic everywhere in the case of countable state Markov shifts. Iommi proves an expression for the multifractal spectrum in the setting of a countable Markov shift. Using Iommi's result, our paper proves that the multifractal spectrum has up to infinitely many phase transitions or non-analytic points. Before outlining our results, we give some background and definitions in the area of multifractal analysis. Consider the countable state shift \(\Sigma\) satisfying topological mixing and the Big Images Property. We take the measure \(\mu\) on our countable shift to be a Gibbs state and \(\alpha \in \mathbb{R}.\) The sequence \(x\) is said to have pointwise dimension \(\alpha\) if \[d_{\mu}(x):=\lim\limits_{m \rightarrow \infty}\frac{\log\mu([x_{1}...,x_{m}])}{\log|[x_{1}...,x_{m}]|}=\alpha.\] Let the set of sequences \(x\) in \(\Sigma\) such that \(d_{\mu}(x)\) does not exist be denoted as \(X'\) and let the set of sequences with pointwise dimension \(\alpha\) be denoted as \(X_{\alpha}^{s}.\) Hence, we can decompose \(\Sigma\) into a disjoint union called the multifractal decomposition based on the pointwise dimension of each \(x\) in our countable shift. We will consider a function that gives the Hausdorff dimension of sets that have a pointwise dimension. The multifractal spectrum with respect to \(\mu\) is the function \[f_{\mu}(\alpha)=\dim_{H}(X_{\alpha}^{s}).\] We remark that the multifractal spectrum is similarly defined for finite state Markov shifts. We first discuss historical results about the multifractal spectrum in the setting of finite state Markov shifts. Rand considers a cookie-cutter, which is a uniformly hyperbolic map. He uses thermodynamic formalism on the finite state Markov shift to prove that the multifractal spectrum is everywhere analytic. Next, we discuss the work of Cawley and Mauldin. They consider a fractal constructed by taking an iterated function system on a self-similar set (which yields a self-similar measure). Falconer gives more details on the construction of such a fractal. Cawley and Mauldin essentially model an iterated function system, based on countably many contractions, with a finite state Markov shift. Their methodology to prove that the multifractal spectrum is everywhere analytic involves geometric arguments. Pesin and Weiss consider a uniformly expanding map and prove an expression for the multifractal spectrum. They use a combination of thermodynamic formalism and a covering argument to prove that the multifractal spectrum is analytic everywhere. Analysing the multifractal spectrum involves using thermodynamic formalism differently in the case of measures on the countable full shift \(\Sigma.\) Iommi proves a general formula for the multifractal spectrum with respect to a measure on the countable shift and analyses when the multifractal spectrum is non-analytic. Hanus, Mauldin, and Urbański also consider the multifractal spectrum in the setting of the countable conformal iterated function system modeled by a countable Markov shift. Complementary to Iommi's result about the multifractal spectrum's non-analyticity, their paper gives additional conditions to prove that the multifractal spectrum is analytic. Sarig has developed the thermodynamic formalism on countable Markov shifts \(\Sigma.\) His work has established criteria for the existence of Gibbs and equilibrium states for potentials. This is significant because the existence of equilibrium states, which are important for topological pressure, is not guaranteed for potentials on countable Markov shifts. These measures are critical in our paper to prove results about the non-analyticity points of \(f_{\mu}(\alpha)\) in the setting of a countable Markov shift. For a thorough discussion of Sarig's work, please refer to Sarig's survey. We remark that Iommi and Jordan's work analyses the phase transitions of the pressure function and the multifractal spectrum in a similar setting to our paper. Their paper, assumes that the two potentials \(\phi\) and \(\psi\) are bounded and \(\lim\limits_{x \rightarrow 0}\frac{\psi(x)}{\log|T'(x)|}=\infty\) (in contrast, we assume that the limit is equal to \(1\)). Their result is that the multifractal spectrum has \(0\) to \(2\) phase transitions. In the paper, they take \(g\) to be a continuous function defined on the range of the suspension flow. Then, they prove that the map \(t \mapsto \mathcal{P}(tg)\) has \(0\) to \(1\) phase transition when the roof function dominates the floor function. Using these results, Iommi and Jordan prove that the multifractal spectrum has \(0\) to \(2\) phase transitions in their paper. Their paper, considers level sets of \([0,1]\) generated by Birkhoff averages and expanding interval maps that have countably many branches. This paper uses results from and. Their results include a variational characterisation of the multifractal spectrum and the existence of \(0\) to \(2\) phase transitions for \(f_{\mu}(\alpha)\) when \(\alpha_{\lim},\) which is a ratio involving the potentials \(\phi\) and \(\psi,\) equals \(0.\) In contrast, our paper assumes that \(0 < \alpha_{\lim} \le \infty\) or does not exist. Their paper also proves results on the multifractal analysis of suspension flows. Let \(T\) be an expanding interval map and \(g\) be a continuous function defined on the range of the suspension flow. Iommi and Jordan prove that the Birkhoff spectrum with respect to \(g\) has two phase transitions if the roof function dominates the geometric potential \(\log|T'|.\) We briefly discuss different examples of the multifractal spectrum's phase transitions in other settings. Researchers have studied phase transitions for non-uniformly expanding interval maps that have neutral fixed points. They use thermodynamical formalism and they respectively prove explicit formulae for the multifractal spectrum. Olivier's paper considers a cookie cutter on \([0,1]\) and in turn, takes an induced map defined on a Cantor set generated by this cookie cutter. Nakaishi's paper considers piecewise interval maps on \([0,1]\) and an induced transformation generated by these maps. We note that Nakaishi's paper is related to a paper by Pollicott and Weiss. The multifractal spectrum with respect to Bernoulli convolutions for algebraic paramters has been analysed in an example in Feng and Olivier's paper. Now that we have given some background into multifractal analysis, we outline our paper's results and methodology to obtain these results. This paper is about multifractal analysis in the setting of the countable Markov shift \(\Sigma\) and the interval \([0,1].\) In our analysis of \(f_{\mu}(\alpha),\) we use thermodynamic formalism. We take the locally Hölder potential functions \(\phi:\Sigma \rightarrow \mathbb{R}^{-}\) with Gibbs measure \(\mu\) and the metric potential \(\psi:\Sigma \rightarrow \mathbb{R}^{+}\) such that \(\psi=\log|G'|\) for some expanding map \(G.\) We use results from Sarig about Gibbs states on countable Markov shifts to prove the existence of Gibbs states for potential functions related to \(\phi\) and \(\psi.\) These Gibbs states help us show that the multifractal spectrum has phase transitions. Proving the following theorem about the analyticity of the multifractal spectrum requires using results from Sarig, Mauldin, Urbański, and Iommi. We define \(\alpha_{\lim}\) as follows: Let \(\bar{i} \in \Sigma\) be such that \(\bar{i}=(i,i...).\) Take \[\alpha_{\lim}:=\lim\limits_{i \rightarrow \infty}\frac{\phi(\bar{i})}{-\psi(\bar{i})}\] if the limit exists. We also let \(\alpha_{\inf}:=\inf\{d_{\mu}(x): x \in \Sigma\}\) and \(\alpha_{\sup}:=\sup\{d_{\mu}(x): x \in \Sigma\}.\) We remark that we will introduce a different definition of \(\alpha_{\inf}\) and \(\alpha_{\sup}\) in our paper and then prove the equivalence of both definitions of \(\alpha_{\inf}\) and \(\alpha_{\sup}.\) We prove a theorem about the behaviour of the multifractal spectrum when \(\alpha_{\lim}\) is infinite. Since some of the earlier results about the analyticity of the mutlfractal spectrum are still true for \(\alpha_{\lim}=\infty,\) we prove a theorem stating that the multifractal spectrum has \(0\) to \(1\) phase transition as provided below. Finally, we apply Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} and Theorem [\[thm:M2\]](#thm:M2){reference-type="ref" reference="thm:M2"} to the Gauss map \(G\) by defining our locally Hölder potential as \(\psi(x)=\log|G'(\pi(x))|\) with respect to the coding map \(\pi:\Sigma \rightarrow [0,1].\) We provide examples that apply Theorems [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} and [\[thm:M2\]](#thm:M2){reference-type="ref" reference="thm:M2"} to show that the multifractal spectrum has up to three phase transitions. Also, we provide an example in which the multifractal spectrum has infinitely many phase transitions. In these examples, we will use locally Hölder potentials to estimate \(\phi\) and \(\psi.\) Now, we will define some notation and concepts from thermodynamic formalism. # Thermodynamic Formalism We now introduce some concepts from thermodynamic formalism. ## Introductory Definitions from Thermodynamic Formalism Before proving our result about the phase transitions of the multifractal spectrum, we introduce some definitions from thermodynamic formalism. Denote \(S\) as our countable state space and \(A=(a_{lm})_{SxS}\) as our transition matrix of zeroes and ones. We will be treating \(S\) as \(\mathbb{N}.\) Note that \(A\) can be represented by a directed graph. We let \[\Sigma_{A}:=\{x \in S^{\mathbb{N}}: a_{x_{i}x_{i+1}}=1 \text{ for every } i \ge 1\}.\] We take \(\sigma: \Sigma_{A} \rightarrow \Sigma_{A}\) to be the standard left shift. We now define the topology for our countable state Markov shift \(\Sigma_{A}.\) These cylinder sets form the topology for \(\Sigma_{A}.\) Two important assumptions for our countable Markov shift are defined below. We provide the following remark. Since \(\Sigma_{A}\) is a non-compact shift space, the existence of topological mixing alone for \(\Sigma\) is not enough for a Gibbs measure to exist. For this reason, Sarig proves that the combination of topologically mixing and the BIP (big images and pre-images property) is sufficient and necessary for a Gibbs measure to exist. Denote \(\Sigma_{A}\) as \(\Sigma.\) We assume that \(\Sigma\) satisfies the big images and pre-images property and it is topological mixing. Now, we must define a property for functions on \(\Sigma.\) Local Hölder continuity is important in the definition of our metric for \(\Sigma.\) Consider the following metric \(d\) on the countable shift \(\Sigma.\) Take the sequences \(x=(x_{1},x_{2}...), y=(y_{1},y_{2}...) \in \Sigma.\) Find the first common, starting subword in which \(x\) and \(y\) agree. In other terms, take \(x \wedge y=x_{1}...,x_{k}\) such that \(k=\max\{m \in \mathbb{N}:x_{i}=y_{i} \text{ for all } 1 \le i \le m\}.\) For each \(i \in \mathbb{N},\) let \(0<r_{i}<1.\) Then, given \(x\) and \(y,\) we define our metric as \[d(x,y)=r_{x_{1}} \cdots r_{x_{k}}.\] We can define a potential \(\psi\) connected to this metric. In many of our examples, we will define a potential \(\psi: \Sigma \rightarrow \mathbb{R}^{+}\) as a locally constant function: \[\psi(x)=\log{r_{x_{1}}^{-1}}=-\log{r_{x_{1}}}.\] Then, \[\label{eq:B} 0 \le \prod_{j=0}^{n-1}(\exp(\psi(\sigma^{j}(x))))^{-1}=(r_{x_{1}}^{-1}r_{x_{2}}^{-1}\cdots r_{x_{n}}^{-1})^{-1}=r_{x_{1}}r_{x_{2}} \cdots r_{x_{n}}.\] Given our metric, \[\label{eq:C} |[x_{1}...,x_{n}]|=\sup\limits_{x,y \in [x_{1}...,x_{n}]}d(x,y)=r_{y_{1}}r_{y_{2}} \cdots r_{y_{n}}.\] Therefore, by Equations ([\[eq:B\]](#eq:B){reference-type="ref" reference="eq:B"}) and ([\[eq:C\]](#eq:C){reference-type="ref" reference="eq:C"}), \[\prod_{j=0}^{n-1}(\exp(\psi(\sigma^{j}(x))))^{-1} =r_{x_{1}}r_{x_{2}} \cdots r_{x_{n}}.\] In fact, the equation above relates the diameter of the cylinder set \([x_{1}...,x_{n}]\) to the potential \(\psi.\) Potentials satisfying such an inequality are called metric potentials, which are defined as follows. Thus, \(\psi(x)=\log{r_{x_{1}}^{-1}}\) is a metric potential with respect to the chosen metric \(d\) on \(\Sigma\) and the constant \(C=1.\) We can define more general metric potentials on different metrics on \(\Sigma.\) We remark that it is important for \(\psi\) to be a metric potential defined by a hyperbolic map. For instance, take a general expanding map \(T:[a,b] \rightarrow [a,b]\) for \(b > a \ge 0\) and then, assuming that \(\psi(x)=\log|T'(\pi(x))|,\) we find that \[-C\exp\left(\sum_{j=0}^{n-1}-\log|T'(\pi(x))|\right) \le |[x_{1},x_{2}...,x_{n}]| \le C\exp\left(\sum_{j=0}^{n-1}-\log|T'(\pi(x))|\right).\] Hence, given the expanding map \(T\) and metric potential \(\psi,\) we are able to relate the diameter of cylinder sets on \(\Sigma\) to the radii of balls on an interval. Defining our metric by using our potential also gives an additional expression for the local dimension of sets in \(\Sigma.\) Furthermore, we would be able to calculate the Hausdorff dimension of sets with local dimension \(\alpha,\) so we must give the following definition. We also define another form of pressure. Because \(\Sigma\) is topologically mixing, the Gurevich pressure does not depend on \(x_{1}\). Sarig proves that the Gurevich pressure is equivalent to topological pressure when \(\Sigma\) is BIP. We provide this result by Sarig (Pg 1571, Theorem 3), Mauldin, and Urbański (Pg 11, Theorem 2.1.8) below. We remark that Iommi, Jordan, and Todd (Pg 8, Theorem 2.10) proved that \(\sup\psi<\infty\) is an unnecessary condition for the previous proposition. We will provide a result that gives another way to calculate the topological pressure of a potential. Let \(\psi\) and \(\psi\) be locally Hölder. We can approximate the topological pressure of a potential \(q\phi-t\psi\) (with \(q,t \in \mathbb{R}\)) on the shift \(\Sigma\) by restricting the potential to a compact, invariant subset \(K \subset \Sigma.\) For such a set \(K,\) denote \[\mathcal{P}_{K}(q\phi-t\psi):=P((q\phi-t\psi)|_{K})\] as the restriction of the topological pressure of \(q\phi-t\psi\) to \(K.\) Sarig (Pg 1570, Theorem 2), Mauldin, and Urbański (Pg 8, Theorem 2.1.5) have proven the following proposition. Compact subsets of our countable Markov shift \(\Sigma\) include finite state Markov shifts \(\Sigma_{n}.\) We will later use a nested sequence of finite state full shifts to approximate the topological pressure of \(q\phi-t\psi\) on \(\mathbb{N}^{\mathbb{N}}.\) In fact, Mauldin and Urbański (Pg 13, Proposition 2.2.2) proved that \(P=\mathcal{P}(\phi).\) Now, we will define the locally Hölder potentials needed for our analysis of the multifractal spectrum. ## The Potentials \(\phi\) and \(\psi\) Throughout this paper, let \(\phi: \Sigma \rightarrow \mathbb{R}^{-}\) be a locally Hölder potential such that \(0 \le \mathcal{P}(\phi)<\infty.\) We assume that \(\mu\) is the Gibbs measure for \(\phi.\) We will show the existence of \(\mu\) later. Also, let \(\psi: \Sigma \rightarrow \mathbb{R}^{+}\) be a locally Hölder, metric potential with respect to an appropriate metric. We assume that \(\phi\) and \(\psi\) are non-cohomologous to each other. For brevity, we will instead state that \(\phi\) and \(\psi\) are non-cohomologous. For \(q, t \in \mathbb{R},\) consider the family of potentials \(q\phi-t\psi.\) Given this family of potentials, we will analyse the behaviour and phase transitions of a function \(T(q)\) dependent on \(q.\) This analysis is needed for our result about the multifractal spectrum's phase transitions. ## The Functions \(T(q)\) and \(\tilde{t}(q)\) and the Limit \(\alpha_{\lim}\) We will later prove that the multifractal spectrum's phase transitions are closely related to the behaviour of a family of potentials \(q\phi-t\psi.\) To follow this argument, we must define the following function. Proposition 4.3 on Pg 1892 of Iommi states the following important result. We also define a function \(\tilde{t}(q)\) which is similar to \(T(q).\) We will explain the significance of the family of potentials \(q\phi-\tilde{t}(q)\psi\) later. We need the following limit to get an expression for \(\tilde{t}(q).\) Assume that \[t_{\infty}:=\inf\{t \in \mathbb{R}: \mathcal{P}(-t\psi)<\infty\}\] exists and it is finite. When \(\alpha_{\lim}\) exists, we will prove that \[\tilde{t}(q)=-\alpha_{\lim}q+t_{\infty}.\] The following lemma is a modified version of Proposition 2.1.9 on Pg 11 in Mauldin and Urbański. We use this lemma in the proof of the following proposition. The function \(\tilde{t}(q)\) is connected to the following set \(Q.\) ## The Set \(Q\) and The Function \(\alpha(q)\) Let \(Q=\{q \in \mathbb{R}:T(q)=\tilde{t}(q)\}.\) Since \(T(q)\) is strictly convex and \(\tilde{t}(q)\) is linear, the following proposition is immediate. The following proposition immediately follows from Propositions [\[prop:CD1\]](#prop:CD1){reference-type="ref" reference="prop:CD1"}, [\[prop:L1\]](#prop:L1){reference-type="ref" reference="prop:L1"}, and [\[prop:Q1\]](#prop:Q1){reference-type="ref" reference="prop:Q1"}. We remark that Hanus, Mauldin, and Urbański considered the families of potentials \(f^{(i)},\) which is strongly Hölder, and \(\log|\phi_{i}^{'}|\), which is defined by a regular, conformal iterated function system \(\{\phi_{i}\}\) that satisfy the open set condition (both the terms regular conformal iterated function system and the open set condition are defined in Chapter 4 of ). Their potentials give that \(Q=\emptyset.\) Hence, the multifractal spectrum was analytic in their case. An important function used in multifractal analysis is \(\alpha(q),\) which is connected to \(Q.\) By Proposition [\[prop:Q1\]](#prop:Q1){reference-type="ref" reference="prop:Q1"}, \(Q=[q_{0}, q_{1}]\) for some \(q_{0},q_{1} \in \mathbb{R} \cup \{-\infty,\infty\}.\) If \(Q\) is a singleton, then our preceding definition applies for \(q_{0}=q_{1}<\infty.\) In that case, \(\alpha^{-}=\alpha^{+}=\alpha(q_{0}).\) Our analysis of \(\alpha(q)\) also depends on its extreme values. We remark that the extreme values of \(\alpha(q)\) are the supremum and infimum of the possible local dimension, which we will define later, of any sequence in \(\Sigma.\) We prove the following formula connecting \(\alpha_{\inf}\) and \(\alpha_{\sup}\) to a ratio involving \(\phi\) and \(\psi.\) Note that \(\alpha_{\lim} \in [\alpha_{\inf},\alpha_{\sup}]\) if it exists. Since \(Q=[q_{0},q_{1}]\) for some \(q_{0},q_{1} \in \mathbb{R},\) we immediately get the following decomposition from the definition of \(\alpha(q):\) \[(\alpha_{\inf},\alpha_{\sup})=\{\alpha(q):q\in Q^{\complement}\} \cup (\alpha^{+},\alpha_{\lim}) \cup \{\alpha_{\lim}\} \cup (\alpha_{\lim},\alpha^{-}) \cup \{\alpha^{-}\} \cup \{\alpha^{+}\}.\] The function \(\alpha(q)\) is always positive for every \(q \in \mathbb{R}\) because Iommi (Pg 1892, Proposition 4.3) proved that \(T(q)\) is a decreasing function of \(q.\) We will soon notice that the multifractal spectrum is connected to the functions \(T(q)\) and \(\alpha(q).\) ## The Multifractal Spectrum Before defining the multifractal spectrum we define symbolic dimension and the set \(X_{\alpha}^{s}.\) We use the Gibbs measure \(\mu\) for \(\phi.\) We now consider sets with local dimension \(\alpha\) as follows. Similarly, the local dimension of \(x \in \Sigma\) is the limit \[\lim\limits_{r \rightarrow 0}\frac{\log\mu(B(x,r))}{\log{r}}.\] We now prove a proposition that the pointwise and symbolic dimensions are equal a.e. \(x \in \Sigma.\) We give the following result as an alternate characterisation of \(\alpha_{\inf}\) and \(\alpha_{\sup}\) because Proposition [\[prop:AAA\]](#prop:AAA){reference-type="ref" reference="prop:AAA"} can be applied to all \(q \in \mathbb{R}.\) As Proposition [\[prop:AAA\]](#prop:AAA){reference-type="ref" reference="prop:AAA"} infers, we can consider the set \[X_{\alpha}:=\{x \in \Sigma: \lim\limits_{r \rightarrow 0}\frac{\log\mu(B(x,r))}{\log{r}}=\alpha\}.\] Since \(X_{\alpha}=X_{\alpha}^{s}\) a.e., we use \(X_{\alpha}^{s}\) in our definition of the multifractal spectrum. We remark that the multifractal spectrum depends on the measure \(\mu.\) Iommi's work has a theorem stating that the multifractal spectrum is a Legendre transform. Hence, we will define the concepts of Fenchel and Legendre transforms. We provide Iommi's theorem from Pg 11, Theorem 4.1 of , which proves that the multifractal spectrum is a Legendre transform. We provide the following remark. Iommi proves that \(T(q)\) is a convex function. If we take \(x=q,\) \(h(q)=T(q),\) \(h'(q)=-\alpha(q),\) and \(p=-\alpha,\) \[f_{\mu}(\alpha)=g(p)=\sup\limits_{x \in \mathbb{R}}\{px-h(x)\}=\sup\limits_{q \in \mathbb{R}}\{-q\alpha-T(q)\}=\inf\limits_{q\in\mathbb{R}}\{T(q)+q\alpha\}.\] Hence, \((T,f_{\mu})\) form a Fenchel pair and \(f_{\mu}(\alpha)\) is a Lengendre transform. Thus, we will use the following form of Iommi's theorem to prove Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"}. The following proposition is immediate. Our paper is on the multifractal spectrum's phase transitions; however, we gave the preceding proposition for completeness. In order to analyse the phase transitions of the multifractal spectrum, we must analyse the phase transitions of \(T(q).\) To do this, we must give criteria for the existence of special measures for our potential. ## Measures for Our Potentials In this section, we provide criteria for the existence and uniqueness of special measures for \(\phi\) and \(q\phi-T(q)\psi.\) We have the following (modified) result by Sarig (Pg 2, Theorem 1) and Mauldin and Urbański (Pg 14, Theorem 2.2.4). Thus, by Sarig, Mauldin, and Urbański, \(\phi\) has a corresponding unique Gibbs measure \(\mu.\) We now consider the family of potentials \(q\phi-T(q)\psi.\) An important assumption for our potentials \(-\phi,\psi\) is that they are in \(\mathcal{L}^{1}(\mu_{q})\) for each \(q \in Q^{\complement}.\) Before we provide a corollary complementing Theorem [\[thm:MR1\]](#thm:MR1){reference-type="ref" reference="thm:MR1"}, we give a definition from Pg 6 of Sarig. This definition will help us form a corollary about the function \(\alpha(q)\) with respect to \(\mu_{q}.\) Since \(\psi\) is locally Hölder and there exist \(\tilde{t}(q)<t<T(q)\) such that \(\mathcal{P}(q\phi-t\psi)<\infty\) for \(q \in Q^{\complement},\) it follows that \(\psi \in Dir(\phi).\) Now, we provide a proposition (which is a modified version of Corollary 4 on Pg 6 of Sarig ). Hence, we have that \(t \mapsto \mathcal{P}(q\phi-t\psi)\) is real analytic on \((T(q)-\varepsilon_{0}, T(q)+\varepsilon_{0}).\) We will use this important fact in the next section. Note that we can also apply Sarig's results to prove that \(\mu\) is a Gibbs measure for \(\phi.\) Furthermore, by Theorem [\[thm:S2\]](#thm:S2){reference-type="ref" reference="thm:S2"}, \(\mu\) is the unique, ergodic Gibbs state for \(\phi.\) Since we now have definitions and necessary results from thermodynamic formalism, we can now prove Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} and Theorem [\[thm:M2\]](#thm:M2){reference-type="ref" reference="thm:M2"}, which are about the phase transitions of the multifractal spectrum. # Proof of Theorems 1.1 and 1.2 We will prove that the multifractal spectrum is analytic by taking advantage of a decomposition for \((\alpha_{\inf},\alpha_{\sup}).\) We take these steps in our proof. 1. Let \(Q=[q_{0},q_{1}]\) for some \(q_{0},q_{1} \in \mathbb{R} \cup \{-\infty,\infty\}.\) We prove that the functions \(t \mapsto \mathcal{P}(q\phi-t\psi),\) \(T(q),\) and \(\alpha(q)\) are analytic on open subintervals of \(Q^{\complement}.\) Then, we prove that the multifractal spectrum is analytic on open subintervals of \(\{\alpha(q): q \in Q^{\complement}\}.\) 2. We prove that the multifractal spectrum is analytic on \((\alpha^{+},\alpha_{\lim})\) and \((\alpha_{\lim},\alpha^{-}).\) 3. Finally, we assume that \(Q=[q_{0},q_{1}]\) such that \(0<q_{0}<q_{1}<\infty\) and \(0< \alpha_{\lim} \le \infty\) exists. Our result about multifractal spectrum's phase transitions follows. ## The Set \(Q^{\complement}\) and Its Connection to \(T(q)\) First, we recall the definition of \(Q:\) \[Q:=\{q \in \mathbb{R}: T(q)=\tilde{t}(q)\}.\] We will prove that the functions \(f_{\mu}(\alpha),\) \(\alpha(q),\) and \(T(q)\) are analytic on sets related to \(Q^{\complement}.\) Fix \(q \in Q^{\complement}.\) By Proposition [\[cor:AN1\]](#cor:AN1){reference-type="ref" reference="cor:AN1"}, \(\mathcal{P}(q\phi-t\psi)\) is an analytic function of \(t\) in a neighbourhood around \(T(q).\) Hence, we can now prove that \(\alpha(q)\) is analytic on open subintervals of \(Q^{\complement}.\) We now prove some results about the multifractal spectrum. This lemma gives us a formula we need for the proof of the following proposition. The proof of Proposition [\[prop:A1\]](#prop:A1){reference-type="ref" reference="prop:A1"} also establishes two further results about \(\alpha(q)\) and \(f_{\mu}(\alpha)\) respectively. In summary, we proved that if \(T(q)\) is analytic on open subintervals of \(Q^{\complement},\) then \(\alpha(q)\) is analytic on open subintervals of \(Q^{\complement}.\) In turn, this gives us that \(f_{\mu}(\alpha)\) is analytic on open subintervals of \(\{\alpha(q):q\in Q^{\complement}\}.\) We also find that increasing and decreasing behaviour on open subintervals of \(\{\alpha(q):q\in Q^{\complement}\}\) is based on the sign of each \(q\in Q^{\complement}.\) ## The Intervals \((\alpha^{+},\alpha_{\lim})\) and \((\alpha_{\lim},\alpha^{-})\) Now that we have proven that the multifractal spectrum is analytic on open subintervals of \(\{\alpha(q):q\in Q^{\complement}\},\) we must consider the behaviour of the multifractal spectrum on other open subintervals of \((\alpha_{\inf},\alpha_{\sup}).\) Without loss of generality, we assume that \(Q=[q_{0},q_{1}]\) for \(-\infty<0<q_{0}<q_{1}<\infty.\) As stated earlier, \(\alpha^{-}=\alpha(q_{0})\) and \(\alpha^{+}=\alpha(q_{1}).\) Using our decomposition of \((\alpha_{\inf},\alpha_{\sup}),\) we know that we must prove that the multifractal spectrum is analytic on \((\alpha^{+},\alpha_{\lim})\) and \((\alpha_{\lim},\alpha^{-}).\) The multifractal spectrum is linear on \((\alpha_{\lim},\alpha^{-})\) and \((\alpha^{+},\alpha_{\lim})\) because of the endpoints of \(Q.\) We can finally prove our main theorems. ## Phase Transitions-Intervals and Points of \(Q\) Now that we have proven results relating the analyticity of \(T(q)\) to the analyticity of \(f_{\mu}(\alpha),\) we use Propositions [\[prop:Q1\]](#prop:Q1){reference-type="ref" reference="prop:Q1"}, [\[prop:A1\]](#prop:A1){reference-type="ref" reference="prop:A1"}, and [\[prop:A2\]](#prop:A2){reference-type="ref" reference="prop:A2"} to prove that the multifractal spectrum has \(0\) to \(3\) phase transitions when \(\alpha_{\lim}\) exists and \(0<\alpha_{\lim}<\infty.\) Furthermore, we prove a complementary result when \(\alpha_{\lim}=\infty.\) We provide the subsequent proposition to remind the reader about the forms \(Q\) can take. Analysing \(Q\) gives us information about the number of phase transitions for the multifractal spectrum. We give a thorough discussion of the multifractal spectrum's phase transitions in the case that \(Q\) is a closed interval with positive endpoints. ### Positive Closed Interval We assume that \(Q=[q_{0}, q_{1}]\) for some \(q_{0}, q_{1} \in \mathbb{R}\) such that \(0<q_{0}<q_{1}<\infty.\) The proofs for the following other cases for \(Q=[q_{0}, q_{1}]\) are proved in the same way as Proposition [\[prop:PCD1\]](#prop:PCD1){reference-type="ref" reference="prop:PCD1"}. In all of the following cases for \(Q,\) the multifractal spectrum ranges from having no phase transitions if \(\alpha^{-}=\alpha_{\lim}=\alpha^{+}\) to three phase transitions if \(\alpha^{-}>\alpha_{\lim}>\alpha^{+}.\) Using the same techniques as the proof to Proposition [\[prop:PCD1\]](#prop:PCD1){reference-type="ref" reference="prop:PCD1"}, we get the behaviour as outlined below. 1. If \(Q\) is a closed interval, then the multifractal spectrum has \(0\) to \(3\) phase transitions. 2. If \(Q\) is a point \(q \in \mathbb{R},\) then \(\alpha^{+}=\alpha^{-}.\) Hence, the multifractal spectrum has \(0\) to \(2\) phase transitions. 3. If \(Q\) is the half-open interval \((-\infty,q_{1}),\) then \(\alpha_{\lim} \ge \alpha^{+}.\) The multifractal spectrum would then have \(0\) to \(1\) phase transition. 4. If \(Q\) is the half-open interval \((q_{0},\infty),\) then \(\alpha^{-} \ge \alpha_{\lim}.\) The multifractal spectrum would then have \(0\) to \(1\) phase transition. 5. If \(Q\) is the open interval \((-\infty,\infty),\) then \(\alpha^{+}=\alpha^{-}=\alpha_{\lim}=\alpha_{\inf}=\alpha_{\sup}.\) Then, the multifractal spectrum is constant because \(f_{\mu}(\alpha)=t_{\infty}.\) Therefore, we have proven Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"}. Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} tells us that when \(0<\alpha_{\lim}<\infty,\) the multifractal spectrum has \(0\) to \(3\) phase transitions. We repeat the theorem for completeness: We now analyse the case when \(\alpha_{\lim}=\infty.\) We give a map that generates examples of phase transitions for \(f_{\mu}(\alpha).\) # Adaptation For the Gauss Map Up until this point, we have only considered the multifractal spectrum with respect to a locally Hölder function, defined by a general expanding map. We give a specific expanding map as follows. ## The Gauss Map The inverse branches of \(G\) are similar to a translated Gauss map. From this point, we use the full shift \(\Sigma=\mathbb{N}^{\mathbb{N}}.\) The coding map we take between \(\Sigma\) and \([0,1]\) is the continued fraction map. We use the Gauss map to define the potential \(\psi.\) ## Thermodynamic Formalism By symbolic coding, we provide definitions for our locally Hölder potentials \(\phi\) and \(\psi.\) Again, we assume that \(\phi: \Sigma \rightarrow \mathbb{R}^{-}\) is a locally Holder potential on \(\Sigma\) such that \(0 \le \mathcal{P}(\phi)<\infty.\) Again, we assume that \(\phi\) and \(\psi\) are non-cohomologous to each other. Using the potential \(\psi,\) the following lemma relates the Gauss map to the diameter of any cylinder in \(\Sigma.\) We use the following result to prove our main theorem. Since we have an expression for \(\log{|[x_{1}...,x_{n}]|},\) let us again consider subsets of \(\Sigma\) with symbolic dimension \(\alpha.\) # The Multifractal Spectrum and \(X_{\alpha}\) We now revisit the set \(X_{\alpha}^{s}\) and recall the definition of symbolic dimension. Again, we take \(\mu\) as the Gibbs measure for the potential \(\phi: \Sigma \rightarrow \mathbb{R}^{-}.\) There is also a measure \(\bar{\mu}=\mu \circ \pi^{-1}\) on \([0,1].\) Define the level set \[\pi(X_{\alpha}^{s})=\pi\left(\left\{x \in \Sigma: \lim\limits_{n \rightarrow \infty}\frac{\log{\mu([x_{1}...,x_{n}])}}{\log{|[x_{1}...,x_{n}]|}}=\alpha\right\}\right).\] There is a similar type of local dimension for \([0,1].\) Define the set \[X_{\alpha}=\{x \in [0,1]\setminus{\mathbb{Q}}: \lim\limits_{r \rightarrow 0}\frac{\log{\bar{\mu}(B(x,r))}}{\log{r}}=\alpha\}.\] We now give an important result by Iommi (Page 1891, Theorem 3.7). We have a similar result for \(X_{\alpha}.\) We outline the proof of Proposition [\[thm:G1\]](#thm:G1){reference-type="ref" reference="thm:G1"} as follows. To prove \(\dim_{H}(X_{\alpha}) \ge f_{\mu}(\alpha),\) we first notice that the geometric structure of the Gauss map makes it difficult to cover cylinder sets with neighbourhoods and vice versa. Define \(T_{n}(q)\) as the temperature function in the setting of \(\Sigma_{n}.\) Take \(\mu_{q_{n}}\) as the equilibrium state for \(q\phi-T_{n}(q)\psi.\) We provide a condition involving the hitting times of a \(\mu_{q_{n}}-\)typical cylinder set. This gives us that \(X_{\alpha}^{s} \subset \pi^{-1}(X_{\alpha}).\) Then, we use an inequality proven by Pesin and Weiss involving the pointwise dimension of \(x \in \Sigma_{n}\) and the multifractal spectrum \(f_{\mu,n}(\alpha).\) To apply these results to \(\Sigma,\) we approximate pressure in \(\Sigma\) with pressure in \(\Sigma_{n}.\) This gives us a monotone convergence argument for \(T_{n}(q)\) and it induces a similar argument involving Pesin and Weiss's inequality. Hence, the result follows. To prove \(\dim_{H}(X_{\alpha}) \le f_{\mu}(\alpha),\) we again use the behaviour of the Gauss map. Neighbouring \(m-\)cylinders nearly have equal diameters for large \(m \in \mathbb{N}.\) We exploit this fact by slightly increasing the size of cylinders around each \(x \in X_{\alpha}^{s}.\) This creates a Hausdorff cover for \(\pi^{-1}(X_{\alpha}),\) which we use in an argument for bounding the Hausdorff measure of \(X_{\alpha}.\) This gives the appropriate upper bound for \(\dim_{H}(X_{\alpha}).\) Through Proposition [\[thm:G1\]](#thm:G1){reference-type="ref" reference="thm:G1"}, we can apply the results of Theorems [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} and [\[thm:M2\]](#thm:M2){reference-type="ref" reference="thm:M2"} to the potential \(\psi=\log|G'|.\) Now, we consider examples in the next section. # Examples of Phase Transitions for the Gauss Map We consider the geometric potential \(\psi=\log|G'|\) and provide examples in which the multifractal spectrum has up to infinitely many phase transitions. Also, we apply the results of Theorems [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} and [\[thm:M2\]](#thm:M2){reference-type="ref" reference="thm:M2"} as well as provide an example when \(\alpha_{\lim}\) does not exist. Furthermore, we remark that by estimating \(\psi(x)\) for \(x=(x_{1},x_{2}...)\) with a locally Hölder potential \(2\log(x_{1}),\) we find that \(\psi\) is unbounded. Now, we consider the first of our three cases: \(\alpha_{\lim}<\infty.\) ## The Case \(\alpha_{\lim}<\infty\) As reflected by Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"}, we provide examples of potentials such that \(f_{\mu}(\alpha)\) has zero to three phase transitions. The easiest way to show this is by approximating \(\phi\) and \(\psi\) with locally constant potentials. ### Approximation with Locally Constant Potentials Since \(\phi\) and \(\psi\) are locally Hölder, we will give some of our estimates in terms of one periodic sequences. Now, we provide a critical technique used in our examples. Locally Hölder functions can be approximated using locally constant functions as follows. Let \(\Sigma=\mathbb{N}^{\mathbb{N}},\) take \(\psi(x)=\log|G'(\pi(x))|\) for each \(x \in \Sigma,\) and assume that \(\tilde{\psi}(x)\) is a locally constant function. Then, consider \(\psi-\bar{\psi}.\) This potential is locally Hölder because it is a difference of locally Hölder functions. For any arbitrary \(n-\)cylinder \([x_{1}...,x_{n}]\) in \(\Sigma\) and \(x \in [x_{1}...,x_{n}],\) \[0 \le |\psi(x)-\bar{\psi}(x)| \le V_{n}(\psi-\tilde{\psi}) \le C\theta^{n}\] for some \(C>0\) and \(\theta \in (0,1).\) Let \(\varepsilon>0.\) Since \(C>0\) is fixed, we have that for \(n \ge N\) such that \(N \in \mathbb{N}\) is large, \[0 \le |\psi(x)-\bar{\psi}(x)| \le V_{n}(\psi-\bar{\psi}) \le C\theta^{N} \le \varepsilon.\] It follows that we can approximate \(\psi\) with a locally Hölder and locally constant potential \(\bar{\psi}(x)=-\log(\frac{6}{\pi^{2}x_{1}^{2}}).\) Furthermore, we provide a formula used to calculate the topological pressure of \(q\bar{\phi}-t\bar{\psi}.\) Let \(0<p_{i}<1\) and \(0<s_{i}<1\) for each \(i \in \mathbb{N}.\) We will usually take \(\bar{\phi}(x)=\log{p_{x_{1}}}\) and \(\bar{\psi}(x)=\log{s_{x_{1}}^{-1}}.\) For most of our examples, \(s_{i}=\frac{6}{\pi^{2}i^{2}}.\) Then, for \(\varepsilon>0,\) \[\log\left(\sum_{i=1}^{\infty}p_{i}^{q}s_{i}^{t}\right)-\varepsilon \le \mathcal{P}(q\bar{\phi}-t\bar{\psi}) \le \log\left(\sum_{i=1}^{\infty}p_{i}^{q}s_{i}^{t}\right) + \varepsilon.\] We now proceed with our examples. ### Example of Zero Phase Transitions For each \(i \in \mathbb{N},\) let \(0<p_{i}<1.\) Take \(p_{i}=\frac{C}{(i+1)^{3}}\) with \(C \approx 4.9491\) chosen so that \(\sum_{i=1}^{\infty}p_{i}=1.\) For each \(x=(x_{1}...) \in \Sigma,\) define the potentials \(\phi:\Sigma\rightarrow\mathbb{R}^{-}\) and \(\psi:\Sigma\rightarrow\mathbb{R}^{+}\) as follows: \[\phi(x)=\log{p_{x_{1}}}=\log\left(\frac{4.9491}{(x_{1}+1)^{3}}\right)\] and \[\psi(x)=\log|G'(\pi(x))|.\] Throughout this example, we will estimate \(\psi(x)\) with \(\tilde{\psi}(x)=-\log(\frac{6}{\pi^{2}(x_{1})^{2}})\) and take \(\tilde{\phi}=\phi.\) We now consider the potential \(q\tilde{\phi}-t\tilde{\psi}.\) As noticed earlier, the number of phase transitions is determined by the relationship between \(T(q)\) and \(\tilde{t}(q).\) We found that \(\tilde{t}(q)=-\alpha_{\lim}q+t_{\infty}.\) The calculations for \(\alpha_{\lim}\) are as follows: Let \(\bar{j}=(j,j...)\) such that \(j \in \mathbb{N}.\) Then, \[\label{eq:K} \alpha_{\lim}=\lim\limits_{j\rightarrow\infty}\frac{\tilde{\phi}(\bar{j})}{-\tilde{\psi}(\bar{j})}=\lim\limits_{j\rightarrow\infty}\frac{-3\log{j}}{-2\log{j}}=\frac{3}{2}.\] The value \(t_{\infty}\) can be found using Lemma [\[prop:L1\]](#prop:L1){reference-type="ref" reference="prop:L1"}. Remember that \[t_{\infty}=\inf\{t\in\mathbb{R}:\mathcal{P}(-t\tilde{\psi})<\infty\}= \inf\{t\in\mathbb{R}:Z_{1}(-t\tilde{\psi})<\infty\}.\] Let \(j \in \mathbb{N}.\) We let \(V_{1,j}(\tilde{\psi})\) be the Hölder constant for \(\tilde{\psi}\) such that for every \(i,x \in [j],\) \[|\tilde{\psi}(i)-\tilde{\psi}(x)| \le V_{1,j}(\tilde{\psi}).\] Thus, we must find the infimum of all \(t\in\mathbb{R}\) such that \[\begin{aligned} \label{eq:L} Z_{1}(-t\tilde{\psi}) &=& \sum_{x_{1}=1}^{\infty}\exp\sup\limits_{x\in[x_{1}]}\left(-2t\log{x_{1}}-t\log\left(\frac{6}{\pi^{2}}\right)\right) \nonumber \\ &\le& \sum_{j=1}^{\infty}\exp\left(\log(j^{-2t})-t\log\left(\frac{6}{\pi^{2}}\right)+tV_{1,j}(\tilde{\psi})\right)<\infty \end{aligned}\] for \(t>\frac{1}{2}\) because \(V_{1,j}(\tilde{\psi}) \rightarrow 0\) as \(j \rightarrow \infty.\) Hence, \(t_{\infty}=\frac{1}{2}.\) Thus, by Equations ([\[eq:K\]](#eq:K){reference-type="ref" reference="eq:K"}) and ([\[eq:L\]](#eq:L){reference-type="ref" reference="eq:L"}), \[\tilde{t}(q)=-\frac{3}{2}q+\frac{1}{2}.\] By definition, we must have that \(Z_{n}(q\tilde{\phi}-t\tilde{\psi})<1\) in order for \(\mathcal{P}(q\tilde{\phi}-t\tilde{\psi})<0.\) As noticed earlier, we get \(Q\) by considering the values of \(q\in\mathbb{R}\) satisfy \(T(q)=\tilde{t}(q).\) Hence, we must find \(q\in\mathbb{R}\) such that \(Z_{n}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})<1.\) It is enough to consider \(q\in\mathbb{R}\) such that \(Z_{1}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})<1.\) Let \(j \in \mathbb{N}.\) We let \(V_{1,j}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})\) be the Hölder constant for \(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi}\) such that for every \(i,x \in [j],\) \[|(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})(i)-(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})(x)| \le V_{1,j}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi}).\] Furthermore, for each \(x \in [j]\) and \(\bar{j}=(j,j,j....),\) \[|\sup\limits_{x\in[x_{1}]}q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})(x)-(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})(\bar{j})|=C_{1,j}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})\] such that \(C_{1,j}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})>0.\) Hence, \[\begin{aligned} Z_{1}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi}) &=& \sum_{x_{1}=1}^{\infty}\exp\sup\limits_{x\in[x_{1}]}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})(x) \nonumber \\ &=& \sum_{j=1}^{\infty}\exp((q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})(\bar{j})+C_{1,j}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi})) \nonumber \\ &=& \sum_{j=1}^{\infty}\frac{(4.9491)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{2}q+\frac{1}{2}}}{j^{3q+2\tilde{t}(q)}}\exp(C_{1,j}(q\tilde{\phi}-\tilde{t}(q)\tilde{\psi}))=\infty. \nonumber \end{aligned}\] Hence, \(T(q)\not=\tilde{t}(q)\) for any \(q\in\mathbb{R}.\) Thus, \(Q=\emptyset.\) Therefore, by Proposition 6.3, the multifractal spectrum has no phase transitions. ### One Phase Transition We first provide pictures of the multifractal spectrum and \(T(q)\) for the following example. For each \(i \in \mathbb{N},\) let \(0<p_{i}<1.\) Take \(p_{i}=\frac{C}{(i)^{\frac{6}{5}}(\log(i+2))^{2}}\) with \(C \approx 0.67569\) chosen so that \(\sum_{i=1}^{\infty}p_{i}=1.\) For each \(x=(x_{1}...) \in \Sigma,\) define the potentials \(\phi:\Sigma\rightarrow\mathbb{R}^{-}\) and \(\psi:\Sigma\rightarrow\mathbb{R}^{+}\) as follows: \[\phi(x)=\log{p_{x_{1}}}=\log\left(\frac{0.67569}{(i)^{\frac{6}{5}}(\log(i+2))^{2}}\right)\] and \[\psi(x)=\log|G'(\pi(x))|.\] Hence, we define the locally constant potentials \(\bar{\phi}\) and \(\bar{\psi}\) as follows: \[\bar{\phi}(x):=\log{p_{x_{1}}} \text{ and } \bar{\psi}(x):=-\log\left({\frac{6}{\pi^{2}{x_{1}}^{2}}}\right)\] such that \[p_{i}=\frac{0.67569}{(i)^{\frac{6}{5}}(\log(i+2))^{2}}\] for each \(i \in \mathbb{N}.\) We will again find an explicit expression for \(\tilde{t}(q).\) Let \(\bar{j}=(j,j...)\) for any \(j \in \mathbb{N}.\) We have \[\alpha_{\lim}=\lim\limits_{j\rightarrow\infty}\frac{\bar{\phi}(\bar{j})}{-\bar{\psi}(\bar{j})}=\lim\limits_{j\rightarrow\infty}\frac{-\frac{6}{5}\log{j}}{-\log|G'(\pi(\bar{j}))|}=\lim\limits_{j\rightarrow\infty}\frac{-\frac{6}{5}\log{j}}{-2\log{j}}=\frac{3}{5}.\] Again, we have that \(Z_{1}(-t\bar{\psi})<\infty\) for \(t>\frac{1}{2},\) so \(t_{\infty}=\frac{1}{2}.\) Hence, we get that \[\tilde{t}(q)=-\frac{3}{5}q+\frac{1}{2}.\] We prove that it is possible for the set \(Q\) to equal \([q_{0},\infty)\) for some \(q_{0}>\frac{6}{5}.\) This involves using partial sums of \(\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})\) to bound \(\mathcal{P}(q\phi-\tilde{t}(q)\psi)\) below for some \(q<\frac{6}{5}\) and \(Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})\) to estimate \(\mathcal{P}(q\phi-\tilde{t}(q)\psi)\) above for a fixed \(q>\frac{6}{5}.\) First, we need to prove that there exists a \(q>\frac{6}{5}\) such that \(Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})<1.\) Again, we let \(V_{1,j}(q\bar{\psi}+-\tilde{t}(q)\bar{\psi})\) be the Hölder constant for \(q\bar{\phi}-\tilde{t}(q)\bar{\psi}\) such that for every \(i,x \in [j],\) \[|(q\bar{\phi}-\tilde{t}(q)\bar{\psi})(i)-(q\bar{\phi}-\tilde{t}(q)\bar{\psi})(x)| \le V_{1,j}(q\bar{\phi}-\tilde{t}(q)\bar{\psi}).\] The preceding estimate gives us that \[\begin{aligned} Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi}) &=& \sum_{x_{1}=1}^{\infty}\exp\sup\limits_{x\in[x_{1}]}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})(x) \nonumber \\ & \le & \sum_{j=1}^{\infty}\exp(q\bar{\phi}-\tilde{t}(q)\bar{\psi})(\bar{j})+V_{1,j}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})) \nonumber \\ &<& \sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\exp(V_{1,j}(q\bar{\phi}-\tilde{t}(q)\bar{\psi}))). \nonumber \end{aligned}\] We notice the following: Let \(\varepsilon>0\) and \(N \in \mathbb{N}\) be large. Then, \[\sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}=\sum_{j=1}^{N-1}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}+\sum_{j=N}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}.\] Let \(f(q)=\sum_{j=N}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}=\sum_{j=N}^{\infty}\frac{((0.67569)(\frac{6}{\pi^{2}})^{-\frac{3}{5}})^{q}(\frac{6}{\pi^{2}})^{\frac{1}{2}}}{(j)(\log(j+2))^{2q}}.\) We have that \(f'(q)=\) \[\sum_{j=N}^{\infty}\frac{((0.67569)(\frac{6}{\pi^{2}})^{-\frac{3}{5}})^{q}(\frac{6}{\pi^{2}})^{\frac{1}{2}}\log((0.67569)(\frac{6}{\pi^{2}})^{-\frac{3}{5}})}{(j)(\log(j+2))^{2q}}-2\sum_{j=N}^{\infty}\frac{((0.67569)(\frac{6}{\pi^{2}})^{-\frac{3}{5}})^{q}(\frac{6}{\pi^{2}})^{\frac{1}{2}}}{j\log\log(j+2)(\log(j+2))^{2q}}\] which is less than \(\varepsilon\) for a fixed \(q > \frac{6}{5}.\) Hence, \(\sum_{j=N}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\) is a decreasing function with respect to \(q.\) Let \(g(q)=\sum_{j=1}^{N-1}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}.\) Similar to what was shown earlier, \(g'(q)<0.\) Hence, \[\sum_{j=1}^{N-1}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}} \text{ and } \sum_{j=N}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\] are decreasing functions with respect to \(q.\) Then, for a fixed \(q>\frac{6}{5},\) \[\label{eq:N} \sum_{j=1}^{N-1}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}} \le \int_{1}^{N-1} \frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\, \mathrm{d}\mu \le 1-\varepsilon\] \[\label{eq:P} \text{and} \sum_{j=N}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}} \le \int_{N}^{\infty} \frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\, \mathrm{d}\mu \le \varepsilon.\] Therefore, by Equations ([\[eq:N\]](#eq:N){reference-type="ref" reference="eq:N"}) and ([\[eq:P\]](#eq:P){reference-type="ref" reference="eq:P"}), \[Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})<\sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\exp(V_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi}))) \le 1\] for a fixed \(\widehat{q}>\frac{6}{5}\) (with \(\widehat{q}\) not necessarily equal to \(q_{0}\)). Since \({\bar{\phi}}\) and \({\bar{\psi}}\) approximate \(\phi\) and \(\psi\) respectively, we will prove that there exists a value \(1<q<\frac{6}{5}\) such that \[\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})>0.\] Let \(\varepsilon>0.\) It follows that \[\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})>\sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}-\varepsilon\] because \(\bar{\psi}\) and \(\bar{\phi}\) are locally constant. We will estimate \(\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})\) by using \[\sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}.\] In particular, we get that \[\sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}>\sum_{j=1}^{25}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}>1\] for \(q=1.15.\) Since \[\label{eq:partialsum} \sum_{j=1}^{N}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{-\frac{3}{5}q+\frac{1}{2}}}{(j)(\log(j+2))^{2q}}\] is a decreasing function with respect to \(q\) and the value of sum ([\[eq:partialsum\]](#eq:partialsum){reference-type="ref" reference="eq:partialsum"}) increases as \(N \rightarrow \infty,\) there exists a value \(1<q<\frac{6}{5}\) such that \[\mathcal{P}(q{\bar{\phi}}+-\tilde{t}(q){\bar{\psi}})>0.\] Since \({\bar{\phi}}\) and \({\bar{\psi}}\) approximate \(\phi\) and \(\psi\) respectively, we get that there exists a value \(1<q<\frac{6}{5}\) such that \[\mathcal{P}(q\phi-\tilde{t}(q)\psi)>0.\] This means that \(T(q)=\tilde{t}(q)\) for all \(\widehat{q}>\frac{6}{5}\) and \(T(q)>\tilde{t}(q)\) for some \(q<\frac{6}{5}.\) Thus, we have established that \(Q=[q_{0},\infty)\) for \(q_{0}>\frac{6}{5}.\) Now, we can consider \(\alpha(q_{0})\) for \(q_{0}>\frac{6}{5}.\) Using our earlier estimates, we have that \[\begin{aligned} \int-\phi\,\mathrm{d}\mu_{q} &\le& \int-\bar{\phi}\,\mathrm{d}\mu_{q}=\sum_{j=1}^{\infty}-\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{T(q)}}{j^{\frac{6}{5}q+2T(q)}(\log(j+2))^{2q}}\log\left(\frac{0.67569}{j^{\frac{6}{5}}(\log(j+2))^{2}}\right)\nonumber \\ &<&\sum_{j=1}^{\infty}\frac{1}{j^{\frac{6}{5}q+2T(q)-1}(\log(j+2))^{2q-1}}<\infty \nonumber \end{aligned}\] and \[\begin{aligned} \int \psi\, \mathrm{d}\mu_{q} &\le& \int \bar{\psi}\, \mathrm{d}\mu_{q} =-\sum_{j=1}^{\infty}\frac{(0.67569)^{q}(\frac{6}{\pi^{2}})^{T(q)}}{j^{\frac{6}{5}q+2T(q)}(\log(j+2))^{2q}}\log\left(\frac{6}{\pi^{2}j^{2}}\right) \nonumber \\ &<&\sum_{j=1}^{\infty}\frac{1}{j^{\frac{6}{5}q+2T(q)-1}(\log(j+2))^{2q-1}}<\infty \nonumber \end{aligned}\] when \(q>\frac{6}{5}.\) We remark that both integrals are infinite if \(q<\frac{5}{6}.\) Hence, if \(q_{0}>q>\frac{6}{5},\) \[\alpha(q_{0})=\lim\limits_{q\rightarrow q_{0}^{-}}\frac{\int \phi\, \mathrm{d}\mu_{q}}{\int-\psi\, \mathrm{d}\mu_{q}}>\alpha_{\lim}.\] Using the techniques from the proof of Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"}, we get the following. Given that \(q_{0}>\frac{6}{5},\) \(\alpha(q_{0}) \ge \alpha_{\lim}=\alpha_{\inf}.\) In this case, \(f_{\mu}(\alpha)\) is analytic on \((\alpha_{\inf}, \alpha(q_{0}))\) and \((\alpha(q_{0}),\alpha_{\sup}).\) The multifractal spectrum is an increasing linear function and equals \(T(q_{0})+q_{0}\alpha\) on \((\alpha_{\inf},\alpha(q_{0})),\) is strictly concave on \((\alpha(q_{0}),\alpha_{\sup}),\) and has its maximum at \(\alpha(0).\) The multifractal spectrum has its only phase transition at \(\alpha(q_{0}).\) ### Two and Three Phase Transitions For each \(j \in \mathbb{N},\) let \(0<p_{j}<1.\) Pick a \(k\in\mathbb{N}\) and let \[p_{k}=\frac{C_{k}}{(k)^{\frac{6}{5}}(\log(k+2))^{2}}\] such that \(C_{k}>1.\) For each \(i \not= k \in \mathbb{N},\) take \[p_{i}=\frac{C}{(i)^{\frac{6}{5}}(\log(i+2))^{2}}\] with \(0<C<1\) chosen so that \(\sum_{i=1}^{\infty}p_{i}=1.\) For simplicity, we take \(C=\frac{3}{5}\) because the introduction of \(C_{k}\) must mean that \(C<0.67569.\) Note that \(p_{k}=\frac{C_{k}}{C}\frac{C}{k^{\frac{6}{5}}(\log(k+2))^{2}}.\) For each \(x=(x_{1}...) \in \Sigma,\) define the potentials \(\phi:\Sigma\rightarrow\mathbb{R}^{-}\) and \(\psi:\Sigma\rightarrow\mathbb{R}^{+}\) as follows: \[\phi(x)=\log{p_{x_{1}}}=\log\left(\frac{C}{(i)^{\frac{6}{5}}(\log(i+2))^{2}}\right)\] if \(x_{1}=i\not=k,\) \[\phi(x)=\log{p_{x_{1}}}=\log\left(\frac{C_{k}}{(k)^{\frac{6}{5}}(\log(k+2))^{2}}\right)\] if \(x_{1}=k,\) and \[\psi(x)=\log|G'(\pi(x))|.\] Let \(\bar{j}=(j,j...)\) for any \(j \in \mathbb{N}.\) Again, we approximate \(\phi\) and \(\psi\) with locally constant functions \(\bar{\phi}\) and \(\bar{\psi}.\) We let \(\bar{\phi}(x)=\phi(x)\) and \(\bar{\psi}(x)=-\log\left(\frac{6}{\pi^{2}x_{1}^{-2}}\right)\) for each \(x \in \Sigma.\) We again have that \(Z_{1}(-t\psi)<\infty\) for \(t>\frac{1}{2},\) so \(\alpha_{\lim}=\frac{3}{5}\) and \(t_{\infty}=\frac{1}{2}.\) Hence, \[\tilde{t}(q)=-\frac{3}{5}q+\frac{1}{2}.\] We will try to prove that it is possible for the set \(Q\) to equal \([q_{0},q_{1}]\) for \(q_{0}>\frac{5}{6}.\) Since the arguments are nearly identical to the previous example, we instead give an outline. This involves using \(\bar{\phi}\) and \(\bar{\psi}\) to estimate \(\mathcal{P}(q\phi-\tilde{t}(q)\psi)\) below for \(q<\frac{5}{6},\) \(Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})\) to estimate \(\mathcal{P}(q\phi-\tilde{t}(q)\psi)\) above for \(q_{0}>q>\frac{5}{6},\) and again \(\bar{\phi}\) and \(\bar{\psi}\) to estimate \(\mathcal{P}(q\phi-\tilde{t}(q)\psi)\) below for \(\frac{5}{6}<q<q_{1}<\infty.\) In the case that \(Q=[q_{0},q_{1}]\) such that \(q_{0}<\frac{5}{6}<q_{1},\) the only modification to the previous argument is the need to use \(Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})\) to estimate \(\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})\) above for \(q<\frac{5}{6}.\) For some \(q<\frac{5}{6},\) \(Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})<1\) so \(\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})<0.\) To prove that \(T(q)\) has a second phase transition at \(q_{1},\) we consider the following. Using simple analysis, we find that we need to satisfy \[\frac{\log{p_{k}}}{-2\log(k)}>\alpha_{\lim}\] in order for \(q_{1}\) to exist (because this would mean that there exists \(q>q_{1}\) such that \(\alpha(q)>\alpha_{\lim}\)). Let \(\varepsilon>0.\) Since \(\alpha_{\lim}=\lim\limits_{j \rightarrow \infty}\frac{\log{p_{j}}}{\log|G'(\pi(\bar{j}))|}=\lim\limits_{j \rightarrow \infty}\frac{\log{p_{j}}}{2\log(j)},\) it follows that for each \(j \ge J\) for some large \(J \in \mathbb{N},\) \[\alpha_{\lim}-\varepsilon \le \frac{\log{p_{j}}}{-2\log{j}} \le \alpha_{\lim}+\varepsilon.\] Thus, for \(j \ge k,\) we must satisfy \[\frac{\log{p_{k}}}{-2\log(k)}> \frac{\log{p_{j}}}{-2\log{j}}.\] Hence, choosing \(C_{k}>1\) gives us the second phase transition for \(T(q).\) Now, we must analyse the behaviour of \(\alpha(q)\) at \(q_{0}\) and \(q_{1}.\) The work to show that \[\int-\bar{\phi}\, \mathrm{d}\mu_{q}<\infty \text{ and} \int \bar{\psi}\, \mathrm{d}\mu_{q}<\infty\] for \(q>\frac{5}{6}\) is identical to the previous example. Hence, if \(q<q_{0}\le \frac{5}{6},\) \[\int-\bar{\phi}\, \mathrm{d}\mu_{q}=\infty \text{ and} \int \bar{\psi}\, \mathrm{d}\mu_{q}=\infty.\] The same results are true for \(q_{1}.\) Therefore, using the techniques from the proof of Theorem [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"}, we get either two or three phase transitions for our chosen \(\phi\) and \(\psi,\) depending on the values of \(q_{0}\) and \(q_{1}:\) 1. If \(q_{0} \le \frac{5}{6} < q_{1},\) \(\alpha(q_{0})=\alpha_{\lim}>\alpha(q_{1}).\) The multifractal spectrum is analytic on \((\alpha_{\inf},\alpha(q_{1})),\) \((\alpha(q_{1}),\alpha(q_{0})),\) and \((\alpha(q_{0}),\alpha_{\sup}).\) Furthermore, \(f_{\mu}(\alpha)\) is strictly concave on \((\alpha_{\inf},\alpha(q_{1}))\) as well as \((\alpha(q_{0}),\alpha_{\sup}),\) is linear and equals \(T(q_{1})+q_{1}\alpha\) on \((\alpha(q_{1}),\alpha(q_{0})),\) and has its maximum at \(\alpha(0).\) The multifractal spectrum has phase transitions at \(\alpha_{\lim}\) and \(\alpha(q_{1})\) in this case. 2. If \(\frac{5}{6}<q_{0}<q_{1},\) \(\alpha(q_{0})>\alpha_{\lim}>\alpha(q_{1}).\) The multifractal spectrum is analytic on \((\alpha_{\inf},\alpha(q_{1})),\) \((\alpha(q_{1}),\alpha_{\lim}),\) \((\alpha_{\lim},\alpha(q_{0})),\) and \((\alpha(q_{0}),\alpha_{\sup}).\) Furthermore, \(f_{\mu}(\alpha)\) is strictly concave on \((\alpha_{\inf},\alpha(q_{1}))\) as well as \((\alpha(q_{0}),\alpha_{\sup}),\) is linear and equals \(T(q_{1})+q_{1}\alpha\) on \((\alpha(q_{1}),\alpha_{\lim}),\) is linear and equals \(T(q_{0})+q_{0}\alpha\) on \((\alpha_{\lim},\alpha(q_{0})),\) and has its maximum at \(\alpha(0).\) Then, the multfractal spectrum has phase transitions at \(\alpha(q_{0}), \alpha_{\lim}, \text{ and } \alpha(q_{1})\) in this case. In fact, Proposition [\[prop:PCD1\]](#prop:PCD1){reference-type="ref" reference="prop:PCD1"} gives these results. ## The Case \(\alpha_{\lim}=\infty\) We provide the following example in which \(\alpha_{\lim}=\infty.\) Let \(\Sigma=\mathbb{N}^{\mathbb{N}}.\) Define \(\phi:\Sigma \rightarrow \mathbb{R}^{-}\) and \(\psi:\Sigma \rightarrow \mathbb{R}^{+}\) as follows. For each \(i \in \mathbb{N},\) let \(p_{i}=\left(\frac{1}{2}\right)^{i}.\) We remark that \(p_{i}\) is exactly the Minkowski \(?-\)function. Clearly, \(\sum_{i=1}^{\infty}p_{i}=1.\) Again, we approximate \(\psi(x)=\log|G'(\pi(x))|\) with \(\bar{\psi}(x)=-\log\left(\frac{6}{\pi^{2}x_{1}^{2}}\right)\) and we let \[\phi(x)=\bar{\phi}(x)=\log{p_{x_{1}}}=\log\left(\left(\frac{1}{2}\right)^{x_{1}}\right).\] for each \(x \in \Sigma.\) Let \(\bar{j}=(j,j,j...)\) such that \(j \in \mathbb{N}.\) Since \(\bar{\phi}\) and \(\bar{\psi}\) are locally constant, we can approximate \(\mathcal{P}(q\phi-t\psi)\) with \[\mathcal{P}(q\bar{\phi}-t\bar{\psi})=\log{Z_{1}(q\bar{\phi}-t\bar{\psi})}=\log\left(\sum_{j=1}^{\infty}\frac{(\frac{6}{\pi^{2}})^{t}}{2^{jq}(j)^{2t}}\right).\] We notice that 1. For each \(q<0,\) \[\log\left(\sum_{j=1}^{\infty}\frac{(\frac{6}{\pi^{2}})^{t}}{2^{jq}(j)^{2t}}\right)=\infty\] independent of the choice of \(t\in\mathbb{R}.\) Then, \(T(q)=\tilde{t}(q)=\infty\) for these \(q.\) 2. If \(q=0,\) \[\log\left(\sum_{j=1}^{\infty}\frac{(\frac{6}{\pi^{2}})^{t}}{(j)^{2t}}\right)<\infty\] for every \(t>\frac{1}{2}\in\mathbb{R}.\) Hence, \(\tilde{t}(0)=\frac{1}{2}<T(0).\) In fact, \(T(0)=1.\) 3. For each \(q>0,\) \[\log\left(\sum_{j=1}^{\infty}\frac{(\frac{6}{\pi^{2}})^{t}}{2^{jq}(j)^{2t}}\right)<\infty\] independent of the choice of \(t\in\mathbb{R}.\) Then, \(-\infty=\tilde{t}(q)<T(q)\) for these \(q.\) By Lemma [\[lemma:MU1\]](#lemma:MU1){reference-type="ref" reference="lemma:MU1"}, we get that \[\tilde{t}(q) = \left\{ \begin{array}{ll} \infty & \mbox{if } q < 0 \\ \frac{1}{2} & \mbox{if } q = 0 \\ -\infty & \mbox{if } q > 0. \end{array} \right.\] It follows that \[Q=\{q \in \mathbb{R}: T(q)=\tilde{t}(q)\}=(-\infty,0).\] By convention, \[\alpha(0)=\lim\limits_{q \rightarrow 0^{+}}\alpha(q).\] Also, notice that \(\alpha_{\lim}=\lim\limits_{i \rightarrow \infty}\frac{\log{p_{i}}}{\log{s_{i}}}=\infty.\) Then, \(\alpha(0)<\alpha_{\lim}=\infty=\alpha_{\sup}\) because \(T(0)=1>\tilde{t}(0)\) gives us that \[\lim\limits_{q \rightarrow 0^{+}}\int-\phi\, \mathrm{d}\mu_{q} \le \lim\limits_{q \rightarrow 0^{+}}\sum_{j=1}^{\infty}\frac{(\frac{6}{\pi^{2}})^{T(q)}j\log(2)}{2^{jq}(j)^{2T(q)}} \in (-\infty,\infty)\] \[\text{and} \lim\limits_{q \rightarrow 0^{+}}\int\psi\, \mathrm{d}\mu_{q}\le \lim\limits_{q \rightarrow 0^{+}}\sum_{j=1}^{\infty}\frac{(\frac{6}{\pi^{2}})^{T(q)}[2\log(j)-\log(\frac{6}{\pi^{2}})]}{2^{jq}(j)^{2T(q)}} \in (-\infty,\infty).\] Hence, we have the following analysis for the behaviour of the multifractal spectrum. ## The Case When \(\alpha_{\lim}\) Does Not Exist When \(\alpha_{\lim}\) does not exist, the multifractal spectrum has up to infinitely many phase transitions. We remark that Iommi and Jordan create a similar example in the setting of a suspension flow. Now, we roughly outline the procedure for creating such an example in our setting. First, we take the locally Hölder potentials \(\phi\) and \(\psi\) such that \[\phi(x)=\log{p_{x_{1}}} \text{ and } \psi(x)=\log|G'(\pi(x))|\] for each \(x=(x_{1},x_{2}...)\in\Sigma.\) Then, to define the \(p_{i}\) for each \(i \in \mathbb{N},\) we partition the natural numbers as follows. Let \(r_{0}=0\) and \(r_{1}=1.\) Consider the infinite sequence \(\{r_{k}\}_{k=2}^{\infty}\) of primes \(\{2,3,5,7,11...\}.\) We define the sets \(\{I_{k}\}_{k \in \mathbb{N}_{0}}\) as follows: \[I_{0}:=\{m \in \mathbb{N} \text{ such that } m \text{ cannot be written as any prime power of any } n \in \mathbb{N}\}.\] \[I_{1}:=\{m \in \mathbb{N} \text{ that can be written as the } 2\text{nd power of some } n \in \mathbb{N} \}\] In general, \[I_{k}:=\{m \in \mathbb{N} \text{ that can be written as the } r_{k+1}\text{st power of some } n \in \mathbb{N} \}.\] For each \(m \in \mathbb{N},\) we get that \[p_{m} =\frac{C_{k}}{m^{l_{k}}(\log(m+2))^{M_{k}}}\] if \(k \in\mathbb{N}_{0}\) and \(m \in I_{k}.\) We have increasing sequences of constants \(\{C_{k}\}_{k \in \mathbb{N}_{0}}\) and \(\{M_{k}\}_{k \in \mathbb{N}_{0}}.\) The terms of both sequences are chosen such that \(\sum_{m=1}^{\infty}p_{m}=1.\) For each \(k \in \mathbb{N} \cup \{0\},\) we have a recursive relation for the values of the sequence \(\{l_{k}\}\) stated in the expression for \(p_{m}.\) We remind the reader that locally Hölder potentials can be approximated by locally constant potentials. For each \(a_{1}(x)=m \in \mathbb{N},\) define \[s_{m}:= \frac{6}{\pi^{2}m^{2}}\] for each \(m \in \mathbb{N};\) hence, we can estimate \(\psi\) with \(\bar{\psi}(x)=\log{s_{x_{1}}^{-1}}.\) For each \(I_{k},\) we have a function \(\tilde{t}_{k}(q)\) as follows: \[\tilde{t}_{k}(q):=\inf\{t \in \mathbb{R}: \sum_{I_{k}}p_{m}^{q}s_{m}^{\tilde{t}_{k}(q)}<\infty\}.\] We define \[\tilde{t}(q)=\left\{\sup_{q \in \mathbb{R}}\tilde{t}_{k}(q): k\in \mathbb{N} \cup \{0\}\right\}.\] By construction, \(|\tilde{t}_{k}'|>|\tilde{t}_{k+1}'|\) and each \(\tilde{t}_{k}\) is linear. We get that the phase transitions for \(\tilde{t}(q)\) occur at values of \(q\) such that \(\tilde{t}_{k}(q)=\tilde{t}_{k+1}(q).\) Proceeding with the computation of these \(q,\) we find that they occur at each \(q \in \mathbb{N}.\) Hence, \(\tilde{t}(q)\) has infinitely many phase transitions. Finally, we prove that \(Z_{1}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})<1\) for each \(q \in (k,k+1) \cup (a,1)\) for some \(a \in \mathbb{R}\) and for each \(k \in \mathbb{N}.\) This gives us that \(\mathcal{P}(q\bar{\phi}-\tilde{t}(q)\bar{\psi})<1\) for those \(q;\) hence, \(T(q)=\tilde{t}(q)\) for all \(q \ge \bar{q} \ge a\) (for some \(1> \bar{q} \ge a\)). Using techniques from the previous example, we get results for a possible phase transition for the multifractal spectrum at \(\alpha(\bar{q}).\) Without loss of generality, let us assume that there is no phase transition at \(\alpha(\bar{q}).\) Hence, we have the following behaviour. \(f_{\mu}(\alpha)\) is analytic on \((\alpha(0),\alpha_{\sup}),\) \((\alpha(1),\alpha(0)),\) \((\alpha(2),\alpha(1)),\)\...,\((\alpha(N),\alpha(N-1)),\)\.... The phase transitions for \(f_{\mu}(\alpha)\) are at \(\alpha(1),\) \(\alpha(2),\)\...\(,\alpha(N),\)\.... The multifractal spectrum increases and is piecewise linear on \((\alpha_{\inf},\alpha(N)),\)\..., equals \(T(3)+3\alpha\) and increases on \((\alpha(3),\alpha(2)),\) equals \(T(2)+2\alpha\) and increases on \((\alpha(2),\alpha(1)),\) equals \(\alpha\) and increases on \((\alpha(1),\alpha(0)),\) and finally, decreases on \((\alpha(0),\alpha_{\sup}).\) Finally, we remark that the existence of \(\alpha_{\lim}\) is absolutely necessary for Theorems [\[thm:M1\]](#thm:M1){reference-type="ref" reference="thm:M1"} and [\[thm:M2\]](#thm:M2){reference-type="ref" reference="thm:M2"} to be true.

{'timestamp': '2016-07-19T02:05:17', 'yymm': '1607', 'arxiv_id': '1607.04786', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04786'}

# Conclusion {#sec:conclusion} We have described , a system for proving equivalence of SQL rewrite rules. In support of , we defined a formal language , following closely SQL's syntax. Our semantics extends that of SQL from finite relations to infinite relations, and uses univalent types from Homotopy Type Theory to represent and prove equality of cardinal numbers (finite and infinite). We have demonstrated the power and flexibility of by proving the correctness of several powerful optimization rules found in the database literature. # Denotation of {#sec:denotation} In this section we define the denotational semantics of . We first discuss the translation of constructs into . Then, in , we describe how advanced features of SQL (such as integrity constraints and indexes) can be expressed using and subsequently translated. Figure [\[fig:denote-query\]](#fig:denote-query){reference-type="ref" reference="fig:denote-query"} shows the translation of to . The translation rules make use of contexts. A *context schema* \(\Gamma\) is a schema (see the definition of \(\texttt{Schema}\) in Fig. [\[fig:data-model\]](#fig:data-model){reference-type="ref" reference="fig:data-model"}); a *context* \(g\) is a tuple of type \(\texttt{Tuple}\; \Gamma\) associated to that schema. Intuitively, the context consists of the concatenation of all tuple variables occurring in a surrounding scope. For example, consider the query with correlated subqueries in  where path expressions are used to refer to relations in predicates, as discussed in . As in standard SQL, evaluation proceeds from the outermost to the innermost query, starting with query \(q_1\). After the `FROM` clause in \(q_1\) is processed, the context consists of \(R_1\)'s schema (\(\sigma_{R_1}\)), which is then passed to the query \(q_2\). In turn, \(q_2\) then processes its `FROM` clause, and appends the schema of \(R_2\) to the context (\(\node{\sigma_{R_1}}{\sigma_{R_2}}\)), and this context is used to evaluate the path expression \(q_2\)'s predicate (`right.k = left.k`, i.e., `R_1.k = R_2.k`), and similarly when \(q_3\) evaluates its predicate. In our system, contexts are implemented as tuples. To make passing of contexts explicit, in the following, each construct takes in a context tuple (represented by `Tuple \Gamma`), and is translated to functions that take in both a tuple (\(t\)) and a context tuple (\(g\)). ## Denoting Basic HoTTSQL Constructs #### Queries A query \(q\) is denoted to a function from \(q\)'s context tuple (of type \(\texttt{Tuple} \; \Gamma\)) to a HoTT-Relation (of type \(\texttt{Tuple} \; \sigma \rightarrow \mathcal{U}\)): \[\denoteQuery{\Gamma}{q}{\sigma}: Tuple \; \Gamma \rightarrow Tuple \; \sigma \rightarrow \mathcal{U}\] The `FROM` clause is recursively denoted to a series of cross products of HoTT-Relations. Each cross product is denoted using \(\times\) as shown in Section [\[sec:target-lang\]](#sec:target-lang){reference-type="ref" reference="sec:target-lang"}. For example: \[\begin{array}{ll} \denoteQuery{\Gamma}{\FROM{q_1, \; q_2}}{\sigma} & \triangleq \\ \lambda \; \context \; \tuple. \; (\denoteQuery{\Gamma}{q_1}{\sigma} \; g \; \fst{t}) \times (\denoteQuery{\Gamma}{q_2}{\sigma} \; g \; \snd{t}) \end{array}\] where \(\fst{t}\) and \(\snd{t}\) indexes into the context tuple \(\Gamma\) to retrieve the schemas of \(q_1\) and \(q_2\) respectively. Note the manipulation of the context tuple in the denotation of `WHERE`: for each tuple \(t\), we first evaluate \(t\) against the query before `WHERE`, using the context tuple \(g\). After that, we evaluate the predicate \(b\) by first constructing a new context tuple as discussed (namely, by concatenating \(\Gamma\) and \(\sigma\), the schema of \(q\)), passing it the combined tuple \((g,t)\). The combination is needed as \(t\) has schema \(\sigma\) while the predicate \(b\) is evaluated under the schema \(\node{\Gamma}{\sigma}\), and the combination is easily accomplished as \(g\), the context tuple, has schema \(\Gamma\). `UNION ALL`, `EXCEPT`, and `DISTINCT` are denoted using +, negation (\(\negate{n}\)) and merely (\(\merely{n}\)) on univalent types as shown in Section [\[sec:target-lang\]](#sec:target-lang){reference-type="ref" reference="sec:target-lang"}. #### Predicates A predicate \(b\) is denoted to a function from a tuple (of type \(\texttt{Tuple} \; \Gamma\)) to a univalent type (of type \(\mathcal{U}\)): \[\denotePred{\Gamma}{b}: \texttt{Tuple} \; \Gamma \rightarrow \mathcal{U}\] More specifically, the return type \(\mathcal{U}\) must be a *squash type* . A squash type can only be a type of 1 element, namely **1**, and a type of 0 element, namely **0**. program with the form `q WHERE b` is denoted to the Cartesian product between a univalent type and a mere proposition. As an example, suppose a particular tuple \(t\) has multiplicity 3 in query \(q\), i.e., \(q\; t = \denote{R}{t} = {\bf 3}\), where \({\bf 3}\) is a univalent type. Since predicates are denoted to propositions, applying the tuple to the predicate returns either \({\bf 1}\) or \({\bf 0}\), and the overall result of the query for tuple \(t\) is then either \({\bf 3} \times {\bf 0} = {\bf 0}\), or \({\bf 3} \times {\bf 1} = {\bf 1}\), i.e., a squash type. #### Expressions and Projections A value expression \(e\) is denoted to a function from a tuple (of type \(\texttt{Tuple} \; \Gamma\)) to its data type, such as `int` and `bool`  (\(\denote{\tau}\)): \[\denoteExpr{\Gamma}{e}{\tau}: \texttt{Tuple} \; \Gamma \rightarrow \denote{\tau}\] A projection \(p\) from \(\Gamma\) to \(\Gamma'\) is denoted to a function from a tuple of type \(\texttt{Tuple} \; \Gamma\) to a tuple of type \(\texttt{Tuple} \; \Gamma'\). \[\denoteProj{p}{\Gamma}{\Gamma'}: \texttt{Tuple} \; \Gamma \rightarrow \texttt{Tuple} \; \Gamma'\] Projections are recursively defined. A projection can be composed by two projections using "`.`". The composition of two projection, "\(\Compose{p_1}{p_2}\)", where \(p_1\) is a projection from \(\Gamma\) to \(\Gamma'\) and \(p_2\) is a projection from \(\Gamma'\) to \(\Gamma''\), is denoted to a function from a tuple of type \(\texttt{Tuple} \; \Gamma\) to a tuple of type \(\texttt{Tuple} \; \Gamma''\) as follows: \[\lambda \; \context. \; \denoteProj{p_2}{\Gamma'}{\Gamma''}\;(\denoteProj{p_1}{\Gamma}{\Gamma'} \; g)\] We apply the denotation of \(p_1\), which is a function of type \(\texttt{Tuple}\;\Gamma \rightarrow \texttt{Tuple} \; \Gamma'\), to the argument of composed projection \(g\), then apply the denotation of \(p_2\) to the result of application. A projection can also be combined by two projections using "`,`". The combining of two projection, \(\Duplicate{p_1}{p_2}\), is denoted to: \[\lambda \; \context. \; (\denoteProj{p_1}{\Gamma}{\Gamma_0} \; g, \; \denoteProj{p_2}{\Gamma}{\Gamma_1} \; g)\]

{'timestamp': '2016-08-09T02:00:58', 'yymm': '1607', 'arxiv_id': '1607.04822', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04822'}

# Introduction #### **Hard lattice problems.** Lattices are discrete subgroups of \(\mathbb{R}^d\). More concretely, given a basis \(B = \{\vc{b}_1, \dots, \vc{b}_d\} \subset \mathbb{R}^d\), the lattice \(\mathcal{L} = \mathcal{L}(B)\) generated by \(B\) is defined as \(\mathcal{L}(B) = \left\{\sum_{i=1}^d \lambda_i \vc{b}_i: \lambda_i \in \mathbb{Z}\right\}\). Given a basis of a lattice \(\mathcal{L}\), the Shortest Vector Problem (SVP) asks to find a shortest non-zero vector in \(\mathcal{L}\) under the Euclidean norm, i.e., a non-zero lattice vector \(\vc{s}\) of norm \(\|\vc{s}\| = \lambda_1(\mathcal{L}) := \min_{\vc{v} \in \mathcal{L} \setminus \{\vc{0}\}} \|\vc{v}\|\). Given a basis of a lattice and a target vector \(\vc{t} \in \mathbb{R}^d\), the Closest Vector Problem (CVP) asks to find a vector \(\vc{s} \in \mathcal{L}\) closest to \(\vc{t}\) under the Euclidean distance, i.e. such that \(\|\vc{s}-\vc{t}\| = \min_{\vc{v} \in \mathcal{L}} \|\vc{v}-\vc{t}\|\). These two hard problems are fundamental in the study of lattice-based cryptography, as the security of these schemes is directly related to the hardness of SVP and CVP in high dimensions. Various other hard lattice problems, such as Learning With Errors (LWE) and the Shortest Integer Solution (SIS) problem are closely related to SVP and CVP, and many reductions between these and other hard lattice problems are known; see e.g. or for an overview. These reductions show that being able to solve CVP efficiently implies that almost all other lattice problems can also be solved efficiently in the same dimension, which makes the study of the hardness of CVP even more important for choosing parameters in lattice-based cryptography. #### **Algorithms for SVP and CVP.** Although SVP and CVP are both central in the study of lattice-based cryptography, algorithms for SVP have received somewhat more attention, including a benchmarking website to compare different algorithms . Various SVP methods have been studied which can solve CVP as well, such as enumeration (see e.g. ), discrete Gaussian sampling , constructing the Voronoi cell of the lattice , and using a tower of sublattices . On the other hand, for the asymptotically fastest method in high dimensions for SVP[^1], lattice sieving, it is not known how to solve CVP with similar costs as SVP. After a series of theoretical works on constructing efficient heuristic sieving algorithms  as well as practical papers studying how to speed up these algorithms even further , the best time complexity for solving SVP currently stands at \(2^{0.292d + o(d)}\) . Although for various other methods the complexities for solving SVP and CVP are similar , one can only guess whether the same holds for lattice sieving methods. To date, the best heuristic time complexity for solving CVP in high dimensions stands at \(2^{0.377d + o(d)}\), due to Becker--Gama--Joux . #### **Adaptive sieving.** In *adaptive sieving*, we adapt the entire sieving algorithm to the problem instance, including the target vector. As the resulting algorithm is tailored specifically to the given CVP instance, this leads to the best asymptotic complexity for solving a single CVP instance out of our two proposed methods: \(2^{0.292d + o(d)}\) time and space. This method is very similar to solving SVP with lattice sieving, and leads to equivalent asymptotics on the space and time complexities as for SVP. The corresponding space-time tradeoff is illustrated in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}, and equals that of for solving SVP. #### **Non-adaptive sieving.** Our main contribution, *non-adaptive sieving*, takes a different approach, focusing on cases where several CVP instances are to be solved on the same lattice. The goal here is to minimize the costs of computations depending on the target vector, and spend more time on preprocessing the lattice, so that the amortized time complexity per instance is smaller when solving many CVP instances on the same lattice. This is very closely related to the Closest Vector Problem with Preprocessing (CVPP), where the difference is that we allow for exponential-size preprocessed space. Using nearest neighbor techniques with a balanced space-time tradeoff, we show how to solve CVPP with \(2^{0.636d + o(d)}\) space and preprocessing, in \(2^{0.136d + o(d)}\) time. A continuous tradeoff between the two complexities can be obtained, where in the limit we can solve CVPP with \((1/\varepsilon)^{O(d)}\) space and preprocessing, in \(2^{\varepsilon d + o(d)}\) time. This tradeoff is depicted in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. A potential application of non-adaptive sieving is as a subroutine within enumeration methods. As described in e.g.  at any given level in the enumeration tree, one is attempting to solve a CVP instance in a lower-dimensional sublattice of \(\mathcal{L}\), where the target vector is determined by the path chosen from the root to the current node in the tree. That means that if we can preprocess this sublattice such that the amortized time complexity of solving CVPP is small, then this could speed up processing the bottom part of the enumeration tree. This in turn might help speed up the lattice basis reduction algorithm BKZ , which commonly uses enumeration as its SVP subroutine, and is key in assessing the security of lattice-based schemes. As the preprocessing needs to be performed once, CVPP algorithms with impractically large preprocessing costs may not be useful, but we show that with sieving the preprocessing costs can be quite small. #### **Outline.** The remainder of the paper is organized as follows. In Section [2](#sec:pre){reference-type="ref" reference="sec:pre"} we describe some preliminaries, such as sieving algorithms and a useful result on nearest neighbor searching. Section [3](#sec:ad){reference-type="ref" reference="sec:ad"} describes adaptive sieving and its analysis for solving CVP without preprocessing. Section [4](#sec:non){reference-type="ref" reference="sec:non"} describes the preprocessing approach to solving CVP, with complexity analyses for exact CVP and some of its relaxations. # Preliminaries {#sec:pre} ## Lattice sieving for solving SVP Heuristic lattice sieving algorithms for solving the shortest vector problem all use the following basic property of lattices: if \(\vc{v}, \vc{w} \in \mathcal{L}\), then their sum/difference \(\vc{v} \pm \vc{w} \in \mathcal{L}\) is a lattice vector as well. Therefore, if we have a long list \(L\) of lattice vectors stored in memory, we can consider combinations of these vectors to obtain new, shorter lattice vectors. To make sure the algorithm makes progress in finding shorter lattice vectors, \(L\) needs to contain a lot of lattice vectors; for vectors \(\vc{v}, \vc{w} \in \mathcal{L}\) of similar norm, the vector \(\vc{v}-\vc{w}\) is shorter than \(\vc{v}, \vc{w}\) iff the angle between \(\vc{v}, \vc{w}\) is smaller than \(\pi/3\), which for random vectors \(\vc{v}, \vc{w}\) occurs with probability \((3/4)^{d/2 + o(d)}\). The expected space complexity of heuristic sieving algorithms follows directly from this observation: if we draw \((4/3)^{d/2 + o(d)}\) random vectors from the unit sphere, we expect a large number of pairs of vectors to have angle less than \(\pi/3\), leading to many short difference vectors. This is exactly the heuristic assumption used in analyzing these sieving algorithms: when normalized, vectors in \(L\) follow the same distribution as vectors sampled uniformly at random from the unit sphere. Therefore, if we start by sampling a list \(L\) of \((4/3)^{d/2 + o(d)}\) long lattice vectors, and iteratively consider combinations of vectors in \(L\) to find shorter vectors, we expect to keep making progress. Note that naively, combining pairs of vectors in a list of size \((4/3)^{d/2 + o(d)} \approx 2^{0.208d + o(d)}\) takes time \((4/3)^{d + o(d)} \approx 2^{0.415d + o(d)}\). #### **The Nguyen-Vidick sieve.** The heuristic sieve algorithm of Nguyen and Vidick  starts by sampling a list \(L\) of \((4/3)^{d/2 + o(d)}\) long lattice vectors, and uses a *sieve* to map \(L\), with maximum norm \(R := \max_{\vc{v} \in L} \|\vc{v}\|\), to a new list \(L'\), with maximum norm at most \(\gamma R\) for \(\gamma < 1\) close to \(1\). By repeatedly applying this sieve, after \(\operatorname{poly}(d)\) iterations we expect to find a long list of lattice vectors of norm at most \(\gamma^{\operatorname{poly}(d)} R = O(\lambda_1(\mathcal{L}))\). The final list is then expected to contain a shortest vector of the lattice. Algorithm [\[alg:nv\]](#alg:nv){reference-type="ref" reference="alg:nv"} in Appendix [\[app:alg\]](#app:alg){reference-type="ref" reference="app:alg"} describes a sieve equivalent to Nguyen-Vidick's original sieve, to map \(L\) to \(L'\) in \(|L|^2\) time. #### **Micciancio and Voulgaris' GaussSieve.** Micciancio and Voulgaris used a slightly different approach in the GaussSieve . This algorithm reduces the memory usage by immediately *reducing* all pairs of lattice vectors that are sampled. The algorithm uses a single list \(L\), which is always kept in a state where for all \(\vc{w}_1, \vc{w}_2 \in L\), \(\|\vc{w}_1 \pm \vc{w}_2\| \geq \|\vc{w}_1\|, \|\vc{w}_2\|\), and each time a new vector \(\vc{v} \in \mathcal{L}\) is sampled, its norm is reduced with vectors in \(L\). After the norm can no longer be reduced, the vectors in \(L\) are reduced with \(\vc{v}\). Modified list vectors are added to a stack to be processed later (to maintain the pairwise reduction-property of \(L\)), and new vectors which are pairwise reduced with \(L\) are added to \(L\). Immediately reducing all pairs of vectors means that the algorithm uses less time and memory in practice, but at the same time Nguyen and Vidick's heuristic proof technique does not apply here. However, it is commonly believed that the same bounds \((4/3)^{d/2 + o(d)}\) and \((4/3)^{d + o(d)}\) on the space and time complexities hold for the GaussSieve. Pseudocode of the GaussSieve is given in Algorithm [\[alg:gauss\]](#alg:gauss){reference-type="ref" reference="alg:gauss"} in Appendix [\[app:alg\]](#app:alg){reference-type="ref" reference="app:alg"}. ## Nearest neighbor searching Given a data set \(L \subset \mathbb{R}^d\), the nearest neighbor problem asks to preprocess \(L\) such that, when given a query \(\vc{t} \in \mathbb{R}^d\), one can quickly return a nearest neighbor \(\vc{s} \in L\) with distance \(\|\vc{s}-\vc{t}\| = \min_{\vc{w} \in L} \|\vc{w}-\vc{t}\|\). This problem is essentially identical to CVP, except that \(L\) is a finite set of unstructured points, rather than the infinite set of all points in a lattice \(\mathcal{L}\). #### **Locality-Sensitive Hashing/Filtering (LSH/LSF).** A celebrated technique for finding nearest neighbors in high dimensions is Locality-Sensitive Hashing (LSH) , where the idea is to construct many random partitions of the space, and store the list \(L\) in hash tables with buckets corresponding to regions. Preprocessing then consists of constructing these hash tables, while a query \(\vc{t}\) is answered by doing a lookup in each of the hash tables, and searching for a nearest neighbor in these buckets. More details on LSH in combination with sieving can be found in e.g.Ṡimilar to LSH, Locality-Sensitive Filtering (LSF)  divides the space into regions, with the added relaxation that these regions do not have to form a partition; regions may overlap, and part of the space may not be covered by any region. This leads to improved results compared to LSH when \(L\) has size exponential in \(d\) . Below we restate one of the main results of  for our applications. The specific problem considered here is: given a data set \(L \subset \mathcal{S}^{d-1}\) sampled uniformly at random, and a random query \(\vc{t} \in \mathcal{S}^{d-1}\), return a vector \(\vc{w} \in L\) such that the angle between \(\vc{w}\) and \(\vc{t}\) is at most \(\theta\). The following result further assumes that the list \(L\) contains \(n = (1 / \sin \theta)^{d + o(d)}\) vectors. Applying this result to sieving for solving SVP, where \(n = \sin(\frac{\pi}{3})^{-d + o(d)}\) and we are looking for pairs of vectors at angle at most \(\frac{\pi}{3}\) to perform reductions, this leads to a space and preprocessing complexity of \(n^{0.292d + o(d)}\), and a query complexity of \(2^{0.084d + o(d)}\). As the preprocessing in sieving is only performed once, and queries are performed \(n \approx 2^{0.208d + o(d)}\) times, this leads to a reduction of the complexities of sieving (for SVP) from \(2^{0.208d + o(d)}\) space and \(2^{0.415d + o(d)}\) time, to \(2^{0.292d + o(d)}\) space and time . # Adaptive sieving for CVP {#sec:ad} We present two methods for solving CVP using sieving, the first of which we call *adaptive sieving*--we adapt the entire sieving algorithm to the particular CVP instance, to obtain the best overall time complexity for solving one instance. When solving several CVP instances, the costs roughly scale linearly with the number of instances. ### Using one list. The main idea behind this method is to translate the SVP algorithm by the target vector \(\vc{t}\); instead of generating a long list of lattice vectors reasonably close to \(\vc{0}\), we generate a list of lattice vectors close to \(\vc{t}\), and combine lattice vectors to find lattice vectors even closer vectors to \(\vc{t}\). The final list then hopefully contains a closest vector to \(\vc{t}\). One quickly sees that this does not work, as the fundamental property of lattices does not hold for the lattice coset \(\vc{t} + \mathcal{L}\): if \(\vc{w}_1, \vc{w}_2 \in \vc{t} + \mathcal{L}\), then \(\vc{w}_1 \pm \vc{w}_2 \notin \vc{t} + \mathcal{L}\). In other words, two lattice vectors close to \(\vc{t}\) can only be combined to form lattice vectors close to \(\vc{0}\) or \(2 \vc{t}\). So if we start with a list of vectors close to \(\vc{t}\), and combine vectors in this list as in the Nguyen-Vidick sieve, then after one iteration we will end up with a list \(L'\) of lattice vectors close to \(\vc{0}\). ### Using two lists. To make the idea of translating the whole problem by \(\vc{t}\) work for the Nguyen-Vidick sieve, we make the following modification: we keep track of two lists \(L = L_{\vc{0}}\) and \(L_{\vc{t}}\) of lattice vectors close to \(\vc{0}\) and \(\vc{t}\), and construct a sieve which maps two input lists \(L_{\vc{0}}, L_{\vc{t}}\) to two output lists \(L_{\vc{0}}', L_{\vc{t}}'\) of lattice vectors slightly closer to \(\vc{0}\) and \(\vc{t}\). Similar to the original Nguyen-Vidick sieve, we then apply this sieve several times to two initial lists \((L_{\vc{0}}, L_{\vc{t}})\) with a large radius \(R\), to end up with two lists \(L_{\vc{0}}\) and \(L_{\vc{t}}\) of lattice vectors at distance at most approximately \(\sqrt{4/3} \cdot \lambda_1(\mathcal{L})\) from \(\vc{0}\) and \(\vc{t}\)[^2]. The argumentation that this algorithm works is almost identical to that for solving SVP, where we now make the following slightly different heuristic assumption. The resulting algorithm, based on the Nguyen-Vidick sieve, is presented in Algorithm [\[alg:nv-adaptive\]](#alg:nv-adaptive){reference-type="ref" reference="alg:nv-adaptive"}. An important open question is whether these techniques can also be applied to the faster GaussSieve algorithm to solve CVP. The GaussSieve seems to make even more use of the property that the sum/difference of two lattice vectors is also in the lattice, and operations in the GaussSieve in \(\mathcal{L}\) cannot as easily be *mimicked* for the coset \(\vc{t} + \mathcal{L}\). Solving CVP with the GaussSieve with similar complexities is left as an open problem. # Non-adaptive sieving for CVPP {#sec:non} Our second method for finding closest vectors with heuristic lattice sieving follows a slightly different approach. Instead of focusing only on the total time complexity for one problem instance, we split the algorithm into two phases: - Phase 1: Preprocess the lattice \(\mathcal{L}\), without knowledge of the target \(\vc{t}\); - Phase 2: Process the query \(\vc{t}\) and output a closest lattice vector \(\vc{s} \in \mathcal{L}\) to \(\vc{t}\). Intuitively it may be more important to keep the costs of Phase 2 small, as the preprocessed data can potentially be reused later for other instances on the same lattice. This approach is essentially equivalent to the Closest Vector Problem with Preprocessing (CVPP): preprocess \(\mathcal{L}\) such that when given a target vector \(\vc{t}\) later, one can quickly return a closest vector \(\vc{s} \in \mathcal{L}\) to \(\vc{t}\). For CVPP however the preprocessed space is usually restricted to be of polynomial size, and the time used for preprocessing the lattice is often not taken into account. Here we will keep track of the preprocessing costs as well, and we do not restrict the output from the preprocessing phase to be of size \(\operatorname{poly}(d)\). ### Algorithm description. To minimize the costs of answering a query, and to do the preprocessing independently of the target vector, we first run a standard SVP sieve, resulting in a large list \(L\) of almost all short lattice vectors. Then, after we are given the target vector \(\vc{t}\), we use \(L\) to reduce the target. Finally, once the resulting vector \(\vc{t}' \in \vc{t} + \mathcal{L}\) can no longer be reduced with our list, we hope that this reduced vector \(\vc{t}'\) is the shortest vector in the coset \(\vc{t} + \mathcal{L}\), so that \(\vc{0}\) is the closest lattice vector to \(\vc{t}'\) and \(\vc{s} = \vc{t}-\vc{t}'\) is the closest lattice vector to \(\vc{t}\). The first phase of this algorithm consists in running a sieve and storing the resulting list in memory (potentially in a nearest neighbor data structure for faster lookups). For this phase either the Nguyen-Vidick sieve or the GaussSieve can be used. The second phase is the same for either method, and is described in Algorithm [\[alg:nonadaptive\]](#alg:nonadaptive){reference-type="ref" reference="alg:nonadaptive"} for the general case of an input list essentially consisting of the \(\alpha^{d + o(d)}\) shortest vectors in the lattice. Note that a standard SVP sieve would produce a list of size \((4/3)^{d/2 + o(d)}\) corresponding to \(\alpha = \sqrt{4/3}\). Under Heuristic [\[heur:1\]](#heur:1){reference-type="ref" reference="heur:1"}, we then obtain a relation between the choice of \(\alpha\) for the input list and the expected norm of the reduced vector \(\vc{t}'\) as follows. Note that in our applications, we do not just have a list of \(\alpha^{d + o(d)}\) lattice vectors of norm \(\alpha \cdot \lambda_1(\mathcal{L})\); for any \(\alpha_0 \in [1, \alpha]\) we expect \(L\) to contain \(\alpha_0^{d + o(d)}\) lattice vectors of norm at most \(\alpha_0 \cdot \lambda_1(\mathcal{L})\). To obtain a reduced vector \(\vc{t}'\) of norm \(\beta \cdot \lambda_1(\mathcal{L})\), we therefore obtain the condition that for *some* value \(\alpha_0 \in [1, \alpha]\), it must hold that \(\alpha_0^4-4 \beta^2 \alpha_0^2 + 4\beta_0^2 < 0\). From [\[eq:a\]](#eq:a){reference-type="eqref" reference="eq:a"} it follows that \(p(\alpha^2) = \alpha^4-4 \beta^2 \alpha^2 + 4\beta^2\) has two roots \(r_1 < 2 < r_2\) for \(\alpha^2\), which lie close to \(2\) for \(\beta \approx 1\). The condition that \(p(\alpha_0^2) < 0\) for some \(\alpha_0 \leq \alpha\) is equivalent to \(\alpha > r_1\), which for \(\beta = 1 + o(1)\) implies that \(\alpha^2 \geq 2 + o(1)\). This means that asymptotically we must set \(\alpha = \sqrt{2}\), and use \(n = 2^{d/2 + o(d)}\) input vectors, to guarantee that w.h.p. the algorithm succeeds. A sketch of the situation is also given in Figure [\[fig:2a\]](#fig:2a){reference-type="ref" reference="fig:2a"}. ### Modifying the first phase. As we will need a larger list of size \(2^{d/2 + o(d)}\) to make sure we can solve CVP exactly, we need to adjust Phase 1 of the algorithm as well. Recall that with standard sieving, we reduce vectors iff their angle is at most \(\theta = \frac{\pi}{3}\), resulting in a list of size \((\sin \theta)^{-d + o(d)}\). As we now need the output list of the first phase to consist of \(2^{d/2 + o(d)} = (\sin \theta')^{-d + o(d)}\) vectors for \(\theta' = \frac{\pi}{4}\), we make the following adjustment: only reduce \(\vc{v}\) and \(\vc{w}\) if their common angle is less than \(\frac{\pi}{4}\). For unit length vectors, this condition is equivalent to reducing \(\vc{v}\) with \(\vc{w}\) iff \(\|\vc{v}-\vc{w}\|^2 \leq (2-\sqrt{2}) \cdot \|\vc{v}\|^2\). This further accelerates nearest neighbor techniques due to the smaller angle \(\theta\). Pseudocode for the modified first phase is given in Appendix [\[app:alg2\]](#app:alg2){reference-type="ref" reference="app:alg2"} ### Main result. With the algorithm in place, let us now analyze its complexity for solving CVP. The first phase of the algorithm generates a list of size \(2^{d/2 + o(d)}\) by combining pairs of vectors, and naively this can be done in time \(\mathrm{T}_1 = 2^{d + o(d)}\) and space \(\mathrm{S} = 2^{d/2 + o(d)}\), with query complexity \(\mathrm{T}_2 = 2^{d/2 + o(d)}\). Using nearest neighbor searching (Lemma [\[lem:nns\]](#lem:nns){reference-type="ref" reference="lem:nns"}), the query and preprocessing complexities can be further reduced, leading to the following result. To illustrate the time and space complexities of Theorem [\[thm:2\]](#thm:2){reference-type="ref" reference="thm:2"}, we highlight three special cases \(u\) as follows. The full tradeoff curve for \(u \in (\frac{1}{2} \sqrt{2}, \sqrt{2})\) is depicted in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}. - Setting \(u = \frac{1}{2} \sqrt{2}\), we obtain \(\mathrm{S} = \mathrm{T}_1 = 2^{d/2 + o(d)}\) and \(\mathrm{T}_2 \approx 2^{0.2925d + o(d)}\). - Setting \(u = 1\), we obtain \(\mathrm{S} = \mathrm{T}_1 \approx 2^{0.6358 d + o(d)}\) and \(\mathrm{T}_2 \approx 2^{0.1358 d + o(d)}\). - Setting \(u = \frac{1}{2}(\sqrt{2} + 1)\), we get \(\mathrm{S} = \mathrm{T}_1 = 2^{d + o(d)}\) and \(\mathrm{T}_2 \approx 2^{0.0594 d + o(d)}\). The first result shows that the query complexity of non-adaptive sieving is never worse than for adaptive sieving; only the space and preprocessing complexities are worse. The second and third results show that CVP can be solved in significantly less time, even with preprocessing and space complexities bounded by \(2^{d + o(d)}\). #### **Minimizing the query complexity.** As \(u \to \sqrt{2}\), the query complexity keeps decreasing while the memory and preprocessing costs increase. For arbitrary \(\varepsilon > 0\), we can set \(u = u_\varepsilon \approx \sqrt{2}\) as a function of \(\varepsilon\), resulting in asymptotic complexities \(\mathrm{S} = \mathrm{T}_1 = (1/\varepsilon)^{O(d)}\) and \(\mathrm{T}_2 = 2^{\varepsilon d + o(d)}\). This shows that it is possible to obtain a slightly subexponential query complexity, at the cost of superexponential space, by taking \(\varepsilon = o(1)\) as a function of \(d\). Being able to solve CVPP in subexponential time with superexponential preprocessing and memory is neither trivial nor quite surprising. A naive approach to the problem, with this much memory, could for instance be to index the entire fundamental domain of \(\mathcal{L}\) in a hash table. One could partition this domain into small regions, solve CVP for the centers of each of these regions, and store all the solutions in memory. Then, given a query, one looks up which region \(\vc{t}\) is in, and returns the answer corresponding to that vector. With a sufficiently fine-grained partitioning of the fundamental domain, the answers given by the look-ups are accurate, and this algorithm probably also runs in subexponential time. Although it may not be surprising that it is possible to solve CVPP in subexponential time with (super)exponential space, it is not clear what the complexities of other methods would be. Our method presents a clear tradeoff between the complexities, where the constants in the preprocessing exponent are quite small; for instance, we can solve CVPP in time \(2^{0.06d + o(d)}\) with less than \(2^{d + o(d)}\) memory, which is the same amount of memory/preprocessing of the best provable SVP and CVP algorithms . Indexing the fundamental domain may well require much more memory than this. ## Bounded Distance Decoding with Preprocessing We finally take a look at specific instances of CVP which are easier than the general problem, such as when the target \(\vc{t}\) lies unusually close to the lattice. This problem naturally appears in practice, when a private key consists of a *good basis* of a lattice with short basis vectors, and the public key is a *bad basis* of the same lattice. An encryption of a message could then consist of the message being mapped to a lattice point \(\vc{v} \in \mathcal{L}\), and a small error vector \(\vc{e}\) being added to \(\vc{v}\) (\(\vc{t} = \vc{v} + \vc{e}\)) to hide \(\vc{v}\). If the noise \(\vc{e}\) is small enough, then with a good basis one can decode \(\vc{t}\) to the closest lattice vector \(\vc{v}\), while someone with the bad basis cannot decode correctly. As decoding for arbitrary \(\vc{t}\) (solving CVP) is known to be hard even with knowledge of a good basis , \(\vc{e}\) needs to be very short, and \(\vc{t}\) must lie unusually close to the lattice. So instead of assuming target vectors \(\vc{t} \in \mathbb{R}^d\) are sampled at random, suppose that \(\vc{t}\) lies at distance at most \(\delta \cdot \lambda_1(\mathcal{L})\) from \(\mathcal{L}\), for \(\delta \in (0,1)\). For adaptive sieving, recall that the list size \((4/3)^{d/2 + o(d)}\) is the minimum initial list size one can hope to use to obtain a list of short lattice vectors; with fewer vectors, one would not be able to solve SVP.[^3] For non-adaptive sieving however, it may be possible to reduce the list size below \(2^{d/2 + o(d)}\). ### List size. Let us again assume that the preprocessed list \(L\) contains almost all \(\alpha^{d + o(d)}\) lattice vectors of norm at most \(\alpha \cdot \lambda_1(\mathcal{L})\). The choice of \(\alpha\) implies a maximum norm \(\beta_{\alpha} \cdot \lambda_1(\mathcal{L})\) of the reduced vector \(\vc{t}'\), as described in Lemma [\[lem:2\]](#lem:2){reference-type="ref" reference="lem:2"}. The nearest lattice vector \(\vc{s} \in \mathcal{L}\) to \(\vc{t}'\) lies within radius \(\delta \cdot \lambda_1(\mathcal{L})\) of \(\vc{t}'\), and w.h.p. \(\vc{s}-\vc{t}'\) is approximately orthogonal to \(\vc{t}'\); see Figure [\[fig:2b\]](#fig:2b){reference-type="ref" reference="fig:2b"}, where the shaded area is asymptotically negligible. Therefore w.h.p. \(\vc{s}\) has norm at most \((\sqrt{\beta_{\alpha}^2 + \delta^2}) \cdot \lambda_1(\mathcal{L})\). Now if \(\sqrt{\beta_{\alpha}^2 + \delta^2} \leq \alpha\), then we expect the nearest vector to be contained in \(L\), so that ultimately \(\vc{0}\) is nearest to \(\vc{t}'\). Substituting \(\alpha^4-4 \beta^2 \alpha^2 + 4 \beta^2 = 0\) and \(\beta^2 + \delta^2 \leq \alpha^2\), and solving for \(\alpha\), this leads to the following condition on \(\alpha\). \[\begin{aligned} \alpha^2 \geq \tfrac{2}{3} (1 + \delta^2) + \tfrac{2}{3} \sqrt{(1 + \delta^2)^2-3 \delta^2} \,. \label{eq:a2} \end{aligned}\] Taking \(\delta = 1\), corresponding to exact CVP, leads to the condition \(\alpha \geq \sqrt{2}\) as expected, while in the limiting case of \(\delta \to 0\) we obtain the condition \(\alpha \geq \sqrt{4/3}\). This matches experimental observations using the GaussSieve, where after finding the shortest vector, newly sampled vectors often cause *collisions* (i.e. being reduced to the \(\vc{0}\)-vector). In other words, Algorithm [\[alg:nonadaptive\]](#alg:nonadaptive){reference-type="ref" reference="alg:nonadaptive"} often reduces target vectors \(\vc{t}\) which essentially lie *on* the lattice (\(\delta \to 0\)) to the \(\vc{0}\)-vector when the list has size \((4/3)^{d/2 + o(d)}\). This explains why collisions in the GaussSieve are common when the list size grows to size \((4/3)^{d/2 + o(d)}\). ### Main result. To solve BDD with a target \(\vc{t}\) at distance \(\delta \cdot \lambda_1(\mathcal{L})\) from the lattice, we need the preprocessing to produce a list of almost all \(\alpha^{d + o(d)}\) vectors of norm at most \(\alpha \cdot \lambda_1(\mathcal{L})\), with \(\alpha\) satisfying [\[eq:a2\]](#eq:a2){reference-type="eqref" reference="eq:a2"}. Similar to the analysis for CVP, we can produce such a list by only doing reductions between two vectors if their angle is less than \(\theta\), where now \(\theta = \arcsin(1 / \alpha)\). Combining this with Lemma [\[lem:1\]](#lem:1){reference-type="ref" reference="lem:1"}, we obtain the following result. Theorem [\[thm:BDD\]](#thm:BDD){reference-type="ref" reference="thm:BDD"} is a generalization of Theorem [\[thm:2\]](#thm:2){reference-type="ref" reference="thm:2"}, as the latter can be derived from the former by substituting \(\delta = 1\) above. To illustrate the results, Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"} considers two special cases: - For \(\delta = \frac{1}{2}\), we find \(\alpha \approx 1.1976\), leading to \(\mathrm{S} \approx 2^{0.2602d + o(d)}\) and \(\mathrm{T}_2 = 2^{0.1908d + o(d)}\) when minimizing the space complexity. - For \(\delta \to 0\), we have \(\alpha \to \sqrt{4/3} \approx 1.1547\). The minimum space complexity is therefore \(\mathrm{S} = (4/3)^{d/2 + o(d)}\), with query complexity \(\mathrm{T}_2 = 2^{0.1610d + o(d)}\). In the limit of \(u \to \sqrt{\frac{\alpha^2}{\alpha^2-1}}\) we need superexponential space/preprocessing \(\mathrm{S}, \mathrm{T}_1 \to 2^{\omega(d)}\) and a subexponential query time \(\mathrm{T}_2 \to 2^{o(d)}\) for all \(\delta > 0\). ## Approximate Closest Vector Problem with Preprocessing Given a lattice \(\mathcal{L}\) and a target vector \(\vc{t} \in \mathbb{R}^d\), approximate CVP with approximation factor \(\kappa\) asks to find a vector \(\vc{s} \in \mathcal{L}\) such that \(\|\vc{s}-\vc{t}\|\) is at most a factor \(\kappa\) larger than the real distance from \(\vc{t}\) to \(\mathcal{L}\). For random instances \(\vc{t}\), by the Gaussian heuristic this means that a lattice vector counts as a solution iff it lies at distance at most \(\kappa \cdot \lambda_1(\mathcal{L})\) from \(\vc{t}\). ### List size. Instead of reducing \(\vc{t}\) to a vector \(\vc{t}'\) of norm at most \(\lambda_1(\mathcal{L})\) as is needed for solving exact CVP, we now want to make sure that the reduced vector \(\vc{t}'\) has norm at most \(\kappa \cdot \lambda_1(\mathcal{L})\). If this is the case, then the vector \(\vc{t}-\vc{t}'\) is a lattice vector lying at distance at most \(\kappa \cdot \lambda_1(\mathcal{L})\), which w.h.p. qualifies as a solution. This means that instead of substituting \(\beta = 1\) in Lemma [\[lem:2\]](#lem:2){reference-type="ref" reference="lem:2"}, we now substitute \(\beta = \kappa\). This leads to the condition that \(\alpha_0^4-4\kappa^2 \alpha_0^2 + 4 \beta^2 < 0\) for some \(\alpha_0 \leq \alpha\). By a similar analysis \(\alpha^2\) must therefore be larger than the smallest root \(r_1 = 2\kappa (\kappa-\sqrt{\kappa^2-1})\) of this quadratic polynomial in \(\alpha^2\). This immediately leads to the following condition on \(\alpha\): \[\begin{aligned} \alpha^2 \geq 2 \kappa \left(\kappa-\sqrt{\kappa^2-1}\right). \label{eq:a3} \end{aligned}\] A sanity check shows that \(\kappa = 1\), corresponding to exact CVP, indeed results in \(\alpha \geq \sqrt{2}\), while in the limit of \(\kappa \to \infty\) a value \(\alpha \approx 1\) suffices to obtain a vector \(\vc{t}'\) of norm at most \(\kappa \cdot \lambda_1(\mathcal{L})\). In other words, to solve approximate CVP with very large (constant) approximation factors, a preprocessed list of size \((1 + \varepsilon)^{d + o(d)}\) suffices. ### Main result. Similar to the analysis of CVPP, we now take \(\theta = \arcsin(1/\alpha)\) as the angle with which to reduce vectors in Phase 1, so that the output of Phase 1 is a list of almost all \(\alpha^{d + o(d)}\) shortest lattice vectors of norm at most \(\alpha \cdot \lambda_1(\mathcal{L})\). Using a smaller angle \(\theta\) for reductions again means that nearest neighbor searching can speed up the reductions in both Phase 1 and Phase 2 even further. The exact complexities follow from Lemma [\[lem:nns\]](#lem:nns){reference-type="ref" reference="lem:nns"}. Note that only the dependence of \(\alpha\) on \(\kappa\) is different, compared to the dependence of \(\alpha\) on \(\delta\) for bounded distance decoding. The complexities for \(\kappa\)-CVP arguably decrease *faster* than for \(\delta\)-BDD: for instance, for \(\kappa \approx 1.0882\) we obtain the same complexities as for BDD with \(\delta = \frac{1}{2}\), while \(\kappa = \sqrt{4/3} \approx 1.1547\) leads to the same complexities as for BDD with \(\delta \to 0\). Two further examples are illustrated in Figure [\[fig:1\]](#fig:1){reference-type="ref" reference="fig:1"}: - For \(\kappa = 2\), we have \(\alpha \approx 1.1976\), which for \(u \approx 0.5503\) leads to \(\mathrm{S} = \mathrm{T}_1 = 2^{0.2602 d + o(d)}\) and \(\mathrm{T}_2 = 2^{0.1908 d + o(d)}\), and for \(u = 1\) leads to \(\mathrm{S} = \mathrm{T}_1 = 2^{0.3573 d + o(d)}\) and \(\mathrm{T}_2 = 2^{0.0971 d + o(d)}\). - For \(\kappa \to \infty\), we have \(\alpha \to 1\), i.e. the required preprocessed list size approaches \(2^{o(d)}\) as \(\kappa\) grows. For sufficiently large \(\kappa\), we can solve \(\kappa\)-CVP with a preprocessed list of size \(2^{\varepsilon d + o(d)}\) in at most \(2^{\varepsilon d + o(d)}\) time. The preprocessing time is given by \(2^{0.2925 d + o(d)}\). The latter result shows that for any superconstant approximation factor \(\kappa = \omega(1)\), we can solve the corresponding approximate closest vector problem with preprocessing in subexponential time, with an exponential preprocessing time complexity \(2^{0.292d + o(d)}\) for solving SVP and generating a list of short lattice vectors, and a subexponential space complexity required for Phase 2. In other words, even without superexponential preprocessing/memory we can solve CVPP with large approximation factors in subexponential time. To compare this result with previous work, note that the lower bound on \(\alpha\) from [\[eq:a3\]](#eq:a3){reference-type="eqref" reference="eq:a3"} tends to \(1 + 1/(8 \kappa^2) + O(\kappa^{-4})\) as \(\kappa\) grows. The query space and time complexities are further both proportional to \(\alpha^{\Theta(d)}\). To obtain a polynomial query complexity and polynomial storage after the preprocessing phase, we can solve for \(\kappa\), leading to the following result. Apart from the heuristic assumptions, this approximation factor is equivalent to Aharonov and Regev , who showed that the decision version of CVPP with approximation factor \(\kappa = \Omega(\sqrt{d / \log d})\) can provably be solved in polynomial time. This further (heuristically) improves upon results of , who are able to solve search-CVPP with polynomial time and space complexities for \(\kappa = O(d^{3/2})\) and \(\kappa = \Omega(d / \sqrt{\log d})\) respectively. Assuming the heuristic assumptions are valid, Corollary [\[cor:acvpp-poly\]](#cor:acvpp-poly){reference-type="ref" reference="cor:acvpp-poly"} closes the gap between these previous results for decision-CVPP and search-CVPP with a rather simple algorithm: (1) preprocess the lattice by storing all \(d^{O(1)}\) shortest vectors of the lattice in a list; and (2) apply Algorithm [\[alg:nonadaptive\]](#alg:nonadaptive){reference-type="ref" reference="alg:nonadaptive"} to this list and the target vector to find an approximate closest vector. Note that nearest neighbor techniques only affect leading constants; even without nearest neighbor searching this would heuristically result in a polynomial time and space algorithm for \(\kappa\)-CVPP with \(\kappa = \Omega(\sqrt{d / \log d})\). An interesting open problem would be to see if this result can be made provable for arbitrary lattices, without any heuristic assumptions.

{'timestamp': '2016-07-19T02:05:21', 'yymm': '1607', 'arxiv_id': '1607.04789', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04789'}

# Introduction Exciton mediated many-body interactions give rise to a host of physical effects that determine the opto-electronic properties of low dimensional transition metal dichalcogenides, MX\(_2\) (M = Mo, W, Nb and X = S, Se) [@scharf2016probing; @rama; @ugeda2014giant; @wang2012electronics; @makatom; @qiu2013; @splen; @komsa2012effects; @kormanyos2013monolayer], with important consequences for fundamental and applied research. The confinement of correlated charge carriers or excitons to a narrow region of space in low dimensional transition metal dichalcogenides (TMDCs) leads to unique photoluminescence properties that are otherwise absent in the bulk configurations. The availability of state-of-the art exfoliation techniques enable fabrication of low dimensional transition metal dichalcogenides that is useful for applications [@ou2014ion; @li2016charge; @perebeinos2015metal; @beck2000; @tsai2013few; @bernardi2013extraordinary; @he2012fabrication; @wi2014enhancement; @bertolazzi2013nonvolatile; @ji2013epitaxial; @yu2016evaluation; @park2016mos2; @eda2013two; @radisavljevic2011integrated; @lembke2015single; @pospischil2014solar; @zhang2014m; @yoon2011good]. The excitonic processes that determine the performance of TMDC-based electronic devices include defect assisted scattering and trapping by surface states, decay via exciton-exciton annihilation , phonon assisted relaxation, and capture by mid-gap defects through Auger processes . Excitonic processes that result in the formation of complex trions [@mak; @berkelbach2013theory; @thiltrion] and electron-hole recombination with generation of hot carriers are also of importance in device performances. Dynamical processes incorporating exciton-phonon interactions underlie the opto-electronic properties of monolayer transition metal dichalcogenides . The strength of the interactions between coupled charge carriers and phonons is deduced from experimental measurements of the dephasing times , exciton linewidths, photoluminescence, and other parameters such as the exciton mobility and luminescence rise times. The exciton formation time is determined by a complicated interplay of various dynamical processes in the picosecond time scale and is linked to the efficient operation of optoelectronic devices. To this end, a comprehensive understanding of how newly generated electron-hole pairs relax energetically to form excitons still remain unclear. Recently decay times of \(\approx\) 0.3 ps of the transient absorption signal subsequent to the interband excitation of the monolayer WSe\(_2\), MoS\(_2\), and MoSe\(_2\) was recorded in time-resolved measurements. The ultrafast decay times were deduced as the exciton formation times from electron-hole pairs in monolayer systems. Motivated by these considerations, we examine a mechanism by which excitons are formed from an initial state of unbound electron-hole pairs to account for the observed short exciton formation time in common TMDCs (MoS\(_2\), MoSe\(_2\), WS\(_2\), and WSe\(_2\)). While the focus of this paper is on the theoretical aspects of excitonic interactions, the overall aim is to seek an understanding of the critical factors that limit the exciton formation time which is of relevance to experimental investigations involving device applications. The unequal charges of the basis atoms in polar crystals allow a moving electron to polarize the electric field of its surrounding medium. The polarization effects displaces the ions giving rise to lattice vibrations of a optical phonon frequency in resonance with the polarization field, and enable direct Fröhlich coupling between phonons and charge carriers. In this work we consider that the excitons are created via the two-dimensional Fröhlich interaction which provides a critical pathway by which charge carriers undergo energy loss to optical phonons at elevated temperatures in the monolayers MoS\(_2\) and other transition-metal dichalcogenides. The exciton is a neutral quasiparticle, and polarization effects due to the longitudinal optical waves may appear to have less influence than those associated with polarization effects of the individual electron or hole. In reality the internal state of the exciton undergoes dipole type transitions and there occurs measurable effects due to Fröhlich interactions in constrained systems. The focus on LO phonons in the exciton formation process in this study is justified by the large strength of excitonic interactions with high frequency phonons that arise due to the strong confinement of the exciton wave-functions in the real space of monolayer systems. Moreover the exciton-phonon interaction is long ranged due to the existence of polarization effects despite large separations between charge carriers and the ions in the material system. The phonon-limited mobility is largely dominated by polar optical scattering via the Fröhlich interaction at room temperatures. Exciton formation may take place via deformation potential coupling to acoustic phonons , but is likely to occur with less efficiency due to the high exciton binding energies in monolayer dichalcogenides. In conventional semiconductors such as the two band GaAs material system, excitons are formed via the Fröhlich interaction in the picosecond time range . While excitons in GaAs materials are dominantly formed at the center of the Brillouin zone center, the formation process occurs at the non-central points in the momentum space of monolayer TMDCs. This gives rise to quantitative differences in the exciton creation times between GaAs and TMDCs. For excitation energies higher than the band-gap of monolayer systems, the electron-hole pair creates an exciton with a non-zero wavevector associated with its center-of-mass motion. The exciton subsequently relaxes to the zero wavevector state with emission of acoustic or LO phonons before undergoing radiative recombination by emitting a photon. To this end, the formation time of an exciton as a function of exciton wave vector is useful in analyzing the luminescence rise times that can be measured experimentally. In this study we employ the exciton-LO phonon interaction operator to estimate the exciton formation times in monolayer transition metal dichalcogenides. The formation time of excitons is determined using the interaction Hamiltonian which describes the conversion of the photoexcited free electron-hole pair to a final exciton state initiated by exciton-phonon Fröhlich interactions, and accompanied by absorption or emission of phonons. The dependence of the exciton formation time on several parameters such as the temperatures of the crystal lattice, charge carriers and excitons as well as the densities of charge carriers and excitons will be closely examined in this study. # Formation of Excitons in monolayer Molybdenum Disulfide [\[basic\]]{#basic label="basic"} ## Exciton-LO phonon Hamiltonian We project the single monolayer of a hexagonally ordered plane of metal atoms sandwiched between two other hexagon planes of chalcogens onto a quasi two-dimensional space. The motion of the exciton is generally confined to the parallel two-dimensional \(XY\) layers of the atomic planes with restricted electron and hole motion in the \(z\) direction perpendicular to the monolayer plane. The monolayer MoS\(_2\) has nine phonon branches consisting of three acoustic and six optical branches. The two lowest optical branches are weakly coupled to the charge carriers are therefore not expected to play a significant role in the creation of excitons. The next two phonon branches at the \(\Sigma\) point positioned at energies 48 meV are linked to polar optical modes, which play a critical role in the formation of exciton after photoexcitation of the material system. The roles of the homopolar dispersionless mode at 50 meV which typically occurs in layered structures as well as the sixth phonon mode with the highest energy will not be considered here. Due to the large difference in momentum between valleys in TMDCs, we assume that the exciton formation occurs via an LO phonon-assisted intravalley process which preserves the valley polarization in the monolayer system. The Hamiltonian term associated with the interaction between excitons and LO phonons is obtained by summing the electron-LO phonon and hole-LO phonon interaction Hamiltonians as follows H\^op(**r_e,r_h**) &=& \_**q** C , [\[hop\]]{#hop label="hop"}\ C &=& erfc\[ \|\| /2\] [\[coup\]]{#coup label="coup"} where \({\bf{r_e}}=\left(x_e,y_e,z_e\right) = \left(\vec{r_e},z_e\right)\) and \({\bf{r_h}} = \left( x_h,y_h,z_h \right) = \left( \vec{r_h},z_h \right)\) denote the respective space coordinates of the electron and hole, and \(\vec{r_e}\) (or \(\vec{r_h}\)) marked with an arrow denotes the monolayer in-plane coordinates of the electron (or hole). The phonon creation and annihilation operators are denoted by \(b^{\dagger}_{\vec{q},q_z}\) and \(b_{\vec{q},q_z}\), respectively, where \({\bf{q}}=\left(\vec{q}, q_z\right)\) is composed of the in-plane \(\vec{q}\) and perpendicular \(q_z\) components of the phonon wavevector. The term \(\omega_{LO}\) denotes the frequency of the LO phonon, \(\epsilon_o\) is the permittivity of free space, \(V\) is the volume of the crystal. The low-frequency and low-frequency relative dielectric constants are given by \(\kappa_{0}\) and \(\kappa_{\infty}\), respectively. The inclusion of the complementary error function \({\rm erfc[q} \sigma/2]\) where \(\sigma\) is the effective width of the electronic Bloch states is based on the constrained interaction of LO phonon with charge carriers in two-dimensional materials. For the monolayer MoS\(_2\), the Fröhlich coupling constant of 98 meV and an effective width \(\sigma\) = 4.41 Å provide good fit to the interaction energies evaluated from first principles in the long-wavelength limit. Due to dielectric screening, the Fröhlich interaction decreases with increase in the phonon momentum, and larger coupling values \(( \ge 330 )\) meV were obtained in the small momentum limit in another study. The Fröhlich coupling constants obtained in earlier works will be used in this study to compute the formation times of excitons. The field operator \(\hat\Psi^\dagger_{_{e-h}}\) of a pair of electron and hole with a centre of mass that moves freely is composed of electron and hole operators as follows [\[field\]]{#field label="field"} \^\_\_e-h(,,z_e,z_h) = \_, e\^-i e\^-i \_e(z_e) \_h(z_h) a\_v,\_h -\^ a\_c,\_e + \^, where \(A\) is the two-dimensional quantization area in the monolayer plane, and \(a_{v,\vec{K}}^{\dagger}\) (\(a_{c,\vec{K}}^{\dagger}\)) are the respective hole and electron creation operators with in-plane wavevector \(\vec{K}\). The center-of-mass wavevector \(\vec{K} = \vec{k_e}+\vec{k_h}\) and the wavevector of the relative motion \(\vec{k} = \alpha_h \vec{k_e}-\alpha_e \vec{k_h}\) where \(\vec{k_e}\) (\(\vec{k_h}\)) is the electron (hole) in-plane wavevector, with \(\alpha_e ={m_e}/(m_e+m_h)\), \(\alpha_h = {m_h}/(m_e+m_h)\) where \(m_e\) (\(m_h\) ) is the effective mass of the electron (hole). In Eq.[\[field\]](#field){reference-type="ref" reference="field"}, the excitonic center of mass coordinate \(\vec{R}\) and relative coordinate \(\vec{r}\) parallel to the monolayer plane are given by [\[coord\]]{#coord label="coord"} &=& \_e + \_h ,\ &=& -. The electron and hole wave functions (\(\psi_e(z_e)\), \(\psi_h(z_h)\)) in the lowest-energy states are given by \(\mathcal{N} \; \cos[\frac{\pi z_j}{L_w}]\) (j = e,h) for \(|z_j| \le \frac{L_w}{2}\), and 0 for \(|z_j| > \frac{L_w}{2}\). The term \(\mathcal{N}\) denotes the normalization constant and \(L_w\) is the average displacement of electrons and holes in the \(z\) direction perpendicular to the monolayer surface. ## Exciton creation Hamiltonian The field operator \(\hat\Psi^\dagger_{ex}\) of an exciton located at \(\left(\vec{R},\vec{r},z_e,z_h \right)\) differs from the field operator \(\hat\Psi^\dagger_{_{e-h}}\) of a free moving pair of electron and hole (see Eq.[\[field\]](#field){reference-type="ref" reference="field"}), and is given by [\[fieldex\]]{#fieldex label="fieldex"} \^\_ex(,,z_e,z_h) = \_ e\^-i \_ex\^() \_e(z_e) \_h(z_h) B\_\^, where \(B_{\vec{K}}^{\dagger}\) is the exciton creation operator with center-of-mass wavevector \(\vec{K}\) parallel to the monolayer plane. The 1s two-dimensional exciton wavefunction \(\rho_{ex}(\vec{r})\) = \(\sqrt{\frac{2 \beta^2}{\pi}} \exp(-\beta |\vec{r}|)\) where \(\beta\) is a variational parameter. Using Eqs. [\[hop\]](#hop){reference-type="ref" reference="hop"}, [\[field\]](#field){reference-type="ref" reference="field"} and [\[fieldex\]](#fieldex){reference-type="ref" reference="fieldex"}, the Hamiltonian associated with the formation of an exciton from an initial state of free electron-hole pair with absorption/emission of an LO phonon appear as [\[int\]]{#int label="int"} H\^F\_I &=& \_,,,q_z \_o F\_-(,,q_z) erfc\[ \] B\_\^ b\_, q_z a\_v,\_h (-)- a\_c,\_e (-) + \ &+& \_o F\_+(,,q_z) erfc\[ \] B\_\^ b\_, q_z\^ a\_v,\_h (+)- a\_c,\_e (+) + ,\ \ [\[ffact\]]{#ffact label="ffact"} F\_(,,q_z) &=& F_e(q_z) G( \_h )-F_h(q_z) G ( \_e ),\ [\[ffact2\]]{#ffact2 label="ffact2"} G( \_i ) &=& d \_ex\^() e\^i( \_i ) ,\ [\[ffact3\]]{#ffact3 label="ffact3"} F_i(q_z) &=& dz_i \|\_i(z_i)\|\^2 e\^i q_z z_j, i = e,h where the coupling constant \(\lambda_o = \sqrt{\frac{e^2 L_m \hbar \omega_{LO}}{2 \epsilon_o A} (\frac{1}{\kappa_{\infty}}-\frac{1}{\kappa_{0}})}\) and \(L_m\) is the monolayer thickness. The form factor \({\mathcal G}\) is evaluated using the explicit form of the two-dimensional exciton wavefunction \(\rho_{ex}(\vec{r})\). Likewise the second form factor \({\mathcal F}\) is computed using the electron wavefunction \(\psi_e(z_e)\) and hole wavefunction \(\psi_h(z_h)\). ## Exciton formation rate For transitions involving a single phonon with wavevector \(\vec{q}\), the formation rate of the exciton with wavevector \(\vec{K}\) is computed by employing the Fermi golden rule and the interaction operator in Eq.[\[int\]](#int){reference-type="ref" reference="int"} as follows [\[rate\]]{#rate label="rate"} W( ,q_z) &=& \|\_o\|\^2 \|F\_(,,q_z)\|\^2 erfc\[ \]\^2 f_h(\_h ( )-) f_e(\_e ( ) + )\ && (f\_ex()+1) (n\_**q** + ) (E\_ex-E\_eh\^\_LO), where the emission (absorption) of phonon is denoted by \(+\) (\(-\)), and the exciton energy \(E_{ex} = E_g + E_{_b}+ \frac{\hbar^2 |\vec{K}|^2}{2 (m_e+m_h)}\) where \(E_{_b}\) is the exciton binding energy. The energies of charge carrier is \(E_{eh}^\pm = E_g + \frac{\hbar^2 |\vec{K} \pm \vec{q}|^2}{2 (m_e + m_h)} + \frac{\hbar^2 |\vec{k}|^2}{2 \mu}\) where \(\mu\), the reduced mass is obtained using \(\frac{1}{\mu} = \frac{1}{m_e} + \frac{1}{m_h}\). At low temperatures of the charge carriers the phonon bath can be considered to be thermal equilibrium with negligible phonon-phonon scatterings and phonon decay processes.The thermalized average occupation of phonons for low temperatures of the charge carriers is given by [\[phon\]]{#phon label="phon"} n\_q = \[()-1\]\^-1, where \(T_l\) is the lattice temperature and \(\hbar \omega_{LO}\) is the energy of the LO phonon that is emitted during the exciton generation process. The relaxation of electrons and holes at high enough temperatures (\(\approx 200 K\)) generally displaces phonons beyond the equilibrium point when phonon-phonon related processes become dominant. The phonon Boltzmann equation takes into account a common temperature that is achieved as a result of equilibrium reached between electrons and phonons. Hot-phonon effects is incorporated by replacing the temperature \(T_l\) in Eq.[\[phon\]](#phon){reference-type="ref" reference="phon"} by an effective lattice temperature \(T_{ph}\). The charge carriers are assumed to be in quasi-thermal equilibrium during the exciton formation process. Consequently the occupation numbers (\(f_h(K)), f_e(K)\)) of hole and electron states in Eq.[\[rate\]](#rate){reference-type="ref" reference="rate"} can be modeled using the Fermi-Dirac distribution [\[fd\]]{#fd label="fd"} f\^i(K_i) &=&\^-1, i = e, h\ [\[fd2\]]{#fd2 label="fd2"} \_i &=& K_B T_i , where the chemical potential \(\mu_i\) is dependent on the temperature \(T_{i}\) and the two-dimensional density \(n_i\) of the charge carriers. When the mean inter-excitonic distance is higher the exciton Bohr radius as considered to be the case in this study, the exciton can be assumed to be an ideal boson with a Bose-Einstein distribution [\[bes\]]{#bes label="bes"} f\^ex(K) &=&\^-1,\ [\[bes2\]]{#bes2 label="bes2"} \_ex &=& K_B T\_ex , where \(\mu_{ex}\) is the exciton chemical potential, \(T_{ex}\) is the exciton temperature and \(n_{ex}\) is the exciton density. The degeneracy factor \(g\) is obtained as the product of the spin and valley degeneracy factors. ## Numerical results of the Exciton formation Time The formation time of an exciton with wavevector \(\vec{K}\), \(T(\vec{K})\) is obtained by summing the wavevectors \((\vec{k},\vec{q},q_z)\) over the rate obtained in Eq.[\[rate\]](#rate){reference-type="ref" reference="rate"} [\[formt\]]{#formt label="formt"} = \_,,q_z W( ,q_z) To obtain quantitative estimates of the exciton formation time using Eq.[\[formt\]](#formt){reference-type="ref" reference="formt"}, we use the monolayer MoS\(_2\) material parameters as \(m_e\) = 0.51 \(m_o\), \(m_h\) = 0.58 \(m_o\) where \(m_o\) is the free electron mass, and the coupling constant \(\alpha_o\)= 330 meV. We set the phonon energy \(\hbar \omega_{LO}\) = 48 meV, and the layer thickness \(h\) = 3.13 Å is used to determine the upper limit of \(\approx\) 6 Å for \(L_w\), the average displacement of electrons and holes in the direction perpendicular to the plane of the monolayer. We fix the effective lattice temperature \(T_{ph}\) = 15 K, but vary the electron and hole temperatures, \(T_e\) and \(T_h\). Fig. [\[formK\]](#formK){reference-type="ref" reference="formK"}a,b show the calculated values (using Eqs. [\[ffact\]](#ffact){reference-type="ref" reference="ffact"}-[\[formt\]](#formt){reference-type="ref" reference="formt"}) of the exciton formation times as a function of exciton wavevector \(\vec{K}\) with emission of an LO phonon at different electron, hole and exciton temperatures and densities, \(n_e\) = \(n_h\) = \(n_{ex}\) = 1 \(\times\) 10\(^{11}\) cm\(^{-2}\) and 5 \(\times\) 10\(^{11}\) cm\(^{-2}\). To obtain the results, we assume the temperatures to be the same for excitons and unbound electron-hole pairs. The results indicate that very fast exciton formation times of less than one picosecond time occurs at charge densities of 5 \(\times\) 10\(^{11}\) cm\(^{-2}\) and carrier temperatures less than 300 K. These ultrafast sub-picosecond exciton formation times are in agreement with recent experimental findings recorded at room temperatures in the monolayer MoS\(_2\). The exciton formation times are increased at the lower carrier densities of 1 \(\times\) 10\(^{11}\) cm\(^{-2}\). The wavevector of exciton states formed due to optical excitation of the ground state of the crystal lies close to zero due to selection rules. The results in Fig. [\[formK\]](#formK){reference-type="ref" reference="formK"}a,b show that while excitons are dominantly created at \(|\vec{K}|\) = 0 at low charge carrier temperatures (\(\approx\) 50 K), exciton formation occurs most rapidly at non-zero exciton center-of-mass wavevectors (\(|\vec{K}|_f \neq\)) at higher temperatures (\(T_e\) = \(T_h\) \(\ge\) 140 K) of the charge carriers. At \(T_e\) = \(T_h\) \(\approx\) 300 K, the shortest exciton formation time occurs at \(|\vec{K}|_f\) = 0.04 \(\times\) 10\(^{10}\) m\(^{-1}\) (about 5.6 meV). The results in Fig. [\[formK\]](#formK){reference-type="ref" reference="formK"}a,b indicate that at exciton wavevectors greater than 0.06 \(\times\) 10\(^{10}\) m\(^{-1}\), there is a notable increase in the exciton formation times linked to low electron-hole plasma temperatures \(T_e = T_h \le\) 80 K. At high carrier temperatures there is likely conversion of newly formed composite bosons such as excitons into fermionic fragment species . The inclusion of considerations of the quantum mechanical crossover of excitons into charge carriers at higher plasma temperatures will add to greater accuracy when computing exciton formation times. This currently lies beyond the scope of this study and will be considered in future investigations where the role of the composite structure of excitons in their formation rate will be examined. The effect of the variations within the electron-hole plasma temperatures or differences between \(T_e\) and \(T_h\) on the exciton formation time is illustrated in Fig. [\[diff\]](#diff){reference-type="ref" reference="diff"}. The formation times are computed at different exciton center-of-mass wavevectors with the electron temperature fixed at \(T_e\) = 250 K, and exciton temperature \(T_{ex}\) = 50 K. The charge densities \(n_e\) = \(n_h\) = \(n_{ex}\) = 5 \(\times\) 10\(^{11}\) cm\(^{-2}\) and all other parameters used are the same as specified in the caption for Fig. [\[formK\]](#formK){reference-type="ref" reference="formK"}. At the larger wavevector \(|\vec{K}|\) = 0.07 \(\times\) 10\(^{10}\) m\(^{-1}\) (\(\approx\) 17.1 meV) the formation time is shortest when the hole temperature \(T_h\) = \(\approx\) 120. With decrease in the center-of-mass wavevector \(|\vec{K}|\), there is corresponding decrease in the formation time when the hole temperature is allowed to decrease further from the electron temperature. At the low exciton wavevector \(|\vec{K}|\) = 0.005 \(\times\) 10\(^{10}\) m\(^{-1}\), the shortest formation time occurs when the difference between \(T_e\) and \(T_h\) reach the maximum possible value. These results demonstrate the interplay of competitive effects of the hole-phonon and the electron-phonon dynamics on a picosecond time scale which results in a non-monotoinic temperature difference dependence \(|T_h-T_e|\) of the exciton formation time. In Fig. [\[concf\]](#concf){reference-type="ref" reference="concf"}, the exciton formation time is plotted as a function of the carrier density \(n_e\) = \(n_h\) at different temperatures, \(T_e\) = \(T_h\) = \(T_{ex}\) (300 K, 200 K, 100 K). The exciton center-of-mass wavevector \(|\vec{K}|\) = 0.03 \(\times\) 10\(^{10}\) m\(^{-1}\). All other parameters used are the same as specified in the caption for Fig. [\[formK\]](#formK){reference-type="ref" reference="formK"}. Using the numerical values of the formation times, we performed numerical fits using the following relation which involve the carrier concentrations [\[fitT\]]{#fitT label="fitT"} T(n_i) = i = e, h where \(B\) and \(p\) are fitting parameters. Using the results used to obtain Fig. [\[concf\]](#concf){reference-type="ref" reference="concf"}, we get \(B\) = 20.64 at \(T_e\) = \(T_h\) = 300 K, \(B\) = 10.35 at \(T_e\) = \(T_h\) = 200 K, \(B\) = 3.56 at \(T_e\) = \(T_h\) = 100 K and \(B\) = 1.54 at \(T_e\) = \(T_h\) = 50 K. The constant \(p \approx\) 2 irrespective of the electron and hole temperatures. This implies an inverse square-law dependence of the exciton formation time on the electron/hole concentration. Consequently a square-law dependence of the photoluminescence on excitation density is expected to arise in the monolayer MoS\(_2\) as well as other monolayer transition metal dichalcogenides. # Exciton formation times for other exemplary monolayer transition metal dichalcogenides The theoretical results obtained in this study for MoS\(_2\) are expected to be applicable to other low dimensional transition metal dichalcogenides. However subtle variations in the exciton formation times are expected due to differences in the exciton-LO coupling strengths and energies of the LO phonon in the monolayer materials. The bare Fröhlich interaction strengths obtained via ab initio techniques give 334 mev (MoS\(_2\)), 500 meV (MoSe\(_2\)), 140 mev (WS\(_2\)) and 276 meV (WSe\(_2\)) , hence the Molybdenum-based TMDCs possess higher exciton-phonon coupling strengths than the Tungsten-based TMDCs. A precise estimate of the exciton binding energy in the monolayer TMDCs is not available, however a range of binding energies (100 to 800 meV) have been reported for the monolayer systems [@hanbicki2015measurement; @olsen2016simple; @makatom; @chei12; @ugeda2014giant; @hill2015observation; @choi2015linear; @chei12; @komsa2012effects; @thiljap]. In order to compare the exciton formation rates between Molybdenum-based TMDCs and Tungsten-based TMDCs, we make use of the effective masses of electron and holes at the \(K\) energy valleys/peak position given in Ref. and the Fröhlich interaction strengths and LO phonon energies given in Ref.. To simplify the numerical analysis, we fix the exciton binding energies at \(\approx\) 330 meV for all the TMDCs under investigation. This assumption is not expected to affect the order of magnitude of the exciton formation times, and also to not detract from the analysis of effects of Fröhlich interaction strengths on the formation times. The results in Fig. [\[comp\]](#comp){reference-type="ref" reference="comp"} a, b show that the exciton formation times of the selenide-based dichalcogenides are smaller than the sulphide-based dichalcogenides at \(T_e\) = \(T_h\) = \(T_{ex}\) = 100 K and 300 K (with \(|\vec{K}| \le\) 0.05 \(\times\) 10\(^{10}\) m\(^{-1}\)). This is due to the comparatively higher Fröhlich interaction strengths and lower LO phonon energies of monolayer MoSe\(_2\) and WSe\(_2\). The results in Fig. [\[comp\]](#comp){reference-type="ref" reference="comp"}a,b also indicate that excitons in the monolayer WS\(_2\) are dominantly created at non-zero center-of-mass wavevectors compared to the other three monolayer dichalcogenide systems. This may be attributed to the comparatively lower effective masses of hole and electron in the monolayer WS\(_2\). It is instructive to compare the exciton formation times in Fig. [\[comp\]](#comp){reference-type="ref" reference="comp"}a,b with the radiative lifetimes of zero in-plane momentum excitons in suspended MoS2 monolayer of \(\approx\) 0.18-0.30 ps at 5 K . The lifetimes of excitons depend linearly on the exciton temperature and increase to the picoseconds range at small temperatures and is larger than 1 ns at the room temperature. This indicates that the exciton formation processes are likely to dominate in the initial period when the TMDCs are optically excited at high exciton temperatures. In the low temperature range (5 K-20 K), an interplay of competing effects of exciton generation and radiative decay are expected to occur on the sub-picosecond time scale. Environmental parameters such as impurity concentration, exciton density and density of excess charge carriers that affect the stability of low dimensional trions will need to be taken into account in order to accurately model the exciton generation process at the low temperature regime. The exciton formation scheme adopted in this study has been parameterized by physical quantities such as the exciton density and charge carrier densities. It is not immediately clear whether these parameters can be extracted directly using ab-initio quantum mechanical and time-dependent density functional theory approaches. Computations based on ab-initio techniques are generally numerically intensive and time consuming which are the main challenges in modeling low dimensional material systems. It is expected that improvements in first principles modeling of anisotropic systems may result in more efficient and rewarding approaches to determining the density functions of excitons and charge carriers in future investigations. The Auger process provides a non-radiative decay channel for electron-hole pair recombination, hence this mechanism must be taken into account for accurate predictions of exciton formation times in future studies. # Conclusion [\[conc\]]{#conc label="conc"} Transition metal chalcogenides have emerged as promising materials in which excitons exist as stable quasi-particles with high binding energies and thus play important roles in the optical processes of monolayer TMDCs. The dynamics of excitons in monolayer transition metal dichalcogenides has been extensively studied over the last five years in terms of both theory and applications. However the formation of excitons from free carriers has only been recently measured, and in this work we develop a model within the framework of Fermi's Golden rule to calculate the formation dynamics of excitons from free carriers. This theoretical study is aimed at providing a fundamental understanding of the exciton generation process in optically excited monolayer transition metal dichalcogenides. We focus on a mechanism by which excitons are generated via the LO (longitudinal optical) phonon-assisted scattering process from free electron-hole pairs in layered structures. The exciton formation time is computed as a function of the exciton center-of-mass wavevector, electron and hole temperatures and densities for known values of the Fröhlich coupling constant, LO phonon energy, lattice temperature and the exciton binding energy. Our results show that the exciton is generated at non-zero wavevectors at higher temperatures (\(\ge\) 120 K) of charge carriers, that is also dependent on the density of the electron and holes. The inverse square-law dependence of the exciton formation time on the density of charge carriers is also demonstrated by the results of this study. For monolayer MoS\(_2\), we obtain exciton formation times on the picosecond time scale at charge densities of 1 \(\times\) 10\(^{11}\) cm\(^{-2}\) and carrier temperatures less than 100 K. The exciton formation times decreases to the sub-picosecond time range at higher densities (5 \(\times\) 10\(^{11}\) cm\(^{-2}\)) and electron-hole plasma temperatures (\(\le\) 300 K). These ultrafast formation times are in agreement with recent experimental results (\(\approx\) 0.3 ps) for WSe\(_2\), MoS\(_2\), and MoSe\(_2\). Due to the comparatively higher Fröhlich interaction strengths and lower LO phonon energies of monolayer MoSe\(_2\) and WSe\(_2\), the exciton formation times of the selenide-based dichalcogenides are smaller than the sulphide-based dichalcogenides at \(T_e\) = \(T_h\) = \(T_{ex}\) = 100 K and 300 K (with \(|\vec{K}| \le\) 0.05 \(\times\) 10\(^{10}\) m\(^{-1}\)). The results of this study is expected to be useful in understanding the role of the exciton formation process in electroluminescence studies and exciton-mediated processes in photovoltaic devices.

{'timestamp': '2016-09-22T02:02:41', 'yymm': '1607', 'arxiv_id': '1607.04856', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04856'}

# A letter from Saratov {#story} During his last trip to Moscow, the second author of this article came into possession of a remarkable mathematical letter. The custodian of the letter, a Russian businessman X who wished to remain anonymous, presented the letter to Tabachnikov at the end of his lecture on configuration theorems in projective geometry (like many successful contemporary Russian entrepreneurs, Mr. X has a degree in mathematics). Mr. X explained that the letter had been written by his great-great-great grandfather, Konstantin Shestakov, shortly after he had been discharged from the Russian Army during the Napoleonic Wars. It appears that K. Shestakov had befriended the famous French mathematician, Jean-Victor Poncelet, then a prisoner of war in the Russian town of Saratov, and had been drawn into geometry by him.[^1] The letter, dated fall of 1814 and apparently never mailed, was addressed to Shestakov's younger brother, Alexander, who had attended the Kazan Gymnasium with N. I. Lobachevsky. We reproduce the first paragraph of the letter in the original Russian (in the modern orthography) and give its full translation into English. > ::: {.otherlanguage} > russian Милый Саша! Ты конечно слышал, что раны мои, полученные при битве под Смоленском, вынудили меня выйти в отставку, и что я постепенно восстанавливаю силы свои в Саратове. Это сонный город на Волге, где окромя роскошных обедов да балов у здешнего губернатора г. Панчулидзева и делать вовсе нечего. Служба моя состоит и в надзоре за здешней тюрьмой, однако заключённые ведут себя смирно и не нуждаются в постоянном присмотре. Пишу сообщить о своём житье-бытье, а также и с просьбой переслать математическую часть сего письма однокашнику твоему по гимназии Николаю Лобачевскому; как слыхал я, он теперь адъюнкт-профессор в Казани\... > ::: > *Dearest Sasha: You doubtless have heard that the wounds sustained at a battle near Smolensk forced me to withdraw from the army and that I have been recovering ever since in Saratov. Saratov is a dreary town on the Volga river which, aside from the several excellent dinner parties and balls thrown by Governor Panchuleedzev, affords us almost nothing interesting to do. I have some hand in overseeing the prison, but the prisoners are peaceful and rarely need attention. I am writing this partly to tell you how I have been keeping myself, but also to ask you the favor of passing on the mathematical part of this letter to your former classmate Nikolai Lobachevsky who, as I understand it, is now an adjunct professor at the Kazan University. I spent much of the winter holed up in my room, trying to avoid the oppressive regimen of drinking and other kinds of enforced merriment. As you know, I prefer a life of quiet contemplation. I had never quite become used to the cruel interruption of my studies brought about by the war. I confess that my bitterness over the situation had been driving me towards madness, but then everything changed. I write to you now with great excitement, as a man given a divine gift. My gift came to me in the form a prisoner, Lieutenant of the French Army Monsieur Poncelet. As luck would have it, Poncelet was assigned to me as an assistant of sorts, though the pedestrian kind of work required of him hardly touched on his enormous talents. It became clear almost immediately that he was nothing at all like our other prisoners. Quiet and unassuming, with a temperament much like my own, Poncelet preferred to keep his own company. The relatively undemanding work he did for me afforded him to pursue his true passion, which was geometry. Poncelet explained to me that he was whiling away the time by reconstructing from first principles the theorems of geometry he had learned during his student days at École Polytechnique in Paris. He said more than once that the great beauty of mathematics is that it can be created whole from very few principles. I knew something of geometry myself and I gradually became engaged in the project of this brilliant man. In the space of several months I went from casual observer to eager student to active participant. You have to understand that Poncelet was no ordinary student! As time passed, I could see that he was not just trying to recreate the lessons of his masters, as I had first thought, but rather that he was striking out into the unknown. My excitement was almost as great as his when he explained to me his astonishing discovery about conic sections. This is the part of the letter I request that you pass on to Lobachevsky. I recall that your mathematics teacher, the venerable Grigorii Ivanovich, always praised Nikolai for his aptitude for mathematics. I trust that you will be able to follow this but, in any case, Lobachevsky certainly will, and I hope that it will interest him. Suppose that \(A\) and \(B\) are elliptical conic sections with \(A\) surrounding \(B\). Imagine that one has a point \(P_1\) on \(A\), then draws a segment \(P_1P_2\) which is tangent to \(B\) and connects \(P_1\) to \(P_2\), also on \(A\). Now repeat the construction, drawing the segment \(P_2 P_3\) tangent to \(B\) so that \(P_3\) is on \(A\). Imagine that this construction is repeated some number of times, say 100, so that \(P_{101}\) is the same point as \(P_1\). One has produced a kind of closed polygonal figure involving points \(P_1, P_2\), etc., which ends up where it starts after 100 steps. Poncelet's great discovery is that the same construction, starting with a different point on \(A\), will also repeat after 100 steps. That is, if one moves \(P_1\) to the new point \(P'_1\) on \(A\), and then produces points \(P'_2, P'_3\), etc. then it will again happen that \(P'_{101} = P'_1\)! You have to understand, finally, that I have only used 100 as an example. Whether the figure repeats itself or not, and after how many steps, depends on the choice of \(A\) and \(B\). Poncelet encouraged me to think about this construction, as he did, as a kind of spinning polygon. He imagined moving the point \(P_1\) continuously, so that the other points would move as well. Poncelet's great discovery is that this configuration of (say) 100 points and 100 segments spins around the conics \(A\) and \(B\) and remains intact. Poncelet left the camp in June and I do not know what became of him, but I continued his work. I could not get the idea of the spinning configuration out of my mind and eventually I thought to ask some questions of my own. For simplicity I considered the case when \(A\) and \(B\) were both circles, with \(A\) surrounding \(B\) and \(B\) being set in a position that was somewhat offset from the center of \(A\). By adjusting the center and position of \(B\) I could control the particulars of Poncelet's construction. After some amount of fussing around, I settled on a location and size of \(B\) which caused the configuration to close up after 5 steps, making a star-like pattern. I became interested in the question of how the center of gravity of the figure changed as it spun around. You might say that I approached this in an experimental way, drawing the figure on the page and then approximating the center of gravity by a kind of trick of drawing the points together. I wish that I could include diagrams of this laborious method, but perhaps they would mean nothing to you. Suffice it to say that my calculations seemed to show a promising result but were not accurate enough to convince me. You probably remember my passion for precise reasoning! Eventually I joined many sheets together, making a kind of poster on which to do my calculations. I remember spending the bulk of a week scribbling on this poster, night after night, as gradually the beautiful answer revealed itself. It is yet another circle! As the great Poncelet figure spins around, its center of gravity traces out a circle as well! Understand that there is nothing special about the star configuration that I settled upon; I expect that the principle should be completely general. I am convinced that this is a first-rate discovery to rival Poncelet's own. I regret that I have so far not hit upon a proof of my discovery, in the same way that Poncelet was able to find a proof for his. Alas, perhaps if my friend were still here we could find it together. Nonetheless, I have calculated things out to such a precision that I am confident in proclaiming this as a theorem. I would be grateful if Lobachevsky could look into the matter. Your brother Konstantin.* To put this letter into a historical context, Jean-Victor Poncelet participated in Napoleon's invasion of Russia and was wounded at the Battle of Krasnoi during Napoleon's retreat from Moscow in November of 1812. He spent more than a year in Russian captivity in Saratov. There he developed projective geometry. Later, his *Saratov notebooks* became part of his *Traité des propriétés projectives des figures* (1822) and *Applications d'analyse et de géométrie* (1862). In particular, Poncelet discovered his celebrated porism during the Saratov exile. N. I. Lobachevsky grew up in Kazan where he attended gymnasium[^2] and then the newly established Kazan University. The whole career of Lobachevsky was spent at this university where he served from 1814, as an adjunct professor, until 1846 (elected Chancellor in 1827). The literature on the Poncelet theorem and its ramifications is vast; see, e.g., and the references therein. Still, after having mulled over Shestakov's letter at some length, we feel that his result appears to be a new and surprising addition to the Poncelet theorem. What makes the result especially surprising is that it makes a statement about the center of mass--an affine geometry construction--in the context of projective geometry. The one result we know in this direction is Weill's Theorem, which deals with the centers of mass of Poncelet polygons in a special case. One could view Shestakov's Theorem as a generalization of Weill's Theorem, though Shestakov's result does not quite imply Weill's result. In what follows, we will present and sketch modern proofs for two versions of Shestakov's result. For reasons we will explain, we think that he most likely had the first statement in mind, but it seems reasonable to present the second statement as well, which is a variant. At the end, and for the sake of completeness, we will sketch a proof of Weill's Theorem. # The theorem {#mainform} Figure [\[pentagon\]](#pentagon){reference-type="ref" reference="pentagon"} features an instance of the Poncelet porism--exactly the kind Shestakov said that he had "fussed around to arrange". The Poncelet theorem asserts that every point of the outer conic is a vertex of such a polygon. The interested reader can use[^3] our program and see the kind of animation envisioned by Poncelet. Let \(P=(A_1\ldots A_n)\) be a Poncelet \(n\)-gon, and let \((x_i,y_i)\) be the Cartesian coordinates of the vertex \(A_i\), \(i=1,\ldots,n\). There are really \(3\) natural interpretations for the center of mass of \(P\): - The center of mass of the vertices. We denote this by \(CM_0(P)\). - The center of mass of the edges, when the edges are given a uniform density. We denote this by \(CM_1(P)\). - The center of mass of the \(P\) when \(P\) is considered as a "homogeneous lamina". We denote this by \(CM_2(P)\). When \(P\) is not embedded, \(CM_2(P)\) requires the interpretation we give in Lemma [\[CMcoord\]](#CMcoord){reference-type="ref" reference="CMcoord"} below. We provide formulas for the three kinds of centers below. We will present version of Weill's Theorem for \(CM_0\) and \(CM_2\). The result fails for \(CM_1\) for reasons we will mention briefly below. We think that Shestakov most likely had \(CM_0\) in mind from his description of "drawing the points together". This is the sort of thing one would do when computing \(CM_0(P)\). The coordinates of these centroids are given in the next lemma. Let \[\ell_i=\sqrt{(x_{i+1}-x_i)^2 + (y_{i+1}-y_i)^2}\] be the length of \(i\)th side of the polygon, and let \(d_i = (x_i y_{i+1}-x_{i+1} y_i).\) Then \(L(P)=\sum_i \ell_i\) is the perimeter of \(P\), and \(A(P)=(1/2) \sum_i d_i\) is its area, counted with sign and multiplicity. #### Proof. Consider the triangle \(O A_i A_{i+1}\) where \(O\) is the origin. The centroid of this triangle is at point \((A_i + A_{i+1})/3\) and it has the area \(d_i/2\). Summing up over \(i\) and dividing by the total area of \(P\) yields the result for \(CM_2(P)\). Likewise for \(CM_0(P)\) and \(CM_1(P)\). \(\Box\)Here is (our interpretation of) Shestakov's Theorem. As we mentioned above, Shestakov most likely had \(CM_0(P)\) in mind. Also, he has normalized so that the outer ellipse is a circle; in this case the loci \(CM_0(P_t)\) and \(CM_2(P_t)\) are circles. For the proof we will also normalize this way. Our proof relies on the fact that \(CM_0(P)\) and \(CM_2(P)\) are rational expressions in the coordinates of \(P\). This is not true for \(CM_1(P)\) and, as we mentioned above, the result fails for \(CM_1(P)\). The locus in this case is not generally a conic section. We wrote a computer program which tests Theorem [\[main\]](#main){reference-type="ref" reference="main"}. Figure [\[experiment\]](#experiment){reference-type="ref" reference="experiment"} shows two pictures from this program. The left side deals with \(CM_0\) and the right side deals with \(CM_2\). In both cases, the loci \(CM_0(P_t)\) and \(CM_2(P_t)\) are black circles. The two Poncelet polygons shown belong to the same Poncelet family, but they are not the same polygon. The two circles \(CM_0(P_t)\) and \(CM_2(P_t)\) are different circles. We are not sure how they are related to each other. # A proof {#mainpf} Now we sketch a proof of Theorem [\[main\]](#main){reference-type="ref" reference="main"}. Let us start with the algebraic geometry proof of the Poncelet porism, see, or, for a detailed account. One starts with complexifying and projectivizing: we assume that \(\gamma\) and \(\Gamma\) are complex conics in general position in the complex projective plane \({\mathbb {CP}}^2\) obtained from the affine plane \({\mathbb C}^2\) by adding a line at infinity. A complex conic is isomorphic to Riemann sphere, \({\mathbb {CP}}^1\). Let \(E\) be the set of flags \((x,L)\) consisting of a point \(x \in \Gamma\) and a line \(L\) through \(x\) that is tangent to \(\gamma\). The space \(E\) is naturally a Riemann surface, and the projection \(p:E \to \Gamma\) that takes \((x,L)\) to point \(x\) is a 2-fold branched covering with four branch points. These branch points, invisible in Figure 5, are the four intersection points of \(\Gamma\) and \(\gamma\). One easily computes that the Euler characteristic of \(E\) is zero. Thus \(E\) is a Riemann surface which, topologically, is a torus. One has two involutions on \(E\) depicted in Figure [\[involutions\]](#involutions){reference-type="ref" reference="involutions"}. \[\sigma (x,L) = (x',L),\ \ \tau(x',L) = (x',L').\] The Poncelet map \(T=\tau \circ \sigma\) is a translation of \(E\). If a translation of a torus has a periodic point of order \(n\), then \(T^n\) is the identity, and this proves the Poncelet theorem. The two claims of Theorem [\[main\]](#main){reference-type="ref" reference="main"} are proved similarly, so we let \(CM\) mean either of the two centers of mass involved. The group generated by \(\sigma\) and \(\tau\) is the dihedral group \(D_n\). The coordinates \((x,y)\) of the center of mass \(CM\) of the Poncelet polygons are \(D_n\)-invariant meromorphic functions on \(E\). Indeed these coordinates are rational functions of the coordinates of the vertices of the polygons (Lemma [\[CMcoord\]](#CMcoord){reference-type="ref" reference="CMcoord"}), which, in turn, are rational functions of the rational parameter on the conic, which is a meromorphic function \(p\) on \(E\). Without loss of generality, assume that \(\Gamma\) is a circle. From the complex projective point of view, a circle is a conic that passes through the two circular points \((1:\pm i:0) \in {\mathbb {CP}}^2\) at infinity. When does either coordinate \(x\) or \(y\) of \(CM\) go to infinity? It happens exactly when one of the vertices of the Poncelet \(n\)-gons coincides with one of the two circular points. This occurs twice for each vertex, with multiplicity one,[^4] and then both \(x\) and \(y\) go to infinity. Therefore the functions \(x\) and \(y\) have \(4n\) simple poles on \(E\), and these poles comprise two orbits of the group \(D_n\), corresponding to the two circular points. Let \(A\) be a pole corresponding to point \((1:i:0)\), and let \(z\) be a local holomorphic parameter on \(E\) at \(A\). Then, at \(A\), \[x(z) = \frac{a_1}{z}+ a_2 + \ldots,\quad y(z) = \frac{ia_1}{z}+ a_3 + \ldots\] where \(a_i \in {\mathbb C}\), and the dots denote terms of degree one or higher in \(z\). Likewise, let \(B\) be a pole corresponding to \((1:-i:0)\). We have, in a local parameter \(w\) at \(B\): \[x(w) = \frac{b_1}{w}+b_2 + \ldots,\quad y(w) = \frac{-ib_1}{w}+ b_3 + \ldots\] We want to find \(u,v\in{\mathbb C}\) such that the function \(F=(x-u)^2+(y-v)^2\) has no poles at \(A\) and \(B\). Expanding out \(F\) at \(A\) and \(B\) yields: \[F=\frac{2a_1}{z}\left((a_2-u)+i(a_3-v)\right) + \ldots \ {\rm and}\ \ F=\frac{2b_1}{w}\left((b_2-u)-i(b_3-v)\right) + \ldots\] Equating both parentheses to zero yields two linear equations on \(u\) and \(v\) which we can easily solve. Consider the function \(F\) when \(u\) and \(v\) are the solutions to these simultaneous equations. The function \(F\) is \(D_n\)-invariant and therefore has no poles at all. Hence \(F\) is holomorphic on \(E\). But the only holomorphic functions on a compact Riemann surface are constants. So, \((X-u)^2+(Y-v)^2=r\), where \(r\) is some constant. Since \(CM(P_t)\) contains a curve of points in \({\mathbb R}^2\), the three constands \(u,v,r\) must be real. In particular, the complexified locus intersects \({\mathbb R}^2\) in the circle of radius \(r\) centered at \((u,v)\). # Weill's Theorem {#Weill} Here is Weill's Theorem, stated in the language of Theorem [\[main\]](#main){reference-type="ref" reference="main"}. We present a proof adapted from. We start with Bertrand's proof of the Poncelet theorem in the case when both conics are circles (see, e.g., or ).) There is a subtlety here that we want to emphasize. For the proof of Poncelet's theorem, which belongs entirely to the domain of projective geometry, one can always normalize that both conics are circles. However, for Theorem [\[main\]](#main){reference-type="ref" reference="main"} one cannot do this because the various centers of mass are not projectively invariant. Consider Figure [\[Bertrand\]](#Bertrand){reference-type="ref" reference="Bertrand"}. Call the map \(A_i \mapsto A_{i+1}\) the Poncelet map. The proof consists of constructing a cyclic coordinate \(t\) on the outer circle such that, in this coordinate, the Poncelet map is \(t \mapsto t+c\), where the constant \(c\) depends on the mutual position of the circles. The periodicity property of such a map depends only on \(c\), and not on the starting point. What follows is, essentially, the argument from Theorem XXX, figure 102, in I. Newton's "Principia\" where Newton studies the gravitational attraction of spherical bodies. Assume that points \(A_i\) and \(B_i\) are infinitesimally close (we keep them apart to avoid cluttering the figure). The triangles \(A_iB_iX\) and \(A_{i+1} B_{i+1}X\) are similar, and hence \[\frac{|A_i B_i| }{|A_iX|} = \frac{|A_{i+1} B_{i+1}|}{|B_{i+1}X|}.\] Denote by \(F(A)\) the length of the tangent segment from a point \(A\) on the outer circle \(\Gamma\) to the inner circle \(\gamma\). Then, in the limit \(B_i \to A_i\), \[\frac{dA_i}{F(A_i)} = \frac{dA_{i+1}}{F(A_{i+1})},\] where \(d A\) is the angular measure on the circle. It follows that the measure \(dt(A) := dA/F(A)\) is invariant under the Poncelet map. Hence the map is a translation in the \(t\)-coordinate on \(\Gamma\). Now we are ready to prove Weill's Theorem. Denote the infinitesimal angle \(A_iXB_i\) by \(\varepsilon\). Up to infinitesimals, the angles made by the segments \(A_i A_{i+1}\) and \(B_i B_{i+1}\) with the circle are equal; let \(\varphi\) denote this angle. Let the outer circle be unit, and the inner have radius \(R\). Then, in the limit \(B_i \to A_i\), \[\frac{|C_iD_i|}{|A_iA_{i+1}|} = \frac{\varepsilon R}{2\sin\varphi} = \frac{R}{2} \frac{|A_iB_i|}{|A_iX|} = \frac{R}{2}\, dt(A_i),\] where the second equality follows from the Sine Rule for triangle \(A_iXB_i\). Consider an infinitesimal motion of the Poncelet polygon \(P=(\ldots A_i A_{i+1}\ldots)\) to \(P'=(\ldots B_i B_{i+1} \ldots)\). The center of mass of the points \(C_i\) is displaced by the infinitesimal vector \[\sum_i C_i D_i = \frac{R}{2}\, dt\, \sum_i A_i A_{i+1},\] and the latter sum is zero because the polygon \(P\) is closed. Therefore the velocity of the center of mass vanishes, and it remains stationary. . We are deeply grateful to Mr. X for sharing with us the historical letter from his family archive. Many thanks to A. Akopyan, I. Dolgachev, D. Ivanov, J. Jeronimo, S. Lvovsky, S. Markelov for useful discussions, and to A. Zaslavsky for making available their harder-to-find article. The authors were inspired by Martin Gardner's essay. The first author was supported by NSF grant DMS-1204471. The second author was supported by NSF grants DMS-1105442 and DMS-1510055. [^1]: According to the family history, Shestakov was well educated: he attended Saint Petersburg Engineering School. [^2]: His teacher of mathematics was G. I. Kartashevsky. [^3]: You can download the program at http://www.math.brown.edu/\(\sim\)res/Java/PORISM.tar [^4]: We assume that the conics are in general position. [^5]: We thank A. Akopyan for this argument.

{'timestamp': '2016-07-19T02:04:54', 'yymm': '1607', 'arxiv_id': '1607.04766', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04766'}

# Introduction The idea of optimally controlling the dynamical behavior of microsystems described by quantum mechanics has been an explicit dream for more than \(50\) years. Among all the approaches to solving the resulting optimization problems, model-based optimal control relies on computer simulations of quantum dynamics and numerical optimization algorithms. The early attempt to seek the counterpart of the classical optimal control theory in quantum systems may be traced to Rabitz *et al.*, in which they explicated basic concepts of quantum optimal control theory which paved the way for later approaches. From then on, the studies have been enriched by numerous optimization algorithms with various applications. Generally, a mathematical model of the underlying quantum system is required for designing optimal controls, and the strategies for solving this model-based optimization problem can be roughly divided into two classes. One approach is based on the necessary conditions derived from Pontryagin's Maximum Principle, i.e. \"considering a pointer moving back and forth within the time interval\". For that, Zhu, Sundermann, and Ho *et al.* proposed several quadratically convergent iterative algorithms; Sklarz, Reich, and Schirmer *et al.* developed different kinds of Krotov algorithms. The other approach is to eliminate the \"Quantum Constrains\", i.e. eliminating the equality constrain of Schrödinger equation (SE) and apply classical optimization algorithms. The quantum gradient algorithm was first introduced by Shi *et al.* based on difference method, and it was improved by Khaneja *et al.* through a much more convenient way for calculating the gradient, saying the Gradient Ascent Pulse Engineering (GRAPE) method; besides, algorithms such as the Newton algorithm, quasi-Newton algorithm, and various evolutionary algorithms were also proposed. The efficiency of model-based optimizations is usually determined by two factors. One is how fast the optimization algorithm can converge to the desired solution, and the other is how fast the dynamical evolution under a given control protocol can be calculated. In order to reduce the number of iterations, Ho *et al.* proposed the two-point boundary-value quantum control paradigm (TBQCP), Eitan *et al.* combined the Krotov and quasi-Newton methods, and Watts *et al.* constructed a easily computable mathematical expression for the target function. More importantly, the simulation of quantum dynamical evolution need to be fast. Palao *et al.* restricted the objective to only the states used directly as registers, Yip *et al.* proposed the concept of \"propagation toolkit\" to simplify control functions, Arai *et al.* introduced the reduced time-evolution operators. In our previous work, we proposed the Pulse Width Modulation (PWM) method for efficiently simulating time-dependent systems. It transforms an arbitrary bandwidth-limited Hamiltonian into a sequence of on and off \"Hamiltonian Pulses\" (called Hamiltonian in PWM form), and reduces the simulation time by a large extent. In this paper, we design a gradient algorithm based on PWM and study the resulting improvement on the optimization efficiency with respect to multi-qubit quantum systems. The rest of this paper is organised as follows. Section II briefly reviews the PWM method, and Sec. III presents the problem, derives the optimization algorithm for one-control and multi-control cases. We apply the algorithm to \(3\)-qubit and \(6\)-qubit systems, and compare computational efficiency with the GRAPE algorithm in Sec. IV. We show how to smooth the optimized control function by example of transforming rectangular pulses into Gaussian pulse train in Sec. V. Finally, we summarise all the results and draw the conclusions in Sec. VI. In addition, App. A shows different realizations of PWM transformation and compares PWM with \(2\)nd-order Split Operator Method (SPO), and App. B presents the detailed Hamiltonian for the D-Norleucine we used for numerical simulation. # Brief review of PWM {#PWM} Pulse Width Modulation (PWM) is a well accepted concept in control technology in power electronics, which is based on the fact that any function \(u(t)\) with finite frequency bandwidth can be well approximated at any precision by a sequence of rectangular pulses with unanimous amplitude \(\varepsilon=\max_{t}{|u(t)|}\) but different pulse width \(t_p\). To apply PWM to a quantum system, we define the Frequency Band \(\Delta=[\omega_{min},\omega_{max}]\) as the interval from the minimum to the maximum frequency of all the time-dependent variables in the Hamiltonian, and the Frequency Scope \(\Omega= M\omega_{min}\) as the frequency limit one should care about, where \(M\) is called the Pulse Number within every time interval \({2\pi}/{\omega_{min}}\). Consider an arbitrary time-dependent Hamiltonian in the semiclassical form \[\label{TD} \hat H(t)=\hat H_0+\sum_{i=1}u_{i}(t)\hat H_{i},\] where \(H_0\) and \(H_i\)'s are time-independent Hamiltonians, \(u_{i}(t)\)'s are time-dependent real functions with finite bandwidth. The procedure of applying PWM in quantum system is based on Equal Integral Area Principle (EAP); in short, we divide every time interval \(2\pi/\omega_{min}\) into \(M\) pieces with equal length \(\tau\), in each we replace \(u_i(t)\) by a rectangular function with amplitude \(\varepsilon_i\) and width \(t_{p,i}^{(k)}\) which contains equal integral area. In this way, the time-dependent Hamiltonian is transformed into the PWM form \[\label{TID} \hat H(t)={\hat H}_0+ \sum_{i=1}s_{i}(t)\varepsilon_i{\hat H}_i,\] where \(s_{i}(t)\) is a sign function which switches among \(0,\pm1\); \(\varepsilon_i\geq\max_{t}{|u_i(t)|}\) is a constant. We further define anti-PWM transformation as transforming Eq. [\[TID\]](#TID){reference-type="eqref" reference="TID"} into Eq. [\[TD\]](#TD){reference-type="eqref" reference="TD"}, which replaces rectangular pulse in each time interval \(\tau\) by a constant with the same integral area; moreover, the Piecewise Constant (PWC) result may be further smoothed through EAP, which may be desired in many cases. However, as we will see the anti-PWM transform cannot always work well when one sets a relatively large time interval \(\tau\) for simplifying the optimization in complex systems, how to smooth PWM pulses through smooth pulses would be discussed in Sec. V. # Algorithm design {#DES} Consider an arbitrary time-dependent Hamiltonian in the form of Eq. [\[TD\]](#TD){reference-type="eqref" reference="TD"}, we name \(u_i(t)\) as the \(i\)-th control signal, \(\varepsilon_i\) as the bond on the \(i\)-th control field. Our aim is to tailor \(u_i(t)\) to minimize the following defined objective \(\rm IF\) at time \(t=T\) \[\label{IF} {\rm IF}=1-\frac{1}{2^m}Re{\left\{tr\left[{W}^{\dagger}{U}(T,0)\right]\right\}},\] where \({W}\) is the target operation, \(\rm IF\) is called the \"Infidelity\" in quantum information. Since optimizing the target Eq. [\[IF\]](#IF){reference-type="eqref" reference="IF"} relays highly on the dynamics of the system, an efficient propagation scheme is crucial for quantum optimization. The broadly used approach is to break up the total evolution operator into small increments of duration \(\tau\) in which the variation of the Hamiltonian operator is negligible, for example \[\label{OLD} \hat U(t+\tau,t)=\exp{\Big[-i\tau(H_0+u(t+\frac{\tau}{2}){\hat H}_1)\Big]},\] where the sampling point \(t+\frac{\tau}{2}\) is decided by the Exponential Midpoint Rule that usually guarantees a result with the highest precision. This approach is called the Piecewise Constant Scheme (PWC), which applies to nearly all the existing numerical quantum algorithms, e.g. the basic GRAPE algorithm. However, PWM provides an alternative approach for solving SE, with the same simplicity but higher efficiency. Based on Eq. [\[TID\]](#TID){reference-type="eqref" reference="TID"}, one can construct a corresponding time-independent SE in each interval \(\tau\) and combine the results to calculate the propagator, for example \[\begin{aligned} \hat U_M(t+\tau,t)&=&\exp{\Big[-it_f{\hat H}_0\Big]} \exp{\Big[-it_{p}(H_0\pm &\times&\exp{\Big[-it_f{\hat H}_0\Big]}, \end{aligned}\] where \(t_{p}=\varepsilon^{-1}\int_t^{t+\tau}dt'u(t')\), \(t_{f}={\left(\tau-t_{p}\right)}/2\), subscript \(M\) means it is calculated by PWM transformation. Instead of using the amplitude of \(u_i(t)\) as the decision variables of the system in GRAPE and all the other algorithms, the PWM approach utilizes the time intervals of rectangular pulses \(t_{p,i}^{(k)}\) to be the decision variables. In the following, we derive the gradient formulas for combining PWM with the gradient descent algorithm. ## Gradient formula for one-control case Consider a Hamiltonian with only one control function \[{H}(t)={H}_0+u_x(t){H}_x~|u_x(t)|\leq\varepsilon_x,\] we apply PWM transformation to the Hamiltonian and obtain the corresponding propagator of time \(t=T\) where \[g^{(k)}=\left\{ \begin{array}{ll} +\varepsilon_x, & \hbox{if the \(k\)-th pulse is positive} \\ -\varepsilon_x, & \hbox{otherwise} \end{array} \right.\] is a hidden binary function which indicates whether the \(k\)-th pulse is positive or negative, \(t_{p}^{(k)}\) is the time duration of the \(k\)-th pulses, and \(t_{f}^{(k)}=\left(\tau-t_{p}^{(k)}\right)/2\). Note that the superscript turns out to be either \(H_0\) or \(H_0+g^{(k)}H_x\), in order to save computational resource we define \[\begin{aligned} \label{ONEPRE} H_0&=&D_0\Lambda_0D_0^{\dagger},\\\nonumber H_0+g^{(k)}H_x&=&D_x\Lambda_xD_x^{\dagger}, \end{aligned}\] \[\begin{aligned} \label{ONEV} V_k=D_{x}\exp{\left[-it_p^{(k)}\Lambda_x\right]}D_{x}^{\dagger}, \end{aligned}\] \[\begin{aligned} \label{ONEU} &U_{1}=D_{0}\exp{\left[-it_f^{(1)}\Lambda_0\right]}D_0^{\dagger}\\\nonumber &U_{k+1}=\left\{D_{0}\exp{\left[-i{(t_f^{(k)}+t_f^{(k+1)})}\Lambda_0\right]} D_{0}^{\dagger}V_k\right\}U_{k}, \end{aligned}\] where \(k=1,\cdots,M-1\). Thus, we obtain the formula of the propagator and the corresponding gradient \[\begin{aligned} \label{ONECON} U(T,0)&=&D_0\exp{\left[-it_f^{(k)}\Lambda_0\right]}D_{0}^{\dagger}V_{k}U_{k}\\\nonumber \frac{\partial U(T,0)}{\partial t_p^{(k)}} &=&\frac{-i}{2}U(T,0)\left(\left(V_{k}U_{k}\right)^{\dagger}H_x\left(V_{k}U_{k}\right)+U_k^{\dagger}H_xU_k\right) \end{aligned}\] ## Gradient formula for multi-control case Consider a more general case where the Hamiltonian contains multiple control functions, e.g. the following \(2\)-control case \[{H}(t)={H}_0+u_x(t){H}_x+u_y(t){H}_y,~|u_{x,y}(t)|\leq\varepsilon_{x,y}.\] We define: if \(t_{p,x}^{(k)}-t_{p,y}^{(k)}\geq0\): \[\begin{aligned} &t_{a}^{(k)}=t_{p,x}^{(k)}, t_{b}^{(k)}=t_{p,y}^{(k)}, t_{d}^{(k)}=\frac{t_{p,x}^{(k)}-t_{p,y}^{(k)}}{2}, t_{f}^{(k)}=t_{f,x}^{(k)},\\ &H_{a}^{(k)}=g_{x}^{(k)}H_{x}, H_{b}^{(k)}=g_{y}^{(k)}H_{y}; \end{aligned}\] if \(t_{p,x}^{(k)}-t_{p,y}^{(k)}\leq0\): \[\begin{aligned} &t_{a}^{(k)}=t_{p,y}^{(k)}, t_{b}^{(k)}=t_{p,x}^{(k)}, t_{d}^{(k)}=\frac{t_{p,y}^{(k)}-t_{p,x}^{(k)}}{2}, t_{f}^{(k)}=t_{f,y}^{(k)},\\ &H_{a}^{(k)}=g_{y}^{(k)}H_{y}, H_{b}^{(k)}=g_{x}^{(k)}H_{x}; \end{aligned}\] where \[g_{i}^{(k)}=\left\{ \begin{array}{lll} +\varepsilon_i, & \hbox{if the \(k\)-th pulse of \(i\)-th control is positive}\\ -\varepsilon_i, & \hbox{otherwise} \end{array} \right.\] is a hidden binary function which indicates whether the \(k\)-th pulse of \(i\)-th control is positive or negative, \(t_{p,i}^{(k)}\) for \(i=x,y\) is the time duration of the \(k\)-th pulse of the \(i\)-th control, and \(t_{f,i}^{(k)}=\left(\tau-t_{p,i}^{(k)}\right)/2\). In the similar way of one-control case, we obtain the corresponding propagator of time \(t=T\) Similarly, we define \[\begin{aligned} \label{TWOPRE} H_0&=&D_0\Lambda_0D_0^{\dagger},\\\nonumber H_0+H_a&=&D_a\Lambda_aD_a^{\dagger},\\\nonumber H_0+H_a+H_b&=&D_b\Lambda_bD_b^{\dagger}, \end{aligned}\] \[\begin{aligned} \label{TWOV} V_{k,1}&=&D_{a}\exp{\left[-i{t_{d}^{(k)}}\Lambda_a\right]}D_{a}^{\dagger},\\\nonumber V_{k,2}&=&D_{b}\exp{\left[-i{t_{b}^{(k)}}\Lambda_b\right]}D_{b}^{\dagger}V_{k,1},\\\nonumber V_{k,3}&=&V_{k,1}V_{k,2}; \end{aligned}\] \[\begin{aligned} \label{TWOU} &U_{1}=D_{0}\exp{\left[-i{t_{f}^{(1)}}\Lambda_0\right]}D_{0}^{\dagger},\\\nonumber &U_{k+1}=\left\{ D_0\exp{\left[-i\left(t_{f}^{(k)}+t_{f}^{(k+1)}\right)\Lambda_0\right]}D_{0}^{\dagger} V_{k,3}\right\}U_{k}, \end{aligned}\] where \(k=1,\cdots,M-1\). Thus we obtain the formula of the propagator and the corresponding gradient \[\begin{aligned} \label{TWOCON} U(T,0)&=&D_0\exp{\left[-i{t_{f}^{(k)}}\Lambda_0\right]}D_{0}^{\dagger}V_{m,3}U_m\\\nonumber \frac{\partial U(T,0)}{\partial t_{a}^{(k)}} &=&\frac{-i}{2}U(T,0)\left(\left(V_{k,3}U_{k}\right)^{\dagger}H_a\left(V_{k,3}U_{k}\right) +U_k^{\dagger}H_aU_k\right)\\\nonumber \frac{\partial U(T,0)}{\partial t_{b}^{(k)}} &=&\frac{-i}{2}U(T,0)\Big(\left(V_{k,2}U_{k}\right)^{\dagger}H_b\left(V_{k,2}U_{k}\right)\\\nonumber &&+\left(V_{k,1}U_{k}\right)^{\dagger}H_b\left(V_{k,1}U_{k}\right)\Big) \end{aligned}\] ## Algorithm design {#EFF} Without loss of generality, we present the whole procedure for the algorithm according to the one-control case:\ (0) Diagonalize the Hamiltonian \(H_0\) and \(H_0+g^{(k)}H_x\) using Eq. [\[ONEPRE\]](#ONEPRE){reference-type="eqref" reference="ONEPRE"}, and guess the initial pulse intervals \(t_p^{(k)}\). (1) Calculate the exponents \(V_k\) and \(U_k\) using Eq. [\[ONEV\]](#ONEV){reference-type="eqref" reference="ONEV"} and [\[ONEU\]](#ONEU){reference-type="eqref" reference="ONEU"}, the propagator and the gradient using Eq. [\[ONECON\]](#ONECON){reference-type="eqref" reference="ONECON"}. (2) Evaluate \({\partial U(T,0)}/{\partial t_p^{(k)}}\) and update \(t_p^{(k)}\) by gradient descent method (or other descent methods ). (3) Evaluate the infidelity \(\rm IF\) in Eq. [\[IF\]](#IF){reference-type="eqref" reference="IF"} with these new \(t_p^{(k)}\), repeat step (1)\(\sim\)(3) if the result is not satisfactory. (4) Using anti-PWM transformation or smooth pulses to generate a smooth result.\ Since Eq. [\[ONEPRE\]](#ONEPRE){reference-type="eqref" reference="ONEPRE"} and [\[TWOPRE\]](#TWOPRE){reference-type="eqref" reference="TWOPRE"} allow one to diagonalize and calculate the exponents previously (offline), it transforms the large number of calculations of exponents with matrices on the shoulder Eq. [\[OLD\]](#OLD){reference-type="eqref" reference="OLD"} into exponents with only scalars on the shoulder Eq. [\[ONECON\]](#ONECON){reference-type="eqref" reference="ONECON"} and [\[TWOCON\]](#TWOCON){reference-type="eqref" reference="TWOCON"}, which may simplify the computational complexity and accelerate the simulation of quantum dynamics by a large extent; moreover, Eq. [\[ONEPRE\]](#ONEPRE){reference-type="eqref" reference="ONEPRE"}-[\[ONEU\]](#ONEU){reference-type="eqref" reference="ONEU"} and Eq. [\[TWOPRE\]](#TWOPRE){reference-type="eqref" reference="TWOPRE"}-[\[TWOU\]](#TWOU){reference-type="eqref" reference="TWOU"} can be recycled in Eq. [\[ONECON\]](#ONECON){reference-type="eqref" reference="ONECON"} and [\[TWOCON\]](#TWOCON){reference-type="eqref" reference="TWOCON"} for calculating the time evolution and the corresponding gradient. Thus, we may expected that PWM should be very efficient for solving optimization problems, especially in complex situations. # Optimization Examples {#NUM} Consider the example of D-Norleucine molecule which contains six homonuclear carbon spins that can be encoded as qubits for Nuclear Magnetic Resonance (NMR) quantum computation. The system's Hamiltonian reads \[{H}(t)={H}_0+u_x(t){H}_x+u_y(t){H}_y,~|u_{x,y}(t)|\leq\varepsilon\] where the complete form of \(H_0\) and \(H_{x,y}\) are given in App. [\[HAM\]](#HAM){reference-type="ref" reference="HAM"}, \(u_{x,y}(t)\) are the control functions implemented by radio-frequency magnetic fields, \(\varepsilon=2\rm MHz\) is the bound on the amplitude of control fields. ## 3-qubit example with single control {#THREE} To illustrate how the PWM algorithm works, we choose the first three nuclears \(C_1\), \(C_2\), \(C_3\) to seek an optimal solution. The target operation is to realize an \(\frac{\pi}{2}\)-rotation on \(C_1\) around the \(\langle\sigma_x\rangle\)-axis at time \(T=10 ms\) while remain the rest spins unchanged, saying \[{W}=\exp{\left(-i\frac{\pi}{4}\sigma_x\right)}\otimes I^2,\] where \(I\) is identity. We start from \(10\) randomly chosen initial fields of \(u(0)\), set \(M=1\times10^5\) to transform the Hamiltonian into the PWM form, and apply the gradient optimization algorithm to look for optimal solutions. Since there is no guarantee that the algorithm will converge to the global minimum, we add some random perturbations on the gradient to try to avoid potential trap. Figure [\[E1\]](#E1){reference-type="ref" reference="E1"} shows the corresponding results, where all the \(10\) cases achieve the infidelity requirement \(\rm IF\leq1\times10^{-3}\) within \(210\) iterations. The blue line corresponds to a typical optimization procedure that achieves \(\rm IF=0.84\times10^{-3}\) after \(164\) iterations, where the infidelity could be extremely close to zero under the same condition when we allow more iterations, or when we apply a more sophisticated algorithm (such as the genetic algorithm). This indicates that the PWM algorithm is capable of finding highly accurate optimal solutions. Since a continuous control field is more acceptable than a rectangular one in laboratory, we reverse the PWM transformation procedure to convert the optimized PWM control field back to the continuous field. Figure [\[E2\]](#E2){reference-type="ref" reference="E2"} and [\[E3\]](#E3){reference-type="ref" reference="E3"} respectively show the corresponding continuous control field \(u(t)\) in time domain and frequency domain, which slightly increases the infidelity but still guarantees a rather high control precision \(\rm IF=0.95\times10^{-3}\). To test the performance of the continuous control fields transformed from PWM pulses with different parameter \(M\), we vary \(M\) from \(50\) to \(2.5\times10^{5}\) to optimize the same problem and anti-transformed the optimized PWM pulses into continuous fields. Figure [\[E4\]](#E4){reference-type="ref" reference="E4"} reveals the relation between the performance of the continuous control field and the pulse number \(M\) of the optimization process. We found that the anti-transformed fields could still provide a rather accurate result if the PWM time interval \(\tau\leq1\mu s\) (\(M\geq1\times10^{4}\)) for this problem; however, it is very difficult to obtain a satisfactory continuous field when \(\tau\geq1\mu s\) (\(M\leq1\times10^{4}\)). This can be explained by that EAP works under the condition of small time interval \(\tau\), for it is equivalently using only the \(1\)st-order term of Magnus expansion. Nevertheless, since in the latter case the time interval \(\tau\) for preparing each pulse is relatively long, directly using on and off pulses may be applicable in contemporary laboratory conditions. On this regard, one may think the results for small pulse number \(M\) is still acceptable, which further saves the optimization procedure by a large extent. ## 6-qubit example with 2 controls We consider a larger system with two control functions for manipulating all the six spins in the molecule, which is regarded to be very difficult for seeking optimal solutions. The target operation is to realize the CNOT gate on \(C_1\) and \(C_2\) at time \(T=10s\) while remain the rest spins unchanged, saying \[{W}={U}_{CNOT}\otimes I^4,\] where \(I\) is identity. The time duration may touch the limit of the dephasing time in many NMR systems; nevertheless, it can be achieved in well-prepared samples and in a good experimental apparatus at reasonably high magnetic fields. In this regard, we still focus on seeking optimal control solutions in closed systems. We start from a randomly chosen initial fields of \(u(0)\), set \(M=1\times10^5\) to transform the Hamiltonian into the PWM form, and apply the gradient optimization algorithm to look for optimal solutions. Figure [\[F1\]](#F1){reference-type="ref" reference="F1"} shows the optimization procedure which achieves the infidelity \(\rm IF=3.89\times10^{-4}\) after \(24\) iterations. We further changed other target operations and also met no any trap in the optimization process, which is consistent with the landscape of quantum system from the practical perspective, saying the traps of optimization are usually avoidable when the time duration is very long. On the other hand, the PWM algorithm achieves a very high precision with only \(24\) iterations. It reveals that PWM is highly efficient to handle high dimension, high accuracy, and long time interval problems which are regarded as rather difficult for other algorithms. ## Efficiency analysis PWM speeds up quantum optimal control in accelerating the simulation of quantum dynamical evolution. To quantitatively compare the optimization efficiency of basic GRAPE algorithm and PWM based gradient algorithm, we suppose the iterations before achieving the target are of the same scale for the two algorithms, and compare the CPU time \(t_c\) of the two method for calculating a desired propagator \(U(T,0)\) [^1]. For simplicity, we consider the \(1\)-control case, use the Hamiltonian of \(6\)-qubit example, and set the desired control function \(u(t)=\sin{(\omega t)}\) for \(\omega=50\rm KHz\). Figure [\[CPU\]](#CPU){reference-type="ref" reference="CPU"} displays the relation among the frequency scope \(\Omega\) (which corresponds to pulse number \(M\)), time duration \(t\), and the computational efficiency (CPU time \(t_c\)). Roughly speaking, PWM shows a distinct advantage in saving computational resources, at \(T=20\mu s\), \(\rm IF=1\times10^{-1}\) the PWM algorithm saves \(29.6\%\) computational time for corresponding evolution; at \(T=200\mu s\), \(\rm IF=1\times10^{-8}\) the PWM algorithm saves \(29.7\%\) computational time. Though compared with the low-dimensional results the computation advantage becomes more and more small with the increase of system dimension, PWM still shows a nearly \(30\%\) time saving in the \(6\)-qubit system. Since a quantum optimization problem usually requires solving SE many times, and solving that for high-dimensional problems needs extremely long computation time, the \(30\%\) time saving in fact saves a large quantity of computational resources and makes it possible for optimizing complex systems. # Smooth PWM Pulses {#GAU} Though we have demonstrated that PWM has the ability to efficiently find optimal controls in very complex systems, a smooth control is usually preferable in experimental realization such as superconducting control system or laser control system. Typically, this mission is implemented by introducing a smooth error function into the objective, and re-optimize the control problems; however, the way to smooth PWM pulses is more straight forward, and this approach may be also applied to smoothing results from other algorithms. The first approach is using anti-PWM transformation, which replaces rectangular pulse by a continuous function with the same integral area in each time interval \(\tau\). This approach also meets the conclusion in, where the author found \"the target fidelity does not depend on the shape of the pulse\" for the subspace-selective self-adaptive differential evolution algorithm (SuSSADE). Since the number of control parameters are equal, one may smooth the piecewise constant control field by using EAP and obtain the same continuous field, rather than sacrifice the efficiency by introducing the error function. In fact, as we demonstrated in Sec. II the anti-transformation does not always work when one sets a relatively large \(\tau\). Here, we propose another approach for smoothing PWM pulses; that is, we transform the rectangular pulses into continuous pulses, e.g. Gaussian pulses. Detail discussion of this can be found in; in short, the rectangular pulses can be well approximated by a sequence of Gaussian pulses \(\exp{\Big[-{\pi t^2}/{t_{p}^2}\Big]}\) with the same value of \(t_p\). Figure [\[G1\]](#G1){reference-type="ref" reference="G1"} shows the relation between the rectangular pulses and the corresponding Gaussian pulses in time domain, Fig. [\[G2\]](#G2){reference-type="ref" reference="G2"} reveals the properties of the twos in frequency domain. It demonstrates the Gaussian pulses can well approximate the PWM pulses within a large frequency band. However, though the Gaussian pulse works for a relatively larger \(\tau\), it does not always work also. How to effectively smooth PWM pulses is still an open problem for further discussions. # Conclusions {#CON} Combining Pulse Width Modulation (PWM) and gradient algorithm, we applied PWM into quantum optimal control problems. Compared with the basic GRAPE algorithm, the our algorithm raised the efficiency by a large extant, which provides us a highly efficient way for solving various optimization problems. In fact, PWM can also be imbedded in other descendent algorithms, such as Newton algorithm, quasi-Newton algorithm, and various evolutionary algorithms. In addition, since we have demonstrated in that the rectangular pulses generated by PWM are rather robust with noise and can also be implemented by other functions (e.g. the Gaussian pulse train), it indicates the result of PWM algorithm may be directly applied into experiments. Thus, the PWM algorithm may highly simplify the current experiment relying on subtle control of shaped pulses, and provide a very robust and reliable quantum optimal control design.

{'timestamp': '2016-07-19T02:02:25', 'yymm': '1607', 'arxiv_id': '1607.04694', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04694'}

# Introduction {#SEC:intro} Let \(p\) be a prime and let \(G\) be a torsion-free abelian group. The subject of this work is the group \(Hom (G,\mathbb{Z}_p)\) of \(p\)-adic functionals on \(G\). This group is viewed as a \(p\)-adic dual of \(G\), and is studied using the techniques of functional analysis. This is a novel approach that adds clarity to a difficult area of study. To emphasize this approach, the notation \(G^{*p}\) is used instead of \(Hom (G,\mathbb{Z}_p)\). It is shown that an analogue of the Hahn--Banach Theorem (Proposition [\[PROP:all-states-extend\]](#PROP:all-states-extend){reference-type="ref" reference="PROP:all-states-extend"}) holds in this setting; broadly speaking, this means that \(G^{*p}\) contains enough functionals to be interesting. There is a natural homomorphism of \(G\) into its \(p\)-adic double dual, \(G^{*p*p}\), and one of the main results of this work is that the image of \(G\) is dense in the relative topology that the double dual inherits as a subspace of \(\mathbb{Z}_p^{G^{*p}}\) (Theorem [\[THM:density\]](#THM:density){reference-type="ref" reference="THM:density"}). As a submodule of the free \(\mathbb{Z}_p\)-module \(\mathbb{Z}_p^{G^{*p}}\), \(G^{*p*p}\) is also a free \(\mathbb{Z}_p\)-module. Thus, in particular, it is a pro-\(p\) group, and the pro-\(p\) topology agrees with the relative topology. There is another pro-\(p\) group, called the pro-\(p\) completion of \(G\), that shares this same property of having a homomorphism from \(G\) into it, the image of which is dense. Moreover, the pro-\(p\) completion satisfies a universal property, which implies that there exists a continuous homomorphism from the pro-\(p\) completion to the \(p\)-adic double dual. The main result of Section [4](#SEC:completions){reference-type="ref" reference="SEC:completions"} is Theorem [\[THM:pro-p\]](#THM:pro-p){reference-type="ref" reference="THM:pro-p"}, which states that this homomorphism has a continuous inverse, so the pro-\(p\) completion and the \(p\)-adic double dual of \(G\) are continuously isomorphic (in a way that preserves the image of \(G\) in each of them). The \(\mathbb{Z}_p\)-modules \(G^{*p}\), as \(p\) ranges over the set of all primes, give enough information to reconstruct the isomorphism class of the torsion-free abelian group \(G\). Section [5](#SEC:classification){reference-type="ref" reference="SEC:classification"} describes how this is done in the finite-rank case; in particular, if \(G\) has finite rank, then Theorem [\[THM:classification-3\]](#THM:classification-3){reference-type="ref" reference="THM:classification-3"} provides an isomorphism invariant for \(G\). (For a torsion-free abelian group \(G\), let us use the term *rank* to refer to the torsion-free rank of \(G\), that is, the size of any maximal independent subset.) This invariant is a matrix description of \(G\) that is related to others due to Kurosch and Malcev from 1937 and 1938 respectively (see and, or for a more modern description of the latter, with applications to developments since then). ## History {#SUBSEC:history} This work contributes to the classification and description of torsion-free abelian groups, which is an area with a long history---an excellent overview of the theory is given in. In 1937, Baer classified the torsion-free abelian groups of rank one using an invariant called the *type* of the group, which is an equivalence class of supernatural numbers. At the same time, Kurosch and Malcev developed matrix descriptions of a finite-rank torsion-free abelian group. Definitions [\[DEF:group-determined\]](#DEF:group-determined){reference-type="ref" reference="DEF:group-determined"} and [\[DEF:factored\]](#DEF:factored){reference-type="ref" reference="DEF:factored"} and Theorem [\[THM:classification-3\]](#THM:classification-3){reference-type="ref" reference="THM:classification-3"} from Section [5](#SEC:classification){reference-type="ref" reference="SEC:classification"} provide a matrix description that is related to these. The results here are most closely related to work of Warfield from 1968, in which, for a rank-\(n\) group \(M\) and a rank-\(1\) group \(G\), Warfield considers the dual \(Hom (M,G)\) and the double dual \(Hom(\) \(Hom (\) \(M,G),G)\), and gives necessary and sufficient conditions for the natural homomorphism from \(M\) to the double dual to be an isomorphism. The functional analysis approach that is used here is modelled after the theory of functional analysis on partially ordered torsion-free abelian groups, an excellent reference for which is the book. In particular, the \(p\)-adic Hahn--Banach Theorem that is proved here is modelled after Lemma 4.1 and Proposition 4.2 of (originally from ). The matrix description of finite-rank torsion-free abelian groups that appears in Section [5](#SEC:classification){reference-type="ref" reference="SEC:classification"} is one of the main results of this work (although by no means is it the only one). Given that it is closely related to the matrix descriptions of Kurosch and Malcev, it would be natural to ask: What is new in this work that is not already present in the works of those authors? In response to this question, I would point to the epigraph that precedes this introduction. This epigraph is a quotation from Kurosch, in which he acknowledges that his matrices arise from a difficult sequence of calculations, and calls for a better theoretical understanding of the relationship between torsion-free abelian groups and \(p\)-adic integers. The theory of \(p\)-adic functionals that is introduced here subsumes most of the difficult calculations, making them invisible, and so represents a clear answer to Kurosch's entreaty. There is the added bonus that the matrices in the present work are simpler than Kurosch's or Malcev's in that they have fewer rows. Furthermore Theorem [\[THM:pro-p\]](#THM:pro-p){reference-type="ref" reference="THM:pro-p"}, about the pro-\(p\) completion of a torsion-free abelian group, is another consequence of this theory, and to my knowledge this result is entirely new. ## Notation {#SUBSEC:notation} Let \(p\) be a prime. \(\mathbb{Z}_p\) denotes the ring of \(p\)-adic integers and \(\mathbb{Q}_p\) denotes its field of fractions. The \(p\)-adic absolute value of an element \(\alpha\) of \(\mathbb{Q}_p\) or \(\mathbb{Z}_p\) is denoted by \(|\alpha |_p\). If \(n\in\mathbb{N}\) and \(R\) is a ring, then \(R^n\) denotes the module consisting of \(n\times 1\) columns with entries in \(R\), while \((R^n)^*\) denotes its dual consisting of \(1\times n\) rows. If \(m,n\in\mathbb{N}\), then \(M_{m\times n}(R)\) denotes the module of \(m\times n\) matrices over \(R\). For \(r\in R^n\), \(r^t\) denotes its transpose in \((R^n)^*\). For \(A\in M_{m\times n}(R)\), \(A^t\) denotes its transpose in \(M_{n\times m}(R)\). For \(v = (v_1,\ldots, v_n)^t \in \mathbb{Z}_p^n\), the norm \(\| v\|_p\) is defined by \(\| v\|_p := \max_{1\leq i\leq n} \{ \| v_i\|_p\}\). Let \(G\) be a torsion-free abelian group and \(S\) a subset of \(G\). Then \(\langle S\rangle_*\) denotes the pure subgroup of \(G\) generated by \(S\); that is, \(\langle S\rangle_* := \{ g\in G: ng \in \langle S\rangle\) for some \(n\in\mathbb{N}\}\). ## Outline {#SUBSEC:outline} Section [2](#SEC:examples){reference-type="ref" reference="SEC:examples"} contains examples of torsion-free abelian groups with calculations of their \(p\)-adic functionals. Section [3](#SEC:states){reference-type="ref" reference="SEC:states"} is the biggest section, and introduces the notation and terminology of \(p\)-adic functionals. It includes a discussion of the \(p\)-adic topology on a torsion-free abelian group \(G\) (Section [3.1](#SUBSEC:pseudometric){reference-type="ref" reference="SUBSEC:pseudometric"}); the statement and proof of an analogue of the Hahn--Banach Theorem (Proposition [\[PROP:all-states-extend\]](#PROP:all-states-extend){reference-type="ref" reference="PROP:all-states-extend"} in Section [3.2](#SUBSEC:functionals){reference-type="ref" reference="SUBSEC:functionals"}); results about the dual space of all \(p\)-adic functionals on \(G\) (Section [3.3](#SUBSEC:state-spaces){reference-type="ref" reference="SUBSEC:state-spaces"}); and results about double dual spaces, including the fact that the natural homomorphism from \(G\) to its double dual has dense image (Theorem [\[THM:density\]](#THM:density){reference-type="ref" reference="THM:density"} in Section [3.4](#SUBSEC:double-duals){reference-type="ref" reference="SUBSEC:double-duals"}). Section [4](#SEC:completions){reference-type="ref" reference="SEC:completions"} contains a proof that the \(p\)-adic double dual of a torsion-free abelian group \(G\) can be naturally identified with the pro-\(p\) completion of \(G\) (Theorem [\[THM:pro-p\]](#THM:pro-p){reference-type="ref" reference="THM:pro-p"}). Section [5](#SEC:classification){reference-type="ref" reference="SEC:classification"} uses the theory of \(p\)-adic functionals to classify finite-rank torsion-free abelian groups up to isomorphism; this is the content of Theorem [\[THM:classification-3\]](#THM:classification-3){reference-type="ref" reference="THM:classification-3"}. The classifying invariant is a sequence of matrices that is related to Malcev's matrices, and is subject to the same equivalence relation as Malcev's matrices, although these matrices have fewer rows. # Examples {#SEC:examples} This section contains two examples to illustrate the theory. The groups in both of these examples are inductive limits of stationary inductive sequences of finitely-generated free abelian groups---see for previous results about these groups, including a classification up to quasi-isomorphism. There has already been work done on the real-valued functionals on *ordered* abelian groups that are inductive limits of stationary inductive sequences of simplicial groups (see and ). The \(p\)-adic theory is similar to the theory in the ordered case: in particular, \(p\)-adic functionals come from rows that are combinations of certain left eigenvectors of a matrix representation of the homomorphism from the stationary inductive sequence. The group in the next example admits two independent \(3\)-adic functionals. # \(p\)-adic functionals {#SEC:states} The goal of this section is to study the group \(Hom (G,\mathbb{Z}_p)\), of homomorphisms from a torsion-free abelian group \(G\) into \(\mathbb{Z}_p\), from a functional analysis point of view. To emphasize this point of view, these homomorphisms will be referred to as *\(p\)-adic functionals* on \(G\). The starting point of this work is Proposition [\[PROP:all-states-extend\]](#PROP:all-states-extend){reference-type="ref" reference="PROP:all-states-extend"}, which is an analogue of the Hahn--Banach Theorem for the setting of torsion-free abelian groups. Proposition [\[PROP:all-states-extend\]](#PROP:all-states-extend){reference-type="ref" reference="PROP:all-states-extend"} says that if \(f\) is a \(\mathbb{Z}_p\)-valued group homomorphism on a subgroup \(H\) of \(G\) and \(f\) is *\(t\)-contractive* for some *\(p\)-adic sublinear pseudometric* \(t\) (Definition [\[DEF:t-contractive\]](#DEF:t-contractive){reference-type="ref" reference="DEF:t-contractive"}), then \(f\) extends to a \(p\)-adic functional on \(G\) that is also \(t\)-contractive. This allows us to establish the existence of \(p\)-adic functionals on \(G\) with desired properties by first finding functionals with those properties defined on some subgroup and then extending them to all of \(G\). A consequence of this theory is that the group of all \(\mathbb{Z}_p\)-valued homomorphisms on \(G\) (which will be denoted by \(G^{*p}\) to emphasize the functional analysis point of view) contains enough functionals to be interesting. Section [3.3](#SUBSEC:state-spaces){reference-type="ref" reference="SUBSEC:state-spaces"} deals with this group, both as a topological space and as a torsion-free abelian group. The elements of \(G\) are themselves \(p\)-adic functionals on \(G^{*p}\), and there is a homomorphism from \(G\) into the double dual group, which consists of \(p\)-adic functionals on the group of \(p\)-adic functionals on \(G\). It will be shown in Section [4](#SEC:completions){reference-type="ref" reference="SEC:completions"} that this double dual group \(G^{*p*p}\) coincides exactly with the pro-\(p\) completion of \(G\). This result provides a means of describing the pro-\(p\) completion of \(G\), which is interesting in itself. \(p\)-adic functionals are a \(\mathbb{Z}_p\)-submodule of the set of \(\mathbb{Q}_p\)-valued group homomorphisms on \(G\) that are contractive with respect to a certain pseudometric; let us describe this pseudometric now. ## The \(p\)-adic pseudometric {#SUBSEC:pseudometric} If \(g\) is an element of a torsion-free abelian group \(G\) and \(n\in\mathbb{N}\), let us say that \(g\) is *divisible by \(n\)* or *\(n\)-divisible in \(G\)* if there exists \(h\in G\) such that \(nh = g\). Let us say that \(G\) is \(n\)-divisible if each element of \(G\) is \(n\)-divisible. The reader is invited to verify that \(d_p\) satisfies the definition of a pseudometric; indeed, it satisfies the strong triangle inequality: \(d_p(g,h) \leq \max \{ d_p(g,z),d_p(z,h)\}\), where \(g,h,\) and \(z\in G\). \(d_p\) also satisfies an absolute scalability condition: if \(g,h\in G\) and \(n\in\mathbb{Z}\), then \(d_p(ng,nh) = |n|_pd_p(g,h)\). \(d_p\) is a metric if and only if \(0 \in G\) is the only element that is divisible by arbitrarily high powers of \(p\); in this case, let us refer to it as the *\(p\)-adic metric*. \(d_p\) is bounded: \(d_p(g,h)\leq 1\) for all \(g,h\in G\). The topology arising from the \(p\)-adic pseudometric is called the \(p\)-adic topology in. Let us use the symbol \(G_p\) to denote the subgroup of \(G\) containing all elements that are divisible by arbitrarily high powers of \(p\). It is not difficult to verify that \(G_p\) is a pure subgroup. Given a subgroup \(H\) of \(G\), let \(\overline{H}^p\) denote the closure of \(H\) in the topology arising from \(d_p\). Note that \(\overline{\{ 0\}}^p = G_p\), so \(\overline{H}^p\) contains \(G_p\) for all subgroups \(H\). The following lemma will be useful later, when proving Lemma [\[LEM:quotient\]](#LEM:quotient){reference-type="ref" reference="LEM:quotient"}, which is used to prove Theorem [\[THM:pro-p\]](#THM:pro-p){reference-type="ref" reference="THM:pro-p"}, one of the main theorems. It says that, if \(H\) is a subgroup of \(G\) such that any element of \(G/H\) has order dividing \(p^m\), then any element \(g\in G\) that is not in \(H\) must lie at a distance of at least \(p^{-m}\) from \(H\). In particular, this implies that if \([G:H] = p^m\), then \(H = \overline{H}^p\). The following lemma establishes that the closure of a pure subgroup is pure. Closed pure subgroups will play a role in what follows, so it will be useful to give a name to such subgroups. The group \(G\) in Example [\[EX:favourite\]](#EX:favourite){reference-type="ref" reference="EX:favourite"} is \(3\)-simple: it has many proper non-trivial pure subgroups, each of which is isomorphic to \(\mathbb{Z}\), but the closure of any of them in the metric \(d_3\) is all of \(G\). Given a family \((J_\lambda)_{\lambda\in\Lambda}\) of \(p\)-ideals of \(G\), the intersection \(\bigcap_{\lambda\in\Lambda}J_\lambda\) is again a \(p\)-ideal. The sum of the \(J_\lambda\)s is not necessarily pure or closed, but \(\overline{\langle \bigcup_{\lambda\in\Lambda}J_\lambda \rangle_*}^p\) is a \(p\)-ideal, and is clearly the smallest one that contains each \(J_\lambda\). Thus the set of \(p\)-ideals of \(G\) forms a complete lattice where the partial order is given by inclusion. The minimal \(p\)-ideals---namely, the elements of this lattice that have nothing except \(G_p\) below them---are of particular interest. Lemma [\[LEM:singly-generated-ideal\]](#LEM:singly-generated-ideal){reference-type="ref" reference="LEM:singly-generated-ideal"}, below, establishes that a singly-generated \(p\)-ideal is minimal. The proof of this fact uses the following lemma, which will also be used later to show that a \(p\)-ideal can be separated from an element not in its closure using a \(p\)-adic functional (Corollary [\[COR:separating-state\]](#COR:separating-state){reference-type="ref" reference="COR:separating-state"}). An immediate consequence of Lemma [\[LEM:singly-generated-ideal\]](#LEM:singly-generated-ideal){reference-type="ref" reference="LEM:singly-generated-ideal"} is the following. ## A Hahn--Banach Theorem for \(p\)-adic functionals {#SUBSEC:functionals} Let \(p\) be a prime, let \(G\) be a torsion-free abelian group with \(p\)-adic pseudometric \(d_p\), and let \(f:G\to \mathbb{Z}_p\) be a group homomorphism (i.e., a \(p\)-adic functional). Then \(f\) is automatically contractive (and hence continuous) with respect to the \(p\)-adic pseudometric \(d_p\) on \(G\), meaning that \(|f(g_1)-f(g_2)|_p\leq d_p(g_1,g_2)\) for all \(g_1,g_2\in G\). To see this, it is enough to observe that, if \(g\in G\) is \(p^n\)-divisible---say \(g = p^nh\) for some \(h\in G\)---then \(|f(g)|_p = |p^nf(h)|_p = p^{-n}|f(h)|_p\leq p^{-n}\), as \(|f(h)|_p\leq 1\). If instead \(f\) is only defined on a subgroup \(H\) of \(G\), then the above argument shows that, in order for there to exist an extension of \(f\) to all of \(G\), it is necessary that \(f\) be contractive with respect to \(d_p\). It turns out that this condition is also sufficient (Proposition [\[PROP:all-states-extend\]](#PROP:all-states-extend){reference-type="ref" reference="PROP:all-states-extend"}). But we can say more: if \(f: H\to \mathbb{Z}_p\) is contractive with respect to some *\(p\)-adic sublinear pseudometric* \(t\) (Definition [\[DEF:t-contractive\]](#DEF:t-contractive){reference-type="ref" reference="DEF:t-contractive"}), then \(f\) extends to a \(p\)-adic functional on \(G\) that is also contractive with respect to \(t\). Note in particular that \(d_p\) itself is a \(p\)-adic sublinear pseudometric, and, if \(t\) is another \(p\)-adic sublinear pseudometric, then any \(t\)-contractive functional is also a \(d_p\)-contractive functional. Most of the results in this section are statements about the existence of \(p\)-adic functionals; these functionals are extensions of \(t\)-contractive functionals for some \(p\)-adic sublinear pseudometric \(t\). For most of the proofs, \(t\) can be taken to be \(d_p\); however, Proposition [\[PROP:separate-from-subgroup\]](#PROP:separate-from-subgroup){reference-type="ref" reference="PROP:separate-from-subgroup"} establishes the existence of \(p\)-adic functionals with very specific properties, and for this it will be necessary to use a more restrictive \(t\) to get sharper information. Note that, if \(H\) is a subgroup of \(G\) and \(d_p\) is the \(p\)-adic pseudometric on \(G\), then the \(p\)-adic pseudometric on \(H\) is bounded below by the restriction of \(d_p\) to \(H\), and this bound may be strict. Thus to say that \(f\) is a \(d_p\)-contractive functional on \(H\) is stronger than saying that \(f\) is a \(p\)-adic functional on \(H\). Of course, if \(H\) is a pure subgroup of \(G\), then the \(p\)-adic pseudometric on \(H\) coincides with the restriction to \(H\) of that on \(G\). It is clear that the trivial homomorphism is a \(p\)-adic functional, and, if \(G\) is \(p\)-divisible, then this is the only one. It is natural to ask whether there exist any non-trivial \(p\)-adic functionals on \(G\) if \(G\) is not \(p\)-divisible. Corollary [\[COR:states-exist\]](#COR:states-exist){reference-type="ref" reference="COR:states-exist"} shows that, if \(G\) is not \(p\)-divisible, then non-zero \(p\)-adic functionals on \(G\) do exist. The following is the key lemma that makes everything else work. Lemma [\[LEM:extend\]](#LEM:extend){reference-type="ref" reference="LEM:extend"} allows us to extend a \(t\)-contractive functional on a subgroup of \(G\) to one on a larger subgroup, but to show that non-trivial \(p\)-adic functionals exist, we need to start with a non-trivial \(t\)-contractive functional on some subgroup of \(G\). The natural place to begin is with a cyclic subgroup. The following corollary, which deals with functionals on cyclic subgroups of \(G\), follows from Lemma [\[LEM:extend\]](#LEM:extend){reference-type="ref" reference="LEM:extend"} by taking \(H = \{ 0\}\) and \(t = d_p\). The next proposition is an analogue of the Hahn--Banach Theorem for the setting of torsion-free abelian groups. Proposition [\[PROP:all-states-extend\]](#PROP:all-states-extend){reference-type="ref" reference="PROP:all-states-extend"} can be combined with Corollary [\[COR:rank-one-seed\]](#COR:rank-one-seed){reference-type="ref" reference="COR:rank-one-seed"} and Lemma [\[LEM:closure\]](#LEM:closure){reference-type="ref" reference="LEM:closure"} to produce the following corollaries, which say that any \(g\in G\) realizes its maximum possible value on some functional, and \(p\)-adic functionals can separate a pure subgroup of \(G\) from an element \(g\) not in its closure. The following proposition will be useful in Section [4](#SEC:completions){reference-type="ref" reference="SEC:completions"}, when we consider finite-index subgroups of a torsion-free abelian group \(G\). ## Spaces of functionals {#SUBSEC:state-spaces} \(G^{*p}\) is a torsion-free abelian group, and as it is a subgroup of \(\mathbb{Z}_p^G\), it contains no non-zero elements that are divisible by arbitrarily high powers of \(p\), so the \(p\)-adic pseudometric on \(G^{*p}\) is in fact a metric. In what follows, it will be useful to view \(G^{*p}\) as a topological space. It turns out that the most useful topology is not the metric topology arising from the \(p\)-adic metric on \(G^{*p}\), but rather the topology that \(G^{*p}\) inherits from \(\mathbb{Z}_p^{G}\). \(\mathbb{Z}_p^G\) is a product of compact spaces, and hence is compact by Tychonoff's Theorem. It is easy to verify that \(G^{*p}\) is closed in the relative topology, and hence is compact. \(\mathbb{Z}_p\) is a principal ideal domain, and \(G^{*p}\) is a submodule of the \(\mathbb{Z}_p\)-module \(\mathbb{Z}_p^G\), which is free, so \(G^{*p}\) is free as well. Thus is makes sense to speak of a \(\mathbb{Z}_p\)-basis and the \(\mathbb{Z}_p\)-rank of \(G^{*p}\). The metric topology on \(G^{*p}\) is finer than the relative topology in general, and if the \(\mathbb{Z}_p\)-rank of \(G^{*p}\) is finite, then the two topologies agree, but this need not be true if \(G^{*p}\) has infinite rank, as the following example illustrates. Note that any \(p\)-adic functional on \(G\) is entirely determined by its values on a maximal independent subset of \(G\), which means that, if \(G\) has finite rank, then the \(\mathbb{Z}_p\)-rank of \(G^{*p}\) cannot exceed the rank of \(G\)---more precisely, it cannot exceed the rank of \(G\) minus the rank of \(G_p\). Proposition [\[PROP:simple\]](#PROP:simple){reference-type="ref" reference="PROP:simple"}, below, says that \(G\) is \(p\)-simple but not \(p\)-divisible if and only if the \(\mathbb{Z}_p\)-rank of \(G^{*p}\) is \(1\). The proof of this fact uses the following easy lemma. The next lemma characterizes the \(p\)-adic metric on \(G^{*p}\) in terms of the absolute values of the images of elements of \(G\). The following corollary is an immediate consequence of Lemma [\[LEM:state-space-metric\]](#LEM:state-space-metric){reference-type="ref" reference="LEM:state-space-metric"}. The goal of the next sequence of results is to characterize the subsets of \(G^{*p}\) that are sufficiently rich to contain, for each \(g\in G\), a \(p\)-adic functional that achieves the maximum possible absolute value on \(g\). This is the content of Proposition [\[PROP:spanning-states\]](#PROP:spanning-states){reference-type="ref" reference="PROP:spanning-states"}, which says that this is occurs for subsets of \(G^{*p}\) that have \(\mathbb{Z}_p\)-spans that are dense in \(G^{*p}\) in the relative topology. The proof of this result uses Lemma [\[LEM:independence\]](#LEM:independence){reference-type="ref" reference="LEM:independence"}, which in turn uses the following technical lemma. The next lemma says that a functional \(f\in G^{*p}\) can be separated from a subset \(F\subseteq G^{*p}\) by an element of \(G\) if and only if it is not in the relative-topology closure of the span of \(F\). The next result describes the subsets of \(G^{*p}\), the relative-topology closures of the the spans of which are all of \(G^{*p}\). ## Double duals {#SUBSEC:double-duals} If \(p\) is a prime and \(G\) is a torsion-free abelian group, then the dual space \(G^{*p}\) of \(p\)-adic functionals is also a torsion-free abelian group, and so has its own dual space \(G^{*p*p}\). There is a natural homomorphism \(\Phi^p\) from \(G\) to \(G^{*p*p}\) (Definition [\[DEF:phi-p\]](#DEF:phi-p){reference-type="ref" reference="DEF:phi-p"}, below), the image of which is dense in \(G^{*p*p}\) (Theorem [\[THM:density\]](#THM:density){reference-type="ref" reference="THM:density"}, below). \(\hat{g}^p\) is clearly a group homomorphism into \(\mathbb{Z}_p\), so is an element of \(G^{*p*p}\). Note that \(\Phi^p\) is a group homomorphism: \(\widehat{g+h}^p(f) = f(g+h) = f(g) + f(h) = \hat{g}^p(f) + \hat{h}^p(f)\). Its kernel certainly contains \(G_p\), and, by Corollary [\[COR:state-hits-distance\]](#COR:state-hits-distance){reference-type="ref" reference="COR:state-hits-distance"}, \(G_p\) is the entire kernel. Let \(\hat{G}^p\) denote the image of \(\Phi^p\): \(\hat{G}^p := \{ \hat{g}^p: g\in G\}\subset G^{*p*p}\). The main result of this section is the following. # Pro-\(p\) completions {#SEC:completions} Readers who are familiar with profinite groups might have noticed that the double dual \(G^{*p*p}\) of a torsion-free abelian group \(G\) shares important features with the pro-\(p\) completion of \(G\). This is not a coincidence---the two objects are, in fact, the same. The topology in Definition [\[DEF:pro-p\]](#DEF:pro-p){reference-type="ref" reference="DEF:pro-p"} may require some explanation. Any inverse limit of an inverse system of groups has a standard realization as a subgroup of the direct product: \[\varprojlim (G_\lambda)_{\lambda\in\Lambda} = \{ (x_\lambda)_{\lambda\in\Lambda} \in \prod_{\lambda\in\Lambda} G_\lambda: \phi_{\mu\lambda}(x_\lambda) = x_\mu \text{ for all } \lambda\leq \mu\},\] where, for \(\lambda\leq \mu\), \(\phi_{\mu\lambda}: G_\lambda\to G_\mu\) is the corresponding map in the inverse system. In the pro-\(p\) case, the groups \(G_\lambda\) in the direct product are finite \(p\)-groups, and hence are topological groups with the discrete topology. The topology on the inverse limit is the relative topology that it inherits from the direct product. Given an arbitrary group \(G\), consider the set of all normal subgroups \(N\) of \(G\) for which \(G/N\) is a finite \(p\)-group. This is a partially ordered set under inclusion, which means that there is an inverse system of natural homomorphisms \(\phi_{\mu\lambda}: G/N_\lambda \to G/N_\mu\) between the quotients, where \(N_\mu\leq N_\lambda\). The *pro-\(p\) completion* \(\mathcal{K}_p(G)\) of \(G\) is the inverse limit of this inverse system. There is a natural embedding \(\Psi^p: G\to \mathcal{K}_p(G)\), which sends \(g\in G\) to \((gN_\lambda)_{\lambda\in\Lambda}\), and the image of \(G\) under this embedding is a dense subgroup. \(\mathcal{K}_p(G)\) is clearly a pro-\(p\) group. Moreover it satisfies the following universal property with respect to \(G\): if there exists a pro-\(p\) group \(H\) and a group homomorphism \(\theta: G\to H\), then there exists a continuous homomorphism \(\eta: \mathcal{K}_p(G)\to H\) such that \(\theta = \eta\circ\Psi^p\). If \(G\) is a torsion-free abelian group, then \(G^{*p*p}\) is a free \(\mathbb{Z}_p\)-module, and hence is a pro-\(p\) group (and the pro-\(p\) topology agrees with the relative topology on \(G^{*p*p}\) ). Thus the universal property guarantees the existence of a continuous homomorphism \(\eta: \mathcal{K}_p(G)\to G^{*p*p}\) such that \(\eta(\Psi^p(g)) = \hat{g}^p\) for all \(g\in G\). Then in order to show that \(G^{*p*p}\) is the pro-\(p\) completion, it will be necessary to find a continuous homomorphism in the other direction. Let us do so by finding a compatible system of homomorphisms from \(G^{*p*p}\) into all of the finite \(p\)-group quotients of \(G\), and then use the universal property of the inverse limit to construct the required homomorphism. # Classification {#SEC:classification} The theory of \(p\)-adic functionals can be used to classify finite-rank torsion-free abelian groups. Specifically, the collection of homomorphisms \(\Phi^p: G\to G^{*p*p}\) for a torsion-free abelian group \(G\), as \(p\) ranges over the set of all primes, gives enough information to recover the isomorphism class of \(G\). Let \(P\) denote the set consisting of all integer primes and let \(G\) be a torsion-free abelian group. Pick a maximal independent subset \(\mathcal{B}\) of \(G\). For an element \((\alpha_\lambda)_{\lambda\in\Lambda} \in \mathbb{Z}_p^{\Lambda}\), define the sup-norm \(\| (\alpha_\lambda)_{\lambda\in\Lambda} \|_p = \sup_{\lambda\in\Lambda} \{ |\alpha_\lambda|_p\}\). It is easy to verify that \(\|\cdot \|_p\) is indeed a norm, and that it satisfies the strong triangle inequality. \(G^{*p*p}\) is isomorphic to \(\mathbb{Z}_p^{\Lambda_p}\) for some index set \(\Lambda_p\); if we identify \(G^{*p*p}\) with this latter group, then the homomorphism \(\Phi^p\) sends elements of \(G\) to \(\Lambda_p\)-tuples in \(\mathbb{Z}_p^{\Lambda_p}\). Moreover, by Theorem [\[THM:density\]](#THM:density){reference-type="ref" reference="THM:density"}, \(\Phi^p\) is a pseudo-isometry, meaning that \(d_p(g,0) = \| \Phi^p(g)\|_p\). This is a powerful fact: it implies that \(G\) is entirely determined by the values \((\Phi^p(g))_{p\in P}\) as \(g\) ranges over the elements of \(\mathcal{B}\). Indeed, \(G\) is isomorphic to a subgroup of \(\langle\mathcal{B}\rangle\otimes \mathbb{Q}\), so in order to specify \(G\) it is enough to be able to say, for each element \(g = a_1g_1+\cdots +a_kg_k\in \langle\mathcal{B}\rangle\subseteq G\) and \(m\in\mathbb{N}\), whether or not \(g\) is in \(mG\). But \(g\in mG\) if and only if \(|m|_p \geq d_p(g,0) = \| \Phi^p(g)\|_p = \| a_1\Phi^p(g_1) + \cdots + a_k\Phi^p(g_k)\|_p\) for all \(p\in P\). This is summarized in the following proposition. In the case that \(G\) has finite rank, the data that specify \(G\) can be stated more explicitly in terms of the following definition. Thus the finite-rank torsion-free abelian group \(G\) is completely described by a sequence, indexed by the prime numbers, of matrices \(A_p\) with entries in \(\mathbb{Z}_p\). The rows of the \(p\)th matrix correspond to elements of a basis for \(G^{*p}\). This fact places a restriction on the structure of \(A_p\); this restriction is expressed in terms of the following definition. Then, by Theorem [\[THM:density\]](#THM:density){reference-type="ref" reference="THM:density"}, the matrices \(A_p\) in Proposition [\[PROP:classification-2\]](#PROP:classification-2){reference-type="ref" reference="PROP:classification-2"} have the property that the column span \(C(A_p)\) of \(A_p\) is dense in \(\mathbb{Z}_p^{n_p}\). This implies in particular that \(n_p\leq n\) for all primes \(p\). It is possible to characterize the homomorphisms between factored forms. We are now in a position to state Theorem [\[THM:classification-3\]](#THM:classification-3){reference-type="ref" reference="THM:classification-3"}, the main classification result. It says that every finite-rank torsion-free abelian group has a factored form, and gives a characterization of isomorphisms between factored forms. The first statement is a consequence of Proposition [\[PROP:classification-2\]](#PROP:classification-2){reference-type="ref" reference="PROP:classification-2"} and the discussion following it, while the second statement is a consequence of Proposition [\[PROP:factored-homomorphisms\]](#PROP:factored-homomorphisms){reference-type="ref" reference="PROP:factored-homomorphisms"}. The matrix \(V\) in Theorem [\[THM:classification-3\]](#THM:classification-3){reference-type="ref" reference="THM:classification-3"} should be thought of as a change of basis matrix that represents a new maximal independent subset in terms of the old one. The matrices \(U_p\) should be thought of as change of basis matrices that represent a new basis of the \(p\)-adic dual group in terms of the old basis.

{'timestamp': '2016-08-10T02:06:59', 'yymm': '1607', 'arxiv_id': '1607.04586', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04586'}

# Introduction {#sec1} Let \(\mathbb{Z}\) be a finite dimensional vector space, and \(\mathbb{R}^{m\times n}\) be the vector space of all \(m\!\times\!n\) real matrices equipped with the trace inner product \(\langle \cdot,\cdot\rangle\) and its induced norm \(\|\cdot\|_F\). Denote by \(\|X\|_*\) the nuclear norm of a matrix \(X\in\mathbb{R}^{m\times n}\), i.e., the sum of all singular values of \(X\), and by \(\|X\|\) the spectral norm of \(X\). Consider the following optimization problem \[\label{prob} \min_{z\in\mathbb{Z}}\Big\{f(z)\!:\ g(z)\in K,\ (G(z),H(z))\in{\rm gph}\,\partial\|\cdot\|_*\Big\},\] where \(f\!:\mathbb{Z}\to\mathbb{R}\) is a locally Lipschitz function, \(g\!:\mathbb{Z}\to\mathbb{R}^l\times\mathbb{R}^{m\times n}\) and \(G,H\!:\mathbb{Z}\to\mathbb{R}^{m\times n}\) are continuously differentiable mappings, \(K\) is a simple closed convex set of \(\mathbb{R}^l\times\mathbb{R}^{m\times n}\), and \({\rm gph}\,\partial\|\cdot\|_*\) denotes the graph of the subdifferential mapping of the nuclear norm. By the proof of Lemma [\[graph-subdiff\]](#graph-subdiff){reference-type="ref" reference="graph-subdiff"} below, we know that \((G(z),H(z))\in{\rm gph}\,\partial\|\cdot\|_*\) if and only if \(H(z)\!\in\mathop{\arg\max}_{Y\in\mathbb{B}}\langle G(z),Y\rangle\), where \(\mathbb{B}\!:=\!\big\{Z\in\mathbb{R}^{m\times n}\ |\ \|Z\|\le 1\big\}\) in this paper. This shows that the constraint \((G(z),H(z))\!\in{\rm gph}\,\partial\|\cdot\|_*\) represents a kind of optimality conditions. Therefore, problem [\[prob\]](#prob){reference-type="eqref" reference="prob"} is a mathematical program with a matrix equilibrium constraint \((G(z),H(z))\!\in{\rm gph}\,\partial\|\cdot\|_*\), which extends the optimization problems with polyhedral variational inequality constraints , second-order cone complementarity constraints , or positive semidefinite (PSD) complementarity constraints to those with general matrix equilibrium constraints. Our interest in [\[prob\]](#prob){reference-type="eqref" reference="prob"} comes from the fact that it covers an equivalent reformulation of low-rank optimization problems. Indeed, for the following rank minimization problem \[\label{rank-min} \min_{X\in\mathbb{R}^{m\times n}}\!\Big\{{\rm rank}(X)\!:\ \|\mathcal{A}(X)-b\|\le\delta,\,X\in\Omega\Big\},\] from and Lemma [\[graph-subdiff\]](#graph-subdiff){reference-type="ref" reference="graph-subdiff"} in Section [2](#sec2){reference-type="ref" reference="sec2"} we know that it can be reformulated as \[\label{equiv-rank} \min_{X,Y\in\mathbb{R}^{m\times n}}\!\Big\{\|Y\|_*\!:\ \|\mathcal{A}(X)-b\|\le\delta,\, X\in\Omega,\,(X,Y)\in{\rm gph}\,\partial\|\cdot\|_*\Big\},\] where \(\mathcal{A}\!: \mathbb{R}^{m\times n}\to\mathbb{R}^{N}\) is a sampling operator, \(b\in\mathbb{R}^{N}\) is a noisy observation vector, \(\delta>0\) is a constant related to the noise level, and \(\Omega\subseteq\mathbb{R}^{m\times n}\) is a closed convex set. Clearly, problem [\[equiv-rank\]](#equiv-rank){reference-type="eqref" reference="equiv-rank"} is a special case of [\[prob\]](#prob){reference-type="eqref" reference="prob"} with \(K=\{x\in\mathbb{R}^N\!:\|x\|\le\delta\}\times\Omega\), and \(g(z)=(\mathcal{A}(X)-b; X)\) and \((G(z),H(z))=(X,Y)\) for \(z=(X,Y)\in\mathbb{R}^{m\times n}\times\mathbb{R}^{m\times n}\). As it is well known, low-rank optimization problems have wide applications in many fields such as statistics, system identification and control, signal and image processing , machine learning, finance, quantum tomography , and so on. This motivates us to develop the optimality conditions and stability results for problem [\[prob\]](#prob){reference-type="eqref" reference="prob"}, especially the equivalent reformulation [\[equiv-rank\]](#equiv-rank){reference-type="eqref" reference="equiv-rank"} of the rank minimization problem [\[rank-min\]](#rank-min){reference-type="eqref" reference="rank-min"}. To achieve this goal, an essential step is to provide the characterization for the regular and limiting normal cones to the graph of the subdifferential mapping of the nuclear norm. In this work, we shall resolve this critical problem and as a byproduct establish the (regular) coderivative of the projection operator onto the unit spectral norm ball. Throughout this paper, we stipulate \(m\le n\). Let \(\mathbb{S}^m\) be the space of all \(m\times m\) real symmetric matrices and \(\mathbb{S}_{+}^m\) be the cone of all PSD matrices from \(\mathbb{S}^m\). Let \(\mathbb{O}^{n_1\times n_2}\) be the set of all \(n_1\times n_2\) real matrices with orthonormal columns, and \(\mathbb{O}^{n_1}\) be the set of all \(n_1\times n_1\) real orthogonal matrices. For \(Z\in\!\mathbb{R}^{m\times n}\), \(Z_{\alpha\beta}\) denotes the submatrix consisting of those \(Z_{ij}\) with \((i,j)\in\alpha\times\beta\). Let \(e\) and \(E\) be the vector and the matrix of all ones whose dimensions are known from the context, and for a vector \(z\), \({\rm Diag}(z)\) denotes a diagonal matrix which may be square or rectangular. For a given set \(S\), \(\delta_S(\cdot)\) means the indicator function over \(S\); \(\mathcal{T}_S(x)\) and \(\mathcal{T}_S^{i}(x)\) denote the tangent cone and the inner tangent cone to \(S\) at \(x\), respectively; and \(\mathcal{N}^{\pi}_S(x)\), \(\widehat{\mathcal{N}}_S(x)\) and \(\mathcal{N}_S(x)\) denote the proximal normal cone, the regular normal cone and the limiting normal cone to \(S\) at \(x\), respectively (see ). # Preliminaries {#sec2} This section includes three technical lemmas used for the subsequent analysis. The first one gives some characterizations for the graph of the subdifferential mapping \(\partial\|\cdot\|_*\). Lemma [\[graph-subdiff\]](#graph-subdiff){reference-type="ref" reference="graph-subdiff"} shows that the graph of \(\partial\|\cdot\|_*\) has an intimate link with \(\Pi_{\mathbb{B}}(\cdot)\). Motivated by this, we here establish an important property for the projection operator \(\Pi_{\mathbb{B}}(\cdot)\). Next we give the expression of the directional derivative of \(\Pi_{\mathbb{B}}\) by. Let \(\Omega_1,\Omega_2\in\mathbb{S}^{m}\) and \(\Omega_3\in\mathbb{R}^{m\times (n-m)}\) be the matrices associated with \(\sigma(\overline{Z})\), defined by \[\begin{aligned} \label{Omega1} \big(\Omega_1\big)_{ij} &:=\left\{\begin{array}{cl} \frac{\min(1,\sigma_i(\overline{Z}))-\min(1,\sigma_j(\overline{Z}))} {\sigma_i(\overline{Z})-\sigma_j(\overline{Z})}& {\rm if}\ \sigma_i(\overline{Z})\ne\sigma_j(\overline{Z}),\\ 0 & {\rm otherwise}\\ \end{array}\right.\ \ i,j\in\{1,2,\ldots,m\},\\ \label{Omega2} \big(\Omega_2\big)_{ij} &:=\left\{\begin{array}{cl} \frac{\min(1,\sigma_i(\overline{Z}))+\min(1,\sigma_j(\overline{Z}))} {\sigma_i(\overline{Z})+\sigma_j(\overline{Z})}& {\rm if}\ \sigma_i(\overline{Z})+\!\sigma_j(\overline{Z})\ne 0,\\ 0 & {\rm otherwise}\\ \end{array}\right.\ \ i,j\in\{1,2,\ldots,m\},\\ \label{Omega3} \big(\Omega_3\big)_{ij} &:=\left\{\begin{array}{cl} \frac{\min(1,\sigma_i(\overline{Z}))}{\sigma_i(\overline{Z})} & {\rm if}\ \sigma_i(\overline{Z})\ne 0,\\ 0 & {\rm otherwise}\\ \end{array}\right.\ \ i\in\{1,\ldots,m\},j\in\{1,\ldots,n\!-\!m\}. \end{aligned}\] Then, for any \(H\in\mathbb{R}^{m\times n}\), with \(\widetilde{H}_1=\overline{U}^{\mathbb{T}}H\overline{V}_1\) and \(\widetilde{H}=[\overline{U}^{\mathbb{T}}H\overline{V}_1\ \ \overline{U}^{\mathbb{T}}H\overline{V}_2]\) it holds that \[\begin{aligned} \Pi_{\mathbb{B}}'(\overline{Z};H) =\!\overline{U}\left[\begin{matrix} (\Omega_2)_{\alpha\alpha}\circ \big(\mathcal{X}(\widetilde{H}_1)\big)_{\alpha\alpha} &(\Omega_2)_{\alpha\beta}\!\circ\big(\mathcal{X}(\widetilde{H}_1)\big)_{\alpha\beta}& \widehat{H}_{\alpha\gamma} &(\Omega_3)_{\alpha c}\!\circ\!\widetilde{H}_{\alpha c}\\ (\Omega_2)_{\beta\alpha }\!\circ\! \big(\mathcal{X}(\widetilde{H}_1)\big)_{\beta\alpha }&\widetilde{H}_{\beta\beta} \!-\!\Pi_{\mathbb{S}_{+}^{|\beta|}}(\mathcal{S}(\widetilde{H}_{\beta\beta}))& \widetilde{H}_{\beta\gamma}&\widetilde{H}_{\beta c}\\ \widehat{H}_{\gamma\alpha}&\widetilde{H}_{\gamma\beta}&\widetilde{H}_{\gamma\gamma} &\widetilde{H}_{\gamma c}\\ \end{matrix}\right]\overline{V}^{\mathbb{T}}, \end{aligned}\] where \(\widehat{H}_{ij}=(\Omega_1)_{ij}\big(\mathcal{S}(\widetilde{H}_1)\big)_{ij}+(\Omega_2)_{ij}\big(\mathcal{X}(\widetilde{H}_1)\big)_{ij}\) for \((i,j)\in\alpha\times\gamma\) or \((i,j)\in\gamma\times\alpha\), and \(\mathcal{S}\!:\mathbb{R}^{m\times m}\to\mathbb{S}^m\) and \(\mathcal{X}\!:\mathbb{R}^{m\times m}\to\mathbb{R}^{m\times m}\) are two linear mappings defined by \[\label{ST-oper} \mathcal{S}(Z):=(Z\!+\!Z^{\mathbb{T}})/2\ \ {\rm and}\ \ \mathcal{X}(Z):=(Z\!-\!Z^{\mathbb{T}})/2 \quad\ \forall Z\in\mathbb{R}^{m\times m}.\] Now we are in a position to establish the coincidence between the regular normal cone to \({\rm gph}\,\partial\|\cdot\|_*\) and the proximal normal cone to \({\rm gph}\,\partial\|\cdot\|_*\). Proposition [\[prop-relation\]](#prop-relation){reference-type="ref" reference="prop-relation"} shows that, to characterize the regular normal cone to \({\rm gph}\,\partial\|\cdot\|_*\), one only needs to characterize its proximal normal cone. Next we shall employ Lemma [\[prox-normalcone\]](#prox-normalcone){reference-type="ref" reference="prox-normalcone"} and Lemma [\[dir-derivative\]](#dir-derivative){reference-type="ref" reference="dir-derivative"} to derive the expression of the proximal normal cone to \({\rm gph}\,\partial\|\cdot\|_*\). Then \((X^*,Y^*)\in \mathcal{N}_{{\rm gph}\,\partial\|\cdot\|_*}^{\pi}(X,Y)\) if and only if \((X^*,Y^*)\) satisfies the following conditions where \(\widetilde{X}_1^*=\overline{U}^{\mathbb{T}}X^*\overline{V}_1,\) \(\widetilde{Y}_1^*=\overline{U}^{\mathbb{T}}Y^*\overline{V}_1\), \(\widetilde{X}^*=\overline{U}^{\mathbb{T}}X^*\overline{V}\) and \(\widetilde{Y}^*=\overline{U}^{\mathbb{T}}Y^*\overline{V}\). To sum up, the fact that inequality [\[main-ineq1\]](#main-ineq1){reference-type="eqref" reference="main-ineq1"} holds for any \(H\in\mathbb{R}^{m\times n}\) implies that \[\begin{aligned} \label{main-equa2} \widetilde{X}_{\alpha\alpha}^*+\mathcal{X}(\widetilde{Y}_{\alpha\alpha}^*\!-\!\widetilde{X}_{\alpha\alpha}^*)\circ(\Omega_2)_{\alpha\alpha}=0,\ \ \widetilde{X}_{\beta\beta}^*\preceq 0,\ \widetilde{Y}_{\beta\beta}^*\succeq 0,\qquad\nonumber\\ \widetilde{X}_{\alpha\beta}^*+\big(\mathcal{X}(\widetilde{Y}_1^*\!-\!\widetilde{X}_1^*)\big)_{\alpha\beta}\circ (\Omega_2)_{\alpha\beta}=0,\, \widetilde{X}_{\beta\alpha}^*+\big(\mathcal{X}(\widetilde{Y}_1^*\!-\!\widetilde{X}_1^*)\big)_{\beta\alpha}\circ (\Omega_2)_{\alpha\beta}=0,\nonumber\\ \widetilde{X}_{\alpha\gamma}^*+\big(\mathcal{S}(\widetilde{Y}_1^*\!-\!\widetilde{X}_1^*)\big)_{\alpha\gamma} \circ(\Omega_1)_{\alpha\gamma}+\big(\mathcal{X}(\widetilde{Y}_1^*\!-\!\widetilde{X}_1^*)\big)_{\alpha\gamma} \circ(\Omega_2)_{\alpha\gamma}=0,\qquad\\ \widetilde{X}_{\gamma\alpha}^*+\big(\mathcal{S}(\widetilde{Y}_1^*\!-\!\widetilde{X}_1^*)\big)_{\gamma\alpha} \circ(\Omega_1)_{\gamma\alpha}+\big(\mathcal{X}(\widetilde{Y}_1^*\!-\!\widetilde{X}_1^*)\big)_{\gamma\alpha} \circ(\Omega_2)_{\gamma\alpha}=0,\,\widetilde{Y}_{\beta\gamma}^*=0,\qquad\nonumber\\ \widetilde{Y}_{\gamma\beta}^*=0,\widetilde{Y}_{\gamma\gamma}^*=0,\widetilde{Y}_{\beta c}^*=0,\widetilde{Y}_{\gamma c}^*=0,\, \widetilde{X}_{\alpha c}^*\circ (E_{\alpha c}-(\Omega_3)_{\alpha c})+\widetilde{Y}_{\alpha c}^*\circ (\Omega_3)_{\alpha c}=0.\nonumber \end{aligned}\] By the definitions of \(\Theta_1,\Theta_2\) and \(\Sigma_1,\Sigma_2\), equation [\[main-equa2\]](#main-equa2){reference-type="eqref" reference="main-equa2"} can be compactly written as [\[result-equa1\]](#result-equa1){reference-type="eqref" reference="result-equa1"}-[\[result-equa3\]](#result-equa3){reference-type="eqref" reference="result-equa3"}. Conversely, it is easy to check that if \((\widetilde{X}^*,\widetilde{Y}^*)\) satisfies [\[main-equa2\]](#main-equa2){reference-type="eqref" reference="main-equa2"} or its compact form [\[result-equa1\]](#result-equa1){reference-type="eqref" reference="result-equa1"}-[\[result-equa3\]](#result-equa3){reference-type="eqref" reference="result-equa3"}, then [\[main-ineq1\]](#main-ineq1){reference-type="eqref" reference="main-ineq1"} holds for any \(H\in\mathbb{R}^{m\times n}\), i.e., \((X^*,Y^*)\in\!\mathcal{N}_{{\rm gph}\,\partial\|\cdot\|_*}^{\pi}(X,Y)\). ◻ ## Limiting normal cone {#subsec3.2} Let \(\beta\) be a nonempty index set and denote the set of all partitions of \(\beta\) by \(\mathscr{P}(\beta)\). Write \(\mathbb{R}_{>}^{|\beta|}:=\big\{z\in\mathbb{R}^{|\beta|}\!:\ z_1\ge\cdots\ge z_{|\beta|}>0\big\}.\) For any \(z\in\mathbb{R}_{>}^{|\beta|}\), let \(D(z)\in\mathbb{S}^{|\beta|}\) denote the generalized first divided difference matrix of \(h(t)=\min(1,t)\) at \(z\), which is defined as \[\label{Ddiff-matrix} (D(z))_{ij}:=\left\{\begin{array}{cl} \!\frac{\min(1,z_i)-\min(1,z_j)}{z_i-z_j}\in[0,1]&{\rm if}\ z_i\ne z_j,\\ 0 &{\rm if}\ z_i=z_j\ge 1,\\ 1 & {\rm otherwise}. \end{array}\right.\] Write \(\mathcal{U}_{|\beta|} :=\big\{\overline{\Omega}\in\mathbb{S}^{|\beta|}\!:\ \overline{\Omega}=\lim_{k\to\infty}D(z^k),\, z^k\to e_{|\beta|},\, z^k\in\mathbb{R}_{>}^{|\beta|}\big\}.\) For each \(\Xi_1\in\mathcal{U}_{|\beta|}\), by equation [\[Ddiff-matrix\]](#Ddiff-matrix){reference-type="eqref" reference="Ddiff-matrix"} there exists a partition \((\beta_{+},\beta_{0},\beta_{-})\in\mathscr{P}(\beta)\) such that \[\label{Xi1-matrix} \Xi_1=\left[\begin{matrix} 0_{\beta_{+}\beta_{+}}& 0_{\beta_{+}\beta_{0}}&(\Xi_1)_{\beta_{+}\beta_{-}}\\ 0_{\beta_{0}\beta_{+}}& 0_{\beta_{0}\beta_{0}}& E_{\beta_{0}\beta_{-}}\\ (\Xi_1)_{\beta_{+}\beta_{-}}^{\mathbb{T}}&E_{\beta_{-}\beta_{0}}&E_{\beta_{-}\beta_{-}} \end{matrix}\right],\] where each entry of \((\Xi_1)_{\beta_{+}\beta_{-}}\) belongs to \([0,1]\). Let \(\Xi_2\) be the matrix associated to \(\Xi_1\): \[\label{Xi2-matrix} \Xi_2=\left[\begin{matrix} E_{\beta_{+}\beta_{+}}& E_{\beta_{+}\beta_{0}}& E_{\beta_{+}\beta_{-}}\!-\!(\Xi_1)_{\beta_{+}\beta_{-}}\\ E_{\beta_{0}\beta_{+}}& 0_{\beta_{0}\beta_{0}}& 0_{\beta_{0}\beta_{-}}\\ E_{\beta_{-}\beta_{+}}\!-\!(\Xi_1)_{\beta_{+}\beta_{-}}^{\mathbb{T}}&0_{\beta_{-}\beta_{0}}&0_{\beta_{-}\beta_{-}} \end{matrix}\right].\] With the above notations, we shall provide the exact formula for the limiting normal cone to \({\rm gph}\,\partial \|\cdot\|_*\) in the following theorem, whose proof is included in Appendix. where \(\widetilde{G}_1=\overline{U}^{\mathbb{T}}G\overline{V}_1,\widetilde{H}_1\!=\overline{U}^{\mathbb{T}}H\overline{V}_1\), and \(\Theta_1,\Theta_2,\Sigma_1,\Sigma_2\) are the same as those before. ::: To close this paper, we point out that Theorem [\[normal-theorem\]](#normal-theorem){reference-type="ref" reference="normal-theorem"} also provides the characterization for the coderivative of \(\Pi_\mathbb{B}\). Indeed, by Lemma [\[graph-subdiff\]](#graph-subdiff){reference-type="ref" reference="graph-subdiff"}, \({\rm gph}\,\Pi_{\mathbb{B}}=\mathcal{L}^{-1}({\rm gph}\,\partial\|\cdot\|_*)\) with \[\mathcal{L}(X,Y):=(X\!-\!Y,Y)\ \ {\rm for}\ \ (X,Y)\in\mathbb{R}^{m\times n}\times\mathbb{R}^{m\times n}.\] Since the linear map \(\mathcal{L}\!:\mathbb{R}^{m\times n}\times\mathbb{R}^{m\times n}\to \mathbb{R}^{m\times n}\times\mathbb{R}^{m\times n}\) is onto, from \[\mathcal{N}_{{\rm gph}\,\Pi_{\mathbb{B}}}(X,Y)\!=\!\mathcal{L}^*(\mathcal{N}_{{\rm gph}\,\partial\|\cdot\|_*}(\mathcal{L}(X,Y))),\] where \(\mathcal{L}^*\) is the adjoint of \(\mathcal{L}\). By the definition of coderivative (see ), \[W\in D^*\Pi_\mathbb{B}(X,Y)(S)\Longleftrightarrow (W,W\!-\!S)\in\mathcal{N}_{{\rm gph}\,\partial\|\cdot\|_*}(X\!-\!Y,Y).\] Similarly, Theorem [\[proxnormal-theorem\]](#proxnormal-theorem){reference-type="ref" reference="proxnormal-theorem"} also provides the characterization for the regular coderivative: \[W\in \widehat{D}^*\Pi_\mathbb{B}(X,Y)(S)\Longleftrightarrow (W,W\!-\!S)\in\widehat{\mathcal{N}}_{{\rm gph}\,\partial\|\cdot\|_*}(X\!-\!Y,Y).\] In addition, with the help of Theorem [\[normal-theorem\]](#normal-theorem){reference-type="ref" reference="normal-theorem"} and the Mordukhovich criterion on the Aubin property of a multifunction, one may easily obtain the practical conditions for the Aubin property of \(\partial\|\cdot\|_*\). In our future work, we shall use Theorem [\[proxnormal-theorem\]](#proxnormal-theorem){reference-type="ref" reference="proxnormal-theorem"} and [\[normal-theorem\]](#normal-theorem){reference-type="ref" reference="normal-theorem"} to derive the optimality conditions of the rank minimization problem [\[rank-min\]](#rank-min){reference-type="eqref" reference="rank-min"}.

{'timestamp': '2017-01-23T02:03:03', 'yymm': '1607', 'arxiv_id': '1607.04826', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04826'}

# Introduction Nowadays, millions of videos are being uploaded to the Internet every day. These videos capture all aspects of multimedia content about their uploader's daily life. These explosively growing user generated content videos online are becoming an crucial source of video data. Automatically categorizing videos into concepts, such as people actions, objects, etc., has become an important research topic. Recently many work have been proposed to tackle with building concept detectors both in image domain and video domain. However, the need for manual labels by human annotators has become one of the major important limitations for large-scale concept learning. It is even more so in video domain, since training concept detectors on videos is more challenging than on still images. Many image datasets such as ImageNet , CIFAR , PASCAL VOC, MS COCO  and Caltech  have been collected and manually labeled. In video domain, some largest datasets such as UCF-101, MCG-WEBV , TRECVID MED  and FCVID  are popular benchmark datasets for video classification. Collecting such datasets requires a large amount of human effort that can take thousands of man hours. In addition, manually labeling video requires playing back the video, which is more time consuming and expensive than labeling still images. As a result, the largest labeled video collection, FCVID , only contains about 0.09 million labels with 239 concept classes, much less than the 14 million labels with over 20,000 classes in the image collection ImageNet . Many state-of-the-art models in visual classification are based on the neural networks . As the architecture gets deeper, the neural network would need more data to train in order to get better performance. However, more data needs more human supervision which are more expensive to acquire in the video domain. Videos are available on the web and contain rich contextual information with a weak annotation about their content, such as their titles, descriptions and surrounding text. These webly-labeled data are orders of magnitude larger than that of any manually-labeled collections. Moreover, automatically extracted features from multiple modalities such as existing still image classification models, automatic speech recognition and optical character recognition tools can be useful additional information for the content of the video. Figure [\[cur-example\]](#cur-example){reference-type="ref" reference="cur-example"} shows an example of webly-labeled video for walking with a dog. As we see, the textual metadata we get from the web videos contain useful but very noisy information. The multi-modal prior information we get is correlated across modalities, as the image classification results and speech transcript show high probability of dog appearance, while the textual metadata indicates the same content. Some of the videos (about 20% in the FCVID dataset) have very little textual metadata and we can only obtain web labels via other modalities. To address the problem of learning detectors from the big web data, in this paper, we utilize multi-modal information to harness prior knowledge from the web video without any manual efforts. Existing methods on learning from noisy webly-labeled data has mainly focused on the image domain. Existing studies demonstrated promising results in this direction. However, these methods are primarily based on some heuristic methods. It is not clear what objective is being optimized and where or even whether the learning process will converge. Moreover, these methods only utilize a single text modality in the image domain. It is unclear how to exploit the multi-modal prior knowledge for concept learning from the rich context of Internet video. To utilize the large amount of webly-labeled video data for concept learning, we propose a learning framework called **WEbly-Labeled Learning (WELL)**. It is established on the theories called *curriculum learning*  and *self-paced learning* . The learning framework is motivated by human learning, where people generally start learning easier aspects of a concept, and then gradually take more complex examples into the learning process. Following this idea, WELL learns a concept detector iteratively from first using a few samples with more confident labels (more related to the concept), then gradually incorporate more video samples with noisier labels. The algorithm combines the prior knowledge, called learning curriculum, extracted from the webly-labeled data with the dynamic information learned from the statistical model (self-paced) to determine which video samples to learn in the next iteration. This idea of easy-to-hard learning paradigm has been adopted for learning in noisy web data  and has been proved to be efficient to deal with noise and outliers. Our proposed method generalizes such learning paradigm using a clear objective function. It is proved to be convex and is a also general framework that can incorporate state-of-the-art deep learning methods to learn robust detectors from noisy data. Our framework fundamentally changes self-paced learning and allows learning for video concept detectors at unlimited scale. Figure [\[well-cnn\]](#well-cnn){reference-type="ref" reference="well-cnn"} shows the architecture of the proposed method. We extract keyframe-level convolutional neural network features and feed them into WELL layer with average pooling and iterative learning process. We have also tried using other video features such as motion features and audio MFCC features. Our contributions are threefold. First, we address the problem of learning robust video concept detectors from noisy web data through a general framework with solid theoretical justifications. We show that WELL not only outperforms state-of-the-art learning methods on noisy labels, but also, notably, achieves comparable results with state-of-the-art models trained using manual annotation on one of the largest video dataset. Second, we provide detailed comparison of different approaches to exploit multi-modal curriculum from noisy labels and verify that our method is robust against certain level of noisiness in the video data. Finally, the efficacy and the scalability have been empirically demonstrated on two public benchmarks, including by far the largest manually-labeled video set called FCVID  and the largest multimedia dataset called YFCC100M . The promising results suggest that detectors trained on sufficient webly-labeled videos may outperform detectors trained on any existing manually-labeled datasets. # Related Work **Curriculum and Self-paced Learning**: Recently a learning paradigm called *curriculum learning* (CL) was proposed by Bengio et al., in which a model is learned by gradually incorporating from easy to complex samples in training so as to increase the entropy of training samples . A curriculum determines a sequence of training samples and is often derived by predetermined heuristics in particular problems. For example, Chen et al. designed a curriculum where images with clean backgrounds are learned before the images with noisy backgrounds , i.e. their method first builds a feature representation by a Convolutional Neural Network (CNN) on images with clean background and then they fine tune the models on images with noisy background. In , the authors approached grammar induction, where the curriculum is derived in terms of the length of a sentence. Because the number of possible solutions grows exponentially with the length of the sentence, and short sentences are easier and thus should be learn earlier. The heuristic knowledge in a problem often proves to be useful. However, the curriculum design may lead to inconsistency between the fixed curriculum and the dynamically learned models. That is, the curriculum is predetermined a prior and cannot be adjusted accordingly, taking into account the feedback about the learner. To alleviate the issue of CL, Kumar et al. designed a learning paradigm, called *self-paced learning* (SPL) . SPL embeds curriculum design as a regularizer into the learning objective. Compared with CL, SPL exhibits two advantages: first, it jointly optimizes the learning objective with the curriculum, and thus the curriculum and the learned model are consistent under the same optimization problem; second, the learning is controlled by a regularizer which is independent of the loss function in specific problems. This theory has been successfully applied to various applications, such as matrix factorization , action/event detection , domain adaption , tracking  and segmentation , reranking , etc. **Learning Detectors in Web Data**: Many recent studies have been proposed to utilize the large amount of noisy data from the Internet. For example, proposed a Never-Ending Language Learning (NELL) paradigm and built adaptive learners that makes use of the web data by learning different types of knowledge and beliefs continuously. Such learning process is mostly self-supervised, and previously learned knowledge enables learning further types of knowledge. In the image domain, existing methods try to tackle the problem of constructing qualified training sets based on the search results of text or image search engines . For example, extended the probabilistic Latent Semantic Analysis in visual domain and learned object categories using results from image search engines. proposed an incremental learning paradigm that initialized from a few seed images and repeatedly trained models to refine the collected image dataset from the Internet. NEIL  followed the idea of NELL and learned from web images to form a large collection of concept detectors iteratively via a semi-supervised fashion. By combining the classifiers and the inter-concept relationships it learned, NEIL can be used for scene classification and object detection task. tried to learn robust classifiers by considering the noisy textual information accompanied with web images. However, the down side is that the portion of the true positive samples has to be determined via prior knowledge, where in fact it is not accurate to assume the same number of true positive samples for any targeted concept. introduced a webly-supervised visual concept learning method that automatically learns large amount of models for a wide range of variations within visual concepts. They discovered concept variances through vocabulary of online books, and then downloaded images based on text-search from the web to train object detection and localization models. presented a weakly-supervised method called Baby Learning for object detection from a few training images and videos. They first embed the prior knowledge into a pre-trained CNN. When given very few samples for a new concept, a simple detector is constructed to discover much more training instances from the online weakly labeled videos. As more training samples are selected, the concept detector keeps refining until a mature detector is formed. Another recent work in image domain proposed a webly supervised learning of Convolutional Neural Network. They utilized easy images from search engine like Google to bootstrap a first-stage network and then used noisier images from photo-sharing websites like Flickr to train an enhanced model. In video domain, only few studies have been proposed for noisy data learning since training robust video concept detectors is more challenging than the problem in the image domain. tackled visual event detection problem by using SVM based domain adaptation method in web video data. described a fast automatic video retrieval method using web images. Given a targeted concept, compact representations of web images obtained from search engines like Google, Flickr are calculated and matched to compact features of videos. Such method can be utilized without any pre-defined concepts. discussed a method that exploits the YouTube API to train large scale video concept detectors on YouTube. The method utilized a calibration process and hard negative mining to train a second order mixture of experts model in order to discover correlations within the labels. Most of the existing methods are heuristic approaches as it is unclear what objective is being optimizing on the noisy data. Moreover, results obtained from the web search results is just one approach to acquire prior knowledge or curriculum. To the best of our knowledge, there have been no systematical studies on exploiting the multi-modal prior knowledge in video concept learning on noisy data. Since search engine algorithm is changing rapidly, it is unclear that how noisy the web labels are and how the level of noisiness in the data will affect performance. In this paper, we proposed a theoretically justified method with clear framework for curriculum constructing and model learning. We also empirically demonstrate its superior performance over representative existing methods and systemically verify that WELL is robust against the level of noisiness of the video data. # WEbly-Labeled Learning (WELL) ## Problem Description In this paper, following , we consider a concept detector as a classifier and our goal is to train concept detectors from webly-labeled video data without any manual labeling effort. Given a noisy web video training set and a target concept set, we do not assume any distribution of the noise. Formally, we represent the training set as \(\mathcal{D} = \{(\mathbf{x}_i,\mathbf{z}_i,\mathbf{\tilde{y}}_i)\}_{i=1}^n\) where \(\mathbf{x}_i \in \mathbb{R}^m\) denotes the feature for the \(i^{th}\) observed sample, and \(\mathbf{z}_i\) represents its noisy web label, which generally means the prior knowledge we can get from the web without additional human effort that not only includes textual information provided by the uploaders in the video metadata but also includes prior knowledge from other modalities using existing tools like pre-trained Convolutional Neural Network image detector, Automatic Speech Recognition and Optical character recognition. The \(\mathbf{\tilde{y}}_i \subset \mathcal{Y}\) is the inferred concept label set for the \(i^{th}\) observed sample based on its noisy web label, and \(\mathcal{Y}\) denotes the full set of target concepts. In our experiment, to simplify the problem, we apply our method on binary classification and infer binary labels \(\tilde{y}_i\) from the noisy web labels. The noisy web labels can be used to automatically infer concept labels by matching the concept name to the video textual metadata. For example, a video may be inferred to the concept label "cat" as its textual title contains cat. utilizes the YouTube topic API, which is derived from the textual metadata, to automatically get concept labels for videos. The web labels are quite noisy as the webly-labeled concepts may not present in the video content whereas the concepts not in the web label may well appear. ## Model and Algorithm ### Objective Function To leverage the noisy web labels in a principled way, we propose WEbly-Labeled Learning (WELL). Formally, given a training set \(\mathcal{D}\) as described before, Let \(L(\tilde{y}_i,g(\mathbf{x}_i,\mathbf{w}))\), or \(\ell_i\) for short, denote the loss function which calculates the cost between the inferred label \(\tilde{y}_i\) and the estimated label \(g(\mathbf{x}_i,\mathbf{w})\). Here \(\mathbf{w}\) represents the model parameter inside the decision function \(g\). For example, in our paper, \(\mathbf{w}\) represents the weight parameters in the Convolutional Neural Network (CNN) and the Support Vector Machine (SVM). Our objective function is to jointly learn the model parameter \(\mathbf{w}\) and the latent weight variable \(\mathbf{v}= [v_1,\cdots,v_n]^T\) by: \[\label{eq:spcl_obj} \begin{split} \!\min_{\mathbf{w},\mathbf{v}\in \lbrack 0,\!1]^{n}}\!\!\mathbb{E}(\mathbf{w},\!\mathbf{v}\!;\lambda,\!\Psi\!) \!=\! \sum_{i=1}^n v_i L(\tilde{y}_i,\!g(\mathbf{x}_i,\!\mathbf{w})) \!+\! f(\mathbf{v};\! \lambda), \\ \text{ subject to } \mathbf{v} \in \Psi \end{split}\] where \(\mathbf{v=[}v_{1},v_{2},\cdots ,v_{n}\mathbf{]}^{T}\) denote the latent weight variables reflecting the inferred labels' confidence. The weights determine a learning sequence of samples, where samples with greater weights tend to be learned earlier. Our goal is to assign greater weights to the samples with more confident labels whereas smaller or zero weights to the samples with noisy labels. To this end, we employ the self-paced regularizer \(f\), which controls the learning process of the model. We consider the linear regularizer Eq. [\[eq:linear_scheme\]](#eq:linear_scheme){reference-type="eqref" reference="eq:linear_scheme"} proposed in : \[\label{eq:linear_scheme} f(\mathbf{v};\lambda) = \frac{1}{2} \lambda \sum_{i=1}^n ( v_i^2-2v_i ). \vspace{-1mm}\] Generally, a self-paced regularizer determines the scheme for penalizing the latent weight variables. Physically it resembles the learning schemes human used in understanding new concepts. The linear scheme corresponds to a prudent strategy, which linearly penalizes the samples that are different to what the model has already learned (see Eq. [\[eq:linear_closedform\]](#eq:linear_closedform){reference-type="eqref" reference="eq:linear_closedform"}). The hyper-parameter \(\lambda\) \((\lambda > 0)\) is called "model age", which controls the pace at which the model learns new samples. When \(\lambda\) is small only samples of with small loss will be considered. As \(\lambda\) grows, more samples with larger loss will be gradually appended to train a "mature" mode. \(\Psi\) in Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"} is a curriculum region derived from noisy web labels \(\mathbf{z}\) that incorporates the prior knowledge extracted from the webly-labeled data as a convex feasible region for the weight variables. The shape of the region weakly implies a prior learning sequence of samples, where the expected values for favored samples are larger. The curriculum region can be derived in a variety of ways that make use of different modalities. We will discuss this topic in details in following section. A straightforward approach is by counting the term frequency in the video's textual metadata. That is, for example, the chance of a video containing the concept "cat" become higher when it has more word "cat" in its title, description or tags. Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"} represents a concise and general optimization model  . It combines the prior knowledge extracted from the noisy webly-labeled data (as the curriculum region) and the information dynamically learned during the training (via the self-paced regularizer). Intuitively, the prior knowledge serves as an instructor providing a guidance on learning the latent weights, but it leaves certain freedom for the model (the student) to adjust the actual weights according to its learning pace. Experimental results in Section [9](#sec:experiments){reference-type="ref" reference="sec:experiments"} demonstrate the learning paradigm can better overcome the noisy labels than heuristic approaches. Figure [\[well-cnn\]](#well-cnn){reference-type="ref" reference="well-cnn"} shows the learning process of our method. Following , we employ the alternative convex search algorithm to solve Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"}. Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} takes the input of a curriculum region, an instantiated self-paced regularizer and a step size parameter; it outputs an optimal model parameter \(\mathbf{w}\). First of all, it initializes the latent weight variables in the feasible region. Then it alternates between two steps until it finally converges: Step 3 learns the optimal model parameter with the fixed and most recent \(\mathbf{v}^*\); Step 5 learns the optimal weight variables with the fixed \(\mathbf{w}^*\). In the beginning, the model "age" is gradually increased so that more noisy samples will be gradually incorporated in the training. Step 3 can be conveniently implemented by existing off-the-shelf supervised learning methods such as the back propagation. Gradient-based methods can be used to solve the convex optimization problem in Step 4. According to , the alternative search in Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} converges as the objective function is monotonically decreasing and is bounded from below. At an early age when \(\lambda\) is small, Step 4 in Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} has an evident suppressing effect over noisy samples that have greater loss to the already learned model. For example, with a fixed \(\mathbf{w}\), the unconstrained close-formed solution for the regularizer in Eq. [\[eq:linear_scheme\]](#eq:linear_scheme){reference-type="eqref" reference="eq:linear_scheme"} equals \[\label{eq:linear_closedform} v_i^* =\begin{cases}-\frac{1}{\lambda} \ell_i +1 & \ell_{i} < \lambda\\ 0 & \ell_{i} \ge \lambda \end{cases},\] where \(v_i\) represents the \(i\)th element in the optimal solution \(\mathbf{v}^* = [v_1^*, \cdots, v_n^*]^T\). Eq. [\[eq:linear_closedform\]](#eq:linear_closedform){reference-type="eqref" reference="eq:linear_closedform"} called linear regularizer indicates the latent weight is proportional to the negative sample loss, and the sample whose loss is greater or equals to \(\lambda\) will have zero weights and thus will not affect the training of the next model. As the model age grows, the hyper-parameter \(\lambda\) increases, and more noisy samples will be used into training. The prior knowledge embedded in the curriculum region \(\Psi\) is useful as it suggests a learning sequence of samples for the "immature" model. theoretically proves that the iterative learning process is identical to optimizing a robust loss function on the noisy data. If we keep increasing \(\lambda\), the model will ultimately use every sample in the noisy data, which is undesirable as the labels of some noisy samples are bound to be incorrect. To this end, we stop increasing the age \(\lambda\) after about a certain number of iterations (early stopping). The exact stopping iteration for each detector is automatically tuned in terms of its performance on a small validation set. ### Model Details {#sec:partial_curriculum} In this section we discuss further details of the curriculum region \(\Psi\) and the self-paced regularizer \(f(\mathbf{v};\lambda)\). \(\Psi\) is a feasible region that embeds the prior knowledge extracted from the webly-labeled data. It physically corresponds to a convex search region for the latent weight variable. Given a set of training samples \(\mathbf{X}=\{\mathbf{x}_{i}\}_{i=1}^{n}\), we utilize the partial-order curriculum which generalizes the total-order curriculum by incorporating the incomplete prior over groups of samples. Samples in the confident groups should be learned earlier than samples in the less confident groups. It imposes no prior over the samples within the same group nor the samples not in any group. Formally, we define a partial order relation \(\preceq\) such that \(x_i \preceq x_j\) indicates that the sample \(x_i\) should be learned no later than \(x_j\) (\(i,j \in [1,n]\)). Similarly given two sample subsets \(\mathbf{X}_{a} \preceq \mathbf{X}_{b}\) denotes the samples in \(\mathbf{X}_{a}\) should be learned no later than the samples in \(\mathbf{X}_{b}\). In our problem, we extract the partial-order curriculum in the webly-labeled data in the following ways: we only distinguish the training order for groups of samples. Information from different modalities of the web labels can be used for curriculum design. A straightforward way is to directly utilize the textual descriptions of the videos generated by the uploaders. We compare common ways to extract curriculum from web data for concept learning to the proposed novel method that utilize state-of-the-art topic modeling techniques in natural language processing. In the following methods (Exact & Stem Matching, Word Embedding and Latent Topic with Word Embedding), we first extract bag-of-words features from different modalities and then match them using specific matching methods to the concept words. Each video will then come with a matching score to each concept. In our experiment, we divide the data into two partial-order curriculum groups, where the videos with matching scores larger than zero will be in one group while others will be in the other group. **Exact & Stem Matching** We build curriculum directly using exact word matching or stemmed word matching between the textual metadata of the noisy videos to the targeted concept names. **YouTubeTopicAPI** We directly utilize the YouTube topic API to search for videos that are related to the concept words. The topic API utilizes textual information of the uploaded videos to obtain related topics of the videos from Freebase. **SearchEngine** We build curriculum using the search result from a text-based search engine. It is similar to related web-search based methods. **Word Embedding** We use word embedding to match words in metadata to targeted concept words in order to deal with synonyms and related concepts. The word embedding is trained using Google News data. **Latent Topic** We build curriculum based on the latent topic we learned from the noisy label. We incorporate Latent Dirichlet Allocation (LDA)  to determine how each noisy labeled video is related to each target concept. The basic idea is that each web video consists of mixtures of topics (concepts), and each topic is characterized by a distribution of words. Formally, given all the noisy information extracted from a web video and collected them as a document \(\mathbf{d_i}\), which combines into a corpus \(\mathbf{d}\), we have a target set of \(\mathbf{k}\) topics, then the key inferential problem that we are going to solve is that of computing the posterior distribution of the latent topics given a corpus (how likely the videos are related to each target concept given the noisy information): \[\label{eq:lda} p(\theta,\mathbf{t}|\mathbf{d},\alpha,\beta) = \frac{p(\theta,\mathbf{t},\mathbf{d}|\alpha,\beta)}{p(\mathbf{d}|\alpha,\beta)}\] where \(\theta\) is the topic distribution variable for the corpus, \(\theta\) \(\sim\) Dir(\(\alpha\)), in which Dir(\(\alpha\)) represents a uniform Dirichlet distribution with scaling parameter \(\alpha\). The \(\mathbf{t}\) is the topic assignment variable that indicates which topic each word belong to in the document. \(\beta\) is the Dirichlet prior on the per-topic word distribution, in which we impose asymmetric priors over the word distribution so that each learned topic will be seeded with particular words in our target concept. For example, a topic will be seeded with words \"walk, dog\" for the target concept \"WalkingWithDog\". The parameter estimation in Eq [\[eq:lda\]](#eq:lda){reference-type="eqref" reference="eq:lda"} can be done via Bayes methods. However, Eq. [\[eq:lda\]](#eq:lda){reference-type="eqref" reference="eq:lda"} is intractable to compute since \(p(\mathbf{d}|\alpha,\beta)\) is intractable due to the coupling between \(\theta\) and \(\beta\). To solve this problem, approximate inference algorithms are introduced and we use the online variational inference algorithm from. The true posterior is approximated by a simpler distribution: \[\label{eq:lda-vi} q(\theta,\mathbf{t}|\gamma,\phi) = q(\theta|\gamma)\prod_{i=1}^{N}{q(t_i|\phi_i)}\] where the Dirichlet parameter \(\gamma\) and the multinomial parameters (\(\phi_1...,\phi_N\)) are the free variational parameters. Thus the maximization problem is equivalent to minimizing the Kullback-Leibler(KL) divergence between \(q(\theta,\mathbf{t}|\gamma,\phi)\) and the posterior \(p(\theta,\mathbf{t}|\mathbf{d},\alpha,\beta)\). The optimization problem can then be solve using Expectation-Maximization algorithm. The estimated parameters \(\theta,t\) in Eq. [\[eq:lda\]](#eq:lda){reference-type="eqref" reference="eq:lda"} is then used for constructing curriculum region in Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"}. **Latent Topic with Word Embedding (LT+WE)** We first learn latent topics using LDA to replace the concept words with a topic word distribution and then match the latent topic words to the web label's bag-of-words features by using the word embeddings. We compare this method to the others using only textual information from the web videos. We also use this method to get curriculum from other modalities such as Automatic Speech Recognition (ASR), Optical Character Recognition (OCR) and basic image detector pre-trained on still images  (in this paper we use VGG net , extract keyframe-level image classification results and average them to get video-level results.). We extract bag-of-word features from them and combine them with linear weights. Detailed fusion experiments can be found in Section [9](#sec:experiments){reference-type="ref" reference="sec:experiments"}. We empirically set OCR's weight to be small as the results are much noisier than other features. Figure [\[curriculum_example\]](#curriculum_example){reference-type="ref" reference="curriculum_example"} shows an example of the noisy web video data and how the curriculum is extracted with different methods. Our method can utilize information from different modalities while common methods like search engine only consider textual information. We compare the performance of different ways of curriculum design by training detectors directly in Section 4. The labels in webly-labeled data are much noisier than manually-labeled data, and as a result, we found that the learning is prone to overfitting the noisy labels. To address this issue, inspired by the dropout technique in deep learning , we use a dropout strategy for webly-labeled learning . It is implemented in the self-paced regularizer discussed in Section 3. With the dropout, the regularizers become: \[\label{eq:dropout} \vspace{-2mm} \begin{split} r_i(p) \sim \text{Bernoulli}(p) + \epsilon, (0 < \epsilon \ll 1)\\ f(\mathbf{v};\lambda, p) = \frac{1}{2} \lambda \sum_{i=1}^n (\frac{1}{r_i} v_i^2-2v_i), \end{split} \vspace{-2mm}\] where \(\mathbf{r}\) is a column vector of independent Bernoulli random variables with the probability \(p\) of being 1. Each of the element equals the addition of \(r_i\) and a small positive constant \(\epsilon\). Denote \(\mathbb{E}_{\mathbf{w}} = \sum_{i=1}^n v_i \ell_i + f(\mathbf{v};\lambda)\) as the objective with the fixed model parameters \(\mathbf{w}\) without any constraint, and the optimal solution \(\mathbf{v}^* = [v_1^*, \cdots, v_n^*]^T= \argmin_{\mathbf{v}\in[0,1]^n} \mathbb{E}_{\mathbf{w}}\). We have: \[\vspace{-2mm} \label{eq:linear_dropout} \begin{split} \mathbb{E}_{\mathbf{w}} = \sum_{i=1}^n \ell_i v_i + \lambda ( \frac{1}{2r_i} v_i^2-v_i );\\ \frac{\partial \mathbb{E}_{\mathbf{w}}}{\partial v_i} = \ell + \lambda v_i/r_i -\lambda = 0;\\ \Rightarrow v_i^* =\begin{cases} r_i(-\frac{1}{\lambda} \ell_i +1) & \ell_{i} < \lambda\\ 0 & \ell_{i} \ge \lambda \end{cases}. \end{split} \vspace{-2mm}\] The dropout effect can be demonstrated in the closed-form solutions in Eq. [\[eq:linear_dropout\]](#eq:linear_dropout){reference-type="eqref" reference="eq:linear_dropout"}: with the probability \(1-p\), \(v_i^*\) in both the equations approaches 0; with the probability \(p\), \(v_i^*\) approaches the solution of the plain regularizer discussed in Eq. [\[eq:linear_scheme\]](#eq:linear_scheme){reference-type="eqref" reference="eq:linear_scheme"}. Recall the self-paced regularizer defines a scheme for learning samples. Eq. [\[eq:linear_dropout\]](#eq:linear_dropout){reference-type="eqref" reference="eq:linear_dropout"} represent the new dropout learning scheme. When the base learner is neural networks, the proposed dropout can be used combined with the classical dropout in . The term dropout in this paper refers to dropping out samples in the iterative learning. By dropping out a sample, we drop out its update to the model parameter, which resembles the classical dropout used in neural networks. It operates on a more coarse-level which is useful for noisy data. When samples with incorrect noisy labels update a model, it will encourage the model to select more noisy labels. The dropout strategy prevents overfitting to noisy labels. It provides a way of combining many different sample subsets in different iterations in order to help avoid bad local minima. Experimental results substantiate this argument. In practice, we recommend setting two Bernoulli parameters for positive and negative samples on imbalanced data. # Experiments {#sec:experiments} In this section, we evaluate our method WELL for learning video detectors on noisy labeled data. The experiments are conducted on two major public benchmarks: FCVID and YFCC100M, where FCVID is by far one of the biggest manually annotated video dataset, and the YFCC100M dataset is the largest multimedia benchmark. ## Experimental Setup **Datasets, Features and Evaluation Metrics** Fudan-columbia Video Dataset (FCVID) contains 91,223 YouTube videos (4,232 hours) from 239 categories. It covers a wide range of concepts like activities, objects, scenes, sports, DIY, etc. Detailed descriptions of the benchmark can be found in . Each video is manually labeled to one or more categories. In our experiments, we do not use the manual labels in training, but instead we automatically generate the web labels according to the concept name appearance in the video metadata. The manual labels are used only in testing to evaluate our and the baseline methods. Following , the standard train/test split is used. The second set is YFCC100M which contains about 800,000 videos on Yahoo! Flickr with metadata such as the title, tags, the uploader, etc. There are no manual labels on this set and we automatically generate the curriculum from the metadata in a similar way. Since there are no annotations, we train the concept detectors on the most 101 frequent latent topics found in the metadata. There are totally 47,397 webly labeled videos on the 101 concepts for training. On FCVID, as the manual labels are available, the performance is evaluated in terms of the precision of the top 5 and 10 ranked videos (P\@5 and P\@10) and mean Average Precision (mAP) of 239 concepts. On YFCC100M, since there are no manual labels, for evaluation, we apply the detectors to a third public video collection called TRECVID MED which includes 32,000 Internet videos . We apply the detectors trained on YFCC100M to the TRECVID videos and manually annotate the top 10 detected videos returned by each method for 101 concepts. **Implementation Details** We build our method on top of a pre-trained convolutional neural network as the low-level features (VGG network ). We extract the key-frame level features and create a video feature by the average pooling. The same features are used across different methods on each dataset. The concept detectors are trained based on a hinge loss cost function. Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} is used to train the concept models iteratively and the \(\lambda\) stops increasing after 100 iterations. We automatically generate curriculum labels based on the video metadata, ASR, OCR and VGG net 1,000 classification results using latent topic modeling with word embedding matching as shown in Section 3, and derive a partial-order curriculum using the method discussed in Section 3.2.2. **Baselines** The proposed method is compared against the following five baseline methods which cover both the classical and the recent representative learning algorithms on webly-labeled data. *BatchTrain* trains a single SVM model using all samples in the multi-modal curriculum built as described in section 3.2.2 LT+WE. *Self-Paced Learning (SPL)* is a classical method where the curriculum is generated by the learner itself . *BabyLearning* is a recent method that simulates baby learning by starting with few training samples and fine-tuning using more weakly labeled videos crawled from the search engine. *GoogleHNM* is a hard negative mining method proposed by Google. It utilizes hard negative mining to train a second order mixture of experts model according to the video's YouTube topics. *FastImage* is a video retrieval method that utilizes web images from search engine to match to the video with re-ranking. *WELL* is the proposed method. The hyper-parameters of all methods including the baseline methods are tuned on the same validation set. On FCVID, the set is a standard development set with manual labels randomly selected from 10% of the training set (No training was done using ground truth labels) whereas on YFCC100M it is also a 10% proportion of noisy training set. ## Experiments on FCVID **Curriculum Comparison** As disscussed in Section 3.2.2, we compare different ways to build curriculum for noisy label learning. Here we also compare their effectiveness by training concept detectors directly using the curriculum labels. The batch train model is used for all generated cirriculumn labels. In Table [9](#exp-cur){reference-type="ref" reference="exp-cur"} we show the batch trained models' precision at 5, 10 and mean average precision on the test set of FCVID. For LT+WE (Multi-modal), we extract curriculum from different modalities as shown in Section 3.2.2, and combine them using linear weights. The weights are hyper-parameters that are tuned on the validation set, and the optimal weights for textual metadata, ASR, image classification and OCR results are 1.0, 0.5, 0.5 and 0.05, respectively. Results show that the curriculum generated by combining latent topic modeling and word embedding using multi-modal prior knowledge is the most accurate, which indicates our claim of exploiting multi-modal information is beneficial, and we use this method in WELL for the rest of the experiments. **Baseline Comparison** Table [10](#exps-basline){reference-type="ref" reference="exps-basline"} compares the precision and mAP of different methods where the best results are highlighted. As we see, the proposed WELL significantly outperforms all baseline methods, with statistically significant difference at \(p\)-level of 0.05. Comparing WELL with SPL, the effect of curriculum learning and dropout makes a significant difference in terms of performance, which suggests the importance of prior knowledge and preventing over-fitting in webly learning. The promising experimental results substantiate the efficacy of the proposed method. **Robustness to Noise Comparison** In this comparison we manually control the noisiness of the curriculum in order to systematically verify how our methods would perform with respect to the noisiness within the web data. The experimental results indicate the robustness of our method towards noisy labels. To this end, we randomly select video samples with ground truth labels for each concept, so that the precision of the curriculum labels are set at 20%, 40%, 60%, 80% and we fix the recall of all the labels. We then train WELL using such curriculum and test them on the FCVID testing set. We also compare WELL to three other methods with the same curriculum, among them *GoogleHNM* is a recent method to train video concept detector with large-scale data. We exclude *BabyLearning*, which relies on the returned results by the search engine, since in this experiment the curriculum is fixed. As shown in Table [12](#exp-noise){reference-type="ref" reference="exp-noise"}, as the noisiness of the curriculum grows (the precision drops), WELL maintains its performance while other methods drop significantly. Specifically, when the precision of the curriculum drops from 40% to 20%, other methods' mAP averagely drops 46.5% while WELL's mAP only drops 19.1% relatively. It shows that WELL is robust against different level of noise, which shows great potential in larger scale webly-labeled learning as the dataset gets bigger, the noisier it may become. [\[exp-noise\]]{#exp-noise label="exp-noise"} [\[exp-noise\]]{#exp-noise label="exp-noise"} **Ground-truth Training Comparison** In this part, we also compare our method with the state-of-the-art method trained using ground truth labels on FCVID (rDNN) . We compare WELL trained using the static CNN features, the standard features provided by the authors , and we also compare WELL using the late (average) fusion with CNN, motion and audio MFCC features (WELL-MM) to the method that achieves the best result on FCVID trained using the same multi-modal features. WELL-MM uses CNN, Motion and MFCC features, which is the same set of features as rDNN-F . Noted that the state-of-the-art method uses the ground truth labels to train models, which includes 42,223 videos with manual labels, while our proposed method uses none of the human annotation into training but still be able to outperform one of the state-of-the-art results. [\[exps-sa\]]{#exps-sa label="exps-sa"} **Noisy Dataset Size Comparison** To investigate the potential of concept learning on webly-labeled video data, we apply the methods on different sizes of subsets of the data. Specifically, we randomly split the FCVID training set into several subsets of 200, 500, 1,000, and 2,000 hours of videos, and train the models on each subset without using manual annotations. The models are then tested on the same test set. Table [14](#exps-small){reference-type="ref" reference="exps-small"} lists the average results of each type of subsets. As we see, the accuracy of WELL on webly-labeled data increases along with the growth of the size of noisy data while other webly learning methods' performance tend to be saturated. Comparing to the methods trained using ground truth, In Table [14](#exps-small){reference-type="ref" reference="exps-small"}, WELL-MM trained using the whole dataset (2000h) outperforms rDNN-F(trained using manual labels) trained using around 1200h of data. And since the incremental performance increase of WELL-MM is close to linear, we conclude that with sufficient webly-labeled videos WELL-MM will be able to outperform the rDNN-F trained using 2000h of data, which is currently the largest manual labeled dataset. [\[exps-small\]]{#exps-small label="exps-small"} ## Experiments on YFCC100M In the experiments on YFCC100M, we train 101 concept detectors on YFCC100M and test them on the TRECVID MED dataset which includes 32,000 Internet videos. Since there are no manual labels, to evaluate the performance, we manually annotate the top 10 videos in the test set and report their precisions in Table [15](#exps-yfcc){reference-type="ref" reference="exps-yfcc"}. The MED evaluation is done by four annotators and the final results are averaged from all annotations. The Fleiss' Kappa value for these four annotators is 0.64. A similar pattern can be observed where the comparisons substantiate the rationality of the proposed webly learning framework. Besides, the promising results on the largest multimedia set YFCC100M verify the scalability of the proposed method. ## Time Complexity Comparison The computation complexity of WELL is comparable to existing methods. For a single class, the complexity for our model and baseline models is \(O(r \times n \times m)\), where \(r\) is the number of iterations to converge, \(n\) is the number of training samples and \(m\) is the feature dimension. Theoretically, the complexity is comparable to the baseline models that have different \(r\). In practice, on a 40 core-CPU machine, WELL and SPL takes 7 hours to converge (100 iterations) on FCVID with 239 concepts, whereas GoogleHNM and BabyLearning take around 5 hours. In Table [16](#time){reference-type="ref" reference="time"}, we show the theoretical and actual run time for all methods. ## Qualitative Analysis In this section we show training examples of WELL. In Figure [\[well-pick-all\]](#well-pick-all){reference-type="ref" reference="well-pick-all"}, we demonstrate the positive samples that WELL select at different stage of training the concept \"baseball\", \"forest\", \"birthday\" and \"cow\". For the concept \"baseball\", at early stage (1/93, 25/93), WELL selects easier and clearer samples such as those camera directly pointing at the playground, while at later stage (75/93, 93/93) WELL starts to train with harder samples with different lighting conditions and untypical samples for the concept. For the concept \"birthday\", as we see, at later stage of the training, complex samples for birthday event like a video with two girl singing birthday song (75/84) and a video of celebrating birthday during hiking (84/84) are included in the training. For the concept \"forest\", at the final iteration (95/95), a video of a man playing nunchaku is included, as the video title contains \"Air Forester Nunchaku Freestyle\" and it is included in the curriculum, which is reasonable as the curriculum is noisy and may contain false positive. Since WELL is able to leave outliers in later stage of the training, the affection of the false positives in the curriculum can be alleviated by early stopping. In our experiments, we stop increasing \(\lambda\) after 100 iterations so in the final model, only a subset of samples are used. # Conclusions In this paper, we proposed a novel method called WELL for webly labeled video data learning. WELL extracts multi-modal informative knowledge from noisy weakly labeled video data from the web through a general framework with solid theoretical justifications. WELL achieves the best performance only using webly-labeled data on two major video datasets. The comprehensive experimental results demonstrate that WELL outperforms state-of-the-art studies by a statically significant margin on learning concepts from noisy web video data. In addition, the results also verify that WELL is robust to the level of noisiness in the video data. The result suggests that with more webly-labeled data, which is not hard to obtain, WELL can potentially outperform models trained on any existing manually-labeled data. # Introduction Nowadays, millions of videos are being uploaded to the Internet every day. These videos capture all aspects of multimedia content about their uploader's daily life. These explosively growing user generated content videos online are becoming an crucial source of video data. Automatically categorizing videos into concepts, such as people actions, objects, etc., has become an important research topic. Recently many work have been proposed to tackle with building concept detectors both in image domain and video domain. However, the need for manual labels by human annotators has become one of the major important limitations for large-scale concept learning. It is even more so in video domain, since training concept detectors on videos is more challenging than on still images. Many image datasets such as ImageNet , CIFAR , PASCAL VOC, MS COCO  and Caltech  have been collected and manually labeled. In video domain, some largest datasets such as UCF-101, MCG-WEBV , TRECVID MED  and FCVID  are popular benchmark datasets for video classification. Collecting such datasets requires a large amount of human effort that can take thousands of man hours. In addition, manually labeling video requires playing back the video, which is more time consuming and expensive than labeling still images. As a result, the largest labeled video collection, FCVID , only contains about 0.09 million labels with 239 concept classes, much less than the 14 million labels with over 20,000 classes in the image collection ImageNet . Many state-of-the-art models in visual classification are based on the neural networks . As the architecture gets deeper, the neural network would need more data to train in order to get better performance. However, more data needs more human supervision which are more expensive to acquire in the video domain. Videos are available on the web and contain rich contextual information with a weak annotation about their content, such as their titles, descriptions and surrounding text. These webly-labeled data are orders of magnitude larger than that of any manually-labeled collections. Moreover, automatically extracted features from multiple modalities such as existing still image classification models, automatic speech recognition and optical character recognition tools can be useful additional information for the content of the video. Figure [\[cur-example\]](#cur-example){reference-type="ref" reference="cur-example"} shows an example of webly-labeled video for walking with a dog. As we see, the textual metadata we get from the web videos contain useful but very noisy information. The multi-modal prior information we get is correlated across modalities, as the image classification results and speech transcript show high probability of dog appearance, while the textual metadata indicates the same content. Some of the videos (about 20% in the FCVID dataset) have very little textual metadata and we can only obtain web labels via other modalities. To address the problem of learning detectors from the big web data, in this paper, we utilize multi-modal information to harness prior knowledge from the web video without any manual efforts. Existing methods on learning from noisy webly-labeled data has mainly focused on the image domain. Existing studies demonstrated promising results in this direction. However, these methods are primarily based on some heuristic methods. It is not clear what objective is being optimized and where or even whether the learning process will converge. Moreover, these methods only utilize a single text modality in the image domain. It is unclear how to exploit the multi-modal prior knowledge for concept learning from the rich context of Internet video. To utilize the large amount of webly-labeled video data for concept learning, we propose a learning framework called **WEbly-Labeled Learning (WELL)**. It is established on the theories called *curriculum learning*  and *self-paced learning* . The learning framework is motivated by human learning, where people generally start learning easier aspects of a concept, and then gradually take more complex examples into the learning process. Following this idea, WELL learns a concept detector iteratively from first using a few samples with more confident labels (more related to the concept), then gradually incorporate more video samples with noisier labels. The algorithm combines the prior knowledge, called learning curriculum, extracted from the webly-labeled data with the dynamic information learned from the statistical model (self-paced) to determine which video samples to learn in the next iteration. This idea of easy-to-hard learning paradigm has been adopted for learning in noisy web data  and has been proved to be efficient to deal with noise and outliers. Our proposed method generalizes such learning paradigm using a clear objective function. It is proved to be convex and is a also general framework that can incorporate state-of-the-art deep learning methods to learn robust detectors from noisy data. Our framework fundamentally changes self-paced learning and allows learning for video concept detectors at unlimited scale. Figure [\[well-cnn\]](#well-cnn){reference-type="ref" reference="well-cnn"} shows the architecture of the proposed method. We extract keyframe-level convolutional neural network features and feed them into WELL layer with average pooling and iterative learning process. We have also tried using other video features such as motion features and audio MFCC features. Our contributions are threefold. First, we address the problem of learning robust video concept detectors from noisy web data through a general framework with solid theoretical justifications. We show that WELL not only outperforms state-of-the-art learning methods on noisy labels, but also, notably, achieves comparable results with state-of-the-art models trained using manual annotation on one of the largest video dataset. Second, we provide detailed comparison of different approaches to exploit multi-modal curriculum from noisy labels and verify that our method is robust against certain level of noisiness in the video data. Finally, the efficacy and the scalability have been empirically demonstrated on two public benchmarks, including by far the largest manually-labeled video set called FCVID  and the largest multimedia dataset called YFCC100M . The promising results suggest that detectors trained on sufficient webly-labeled videos may outperform detectors trained on any existing manually-labeled datasets. # Related Work **Curriculum and Self-paced Learning**: Recently a learning paradigm called *curriculum learning* (CL) was proposed by Bengio et al., in which a model is learned by gradually incorporating from easy to complex samples in training so as to increase the entropy of training samples . A curriculum determines a sequence of training samples and is often derived by predetermined heuristics in particular problems. For example, Chen et al. designed a curriculum where images with clean backgrounds are learned before the images with noisy backgrounds , i.e. their method first builds a feature representation by a Convolutional Neural Network (CNN) on images with clean background and then they fine tune the models on images with noisy background. In , the authors approached grammar induction, where the curriculum is derived in terms of the length of a sentence. Because the number of possible solutions grows exponentially with the length of the sentence, and short sentences are easier and thus should be learn earlier. The heuristic knowledge in a problem often proves to be useful. However, the curriculum design may lead to inconsistency between the fixed curriculum and the dynamically learned models. That is, the curriculum is predetermined a prior and cannot be adjusted accordingly, taking into account the feedback about the learner. To alleviate the issue of CL, Kumar et al. designed a learning paradigm, called *self-paced learning* (SPL) . SPL embeds curriculum design as a regularizer into the learning objective. Compared with CL, SPL exhibits two advantages: first, it jointly optimizes the learning objective with the curriculum, and thus the curriculum and the learned model are consistent under the same optimization problem; second, the learning is controlled by a regularizer which is independent of the loss function in specific problems. This theory has been successfully applied to various applications, such as matrix factorization , action/event detection , domain adaption , tracking  and segmentation , reranking , etc. **Learning Detectors in Web Data**: Many recent studies have been proposed to utilize the large amount of noisy data from the Internet. For example, proposed a Never-Ending Language Learning (NELL) paradigm and built adaptive learners that makes use of the web data by learning different types of knowledge and beliefs continuously. Such learning process is mostly self-supervised, and previously learned knowledge enables learning further types of knowledge. In the image domain, existing methods try to tackle the problem of constructing qualified training sets based on the search results of text or image search engines . For example, extended the probabilistic Latent Semantic Analysis in visual domain and learned object categories using results from image search engines. proposed an incremental learning paradigm that initialized from a few seed images and repeatedly trained models to refine the collected image dataset from the Internet. NEIL  followed the idea of NELL and learned from web images to form a large collection of concept detectors iteratively via a semi-supervised fashion. By combining the classifiers and the inter-concept relationships it learned, NEIL can be used for scene classification and object detection task. tried to learn robust classifiers by considering the noisy textual information accompanied with web images. However, the down side is that the portion of the true positive samples has to be determined via prior knowledge, where in fact it is not accurate to assume the same number of true positive samples for any targeted concept. introduced a webly-supervised visual concept learning method that automatically learns large amount of models for a wide range of variations within visual concepts. They discovered concept variances through vocabulary of online books, and then downloaded images based on text-search from the web to train object detection and localization models. presented a weakly-supervised method called Baby Learning for object detection from a few training images and videos. They first embed the prior knowledge into a pre-trained CNN. When given very few samples for a new concept, a simple detector is constructed to discover much more training instances from the online weakly labeled videos. As more training samples are selected, the concept detector keeps refining until a mature detector is formed. Another recent work in image domain proposed a webly supervised learning of Convolutional Neural Network. They utilized easy images from search engine like Google to bootstrap a first-stage network and then used noisier images from photo-sharing websites like Flickr to train an enhanced model. In video domain, only few studies have been proposed for noisy data learning since training robust video concept detectors is more challenging than the problem in the image domain. tackled visual event detection problem by using SVM based domain adaptation method in web video data. described a fast automatic video retrieval method using web images. Given a targeted concept, compact representations of web images obtained from search engines like Google, Flickr are calculated and matched to compact features of videos. Such method can be utilized without any pre-defined concepts. discussed a method that exploits the YouTube API to train large scale video concept detectors on YouTube. The method utilized a calibration process and hard negative mining to train a second order mixture of experts model in order to discover correlations within the labels. Most of the existing methods are heuristic approaches as it is unclear what objective is being optimizing on the noisy data. Moreover, results obtained from the web search results is just one approach to acquire prior knowledge or curriculum. To the best of our knowledge, there have been no systematical studies on exploiting the multi-modal prior knowledge in video concept learning on noisy data. Since search engine algorithm is changing rapidly, it is unclear that how noisy the web labels are and how the level of noisiness in the data will affect performance. In this paper, we proposed a theoretically justified method with clear framework for curriculum constructing and model learning. We also empirically demonstrate its superior performance over representative existing methods and systemically verify that WELL is robust against the level of noisiness of the video data. # WEbly-Labeled Learning (WELL) ## Problem Description In this paper, following , we consider a concept detector as a classifier and our goal is to train concept detectors from webly-labeled video data without any manual labeling effort. Given a noisy web video training set and a target concept set, we do not assume any distribution of the noise. Formally, we represent the training set as \(\mathcal{D} = \{(\mathbf{x}_i,\mathbf{z}_i,\mathbf{\tilde{y}}_i)\}_{i=1}^n\) where \(\mathbf{x}_i \in \mathbb{R}^m\) denotes the feature for the \(i^{th}\) observed sample, and \(\mathbf{z}_i\) represents its noisy web label, which generally means the prior knowledge we can get from the web without additional human effort that not only includes textual information provided by the uploaders in the video metadata but also includes prior knowledge from other modalities using existing tools like pre-trained Convolutional Neural Network image detector, Automatic Speech Recognition and Optical character recognition. The \(\mathbf{\tilde{y}}_i \subset \mathcal{Y}\) is the inferred concept label set for the \(i^{th}\) observed sample based on its noisy web label, and \(\mathcal{Y}\) denotes the full set of target concepts. In our experiment, to simplify the problem, we apply our method on binary classification and infer binary labels \(\tilde{y}_i\) from the noisy web labels. The noisy web labels can be used to automatically infer concept labels by matching the concept name to the video textual metadata. For example, a video may be inferred to the concept label "cat" as its textual title contains cat. utilizes the YouTube topic API, which is derived from the textual metadata, to automatically get concept labels for videos. The web labels are quite noisy as the webly-labeled concepts may not present in the video content whereas the concepts not in the web label may well appear. ## Model and Algorithm ### Objective Function To leverage the noisy web labels in a principled way, we propose WEbly-Labeled Learning (WELL). Formally, given a training set \(\mathcal{D}\) as described before, Let \(L(\tilde{y}_i,g(\mathbf{x}_i,\mathbf{w}))\), or \(\ell_i\) for short, denote the loss function which calculates the cost between the inferred label \(\tilde{y}_i\) and the estimated label \(g(\mathbf{x}_i,\mathbf{w})\). Here \(\mathbf{w}\) represents the model parameter inside the decision function \(g\). For example, in our paper, \(\mathbf{w}\) represents the weight parameters in the Convolutional Neural Network (CNN) and the Support Vector Machine (SVM). Our objective function is to jointly learn the model parameter \(\mathbf{w}\) and the latent weight variable \(\mathbf{v}= [v_1,\cdots,v_n]^T\) by: \[\label{eq:spcl_obj} \begin{split} \!\min_{\mathbf{w},\mathbf{v}\in \lbrack 0,\!1]^{n}}\!\!\mathbb{E}(\mathbf{w},\!\mathbf{v}\!;\lambda,\!\Psi\!) \!=\! \sum_{i=1}^n v_i L(\tilde{y}_i,\!g(\mathbf{x}_i,\!\mathbf{w})) \!+\! f(\mathbf{v};\! \lambda), \\ \text{ subject to } \mathbf{v} \in \Psi \end{split}\] where \(\mathbf{v=[}v_{1},v_{2},\cdots ,v_{n}\mathbf{]}^{T}\) denote the latent weight variables reflecting the inferred labels' confidence. The weights determine a learning sequence of samples, where samples with greater weights tend to be learned earlier. Our goal is to assign greater weights to the samples with more confident labels whereas smaller or zero weights to the samples with noisy labels. To this end, we employ the self-paced regularizer \(f\), which controls the learning process of the model. We consider the linear regularizer Eq. [\[eq:linear_scheme\]](#eq:linear_scheme){reference-type="eqref" reference="eq:linear_scheme"} proposed in : \[\label{eq:linear_scheme} f(\mathbf{v};\lambda) = \frac{1}{2} \lambda \sum_{i=1}^n ( v_i^2-2v_i ). \vspace{-1mm}\] Generally, a self-paced regularizer determines the scheme for penalizing the latent weight variables. Physically it resembles the learning schemes human used in understanding new concepts. The linear scheme corresponds to a prudent strategy, which linearly penalizes the samples that are different to what the model has already learned (see Eq. [\[eq:linear_closedform\]](#eq:linear_closedform){reference-type="eqref" reference="eq:linear_closedform"}). The hyper-parameter \(\lambda\) \((\lambda > 0)\) is called "model age", which controls the pace at which the model learns new samples. When \(\lambda\) is small only samples of with small loss will be considered. As \(\lambda\) grows, more samples with larger loss will be gradually appended to train a "mature" mode. \(\Psi\) in Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"} is a curriculum region derived from noisy web labels \(\mathbf{z}\) that incorporates the prior knowledge extracted from the webly-labeled data as a convex feasible region for the weight variables. The shape of the region weakly implies a prior learning sequence of samples, where the expected values for favored samples are larger. The curriculum region can be derived in a variety of ways that make use of different modalities. We will discuss this topic in details in following section. A straightforward approach is by counting the term frequency in the video's textual metadata. That is, for example, the chance of a video containing the concept "cat" become higher when it has more word "cat" in its title, description or tags. Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"} represents a concise and general optimization model  . It combines the prior knowledge extracted from the noisy webly-labeled data (as the curriculum region) and the information dynamically learned during the training (via the self-paced regularizer). Intuitively, the prior knowledge serves as an instructor providing a guidance on learning the latent weights, but it leaves certain freedom for the model (the student) to adjust the actual weights according to its learning pace. Experimental results in Section [9](#sec:experiments){reference-type="ref" reference="sec:experiments"} demonstrate the learning paradigm can better overcome the noisy labels than heuristic approaches. Figure [\[well-cnn\]](#well-cnn){reference-type="ref" reference="well-cnn"} shows the learning process of our method. Following , we employ the alternative convex search algorithm to solve Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"}. Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} takes the input of a curriculum region, an instantiated self-paced regularizer and a step size parameter; it outputs an optimal model parameter \(\mathbf{w}\). First of all, it initializes the latent weight variables in the feasible region. Then it alternates between two steps until it finally converges: Step 3 learns the optimal model parameter with the fixed and most recent \(\mathbf{v}^*\); Step 5 learns the optimal weight variables with the fixed \(\mathbf{w}^*\). In the beginning, the model "age" is gradually increased so that more noisy samples will be gradually incorporated in the training. Step 3 can be conveniently implemented by existing off-the-shelf supervised learning methods such as the back propagation. Gradient-based methods can be used to solve the convex optimization problem in Step 4. According to , the alternative search in Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} converges as the objective function is monotonically decreasing and is bounded from below. At an early age when \(\lambda\) is small, Step 4 in Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} has an evident suppressing effect over noisy samples that have greater loss to the already learned model. For example, with a fixed \(\mathbf{w}\), the unconstrained close-formed solution for the regularizer in Eq. [\[eq:linear_scheme\]](#eq:linear_scheme){reference-type="eqref" reference="eq:linear_scheme"} equals \[\label{eq:linear_closedform} v_i^* =\begin{cases}-\frac{1}{\lambda} \ell_i +1 & \ell_{i} < \lambda\\ 0 & \ell_{i} \ge \lambda \end{cases},\] where \(v_i\) represents the \(i\)th element in the optimal solution \(\mathbf{v}^* = [v_1^*, \cdots, v_n^*]^T\). Eq. [\[eq:linear_closedform\]](#eq:linear_closedform){reference-type="eqref" reference="eq:linear_closedform"} called linear regularizer indicates the latent weight is proportional to the negative sample loss, and the sample whose loss is greater or equals to \(\lambda\) will have zero weights and thus will not affect the training of the next model. As the model age grows, the hyper-parameter \(\lambda\) increases, and more noisy samples will be used into training. The prior knowledge embedded in the curriculum region \(\Psi\) is useful as it suggests a learning sequence of samples for the "immature" model. theoretically proves that the iterative learning process is identical to optimizing a robust loss function on the noisy data. If we keep increasing \(\lambda\), the model will ultimately use every sample in the noisy data, which is undesirable as the labels of some noisy samples are bound to be incorrect. To this end, we stop increasing the age \(\lambda\) after about a certain number of iterations (early stopping). The exact stopping iteration for each detector is automatically tuned in terms of its performance on a small validation set. ### Model Details {#sec:partial_curriculum} In this section we discuss further details of the curriculum region \(\Psi\) and the self-paced regularizer \(f(\mathbf{v};\lambda)\). \(\Psi\) is a feasible region that embeds the prior knowledge extracted from the webly-labeled data. It physically corresponds to a convex search region for the latent weight variable. Given a set of training samples \(\mathbf{X}=\{\mathbf{x}_{i}\}_{i=1}^{n}\), we utilize the partial-order curriculum which generalizes the total-order curriculum by incorporating the incomplete prior over groups of samples. Samples in the confident groups should be learned earlier than samples in the less confident groups. It imposes no prior over the samples within the same group nor the samples not in any group. Formally, we define a partial order relation \(\preceq\) such that \(x_i \preceq x_j\) indicates that the sample \(x_i\) should be learned no later than \(x_j\) (\(i,j \in [1,n]\)). Similarly given two sample subsets \(\mathbf{X}_{a} \preceq \mathbf{X}_{b}\) denotes the samples in \(\mathbf{X}_{a}\) should be learned no later than the samples in \(\mathbf{X}_{b}\). In our problem, we extract the partial-order curriculum in the webly-labeled data in the following ways: we only distinguish the training order for groups of samples. Information from different modalities of the web labels can be used for curriculum design. A straightforward way is to directly utilize the textual descriptions of the videos generated by the uploaders. We compare common ways to extract curriculum from web data for concept learning to the proposed novel method that utilize state-of-the-art topic modeling techniques in natural language processing. In the following methods (Exact & Stem Matching, Word Embedding and Latent Topic with Word Embedding), we first extract bag-of-words features from different modalities and then match them using specific matching methods to the concept words. Each video will then come with a matching score to each concept. In our experiment, we divide the data into two partial-order curriculum groups, where the videos with matching scores larger than zero will be in one group while others will be in the other group. **Exact & Stem Matching** We build curriculum directly using exact word matching or stemmed word matching between the textual metadata of the noisy videos to the targeted concept names. **YouTubeTopicAPI** We directly utilize the YouTube topic API to search for videos that are related to the concept words. The topic API utilizes textual information of the uploaded videos to obtain related topics of the videos from Freebase. **SearchEngine** We build curriculum using the search result from a text-based search engine. It is similar to related web-search based methods. **Word Embedding** We use word embedding to match words in metadata to targeted concept words in order to deal with synonyms and related concepts. The word embedding is trained using Google News data. **Latent Topic** We build curriculum based on the latent topic we learned from the noisy label. We incorporate Latent Dirichlet Allocation (LDA)  to determine how each noisy labeled video is related to each target concept. The basic idea is that each web video consists of mixtures of topics (concepts), and each topic is characterized by a distribution of words. Formally, given all the noisy information extracted from a web video and collected them as a document \(\mathbf{d_i}\), which combines into a corpus \(\mathbf{d}\), we have a target set of \(\mathbf{k}\) topics, then the key inferential problem that we are going to solve is that of computing the posterior distribution of the latent topics given a corpus (how likely the videos are related to each target concept given the noisy information): \[\label{eq:lda} p(\theta,\mathbf{t}|\mathbf{d},\alpha,\beta) = \frac{p(\theta,\mathbf{t},\mathbf{d}|\alpha,\beta)}{p(\mathbf{d}|\alpha,\beta)}\] where \(\theta\) is the topic distribution variable for the corpus, \(\theta\) \(\sim\) Dir(\(\alpha\)), in which Dir(\(\alpha\)) represents a uniform Dirichlet distribution with scaling parameter \(\alpha\). The \(\mathbf{t}\) is the topic assignment variable that indicates which topic each word belong to in the document. \(\beta\) is the Dirichlet prior on the per-topic word distribution, in which we impose asymmetric priors over the word distribution so that each learned topic will be seeded with particular words in our target concept. For example, a topic will be seeded with words \"walk, dog\" for the target concept \"WalkingWithDog\". The parameter estimation in Eq [\[eq:lda\]](#eq:lda){reference-type="eqref" reference="eq:lda"} can be done via Bayes methods. However, Eq. [\[eq:lda\]](#eq:lda){reference-type="eqref" reference="eq:lda"} is intractable to compute since \(p(\mathbf{d}|\alpha,\beta)\) is intractable due to the coupling between \(\theta\) and \(\beta\). To solve this problem, approximate inference algorithms are introduced and we use the online variational inference algorithm from. The true posterior is approximated by a simpler distribution: \[\label{eq:lda-vi} q(\theta,\mathbf{t}|\gamma,\phi) = q(\theta|\gamma)\prod_{i=1}^{N}{q(t_i|\phi_i)}\] where the Dirichlet parameter \(\gamma\) and the multinomial parameters (\(\phi_1...,\phi_N\)) are the free variational parameters. Thus the maximization problem is equivalent to minimizing the Kullback-Leibler(KL) divergence between \(q(\theta,\mathbf{t}|\gamma,\phi)\) and the posterior \(p(\theta,\mathbf{t}|\mathbf{d},\alpha,\beta)\). The optimization problem can then be solve using Expectation-Maximization algorithm. The estimated parameters \(\theta,t\) in Eq. [\[eq:lda\]](#eq:lda){reference-type="eqref" reference="eq:lda"} is then used for constructing curriculum region in Eq. [\[eq:spcl_obj\]](#eq:spcl_obj){reference-type="eqref" reference="eq:spcl_obj"}. **Latent Topic with Word Embedding (LT+WE)** We first learn latent topics using LDA to replace the concept words with a topic word distribution and then match the latent topic words to the web label's bag-of-words features by using the word embeddings. We compare this method to the others using only textual information from the web videos. We also use this method to get curriculum from other modalities such as Automatic Speech Recognition (ASR), Optical Character Recognition (OCR) and basic image detector pre-trained on still images  (in this paper we use VGG net , extract keyframe-level image classification results and average them to get video-level results.). We extract bag-of-word features from them and combine them with linear weights. Detailed fusion experiments can be found in Section [9](#sec:experiments){reference-type="ref" reference="sec:experiments"}. We empirically set OCR's weight to be small as the results are much noisier than other features. Figure [\[curriculum_example\]](#curriculum_example){reference-type="ref" reference="curriculum_example"} shows an example of the noisy web video data and how the curriculum is extracted with different methods. Our method can utilize information from different modalities while common methods like search engine only consider textual information. We compare the performance of different ways of curriculum design by training detectors directly in Section 4. The labels in webly-labeled data are much noisier than manually-labeled data, and as a result, we found that the learning is prone to overfitting the noisy labels. To address this issue, inspired by the dropout technique in deep learning , we use a dropout strategy for webly-labeled learning . It is implemented in the self-paced regularizer discussed in Section 3. With the dropout, the regularizers become: \[\label{eq:dropout} \vspace{-2mm} \begin{split} r_i(p) \sim \text{Bernoulli}(p) + \epsilon, (0 < \epsilon \ll 1)\\ f(\mathbf{v};\lambda, p) = \frac{1}{2} \lambda \sum_{i=1}^n (\frac{1}{r_i} v_i^2-2v_i), \end{split} \vspace{-2mm}\] where \(\mathbf{r}\) is a column vector of independent Bernoulli random variables with the probability \(p\) of being 1. Each of the element equals the addition of \(r_i\) and a small positive constant \(\epsilon\). Denote \(\mathbb{E}_{\mathbf{w}} = \sum_{i=1}^n v_i \ell_i + f(\mathbf{v};\lambda)\) as the objective with the fixed model parameters \(\mathbf{w}\) without any constraint, and the optimal solution \(\mathbf{v}^* = [v_1^*, \cdots, v_n^*]^T= \argmin_{\mathbf{v}\in[0,1]^n} \mathbb{E}_{\mathbf{w}}\). We have: \[\vspace{-2mm} \label{eq:linear_dropout} \begin{split} \mathbb{E}_{\mathbf{w}} = \sum_{i=1}^n \ell_i v_i + \lambda ( \frac{1}{2r_i} v_i^2-v_i );\\ \frac{\partial \mathbb{E}_{\mathbf{w}}}{\partial v_i} = \ell + \lambda v_i/r_i -\lambda = 0;\\ \Rightarrow v_i^* =\begin{cases} r_i(-\frac{1}{\lambda} \ell_i +1) & \ell_{i} < \lambda\\ 0 & \ell_{i} \ge \lambda \end{cases}. \end{split} \vspace{-2mm}\] The dropout effect can be demonstrated in the closed-form solutions in Eq. [\[eq:linear_dropout\]](#eq:linear_dropout){reference-type="eqref" reference="eq:linear_dropout"}: with the probability \(1-p\), \(v_i^*\) in both the equations approaches 0; with the probability \(p\), \(v_i^*\) approaches the solution of the plain regularizer discussed in Eq. [\[eq:linear_scheme\]](#eq:linear_scheme){reference-type="eqref" reference="eq:linear_scheme"}. Recall the self-paced regularizer defines a scheme for learning samples. Eq. [\[eq:linear_dropout\]](#eq:linear_dropout){reference-type="eqref" reference="eq:linear_dropout"} represent the new dropout learning scheme. When the base learner is neural networks, the proposed dropout can be used combined with the classical dropout in . The term dropout in this paper refers to dropping out samples in the iterative learning. By dropping out a sample, we drop out its update to the model parameter, which resembles the classical dropout used in neural networks. It operates on a more coarse-level which is useful for noisy data. When samples with incorrect noisy labels update a model, it will encourage the model to select more noisy labels. The dropout strategy prevents overfitting to noisy labels. It provides a way of combining many different sample subsets in different iterations in order to help avoid bad local minima. Experimental results substantiate this argument. In practice, we recommend setting two Bernoulli parameters for positive and negative samples on imbalanced data. # Experiments {#sec:experiments} In this section, we evaluate our method WELL for learning video detectors on noisy labeled data. The experiments are conducted on two major public benchmarks: FCVID and YFCC100M, where FCVID is by far one of the biggest manually annotated video dataset, and the YFCC100M dataset is the largest multimedia benchmark. ## Experimental Setup **Datasets, Features and Evaluation Metrics** Fudan-columbia Video Dataset (FCVID) contains 91,223 YouTube videos (4,232 hours) from 239 categories. It covers a wide range of concepts like activities, objects, scenes, sports, DIY, etc. Detailed descriptions of the benchmark can be found in . Each video is manually labeled to one or more categories. In our experiments, we do not use the manual labels in training, but instead we automatically generate the web labels according to the concept name appearance in the video metadata. The manual labels are used only in testing to evaluate our and the baseline methods. Following , the standard train/test split is used. The second set is YFCC100M which contains about 800,000 videos on Yahoo! Flickr with metadata such as the title, tags, the uploader, etc. There are no manual labels on this set and we automatically generate the curriculum from the metadata in a similar way. Since there are no annotations, we train the concept detectors on the most 101 frequent latent topics found in the metadata. There are totally 47,397 webly labeled videos on the 101 concepts for training. On FCVID, as the manual labels are available, the performance is evaluated in terms of the precision of the top 5 and 10 ranked videos (P\@5 and P\@10) and mean Average Precision (mAP) of 239 concepts. On YFCC100M, since there are no manual labels, for evaluation, we apply the detectors to a third public video collection called TRECVID MED which includes 32,000 Internet videos . We apply the detectors trained on YFCC100M to the TRECVID videos and manually annotate the top 10 detected videos returned by each method for 101 concepts. **Implementation Details** We build our method on top of a pre-trained convolutional neural network as the low-level features (VGG network ). We extract the key-frame level features and create a video feature by the average pooling. The same features are used across different methods on each dataset. The concept detectors are trained based on a hinge loss cost function. Algorithm [\[alg:overall\]](#alg:overall){reference-type="ref" reference="alg:overall"} is used to train the concept models iteratively and the \(\lambda\) stops increasing after 100 iterations. We automatically generate curriculum labels based on the video metadata, ASR, OCR and VGG net 1,000 classification results using latent topic modeling with word embedding matching as shown in Section 3, and derive a partial-order curriculum using the method discussed in Section 3.2.2. **Baselines** The proposed method is compared against the following five baseline methods which cover both the classical and the recent representative learning algorithms on webly-labeled data. *BatchTrain* trains a single SVM model using all samples in the multi-modal curriculum built as described in section 3.2.2 LT+WE. *Self-Paced Learning (SPL)* is a classical method where the curriculum is generated by the learner itself . *BabyLearning* is a recent method that simulates baby learning by starting with few training samples and fine-tuning using more weakly labeled videos crawled from the search engine. *GoogleHNM* is a hard negative mining method proposed by Google. It utilizes hard negative mining to train a second order mixture of experts model according to the video's YouTube topics. *FastImage* is a video retrieval method that utilizes web images from search engine to match to the video with re-ranking. *WELL* is the proposed method. The hyper-parameters of all methods including the baseline methods are tuned on the same validation set. On FCVID, the set is a standard development set with manual labels randomly selected from 10% of the training set (No training was done using ground truth labels) whereas on YFCC100M it is also a 10% proportion of noisy training set. ## Experiments on FCVID **Curriculum Comparison** As disscussed in Section 3.2.2, we compare different ways to build curriculum for noisy label learning. Here we also compare their effectiveness by training concept detectors directly using the curriculum labels. The batch train model is used for all generated cirriculumn labels. In Table [9](#exp-cur){reference-type="ref" reference="exp-cur"} we show the batch trained models' precision at 5, 10 and mean average precision on the test set of FCVID. For LT+WE (Multi-modal), we extract curriculum from different modalities as shown in Section 3.2.2, and combine them using linear weights. The weights are hyper-parameters that are tuned on the validation set, and the optimal weights for textual metadata, ASR, image classification and OCR results are 1.0, 0.5, 0.5 and 0.05, respectively. Results show that the curriculum generated by combining latent topic modeling and word embedding using multi-modal prior knowledge is the most accurate, which indicates our claim of exploiting multi-modal information is beneficial, and we use this method in WELL for the rest of the experiments. **Baseline Comparison** Table [10](#exps-basline){reference-type="ref" reference="exps-basline"} compares the precision and mAP of different methods where the best results are highlighted. As we see, the proposed WELL significantly outperforms all baseline methods, with statistically significant difference at \(p\)-level of 0.05. Comparing WELL with SPL, the effect of curriculum learning and dropout makes a significant difference in terms of performance, which suggests the importance of prior knowledge and preventing over-fitting in webly learning. The promising experimental results substantiate the efficacy of the proposed method. **Robustness to Noise Comparison** In this comparison we manually control the noisiness of the curriculum in order to systematically verify how our methods would perform with respect to the noisiness within the web data. The experimental results indicate the robustness of our method towards noisy labels. To this end, we randomly select video samples with ground truth labels for each concept, so that the precision of the curriculum labels are set at 20%, 40%, 60%, 80% and we fix the recall of all the labels. We then train WELL using such curriculum and test them on the FCVID testing set. We also compare WELL to three other methods with the same curriculum, among them *GoogleHNM* is a recent method to train video concept detector with large-scale data. We exclude *BabyLearning*, which relies on the returned results by the search engine, since in this experiment the curriculum is fixed. As shown in Table [12](#exp-noise){reference-type="ref" reference="exp-noise"}, as the noisiness of the curriculum grows (the precision drops), WELL maintains its performance while other methods drop significantly. Specifically, when the precision of the curriculum drops from 40% to 20%, other methods' mAP averagely drops 46.5% while WELL's mAP only drops 19.1% relatively. It shows that WELL is robust against different level of noise, which shows great potential in larger scale webly-labeled learning as the dataset gets bigger, the noisier it may become. [\[exp-noise\]]{#exp-noise label="exp-noise"} [\[exp-noise\]]{#exp-noise label="exp-noise"} **Ground-truth Training Comparison** In this part, we also compare our method with the state-of-the-art method trained using ground truth labels on FCVID (rDNN) . We compare WELL trained using the static CNN features, the standard features provided by the authors , and we also compare WELL using the late (average) fusion with CNN, motion and audio MFCC features (WELL-MM) to the method that achieves the best result on FCVID trained using the same multi-modal features. WELL-MM uses CNN, Motion and MFCC features, which is the same set of features as rDNN-F . Noted that the state-of-the-art method uses the ground truth labels to train models, which includes 42,223 videos with manual labels, while our proposed method uses none of the human annotation into training but still be able to outperform one of the state-of-the-art results. [\[exps-sa\]]{#exps-sa label="exps-sa"} **Noisy Dataset Size Comparison** To investigate the potential of concept learning on webly-labeled video data, we apply the methods on different sizes of subsets of the data. Specifically, we randomly split the FCVID training set into several subsets of 200, 500, 1,000, and 2,000 hours of videos, and train the models on each subset without using manual annotations. The models are then tested on the same test set. Table [14](#exps-small){reference-type="ref" reference="exps-small"} lists the average results of each type of subsets. As we see, the accuracy of WELL on webly-labeled data increases along with the growth of the size of noisy data while other webly learning methods' performance tend to be saturated. Comparing to the methods trained using ground truth, In Table [14](#exps-small){reference-type="ref" reference="exps-small"}, WELL-MM trained using the whole dataset (2000h) outperforms rDNN-F(trained using manual labels) trained using around 1200h of data. And since the incremental performance increase of WELL-MM is close to linear, we conclude that with sufficient webly-labeled videos WELL-MM will be able to outperform the rDNN-F trained using 2000h of data, which is currently the largest manual labeled dataset. [\[exps-small\]]{#exps-small label="exps-small"} ## Experiments on YFCC100M In the experiments on YFCC100M, we train 101 concept detectors on YFCC100M and test them on the TRECVID MED dataset which includes 32,000 Internet videos. Since there are no manual labels, to evaluate the performance, we manually annotate the top 10 videos in the test set and report their precisions in Table [15](#exps-yfcc){reference-type="ref" reference="exps-yfcc"}. The MED evaluation is done by four annotators and the final results are averaged from all annotations. The Fleiss' Kappa value for these four annotators is 0.64. A similar pattern can be observed where the comparisons substantiate the rationality of the proposed webly learning framework. Besides, the promising results on the largest multimedia set YFCC100M verify the scalability of the proposed method. ## Time Complexity Comparison The computation complexity of WELL is comparable to existing methods. For a single class, the complexity for our model and baseline models is \(O(r \times n \times m)\), where \(r\) is the number of iterations to converge, \(n\) is the number of training samples and \(m\) is the feature dimension. Theoretically, the complexity is comparable to the baseline models that have different \(r\). In practice, on a 40 core-CPU machine, WELL and SPL takes 7 hours to converge (100 iterations) on FCVID with 239 concepts, whereas GoogleHNM and BabyLearning take around 5 hours. In Table [16](#time){reference-type="ref" reference="time"}, we show the theoretical and actual run time for all methods. ## Qualitative Analysis In this section we show training examples of WELL. In Figure [\[well-pick-all\]](#well-pick-all){reference-type="ref" reference="well-pick-all"}, we demonstrate the positive samples that WELL select at different stage of training the concept \"baseball\", \"forest\", \"birthday\" and \"cow\". For the concept \"baseball\", at early stage (1/93, 25/93), WELL selects easier and clearer samples such as those camera directly pointing at the playground, while at later stage (75/93, 93/93) WELL starts to train with harder samples with different lighting conditions and untypical samples for the concept. For the concept \"birthday\", as we see, at later stage of the training, complex samples for birthday event like a video with two girl singing birthday song (75/84) and a video of celebrating birthday during hiking (84/84) are included in the training. For the concept \"forest\", at the final iteration (95/95), a video of a man playing nunchaku is included, as the video title contains \"Air Forester Nunchaku Freestyle\" and it is included in the curriculum, which is reasonable as the curriculum is noisy and may contain false positive. Since WELL is able to leave outliers in later stage of the training, the affection of the false positives in the curriculum can be alleviated by early stopping. In our experiments, we stop increasing \(\lambda\) after 100 iterations so in the final model, only a subset of samples are used. # Conclusions In this paper, we proposed a novel method called WELL for webly labeled video data learning. WELL extracts multi-modal informative knowledge from noisy weakly labeled video data from the web through a general framework with solid theoretical justifications. WELL achieves the best performance only using webly-labeled data on two major video datasets. The comprehensive experimental results demonstrate that WELL outperforms state-of-the-art studies by a statically significant margin on learning concepts from noisy web video data. In addition, the results also verify that WELL is robust to the level of noisiness in the video data. The result suggests that with more webly-labeled data, which is not hard to obtain, WELL can potentially outperform models trained on any existing manually-labeled data.

{'timestamp': '2016-07-19T02:05:10', 'yymm': '1607', 'arxiv_id': '1607.04780', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04780'}

# Introduction {#introduction .unnumbered} The minimum energy required to reset one bit of information represents one of the fundamental limits of computation arising when one bit of information is erased or destroyed. Starting with a system encoding two possible states with the same probability and finishing the procedure with only one possible state, the variation of entropy in the system is equal to the difference of entropy between the final state and the initial one, \(\Delta S = k_\textnormal{B} \ln 1-k_\textnormal{B} \ln 2 =-k_\textnormal{B} \ln 2\). On average this reduction of entropy has to be accompanied with an increase of heat in the surrounding environment in order to not violate the second law of thermodynamics, \(Q_\textnormal{L} \geq k_\textnormal{B} T \ln 2\). This limit takes the name of Landauer limit and was theorized by Rolf Landauer in the 60's, and remained untested for over fifty years. The development of computational methods allowed to evaluate tiny amounts of heat exchanged and recent technical advances in micro-and nano-fabrication made possible to recently test the validity of the Landauer principle. In particular, the first experimental verification of the Landauer principle has been carried out by Bérut using a colloidal particle trapped in optical tweezers. More recently Jun presented similar results considering a colloidal particle in a feedback trap. Finally, the last experimental verification of the Landauer limit has been performed by Hong in a nanomagnetic system. In this case information is encoded in the magnetization state of the system while an external magnetic field is used in order to flip between two preferred magnetization states. The evaluation of the heat produced during the reset operation was demonstrated to be compatible with the Landauer limit. Even if there is no doubt nowadays on the validity of the Landauer principle, the test on micro-mechanical systems is still missing. We have recently shown that it is possible to use electro-mechanical devices to accomplish basic and complex logic operations with an arbitrarily low energy expenditure. Therefore the possibility to use this class of devices for memory storage completes the logic architecture for a complete computing device. In the following we show the measurement on heat production when performing the reset operation in a novel memory unit based on a bistable mechanical cantilever at effective temperature. The results are in good agreement with the Landauer limit. The dependence of the dissipation with the error rate is also investigated showing a trend in accordance with the theory. # Results {#results .unnumbered} ## Bistable micro-mechanical system {#bistable-micro-mechanical-system .unnumbered} The mechanical system used to perform the experiment is depicted in Figure [\[f:schematic\]](#f:schematic){reference-type="ref" reference="f:schematic"} (a). A triangular micro-cantilever, 200 μm long, is used to encode one bit on information. In order to obtain two stable states two magnets with opposite magnetization are placed on the tip of the cantilever and on a movable stage facing the cantilever. In this way, depending on the distance between the magnets, \(d\), and the relative lateral alignment, \(\Delta x\), it is possible to induce bistability on the system. Figure [\[f:schematic\]](#f:schematic){reference-type="ref" reference="f:schematic"} (b) shows the potential energy as a function of \(d\) reconstructed from the probability density function of the position of the cantilever at equilibrium, \(\rho (x,d) = A \exp(-U(x,d)/k_\textnormal{B} T)\), which implies that \(U(x,d)=-k_\textnormal{B} T \ln \rho (x,d) + U_\textnormal{0}\). When the magnets are far away the effect of the repulsive force is negligible, the system is then monostable and can be approximated to a linear system. Decreasing the distance the repulsive force between magnets tends to soften the system up to the point where two stable positions appear. The effect of reducing even more \(d\) is to enlarge the separation of the rest states and to increase the potential barrier separating these two wells. Eventually, when the distance between the magnets is small enough, the system remains trapped in one well for a period of time larger than the relaxation time of the system. Logic states are encoded in the position of the cantilever tip: logic 0 for \(x<0\) and logic 1 for \(x>0\). The proposed system presents intrinsic dissipative processes that depend on the maximum displacement of the cantilever tip. The minimum heat produced when performing a physical transformation of the system is proportional to \(x_\textnormal{max}^2\). In our setup it is not possible to reduce the separation between the two potential wells to a value that bounds the heat produced by intrinsic dissipation below the Landauer limit. Increasing the effective temperature increases the value for the Landauer limit making possible to have negligible intrinsic dissipation. A piezoelectric shaker is used to excite the structure with a band limited white Gaussian noise to mimic the effect of an arbitrary temperature. In the present experiment the white noise is limited to 50 kHz, well above the resonance frequency of the free cantilever (\(f_\textnormal{0}\)=5.3 kHz). The dependence of the effective temperature with the root-mean-squared voltage supplied to the shaker is reported in Figure [\[f:schematic\]](#f:schematic){reference-type="ref" reference="f:schematic"} (c). The red dot, corresponding to an effective temperature of \(T_\textnormal{eff}\) = 5 × 10^7^ K, highlights the condition considered in the present case. The solid line represent the expected trend where \(T \propto V_\textnormal{rms}^2\). The effective temperature has been estimated computing the power spectral density (PSD) of the system at various piezoelectric noise excitation voltages. The obtained curves have been fitted with Lorenzian curves taking as reference for the calibration the one at room temperature (\(T\)=300 K). The other curves have been used to extract the only varying parameter \(T_\textnormal{eff}\), corresponding to the effective temperature of the system under external excitation. Finally two electrostatic probes, placed one on the left and the other on the right of the cantilever, are used to apply a negative and positive forces respectively. When a voltage different to zero is applied on one probe the cantilever feels an attractive electrostatic force toward the probe due to the polarization of the cantilever itself. The voltage on the probes, the distance between the magnets and their time evolution are used to specify the protocols used in order to change the bit stored in the system as described in the following subsection. ## Reset protocol {#reset-protocol .unnumbered} Varying the magnets distance \(d\) the barrier separating the two stable wells varies from the minimum value \(B_\textnormal{0}\) for \(d\)=3.65 a.u. to \(B_\textnormal{max}\) for \(d\)=2.8 a.u. Applying a voltage on one probe corresponds to apply a force toward the probe itself. Thus a voltage on the left probe corresponds to a negative force and a voltage on the right probe corresponds to a positive force. Details of the force calibration are presented in methods section. In Figure [\[f:protocol\]](#f:protocol){reference-type="ref" reference="f:protocol"} (a) the protocol followed to reset the bit to the state 0 and 1 is presented. The procedure is similar to the ones presented in Refs. and . Initially the barrier separating the two stable states is removed moving the magnet away (red curve in the first panel) making the system monostable. Once the barrier is removed we apply a negative (positive) force to reset the bit status to 0 (1) applying a finite voltage \(V_\textnormal{L}\) (\(V_\textnormal{R}\)) on the left (right) probe. This is represented by the magenta curve in Figure [\[f:protocol\]](#f:protocol){reference-type="ref" reference="f:protocol"} (a). Once the force is applied we restore the barrier to its original value. Finally, we remove the lateral force finishing in the original parameters configuration. At the end of the operation, if there are no errors the cantilever position encodes the desired bit of information. In order to be sure to perform the operation starting from both initial states, we mimic a statistical ensemble where the initial probability is 50% to start in the left well and 50% in the right well. A trace of the cantilever tip position, \(x\), is shown in Figure [\[f:protocol\]](#f:protocol){reference-type="ref" reference="f:protocol"} (a) (black line), where the dashed blue lines represent the two stable positions once the barrier is restored. When the barrier is removed the system goes from the local prepared state to an undefined state with a free expansion, the entropy on the system thus increases in a irreversible manner. This increment is related to uncontrollable transitions from one well to the other once the barrier height is close to \(k_\textnormal{B} T\). These large excursions of the cantilever can be seen in the time series of position. Figure [\[f:protocol\]](#f:protocol){reference-type="ref" reference="f:protocol"} (b) shows a representation of the time evolution of the potential energy during the set of the bit to the logic state 1. In a first step the barrier is removed allowing the system to oscillate in a monostable potential landscape. Then the potential is slightly tiled and when the barrier is restored the system is confined in the desired state. After these stages the bit is set to the state 1. In order to have a reliable measure of the heat produced during the considered operations, reset protocols are repeated for 800 times in order to have a large enough statistic. ## Heat vs error probability {#heat-vs-error-probability .unnumbered} In the optimal case the initial configuration is a mixed logical 0 and 1 where both states have the same probability while the final configuration is the selected state with a \(100\%\) probability. This corresponds to an entropy variation of \(\Delta S =-k_\textnormal{B} \ln(2)\) and a minimum heat produced of \(Q\ge-T\Delta S = Q_\textnormal{L}\). It has been shown that the bound \(Q_\textnormal{L} =-k_\textnormal{B} T \ln(2)\) applies only for symmetric potentials. Considering asymmetries on the system the minimum produced heat can be lowered below \(Q_\textnormal{L}\). In our setup the system is slightly asymmetric and we have evaluated the variation of entropy from the initial to the final state from the probability density function of the tip position, \(\rho(x)\), being \(\Delta S_\textnormal{G} =-0.61k_\textnormal{B}\) \(\Delta S_\textnormal{S} =-0.68k_\textnormal{B}\) for the Gibbs and Shannon entropy respectively, both close to \(-k_\textnormal{B} \ln(2)\). If we consider the possibility to commit errors during the reset operation the heat produced becomes a function of the probability of success: \[Q(P_\textnormal{s})\ge k_\textnormal{B} T \left[ \ln(2) + P_\textnormal{s} \ln (P_\textnormal{s}) + (1-P_\textnormal{s}) \ln (1-P_\textnormal{s}) \right] \label{e:Q_ps}\] where \(P_\textnormal{s}\) is the probability of success or success rate. When \(P_\textnormal{s}\) is 0.5 no reset operation is performed and thus there is no minimum heat to be produced during the operation. In Figure [\[f:Q_vs_sr\]](#f:Q_vs_sr){reference-type="ref" reference="f:Q_vs_sr"} (a) we present the average heat produced for the reset operation as function of the lateral alignment of the counter magnet \(\Delta x\). When the system is aligned closely to perfection ( \(\Delta x \approx 0\)) we estimate a heat production slightly above \(k_\textnormal{B} T\) and below two times \(Q_\textnormal{L}\). Asymmetrizing the potential, by means of setting \(\Delta x \neq 0\), the heat produced tends to decrease reaching values this time below \(Q_\textnormal{L}\). However, in this conditions the error rate in performing the reset operation have a major role, in fact in this configuration the probability of success, \(P_\textnormal{s}\), decrease rapidly. This is represented by the color map of dots in Figure [\[f:Q_vs_sr\]](#f:Q_vs_sr){reference-type="ref" reference="f:Q_vs_sr"} (a), where green represents higher success rate while blue represents a higher probability of error. In Figure [\[f:Q_vs_sr\]](#f:Q_vs_sr){reference-type="ref" reference="f:Q_vs_sr"} (b) the success rate of the reset operation is reported as function of the lateral alignment. Solid violet circles represent the overall success rate while red and black symbols represent the error rate for resetting to 1 or to 0 respectively. Circles are used to report the error probability for the same initial and final state while crosses are used for 0 to 1 and 1 to 0 transitions. For instance let us consider the case where \(\Delta x <0\): the counter magnet is moved towards the right and as a consequence the 0 state is more favorable respect the 1 state. From Figure [\[f:Q_vs_sr\]](#f:Q_vs_sr){reference-type="ref" reference="f:Q_vs_sr"} (b) we can see that for \(\Delta x <0\) the probability of resetting toward 0 is almost 100% while the probability of resetting toward 1 decreases rapidly reaching values below 50%. The same behavior is present in the case \(\Delta x > 0\), where the counter magnet is moved to the left, where the state 1 is more favorable. We can now correlate the heat produced to the probability of success for resetting to 0 and 1 as presented in Figure [\[f:Q_vs_sr\]](#f:Q_vs_sr){reference-type="ref" reference="f:Q_vs_sr"} (c). Dashed lines represent the Landauer limit for a 100% of success rate (\(\approx 0.7 k_\textnormal{B} T\)). While in both cases the heat produced is above the Landauer limit, in the reset to 0 case the obtained values are very close to \(Q_\textnormal{L}\). As expected, decreasing the success rate the obtained values goes below the Landauer limit for both cases accordingly to Equation [\[e:Q_ps\]](#e:Q_ps){reference-type="ref" reference="e:Q_ps"}.\ As it is well known the adiabatic limit in presence of dissipation mechanisms like viscous damping can be only reached if the operation is performed slowly when these mechanisms are negligible. We increased the protocol time for the reset operation from 0.25 s up to 3.5 s. The results are presented in Figure [\[f:Q_vs_sr\]](#f:Q_vs_sr){reference-type="ref" reference="f:Q_vs_sr"} (d). Increasing the protocol time decreases the heat production reaching values well below the Landauer limit. Notice that in these cases where \(Q < Q_\textnormal{L}\) the \(P_\textnormal{s}\) is well below 1 since the system has more time to relax and therefore tends to thermalize before the reset operation is correctly performed. In fact for protocols lasting more than 1 s the success rate is below 75%. # Discussion {#discussion .unnumbered} We have measured the intrinsic minimum energy dissipation during the reset of one bit of information in a micro-mechanical system. We have considered a completely different physical system respect to the existing literature, micro-electro-mechanical system. To achieve these results we have performed the experiment at an effective temperature of 5 × 10^7^ K in order to make the intrinsic dissipation of the mechanical structure negligible respect to the thermodynamic contribution. In these conditions we have reached values of heat produces consistent with the Landauer limit approaching it closely. Moreover we presented experimental data relating the minimum heat produced with the probability of success of resetting one bit of information. Nowadays where there is a lot of attention on micro-electro-mechanical systems able to perform computation at arbitrary low energy, the achieved results have a significant importance in the development of new computing paradigms based on systems different from the well established CMOS technology. # Methods {#methods .unnumbered} ## Setup preparation and calibration {#setup-preparation-and-calibration .unnumbered} The micro-cantilever used is a commercial atomic force microscopy (AFM) probe (NanoWorld PNP-TR-TL). It is long 200 μm, with a nominal stiffness \(k\)=0.08 N/m and a nominal resonace frequency of 17 kHz. A fragment of NdFeB (neodymium) magnet is attached to the cantilever tip with bi-component epoxy resin. To set the magnetization to a known direction the system is heated up to 670 K, above its Curie temperature in the presence of a strong external magnetic field with the desired orientation. With this additional mass the resonance frequency decreases to 5.3 kHz. The quality factor of the system has been estimated from the power spectral density of the displacement, \(x\), fitted with a Lorenzian curve, giving a value of \(Q_\textnormal{f}=320\). The deflection of the cantilever, \(x\), is determined by means of an AFM-like laser optical lever. A small bend of the cantilever provokes the deflection of a laser beam incident to the cantilever tip that can be detected with a two quadrants photo detector. The laser beam is focused on the cantilever tip with an optical lens (focal length \(f=\)`<!-- -->`{=html}50 mm). For small deflections the response of the photo detector remains linear, thus \(x=r_\textnormal{x}\Delta V_\textnormal{PD}\), where \(V_\textnormal{PD}\) is the voltage difference generated by the two quadrants of the photo detector. In order to determine \(r_\textnormal{x}\) we look at the frequency response of the system as in under thermal excitation. The relation between the measured voltage and the expected displacement gives \(r_\textnormal{x}=\)`<!-- -->`{=html}1.8365 × 10^−5^ m/V. The system is placed in a vacuum chamber and isolated from seismic vibrations to maximize the signal-to-noise ratio. All measurements were performed at pressure \(P=\)`<!-- -->`{=html}4.7 × 10^−2^ mbar. ## Effective temperature estimation {#effective-temperature-estimation .unnumbered} As the system is modeled as a harmonic oscillator with one degree of freedom, according to the Equipartition Theorem the thermal energy present in the system is simply related to the cantilever fluctuations as \(\frac{1}{2}k_\textnormal{B} T = \frac{1}{2} k \langle x^2\rangle\). According to Parseval's theorem \(\langle x^2\rangle =\int_{0}^{\infty}\! \vert G(\omega) \vert ^2 \, \mathrm{d}\omega\) where \(G(\omega)\) stands for the PSD of the system. This takes the form of a Lorenzian function \[\vert G(\omega) \vert= \sqrt{\frac{4k_\textnormal{B} T}{Q_\textnormal{f} k \omega_\textnormal{0}}} \sqrt{\frac{1}{\left(1-\frac{\omega^2}{\omega_\textnormal{0}^2}\right)^2 + \frac{1}{Q_\textnormal{f}^2}\frac{\omega^2}{\omega_{0}^2}}}\] where \(\omega = 2 \pi f\). Once the system has been calibrated at room temperature we estimated the effective temperatures fitting the only free parameter, \(T\), from the measured PSDs to the expected function \(\vert G(\omega) \vert\). ## Force calibration {#force-calibration .unnumbered} A set of two electrodes (see Fig. [\[f:schematic\]](#f:schematic){reference-type="ref" reference="f:schematic"}(a)) is used in order to polarize the cantilever producing a bend on the mechanical structure. Electrostatic forces depend on the voltage applied to the electrodes, i.e. \(V_\textnormal{L}\) and \(V_\textnormal{R}\) for the left and right electrode respectively. Since the restoring force of the cantilever can be expressed as \(F_\textnormal{k}=-k x\), the relation between applied voltage to the probe and the force acting on the cantilever has been estimated in static conditions assuming \(F_\textnormal{k}=F_\textnormal{el}\). The relation between the applied voltages, \(V_\textnormal{L}\) and \(V_\textnormal{R}\), and the electrostatic force \(F_\textnormal{el}\) is then fitted with a \(9^\textnormal{th}\) degree polynomial. ## Heat production evaluation {#heat-production-evaluation .unnumbered} The work performed on the system along a given trajectory \(x(t)\) is given by the integral: \[W=\int_{0}^{\tau_\textnormal{p}}{\sum_{k=1}^{M} \frac{\partial U(x,\boldsymbol{\lambda})}{\partial \lambda_k} \frac{\partial \lambda_k}{\partial t} \textnormal{d}t}\] where \(\tau_\textnormal{p}\) is the protocol time duration, \(U(x,\boldsymbol{\lambda})\) is the total potential energy of the system and \(\boldsymbol{\lambda}\) is a vector containing all the \(M\) control parameters. In our case we have two controls parameter, the voltage applied to the piezoelectric stage to control the energy barrier, and the electrostatic forces. To obtain the heat produced \(Q\) we have to consider the variation in internal energy, \(\Delta E\). Ideally the potential energy at both the bottom wells is the same, however considering asymmetries on the system \(\Delta E\) can be different from zero. The evaluation of the total energy variation \(\Delta E\) is obtained from the reconstructed potential energy, \(U\), and the variation of the kinetic energy. The latter quantity is however negligible even for the shortest \(\tau_p\), where the total kinetic energy variation is one order of magnitude lower than \(Q_\textnormal{L}\) (1.4 × 10^−17^ J versus 5.3 × 10^−16^ J). Finally the heat produced is obtained from \(Q=W-\Delta E\).

{'timestamp': '2016-09-08T02:03:32', 'yymm': '1607', 'arxiv_id': '1607.04802', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04802'}

# Introduction {#sec-intro} In this paper we study sublanguages of Plotkin's functional programming language \({\mathrm{PCF}}\), which we here take to be the simply typed \(\lambda\)-calculus over a single base type \(\nat\), with constants \[\begin{array}{rlcrl} \num{n} &: \nat \mbox{~~for each \(n \in \mathbb{N}\)} \;, & & {\mathit{suc}},{\mathit{pre}} &: \nat\rightarrow\nat \;, \\ {\mathit{ifzero}} &: \nat\rightarrow\nat\rightarrow\nat\rightarrow\nat \;, & & Y_\sigma &: (\sigma\rightarrow\sigma)\rightarrow\sigma \mbox{~~for each type \(\sigma\)} \;. \end{array}\] As usual, we will consider this language to be endowed with a certain (call-by-name) operational semantics, which in turn gives rise to a notion of *observational equivalence* for \({\mathrm{PCF}}\) programs. We define the *level* \({\mathrm{lv}}(\sigma)\) of a type \(\sigma\) inductively by \[{\mathrm{lv}}(\nat) ~=~ 0 \;, ~~~~~~ {\mathrm{lv}}(\sigma\rightarrow\tau) ~=~ \max\,({\mathrm{lv}}(\sigma)+1,{\mathrm{lv}}(\tau)) \;,\] and define the *pure type* \(\pure{k}\) of level \(k \in \mathbb{N}\) by \[\pure{0} ~=~ \nat \;, ~~~~~~ \pure{k+1} ~=~ \pure{k} \rightarrow \nat \;.\] Modifying the definition of \({\mathrm{PCF}}\) so that the constants \(Y_\sigma\) are admitted only for types \(\sigma\) of level \(\leq k\), we obtain a sublanguage \({\mathrm{PCF}}_k\) for any \(k \in \mathbb{N}\). Our main result will be that for each \(k\), the expressive power of \({\mathrm{PCF}}_{k+1}\) strictly exceeds that of \({\mathrm{PCF}}_k\): in particular, there is no closed term of \({\mathrm{PCF}}_k\) that is observationally equivalent to \(Y_{\pure{k+1}}\). (Fortunately, 'observational equivalence' has the same meaning for all the languages in question here, as will be explained in Section [2](#sec-background){reference-type="ref" reference="sec-background"}.) This answers a question posed explicitly by Berger in, but present in the folklore at least since the early 1990s. It is worth remarking that the situation is quite different for various extensions of \({\mathrm{PCF}}\) considered in the literature, in which one may restrict to recursions at level 1 types without loss of expressivity (see Subsection [\[subsec-embed\]](#subsec-embed){reference-type="ref" reference="subsec-embed"}). We can phrase our result more denotationally in terms of the type structure \({\mathsf{SF}}\) of *(PCF-)sequential functionals* or its effective substructure \({\mathsf{SF}}^{\mbox{\scriptsize \rm eff}}\). As will be reviewed in Section [2](#sec-background){reference-type="ref" reference="sec-background"}, the latter may be conveniently characterized up to isomorphism as the closed term model for \({\mathrm{PCF}}\) modulo observational equivalence. Our result can therefore be understood as saying that more elements of \({\mathsf{SF}}^{\mbox{\scriptsize \rm eff}}\) (and hence of \({\mathsf{SF}}\)) are denotable in \({\mathrm{PCF}}_{k+1}\) than in \({\mathrm{PCF}}_k\). From this we may easily infer that there is no finite 'basis' \(B \subseteq {\mathsf{SF}}^{\mbox{\scriptsize \rm eff}}\) relative to which all elements of \({\mathsf{SF}}^{\mbox{\scriptsize \rm eff}}\) are \(\lambda\)-definable (see Corollary [\[no-finite-basis-cor\]](#no-finite-basis-cor){reference-type="ref" reference="no-finite-basis-cor"} below). The models \({\mathsf{SF}}\) and \({\mathsf{SF}}^{\mbox{\scriptsize \rm eff}}\) are *extensional*: elements of type \(\sigma \rightarrow \tau\) can be considered as mathematical functions mapping elements of type \(\sigma\) to elements of type \(\tau\). Whilst our theorem is naturally stated in terms of these extensional models, its proof will make deep use of a more intensional model which yields \({\mathsf{SF}}\) as its extensional quotient. This intensional model of PCF has been considered many times before in the literature, for instance as the *\({\mathrm{PCF}}\) Böhm tree* model of Amadio and Curien, and it is known to be isomorphic to the game models of PCF given by Abramsky, Jagadeesan and Malacaria and by Hyland and Ong. In this paper, we choose to work with the *nested sequential procedure* (NSP) presentation of this model, as studied in detail in the recent book of Longley and Normann . (We will touch briefly on the possible use of other presentations for this purpose in Section [\[sec-further-work\]](#sec-further-work){reference-type="ref" reference="sec-further-work"}.) We shall denote the NSP model by \({\mathsf{SP}}^0\); its construction will be reviewed in Section [2](#sec-background){reference-type="ref" reference="sec-background"}, but in the meantime, let us offer a high-level overview of our proof method without assuming detailed knowledge of this model. As a motivating example, fix \(k \in \mathbb{N}\), and consider the \({\mathrm{PCF}}\) term \[\Phi_{k+1} ~:~ (\nat \rightarrow \pure{k+1} \rightarrow \pure{k+1}) \rightarrow \nat \rightarrow \pure{k+1}\] given informally by \[\Phi_{k+1}\;g\;n ~=~ g\;n\;(\Phi_{k+1}\;g\;(n+1)) ~=~ g\;n\;(g\;(n+1)\;(g\;(n+2)\;(g\;\cdots))) \;,\] or more formally by \[\Phi_{k+1} ~=~ \lambda g n.\; Y_{\nat \rightarrow \pure{k+1}} \, (\lambda f m.\,g\,m\,(f(m+1)))\,n \;,\] where \(f\) has type \(\nat \rightarrow \pure{k+1}\). Clearly, \(\Phi_{k+1}\) is a term of \({\mathrm{PCF}}_{k+1}\); however, we will be able to show that the element of \({\mathsf{SF}}\) that \(\Phi_{k+1}\) defines is not denotable in \({\mathrm{PCF}}_k\). What is the essential feature of this function that puts it beyond the reach of \({\mathrm{PCF}}_k\)? To get a hint of this, we may observe from the informal definition above that \(\Phi_{k+1}\) seems implicitly to involve an infinite nested sequence of calls to its argument \(g\), and indeed the NSP model makes this idea precise. Furthermore, each call to \(g\) involves an argument of type level \(k+1\) resulting from another such call. Broadly speaking, we shall refer to such a sequence of nested calls (subject to certain other conditions) as a *\(k\!+\!1\)-spine* in the NSP associated with \(\Phi_{k+1}\). As a second example, we can see from the natural recursive definition of \(Y_{\pure{k+1}}\) itself that this too involves a spine of this kind: \[Y_{\pure{k+1}} \; h ~=~ h(Y_{\pure{k+1}}\;h) ~=~ h(h(h(\cdots))) \;.\] where \(h\) has type \(\pure{k+1} \rightarrow \pure{k+1}\). In fact, we may view \(Y_{\pure{k+1}}\) as a 'special case' of \(\Phi_{k+1}\), since \(Y_{\pure{k+1}} \,h\) is observationally equivalent to \(\Phi_{k+1} \,(\lambda n.h) \,\num{0}\). A suitable general definition of \(k\!+\!1\)-spine turns out to be quite delicate to formulate; but having done this, it will be possible to prove that no \({\mathrm{PCF}}_k\)-denotable NSP contains a \(k\!+\!1\)-spine (Theorem [\[no-gremlin-thm\]](#no-gremlin-thm){reference-type="ref" reference="no-gremlin-thm"}). This will be proved by induction on the generation of such NSPs: in essence, we have to show that none of the generating operations for the interpretations of \({\mathrm{PCF}}_k\) terms are capable of manufacturing spinal NSPs out of non-spinal ones. This already suffices to show that within the intensional model \({\mathsf{SP}}^0\), the evident procedure for \(Y_{\pure{k+1}}\) is denotable in \({\mathrm{PCF}}_{k+1}\) but not in \({\mathrm{PCF}}_k\). However, this does not yet establish our main theorem, which concerns not \({\mathsf{SP}}^0\) but its extensional quotient \({\mathsf{SF}}\). For this purpose, we undertake a closer analysis of the function \(\Phi_{k+1}\) defined above: we show that not only the NSP arising from the above definition, but any *extensionally equivalent* NSP, must necessarily involve a \(k\!+\!1\)-spine. This shows that, within \({\mathsf{SF}}\) (or \({\mathsf{SF}}^{\mbox{\scriptsize \rm eff}}\)), the functional given by \(\Phi_{k+1}\) is not denotable in \({\mathrm{PCF}}_k\). This establishes Berger's conjecture that the languages \({\mathrm{PCF}}_k\) form a strict hierarchy. Since \(\Phi_{k+1}\) is definable from \(Y_{\nat \rightarrow \pure{k+1}}\), the above shows that the element \(Y_{\nat \rightarrow \pure{k+1}} \in {\mathsf{SF}}\) is not \({\mathrm{PCF}}_k\)-denotable. To complete the picture, however, we would also like to know that the simpler element \(Y_{\pure{k+1}} \in {\mathsf{SF}}\) is not \({\mathrm{PCF}}_k\)-denotable. We show this via a more refined version of the above analysis which is of some interest in its own right. Just as \({\mathrm{PCF}}\) is 'stratified' into sublanguages \({\mathrm{PCF}}_k\), we show that each \({\mathrm{PCF}}_k\) may be further stratified into sublanguages \({\mathrm{PCF}}_{k,1}, {\mathrm{PCF}}_{k,2}, \ldots\) on the basis of the 'width' of the types \(\sigma\) for which \(Y_\sigma\) is permitted (see Definition [\[width-def\]](#width-def){reference-type="ref" reference="width-def"}). An easy adaptation of our earlier proofs then shows that, for any \(l\), there are operators \(Y_\sigma\) in \({\mathrm{PCF}}_{k,l+2}\) that are not denotable in \({\mathrm{PCF}}_{k,l}\) (the appearance of \(l+2\) is admittedly a curiosity here). Since all these \(Y_\sigma\) are themselves readily definable from \(Y_{\pure{k+1}}\), it follows that \(Y_{\pure{k+1}} \in {\mathsf{SF}}\) itself is not denotable in \({\mathrm{PCF}}_{k,l}\) for any \(l\), and hence not in \({\mathrm{PCF}}_k\). This finer analysis illustrates the remarkable richness of structure that \({\mathsf{SF}}\) has to offer. The paper is organized as follows. In Section [2](#sec-background){reference-type="ref" reference="sec-background"} we recall the necessary technical background on \({\mathrm{PCF}}\) and on the models \({\mathsf{SP}}^0\) and \({\mathsf{SF}}\), fleshing out many of the ideas outlined above. In Section [\[sec-denotations\]](#sec-denotations){reference-type="ref" reference="sec-denotations"} we obtain a convenient inductive characterization of the class of procedures denotable in (the 'oracle' version of) \({\mathrm{PCF}}_k\), framed in terms of constructions on the procedures themselves. In Section [\[sec-Y-k+1\]](#sec-Y-k+1){reference-type="ref" reference="sec-Y-k+1"} we introduce the central concept of a \(k\!+\!1\)-spinal procedure, and show using our inductive characterization that no \({\mathrm{PCF}}_k\)-denotable procedure can be \(k\!+\!1\)-spinal (this is the most demanding part of the proof). As noted above, this already shows that \({\mathrm{PCF}}_{k+1}\) denotes more elements of \({\mathsf{SP}}^0\) than \({\mathrm{PCF}}_k\) does. In Section [\[sec-extensional\]](#sec-extensional){reference-type="ref" reference="sec-extensional"} we obtain the corresponding result for \({\mathsf{SF}}\), showing that the element \(\Phi_{k+1} \in {\mathsf{SF}}\), and hence \(Y_{\nat \rightarrow \pure{k+1}} \in {\mathsf{SF}}\), is not \({\mathrm{PCF}}_k\)-denotable. In Section [\[sec-pure-type\]](#sec-pure-type){reference-type="ref" reference="sec-pure-type"} we adapt our methods to the more fine-grained hierarchy of languages \({\mathrm{PCF}}_{k,l}\) that takes account of the widths of types; this enables us to show also that \(Y_{\pure{k+1}} \in {\mathsf{SF}}\) is not \({\mathrm{PCF}}_k\)-denotable. We conclude in Section [\[sec-further-work\]](#sec-further-work){reference-type="ref" reference="sec-further-work"} with a discussion of related and future work. The present paper is a revised, corrected and expanded version of a University of Edinburgh technical report from July 2015, the most significant changes being the addition of the material on the pure type \(\pure{k+1}\) in Section [\[sec-pure-type\]](#sec-pure-type){reference-type="ref" reference="sec-pure-type"}, and a simplified approach to characterizing a suitable substructure of \({\mathsf{SP}}^0\) in Section [\[sec-denotations\]](#sec-denotations){reference-type="ref" reference="sec-denotations"}. I am grateful to Ulrich Berger, Martín Escardó, Dag Normann and Alex Simpson for valuable discussions and correspondence, and to Luke Ong and Colin Stirling for helping me to navigate the existing literature on \(\lambda Y\) and recursion schemes (as discussed in Section [\[sec-further-work\]](#sec-further-work){reference-type="ref" reference="sec-further-work"}). Many of the participants in the Domains XII workshop in Cork and the Galop XI workshop in Eindhoven also offered valuable comments; thanks in particular to Neil Jones for drawing his work to my attention. Finally, I thank the anonymous referees for their careful work on the paper and their valuable suggestions; these have resulted in significant improvements in both the overall architecture and the formal details. # Background {#sec-background} We here summarize the necessary definitions and technical background from , especially from Chapters 6 and 7. ## The language \({\mathrm{PCF}}\) {#subsec-PCF} In, Scott introduced the language LCF for computable functionals of simple type. This language is traditionally called \({\mathrm{PCF}}\) when equipped with a standalone operational semantics as in Plotkin. We will work here with the same version of \({\mathrm{PCF}}\) as in, with the natural numbers as the only base type. Our types \(\sigma\) are thus generated by \[\sigma ~::=~ \nat ~\mid~ \sigma \rightarrow \sigma \;,\] and our terms will be those of the simply typed \(\lambda\)-calculus constructed from the constants \[\begin{array}{rcll} \num{n} &: & \nat & \mbox{for each \(n \in \mathbb{N}\)} \;, \\ \mathit{suc},\;\mathit{pre} &: & \nat\rightarrow\nat \;, \\ \mathit{ifzero} &: & \nat\rightarrow\nat\rightarrow\nat\rightarrow\nat \;, \\ Y_\sigma &: & (\sigma\rightarrow\sigma) \rightarrow \sigma & \mbox{for each type \(\sigma\)} \;. \end{array}\] We often abbreviate the type \(\sigma_0 \rightarrow\cdots\rightarrow \sigma_{r-1} \rightarrow \nat\) to \(\sigma_0,\ldots,\sigma_{r-1} \rightarrow \mathbb{N}\) or just \(\vec{\sigma} \rightarrow \nat\). As usual, we write \(\Gamma \vdash M:\sigma\) to mean that \(M\) is a well-typed term in the environment \(\Gamma\) (where \(\Gamma\) is a finite list of typed variables). Throughout the paper, we shall regard the type of a variable \(x\) as intrinsic to \(x\), and will often write \(x^\sigma\) to indicate that \(x\) carries the type \(\sigma\). For each \(k \in \mathbb{N}\), the sublanguage \({\mathrm{PCF}}_k\) is obtained by admitting the constants \(Y_\sigma\) only for types \(\sigma\) of level \(\leq k\). We endow the class of closed \({\mathrm{PCF}}\) terms with the following small-step reduction rules: \[\begin{array}{rclcrcl} (\lambda x.M)N & \rightsquigarrow & M[x \mapsto N] \;, & & \mathit{ifzero}\;\num{0} & \rightsquigarrow & \lambda xy.x \;, \\ \mathit{suc}\;\num{n} & \rightsquigarrow & \num{n+1} \;, & & \mathit{ifzero}\;\num{n+1} & \rightsquigarrow & \lambda xy.y \;, \\ \mathit{pre}\;\num{n+1} & \rightsquigarrow & \num{n} \;, & & Y_\sigma M & \rightsquigarrow & M(Y_\sigma M) \;. \\ \mathit{pre}\;\num{0} & \rightsquigarrow & \num{0} \;, \end{array}\] We furthermore allow these reductions to be applied in certain term contexts. Specifically, the relation \(\rightsquigarrow\) is inductively generated by the rules above along with the clause: if \(M \rightsquigarrow M'\) then \(E[M] \rightsquigarrow E[M']\), where \(E[-]\) is one of the contexts \[[-]N \;, ~~~~~~~~ {\mathit{suc}}\,[-] \;, ~~~~~~~~ {\mathit{pre}}\,[-] \;, ~~~~~~~~ {\mathit{ifzero}}\,[-] \;.\] We write \(\rightsquigarrow^*\) for the reflexive-transitive closure of \(\rightsquigarrow\). If \(Q\) is any closed \({\mathrm{PCF}}\) term of type \(\nat\), it is easy to check that either \(Q \rightsquigarrow^* \num{n}\) for some \(n \in \mathbb{N}\) or the (unique) reduction path starting from \(Q\) is infinite. This completes the definition of the languages \({\mathrm{PCF}}\) and \({\mathrm{PCF}}_k\). Whilst the language \({\mathrm{PCF}}_0\) is too weak for programming purposes (it cannot even define addition), it is not hard to show that even \({\mathrm{PCF}}_1\) is Turing-complete: that is, any partial computable function \(\mathbb{N} \rightharpoonup \mathbb{N}\) is representable by a closed \({\mathrm{PCF}}_1\) term of type \(\nat \rightarrow \nat\). We will also refer to the non-effective language \({\mathrm{PCF}}^\Omega\) (or *oracle \({\mathrm{PCF}}\)*) obtained by extending the definition of \({\mathrm{PCF}}\) with a constant \(C_f: \nat\rightarrow\nat\) for every set-theoretic partial function \(f: \mathbb{N} \rightharpoonup \mathbb{N}\), along with a reduction rule \(C_f \num{n} \rightsquigarrow \num{m}\) for every \(n,m\) such that \(f(n)=m\). (In \({\mathrm{PCF}}^\Omega\), the evaluation of a closed term \(Q:\nat\) may fail to reach a value \(\num{n}\) either because it generates an infinite computation, or because it encounters a subterm \(C_f(n)\) where \(f(n)\) is undefined.) The languages \({\mathrm{PCF}}^\Omega_k\) are defined analogously. If \(M,M'\) are closed \({\mathrm{PCF}}^\Omega\) terms of the same type \(\sigma\), and \({\mathcal{L}}\) is one of our languages \({\mathrm{PCF}}_k\), \({\mathrm{PCF}}_k^\Omega\), \({\mathrm{PCF}}\), \({\mathrm{PCF}}^\Omega\), we say that \(M,M'\) are *observationally equivalent in \({\mathcal{L}}\)*, and write \(M \simeq_{\mathcal{L}} M'\), if for all closed program contexts \(C[-^\sigma]: \nat\) of \({\mathcal{L}}\) and all \(n \in \mathbb{N}\), we have \[C[M] \rightsquigarrow^* \num{n} \mbox{~~iff~~} C[M'] \rightsquigarrow^* \num{n} \;.\] (It makes no difference to the relation \(\simeq_{\mathcal{L}}\) whether we take \(C[-]\) to range over single-hole or multi-hole contexts.) Fortunately, it is easy to show that all of the above languages give rise to exactly the same relation \(\simeq_{\mathcal{L}}\). Indeed, it is immediate from the definition that if \({\mathcal{L}},{\mathcal{L}}'\) are two of our languages and \({\mathcal{L}} \supseteq {\mathcal{L}}'\), then \(\simeq_{{\mathcal{L}}} \,\subseteq\, \simeq_{{\mathcal{L}}'}\); it therefore only remains to verify that \(M \simeq_{{\mathrm{PCF}}_0} M'\) implies \(M \simeq_{{\mathrm{PCF}}^\Omega} M'\). We may show this by a 'syntactic continuity' argument, exploiting the idea that any of the constants \(Y_\sigma\) or \(C_f\) in \({\mathrm{PCF}}^\Omega\) can be 'approximated' as closely as necessary by terms of \({\mathrm{PCF}}_0\). Specifically, let us write \(\bot\) for the non-terminating program \(Y_0(\lambda x^0.x): \nat\) (a term of \({\mathrm{PCF}}_0\)), and for any type \(\sigma\) write \(\bot_\sigma\) for the term of type \(\sigma\) of the form \(\lambda \vec{x}.\bot\). For any \(j \in \mathbb{N}\), we may then define \({\mathrm{PCF}}_0\) terms \[\begin{aligned} Y_\sigma^{(j)} & = & \lambda f^{\sigma\rightarrow\sigma}.\,f^j(\bot_\sigma) \;, \\ C_f^{(j)} & = & \lambda n.\,\mathit{case}~n~\mathit{of}~ (0 \Rightarrow \num{f(0)} \mid \cdots \mid j-1 \Rightarrow \num{f(j-1)}) \;, \end{aligned}\] where we use some evident syntactic sugar in the definition of \(C_f^{(j)}\). For any \({\mathrm{PCF}}^\Omega\) term \(M\), let \(M^{(j)}\) denote the 'approximation' obtained from \(M\) by replacing all occurrences of constants \(Y_\sigma, C_f\) by \(Y_\sigma^{(j)}, C_f^{(j)}\) respectively. It is then not hard to show that for closed \(Q:\nat\), we have \[Q \rightsquigarrow^* \num{n} \mbox{~~~iff~~~} \exists j.\; Q^{(j)} \rightsquigarrow^* \num{n} \;.\] From this it follows easily that if \(C[-]\) is an observing context of \({\mathrm{PCF}}^\Omega\) that distinguishes \(M,M'\), then some approximation \(C^{(j)}[-]\) (a context of \({\mathrm{PCF}}_0\)) also suffices to distinguish them. This establishes that \(\simeq_{{\mathrm{PCF}}_0} \,\subseteq\; \simeq_{{\mathrm{PCF}}^\Omega}\). We may therefore write \(\simeq\) for observational equivalence without ambiguity. In fact, an even more restricted class of observing contexts suffices for ascertaining observational equivalence of \({\mathrm{PCF}}^\Omega\) terms. The well-known *(equational) context lemma*, due to Milner, states that \(M \simeq M': \sigma_0,\ldots,\sigma_{r-1} \rightarrow \nat\) iff \(M,M'\) have the same behaviour in all *applicative contexts* of \({\mathrm{PCF}}\)---

{'timestamp': '2018-06-20T02:09:57', 'yymm': '1607', 'arxiv_id': '1607.04611', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04611'}

# Introduction In this note, we study the area of a minimal submanifold in the unit ball in \(\mathbb{R}^n\). For a minimal submanifold \(\Sigma\) that passes through the center of the ball, it is well-known that the area of \(\Sigma\) is bounded from below by the area of a flat \(k\)-dimensional disk: Theorem [\[consequence.of.monotonicity\]](#consequence.of.monotonicity){reference-type="ref" reference="consequence.of.monotonicity"} is a direct consequence of the well-known monotonicity formula for minimal submanifolds. This technique is discussed, for example, in and. In 1973, Alexander and Osserman studied a closely related problem. More precisely, they considered a minimal surface in the unit ball in \(\mathbb{R}^3\) which passes through a prescribed point in the interior of the ball (not necessarily the center of the ball). In the special case of disk-type minimal surfaces, they were able to show that the area of the surface is bounded from below by the area of a flat disk. However, their argument does not work for minimal surfaces of other topological types, nor does it generalize to higher dimensions. In 1974, Alexander, Hoffman, and Osserman proved an analogous inequality in higher dimensions, but only in the special case of area-minimizing surfaces. In this note, we completely settle this question for minimal submanifolds of arbitrary dimension and codimension: The proof of Theorem [\[main.theorem\]](#main.theorem){reference-type="ref" reference="main.theorem"} relies on an application of the first variation formula for minimal submanifolds (cf., ) to a carefully chosen vector field in ambient space. In particular, our argument generalizes immediately to the varifold setting. A similar technique was used in to prove a sharp bound for the area of a free-boundary minimal surface in a ball. The main difficulty in this approach is to find the correct vector field. The vector field used in was obtained as the gradient of the Green's function for the Neumann problem on the unit ball. By contrast, the vector field used in the proof of Theorem [\[main.theorem\]](#main.theorem){reference-type="ref" reference="main.theorem"} is not a gradient field, and does not have any obvious geometric interpretation. # Proof of Theorem [\[main.theorem\]](#main.theorem){reference-type="ref" reference="main.theorem"} Let us fix a point \(y \in B^n\). We define a vector field \(W\) on \(B^n \setminus \{y\}\) in the following way: For \(k>2\), we define \[\begin{aligned} W(x) &=-\frac{1}{k} \, \bigg ( \Big ( \frac{1-2 \langle x,y \rangle+|y|^2}{|x-y|^2} \Big )^{\frac{k}{2}}-1 \bigg ) \, (x-y) \\ &+ \frac{1}{k-2} \, \bigg ( \Big ( \frac{1-2 \langle x,y \rangle+|y|^2}{|x-y|^2} \Big )^{\frac{k-2}{2}}-1 \bigg ) \, y. \end{aligned}\] For \(k=2\), we define \[\begin{aligned} W(x) &=-\frac{1}{2} \, \Big ( \frac{1-2 \langle x,y \rangle+|y|^2}{|x-y|^2}-1 \Big ) \, (x-y) \\ &+ \frac{1}{2} \, \log \Big ( \frac{1-2 \langle x,y \rangle+|y|^2}{|x-y|^2} \Big ) \, y. \end{aligned}\] Note that \(1-2 \langle x,y \rangle+|y|^2 \geq |x-y|^2 > 0\) for all points \(x \in B^n \setminus \{y\}\). This shows that \(W\) is indeed a smooth vector field on \(B^n \setminus \{y\}\). **Proof.** We compute \[\begin{aligned} \sum_{i=1}^k \langle D_{e_i} W,e_i \rangle &= 1-\Big ( \frac{1-2 \langle x,y \rangle+|y|^2}{|x-y|^2} \Big )^{\frac{k}{2}} \\ &+ \frac{(1-2 \langle x,y \rangle+|y|^2)^{\frac{k-2}{2}}}{|x-y|^k} \, \sum_{i=1}^k \langle y,e_i \rangle \, \langle x-y,e_i \rangle \\ &+ \frac{(1-2 \langle x,y \rangle+|y|^2)^{\frac{k}{2}}}{|x-y|^{k+2}} \, \sum_{i=1}^k \langle x-y,e_i \rangle^2 \\ &-\frac{(1-2 \langle x,y \rangle+|y|^2)^{\frac{k-4}{2}}}{|x-y|^{k-2}} \, \sum_{i=1}^k \langle y,e_i \rangle^2 \\ &-\frac{(1-2 \langle x,y \rangle+|y|^2)^{\frac{k-2}{2}}}{|x-y|^k} \, \sum_{i=1}^k \langle x-y,e_i \rangle \, \langle y,e_i \rangle \\ &= 1-\frac{(1-2 \langle x,y \rangle+|y|^2)^{\frac{k}{2}}}{|x-y|^{k+2}} \, \Big ( |x-y|^2-\sum_{i=1}^k \langle x-y,e_i \rangle^2 \Big ) \\ &-\frac{(1-2 \langle x,y \rangle+|y|^2)^{\frac{k-4}{2}}}{|x-y|^{k-2}} \, \sum_{i=1}^k \langle y,e_i \rangle^2 \\ &\leq 1. \end{aligned}\] Note that the preceding calculation is valid both for \(k > 2\) and for \(k=2\). This proves the assertion.\ **Proof.** Suppose that \(x \in \partial B^n\). Then \(1-2 \langle x,y \rangle+|y|^2 = |x-y|^2\). This directly implies \(W(x)=0\). Again, this conclusion holds both for \(k>2\) and for \(k=2\).\ **Proof.** By definition of \(W(x)\), we have \[\begin{aligned} W(x) &=-(1-2 \langle x,y \rangle+|y|^2)^{\frac{k}{2}} \, \frac{x-y}{k \, |x-y|^k} + o \Big ( \frac{1}{|x-y|^{k-1}} \Big ) \\ &=-(1-|y|^2)^{\frac{k}{2}} \, \frac{x-y}{k \, |x-y|^k} + o \Big ( \frac{1}{|x-y|^{k-1}} \Big ) \end{aligned}\] as \(x \to y\). This proves the assertion.\ We now describe the proof of Theorem [\[main.theorem\]](#main.theorem){reference-type="ref" reference="main.theorem"}. To that end, we assume that \(\Sigma\) is a minimal surface in \(B^n\) passing through the point \(y\). Since the vector field \(W\) vanishes along the boundary \(\partial \Sigma \subset \partial B^n\), we obtain \[\label{first.variation} \int_{\Sigma \setminus B_r(y)} (1-\text{\rm div}_\Sigma W) = |\Sigma \setminus B_r(y)|-\int_{\Sigma \cap \partial B_r(y)} \langle W,\nu \rangle\] by the divergence theorem. Here, \(\nu\) denotes the inward pointing unit normal to the region \(\Sigma \cap B_r(y)\) within the surface \(\Sigma\). In other words, the vector \(\nu\) is tangential to \(\Sigma\), but normal to \(\Sigma \cap \partial B_r(y)\). It is easy to see that \[\nu =-\frac{x-y}{|x-y|} + o(1)\] for \(x \in \Sigma \cap \partial B_r(y)\). Using Lemma [\[c\]](#c){reference-type="ref" reference="c"}, we obtain \[\langle W,\nu \rangle = (1-|y|^2)^{\frac{k}{2}} \, \frac{1}{k \, r^{k-1}} + o \Big ( \frac{1}{r^{k-1}} \Big )\] for \(x \in \Sigma \cap \partial B_r(y)\). Since \[|\Sigma \cap \partial B_r(y)| = |\partial B^k| \, r^{k-1} + o(r^{k-1}),\] we conclude that \[\label{term.1} \lim_{r \to 0} \int_{\Sigma \cap \partial B_r(y)} \langle W,\nu \rangle = \frac{1}{k} \, |\partial B^k| \, (1-|y|^2)^{\frac{k}{2}} = |B^k| \, (1-|y|^2)^{\frac{k}{2}}.\] Combining ([\[first.variation\]](#first.variation){reference-type="ref" reference="first.variation"}) and ([\[term.1\]](#term.1){reference-type="ref" reference="term.1"}) gives \[\lim_{r \to 0} \int_{\Sigma \setminus B_r(y)} (1-\text{\rm div}_\Sigma W) = |\Sigma|-|B^k| \, (1-|y|^2)^{\frac{k}{2}}.\] On the other hand, by Lemma [\[a\]](#a){reference-type="ref" reference="a"} we have the pointwise inequality \[1-\text{\rm div}_\Sigma W \geq 0.\] Putting these facts together, we obtain \(|\Sigma|-|B^k| \, (1-|y|^2)^{\frac{k}{2}} \geq 0\), as claimed. Finally, we study the case of equality. Suppose that \(|\Sigma|-|B^k| \, (1-|y|^2)^{\frac{k}{2}} = 0\). In this case, we have \[1-\text{\rm div}_\Sigma W = 0\] for each point \(x \in \Sigma \setminus \{y\}\). Hence, if \(x\) is an arbitrary point on \(\Sigma \setminus \{y\}\) and \(\{e_1,\hdots,e_k\}\) is an orthonormal basis of \(T_x \Sigma\), then we have \[|x-y|^2-\sum_{i=1}^k \langle x-y,e_i \rangle^2 = \sum_{i=1}^k \langle y,e_i \rangle^2 = 0.\] This implies that \(\Sigma\) is a flat \(k\)-dimensional disk which is orthogonal to \(y\). This completes the proof of Theorem [\[main.theorem\]](#main.theorem){reference-type="ref" reference="main.theorem"}.

{'timestamp': '2017-01-30T02:05:13', 'yymm': '1607', 'arxiv_id': '1607.04631', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04631'}

# Introduction ## Bloch functions Let \(\D\) be the unit disk in the complex plane and \(\vp\) an analytic function defined in \(\D\). We say that \(\vp\) is a *Bloch function* if [\[Bloch\]]{#Bloch label="Bloch"} () = \_z(1-\|z\|\^2)\|'(z)\|\<. This defines a seminorm, and the Banach space \(\cB\) of all Bloch functions equipped with the norm \(\|\varphi\|_\mathcal B = |\varphi(0)|+\beta(\vp)\) is called *Bloch space*. We refer to, ,, , and for information on the Bloch space. For \(\vp\) analytic in \(\D\) the *schlicht radius* \(d_\vp(z)\) is defined as the radius of the largest disk lying on the Riemann image of \(\vp\) and centered at the point \(\vp(z)\), whenever \(z\) is not a branch point, *i.e.*\(\!\) if \(\vp'(z)\neq 0\). At a branch point of \(\vp\) the schlicht radius is defined as zero. It was shown in that every analytic function satisfies \[d_\vp(z) \, \leq \, (1-|z|^2)|\vp'(z)|, \qquad z\in\D.\] A similar inequality in the reverse direction was also shown in for the case when the schlicht radius is uniformly bounded. Thus, \(\vp\in\cB\) if and only if \(\sup_{z\in\D}d_\vp(z)<\infty\). For univalent functions these inequalities take the simpler form of [\[rad\]]{#rad label="rad"} (1-\|z\|\^2)\|'(z)\|     d\_(z)     (1-\|z\|\^2)\|'(z)\|, zdue to Koebe's \(1/4\)-Theorem. Note that in this case \(d_\vp(z)\) is simply the distance between \(\vp(z)\) and the boundary of \(\vp(\D)\) and, therefore, \(\vp\in\cB\) if and only if \(\vp(\D)\) does not contain arbitrarily large disks. Yet another close connection between Bloch functions and univalent functions was found in. Namely, if \(\psi\) is univalent then \(\beta(\log\psi') \leq 6\) and, conversely, if \(\beta(\vp)\leq 1\) then \(\vp=\log \psi'\) for some univalent function \(\psi\). ## Harmonic mappings A planar harmonic mapping is a complex-valued harmonic function \(f\) defined on a domain \(\Omega\subset\mathbb{C}\). When \(\Omega\) is simply connected, the mapping has a *canonical decomposition* \(f=h+\overline{g}\), where \(h\) and \(g\) are analytic in \(\Omega\). Since the Jacobian of \(f\) is given by \(J_f=|h'|^2-|g'|^2\), it is locally univalent and *orientation-preserving* if and only if \(|g'|< |h'|\), or equivalently, if \(h'(z)\neq0\) and the dilatation \(\omega=g'/h'\) has the property \(|\omega(z)|<1\) in \(\Omega\). We say that \(f\) is *orientation-reversing* if \(\overline{f}\) is orientation-preserving. Since the mid-80s and especially after the work of J. Clunie and T. Sheil-Small [@CS] in 1984, there has been a great interest in trying to extend the classic results of the analytic world to their harmonic analogues. Some work in this direction for the Bloch space was done by F. Colonna, whose point of departure was the *metric* characterization of \(\cB\), namely, \(f\in\cB\) if and only if \(f\) is Lipschitz between \(\D\) endowed with the hyperbolic metric and \(\SC\) endowed with the euclidean metric. For a harmonic mapping \(f=h+\overline g\,\) this Lipschitz condition was proved in to be equivalent to both \(h\) and \(g\) belonging to \(\cB\). The *schlicht radius* \(d_f(z)\) of a harmonic mapping \(f=h+\overline{g}\) is defined as the radius of the largest disk which is the injective image of some subdomain of \(\D\) and is centered at \(f(z)\). We set \(d_f(z)=0\) if no such disk exists. A generalization of the *geometric* definition of Bloch functions would be to ask that \(f\) satisfy \(\sup_{z\in\D}d_f(z)<\infty\). However, we shall prove in Lemma [\[radius\]](#radius){reference-type="ref" reference="radius"} that if \(f\) is univalent and normalized then \[\frac{1}{16}(1-|z|^2) (|h'(z)|-|g'(z)|) \; \leq \; d_f(z) \; \leq \; \frac{\pi}{2} (1-|z|^2) |h'(z)|, \qquad z\in\D.\] It can also be shown that no two of the above three quantities are comparable. Therefore, an analytic characterization of the geometric definition for harmonic mappings is, as far as we know, yet to be found. ## Harmonic Bloch-type functions Our starting point will be the *analytic* definition [\[Bloch\]](#Bloch){reference-type="eqref" reference="Bloch"}. Noting that the Jacobian of an analytic function \(\vp\) is given by \(J_\vp = |\vp'|^2\), we feel justified in introducing the following definition. Indeed, we shall see in Section 2 that this definition gives rise to a class rather than a linear space. However, \(\cB_H\) contains the Bloch space defined in. We shall prove that \(\cB_H\) is both affine and linearly invariant. In Section 3 we show a connection between \(\cB_H\) and univalent harmonic mappings that resembles Pommerenke's theorem. We also study the schlicht radius in \(\cB_H\). In Section 4 we give growth and coefficients estimates for sense-preserving functions in \(\cB_H\). # The class of harmonic Bloch-type mappings Our first task will be to show the affine and linear invariance of \(\cB_H\). Throughout the paper we will denote by \(\vp_\al \, (\al\in\D)\) the disk automorphism given by \(\vp_\al(z)=(\al+z)/(1+\overline{\al}z), z\in\D\). In what follows, Example [\[ex-no-space\]](#ex-no-space){reference-type="ref" reference="ex-no-space"} shows that \(\cB_H\) is not a linear space. It also shows that functions in \(\cB_H\) may grow arbitrarily fast. Hence, in order to get growth and coefficient estimates in Section 4 we shall restrict ourselves to sense-preserving functions in \(\cB_H\). Example [\[ex-extr\]](#ex-extr){reference-type="ref" reference="ex-extr"} shows that the harmonic Bloch space considered in is strictly contained in \(\cB_H\). Recall that in the definition of a Bloch function \(f= h + \overline{g}\) is equivalent to both \(h\) and \(g\) belonging to \(\cB\). # Univalent functions Let \(f=h+\overline{g}\) be a harmonic, univalent and sense-preserving mapping in \(\D\). Let \(\omega = g'/h': \D\to\D\) be its dilatation and write \[h(z)=\sum_{n=0}^\infty a_n z^n \qquad \text{and} \qquad g(z)=\sum_{n=1}^\infty b_n z^n.\] We say that \(f\in S_H\) if it satisfies \(a_0=1-a_1=0\) and that \(f\in S_H^0\) if in addition \(b_1=0\). A simple use of the Schwarz Lemma yields the sharp inequality \(|b_2|\leq 1/2\) for functions in \(S_H^0\). It takes more effort to prove that \(|a_2|<49\) in \(S_H^0\), and still, the best known constant \(49\) is quite distant from the conjectured \(5/2\). For the larger class \(S_H\), we have that \(|b_1|<1\) simply because \(f\) is sense-preserving. Also, it is possible to translate the preceding inequalities by means of an affine transformation. Given \(f\in S_H\), the function [\[aff\]]{#aff label="aff"} f_0 = belongs to \(S_H^0\). This transformation is invertible, so that \(f =f_0 +\overline{b_1}\overline{f_0}\). Hence, it is not difficult to see that [\[a_2\]]{#a_2 label="a_2"} \|a_2\| \< 49 + for functions in \(S_H\). In the recent work a new *Schwarzian* derivative for harmonic locally univalent functions was defined and studied. Also, a *pre-Schwarzian* derivative was defined as \[P_f = \frac{h''}{h'}-\frac{\overline{\omega}\omega'}{1-|\omega|^2}\] and with it the following *Becker-type* criterion for univalence was proved. For any \(\omega:\D\to\D\) analytic we define its *hyperbolic derivative* by \[\omega^*(z) = \frac{\omega'(z) (1-|z|^2)}{1-|\omega(z)|^2}.\] We set \(\|\omega\|_h = sup_{z\in\D}|\omega^*(z)|\) for its *hyperbolic norm*. See. We recall that a sense-preserving homeomorphism \(f\) is called *quasiconformal* if it maps infinitesimal circles onto infinitesimal ellipses having ratio of the major over the minor axis bounded by some constant. This is equivalent to saying that its (second complex) dilatation \(\om=\overline{f_{\overline{z}}}/f_z\) is bounded away from one, that is, \(|\om(z)| \leq k<1\). See. ## Connection between univalent and Bloch-type functions. A well-known theorem of Pommerenke gives yet another characterization of the analytic Bloch space \(\cB\). It states that a function \(f\) is Bloch if and only if there exists a constant \(c>0\) and a univalent function \(g\) such that \(f=c\log g'\). The following theorems show a similar connection between harmonic univalent mappings and the class \(\cB_H\). In the opposite direction we have the following theorem. ## Schlicht radius. A well-known covering theorem states that all functions in \(S_H^0\) contain in their image a disk centered at the origin, having radius \(1/16\). (The conjectured constant is \(1/6\).) Applying as before the affine transformation [\[aff\]](#aff){reference-type="eqref" reference="aff"} it is easy to see that [\[cover\]]{#cover label="cover"} { w: \|w\|\<} f() for every \(f\in S_H\). A result in the opposite direction states that each function in \(S_H\) omits some point on the circle \(|w|=\frac{\pi}{2}\). In other words [\[Hall\]]{#Hall label="Hall"} (\\f()) {\|w\|=} . The constant \(\frac{\pi}{2}\) was given by Hall and is best possible. See also. As mentioned in Section 1, the *schlicht radius* \(d_f(z)\) of a harmonic mapping \(f=h+\overline{g}\) at a point \(z\in\D\) is defined as the radius of the largest disk which is the injective image of some subdomain of \(\D\) and is centered at \(f(z)\). If there is no such disk then we set \(d_f(z)=0\). The existence of a universal lower bound for \(\sup_{z\in\D}d_f(z)\) is commonly refered to as a *Bloch theorem*. It was shown in that openness (*i.e.*\(\!\) the property of mapping open sets to open sets) and the normalization \(g'(0)=1-h'(0)=0\) are sufficient conditions for a Bloch theorem to hold. Moreover, it was shown that the normalization alone is not a sufficient condition. Since here we will be concerned only with univalent functions, the schlicht radius coincides with the distance between \(f(z)\) and the boundary of \(f(\D)\). The following lemma provides us with some estimates. We shall need the following lemma. See, page 3. # Growth and coefficients estimates For a harmonic sense-preserving function \(f = h + \overline{g}\) with dilatation \(\om=g'/h':\D\to\D\), we write the power series \[h(z)=\sum_{n=0}^\infty a_n z^n, \quad g(z)=\sum_{n=1}^\infty b_n z^n \quad \text{and} \quad \om(z)=\sum_{n=0}^\infty c_n z^n.\] Of course \(c_0=b_1/a_1\). We will also make use of the notation \[M_\infty(r,f) = \max_{|z|=r} |f(z)|.\] We now present some growth and coefficients estimates for the class \(\cB_H\). Note, however, that these bounds are not uniform throughout \(\cB_H\), but rather, to each of its subclasses having prescribed \(|c_0|\).

{'timestamp': '2016-12-26T02:05:25', 'yymm': '1607', 'arxiv_id': '1607.04626', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04626'}

# Introduction {#sec:intro} As a key part of the human visual system, visual attention mechanisms filter irrelevant visual stimuli in order to focus more on the important parts. Computational models of attention try to mimic this process through the use of machines and algorithms. These models have gained increasing attention lately. The reason behind this growing interest lies in their use in different computer vision and multimedia problems including but not limited to image retrieval , visual quality assessment  video resizing/summarization , action recognition , event detection  either as a tool for visual feature extraction or as a mechanism for selecting features. These models are also important for generic applications such as advertisement and web design  as attention plays a key role in both user interfaces and human-machine interaction. In the literature, the computational attention models developed so far generally aim to predict where humans fixate their eyes in images . Specifically, they produce the so-called saliency maps from the visual data where a high saliency score at an image location indicates that the point is more likely to be fixated. These models are largely inspired by the hierarchical processing of human visual system  and the theoretical studies like Feature Integration Theory  or Guided Search Model . They consider low-level image features (color, orientation, contrast, motion, etc.) and/or high-level features (pedestrians, faces, text) while predicting saliency maps. In this process, while low-level features are employed to examine how different an image point from its surroundings, high-level features are employed as it is experimentally shown that humans have a tendency to fixate on certain object classes more than others. Another current trend in the existing literature is to detect salient objects  from the images. These models specifically aim at identifying the most prominent objects in an image that attract attention under free-viewing conditions and then segmenting them out from the background. These computational models for visual attention can be further grouped into two according to their inputs as static and dynamic saliency models. While static models work on images, dynamic models take video sequences as input. Saliency prediction from videos leads to great challenges when it is compared to carrying out the same task in still images. The reason is that the dynamic saliency frameworks require taking into account both spatial and temporal characteristics of the video sequences. Static saliency models use features like color, intensity and orientation, however dynamic models need to focus more on the moving objects or image parts as it is shown that humans there is a tendency for humans to look at them while viewing. Hence, the preliminary models proposed for dynamic saliency prediction extend the existing saliency models suggested for still images so that they consider extra motion features . However, more recent works approach the same task from a different point of view and propose novel solutions . ## Overview of our approach Deep learning has been successfully applied to saliency prediction in still images in the last few years, providing state-of-the-art results . The early models utilize pre-trained deep convolutional neural networks (CNNs) proposed to classify images as generic feature extractors and build classifiers on top of those features to classify fixated image regions . Later models, however, approach the problem from an end-to-end perspective and either train networks from scratch or most of the time fine-tune the weights of a pre-trained model . The modifications in the network architectures are usually about integrating multi-scale processing or using different loss functions . It has been investigated that the power of these deep models mainly comes from the property that the features learned by these networks are semantically very rich , capturing high-level factors important for saliency detection. Motivated by the success of these works, in this study, we explore the use of two-stream CNNs for saliency prediction from videos. To the best of our knowledge, our work is the first deep model for dynamic saliency, which is trained in an end-to-end manner, that learns to combine spatial and temporal information in an optimal manner within a two-stream network architecture. ## Our contributions The contributions of our work can be summarized as follows: 1. We study two-stream convolutional neural networks which mimic the visual pathways in the brain and combine networks trained on temporal and spatial information to predict saliency map of a given video frame. Although these network architectures have been previously investigated for some computer vision problems such as video classification  and action recognition , to our knowledge, we are the first to apply two-stream deep models for saliency prediction from videos in the literature. In particular, in our study, we investigate two different fusion strategies, namely element-wise and convolutional fusion strategies, to integrate spatial and temporal streams. 2. We carry out extensive experiments on DIEM   and UCF-Sports  datasets and compare our deep spatio-temporal saliency networks against several state-of-the-art dynamic saliency models. Our evaluation demonstrates that the proposed STSConvNet model outperforms these models in nearly all of the evaluation metrics on these datasets. 3. On a number of challenging still images, we also show that our spatio-temporal saliency network can predict the human fixations better than the state-of-the-art deep static saliency models. The key idea that we follow is to extract optical flow from these static images by using a recently proposed method  and feed them to our network along with the appearance image. # Related Work In this study, we focus on bottom-up modeling of dynamic saliency. Below, we first summarize the existing dynamic saliency models from the literature and then provide a brief overview of the proposed deep-learning based static saliency models which are related to ours. ## Dynamic Saliency Early examples of saliency models for dynamic scenes extend the previously proposed static saliency models which process images in a hierarchical manner by additionally considering features related to motion such as optical flow. For instance, in , Harel *et al.* propose a graph-theoretic solution to dynamic saliency by representing the extracted feature maps in terms of fully connected graphs and by predicting the final saliency map. In , Cui *et al.* extract salient parts of video frames by performing spectral residual analysis on the Fourier spectrum of these frames over the spatial and the temporal domains. In , Guo *et al.* propose a similar spectral analysis based formulation. In , Sultani and Saleemi extend Harel *et al.* 's model by using additional features such as color and motion gradients and by post-processing the predicted maps via a graphical model based on Markov Random Fields. In , Seo and Milanfar employ self similarities of spatio-temporal volumes to predict saliency.  , Mauthner *et al.* also present a video saliency detection method for using as a prior information for activity recognition algorithms. Instead of using a data driven approach they propose an unsupervised algorithm for estimating salient regions of video sequences. Following these early works, other researchers rather take different perspectives and devise novel solutions for dynamic saliency. For example, in , Hou and Zhang consider rarity of visual features while extracting saliency maps from videos and propose an entropy maximization-based model. In , Rahtu *et al.* extract saliency by estimating local contrast between feature distributions. In , Ren *et al.* propose a unified model with the temporal saliency being estimated by sparse and low-rank decomposition and the spatial saliency map being extracted by considering local-global contrast information. In , Mathe *et al.* devise saliency prediction from videos as a classification task where they integrate several visual cues through learning-based fusion strategies. In another study, Rudoy *et al.*  propose another learning based model for dynamic saliency. It differs from Mathe *et al.*'s model  in that they take into account a sparse set of gaze locations thorough which they predict conditional gaze transitions over subsequent video frames. Zhou *et al.*  oversegment video frames and use low-level features from the extracted regions to estimate regional contrast values. Zhao *et al.*  learn a bank of filters for fixations and use it to model saliency in a location-dependent manner. Khatoonabadi *et al.*  propose a saliency model that depends on compressibility principle. In a very recent study, Leboran *et al.* , propose another dynamic saliency model by using the idea that perceptual relevant information is carried by high-order statistical structures. ## Deep Static Saliency In recent years, deep neural networks based models provide state-of-the-art results in many computer vision problems such as image classification , object detection , activity recognition , semantic segmentation  and video classification . These approaches perform hierarchical feature learning specific to a task, and thus gives results better than the engineered features. Motivated by the success of these models, a number of researchers have recently proposed deep learning based models for saliency prediction from images . Vig *et al.*  use an ensemble of CNNs which learns biologically inspired hierarchical features for saliency prediction. Kümmerer *et al.*  employ deep features learned through different layers of the AlexNet  and learn how to integrate them for predicting saliency maps. Kruthiventi et al.  adopt VGGNet  for saliency estimation where they introduce a location-biased convolutional layer to model the center-bias, and train the model on SALICON dataset using Euclidean loss. Jetley *et al.*  also use the VGGNet architecture but they especially concentrate on investigating different kinds of probability distance measures to define the loss function. Pan *et al.*  very recently propose two CNN models having different layer sizes by approaching saliency prediction as a regression task. Li *et al.*  employ a fully convolutional neural network within a multi-task learning framework to jointly detect saliency and perform object class segmentation. It is important to note that all these models are proposed for predicting saliency in still images not videos. Bruce *et al.*  propose yet another fully convolutional network to predict saliency and they try to understand factors and learned representations when training these type of networks for saliency estimation. Motivated by the deep static saliency models, in our paper we investigate the use of two-stream CNNs for saliency prediction from videos. In fact, investigating layered formulations is not new for saliency prediction. As discussed earlier, most of the traditional dynamic saliency models are inspired from the hierarchical processing in the low-level human vision . These models, however, employ hand-crafted features to encode appearance and motion contrast to predict where humans look at in dynamic scenes. Since they depend on low-level cues, they often fail to capture semantics of scenes at its full extent, which is evidently important for gaze prediction. More recent models, on the other hand, employ learning-based formulations to integrate these low-level features with the detection maps for faces, persons, and other objects. This additional supervision boosts the prediction accuracies, however, the performance is limited by the discrimination capability of the considered features and the robustness of the employed detectors. As compared to the previous dynamic saliency models, our deep spatio-temporal saliency networks are trained to predict saliency in an end-to-end manner. This allows us to learn hierarchical features, both low-, mid-and high-level,  that are specialized for the gaze prediction task. For instance, while the early layers learn filters that are sensitive to edges or feature contrasts, the filters in the top layers are responsible from capturing more complex, semantic patterns that are important for the task. In our case, our deep two-stream saliency networks learn multiple layers of spatial and temporal representations and ways to combine them to predict saliency. In particular, in our study we extract temporal information via optical flow between consecutive video frames and investigate different ways to use this additional information in saliency prediction within a deep two-stream spatio-temporal network architecture . These two-stream networks are simple to implement and train, and to our interest, are in line with the hierarchical organization of the human visual system. Specifically, the biological motivation behind these architectures is the so-called two-streams hypothesis  which speculate that human visual cortex is comprised of two distinct streams, namely ventral and the dorsal streams, which are respectively specialized to process appearance and motion information. Here, an alternative deep architecture could be to stack two or more frames together and feeding this input to a deep single-stream CNN, which was investigated in several action recognition networks . In this work, we do not pursue this direction because of two reasons. Firstly, this approach requires learning 3D convolutional filters  in order to capture spatio-temporal regularities among input video frames but using 3D filters highly increases the complexity of the models and these 3D convolutional networks are harder to train with limited training data  (which is the case for the existing dynamic saliency datasets). Secondly, 3D convolutional filters are mainly used for expressing long-range motion patterns which could be important for recognizing an action since they cannot easily be captured by optical-flow based two-stream models. For dynamic saliency prediction though, we believe that such long-range dependencies are minimally important as human attention shifts continuously, and optical flow information is sufficient to establish the link between motion and saliency. # Our Models The aim of our study is to investigate the use of deep architectures for predicting saliency from dynamic scenes. Recently, CNNs provided drastically superior performance in many classification and regression tasks in computer vision. While the lower layers of these networks respond to primitive image features such as edges, corners and shared common patterns, the higher layers extract semantic information like object parts, faces or text . As mentioned before, such low and high-level features are shown to be both important and complementary in estimating visual saliency. Towards this end, we examine two baseline single stream networks (spatial and temporal) given in Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(a) and two two-stream networks  shown in Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(b), which combine spatial and temporal cues via two different integration mechanisms: element-wise max fusion and convolutional fusion, respectively. We describe these models in detail below. ## Spatial Saliency Network For the basic single stream baseline model, we retrain the recently proposed static saliency model in  for dynamic saliency prediction by simply ignoring temporal information and using the input video frame alone. Hence, this model does not consider the inter-frame relationships while predicting saliency for a given video. As shown in the top row of Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(a), this CNN resembles the VGG-M model --the main difference being that the final layer is a deconvolution (fractionally strided convolution) layer to up sample to the original image size. Note that it does not use any temporal information and exploits only appearance information to predict saliency in still video frames. We refer to this network architecture as SSNet. ## Temporal Saliency Network Saliency prediction from videos is inherently different than estimating saliency from still images in that our attention is highly affected by the local motion contrast of the foreground objects. To understand the contribution of temporal information to the saliency prediction, we develop a second single stream baseline. As given in the bottom row of Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(a), this model is just a replica of the spatial stream net but the input is provided in the form of optical flow images, as in , computed from two subsequent frames. We refer to this single stream network architecture as TSNet. ## Spatio-Temporal Saliency Network with Direct Averaging As a baseline model, we define a network model which integrates the responses of the final layers of the spatial and the temporal saliency networks by using direct averaging. Note that this model does not consider a learning strategy on how to combine these two-stream network and consider each one of the single-stream networks equally reliable. We refer to this two-stream network architecture as STSAvgNet. ## Spatio-Temporal Saliency Network with Max Fusion This network model accepts both a video frame and the corresponding optical flow image as inputs and merges together the spatial and temporal single stream networks via element-wise max fusion. That is, given two feature maps \(\mathbf{x}^{s},\mathbf{x}^{t}\in \mathbb{R}^{H\times W \times D}\) from the spatial and temporal streams, with \(W,H,D\) denoting the width, height and the number of channels (filters), max fusion takes the maximum of these two feature maps at every spatial location \(i\) and \(j\), and channel \(d\), as: \[y^{max}_{i,j,d} = \max\left(x^{s}_{i,j,d}, x^{t}_{i,j,d}\right)\;.\] The use of the \(\max\) operation makes the ordering of the channels in a convolutional layer arbitrary. Hence, this fusion strategy assumes arbitrary correspondences between the spatial and temporal streams. That said, this spatio-temporal model seeks filters so that these arbitrary correspondences between feature maps become as meaningful as possible according to the joint loss. After this fusion step, it also uses a deconvolution layer to produce an up-sampled saliency map as the final result as illustrated in the top row of Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(b). We refer to this two-stream network architecture as STSMaxNet. ## Spatio-Temporal Saliency with Convolutional Fusion This network model integrates spatial and temporal streams by applying convolutional fusion. That is, the corresponding feature maps \(\mathbf{x}^{s}\) and \(\mathbf{x}^{t}\) respectively from the spatial and temporal streams are stacked together and then combined as follows: \[\mathbf{y}^{conv} = \left[ \mathbf{x}^{s} \quad\mathbf{x}^{t}\right]*\mathbf{f}+\mathbf{b}\;,\] where \(\mathbf{f}\in\mathbb{R}^{1\times 1 \times 2D \times D}\) denotes a bank of \(1\times 1\) filters, and \(\mathbf{b}\in \mathbb{R}^D\) represents the bias term. The main advantage of the convolutional fusion over the element-wise max fusion is that the filterbank \(\mathbf{f}\) learns the optimal correspondences between the spatial and temporal streams based on the loss function, and reduces the number of channels by a factor of two through the weighted combinations of \(\mathbf{x}^{s}\) and \(\mathbf{x}^{t}\) with weights given by \(\mathbf{f}\) at each spatial location. As demonstrated in the bottom row of Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(b), this is followed by a number of convolutions and a final deconvolution layer to produce the saliency map. We refer to this network architecture as STSConvNet. ## Spatio-Temporal Saliency Network with Direct Fusion Finally, as another baseline model, we design a single stream network in which the appearance and optical flow images are stacked together and fed to the network as input. This model implements an early fusion strategy at the very beginning of the network architecture and can be seen as a special case of STSConvNet. Here, each layer of the network learns a set of filters that directly acts on the given appearance and motion frames. We refer to this model as STSDirectNet. # Implementation Details {#ssec:implementation_details} ## Network Architectures The architecture of our single stream models is the same with that of the deep convolution network proposed in . They take \(320\times 240\times 3\) pixels images and processes them by the following operations: \(C(96,7,3)\) \(\rightarrow\) \(LRN\) \(\rightarrow\) \(P\) \(\rightarrow\) \(C(256,5,2)\) \(\rightarrow\) \(P\) \(\rightarrow\) \(C(512,3,1)\) \(\rightarrow\) \(C(512,5,2)\) \(\rightarrow\) \(C(512,5,2)\) \(\rightarrow\) \(C(256,7,3)\) \(\rightarrow\) \(C(128,11,5)\) \(\rightarrow\) \(C(32,11,5)\) \(\rightarrow\) \(C(1,13,6)\) \(\rightarrow\) \(D\). Here, \(C(d,f,p)\) represents a convolutional layer with \(d\) filters of size \(f\times f\) applied to the input with padding \(p\) and stride 1. \(LRN\) denotes a local response normalization layer that carries out a kind of lateral inhibition, and \(P\) indicates a max pooling layer over \(3\times 3\) regions with stride 2. Finally, \(D\) is a deconvolution layer with filters of size \(8\times 8\times 1\) with stride 4 and padding 2 which upscales the final convolution results to the original size. All convolutional layers except the last one are followed by a ReLU layer. Our spatial and temporal stream models in particular differ from each other in their inputs. While the first one processes still images, the next one accepts optical flow images as input. For the proposed spatio-temporal saliency networks shown in Figure [\[fig:saliency-networks\]](#fig:saliency-networks){reference-type="ref" reference="fig:saliency-networks"}(b), we employ element-wise max and convolutional fusion strategies to integrate the spatial and temporal streams. Performing fusion after the fifth convolutional layer gives the best results for both of these fusion strategies. In STSMaxNet, the single stream networks are combined by applying element-wise max operation, which is followed by the same deconvolution layer in the single stream models. On the other hand, STSConvNet performs fusion by stacking the feature maps together and integrating them by a convolution layer \(C(512,1,0)\) whose weights are initialized with identity matrices. The remaining layers are the same with those of the single stream models. ## Data Preprocessing We employ three publicly available datasets, 1.DIEM (Dynamic Images and Eye Movements) , 2. UCF-Sports  datasets and 3. MIT 300 dataset , which are described in detail in Section [5](#sec:experiments){reference-type="ref" reference="sec:experiments"}, in our experiments. Since our networks accept inputs of size \(320\times 240\times 3\) pixels and outputs saliency maps of the same size, all videos and ground truth fixation density maps are rescaled to this size prior to training. We use the publicly available implementation of DeepFlow  and we additionally extract optical flow information from the rescaled versions of subsequent video frames. Optical flow images are then generated by stacking horizontal and vertical flow components and the magnitude of the flow together. Some example optical flow images are shown in Figure [\[fig:opticalflow\]](#fig:opticalflow){reference-type="ref" reference="fig:opticalflow"}. ## Data Augmentation Data augmentation is a widely used approach to reduce the effect of over-fitting and improve generalization of neural networks. For saliency prediction, however, classical techniques such as cropping, horizontal flipping, or RGB jittering are not very suitable since they alter the visual stimuli used in the eye tracking experiments in collecting the fixation data. Having said that, horizontal flipping is used in  as a data augmentation strategy although there is no theoretical basis for why this helps to obtain better performance. In our study, we propose to employ a new and empirically grounded data augmentation strategy for specifically training saliency networks. In , the authors performed a thorough analysis on how image resolution affects the exploratory behavior of humans through an eye-tracking experiment. Their experiments revealed that humans are quite consistent about where they look on high and low-resolution versions of the same images. Motivated with this observation, we process all video sequences and produce their low-resolution versions by down-sampling them by a factor of 2 and 4, and use these additional images with the fixations obtained from original high-resolution images in training. We note that in reducing the resolution of optical flow images the magnitude should also be rescaled to match with the down-sampling rate. It is worth-mentioning that this new data augmentation strategy can also be used for boosting performances of deep models for static saliency estimation. ## Training We employ the weights of the pretrained CNN model in  to set the initial weights of our spatial and temporal stream networks. In training the models, we use Caffe framework  and employed Stochastic Gradient Descent with Euclidean distance between the predicted saliency map and the ground truth. The networks were trained over 200K iterations where we used a batch size of 2 images, momentum of 0.9 and weight decay of 0.0005, which is reduced by a factor of 0.1 at every 10K iterations. Depending on the network architectures, it takes between 1 day to 3 days to train our models on the DIEM dataset by using a single 2GB GDDR5 NVIDIA GeForce GTX 775M GPU on a desktop PC equipped with 4-core Intel i5 (3.4 GHz) Processor and 16 GB memory. At test time, it takes approximately 2 secs to extract the saliency map of a single video frame. # Experimental Results {#sec:experiments} In the following, we first review the datasets on which we perform our experiments and provide the list of state-of-the-art computational saliency models that we compared against our spatio-temporal saliency networks. We then provide the details of the quantitative evaluation metrics that are used to assess the model performances. Next, we discuss our experimental results. ## Datasets To validate the effectiveness of the proposed deep dynamic saliency networks, our first set of experiments is carried out on the DIEM dataset . This dataset is collected for the purpose of investigating where people look at dynamic scenes. It includes 84 high-definition natural videos including movie trailers, advertisements, etc. Each video sequence has eye fixation data collected from approximately 50 different human subjects. In our experiments, we only used the right-eye positions of the human subjects as done in . We perform our second set of experiments on the UCF-Sports  . This dataset is collected from broadcast television channels such as the BBC and ESPN which consists of a set of sport actions . The video sequences are collected from wide range of websites. This dataset contains 150 video sequences with \(720\times 480\) resolution and cover a range of scene and viewpoints. The dataset includes several actions such as diving, golf swing, kicking and lifting, and is used for action recognition. However, recently, additional human gaze annotations were collected in . These fixations are collected over 16 human subjects under task specific and task-independent free viewing conditions. Lastly, for the experiments on predicting eye fixations on still images, we choose a number of images from the MIT 300 dataset . Selected images are the ones especially depicting an action and including objects that are interpreted as in motion. This dataset has eye fixation data collected from 39 subjects under free-viewing conditions for 3 secs for a total of 300 natural images with longest dimension 1024 pixels and the other dimension varied from 457 to 1024 pixels. ## The compared computational saliency models Through our experiments on DIEM and UCF-Sports datasets, we compare our deep network models with eight state-of-the-art dynamic saliency models: GVBS , PQFT , SR , Seo and Milanfar , Rudoy *et al.* , Fang *et al.* , Zhou *et al.* , and DWS  models. Moreover, we compare our two-stream deep models STSMaxNet and STSConvNet to a certain extent with deep static saliency model DeepSal  as the architectures of our TSNet and SSNet models are adapted from this model. ## Evaluation Measures We employ Area Under Curve (AUC), shuffled AUC (sAUC) , Pearson's Correlation Coefficient (CC), Normalized Scanpath Saliency (NSS) , Normalized Cross Correlation (NCC) and \({\chi}^2\) distance throughout our experiments for performance evaluation. We note that NCC and \({\chi}^2\) distance are not widely-used measures but we report them as some recent studies  employ them in their analysis. AUC measure considers the saliency maps as a classification map and uses the receiver operator characteristics curve to estimate the effectiveness of the predicted saliency maps in capturing the ground truth eye fixations. An AUC value of 1 indicates a perfect match with the ground truth while the performance of chance is indicated by a value close to 0.5. The AUC measure cannot account for the tendency of human subjects to look at the image center of the screen, i.e. the so-called center bias. Hence, we also report the results of the shuffled version of AUC (sAUC) where the center bias is compensated by selecting the set of fixations used as the false positives from another randomly selected video frame from the dataset instead of the processed frame. CC treats the predicted saliency and the ground truth human fixation density maps as random variables and measures the strength of a linear relationship between two using a Gaussian kernel density estimator. While a value close to 0 indicates no correlation, a CC value close to +1/-1 demonstrates a good linear relationship. NSS estimates the average normalized saliency score value by examining the responses at the human fixated locations on the predicted saliency map which has been normalized to have zero mean and unit standard deviation. While a NSS value of 0 indicates chance in predicting eye fixtions, a non-negative NSS value, especially that of greater than 1, denotes correspondence between maps above chance. NCC is a general measure used for assessing image similarity. It treats the ground truth fixation map and the predicted saliency map as images and estimates a score with values close to 1 implying high similarity and negative values indicating low similarity. \({\chi}^2\) distance considers the saliency maps as a probability distribution map and compares the predicted map with the ground truth human fixation map accordingly. A perfect prediction model needs to provide a distance close to 0 for the \({\chi}^2\) distance. ## Experiments on DIEM In our analysis, we first both qualitatively and quantitatively compare our proposed deep dynamic saliency networks, SSNet, TSNet, STSDirectNet, STSAvgNet, STSMaxNet and STSConvNet, with each other. Following the experimental setup of , we split the dataset into a training set containing 64 video sequences and a testing set including the remaining 20 representative videos covering different concepts. Specifically, we use the same set of splits used in . As our STSMaxNet and STSConvNet models integrate spatial and temporal streams by respectively using element-wise and convolutional fusion strategies, we perform an extensive set of initial experiments to determine the optimum layers for the fusion process to take place in these two-stream networks. In particular, we train STSMaxNet and STSConvNet models by considering different fusion layers, and test each one of them on the test set by considering sAUC measure. In Table [1](#tab:invest){reference-type="ref" reference="tab:invest"}, we provide these performance comparisons at different fusion layers. Interestingly, as can be seen from the table, fusing the spatial and temporal streams after the fifth convolution layer achieves the best results for both spatio-temporal networks. In the remainder, we use these settings for our STSMaxNet and STSConvNet models. In Figure [\[fig:network-comparisons\]](#fig:network-comparisons){reference-type="ref" reference="fig:network-comparisons"}, for some sample video frames we provide the outputs of the proposed networks along with the corresponding human fixation density maps (the ground truth). The input frames are given as overlayed images by compositing them with their consecutive frames to show the motion in the scenes. Saliency maps are shown as heatmaps superimposed over the original image for visualization purposes. As can be seen from these results, SSNet, which does employ appearance but not motion information, in general provides less accurate saliency maps and misses the foreground objects or their parts that are in motion. TSNet gives relatively better results, but as shown in the second and the third row, it does not identify all of the salient regions as it focuses more on the moving parts of the images. Directly averaging the saliency maps of these two single stream models, referred to as STSAvgNet, does not produce very satisfactory results either. As expected, STSMaxNet is slightly better since max fusion enforces to learn more effective filters in order to combine the spatial and temporal streams. Overall, STSConvNet is the best performing model. This can be rooted in the application of \(1 \times 1\) convolutional filters that learn the optimal weights to combine appearance and motion feature maps. [\[tab:invest\]]{#tab:invest label="tab:invest"} [\[tab:DIEMresults\]]{#tab:DIEMresults label="tab:DIEMresults"} In Table [2](#tab:DIEMresults){reference-type="ref" reference="tab:DIEMresults"}, we present the quantitative results averaged over all video sequences and frames on the test split of the DIEM dataset. Here, we compare and contrast our single-and two-stream saliency networks with eight existing dynamic saliency methods, namely GVBS , SR , PQFT , Seo and Milanfar , Rudoy *et al.* [^1], Fang *et al.* , Zhou *et al.*  and DWS  models. Among our deep saliency networks, we empirically find that STSDirectNet provides the worst quantitative results. This is in line with our observation in Table [1](#tab:invest){reference-type="ref" reference="tab:invest"} that delaying the integration of appearance and motion streams to a certain extent leads to more effective learning of mid and low level features. Secondly, we see that SSNet performs considerably lower than Temporal stream network, which demonstrates that motion is more vital for dynamic saliency. STSMaxNet gives results better than those of the single stream models but our STSConvNet model performs even better. It can be argued that STSConvNet learns more effective filters that combine spatial and temporal streams in an optimal manner. In addition, when we employ the data augmentation strategy proposed in the previous section, it further improves the overall performance of STSConvNet. In the remainder, we refer to this model with data augmentation as STSConvNet\*. When we compare our proposed STSMaxNet, STSConvNet, and STSConvNet\* models to the previous dynamic saliency methods, our results demonstrate the advantages of two-stream deep CNNs that they consistently outperform all those approaches, including the very recently proposed DWS model, according to five out of six evaluation measures. We present some qualitative results in Figure [\[fig:tab:model-comparisons\]](#fig:tab:model-comparisons){reference-type="ref" reference="fig:tab:model-comparisons"} where we again provide the input frames as transparent overlayed images showing the inherent motion. We observe that the proposed STSConvNet\* model localizes the salient regions more accurately than the existing models. For example, for the frame given in the first row, none of the compared models correctly capture the fixations over the painting brush. Similarly, for the second and the third frames, only our spatio-temporal saliency network fixates to the text and the cellular phone in the frames, respectively. ## Experiments on UCF-Sports Learning-based models might sometimes fail to provide satisfactory results for a test sample due to a shift from the training data domain. To validate generalization ability of our best-performing STSConvNet\* model, we perform additional experiments on UCF-Sports dataset. In particular, we do not carry out any training for our model from scratch or fine-tune it on UCF-Sports but rather use the predictions of the model trained only on DIEM dataset. [\[tab:ucf-results\]]{#tab:ucf-results label="tab:ucf-results"} In Table [3](#tab:ucf-results){reference-type="ref" reference="tab:ucf-results"}, we provide the performance of our model compared to the previous dynamic saliency models which are publicly available on the web. As can be seen, our STSConvNet\* model performs better than the state-of-the-art models according to majority of the evaluation measures. It especially outperforms the recently proposed DWS model in terms of all measures. These results suggest that our two-stream network generalizes well beyond the DIEM dataset. In Figure [\[fig:tab:early-ucf\]](#fig:tab:early-ucf){reference-type="ref" reference="fig:tab:early-ucf"}, we present sample qualitative results of Fang *et al.*  and DWS model  (two recently proposed models) and our STSConvNet model on some video frames. For instance, we observe that for the sample frame given in the first row, our model fixates to both the runner and the crowd. For the second and the third sample frames, the compared models do not accurately localize the weight lifter and the cowboys as salient, respectively. Similarly, the proposed STSConvNet\* model predicts the eye fixations better than the competing models for the fourth image containing a guardian walking in a corridor. For the last diving image, STSConvNet\* and Fang *et al.* give results fairly close to the ground truth, while DWS output some spurious regions as salient. ## Experiments on Still Images from MIT300 Deep static saliency networks achieve excellent performances on existing benchmark datasets for static saliency estimation. These models, however, only exploit spatial information captured in still images, but sometimes an image, despite being taken in an instant, might carry plenty of information regarding the inherent motion exist in it. In a recent study by Bylinski *et al.* , the authors demonstrate that the areas showing these kind of activities are indeed evidently important for saliency prediction since humans have tendency to look at the objects that they think in motion or that are in interaction with humans or some other objects. Motivated by these observations, in this section, we present the failures or the shortcomings of the current deep static saliency models through some examples, and show how motion information exist in still images can be utilized to fill in the semantic gap exist in the current static saliency models. Figure [\[fig:motioneffect\]](#fig:motioneffect){reference-type="ref" reference="fig:motioneffect"} presents sample images taken from  and which are from the MIT 300 dataset  where highly fixated regions (which cover the 95th percentile of the human fixation maps) are highlighted with yellow curves. As can be clearly seen from these examples, the state-of-the-art deep static models generally fail to give high saliency values to regions where an action occurs or which contains objects that are interpreted as in motion. To capture those regions, we employ the deep optical flow prediction model  which extracts optical flow from static images. Once we estimate the motion map of a still image, we can exploit this information together with the RGB image as inputs to our spatio-temporal saliency network (STSConvNet) to extract a saliency map. We observe that using these (possibly noisy) motion maps within our framework provides more accurate predictions than the existing deep static saliency models, and even captures the objects of gaze as illustrated in the first two sample images. These experiments reveal that the performances of static saliency networks can be improved by additionally considering motion information inherent in still images. # Conclusion In this work, we have investigated several deep architectures for predicting saliency from dynamic scenes. Two of these deep models are single-stream convolutional networks respectively trained for processing spatial and temporal information. Our proposed spatio-temporal saliency networks, on the other hand, are built based on two-stream architecture and employ different fusion strategies, namely direct averaging, max fusion and convolutional fusion, to integrate appearance and motion features, and they are all trainable in an end-to-end manner. While training these saliency networks, we additionally employ an effective and well-founded data augmentation method that utilizes low-resolution versions of the video frames and the ground truth saliency maps, giving a significant boost in performance. Our experimental results demonstrate that the proposed STSConvNet model achieves superior performance over the state-of-the-art methods on DIEM and UCF-Sports datasets. Lastly, we provide some illustrative example results on a number of challenging still images, which show that static saliency estimation can also benefit from motion information. This is left as an interesting topic for future research.

{'timestamp': '2017-11-16T02:04:49', 'yymm': '1607', 'arxiv_id': '1607.04730', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04730'}

# INTRODUCTION In recent years deep learning methods have demonstrated significant improvements in various tasks. These methods outperform human-level object detection in some tasks, and achieve state-of-the-art results in machine translation and speech processing. Additionally, deep learning combined with reinforcement learning techniques was able to beat human champions in challenging games such as Go. There are three different reasons for the outstanding results of Deep Learning models: 1. Powerful computing resources such as fast GPUs. 2. Utilizing efficiently large collections of datasets, e.g. ImageNet for image processing. 3. Advanced academic research on training methods and network architectures, , ,. Error correcting codes for channel coding are used in order to enable reliable communications at rates close to the Shannon capacity. A well-known family of linear error correcting codes are the low-density parity-check (LDPC) codes. LDPC codes achieve near Shannon channel capacity with the belief propagation (BP) decoding algorithm, but can typically do so for relatively large block lengths. For high density parity check (HDPC) codes, ,, , such as common powerful algebraic codes, the BP algorithm obtains poor results compared to the maximum likelihood decoder. In this work we focus on HDPC codes and demonstrate how the BP algorithm can be improved. The naive approach to the problem is to assume a neural network type decoder without restrictions, and train its weights using a dataset that contains a large amount of codewords. The training goal is to reconstruct the transmitted codeword from a noisy version after transmitting over the communication channel. Unfortunately, when using this approach our decoder is not given any side information regarding the structure of the code. In fact it is even not aware of the fact that the code is linear. Hence we are required to train the decoder using a huge collection of codewords from the code, and due to the exponential nature of the problem, this is infeasible, e.g., for a BCH(63,45) code we need a dataset of \(2^{45}\) codewords. On top of that, the database needs to reflect the variability due to the noisy channel. In order to overcome this issue, our proposed approach is to assign weights to the edges of the Tanner graph that represent the given linear code, thus yielding a "soft" Tanner graph. These edges are trained using deep learning techniques. A well-known property of the BP algorithm is the independence of the performance on the transmitted codeword. A major ingredient in our new method is that this property is preserved by our decoder. Thus it is sufficient to use a single codeword for training the parameters of our decoder. We demonstrate improvements over BP for various high density parity check codes, including BCH(63,36), BCH(63,45), and BCH(127,106). # THE BELIEF PROPAGATION ALGORITHM The renowned BP decoder, can be constructed from the Tanner graph, which is a graphical representation of some parity check matrix that describes the code. In this algorithm, messages are transmitted over edges. Each edge calculates its outgoing message based on all incoming messages it receives over all its edges, except for the message received on the transmitting edge. We start by providing an alternative graphical representation to the BP algorithm with \(L\) full iterations when using parallel (flooding) scheduling. Our alternative representation is a trellis in which the nodes in the hidden layers correspond to edges in the Tanner graph. Denote by \(N\), the code block length (i.e., the number of variable nodes in the Tanner graph), and by \(E\), the number of edges in the Tanner graph. Then the input layer of our trellis representation of the BP decoder is a vector of size \(N\), that consists of the log-likelihood ratios (LLRs) of the channel outputs. The LLR value of variable node \(v\), \(v=1,2,\ldots,N\), is given by \[l_v = \log\frac{\Pr\left(C_v=1 | y_v\right)}{\Pr\left(C_v=0 | y_v\right)}\] where \(y_v\) is the channel output corresponding to the \(v\)th codebit, \(C_v\). All the following layers in the trellis, except for the last one (i.e., all the hidden layers), have size \(E\). For each hidden layer, each processing element in that layer is associated with the message transmitted over some edge in the Tanner graph. The last (output) layer of the trellis consists of \(N\) processing elements that output the final decoded codeword. Consider the \(i\)th hidden layer, \(i=1,2,\ldots,2L\). For odd (even, respectively) values of \(i\), each processing element in this layer outputs the message transmitted by the BP decoder over the corresponding edge in the graph, from the associated variable (check) node to the associated check (variable) node. A processing element in the first hidden layer (\(i=1\)), corresponding to the edge \(e=(v,c)\), is connected to a single input node in the input layer: It is the variable node, \(v\), associated with that edge. Now consider the \(i\)th (\(i>1\)) hidden layer. For odd (even, respectively) values of \(i\), the processing node corresponding to the edge \(e=(v,c)\) is connected to all processing elements in layer \(i-1\) associated with the edges \(e'=(v,c')\) for \(c'\ne c\) (\(e'=(v',c)\) for \(v'\ne v\), respectively). For odd \(i\), a processing node in layer \(i\), corresponding to the edge \(e=(v,c)\), is also connected to the \(v\)th input node. The BP messages transmitted over the trellis graph are the following. Consider hidden layer \(i\), \(i=1,2,\ldots,2L\), and let \(e=(v,c)\) be the index of some processing element in that layer. We denote by \(x_{i,e}\), the output message of this processing element. For odd (even, respectively), \(i\), this is the message produced by the BP algorithm after \(\lfloor (i-1)/2 \rfloor\) iterations, from variable to check (check to variable) node. For odd \(i\) and \(e=(v,c)\) we have (recall that the self LLR message of \(v\) is \(l_v\)), \[x_{i,e=(v,c)} = l_v + \sum_{e'=(v,c'),\: c'\ne c} x_{i-1,e'} \label{eq:x_ie_RB}\] under the initialization, \(x_{0,e'}=0\) for all edges \(e'\) (in the beginning there is no information at the parity check nodes). The summation in [\[eq:x_ie_RB\]](#eq:x_ie_RB){reference-type="eqref" reference="eq:x_ie_RB"} is over all edges \(e'=(v,c')\) with variable node \(v\) except for the target edge \(e=(v,c)\). Recall that this is a fundamental property of message passing algorithms . Similarly, for even \(i\) and \(e=(v,c)\) we have, \[x_{i,e=(v,c)} = 2\tanh^{-1} \left( \prod_{e'=(v',c),\: v'\ne v} \tanh \left( \frac{x_{i-1,e'}}{2} \right) \right) \label{eq:x_ie_LB}\] The final \(v\)th output of the network is given by \[o_v = l_v + \sum_{e'=(v,c')} x_{2L,e'} \label{eq:ov}\] which is the final marginalization of the BP algorithm. # THE PROPOSED DEEP NEURAL NETWORK DECODER We suggest the following parameterized deep neural network decoder that generalizes the BP decoder of the previous section. We use the same trellis representation for the decoder as in the previous section. The difference is that now we assign weights to the edges in the Tanner graph. These weights will be trained using stochastic gradient descent which is the standard method for training neural networks. More precisely, our decoder has the same trellis architecture as the one defined in the previous section. However, Equations [\[eq:x_ie_RB\]](#eq:x_ie_RB){reference-type="eqref" reference="eq:x_ie_RB"}, [\[eq:x_ie_LB\]](#eq:x_ie_LB){reference-type="eqref" reference="eq:x_ie_LB"} and [\[eq:ov\]](#eq:ov){reference-type="eqref" reference="eq:ov"} are replaced by \[x_{i,e=(v,c)} = \tanh \left(\frac{1}{2}\left(w_{i,v} l_v + \sum_{e'=(v,c'),\: c'\ne c} w_{i,e,e'} x_{i-1,e'}\right)\right) \label{eq:x_ie_RB_NN}\] for odd \(i\), \[x_{i,e=(v,c)} = 2\tanh^{-1} \left( \prod_{e'=(v',c),\: v'\ne v}{x_{i-1,e'}}\right) \label{eq:x_ie_LB_NN}\] for even \(i\), and \[o_v = \sigma \left( w_{2L+1,v} l_v + \sum_{e'=(v,c')} w_{2L+1,v,e'} x_{2L,e'} \right) \label{eq:ov_NN}\] where \(\sigma(x) \equiv \left( 1+e^{-x} \right)^{-1}\) is a sigmoid function. The sigmoid is added so that the final network output is in the range \([0,1]\). This makes it possible to train the network using a cross entropy loss function, as described in the next section. Apart of the addition of the sigmoid function at the outputs of the network, one can see that by setting all weights to one, Equations [\[eq:x_ie_RB_NN\]](#eq:x_ie_RB_NN){reference-type="eqref" reference="eq:x_ie_RB_NN"}-[\[eq:ov_NN\]](#eq:ov_NN){reference-type="eqref" reference="eq:ov_NN"} degenerate to [\[eq:x_ie_RB\]](#eq:x_ie_RB){reference-type="eqref" reference="eq:x_ie_RB"}-[\[eq:ov\]](#eq:ov){reference-type="eqref" reference="eq:ov"}. Hence by optimal setting (training) of the parameters of the neural network, its performance can not be inferior to plain BP. It is easy to verify that the proposed message passing decoding algorithm [\[eq:x_ie_RB_NN\]](#eq:x_ie_RB_NN){reference-type="eqref" reference="eq:x_ie_RB_NN"}-[\[eq:ov_NN\]](#eq:ov_NN){reference-type="eqref" reference="eq:ov_NN"} satisfies the message passing symmetry conditions. Hence, by, when transmitting over a binary memoryless symmetric (BMS) channel, the error rate is independent of the transmitted codeword. Therefore, to train the network, it is sufficient to use a database which is constructed by using noisy versions of a single codeword. For convenience we use the zero codeword, which must belong to any linear code. The database reflects various channel output realizations when the zero codeword has been transmitted. The goal is to train the parameters \(\left \{ w_{i,v},w_{i,e,e'},w_{i,v,e'} \right \}\) to achieve an \(N\) dimensional output word which is as close as possible to the zero codeword. The network architecture is a non-fully connected neural network. We use stochastic gradient descent to train the parameters. The motivation behind the new proposed parameterized decoder is that by setting the weights properly, one can compensate for small cycles in the Tanner graph that represents the code. That is, messages sent by parity check nodes to variable nodes can be weighted, such that if a message is less reliable since it is produced by a parity check node with a large number of small cycles in its local neighborhood, then this message will be attenuated properly. The time complexity of the deep neural network is similar to plain BP algorithm. Both have the same number of layers and the same number of non-zero weights in the Tanner graph. The deep neural network architecture is illustrated in Figure [\[fig:BCH_15_11_arch\]](#fig:BCH_15_11_arch){reference-type="ref" reference="fig:BCH_15_11_arch"} for a BCH(15,11) code. # EXPERIMENTS ## Neural Network Training We built our neural network on top of the TensorFlow framework  and used an NVIDIA Tesla K40c GPU for accelerated training. We applied cross entropy as our loss function, \[L{(o,y)}=-\frac{1}{N}\sum_{v=1}^{N}y_{v}\log(o_{v})+(1-y_{v})\log(1-o_{v}) \label{eq:cross_entropy}\] where \(o_{v}\), \(y_{v}\) are the deep neural network output and the actual \(v\)th component of the transmitted codeword (if the all-zero codeword is transmitted then \(y_{v}=0\) for all \(v\)). Training was conducted using stochastic gradient descent with mini-batches. The mini-batch size was \(120\) examples. We applied the RMSPROP  rule with a learning rate equal to \(0.001\). The neural network has \(10\) hidden layers, which correspond to \(5\) full iterations of the BP algorithm. Each processing element in an odd (even, respectively) indexed hidden layer is described by Equation [\[eq:x_ie_RB_NN\]](#eq:x_ie_RB_NN){reference-type="eqref" reference="eq:x_ie_RB_NN"} (Equation [\[eq:x_ie_LB_NN\]](#eq:x_ie_LB_NN){reference-type="eqref" reference="eq:x_ie_LB_NN"}, respectively). At test time, we inject noisy codewords after transmitting through an AWGN channel and measure the bit error rate (BER) in the decoded codeword at the network output. When computing [\[eq:x_ie_RB_NN\]](#eq:x_ie_RB_NN){reference-type="eqref" reference="eq:x_ie_RB_NN"}, we also clip the input to the tanh function such that the absolute value of the input is always smaller than some positive constant \(A < 10\). This is also required for practical (finite block length) implementations of the BP algorithm, in order to stabilize the operation of the decoder. We trained our decoding network on few different linear codes, including BCH(15,11), BCH(63,36), BCH(63,45) and BCH(127,106). ## Neural Network Training With Multiloss The proposed neural network architecture has the property that after every odd \(i\) layer we can add final marginalization. We can use that property to add additional terms in the loss. The additional terms can increase the gradient update at the backpropagation algorithm and allow learning the lower layers. At each odd \(i\) layer we add the final marginalization to loss: \[L{(o,y)}=-\frac{1}{N}\sum_{i=1,3}^{2L-1}\sum_{v=1}^{N}y_{v}\log(o_{v,i})+(1-y_{v})\log(1-o_{v,i}) \label{eq:multiloss_cross_entropy}\] where \(o_{v,i}\), \(y_{v}\) are the deep neural network output at the odd \(i\) layer and the actual \(v\)th component of the transmitted codeword. This network architecture is illustrated in Figure 2. ## Dataset The training data is created by transmitting the zero codeword through an AWGN channel with varying SNRs ranging from \(1{\rm dB}\) to \(6{\rm dB}\). Each mini batch has \(20\) codewords for each SNR (a total of \(120\) examples in the mini batch). For the test data we use codewords with the same SNR range as in the training dataset. Parity check matrices were taken from. ## Results In this section we present the results of the deep neural decoding networks for various BCH block codes. In each code we observed an improvement compared to the BP algorithm. Note that when we applied our algorithm to the BCH(15,11) code, we obtained close to maximum likelihood results with the deep neural network. For larger BCH codes, the BP algorithm and the deep neural network still have a significant gap from maximum likelihood. The BER figures [\[fig:BCH_63_36_ber\]](#fig:BCH_63_36_ber){reference-type="ref" reference="fig:BCH_63_36_ber"}, [\[fig:BCH_63_45_ber\]](#fig:BCH_63_45_ber){reference-type="ref" reference="fig:BCH_63_45_ber"} and [\[fig:BCH_127_106_ber\]](#fig:BCH_127_106_ber){reference-type="ref" reference="fig:BCH_127_106_ber"} show an improvement of up to \(0.75{\rm dB}\) in the high SNR region. Furthermore, the deep neural network BER is consistently smaller or equal to the BER of the BP algorithm. This result is in agreement with the observation that our network cannot perform worse than the BP algorithm. Figure [\[fig:BCH_63_45_multiloss_ber\]](#fig:BCH_63_45_multiloss_ber){reference-type="ref" reference="fig:BCH_63_45_multiloss_ber"} shows the results of training the deep neural network with multiloss. It shows an improvement of up to \(0.9{\rm dB}\) compared to the plain BP algorithm. Moreover, we can observe that we can achieve the same BER performance of 50 iteration BP with 5 iteration of the deep neural decoder, This is equal to complexity reduction of factor 10. We compared the weights of the BP algorithm and the weights of the trained deep neural network for a BCH(63,45) code. We observed that the deep neural network produces weights in the range from \(-0.8\) to \(2.2\), in contrast to the BP algorithm which has binary \(“1”\) or \(“0”\) weights. Figure [\[fig:weight_hist\]](#fig:weight_hist){reference-type="ref" reference="fig:weight_hist"} shows the weights histogram for the last layer. Interestingly, the distribution of the weights is close to a normal distribution. In a similar way, every hidden layer in the trained deep neural network has a close to normal distribution. Note that Hinton recommends to initialize the weights with normal distribution. In Figures [\[fig:layer4_bp\]](#fig:layer4_bp){reference-type="ref" reference="fig:layer4_bp"} and [\[fig:layer4_dl\]](#fig:layer4_dl){reference-type="ref" reference="fig:layer4_dl"} we plot the weights of the last hidden layer. Each column in the figure corresponds to a neuron described by Equation  [\[eq:x_ie_RB_NN\]](#eq:x_ie_RB_NN){reference-type="eqref" reference="eq:x_ie_RB_NN"}. It can be observed that most of the weights are zeros except the Tanner graph weights which have value of 1 in Figure [\[fig:layer4_bp\]](#fig:layer4_bp){reference-type="ref" reference="fig:layer4_bp"} (BP algorithm) and some real number in Figure [\[fig:layer4_dl\]](#fig:layer4_dl){reference-type="ref" reference="fig:layer4_dl"} for the neural network. In Figure [\[fig:layer4_bp\]](#fig:layer4_bp){reference-type="ref" reference="fig:layer4_bp"} and  [\[fig:layer4_dl\]](#fig:layer4_dl){reference-type="ref" reference="fig:layer4_dl"} we plot a quarter of the weights matrix for better illustration. # CONCLUSIONS In this work we applied deep learning techniques to improve the performance of the BP algorithm. We showed that a "soft" Tanner graph can produce improvements when used in the BP algorithm instead of the standard Tanner graph. We believe that the BER improvement was achieved by properly weighting the messages, such that the effect of small cycles in the Tanner graph was partially compensated. It should be emphasized that the parity check matrices that we worked with were obtained from. We have not evaluated our method on parity check matrices for which an attempt was made to reduce the number of small cycles. A notable property of our neural network decoder is that once we have trained its parameters, we can improve performance compared to plain BP without increasing the required computational complexity. Another notable property of the neural network decoder is that we learn the channel and the linear code simultaneously. We regard this work as a first step in the implementation of deep learning techniques for the design of improved decoders. Our future work include possible improvements in the error rate results by exploring new neural network architectures and combining other decoding methods. Furthermore, we plan to investigate the connection between the parity check matrix and the deep neural network decoding capabilities.

{'timestamp': '2016-10-03T02:05:04', 'yymm': '1607', 'arxiv_id': '1607.04793', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04793'}

# Introduction {#sec:intro} Microquasars are X-ray binaries that exhibit spatially resolved, extended radio emission. The nomenclature is motivated by a structural similarity with the quasar family of active galactic nuclei (AGN). Both object classes are believed to comprise a compact central object embedded in a flow of accreting material, and both exhibit relativistic, collimated jets. In the current paradigm, both microquasars and AGN derive their power from the gravitational potential energy that is liberated as ambient matter falls onto the compact object. Notwithstanding their morphological resemblance, microquasars and radio-loud AGN represent complementary examples of astrophysical jet production on dramatically disparate spatial and temporal scales. Indeed, conditions of accretion and mass provision that pertain to the supermassive (\(10^{6}M_{\odot}\lesssim M_{\rm BH} \lesssim 10^{9}M_{\odot}\)) black holes that power AGN and of the stellar-mass compact primaries of microquasars are markedly different. In the latter, a companion star (or donor) provides the reservoir of matter for accretion onto a compact stellar remnant (or primary), which can be either a neutron star or a black hole. Partial dissipation of the resultant power output occurs in a disk of material surrounding the primary, producing the thermal and non-thermal X-ray emission, which is characteristic of all X-ray binary systems. Microquasars are segregated on the basis of associated non-thermal radio emission, indicative of synchrotron radiation in a collimated outflow, which carries away a sizeable fraction of the accretion luminosity. In AGN, superficially similar jet structures are known to be regions of particle acceleration and non-thermal photon emission. The resulting radiation spectrum can extend from radio wavelengths into the very high-energy (VHE; \(E_{\gamma}>100\) GeV) \(\gamma\)-ray regime. Very high-energy \(\gamma\)-ray emission has been observed from many AGN in the blazar sub-class[^1], where the jet axis is aligned close to the observer line-of-sight, as well as from a few radio galaxies (e.g. M87, [@2003A&A...403L...1A]; Cen A, [@2009ApJ...695L..40A]; NGC 1275, [@Aleksic2012]) and starburst galaxies (e.g. M82, [@Acciari2009]; NGC 253, [@Abramowski2012]). If similar jet production and efficient particle acceleration mechanisms operate in microquasars and AGNs, this might imply that the former object class are plausible sources of detectable VHE \(\gamma\)-ray emission as well, assuming that appropriate environmental conditions prevail. The primarily relevant environmental conditions include the density of nearby hadronic material, which provides scattering targets for inelastic proton scattering interactions; these interactions produce pions that produce \(\gamma\)-rays when they subsequently decay. The ambient magnetic field strength is also important and influences the rate at which electrons lose energy via synchrotron radiation. Synchrotron photons contribute to the reservoir of soft photons that are available for inverse Compton (IC) up-scattering into the VHE \(\gamma\)-ray regime. The argument for phenomenological parity between AGN and microquasars, possibly related to their structural resemblance, has been strengthened in recent years as the spectral properties of both radio and X-ray emission are remarkably similar for both stellar mass and supermassive black holes. In recent years these similarities led to the postulation of a so-called fundamental plane, which describes a three-dimensional, phenomenological correlation between the radio (5 GHz) and X-ray (2--10 keV) luminosities and the black hole mass. However, the fundamental plane does not appear to extend into the TeV band. To date, only one well-established microquasar has been observed to emit in the VHE \(\gamma\)-ray regime. This is the Galactic black hole Cygnus X-1, which was marginally detected (at the \(\sim4\sigma\) level) by the MAGIC telescope immediately prior to a \(2-50\,\)keV X-ray flare observed by the INTEGRAL satellite, the *Swift* Burst Alert Telescope (BAT), and the *RXTE* All-Sky Monitor (ASM). recently identified linear polarized soft \(\gamma\)-ray emission from Cygnus X-1 (see also [@Jourdain2012]), thereby locating the emitter within the jets and identifying their capacity to accelerate particles to high energies (see however [@Romero2014]). Further motivation for observing microquasars in the VHE band arises from the recent identification of the high-mass microquasar Cygnus X-3 as a transient high-energy (HE; 100MeV \(<E_{\gamma}<\) 100 GeV) \(\gamma\)-ray source by the *Fermi* and AGILE satellites. The identification of a periodic modulation of the HE signal is consistent with the orbital frequency of Cygnus X-3 and provides compelling evidence for effective acceleration of charged particles to GeV energies within the binary system. Based on evidence from subsequent reobservations, the HE \(\gamma\)-ray flux from Cygnus X-3 appears to be correlated with transitions observed in X-rays in and out of the so-called ultra-soft state, which exhibits bright soft X-ray emission and low fluxes in hard X-rays and is typically associated with contemporaneous radio flaring activity. Unfortunately, repeated observations of Cygnus X-3 using the MAGIC telescope did not yield a significant detection, despite the inclusion of data that were obtained simultaneously with the periods of enhanced HE emission detected using *Fermi*. However, the intense optical and ultraviolet radiation fields produced by the Wolf-Rayet companion star in Cygnus X-3 imply a large optical depth for VHE \(\gamma\)-ray s due to absorption via \(e^{+}e^{-}\) pair production. Accordingly, particle acceleration mechanisms akin to those operating in Cygnus X-3 may yield detectable VHE \(\gamma\)-ray fluxes in systems with fainter or cooler donors. Mechanisms for \(\gamma\)-ray production in microquasars have been widely investigated, resulting in numerous hadronic and leptonic models, describing the expected fluxes and spectra of microquasars in the GeV-TeV band. In both scenarios, a highly energetic population of the relevant particles is required and, consequently, emission scenarios generally localize the radiating region within the jet structures of the microquasar. Leptonic models rely upon IC scattering of photons from the primary star in the binary system or photons produced through synchrotron emission along the jet to produce VHE \(\gamma\)-ray emission. In this latter scenario, they closely resemble models of extragalactic jets, but typically invoke internal magnetic fields that are stronger by factors \(\sim1000\). Consideration of hadronic models is motivated by the detection of Doppler-shifted emission lines associated with the jets of the microquasar SS 433, indicating that at least some microquasar jets comprise a significant hadronic component. Models of VHE \(\gamma\)-ray production by hadronic particles generally invoke electromagnetic cascades initiated by both neutral and charged pion decays. Electron-positron pair production, \(\gamma\gamma\rightarrow e^{+}e^{-}\) can absorb VHE \(\gamma\)-rays. In the case of 1 TeV \(\gamma\)-rays, the cross section for this process is maximised for ultraviolet target photons (\(E_{\rm ph}\sim10\) eV), where its value may be approximated in terms of the Thomson cross section as \(\sigma_{\gamma\gamma}\approx\sigma_{T}/5\). In high-mass systems, the companion star is expected to produce a dense field of these target photons to interact with the \(\gamma\)-rays. This process can be very significant and probably contributes to the observed orbital modulation in the VHE \(\gamma\)-ray flux from LS 5039. In contrast, the ultraviolet spectrum of low-mass microquasars is likely dominated by the reprocessing of X-ray emission in the cool outer accretion flow, although jet emission might also be significant. Regardless of their origin, the observed optical and ultraviolet luminosities of low-mass X-ray binaries (LMXBs) are generally orders of magnitude lower than those of high-mass systems, and the likelihood of strong \(\gamma\)-ray absorption is correspondingly reduced. However, microquasars may only become visible in the TeV band during powerful flaring events. These transient outbursts, characterised by the ejection of discrete superluminal plasma clouds, are usually observed at the transition between low-and high-luminosity X-ray states. Monitoring black-hole X-ray binaries with radio telescopes and X-ray satellites operating in the last decade enabled a classification scheme of such events to be established. Hardness-intensity diagrams (HIDs) plot the source X-ray intensity against X-ray colour (or hardness) and have subsequently been extensively used to study the spectral evolution of black-hole outbursts. At the transition from the so-called low-hard state to the high-soft states through the hard-to-soft intermediate states, the steady jet associated with the low-hard state is disrupted. These transient ejections, produced once the accretion disk collapses inwards, are more relativistic than the steady low-hard jets. Internal shocks can develop in the outflow, possibly accelerating particles that subsequently give rise to radio optically thin flares observed from black-hole systems; this phenomenological description is also extensible to neutron stars, although in that case jet radio power is lower by a factor 5--30 ([@2006MNRAS.366...79M]). Outburst episodes have also been observed in cases in which the source remained in the hard state without transition to the soft state. The detection (at the \(\sim4\sigma\) level) by the MAGIC telescope of the high-mass, black-hole binary Cygnus X-1 took place during an enhanced \(2-50\,\)keV flux low-hard state as observed with the INTEGRAL satellite, the *Swift* BAT, and the *RXTE* ASM. However, although the source X-ray spectrum remained unchanged throughout the TeV flare, such a bright hard state was unusually long when compared with previous observations of the source. Here we report on contemporaneous observations with H.E.S.S. and *RXTE* of the three microquasars V4641 Sgr, GRS 1915+105, and Circinus X-1. Information on the targets, the H.E.S.S. and *RXTE* observations, and the corresponding trigger conditions are detailed in Sect. [2](#sec:targets){reference-type="ref" reference="sec:targets"}. Analysis results are reported in Sect. [3](#sec:mqs:results){reference-type="ref" reference="sec:mqs:results"} and discussed in Sect. [4](#sec:mqs:context){reference-type="ref" reference="sec:mqs:context"}. In the appendix, detailed information on the X-ray analysis is reported, which in particular includes HIDs corresponding to the time of observations for the three studied sources. # Targets and observations {#sec:targets} ## Observations {#sec:mqs:observations} The H.E.S.S. Imaging Atmospheric Cherenkov Telescope (IACT) array is situated on the Khomas Highland plateau of Namibia (23\(^{\circ}16'18\arcsec\) south, \(16^{\circ}30'00\arcsec\) east), at an elevation of 1800 m above sea level, and is capable of detecting a Crab-like source close to the zenith at the 5\(\sigma\) level within \(<5\) minutes under good observational conditions. The point source sensitivity of H.E.S.S. enables it to detect a \(2.0\times10^{-13}~\rm{cm}^{-2}\rm{s}^{-1}\) \(\gamma\)-ray flux above 1 TeV, at the 5\(\sigma\) level within 25 hours, which, together with a low-energy threshold (\(\sim\)`<!-- -->`{=html}100 GeV), makes H.E.S.S. an invaluable instrument for studying the VHE \(\gamma\)-ray emission from microquasars. A fifth and larger telescope (commissioned in 2013) will allow the energy threshold to be lowered and will further increase the sensitivity of the instrument. For the analysis presented here, H.E.S.S. observations were carried out using the full, original four-telescope array. Owing to the diverse morphologies of the three binary systems, unique observational trigger criteria were established for each target employing various combinations of the observed X-ray state and radio flaring activity. Details are provided in subsequent paragraphs. The *Rossi X-ray Timing Explorer* (*RXTE*) was a space-based X-ray observatory launched on 30 December 30 1995 and decommissioned on 5 January 2012. The primary mission of *RXTE* was to provide astrophysical X-ray data with high timing resolution. This observatory occupied a circular low-earth orbit with an orbital period of \(\sim90\) minutes and carried three separate X-ray telescopes. The *Proportional Counter Array* (PCA) on board *RXTE* comprised five copointing xenon and propane *Proportional Counter Units* (PCUs), which were nominally sensitive in the energy range \(\sim2-60\) keV with an energy resolution of \(<18\%\) at 6 keV. For studies of rapidly varying sources like X-ray binaries, the PCA timing resolution of \(\sim1\;\mu\)s can prove invaluable. However, rapid timing measurements also require a bright source to provide sufficient counting statistics within such short time bins. The *High Energy X-ray-Timing Experiment* (HEXTE) comprised two independent clusters of four phoswich scintillation detectors, which were sensitive to photons in the \(\sim12-250\) keV energy range and had an energy resolution of \(\sim9\) keV at 60 keV. The maximum timing resolution of HEXTE was \(\sim8\;\mu\)s. The *All-Sky Monitor* (ASM) was a wide field-of-view instrument that monitored \(\sim80\%\) of the sky over the course of each \(\sim90\) minute orbit. This instrument consisted of three identical scanning shadow cameras and was designed to provide near-continuous monitoring of bright X-ray sources. Nominally, the ASM was sensitive in the energy range from \(2-10\,\)keV and had a rectangular field of view spanning \(110^{\circ}\times12^{\circ}\). Contemporaneous X-ray (*RXTE*) and VHE \(\gamma\)-ray (H.E.S.S.) observations were performed at the epochs listed in Table [\[table:targets\]](#table:targets){reference-type="ref" reference="table:targets"}. In the following, we briefly review the observational characteristics of the target microquasars, GRS 1915+105, Circinus X-1, and V4641 Sgr. Established system parameters that characterise the three target microquasars are collated in Table [\[tab:target_properties\]](#tab:target_properties){reference-type="ref" reference="tab:target_properties"}. ## GRS 1915+105 {#sec:upperlims:grs1915} GRS 1915+105 is a dynamically established black-hole binary first identified by the WATCH all-sky monitor on board the *GRANAT* satellite. Observations in the optical and near-infrared using the Very Large Telescope succeeded in identifying the stellar companion as a low-mass KM III giant. GRS 1915+105 gained a measure of celebrity as the prototype Galactic superluminal source. In a detailed study of the X-ray light curves of GRS 1915+105, succeeded in identifying 12 distinct variability classes, internally characterised by the duration and juxtaposition of three separate spectral states. Episodes of class \(\chi\) behaviour, belonging to state C and lasting several days, are known as plateaux and are invariably terminated by flaring activity in the radio, infrared, and X-ray bands. In contrast with the evidence for self-absorbed synchrotron radiation seen in the spectrally hard, low-luminosity state C, and often associated with continuous relativistic jets, radio spectra obtained during the end-plateau flaring episodes indicate optically thin synchrotron emission. Occasionally, these flaring episodes are linked to powerful discrete plasma ejections with instantaneous power output reaching \(\gtrsim10^{39}\) erg s\(^{-1}\). Modelling the emission from these discrete relativistic ejecta, showed that inverse Comptonisation of emitted synchrotron photons into the GeV-TeV regime could produce significant and persistent \(\gamma\)-ray fluxes that remain detectable for several days. and reported VHE \(\gamma\)-ray observations of GRS 1915+105; these authors derived integral flux upper limits of \(6.1\times10^{-13}\) cm\(^{-2}\)s\(^{-1}\) above 410 GeV and \(1.17 \times10^{-12}\) cm\(^{-2}\)s\(^{-1}\) above 250 GeV, respectively. For the analysis presented here, GRS 1915+105 was observed by H.E.S.S. between 28 April and 3 May 2004 in response to an apparent decrease in the 15 GHz radio flux, which was monitored by the Ryle Telescope during a \(\sim50\) day plateau state, as shown in Figure [\[evol_GRS1915\]](#evol_GRS1915){reference-type="ref" reference="evol_GRS1915"}, in which coloured markers indicate the H.E.S.S. observation epochs. On the basis of previously observed behaviour, it was thought likely that the observed radio evolution signalled the end of the plateau state and, therefore, that flaring activity would begin within the subsequent 24 hours. The *RXTE* observations of GRS 1915+105 comprised six individual pointings, contributing to accumulated PCA and HEXTE livetimes of \(7.6\,\)ksec and \(5176\,\)s, respectively. Fifteen contemporaneous H.E.S.S. observations were obtained, constituting an overall livetime of 6.9 hours. ## Circinus X-1 {#sec:upperlims:cirx1} Circinus X-1 (hereafter Cir X-1) has been extensively studied since its initial identification, revealing a somewhat confusing collection of complex observational characteristics. Repeated observation of type I X-ray bursts definitively identifies the compact primary in Cir X-1 as a low magnetic field (\(B\lesssim10^{11}\)G) neutron star. Further sub-classification as a Z or atoll source is not possible since Cir X-1 exhibits several confusing spectral and timing properties, subsets of which are characteristic of both source types. Accordingly, established paradigms for disk-jet coupling in X-ray binaries with neutron star primaries cannot be reliably employed. At radio wavelengths, the jets of Cir X-1 display notable structure on arcsecond scales, appearing as a bright core with significant extension along the axial direction of the arcminute jets. In fact, the observed extension is rather asymmetric with a ratio of at least two between the observed fluxes of the two opposing jets. Interpreted as pure relativistic aberration, this asymmetry implies a jet velocity \(\gtrsim0.1c\). Cir X-1 has also been observed to eject condensations of matter with apparently superluminal velocities \(\gtrsim15c\). These observations imply a physical velocity for the ejecta \(v>0.998c\) with a maximum angle between the velocity vector and the line of sight \(\theta<5^{\circ}\).These results identify Cir X-1 as a microblazar, a Galactic, small-scale analog of the blazar class of AGN, several of which are known sources of VHE \(\gamma\)-rays. Definitive classification of the donor star in Cir X-1 is somewhat problematic. The low apparent magnitude of the detected optical counterpart implies a dereddened luminosity consistent with a low-mass or sub-giant companion, implying that Cir X-1 is a LMXB with a high orbital eccentricity \(e\sim0.7-0.9\). Nonetheless, recent near-infrared and I-band optical observations reveal emission features that are consistent with a mid-B supergiant, suggesting a more moderate eccentricity \(e\sim0.45\). Observations of Cir X-1 in the X-ray band reveal a long-term evolution of the average source brightness. Fluxes rose monotonically from near-undetectable in the early 1970s to a peak value of \(\sim 1.5-2\) Crab (\(1.5-10\,\)keV) at the turn of the millennium, before returning over a period of \(\sim4\) years to their pre-rise levels. Various X-ray spectra, obtained during epochs of both high and low flux, display evidence of complex and variable emission and absorption processes. A previous analysis of H.E.S.S. observations of Cir X-1 was presented by, who derived a preliminary upper limit to the \(\gamma\)-ray flux above 1 TeV of \(1.9\times10^{-13}\) cm\(^{-2}\)s\(^{-1}\) corresponding to a detector livetime of 28 hours. The H.E.S.S. observations of Cir X-1 reported here began on 18 June 2004 and were scheduled to coincide with the periastron passage of the binary components. The previous observation of regular radio flares during this orbital interval were thought to provide a good chance of observing during a period of outburst with the associated possibility that superluminal ejections might occur. The *RXTE* observations of Cir X-1 comprised three individual pointings, corresponding to orbital phase intervals \(0.0486\le\phi\le0.0498\), \(0.1104\le\phi\le0.1112,\) and \(0.1718\le\phi\le0.1725\) (using the radio flare ephemeris of ), and contributing to an accumulated PCA livetime of \(2576\,\)s. A data set comprising 12 contemporaneous H.E.S.S. observations yielded a combined livetime of 5.4 hours. ## V4641 Sgr {#sec:upperlims:v4641sgr} V4641 Sgr is the optical designation of the habitually weak X-ray source SAX J1819.3-2525 (XTE J1819-254), which was independently identified using the *BeppoSAX* and *RXTE* satellites. Optical spectroscopic measurements strongly suggest a late B-or early A-type companion with an effective temperature \(T_{\rm{eff}}\approx10500\) K. The mass of the compact primary, \(6.4\pm0.6\, M_{\odot}\), categorises V4641 Sgr as a firm black hole candidate. V4641 Sgr is probably best known for its exhibition of rapid and violent outbursts. Perhaps the most spectacular of these events was the super-Eddington flare detected by the *RXTE ASM* in September 1999. The observed X-ray fluxes (\(2-12\,\)keV) increased sharply, reaching \(\approx12.2\) Crab within eight hours before fading again to below 0.1 Crab in under two hours. Powerful contemporaneous flares were also observed at hard X-ray, optical, and radio wavelengths. In fact, Very Large Array (VLA) radio observations obtained within a day of the X-ray flare resolved a bright jet-like radio structure \(\approx0.25\) arcsec in length. Assuming the most likely hypothesis, i.e., that the ejection is coincident with some phase of the X-ray flare, proper motions in the range \(0.22\lesssim\mu_{\rm{jet}}\lesssim1.1\) arcsec day\(^{-1}\) are derived. At the minimum distance \(d=5.5\) kpc, the implied lower limit to the apparent velocity of the ejecta is \(7c\lesssim v_{\rm{min}}\lesssim35c\), which is comparable with the extragalactic jets seen in blazars. Indeed, the remarkably high apparent velocities imply that V4641 Sgr may be a microblazar with a relativistic jet moving close to the line of sight (\(\theta_{\rm{jet}}\lesssim12^{\circ}\); from ). Subsequent, weaker broadband outbursts have also been observed, suggesting recurrent activity on a timescale \(\sim1-2\) years. Observations of V4641 Sgr with H.E.S.S. were initiated on 7 July 7 2004 (MJD 53193) in response to the source brightening rapidly in the radio, optical, and X-ray bands. The resultant *RXTE* exposure comprised three observations, each contributing to an accumulated PCA livetime of \(5\,\)ksec. Two pairs of \(\sim\)`<!-- -->`{=html}30 minute H.E.S.S. observations were obtained contemporaneously with the final two *RXTE* pointings. In total, the four separate exposures constitute an overall livetime of 1.76 hours. # Analysis and results {#sec:mqs:results} X-ray data reduction with the `FTOOLS 5.3.1` software suite employed the data selection criteria regarding elevation, offset, electron contamination, and proximity to the South Atlantic Anomaly recommended by the *RXTE* Guest Observer Facility website[^2]. For each observation, the PCA STANDARD2 data were extracted from all available PCUs. For all observations, HEXTE Archive mode data for both clusters were extracted following the recommended procedures for time filtering and background estimation. Spectral analysis was carried out using the `XSPEC 12.6.0` package. Spectral fits for GRS 1915+105 use both PCA and HEXTE data, including an energy range of \(3-200\,\)keV. For bright X-ray sources, such as GRS 1915+105, statistical errors on the number of counts per spectral bin become insignificant relative to dominant uncertainties in the instrument response. Accordingly, a 1% systematic error was added to all PCA channels. The remaining sources, Cir X-1 and V4641 Sgr, were not significantly detected by HEXTE and therefore only PCA data in the \(3-20\,\)keV range were considered to ensure good data quality. These targets were sufficiently faint that the spectral bin uncertainties were statistically dominated and the addition of a systematic error component was not required. In the case of GRS 1915+105, power density spectra (PDS) were derived using the ftool `powspec`. For each *RXTE* pointing of GRS 1915+105, individual PDS were extracted from \(128\,\)s intervals comprising \(2^{14}\) bins. The resulting spectra were then averaged to produce a PDS for the total light curve with errors estimated using the standard deviation of the average of the power in each frequency bin. The overall PDS were logarithmically rebinned and normalised to represent the squared fractional RMS in each frequency bin. Corrections for instrument deadtime were applied (although this was found to have a negligible effect in the frequency range under consideration) and the expected white noise level was subtracted. Similar temporal analyses for the remaining targets proved unfeasible because of insufficient count statistics at all but the lowest frequencies. The \(\gamma\)-ray analysis followed the standard point-source procedure described in. The reflected background model was used to derive overall results in conjunction with both the *hard* and *standard* event selection cuts described by. Hard cuts (image size \(\ge200\) photoelectrons) tend to enhance the signal of sources with power-law spectral slopes that are harder in comparison to the dominant cosmic ray background. Standard cuts (image size \(\ge80\) photoelectrons) provide less sensitivity in such cases but allow a lower energy threshold. No significant detection was obtained for any of the three targets. Upper limits to the VHE \(\gamma\)-ray flux above the instrumental threshold energy were therefore derived at the 99% confidence level using the profile likelihood method. ## GRS 1915+105 {#subsec:grs_res} As illustrated by the PCA and ASM light curves shown in Figure [\[fig:grs_lightcurves\]](#fig:grs_lightcurves){reference-type="ref" reference="fig:grs_lightcurves"}, the X-ray count rate was stable to within \(\sim10\%\) during each observation and varied by no more than \(\sim20\%\) between observations. Indeed, the long-term *RXTE* ASM light curve in Figure [\[fig:grs_lightcurves\]](#fig:grs_lightcurves){reference-type="ref" reference="fig:grs_lightcurves"} (top panel) clearly indicates that the H.E.S.S. observation epochs occurred during an extended and relatively faint plateau in the \(2-10\,\)keV flux. The \(3-200\,\)keV X-ray spectra shown in Figure [\[fig:grs1915_spec\]](#fig:grs1915_spec){reference-type="ref" reference="fig:grs1915_spec"} also exhibit remarkable stability between observations. The individual spectra are dominated by a hard non-thermal component and strongly suggest class \(\chi\) (in state C) behaviour, which is confirmed by the location of the observations in the HID of Figure [\[HID_GRS1915\]](#HID_GRS1915){reference-type="ref" reference="HID_GRS1915"}, according to the classification of. Figure [\[evol_GRS1915\]](#evol_GRS1915){reference-type="ref" reference="evol_GRS1915"} shows the contextual X-ray and 15 GHz radio light curves of GRS 1915+105 during a two-month period that brackets the H.E.S.S. observation epochs. It is evident from the figure that H.E.S.S. observed the target during and extended radio-loud plateau (\(\sim\)`<!-- -->`{=html}80 mJy; for historical flux comparison, a three-year monitoring campaign is presented in ). The plateau ended approximately ten days later with a combined radio and X-ray flaring episode. The assertion of radio-loud behaviour at the H.E.S.S. observation epochs is supported by the quasi-periodic oscillation (QPO) analysis presented in Figure [\[fig:grs_powspecs\]](#fig:grs_powspecs){reference-type="ref" reference="fig:grs_powspecs"}. For a detailed discussion see Appendix [\[app:GRS1915state\]](#app:GRS1915state){reference-type="ref" reference="app:GRS1915state"}. In summary, the combined spectral and temporal analyses indicate a robust association of the contemporaneous H.E.S.S. observation with the radio-loud \(\chi\) state, and the presence of steady, mildly relativistic jets at the time of observation may be confidently inferred. The contemporaneous H.E.S.S. observations did not yield a significant VHE \(\gamma\)-ray detection. The significances corresponding to the total H.E.S.S. exposure are computed using Eq. 17 from and are listed in Table [\[tab:mqs:hess_sigmas\]](#tab:mqs:hess_sigmas){reference-type="ref" reference="tab:mqs:hess_sigmas"}. Figure [\[fig:grs_lightcurves\]](#fig:grs_lightcurves){reference-type="ref" reference="fig:grs_lightcurves"} plots runwise 99% confidence level upper limits to the integral VHE \(\gamma\)-ray flux above the instrumental threshold energy and illustrates the overlap between the *RXTE* and H.E.S.S. observations. Integral flux upper limits, which correspond to the overall H.E.S.S. exposure, are listed in Table [\[tab:mqs:hess_uls\]](#tab:mqs:hess_uls){reference-type="ref" reference="tab:mqs:hess_uls"}. ## Cir X-1 {#subsec:cir_res} The ASM light curve shown in Figure [\[fig:cir_lightcurves\]](#fig:cir_lightcurves){reference-type="ref" reference="fig:cir_lightcurves"} reveals that the H.E.S.S. observation epochs occurred during an extended \(\sim4\) day dip in the \(2-10\,\)keV X-ray flux. Additionally, it should be noted that the observations reported here were obtained during an extremely faint episode in the secular X-ray flux evolution of Cir X-1, which is also evident from the HID presented in Figure [\[HID_CirX1\]](#HID_CirX1){reference-type="ref" reference="HID_CirX1"}. As a consequence, the measured X-ray fluxes are significantly lower than most others reported for this source. As illustrated in Figure [\[fig:cir_lightcurves\]](#fig:cir_lightcurves){reference-type="ref" reference="fig:cir_lightcurves"}, the individual PCA light curves obtained during the first two pointings are characterised by a relatively low count rate, which remains approximately constant throughout each observation. In marked contrast, the third observation exhibits clear variability with count rates doubling on timescales of \(\sim50\)s. A detailed analysis of the obtained spectra (see Appendix [\[app:CirState\]](#app:CirState){reference-type="ref" reference="app:CirState"}) reveals that the observed flux variability is accompanied by marked variations in spectral shape. These can be interpreted as hinting towards a strong mass transfer during the periastron passage and subsequent dramatic evolution of the local radiative environment. H.E.S.S. observations obtained contemporaneously with the *RXTE* pointings yield a non-detection that is evident from the significances listed in Table [\[tab:mqs:hess_sigmas\]](#tab:mqs:hess_sigmas){reference-type="ref" reference="tab:mqs:hess_sigmas"}. Figure [\[fig:cir_lightcurves\]](#fig:cir_lightcurves){reference-type="ref" reference="fig:cir_lightcurves"} plots runwise 99% confidence level upper limits to the integral VHE \(\gamma\)-ray flux above the instrumental threshold energy and illustrates complete overlap between the *RXTE* and H.E.S.S. observations. Integral flux upper limits, which correspond to the overall H.E.S.S. exposure, are listed in Table [\[tab:mqs:hess_uls\]](#tab:mqs:hess_uls){reference-type="ref" reference="tab:mqs:hess_uls"}. ## V4641 Sgr {#subsec:v46_res} Figure [\[fig:v4s_lightcurves\]](#fig:v4s_lightcurves){reference-type="ref" reference="fig:v4s_lightcurves"} shows *RXTE* PCA light curves derived from three pointed observations. The individual light curves indicate various degrees of X-ray variability with the clearest evidence for flaring visible as a sharp \(\sim5\)-fold count rate fluctuation during the first observation. In marked contrast, the second observation is uniformly faint with the \(\chi^{2}\) probability of constant count rate \(P_{\rm const}=0.97\) and, hence, this observation is consistent with a period of steady, low-level emission. Subsequently, the third observation reveals a reemergence of mild variability (\(P_{\rm const}=0.07\)) with \(\sim2\)-fold count rate fluctuations occurring on timescales of \(\sim500\)s. Radio data shown on Figure [\[fig:V4641hardness_radio\]](#fig:V4641hardness_radio){reference-type="ref" reference="fig:V4641hardness_radio"} right were obtained using the VLA and Australia Telescope Compact Array (ATCA) between MJD 53190 and MJD 53208. They indicate rapid variability with peak flux densities of \(\sim30\) mJy observed on MJD 53193. An optically thin radio spectrum (\(S_{\nu}\propto \nu^{-0.7}\)) observed on MJD 53191 was interpreted by as the signature of a decaying radio flare. Radio observations were triggered by an optical alert from VSNET (MJD 53190) in combination with a *RXTE* PCA measurement during a Galactic bulge scan (MJD 53189) that revealed a 2-10 keV X-ray flux equivalent to 8.2 mCrab. For comparison, the August 2003 flare of V4641 Sgr reached 66 mCrab, while quiescent fluxes are typically \< 0.5 mCrab. As shown in Figure [\[fig:V4641hardness_radio\]](#fig:V4641hardness_radio){reference-type="ref" reference="fig:V4641hardness_radio"} right, the dedicated *RXTE* PCA observation and H.E.S.S. observations took place between two radio flares, which is consistent with the X-ray variability evolution illustrated in Figure [\[fig:v4s_lightcurves\]](#fig:v4s_lightcurves){reference-type="ref" reference="fig:v4s_lightcurves"}. While V4641 Sgr is evidently the most X-ray-faint binary in the studied sample, it simultaneously exhibits the hardest spectrum, as shown by the hardness values in Figure [\[fig:V4641hardness_radio\]](#fig:V4641hardness_radio){reference-type="ref" reference="fig:V4641hardness_radio"} (left-hand panel). Furthermore, the evolution of the hardness is consistent with contemporaneous observations of rapid flux evolution in the radio band [^3] (Figure [\[fig:V4641hardness_radio\]](#fig:V4641hardness_radio){reference-type="ref" reference="fig:V4641hardness_radio"}). To help place the H.E.S.S. and *RXTE* observations in a historical context, the HID for V4641 Sgr in Figure [\[HID_V4641Sgr\]](#HID_V4641Sgr){reference-type="ref" reference="HID_V4641Sgr"} displays the entire archival *RXTE* PCA data set for this target, and compares the X-ray characteristics corresponding to the H.E.S.S. observation periods (different symbols are used to indicate observations obtained on each day in the range MJD 53193-5) with three flaring episodes observed with *RXTE*. On 15 September 1999 (orange markers in Figure [\[HID_V4641Sgr\]](#HID_V4641Sgr){reference-type="ref" reference="HID_V4641Sgr"}), a 1500s *RXTE* observation revealed a source flux evolution that is characterised by rapid, large-amplitude variability before reverting to a soft, low intensity state after \(\sim1000\)s. An optical flare that preceded the *RXTE* observations likely corresponds with the onset of the short 10-hour outburst, which associated with a low \(\dot{M}\) accretion event. Historically, flaring episodes exhibited by V4641 Sgr are often short, unpredictable, and relatively faint, which implies that many may go unnoticed. Data corresponding to two longer outbursts, spanning the periods 24-26 May 2002 and 5-7 August 2003, are also illustrated in Figure [\[HID_V4641Sgr\]](#HID_V4641Sgr){reference-type="ref" reference="HID_V4641Sgr"}. In the coordinates of the HID, both episodes are topologically similar to the 1999 outburst, but shifted towards fainter harder regions. Evidently, the X-ray fluxes that correspond with H.E.S.S. observation epochs indicated in Figure [\[HID_V4641Sgr\]](#HID_V4641Sgr){reference-type="ref" reference="HID_V4641Sgr"} are substantially fainter than any of these historically remarkable outbursts. In summary, in view of the various multiwavelength data, it seems likely that V4641 Sgr underwent a period of mild activity that spanned the H.E.S.S. observation epochs. The contemporaneous H.E.S.S. data are consistent with non-detection with the corresponding \(\gamma\)-ray significances listed in Table [\[tab:mqs:hess_sigmas\]](#tab:mqs:hess_sigmas){reference-type="ref" reference="tab:mqs:hess_sigmas"}. Technical issues prevented \(\gamma\)-ray data corresponding to the first *RXTE* observation from being obtained. Simultaneous \(\gamma\)-ray observations were obtained corresponding to the second *RXTE* exposure, which showed no indications of X-ray variability. Although the source began to show increased X-ray activity during the third *RXTE* observation, the degree of overlap with the corresponding H.E.S.S. observations was minimal. At radio, optical, and X-ray energies, V4641 Sgr exhibits rapid variability on timescales \(\sim10\) minutes or less. Optimistically, the compelling evidence for mild broadband flaring admits the possibility that the H.E.S.S. observations monitor a transient outburst event. Integral flux upper limits above the instrumental threshold energy, which correspond to the overall H.E.S.S. exposure at the position of V4641 Sgr, are listed in Table [\[tab:mqs:hess_uls\]](#tab:mqs:hess_uls){reference-type="ref" reference="tab:mqs:hess_uls"}. # Discussion {#sec:mqs:context} The principal aim of this investigation was to obtain contemporaneous X-ray and VHE \(\gamma\)-ray observations of three known superluminal microquasars during major flaring events. However, the results presented in \(\S\)[3](#sec:mqs:results){reference-type="ref" reference="sec:mqs:results"} indicate that the interpretation of the VHE \(\gamma\)-ray non-detections cannot proceed under the assumption of energetic flaring or bulk superluminal ejections at the time of observation. Nonetheless, upper limits to the VHE \(\gamma\)-ray flux were derived and an analysis of the contemporaneous *RXTE* observations has helped to reveal the X-ray behaviour corresponding to the H.E.S.S. observation epochs. These data facilitate the straightforwards derivation of constraints on the \(\gamma\)-ray luminosity of the target binary systems. In Table [\[tab:mqs:luminosties\]](#tab:mqs:luminosties){reference-type="ref" reference="tab:mqs:luminosties"} the calculated flux upper limits were used to infer the maximum \(\gamma\)-ray luminosities above the target-specific, instrumental threshold energy for each target binary system by assuming the maximum source distance estimate presented in Table [\[tab:target_properties\]](#tab:target_properties){reference-type="ref" reference="tab:target_properties"}. Analysis of the contemporaneous X-ray and radio observations conclusively places GRS 1915+105 in a radio-loud plateau state at the time of observation. In contrast with the superluminal flaring episodes, this state is characterised by the production of continuous, mildly relativistic radio jets with an estimated power of \(\sim3\times10^{38}\) erg s\(^{-1}\) (, assuming a distance of 11 kpc). Theoretically, it seems unlikely that bright VHE \(\gamma\)-ray emission would be expected from the compact self-absorbed jets, which are typical of the plateau state of GRS 1915+105. For example, a leptonic emission model developed by to simulate the broadband emission of microquasar jets in the low-hard state predicts VHE \(\gamma\)-ray luminosities \(\lesssim10^{33}\) erg s\(^{-1}\) that are consistent with the H.E.S.S. non-detection. Notwithstanding the plausibility of VHE \(\gamma\)-ray emission in the plateau state, a comparison of the estimated jet power with the maximum \(\gamma\)-ray luminosity listed in Table [\[tab:mqs:luminosties\]](#tab:mqs:luminosties){reference-type="ref" reference="tab:mqs:luminosties"} reveals that the jet power conversion efficiency is constrained to be \(\lesssim0.008\%\) for \(\gamma\)-ray production above 562 GeV. For comparison, corresponding efficiency estimates for \(\gamma\)-ray production were derived for the steady, compact jets of other microquasars that were observed in appropriate states. The published MAGIC upper limit on the VHE \(\gamma\)-ray luminosity of Cygnus X-3 during its hard state implies a somewhat larger maximum conversion efficiency of 0.07% and a similar value is obtained from MAGIC upper limits on the steady VHE emission from Cygnus X-1. These efficiencies are inferred from the directly observed jet power, and should be distinguished from the higher jet powers that were indirectly derived from the observation of radio-emitting bubbles inflated by microquasar jets (see e.g. [@Gallo2005] for Cyg X-1, and [@Pakull2010], [@Soria2010] for S26 in NGC 7793). We presented an analysis of the entire H.E.S.S. data set for GRS 1915+105 and we derived an upper limit to the \(\gamma\)-ray flux above 0.41 TeV of \(6.1\times10^{-13}\) cm\(^{-2}\)s\(^{-1}\), corresponding to a detector livetime of 24.1 hours. The somewhat higher upper limits presented in \(\S\)[3.1](#subsec:grs_res){reference-type="ref" reference="subsec:grs_res"} utilise a more limited data set and are therefore consistent with the previously published value. None of the H.E.S.S. observations of GRS 1915+105 coincide with bright flaring episodes at longer wavelengths. Observations of Cir X-1 were obtained during an extended dip in the X-ray flux, at phase intervals close to the periastron passage of the binary components. Spectral analysis of the *RXTE* data showed some evidence for a recent increase in mass transfer, producing strong signatures of X-ray absorption. It was hoped that H.E.S.S. observations would coincide with one of the quasi-regular radio flares, which often accompany periastron passage in Cir X-1. The ephemeris of predicts the onset of a radio flare \(\sim19-20\)h before the first *RXTE* observation. Unfortunately, despite the undoubted occurrence of quasi-periodic radio flares from Cir X-1 near periastron, a robust correlation between the observed X-ray and radio behaviour is yet to be identified. Although rapid brightening of the X-ray continuum might indicate accompanying radio flares, evidence for a definitive association is far from clear. Recent radio observations of Cir X-1 focus primarily on the ultrarelativistic ejection events that manifest as \(\gtrsim3\) day episodes of flaring on timescales of a few hours. In principle, the lack of contemporaneous radio data admits the possibility of such persistent outbursts at the time of observation. By analogy with canonical black hole binaries, it is possible that the inferred variation in the mass accretion rate between the first and second *RXTE* observations also implies an evolution of the jet properties, but this is far from clear in such an unusual system. Moreover, report compelling evidence that prior to 2006, Cir X-1 underwent a \(\sim6\) year episode of unusual radio quiescence, suggesting that jet formation was somewhat suppressed during the epochs of H.E.S.S. observation. Accordingly, without strictly simultaneous radio data indicating otherwise, the most likely scenario is that no outflows were present. In this context the absence of a detectable \(\gamma\)-ray signal is not surprising. As a confirmed high-mass black hole binary, V4641 Sgr is the studied target that most closely resembles the Cygnus X-1 and Cygnus X-3 systems. Moreover, the H.E.S.S. observations were obtained during a period of sporadic broadband flaring, and comparing these observations with the results of, VHE \(\gamma\)-ray emission might have been expected. The detection of Cyg X-1 using the MAGIC telescopes appeared to coincide with the rising part of a strong X-ray flare. In contrast, radio spectra obtained close to the H.E.S.S. observational epochs are indicative of the decay following a flaring episode. Assuming that the \(\gamma\)-ray emission mechanisms operating in Cyg X-1 also occur in V4641 Sgr, the absence of a significant H.E.S.S. detection might be viewed as evidence that production of GeV and TeV photons is a highly transient process. This would further suggest that \(\gamma\)-ray emission originates in a spatially compact region that is at most a few light hours in size. Absorption of \(\gamma\)-rays by pair production is expected to be negligible in GRS 1915+105, since the donor star is too cool and faint to produce a strong ultraviolet photon field. The same is true of Cir X-1 if the conventional assumption of a low-mass companion is adopted. For completeness, Figure [\[fig:mqs:abs\]](#fig:mqs:abs){reference-type="ref" reference="fig:mqs:abs"} plots the level of \(\gamma\)-ray absorption predicted by a numerical implementation of the model presented by, assuming that the companion star in Cir X-1 is in fact a mid-B supergiant as proposed by. The separate curves are representative of the three orbital phase intervals corresponding to the H.E.S.S. observation epochs, and were derived using the system parameters derived by in conjunction with typical values for the temperature (\(T_{\rm eff}\approx20000\) K) and radius (\(R\approx9\) R\(_{\odot}\)) of a mid-B supergiant. It is evident that some non-negligible absorption is expected, particularly during the first observation interval. Nonetheless, it seems unlikely that the expected levels of attenuation (\(\lesssim20\%\)) would suppress an otherwise detectable \(\gamma\)-ray flux sufficiently to yield the low significances listed in Table [\[tab:mqs:hess_sigmas\]](#tab:mqs:hess_sigmas){reference-type="ref" reference="tab:mqs:hess_sigmas"}. The situation with regard to \(\gamma\)-ray absorption is clearer in the case of V4641 Sgr, since the companion has been spectroscopically identified as a late B-or early A-type star. Using the system parameters derived by and assuming a circularised orbit, the model presented by was used to predict the expected levels of \(\gamma\)-ray absorption as a function of orbital phase. As illustrated in Figure [\[fig:mqs:abs\]](#fig:mqs:abs){reference-type="ref" reference="fig:mqs:abs"} (bottom panel), absorption might have an important effect during the first H.E.S.S. observation interval, although as with Cir X-1 the predicted levels of absorption (\(\lesssim25\%\)) would not attenuate a bright \(\gamma\)-ray signal so far below the detection threshold. During the second H.E.S.S. observation interval, when X-ray data show marginal indications of source activity, the predicted absorption due to pair production on the stellar radiation field is negligible. We note however that, as in the case of Cir X-1, the relative inclination of the jets from V4641 Sgr with respect to the accretion disk may be low and, therefore, further absorption of \(\sim100\) GeV--TeV \(\gamma\)-ray photons could occur via interaction with the disk thermal photon field (see e.g. [@1992A&A...264..127C]). It should also be noted that all the confirmed VHE \(\gamma\)-ray binaries lie at distances of \(2-4\) kpc. In contrast, the targets reported here have maximum distances in the range \(7-11\) kpc, resulting in flux dilution factors that are greater by one order of mangnitude. Obviously, this has strong implications for the detectability of any emitted \(\gamma\)-ray signal. # Conclusions {#sec:conclusion} Contemporaneous VHE \(\gamma\)-ray and X-ray observations of GRS 1915+105, Cir X-1, and V4641 Sgr were obtained using H.E.S.S. and *RXTE*. Analysis of the resultant H.E.S.S. data did not yield a significant detection for any of the target microquasars. However, X-ray binaries are dynamic systems and as such are likely to exhibit evolution of their radiative properties, both as a function of orbital phase and also in response to non-deterministic properties. It follows that the non-detections presented in this work do not indicate that the target binary systems do not emit detectable VHE \(\gamma\)-ray emission at phases other than those corresponding to the H.E.S.S. observations. GRS 1915+105 appears to have been observed during an extended plateau state, the archival multiwavelength data suggesting the presence of continuous, mildly relativistic radio jets at the time of observation. The *RXTE* observations of Cir X-1 yield data that are consistent with strongly varying obscuration of the X-ray source shortly after periastron passage, but these data are not indicative of bright flaring during the H.E.S.S. observation epochs. Conversely, V4641 Sgr appears to have been observed during an episode of mild, transient flaring, although rapid source variability, combined with the limited duration of the strictly simultaneous H.E.S.S. and *RXTE* exposure, complicates interpretation. Microquasars continue to be classified as targets of opportunity for IACTs, requiring a rapid response to any external trigger to maximise the likelihood of obtaining a significant detection. These conditions are realised with the commissioning of the H.E.S.S. 28 m telescope, which aims to lower the energy threshold from 100 GeV to about 30 GeV (;; ) while simultaneously enabling very rapid follow-up observations. To exploit these new opportunities and an increasing understanding of the behaviour of microquasars, the triggering strategies for TeV follow-up observations have evolved significantly in recent years. In the future, alternative observational strategies, including continuous monitoring of candidate microquasars in the VHE \(\gamma\)-ray band, may become possible using dedicated sub-arrays of the forthcoming Cherenkov Telescope Array. Irrespective of the non-detections presented herein, the tantalising observations of Cygnus X-3 at GeV energies and Cygnus X-1 by the MAGIC telescope ensures that the motivations for observing microquasars using IACTs remain compelling. Indeed, by further constraining the \(\gamma\)-ray emission properties of microquasars, subsequent observations will inevitably yield an enhanced understanding of astrophysical jet production on all physical scales. More optimistically, the detection of additional \(\gamma\)-ray-bright microquasars would greatly facilitate a comprehensive characterisation of the particle acceleration and radiative emission mechanisms that operate in such systems. [^1]: <http://tevcat.uchicago.edu/> [^2]: http://heasarc.nasa.gov/docs/xte/xhp_proc_analysis.html [^3]: <http://www.ph.unimelb.edu.au/~rsault/astro/v4641/>

{'timestamp': '2016-07-19T02:10:20', 'yymm': '1607', 'arxiv_id': '1607.04613', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04613'}

# Introduction Human Activity Recognition (HAR) is the understanding of human behaviour from data captured by pervasive sensors, such as cameras or wearable devices. It is a powerful tool in medical application areas, where consistent and continuous patient monitoring can be insightful. Wearable devices provide an unobtrusive platform for such monitoring, and due to their increasing market penetration, feel intrinsic to the user. This daily integration into a user's life is crucial for increasing the understanding of overall human health and wellbeing. This is referred to as the "quantified self\" movement. Wearables, such as actigraph accelerometers, generate a continuous time series of a person's daily physical exertion and rest. This ubiquitous monitoring presents substantial amounts of data, which can (*i*)  provide new insights by enriching the feature set in health studies, and (*ii*)  enhance the personalisation and effectiveness of health, wellness, and fitness applications. By decomposing an accelerometer's time series into distinctive activity modes or actions, a comprehensive understanding of an individual's daily physical activity can be inferred. The advantages of longitudinal data are however complemented by the potential of noise in data collection from an uncontrolled environment. Therefore, the data sensitivity calls for robust automated evaluation procedures. In this paper, we present a robust automated human activity recognition (RAHAR) algorithm. We test our algorithm in the application area of sleep science by providing a novel framework for evaluating sleep quality and examining the correlation between the aforementioned and an individual's physical activity. Even though we evaluate the performance of the proposed HAR algorithm on sleep analysis, RAHAR can be employed in other research areas such as obesity, diabetes, and cardiac diseases. # Related Work Human activity recognition (HAR) has been an active research area in computer vision and machine learning for many years. A variety of approaches have been investigated to accomplish HAR ranging from analysis of still images and videos to motion capture and inertial sensor data. Video has been the most widely studied data source in HAR literature. Hence, there exists a wealth of papers in this particular domain. The most recent literature on HAR from videos include trajectory-based descriptors , spatio-temporal feature representations , feature encoding , and deep learning . Reviewing the extensive list of video-based HAR studies, however, goes beyond the scope of this study and we refer the reader to  for a collection of more comprehensive surveys on the topic. Unlike HAR from video, existing approaches for HAR from still images are somewhat limited, and range from histogram-based representations  and color descriptors  to pose-, appearance-and parts-based representations . Guo and Lai recently provided a comprehensive survey of the studies on still image-based HAR in . Several techniques have been proposed, on the other hand, for HAR from 3D data, encompassing representations based on bag-of-words , eigen-joints , sequence of most informative joints , linear dynamical systems , actionlets , Lie algebra embedding , covariance descriptors , hidden Markov models , subspace view-invariant metrics  and occupancy patterns . Aggarwal and Xia presented a recent survey summarizing state-of-the-art techniques in HAR from 3D data . Unlike vision-based HAR systems, sensor-based HAR technologies commonly deal with time series of state changes and/or various parameter values collected from a wide range of sensors such as contact sensors, accelerometers, audio and motion detectors, etc. Chen et al.  and Bulling et al.  present comprehensive reviews of sensor-based activity recognition literature. The most recent work in this domain includes knowledge-based inference , ensemble methods , data-driven approaches , and ontology-based techniques . All of the aforementioned studies investigate recognition/classification of fully observed action or activity, e.g., jumping, walking, running, drinking, etc. (i.e., activities of daily living), using well-curated datasets. However, thanks to the "quantified self\" movement, myriad of consumer-grade wearable devices have become available for individuals who have started monitoring their physical activity on a continuous basis, generating tremendous amount of data. Therefore, there is an urgent need for automatic analysis of data coming from fitness trackers to assess the physical activity levels and patterns of individuals for the ultimate goal of quantifying their overall wellbeing. This task requires understanding of longitudinal, noisy physical activity data at a rather higher (coarser) level than specific action/activity recognition level. Main challenges as well as opportunities of HAR from personalized data and lifelogs have been discussed in several dimensions in . There has been a number of initiatives to overcome the challenge of collecting annotated personalized data to further research on HAR from continuous measurement of real-world physical activities . Even though such systems exhibit a crucial attempt in furthering research in mining personalized data, they have limited practical importance as they rely on manual annotation of the acquired data. There has also been recent attempts to automatically recognize human activities from continuous personalized data . However, most of these studies are designed to recognize only a predefined set of activities, and hence, not comprehensive and robust enough to quantify the physical activity levels for the overall assessment of individuals' wellbeing. # Background Sleep pattern evaluation is a paragon of cumbersome testing and requires extensive manual evaluation and interpretation by clinical experts. Unhealthy sleep habits can impede physical, mental and emotional wellbeing, and lead to exacerbated health consequences . Since patient referral to sleep specialists is often based on self-reported abnormalities, exacerbation often precedes diagnosis. Clinical diagnosis of complex sleep disorders involves a variety of tests, including an overnight lab stay with oxygen and brain wave monitoring (polysomnography and electroencephalogram, respectively), and a daily sleep history log with a subjective questionnaire. The daily sleep logs and questionnaires are often found to be unreliable and inconsistent with actual observed activity. This is especially true in adolescents . The overnight stay allows specialists to manually monitor the patient's sleep period. This requires the active involvement of a clinical sleep specialist. Furthermore, the monitoring is only for one night and in a clinical setting, rather than the patient's own home. Using wearable devices provides both a context-aware and longitudinal monitoring. The inconvenience and inaccuracy of daily logs, coupled with the invasiveness of an overnight lab stay, substantiate the need and adoption of wearable devices for first pass diagnostic screening. More generally, using our HAR approach with a wearable device empowers users to self-monitor their sleep patterns, and reform their activity habits for optimised sleep and an improved quality of life. # Preliminaries In this section we present a description of the dataset and the context-aware definitions used for our application area. ## Data Data was collected as part of a research study to examine the impact of sleep on health and performance in adolescents by Weil COrnell Medical COllege-Qatar. Two international high schools were selected for cohort development. Student volunteers were provided with an actigraph accelerometer, ActiGraph GT3X+[^1], to wear on their non-dominant wrist, continuously throughout the study (i.e. even when sleeping). Deidentified data collected in the study were used in the current analysis. The ActiGraph GT3X+ is a clinical-grade wearable device that has been previously validated against clinical polysomnography . The device samples the user's sleep-wake activity at 30-100 Hertz. Currently sleep experts use this device in conjunction with the accompanying software, ActiLife[^2], to evaluate an individual's sleep period. We evaluate our results side-by-side with ActiLife's results. ## Definitions To apply our methodology to the area of sleep science, it is important to note the definitions mentioned in this section. In traditional sleep study literature, a sleep period is bounded between the sleep-onset-time and sleep-awakening-time . Experts characterise the sleep-onset-time as the first minute after a self-reported bedtime, that is followed by 15 minutes of continuous sleep . We propose a modified definition, that allows for automatic evaluation and deems sleep diaries unnecessary. As a result, we can infer the "bedtime\" of an individual in reverse, based on their sedentary activity before the onset of sleep. Epoch records that contain no triaxial movement, 0 steps taken, and an inclinometer output of not lying down, are candidate sleep records, and are further tested for whether they are a component of the sleep period. We define sleep-onset-time as the first candidate epoch record in a series of 15 continuous candidate sleep minutes. Likewise, the sleep-awakening-time is defined as the last epoch record in a series of 15 continuous candidate sleep minutes, that is followed by 30 continuous non-candidate sleep minutes, (i.e. 30 minutes of active awake time). The sleep period duration can be computed as the time passed between sleep onset and sleep awakening. Within the sleep period, there are periods of unrest or wakefulness. For example, when a user re-adjusts positions, or uses the bathroom. If the duration of movement exceeds 5 consecutive minutes of activity, it is marked as a time of "wakefulness.\" The total sum of all moments of wakefulness is referred to as wake-after-sleep-onset, also known as WASO. Immediately preceding the start of the sleep onset, is the time-in-bed, which quantifies the sedentary time an individual spends before they have fallen asleep. This sedentary time can be observed in the actigraph accelerometer data. The time that the preceding sedentary activity begins until the time of the sleep onset is called the sleep latency. From the aforementioned values, total sleep time and an overall sleep efficiency score can be deduced. Total sleep time covers the defined sleep period, less the wake after sleep onset time and less the latency. Lastly, sleep efficiency is the ratio of total sleep time to total minutes in bed. All of the above definitions are summarised in Table [\[tab:sleep_eqs\]](#tab:sleep_eqs){reference-type="ref" reference="tab:sleep_eqs"}, and visualised in Fig. [\[fig:sleep_example\]](#fig:sleep_example){reference-type="ref" reference="fig:sleep_example"}. In this study, we use sleep efficiency as the metric to measure sleep quality  among other metrics such as latency, wake after sleep onset, awakening index, total sleep time, etc. . # Methodology Our methodology for RAHAR is shown algorithmically in Fig. [\[algo:rahar\]](#algo:rahar){reference-type="ref" reference="algo:rahar"}. We elaborate on the details of our algorithm in the sequel. ## Pre-Processing The accelerometer of choice, Actigraph GT3X+, sampled each person's activity at 30-100 Hertz. The stored data included the triaxial accelerometer coordinates as well as a computed epoch step count based on the vertical axis, and post-processed inclinometer orientation. This raw data was downloaded and aggregated to a minute-by-minute granularity. An epoch of one minute was selected in order to optimise the interpretability of the physical activity , as well as for implementing the state-of-the-art cut point methodology . In other contexts, a different granularity may be sufficient. ## Automated Annotation and Segmentation Due to the context of sleep disorders, sleep periods needed to be annotated within the raw ActiGraph output. Candidate sleep records, epochs with no triaxial movement, 0 steps taken, and an inclinometer output of not lying down, were identified in the time series and tested to find the sleep onset time, and sleep awakening time. The details of this terminology is elaborated in the preliminaries section. All test instances that fell within these two boundary times, were annotated as "Sleep,\" and constituted the sleep period. Whilst analysing the data, we found that several participants had multiple sleep periods in a day, implying that they took daily naps or followed a polyphasic, or biphasic, sleep pattern. Upon closer analysis of the length and time of the sleep period, no discernible patterns were visible. Thus we opted to segment the time series by the end of a sleep period rather than the traditional approach of segmenting by day. Each sleep period was linked to its preceding activity, extending until the previous sleep period. We refer to these segments as sleep-wake segments. The result of this decision is that the activity immediately before each sleep period is used for the correlation analysis for its subsequent sleep period, rather than the total for that day. [\[tab:slp_eff_res\]]{#tab:slp_eff_res label="tab:slp_eff_res"} [\[tab:sens_spec\]]{#tab:sens_spec label="tab:sens_spec"} ::: RAHAR is fundamentally a feature extraction algorithm for HAR in the context of quantifying daily physical activity levels of individuals. We therefore test the quality of activity recognition by RAHAR as compared to an expert-based HAR using a tool on continuous physical activity data from a wearable sensor. Since there is *no* ground truth on human activity in this context, our objective is to evaluate which HAR approach leads to better quality models for sleep research, i.e., models for predicting sleep quality, specifically, sleep efficiency. We selected four models for evaluating the performance of RAHAR against the performance of an expert-based HAR using a tool on the described actigraphy dataset: logistic regression, support vector machines with radial basis function kernel, random forest, and adaboost. - Logistic Regression (LogR): We chose this model because it is an easily interpretable binary classifier. It is also relatively robust to noise, which as explained earlier is a complication on data collected in an uncontrolled environment.[^4] - Support Vector Machine (SVM): This model was selected because it, also, is a binary classifier. We chose a radial basis function kernel, and so it differs from logistic regression in that it does not linearly divide the data. - Random Forest (RF): This model was tested as an alternative because of its easy straightforward interpretation, which is particularly relevant in the healthcare or consumer domains. It also is not restricted to linearly dividing the data. - Adaboost (Ada): Lastly, Adaboost was tested because it is less prone to overfitting than random forest. For comparison purposes, we use the results from a sleep specialist using Actigraph's ActiLife software as a baseline. The sleep specialist segmentation of the ActiLife results uses the preceding day's activity for each sleep period, and aggregates the activity to an epoch of an hour. ActiLife requires the sleep specialist to manually adjust the sleep period boundaries, and then automatically computes the efficiency and other sleep metrics. Figs. [\[fig:roc_RAHAR\]](#fig:roc_RAHAR){reference-type="ref" reference="fig:roc_RAHAR"} and [\[fig:roc_SEAL\]](#fig:roc_SEAL){reference-type="ref" reference="fig:roc_SEAL"} show the ROC curves for RAHAR and the sleep expert using ActiLife software (denoted as "SE+AL\"), respectively, while Table [\[tab:slp_eff_res\]](#tab:slp_eff_res){reference-type="ref" reference="tab:slp_eff_res"} summarises the results for both RAHAR and SE+AL. One of the most important performance measures for HAR is the area under ROC (AU-ROC). Based on AU-ROC scores, both RAHAR and SE+AL performed best with random forest model. Furthermore, SE+AL achieved an average AU-ROC of 0.7246 whereas RAHAR achieved 0.8355, a 15% improvement of AU-ROC score on average by our algorithm as opposed to the sleep expert using ActiLife. With an AU-ROC score of 0.5884 for SE+AL approach, the logistic regression model was, however, unable to stratify the dataset, and so predicted all cases to be in a single class. We considered this to be a failure of the logistic regression model for this problem, and thus, did not include its results in our discussion whenever it was appropriate to do so. For this reason, the misleading results have also been removed from Table [\[tab:slp_eff_res\]](#tab:slp_eff_res){reference-type="ref" reference="tab:slp_eff_res"}. Another important performance measure for HAR is the F1 score, which is computed as the harmonic mean of precision and recall. According to Table [\[tab:slp_eff_res\]](#tab:slp_eff_res){reference-type="ref" reference="tab:slp_eff_res"}, RAHAR performed better than SE+AL in terms of precision and recall for all models, and hence, yielded significantly higher F1 scores. Specifically, F1 score for RAHAR, on average, was 0.7154 whereas it was 0.5393 for SE+AL (excluding logistic regression in both cases), yielding a solid margin of about 0.18 points (i.e., more than 30% improvement). On the other hand, the accuracy scores, on average, were 0.7353 for RAHAR and 0.7266 for SE+AL (again excluding logistic regression), and exhibited a relatively less significant difference still in favor of RAHAR. # Discussion of Results in Medical Context In this section we discuss the results of the best performing model and its broader impact to the area of sleep science. As seen in Fig. [\[fig:roc_curve\]](#fig:roc_curve){reference-type="ref" reference="fig:roc_curve"} random forest and logistic regression were the two best performing models with the RAHAR algorithm. Based on the desired threshold value of true positive rate, TPR, (i.e., sensitivity), either model could be preferred to minimize false positive rate, FPR, (i.e., 1-specificity), which is equivalent to maximizing specificity. Random forest was also the best performing model for the SE+AL approach as mentioned earlier. If we compare the ROC as well as the sensitivity-specificity plots of the best model of each approach (i.e., random forest), we see that RAHAR outperforms SE+AL almost always as illustrated in Fig. [\[fig:rf_comp\]](#fig:rf_comp){reference-type="ref" reference="fig:rf_comp"}. Table [\[tab:sens_spec\]](#tab:sens_spec){reference-type="ref" reference="tab:sens_spec"}, on the other hand, summarises sensitivity and specificity scores for RAHAR and SE+AL. Average sensitivity score for SE+AL and RAHAR across all models except logistic regression were 0.4946 and 0.5965, respectively. In other words, average sensitivity score for RAHAR is 20% higher than that of SE+AL. As for specificity, RAHAR with an average score of 0.9111 outperforms SE+AL with an average score of 0.8505, which corresponds to a 7% improvement. As we seek to determine in our study whether a person had a "good quality sleep\" based on his physical activity levels during awake period prior to sleep, a false positive occurs when the model predicts "good quality sleep\" when the person actually had a "poor quality sleep.\" Therefore, the number of false positives needs to be kept at a minimum for a desired number of true positives. In other words, a high specificity score is sought after while keeping the sensitivity score at the desired level. As can be seen from Fig. [\[fig:rf_ss_comp\]](#fig:rf_ss_comp){reference-type="ref" reference="fig:rf_ss_comp"} with this perspective in mind, for a large range of sensitivity scores, RAHAR achieved higher specificity scores almost all the time than SE+AL did. For example, RAHAR achieved a sensitivity score of 0.9 with a specificity score of 0.8 whereas SE+AL remained at a specificity score of 0.6 for the same sensitivity threshold. In summary, RAHAR outperforms state-of-the-art procedure in sleep research in many aspects. However, its application is not limited to sleep and it can be used for understanding and treatment of other health issues such as obesity, diabetes, or cardiac diseases. Moreover, RAHAR allows for fully automated interpretation without the necessity of manual input or subjective self-reporting. Given the current interest in deep learning, a natural question that may arise is why an approach based on feature extraction and model building has been used instead of using deep learning models directly on the raw sensor data for HAR. In medical community, the explainability of a model is of utmost important as the medical professionals are interested in learning cause-and-effect relationships and using this knowledge to support their decision making processes. In this particular case, for example, sleep experts are interested in understanding how and when certain physical activity levels effect sleep in order to make decisions to improve sleep quality of individuals accordingly. However, it is an interesting idea to explore deep learning to see what is the best model from a model accuracy perspective to understand the limits of the value of continuous monitoring of individuals' physical activity, not only from a medical perspective in particular but also from a "quantified self\" perspective in general. # Conclusion In this paper, we presented a robust automated human activity recognition (RAHAR) algorithm for multi-modal phenomena, and evaluated its performance in the application area of sleep science. We tested the results of RAHAR against the results of a sleep expert using ActiLife for predicting sleep quality, specifically, sleep efficiency. Our model a) automated the activity recognition, and b) improved the current state-of-the-art results, on average, by  15% in terms of AU-ROC and  30% in terms of F1 scores across different models. Automating the human activity recognition puts sleep science evaluation in the hands of wearable device users. This empowers users to self-monitor their sleep-wake habits, and take action to improve the quality of their life. The improved results demonstrate the robustness of RAHAR as well as the capabilities of implementing the algorithm within clinical software such as ActiLife. The application of RAHAR is, however, not limited to sleep science. It can be used to monitor physical activity levels and patterns of individuals with other health issues such as obesity, diabetes, and cardiac diseases. Besides, RAHAR can also be used in the general context of the "quantified self\" movement, and provide individuals actionable information about their overall fitness levels. [^1]: http://actigraphcorp.com/support/activity-monitors/gt3xplus/ [^2]: http://actigraphcorp.com/products-showcase/software/actilife/ [^3]: http://actigraphcorp.com/wp-content/uploads/2015/06/ActiGraph-White-Paper_What-is-a-Count\_.pdf [^4]: Even though we included logistic regression (LogR) in our experiments, it is important to note that LogR model failed to stratify the dataset successfully for the state-of-the-art baseline approach, and predicted all cases to be in a single class. Therefore, we excluded LogR score of RAHAR from analysis whenever corresponding LogR score of the state-of-the-art baseline approach was not available.

{'timestamp': '2016-07-20T02:06:21', 'yymm': '1607', 'arxiv_id': '1607.04867', 'language': 'en', 'url': 'https://arxiv.org/abs/1607.04867'}